2 * SPDX-License-Identifier: BSD-4-Clause
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
10 * Copyright (c) 2003 Peter Wemm
11 * All rights reserved.
12 * Copyright (c) 2005-2010 Alan L. Cox <alc@cs.rice.edu>
13 * All rights reserved.
15 * This code is derived from software contributed to Berkeley by
16 * the Systems Programming Group of the University of Utah Computer
17 * Science Department and William Jolitz of UUNET Technologies Inc.
19 * Redistribution and use in source and binary forms, with or without
20 * modification, are permitted provided that the following conditions
22 * 1. Redistributions of source code must retain the above copyright
23 * notice, this list of conditions and the following disclaimer.
24 * 2. Redistributions in binary form must reproduce the above copyright
25 * notice, this list of conditions and the following disclaimer in the
26 * documentation and/or other materials provided with the distribution.
27 * 3. All advertising materials mentioning features or use of this software
28 * must display the following acknowledgement:
29 * This product includes software developed by the University of
30 * California, Berkeley and its contributors.
31 * 4. Neither the name of the University nor the names of its contributors
32 * may be used to endorse or promote products derived from this software
33 * without specific prior written permission.
35 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
36 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
37 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
38 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
39 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
40 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
41 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
42 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
43 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
44 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
47 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
50 * Copyright (c) 2003 Networks Associates Technology, Inc.
51 * Copyright (c) 2014-2019 The FreeBSD Foundation
52 * All rights reserved.
54 * This software was developed for the FreeBSD Project by Jake Burkholder,
55 * Safeport Network Services, and Network Associates Laboratories, the
56 * Security Research Division of Network Associates, Inc. under
57 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA
58 * CHATS research program.
60 * Portions of this software were developed by
61 * Konstantin Belousov <kib@FreeBSD.org> under sponsorship from
62 * the FreeBSD Foundation.
64 * Redistribution and use in source and binary forms, with or without
65 * modification, are permitted provided that the following conditions
67 * 1. Redistributions of source code must retain the above copyright
68 * notice, this list of conditions and the following disclaimer.
69 * 2. Redistributions in binary form must reproduce the above copyright
70 * notice, this list of conditions and the following disclaimer in the
71 * documentation and/or other materials provided with the distribution.
73 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
74 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
75 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
76 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
77 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
78 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
79 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
80 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
81 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
82 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
86 #define AMD64_NPT_AWARE
88 #include <sys/cdefs.h>
89 __FBSDID("$FreeBSD$");
92 * Manages physical address maps.
94 * Since the information managed by this module is
95 * also stored by the logical address mapping module,
96 * this module may throw away valid virtual-to-physical
97 * mappings at almost any time. However, invalidations
98 * of virtual-to-physical mappings must be done as
101 * In order to cope with hardware architectures which
102 * make virtual-to-physical map invalidates expensive,
103 * this module may delay invalidate or reduced protection
104 * operations until such time as they are actually
105 * necessary. This module is given full information as
106 * to which processors are currently using which maps,
107 * and to when physical maps must be made correct.
111 #include "opt_pmap.h"
114 #include <sys/param.h>
115 #include <sys/bitstring.h>
117 #include <sys/systm.h>
118 #include <sys/kernel.h>
120 #include <sys/lock.h>
121 #include <sys/malloc.h>
122 #include <sys/mman.h>
123 #include <sys/mutex.h>
124 #include <sys/proc.h>
125 #include <sys/rangeset.h>
126 #include <sys/rwlock.h>
127 #include <sys/sbuf.h>
129 #include <sys/turnstile.h>
130 #include <sys/vmem.h>
131 #include <sys/vmmeter.h>
132 #include <sys/sched.h>
133 #include <sys/sysctl.h>
141 #include <vm/vm_param.h>
142 #include <vm/vm_kern.h>
143 #include <vm/vm_page.h>
144 #include <vm/vm_map.h>
145 #include <vm/vm_object.h>
146 #include <vm/vm_extern.h>
147 #include <vm/vm_pageout.h>
148 #include <vm/vm_pager.h>
149 #include <vm/vm_phys.h>
150 #include <vm/vm_radix.h>
151 #include <vm/vm_reserv.h>
154 #include <machine/intr_machdep.h>
155 #include <x86/apicvar.h>
156 #include <x86/ifunc.h>
157 #include <machine/cpu.h>
158 #include <machine/cputypes.h>
159 #include <machine/md_var.h>
160 #include <machine/pcb.h>
161 #include <machine/specialreg.h>
163 #include <machine/smp.h>
165 #include <machine/sysarch.h>
166 #include <machine/tss.h>
169 #define PMAP_MEMDOM MAXMEMDOM
171 #define PMAP_MEMDOM 1
174 static __inline boolean_t
175 pmap_type_guest(pmap_t pmap)
178 return ((pmap->pm_type == PT_EPT) || (pmap->pm_type == PT_RVI));
181 static __inline boolean_t
182 pmap_emulate_ad_bits(pmap_t pmap)
185 return ((pmap->pm_flags & PMAP_EMULATE_AD_BITS) != 0);
188 static __inline pt_entry_t
189 pmap_valid_bit(pmap_t pmap)
193 switch (pmap->pm_type) {
199 if (pmap_emulate_ad_bits(pmap))
200 mask = EPT_PG_EMUL_V;
205 panic("pmap_valid_bit: invalid pm_type %d", pmap->pm_type);
211 static __inline pt_entry_t
212 pmap_rw_bit(pmap_t pmap)
216 switch (pmap->pm_type) {
222 if (pmap_emulate_ad_bits(pmap))
223 mask = EPT_PG_EMUL_RW;
228 panic("pmap_rw_bit: invalid pm_type %d", pmap->pm_type);
234 static pt_entry_t pg_g;
236 static __inline pt_entry_t
237 pmap_global_bit(pmap_t pmap)
241 switch (pmap->pm_type) {
250 panic("pmap_global_bit: invalid pm_type %d", pmap->pm_type);
256 static __inline pt_entry_t
257 pmap_accessed_bit(pmap_t pmap)
261 switch (pmap->pm_type) {
267 if (pmap_emulate_ad_bits(pmap))
273 panic("pmap_accessed_bit: invalid pm_type %d", pmap->pm_type);
279 static __inline pt_entry_t
280 pmap_modified_bit(pmap_t pmap)
284 switch (pmap->pm_type) {
290 if (pmap_emulate_ad_bits(pmap))
296 panic("pmap_modified_bit: invalid pm_type %d", pmap->pm_type);
302 static __inline pt_entry_t
303 pmap_pku_mask_bit(pmap_t pmap)
306 return (pmap->pm_type == PT_X86 ? X86_PG_PKU_MASK : 0);
309 #if !defined(DIAGNOSTIC)
310 #ifdef __GNUC_GNU_INLINE__
311 #define PMAP_INLINE __attribute__((__gnu_inline__)) inline
313 #define PMAP_INLINE extern inline
320 #define PV_STAT(x) do { x ; } while (0)
322 #define PV_STAT(x) do { } while (0)
327 #define pa_index(pa) ({ \
328 KASSERT((pa) <= vm_phys_segs[vm_phys_nsegs - 1].end, \
329 ("address %lx beyond the last segment", (pa))); \
332 #define pa_to_pmdp(pa) (&pv_table[pa_index(pa)])
333 #define pa_to_pvh(pa) (&(pa_to_pmdp(pa)->pv_page))
334 #define PHYS_TO_PV_LIST_LOCK(pa) ({ \
335 struct rwlock *_lock; \
336 if (__predict_false((pa) > pmap_last_pa)) \
337 _lock = &pv_dummy_large.pv_lock; \
339 _lock = &(pa_to_pmdp(pa)->pv_lock); \
343 #define pa_index(pa) ((pa) >> PDRSHIFT)
344 #define pa_to_pvh(pa) (&pv_table[pa_index(pa)])
346 #define NPV_LIST_LOCKS MAXCPU
348 #define PHYS_TO_PV_LIST_LOCK(pa) \
349 (&pv_list_locks[pa_index(pa) % NPV_LIST_LOCKS])
352 #define CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa) do { \
353 struct rwlock **_lockp = (lockp); \
354 struct rwlock *_new_lock; \
356 _new_lock = PHYS_TO_PV_LIST_LOCK(pa); \
357 if (_new_lock != *_lockp) { \
358 if (*_lockp != NULL) \
359 rw_wunlock(*_lockp); \
360 *_lockp = _new_lock; \
365 #define CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m) \
366 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, VM_PAGE_TO_PHYS(m))
368 #define RELEASE_PV_LIST_LOCK(lockp) do { \
369 struct rwlock **_lockp = (lockp); \
371 if (*_lockp != NULL) { \
372 rw_wunlock(*_lockp); \
377 #define VM_PAGE_TO_PV_LIST_LOCK(m) \
378 PHYS_TO_PV_LIST_LOCK(VM_PAGE_TO_PHYS(m))
380 struct pmap kernel_pmap_store;
382 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
383 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
386 SYSCTL_INT(_machdep, OID_AUTO, nkpt, CTLFLAG_RD, &nkpt, 0,
387 "Number of kernel page table pages allocated on bootup");
390 vm_paddr_t dmaplimit;
391 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
394 static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
395 "VM/pmap parameters");
397 static int pg_ps_enabled = 1;
398 SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
399 &pg_ps_enabled, 0, "Are large page mappings enabled?");
401 #define PAT_INDEX_SIZE 8
402 static int pat_index[PAT_INDEX_SIZE]; /* cache mode to PAT index conversion */
404 static u_int64_t KPTphys; /* phys addr of kernel level 1 */
405 static u_int64_t KPDphys; /* phys addr of kernel level 2 */
406 u_int64_t KPDPphys; /* phys addr of kernel level 3 */
407 u_int64_t KPML4phys; /* phys addr of kernel level 4 */
409 static u_int64_t DMPDphys; /* phys addr of direct mapped level 2 */
410 static u_int64_t DMPDPphys; /* phys addr of direct mapped level 3 */
411 static int ndmpdpphys; /* number of DMPDPphys pages */
413 static vm_paddr_t KERNend; /* phys addr of end of bootstrap data */
416 * pmap_mapdev support pre initialization (i.e. console)
418 #define PMAP_PREINIT_MAPPING_COUNT 8
419 static struct pmap_preinit_mapping {
424 } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT];
425 static int pmap_initialized;
428 * Data for the pv entry allocation mechanism.
429 * Updates to pv_invl_gen are protected by the pv list lock but reads are not.
433 pc_to_domain(struct pv_chunk *pc)
436 return (_vm_phys_domain(DMAP_TO_PHYS((vm_offset_t)pc)));
440 pc_to_domain(struct pv_chunk *pc __unused)
447 struct pv_chunks_list {
449 TAILQ_HEAD(pch, pv_chunk) pvc_list;
451 } __aligned(CACHE_LINE_SIZE);
453 struct pv_chunks_list __exclusive_cache_line pv_chunks[PMAP_MEMDOM];
456 struct pmap_large_md_page {
457 struct rwlock pv_lock;
458 struct md_page pv_page;
461 __exclusive_cache_line static struct pmap_large_md_page pv_dummy_large;
462 #define pv_dummy pv_dummy_large.pv_page
463 __read_mostly static struct pmap_large_md_page *pv_table;
464 __read_mostly vm_paddr_t pmap_last_pa;
466 static struct rwlock __exclusive_cache_line pv_list_locks[NPV_LIST_LOCKS];
467 static u_long pv_invl_gen[NPV_LIST_LOCKS];
468 static struct md_page *pv_table;
469 static struct md_page pv_dummy;
473 * All those kernel PT submaps that BSD is so fond of
475 pt_entry_t *CMAP1 = NULL;
477 static vm_offset_t qframe = 0;
478 static struct mtx qframe_mtx;
480 static int pmap_flags = PMAP_PDE_SUPERPAGE; /* flags for x86 pmaps */
482 static vmem_t *large_vmem;
483 static u_int lm_ents;
484 #define PMAP_ADDRESS_IN_LARGEMAP(va) ((va) >= LARGEMAP_MIN_ADDRESS && \
485 (va) < LARGEMAP_MIN_ADDRESS + NBPML4 * (u_long)lm_ents)
487 int pmap_pcid_enabled = 1;
488 SYSCTL_INT(_vm_pmap, OID_AUTO, pcid_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
489 &pmap_pcid_enabled, 0, "Is TLB Context ID enabled ?");
490 int invpcid_works = 0;
491 SYSCTL_INT(_vm_pmap, OID_AUTO, invpcid_works, CTLFLAG_RD, &invpcid_works, 0,
492 "Is the invpcid instruction available ?");
494 int __read_frequently pti = 0;
495 SYSCTL_INT(_vm_pmap, OID_AUTO, pti, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
497 "Page Table Isolation enabled");
498 static vm_object_t pti_obj;
499 static pml4_entry_t *pti_pml4;
500 static vm_pindex_t pti_pg_idx;
501 static bool pti_finalized;
503 struct pmap_pkru_range {
504 struct rs_el pkru_rs_el;
509 static uma_zone_t pmap_pkru_ranges_zone;
510 static bool pmap_pkru_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva);
511 static pt_entry_t pmap_pkru_get(pmap_t pmap, vm_offset_t va);
512 static void pmap_pkru_on_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva);
513 static void *pkru_dup_range(void *ctx, void *data);
514 static void pkru_free_range(void *ctx, void *node);
515 static int pmap_pkru_copy(pmap_t dst_pmap, pmap_t src_pmap);
516 static int pmap_pkru_deassign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva);
517 static void pmap_pkru_deassign_all(pmap_t pmap);
520 pmap_pcid_save_cnt_proc(SYSCTL_HANDLER_ARGS)
527 res += cpuid_to_pcpu[i]->pc_pm_save_cnt;
529 return (sysctl_handle_64(oidp, &res, 0, req));
531 SYSCTL_PROC(_vm_pmap, OID_AUTO, pcid_save_cnt, CTLTYPE_U64 | CTLFLAG_RD |
532 CTLFLAG_MPSAFE, NULL, 0, pmap_pcid_save_cnt_proc, "QU",
533 "Count of saved TLB context on switch");
535 static LIST_HEAD(, pmap_invl_gen) pmap_invl_gen_tracker =
536 LIST_HEAD_INITIALIZER(&pmap_invl_gen_tracker);
537 static struct mtx invl_gen_mtx;
538 /* Fake lock object to satisfy turnstiles interface. */
539 static struct lock_object invl_gen_ts = {
542 static struct pmap_invl_gen pmap_invl_gen_head = {
546 static u_long pmap_invl_gen = 1;
547 static int pmap_invl_waiters;
548 static struct callout pmap_invl_callout;
549 static bool pmap_invl_callout_inited;
551 #define PMAP_ASSERT_NOT_IN_DI() \
552 KASSERT(pmap_not_in_di(), ("DI already started"))
559 if ((cpu_feature2 & CPUID2_CX16) == 0)
562 TUNABLE_INT_FETCH("vm.pmap.di_locked", &tun);
567 sysctl_pmap_di_locked(SYSCTL_HANDLER_ARGS)
571 locked = pmap_di_locked();
572 return (sysctl_handle_int(oidp, &locked, 0, req));
574 SYSCTL_PROC(_vm_pmap, OID_AUTO, di_locked, CTLTYPE_INT | CTLFLAG_RDTUN |
575 CTLFLAG_MPSAFE, 0, 0, sysctl_pmap_di_locked, "",
576 "Locked delayed invalidation");
578 static bool pmap_not_in_di_l(void);
579 static bool pmap_not_in_di_u(void);
580 DEFINE_IFUNC(, bool, pmap_not_in_di, (void))
583 return (pmap_di_locked() ? pmap_not_in_di_l : pmap_not_in_di_u);
587 pmap_not_in_di_l(void)
589 struct pmap_invl_gen *invl_gen;
591 invl_gen = &curthread->td_md.md_invl_gen;
592 return (invl_gen->gen == 0);
596 pmap_thread_init_invl_gen_l(struct thread *td)
598 struct pmap_invl_gen *invl_gen;
600 invl_gen = &td->td_md.md_invl_gen;
605 pmap_delayed_invl_wait_block(u_long *m_gen, u_long *invl_gen)
607 struct turnstile *ts;
609 ts = turnstile_trywait(&invl_gen_ts);
610 if (*m_gen > atomic_load_long(invl_gen))
611 turnstile_wait(ts, NULL, TS_SHARED_QUEUE);
613 turnstile_cancel(ts);
617 pmap_delayed_invl_finish_unblock(u_long new_gen)
619 struct turnstile *ts;
621 turnstile_chain_lock(&invl_gen_ts);
622 ts = turnstile_lookup(&invl_gen_ts);
624 pmap_invl_gen = new_gen;
626 turnstile_broadcast(ts, TS_SHARED_QUEUE);
627 turnstile_unpend(ts);
629 turnstile_chain_unlock(&invl_gen_ts);
633 * Start a new Delayed Invalidation (DI) block of code, executed by
634 * the current thread. Within a DI block, the current thread may
635 * destroy both the page table and PV list entries for a mapping and
636 * then release the corresponding PV list lock before ensuring that
637 * the mapping is flushed from the TLBs of any processors with the
641 pmap_delayed_invl_start_l(void)
643 struct pmap_invl_gen *invl_gen;
646 invl_gen = &curthread->td_md.md_invl_gen;
647 PMAP_ASSERT_NOT_IN_DI();
648 mtx_lock(&invl_gen_mtx);
649 if (LIST_EMPTY(&pmap_invl_gen_tracker))
650 currgen = pmap_invl_gen;
652 currgen = LIST_FIRST(&pmap_invl_gen_tracker)->gen;
653 invl_gen->gen = currgen + 1;
654 LIST_INSERT_HEAD(&pmap_invl_gen_tracker, invl_gen, link);
655 mtx_unlock(&invl_gen_mtx);
659 * Finish the DI block, previously started by the current thread. All
660 * required TLB flushes for the pages marked by
661 * pmap_delayed_invl_page() must be finished before this function is
664 * This function works by bumping the global DI generation number to
665 * the generation number of the current thread's DI, unless there is a
666 * pending DI that started earlier. In the latter case, bumping the
667 * global DI generation number would incorrectly signal that the
668 * earlier DI had finished. Instead, this function bumps the earlier
669 * DI's generation number to match the generation number of the
670 * current thread's DI.
673 pmap_delayed_invl_finish_l(void)
675 struct pmap_invl_gen *invl_gen, *next;
677 invl_gen = &curthread->td_md.md_invl_gen;
678 KASSERT(invl_gen->gen != 0, ("missed invl_start"));
679 mtx_lock(&invl_gen_mtx);
680 next = LIST_NEXT(invl_gen, link);
682 pmap_delayed_invl_finish_unblock(invl_gen->gen);
684 next->gen = invl_gen->gen;
685 LIST_REMOVE(invl_gen, link);
686 mtx_unlock(&invl_gen_mtx);
691 pmap_not_in_di_u(void)
693 struct pmap_invl_gen *invl_gen;
695 invl_gen = &curthread->td_md.md_invl_gen;
696 return (((uintptr_t)invl_gen->next & PMAP_INVL_GEN_NEXT_INVALID) != 0);
700 pmap_thread_init_invl_gen_u(struct thread *td)
702 struct pmap_invl_gen *invl_gen;
704 invl_gen = &td->td_md.md_invl_gen;
706 invl_gen->next = (void *)PMAP_INVL_GEN_NEXT_INVALID;
710 pmap_di_load_invl(struct pmap_invl_gen *ptr, struct pmap_invl_gen *out)
712 uint64_t new_high, new_low, old_high, old_low;
715 old_low = new_low = 0;
716 old_high = new_high = (uintptr_t)0;
718 __asm volatile("lock;cmpxchg16b\t%1"
719 : "=@cce" (res), "+m" (*ptr), "+a" (old_low), "+d" (old_high)
720 : "b"(new_low), "c" (new_high)
723 if ((old_high & PMAP_INVL_GEN_NEXT_INVALID) != 0)
726 out->next = (void *)old_high;
729 out->next = (void *)new_high;
735 pmap_di_store_invl(struct pmap_invl_gen *ptr, struct pmap_invl_gen *old_val,
736 struct pmap_invl_gen *new_val)
738 uint64_t new_high, new_low, old_high, old_low;
741 new_low = new_val->gen;
742 new_high = (uintptr_t)new_val->next;
743 old_low = old_val->gen;
744 old_high = (uintptr_t)old_val->next;
746 __asm volatile("lock;cmpxchg16b\t%1"
747 : "=@cce" (res), "+m" (*ptr), "+a" (old_low), "+d" (old_high)
748 : "b"(new_low), "c" (new_high)
754 static long invl_start_restart;
755 SYSCTL_LONG(_vm_pmap, OID_AUTO, invl_start_restart, CTLFLAG_RD,
756 &invl_start_restart, 0,
758 static long invl_finish_restart;
759 SYSCTL_LONG(_vm_pmap, OID_AUTO, invl_finish_restart, CTLFLAG_RD,
760 &invl_finish_restart, 0,
762 static int invl_max_qlen;
763 SYSCTL_INT(_vm_pmap, OID_AUTO, invl_max_qlen, CTLFLAG_RD,
768 #define di_delay locks_delay
771 pmap_delayed_invl_start_u(void)
773 struct pmap_invl_gen *invl_gen, *p, prev, new_prev;
775 struct lock_delay_arg lda;
783 invl_gen = &td->td_md.md_invl_gen;
784 PMAP_ASSERT_NOT_IN_DI();
785 lock_delay_arg_init(&lda, &di_delay);
786 invl_gen->saved_pri = 0;
787 pri = td->td_base_pri;
790 pri = td->td_base_pri;
792 invl_gen->saved_pri = pri;
799 for (p = &pmap_invl_gen_head;; p = prev.next) {
801 prevl = (uintptr_t)atomic_load_ptr(&p->next);
802 if ((prevl & PMAP_INVL_GEN_NEXT_INVALID) != 0) {
803 PV_STAT(atomic_add_long(&invl_start_restart, 1));
809 prev.next = (void *)prevl;
812 if ((ii = invl_max_qlen) < i)
813 atomic_cmpset_int(&invl_max_qlen, ii, i);
816 if (!pmap_di_load_invl(p, &prev) || prev.next != NULL) {
817 PV_STAT(atomic_add_long(&invl_start_restart, 1));
822 new_prev.gen = prev.gen;
823 new_prev.next = invl_gen;
824 invl_gen->gen = prev.gen + 1;
826 /* Formal fence between store to invl->gen and updating *p. */
827 atomic_thread_fence_rel();
830 * After inserting an invl_gen element with invalid bit set,
831 * this thread blocks any other thread trying to enter the
832 * delayed invalidation block. Do not allow to remove us from
833 * the CPU, because it causes starvation for other threads.
838 * ABA for *p is not possible there, since p->gen can only
839 * increase. So if the *p thread finished its di, then
840 * started a new one and got inserted into the list at the
841 * same place, its gen will appear greater than the previously
844 if (!pmap_di_store_invl(p, &prev, &new_prev)) {
846 PV_STAT(atomic_add_long(&invl_start_restart, 1));
852 * There we clear PMAP_INVL_GEN_NEXT_INVALID in
853 * invl_gen->next, allowing other threads to iterate past us.
854 * pmap_di_store_invl() provides fence between the generation
855 * write and the update of next.
857 invl_gen->next = NULL;
862 pmap_delayed_invl_finish_u_crit(struct pmap_invl_gen *invl_gen,
863 struct pmap_invl_gen *p)
865 struct pmap_invl_gen prev, new_prev;
869 * Load invl_gen->gen after setting invl_gen->next
870 * PMAP_INVL_GEN_NEXT_INVALID. This prevents larger
871 * generations to propagate to our invl_gen->gen. Lock prefix
872 * in atomic_set_ptr() worked as seq_cst fence.
874 mygen = atomic_load_long(&invl_gen->gen);
876 if (!pmap_di_load_invl(p, &prev) || prev.next != invl_gen)
879 KASSERT(prev.gen < mygen,
880 ("invalid di gen sequence %lu %lu", prev.gen, mygen));
881 new_prev.gen = mygen;
882 new_prev.next = (void *)((uintptr_t)invl_gen->next &
883 ~PMAP_INVL_GEN_NEXT_INVALID);
885 /* Formal fence between load of prev and storing update to it. */
886 atomic_thread_fence_rel();
888 return (pmap_di_store_invl(p, &prev, &new_prev));
892 pmap_delayed_invl_finish_u(void)
894 struct pmap_invl_gen *invl_gen, *p;
896 struct lock_delay_arg lda;
900 invl_gen = &td->td_md.md_invl_gen;
901 KASSERT(invl_gen->gen != 0, ("missed invl_start: gen 0"));
902 KASSERT(((uintptr_t)invl_gen->next & PMAP_INVL_GEN_NEXT_INVALID) == 0,
903 ("missed invl_start: INVALID"));
904 lock_delay_arg_init(&lda, &di_delay);
907 for (p = &pmap_invl_gen_head; p != NULL; p = (void *)prevl) {
908 prevl = (uintptr_t)atomic_load_ptr(&p->next);
909 if ((prevl & PMAP_INVL_GEN_NEXT_INVALID) != 0) {
910 PV_STAT(atomic_add_long(&invl_finish_restart, 1));
914 if ((void *)prevl == invl_gen)
919 * It is legitimate to not find ourself on the list if a
920 * thread before us finished its DI and started it again.
922 if (__predict_false(p == NULL)) {
923 PV_STAT(atomic_add_long(&invl_finish_restart, 1));
929 atomic_set_ptr((uintptr_t *)&invl_gen->next,
930 PMAP_INVL_GEN_NEXT_INVALID);
931 if (!pmap_delayed_invl_finish_u_crit(invl_gen, p)) {
932 atomic_clear_ptr((uintptr_t *)&invl_gen->next,
933 PMAP_INVL_GEN_NEXT_INVALID);
935 PV_STAT(atomic_add_long(&invl_finish_restart, 1));
940 if (atomic_load_int(&pmap_invl_waiters) > 0)
941 pmap_delayed_invl_finish_unblock(0);
942 if (invl_gen->saved_pri != 0) {
944 sched_prio(td, invl_gen->saved_pri);
950 DB_SHOW_COMMAND(di_queue, pmap_di_queue)
952 struct pmap_invl_gen *p, *pn;
957 for (p = &pmap_invl_gen_head, first = true; p != NULL; p = pn,
959 nextl = (uintptr_t)atomic_load_ptr(&p->next);
960 pn = (void *)(nextl & ~PMAP_INVL_GEN_NEXT_INVALID);
961 td = first ? NULL : __containerof(p, struct thread,
963 db_printf("gen %lu inv %d td %p tid %d\n", p->gen,
964 (nextl & PMAP_INVL_GEN_NEXT_INVALID) != 0, td,
965 td != NULL ? td->td_tid : -1);
971 static long invl_wait;
972 SYSCTL_LONG(_vm_pmap, OID_AUTO, invl_wait, CTLFLAG_RD, &invl_wait, 0,
973 "Number of times DI invalidation blocked pmap_remove_all/write");
974 static long invl_wait_slow;
975 SYSCTL_LONG(_vm_pmap, OID_AUTO, invl_wait_slow, CTLFLAG_RD, &invl_wait_slow, 0,
976 "Number of slow invalidation waits for lockless DI");
981 pmap_delayed_invl_genp(vm_page_t m)
986 pa = VM_PAGE_TO_PHYS(m);
987 if (__predict_false((pa) > pmap_last_pa))
988 gen = &pv_dummy_large.pv_invl_gen;
990 gen = &(pa_to_pmdp(pa)->pv_invl_gen);
996 pmap_delayed_invl_genp(vm_page_t m)
999 return (&pv_invl_gen[pa_index(VM_PAGE_TO_PHYS(m)) % NPV_LIST_LOCKS]);
1004 pmap_delayed_invl_callout_func(void *arg __unused)
1007 if (atomic_load_int(&pmap_invl_waiters) == 0)
1009 pmap_delayed_invl_finish_unblock(0);
1013 pmap_delayed_invl_callout_init(void *arg __unused)
1016 if (pmap_di_locked())
1018 callout_init(&pmap_invl_callout, 1);
1019 pmap_invl_callout_inited = true;
1021 SYSINIT(pmap_di_callout, SI_SUB_CPU + 1, SI_ORDER_ANY,
1022 pmap_delayed_invl_callout_init, NULL);
1025 * Ensure that all currently executing DI blocks, that need to flush
1026 * TLB for the given page m, actually flushed the TLB at the time the
1027 * function returned. If the page m has an empty PV list and we call
1028 * pmap_delayed_invl_wait(), upon its return we know that no CPU has a
1029 * valid mapping for the page m in either its page table or TLB.
1031 * This function works by blocking until the global DI generation
1032 * number catches up with the generation number associated with the
1033 * given page m and its PV list. Since this function's callers
1034 * typically own an object lock and sometimes own a page lock, it
1035 * cannot sleep. Instead, it blocks on a turnstile to relinquish the
1039 pmap_delayed_invl_wait_l(vm_page_t m)
1043 bool accounted = false;
1046 m_gen = pmap_delayed_invl_genp(m);
1047 while (*m_gen > pmap_invl_gen) {
1050 atomic_add_long(&invl_wait, 1);
1054 pmap_delayed_invl_wait_block(m_gen, &pmap_invl_gen);
1059 pmap_delayed_invl_wait_u(vm_page_t m)
1062 struct lock_delay_arg lda;
1066 m_gen = pmap_delayed_invl_genp(m);
1067 lock_delay_arg_init(&lda, &di_delay);
1068 while (*m_gen > atomic_load_long(&pmap_invl_gen_head.gen)) {
1069 if (fast || !pmap_invl_callout_inited) {
1070 PV_STAT(atomic_add_long(&invl_wait, 1));
1075 * The page's invalidation generation number
1076 * is still below the current thread's number.
1077 * Prepare to block so that we do not waste
1078 * CPU cycles or worse, suffer livelock.
1080 * Since it is impossible to block without
1081 * racing with pmap_delayed_invl_finish_u(),
1082 * prepare for the race by incrementing
1083 * pmap_invl_waiters and arming a 1-tick
1084 * callout which will unblock us if we lose
1087 atomic_add_int(&pmap_invl_waiters, 1);
1090 * Re-check the current thread's invalidation
1091 * generation after incrementing
1092 * pmap_invl_waiters, so that there is no race
1093 * with pmap_delayed_invl_finish_u() setting
1094 * the page generation and checking
1095 * pmap_invl_waiters. The only race allowed
1096 * is for a missed unblock, which is handled
1100 atomic_load_long(&pmap_invl_gen_head.gen)) {
1101 callout_reset(&pmap_invl_callout, 1,
1102 pmap_delayed_invl_callout_func, NULL);
1103 PV_STAT(atomic_add_long(&invl_wait_slow, 1));
1104 pmap_delayed_invl_wait_block(m_gen,
1105 &pmap_invl_gen_head.gen);
1107 atomic_add_int(&pmap_invl_waiters, -1);
1112 DEFINE_IFUNC(, void, pmap_thread_init_invl_gen, (struct thread *))
1115 return (pmap_di_locked() ? pmap_thread_init_invl_gen_l :
1116 pmap_thread_init_invl_gen_u);
1119 DEFINE_IFUNC(static, void, pmap_delayed_invl_start, (void))
1122 return (pmap_di_locked() ? pmap_delayed_invl_start_l :
1123 pmap_delayed_invl_start_u);
1126 DEFINE_IFUNC(static, void, pmap_delayed_invl_finish, (void))
1129 return (pmap_di_locked() ? pmap_delayed_invl_finish_l :
1130 pmap_delayed_invl_finish_u);
1133 DEFINE_IFUNC(static, void, pmap_delayed_invl_wait, (vm_page_t))
1136 return (pmap_di_locked() ? pmap_delayed_invl_wait_l :
1137 pmap_delayed_invl_wait_u);
1141 * Mark the page m's PV list as participating in the current thread's
1142 * DI block. Any threads concurrently using m's PV list to remove or
1143 * restrict all mappings to m will wait for the current thread's DI
1144 * block to complete before proceeding.
1146 * The function works by setting the DI generation number for m's PV
1147 * list to at least the DI generation number of the current thread.
1148 * This forces a caller of pmap_delayed_invl_wait() to block until
1149 * current thread calls pmap_delayed_invl_finish().
1152 pmap_delayed_invl_page(vm_page_t m)
1156 rw_assert(VM_PAGE_TO_PV_LIST_LOCK(m), RA_WLOCKED);
1157 gen = curthread->td_md.md_invl_gen.gen;
1160 m_gen = pmap_delayed_invl_genp(m);
1168 static caddr_t crashdumpmap;
1171 * Internal flags for pmap_enter()'s helper functions.
1173 #define PMAP_ENTER_NORECLAIM 0x1000000 /* Don't reclaim PV entries. */
1174 #define PMAP_ENTER_NOREPLACE 0x2000000 /* Don't replace mappings. */
1177 * Internal flags for pmap_mapdev_internal() and
1178 * pmap_change_props_locked().
1180 #define MAPDEV_FLUSHCACHE 0x00000001 /* Flush cache after mapping. */
1181 #define MAPDEV_SETATTR 0x00000002 /* Modify existing attrs. */
1182 #define MAPDEV_ASSERTVALID 0x00000004 /* Assert mapping validity. */
1184 TAILQ_HEAD(pv_chunklist, pv_chunk);
1186 static void free_pv_chunk(struct pv_chunk *pc);
1187 static void free_pv_chunk_batch(struct pv_chunklist *batch);
1188 static void free_pv_entry(pmap_t pmap, pv_entry_t pv);
1189 static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp);
1190 static int popcnt_pc_map_pq(uint64_t *map);
1191 static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp);
1192 static void reserve_pv_entries(pmap_t pmap, int needed,
1193 struct rwlock **lockp);
1194 static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1195 struct rwlock **lockp);
1196 static bool pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde,
1197 u_int flags, struct rwlock **lockp);
1198 #if VM_NRESERVLEVEL > 0
1199 static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1200 struct rwlock **lockp);
1202 static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va);
1203 static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap,
1206 static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte);
1207 static int pmap_change_props_locked(vm_offset_t va, vm_size_t size,
1208 vm_prot_t prot, int mode, int flags);
1209 static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va);
1210 static boolean_t pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde,
1211 vm_offset_t va, struct rwlock **lockp);
1212 static boolean_t pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe,
1214 static bool pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m,
1215 vm_prot_t prot, struct rwlock **lockp);
1216 static int pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde,
1217 u_int flags, vm_page_t m, struct rwlock **lockp);
1218 static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va,
1219 vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp);
1220 static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte);
1221 static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted);
1222 static void pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva,
1224 static void pmap_invalidate_cache_range_all(vm_offset_t sva,
1226 static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va,
1228 static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode);
1229 static vm_page_t pmap_large_map_getptp_unlocked(void);
1230 static vm_paddr_t pmap_large_map_kextract(vm_offset_t va);
1231 #if VM_NRESERVLEVEL > 0
1232 static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va,
1233 struct rwlock **lockp);
1235 static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva,
1237 static void pmap_pte_props(pt_entry_t *pte, u_long bits, u_long mask);
1238 static void pmap_pti_add_kva_locked(vm_offset_t sva, vm_offset_t eva,
1240 static pdp_entry_t *pmap_pti_pdpe(vm_offset_t va);
1241 static pd_entry_t *pmap_pti_pde(vm_offset_t va);
1242 static void pmap_pti_wire_pte(void *pte);
1243 static int pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
1244 struct spglist *free, struct rwlock **lockp);
1245 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva,
1246 pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp);
1247 static vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va);
1248 static void pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
1249 struct spglist *free);
1250 static bool pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
1251 pd_entry_t *pde, struct spglist *free,
1252 struct rwlock **lockp);
1253 static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va,
1254 vm_page_t m, struct rwlock **lockp);
1255 static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
1257 static void pmap_update_pde_invalidate(pmap_t, vm_offset_t va, pd_entry_t pde);
1259 static vm_page_t _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
1260 struct rwlock **lockp);
1261 static pd_entry_t *pmap_alloc_pde(pmap_t pmap, vm_offset_t va, vm_page_t *pdpgp,
1262 struct rwlock **lockp);
1263 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va,
1264 struct rwlock **lockp);
1266 static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m,
1267 struct spglist *free);
1268 static int pmap_unuse_pt(pmap_t, vm_offset_t, pd_entry_t, struct spglist *);
1270 /********************/
1271 /* Inline functions */
1272 /********************/
1274 /* Return a non-clipped PD index for a given VA */
1275 static __inline vm_pindex_t
1276 pmap_pde_pindex(vm_offset_t va)
1278 return (va >> PDRSHIFT);
1282 /* Return a pointer to the PML4 slot that corresponds to a VA */
1283 static __inline pml4_entry_t *
1284 pmap_pml4e(pmap_t pmap, vm_offset_t va)
1287 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
1290 /* Return a pointer to the PDP slot that corresponds to a VA */
1291 static __inline pdp_entry_t *
1292 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
1296 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
1297 return (&pdpe[pmap_pdpe_index(va)]);
1300 /* Return a pointer to the PDP slot that corresponds to a VA */
1301 static __inline pdp_entry_t *
1302 pmap_pdpe(pmap_t pmap, vm_offset_t va)
1304 pml4_entry_t *pml4e;
1307 PG_V = pmap_valid_bit(pmap);
1308 pml4e = pmap_pml4e(pmap, va);
1309 if ((*pml4e & PG_V) == 0)
1311 return (pmap_pml4e_to_pdpe(pml4e, va));
1314 /* Return a pointer to the PD slot that corresponds to a VA */
1315 static __inline pd_entry_t *
1316 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
1320 KASSERT((*pdpe & PG_PS) == 0,
1321 ("%s: pdpe %#lx is a leaf", __func__, *pdpe));
1322 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
1323 return (&pde[pmap_pde_index(va)]);
1326 /* Return a pointer to the PD slot that corresponds to a VA */
1327 static __inline pd_entry_t *
1328 pmap_pde(pmap_t pmap, vm_offset_t va)
1333 PG_V = pmap_valid_bit(pmap);
1334 pdpe = pmap_pdpe(pmap, va);
1335 if (pdpe == NULL || (*pdpe & PG_V) == 0)
1337 return (pmap_pdpe_to_pde(pdpe, va));
1340 /* Return a pointer to the PT slot that corresponds to a VA */
1341 static __inline pt_entry_t *
1342 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
1346 KASSERT((*pde & PG_PS) == 0,
1347 ("%s: pde %#lx is a leaf", __func__, *pde));
1348 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
1349 return (&pte[pmap_pte_index(va)]);
1352 /* Return a pointer to the PT slot that corresponds to a VA */
1353 static __inline pt_entry_t *
1354 pmap_pte(pmap_t pmap, vm_offset_t va)
1359 PG_V = pmap_valid_bit(pmap);
1360 pde = pmap_pde(pmap, va);
1361 if (pde == NULL || (*pde & PG_V) == 0)
1363 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
1364 return ((pt_entry_t *)pde);
1365 return (pmap_pde_to_pte(pde, va));
1368 static __inline void
1369 pmap_resident_count_inc(pmap_t pmap, int count)
1372 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1373 pmap->pm_stats.resident_count += count;
1376 static __inline void
1377 pmap_resident_count_dec(pmap_t pmap, int count)
1380 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1381 KASSERT(pmap->pm_stats.resident_count >= count,
1382 ("pmap %p resident count underflow %ld %d", pmap,
1383 pmap->pm_stats.resident_count, count));
1384 pmap->pm_stats.resident_count -= count;
1387 PMAP_INLINE pt_entry_t *
1388 vtopte(vm_offset_t va)
1390 u_int64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
1392 KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopte on a uva/gpa 0x%0lx", va));
1394 return (PTmap + ((va >> PAGE_SHIFT) & mask));
1397 static __inline pd_entry_t *
1398 vtopde(vm_offset_t va)
1400 u_int64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
1402 KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopde on a uva/gpa 0x%0lx", va));
1404 return (PDmap + ((va >> PDRSHIFT) & mask));
1408 allocpages(vm_paddr_t *firstaddr, int n)
1413 bzero((void *)ret, n * PAGE_SIZE);
1414 *firstaddr += n * PAGE_SIZE;
1418 CTASSERT(powerof2(NDMPML4E));
1420 /* number of kernel PDP slots */
1421 #define NKPDPE(ptpgs) howmany(ptpgs, NPDEPG)
1424 nkpt_init(vm_paddr_t addr)
1431 pt_pages = howmany(addr, 1 << PDRSHIFT);
1432 pt_pages += NKPDPE(pt_pages);
1435 * Add some slop beyond the bare minimum required for bootstrapping
1438 * This is quite important when allocating KVA for kernel modules.
1439 * The modules are required to be linked in the negative 2GB of
1440 * the address space. If we run out of KVA in this region then
1441 * pmap_growkernel() will need to allocate page table pages to map
1442 * the entire 512GB of KVA space which is an unnecessary tax on
1445 * Secondly, device memory mapped as part of setting up the low-
1446 * level console(s) is taken from KVA, starting at virtual_avail.
1447 * This is because cninit() is called after pmap_bootstrap() but
1448 * before vm_init() and pmap_init(). 20MB for a frame buffer is
1451 pt_pages += 32; /* 64MB additional slop. */
1457 * Returns the proper write/execute permission for a physical page that is
1458 * part of the initial boot allocations.
1460 * If the page has kernel text, it is marked as read-only. If the page has
1461 * kernel read-only data, it is marked as read-only/not-executable. If the
1462 * page has only read-write data, it is marked as read-write/not-executable.
1463 * If the page is below/above the kernel range, it is marked as read-write.
1465 * This function operates on 2M pages, since we map the kernel space that
1468 static inline pt_entry_t
1469 bootaddr_rwx(vm_paddr_t pa)
1473 * The kernel is loaded at a 2MB-aligned address, and memory below that
1474 * need not be executable. The .bss section is padded to a 2MB
1475 * boundary, so memory following the kernel need not be executable
1476 * either. Preloaded kernel modules have their mapping permissions
1477 * fixed up by the linker.
1479 if (pa < trunc_2mpage(btext - KERNBASE) ||
1480 pa >= trunc_2mpage(_end - KERNBASE))
1481 return (X86_PG_RW | pg_nx);
1484 * The linker should ensure that the read-only and read-write
1485 * portions don't share the same 2M page, so this shouldn't
1486 * impact read-only data. However, in any case, any page with
1487 * read-write data needs to be read-write.
1489 if (pa >= trunc_2mpage(brwsection - KERNBASE))
1490 return (X86_PG_RW | pg_nx);
1493 * Mark any 2M page containing kernel text as read-only. Mark
1494 * other pages with read-only data as read-only and not executable.
1495 * (It is likely a small portion of the read-only data section will
1496 * be marked as read-only, but executable. This should be acceptable
1497 * since the read-only protection will keep the data from changing.)
1498 * Note that fixups to the .text section will still work until we
1501 if (pa < round_2mpage(etext - KERNBASE))
1507 create_pagetables(vm_paddr_t *firstaddr)
1509 int i, j, ndm1g, nkpdpe, nkdmpde;
1513 uint64_t DMPDkernphys;
1515 /* Allocate page table pages for the direct map */
1516 ndmpdp = howmany(ptoa(Maxmem), NBPDP);
1517 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
1519 ndmpdpphys = howmany(ndmpdp, NPDPEPG);
1520 if (ndmpdpphys > NDMPML4E) {
1522 * Each NDMPML4E allows 512 GB, so limit to that,
1523 * and then readjust ndmpdp and ndmpdpphys.
1525 printf("NDMPML4E limits system to %d GB\n", NDMPML4E * 512);
1526 Maxmem = atop(NDMPML4E * NBPML4);
1527 ndmpdpphys = NDMPML4E;
1528 ndmpdp = NDMPML4E * NPDEPG;
1530 DMPDPphys = allocpages(firstaddr, ndmpdpphys);
1532 if ((amd_feature & AMDID_PAGE1GB) != 0) {
1534 * Calculate the number of 1G pages that will fully fit in
1537 ndm1g = ptoa(Maxmem) >> PDPSHIFT;
1540 * Allocate 2M pages for the kernel. These will be used in
1541 * place of the first one or more 1G pages from ndm1g.
1543 nkdmpde = howmany((vm_offset_t)(brwsection - KERNBASE), NBPDP);
1544 DMPDkernphys = allocpages(firstaddr, nkdmpde);
1547 DMPDphys = allocpages(firstaddr, ndmpdp - ndm1g);
1548 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
1550 /* Allocate pages */
1551 KPML4phys = allocpages(firstaddr, 1);
1552 KPDPphys = allocpages(firstaddr, NKPML4E);
1555 * Allocate the initial number of kernel page table pages required to
1556 * bootstrap. We defer this until after all memory-size dependent
1557 * allocations are done (e.g. direct map), so that we don't have to
1558 * build in too much slop in our estimate.
1560 * Note that when NKPML4E > 1, we have an empty page underneath
1561 * all but the KPML4I'th one, so we need NKPML4E-1 extra (zeroed)
1562 * pages. (pmap_enter requires a PD page to exist for each KPML4E.)
1564 nkpt_init(*firstaddr);
1565 nkpdpe = NKPDPE(nkpt);
1567 KPTphys = allocpages(firstaddr, nkpt);
1568 KPDphys = allocpages(firstaddr, nkpdpe);
1571 * Connect the zero-filled PT pages to their PD entries. This
1572 * implicitly maps the PT pages at their correct locations within
1575 pd_p = (pd_entry_t *)KPDphys;
1576 for (i = 0; i < nkpt; i++)
1577 pd_p[i] = (KPTphys + ptoa(i)) | X86_PG_RW | X86_PG_V;
1580 * Map from physical address zero to the end of loader preallocated
1581 * memory using 2MB pages. This replaces some of the PD entries
1584 for (i = 0; (i << PDRSHIFT) < KERNend; i++)
1585 /* Preset PG_M and PG_A because demotion expects it. */
1586 pd_p[i] = (i << PDRSHIFT) | X86_PG_V | PG_PS | pg_g |
1587 X86_PG_M | X86_PG_A | bootaddr_rwx(i << PDRSHIFT);
1590 * Because we map the physical blocks in 2M pages, adjust firstaddr
1591 * to record the physical blocks we've actually mapped into kernel
1592 * virtual address space.
1594 if (*firstaddr < round_2mpage(KERNend))
1595 *firstaddr = round_2mpage(KERNend);
1597 /* And connect up the PD to the PDP (leaving room for L4 pages) */
1598 pdp_p = (pdp_entry_t *)(KPDPphys + ptoa(KPML4I - KPML4BASE));
1599 for (i = 0; i < nkpdpe; i++)
1600 pdp_p[i + KPDPI] = (KPDphys + ptoa(i)) | X86_PG_RW | X86_PG_V;
1603 * Now, set up the direct map region using 2MB and/or 1GB pages. If
1604 * the end of physical memory is not aligned to a 1GB page boundary,
1605 * then the residual physical memory is mapped with 2MB pages. Later,
1606 * if pmap_mapdev{_attr}() uses the direct map for non-write-back
1607 * memory, pmap_change_attr() will demote any 2MB or 1GB page mappings
1608 * that are partially used.
1610 pd_p = (pd_entry_t *)DMPDphys;
1611 for (i = NPDEPG * ndm1g, j = 0; i < NPDEPG * ndmpdp; i++, j++) {
1612 pd_p[j] = (vm_paddr_t)i << PDRSHIFT;
1613 /* Preset PG_M and PG_A because demotion expects it. */
1614 pd_p[j] |= X86_PG_RW | X86_PG_V | PG_PS | pg_g |
1615 X86_PG_M | X86_PG_A | pg_nx;
1617 pdp_p = (pdp_entry_t *)DMPDPphys;
1618 for (i = 0; i < ndm1g; i++) {
1619 pdp_p[i] = (vm_paddr_t)i << PDPSHIFT;
1620 /* Preset PG_M and PG_A because demotion expects it. */
1621 pdp_p[i] |= X86_PG_RW | X86_PG_V | PG_PS | pg_g |
1622 X86_PG_M | X86_PG_A | pg_nx;
1624 for (j = 0; i < ndmpdp; i++, j++) {
1625 pdp_p[i] = DMPDphys + ptoa(j);
1626 pdp_p[i] |= X86_PG_RW | X86_PG_V | pg_nx;
1630 * Instead of using a 1G page for the memory containing the kernel,
1631 * use 2M pages with read-only and no-execute permissions. (If using 1G
1632 * pages, this will partially overwrite the PDPEs above.)
1635 pd_p = (pd_entry_t *)DMPDkernphys;
1636 for (i = 0; i < (NPDEPG * nkdmpde); i++)
1637 pd_p[i] = (i << PDRSHIFT) | X86_PG_V | PG_PS | pg_g |
1638 X86_PG_M | X86_PG_A | pg_nx |
1639 bootaddr_rwx(i << PDRSHIFT);
1640 for (i = 0; i < nkdmpde; i++)
1641 pdp_p[i] = (DMPDkernphys + ptoa(i)) | X86_PG_RW |
1645 /* And recursively map PML4 to itself in order to get PTmap */
1646 p4_p = (pml4_entry_t *)KPML4phys;
1647 p4_p[PML4PML4I] = KPML4phys;
1648 p4_p[PML4PML4I] |= X86_PG_RW | X86_PG_V | pg_nx;
1650 /* Connect the Direct Map slot(s) up to the PML4. */
1651 for (i = 0; i < ndmpdpphys; i++) {
1652 p4_p[DMPML4I + i] = DMPDPphys + ptoa(i);
1653 p4_p[DMPML4I + i] |= X86_PG_RW | X86_PG_V | pg_nx;
1656 /* Connect the KVA slots up to the PML4 */
1657 for (i = 0; i < NKPML4E; i++) {
1658 p4_p[KPML4BASE + i] = KPDPphys + ptoa(i);
1659 p4_p[KPML4BASE + i] |= X86_PG_RW | X86_PG_V;
1664 * Bootstrap the system enough to run with virtual memory.
1666 * On amd64 this is called after mapping has already been enabled
1667 * and just syncs the pmap module with what has already been done.
1668 * [We can't call it easily with mapping off since the kernel is not
1669 * mapped with PA == VA, hence we would have to relocate every address
1670 * from the linked base (virtual) address "KERNBASE" to the actual
1671 * (physical) address starting relative to 0]
1674 pmap_bootstrap(vm_paddr_t *firstaddr)
1677 pt_entry_t *pte, *pcpu_pte;
1678 struct region_descriptor r_gdt;
1679 uint64_t cr4, pcpu_phys;
1683 KERNend = *firstaddr;
1684 res = atop(KERNend - (vm_paddr_t)kernphys);
1690 * Create an initial set of page tables to run the kernel in.
1692 create_pagetables(firstaddr);
1694 pcpu_phys = allocpages(firstaddr, MAXCPU);
1697 * Add a physical memory segment (vm_phys_seg) corresponding to the
1698 * preallocated kernel page table pages so that vm_page structures
1699 * representing these pages will be created. The vm_page structures
1700 * are required for promotion of the corresponding kernel virtual
1701 * addresses to superpage mappings.
1703 vm_phys_early_add_seg(KPTphys, KPTphys + ptoa(nkpt));
1706 * Account for the virtual addresses mapped by create_pagetables().
1708 virtual_avail = (vm_offset_t)KERNBASE + round_2mpage(KERNend);
1709 virtual_end = VM_MAX_KERNEL_ADDRESS;
1712 * Enable PG_G global pages, then switch to the kernel page
1713 * table from the bootstrap page table. After the switch, it
1714 * is possible to enable SMEP and SMAP since PG_U bits are
1720 load_cr3(KPML4phys);
1721 if (cpu_stdext_feature & CPUID_STDEXT_SMEP)
1723 if (cpu_stdext_feature & CPUID_STDEXT_SMAP)
1728 * Initialize the kernel pmap (which is statically allocated).
1729 * Count bootstrap data as being resident in case any of this data is
1730 * later unmapped (using pmap_remove()) and freed.
1732 PMAP_LOCK_INIT(kernel_pmap);
1733 kernel_pmap->pm_pml4 = (pdp_entry_t *)PHYS_TO_DMAP(KPML4phys);
1734 kernel_pmap->pm_cr3 = KPML4phys;
1735 kernel_pmap->pm_ucr3 = PMAP_NO_CR3;
1736 CPU_FILL(&kernel_pmap->pm_active); /* don't allow deactivation */
1737 TAILQ_INIT(&kernel_pmap->pm_pvchunk);
1738 kernel_pmap->pm_stats.resident_count = res;
1739 kernel_pmap->pm_flags = pmap_flags;
1742 * Initialize the TLB invalidations generation number lock.
1744 mtx_init(&invl_gen_mtx, "invlgn", NULL, MTX_DEF);
1747 * Reserve some special page table entries/VA space for temporary
1750 #define SYSMAP(c, p, v, n) \
1751 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
1757 * Crashdump maps. The first page is reused as CMAP1 for the
1760 SYSMAP(caddr_t, CMAP1, crashdumpmap, MAXDUMPPGS)
1761 CADDR1 = crashdumpmap;
1763 SYSMAP(struct pcpu *, pcpu_pte, __pcpu, MAXCPU);
1766 for (i = 0; i < MAXCPU; i++) {
1767 pcpu_pte[i] = (pcpu_phys + ptoa(i)) | X86_PG_V | X86_PG_RW |
1768 pg_g | pg_nx | X86_PG_M | X86_PG_A;
1772 * Re-initialize PCPU area for BSP after switching.
1773 * Make hardware use gdt and common_tss from the new PCPU.
1775 STAILQ_INIT(&cpuhead);
1776 wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[0]);
1777 pcpu_init(&__pcpu[0], 0, sizeof(struct pcpu));
1778 amd64_bsp_pcpu_init1(&__pcpu[0]);
1779 amd64_bsp_ist_init(&__pcpu[0]);
1780 __pcpu[0].pc_common_tss.tss_iobase = sizeof(struct amd64tss) +
1782 memcpy(__pcpu[0].pc_gdt, temp_bsp_pcpu.pc_gdt, NGDT *
1783 sizeof(struct user_segment_descriptor));
1784 gdt_segs[GPROC0_SEL].ssd_base = (uintptr_t)&__pcpu[0].pc_common_tss;
1785 ssdtosyssd(&gdt_segs[GPROC0_SEL],
1786 (struct system_segment_descriptor *)&__pcpu[0].pc_gdt[GPROC0_SEL]);
1787 r_gdt.rd_limit = NGDT * sizeof(struct user_segment_descriptor) - 1;
1788 r_gdt.rd_base = (long)__pcpu[0].pc_gdt;
1790 wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[0]);
1791 ltr(GSEL(GPROC0_SEL, SEL_KPL));
1792 __pcpu[0].pc_dynamic = temp_bsp_pcpu.pc_dynamic;
1793 __pcpu[0].pc_acpi_id = temp_bsp_pcpu.pc_acpi_id;
1796 * Initialize the PAT MSR.
1797 * pmap_init_pat() clears and sets CR4_PGE, which, as a
1798 * side-effect, invalidates stale PG_G TLB entries that might
1799 * have been created in our pre-boot environment.
1803 /* Initialize TLB Context Id. */
1804 if (pmap_pcid_enabled) {
1805 for (i = 0; i < MAXCPU; i++) {
1806 kernel_pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN;
1807 kernel_pmap->pm_pcids[i].pm_gen = 1;
1811 * PMAP_PCID_KERN + 1 is used for initialization of
1812 * proc0 pmap. The pmap' pcid state might be used by
1813 * EFIRT entry before first context switch, so it
1814 * needs to be valid.
1816 PCPU_SET(pcid_next, PMAP_PCID_KERN + 2);
1817 PCPU_SET(pcid_gen, 1);
1820 * pcpu area for APs is zeroed during AP startup.
1821 * pc_pcid_next and pc_pcid_gen are initialized by AP
1822 * during pcpu setup.
1824 load_cr4(rcr4() | CR4_PCIDE);
1829 * Setup the PAT MSR.
1838 /* Bail if this CPU doesn't implement PAT. */
1839 if ((cpu_feature & CPUID_PAT) == 0)
1842 /* Set default PAT index table. */
1843 for (i = 0; i < PAT_INDEX_SIZE; i++)
1845 pat_index[PAT_WRITE_BACK] = 0;
1846 pat_index[PAT_WRITE_THROUGH] = 1;
1847 pat_index[PAT_UNCACHEABLE] = 3;
1848 pat_index[PAT_WRITE_COMBINING] = 6;
1849 pat_index[PAT_WRITE_PROTECTED] = 5;
1850 pat_index[PAT_UNCACHED] = 2;
1853 * Initialize default PAT entries.
1854 * Leave the indices 0-3 at the default of WB, WT, UC-, and UC.
1855 * Program 5 and 6 as WP and WC.
1857 * Leave 4 and 7 as WB and UC. Note that a recursive page table
1858 * mapping for a 2M page uses a PAT value with the bit 3 set due
1859 * to its overload with PG_PS.
1861 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |
1862 PAT_VALUE(1, PAT_WRITE_THROUGH) |
1863 PAT_VALUE(2, PAT_UNCACHED) |
1864 PAT_VALUE(3, PAT_UNCACHEABLE) |
1865 PAT_VALUE(4, PAT_WRITE_BACK) |
1866 PAT_VALUE(5, PAT_WRITE_PROTECTED) |
1867 PAT_VALUE(6, PAT_WRITE_COMBINING) |
1868 PAT_VALUE(7, PAT_UNCACHEABLE);
1872 load_cr4(cr4 & ~CR4_PGE);
1874 /* Disable caches (CD = 1, NW = 0). */
1876 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1878 /* Flushes caches and TLBs. */
1882 /* Update PAT and index table. */
1883 wrmsr(MSR_PAT, pat_msr);
1885 /* Flush caches and TLBs again. */
1889 /* Restore caches and PGE. */
1895 * Initialize a vm_page's machine-dependent fields.
1898 pmap_page_init(vm_page_t m)
1901 TAILQ_INIT(&m->md.pv_list);
1902 m->md.pat_mode = PAT_WRITE_BACK;
1905 static int pmap_allow_2m_x_ept;
1906 SYSCTL_INT(_vm_pmap, OID_AUTO, allow_2m_x_ept, CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
1907 &pmap_allow_2m_x_ept, 0,
1908 "Allow executable superpage mappings in EPT");
1911 pmap_allow_2m_x_ept_recalculate(void)
1914 * SKL002, SKL012S. Since the EPT format is only used by
1915 * Intel CPUs, the vendor check is merely a formality.
1917 if (!(cpu_vendor_id != CPU_VENDOR_INTEL ||
1918 (cpu_ia32_arch_caps & IA32_ARCH_CAP_IF_PSCHANGE_MC_NO) != 0 ||
1919 (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1920 (CPUID_TO_MODEL(cpu_id) == 0x26 || /* Atoms */
1921 CPUID_TO_MODEL(cpu_id) == 0x27 ||
1922 CPUID_TO_MODEL(cpu_id) == 0x35 ||
1923 CPUID_TO_MODEL(cpu_id) == 0x36 ||
1924 CPUID_TO_MODEL(cpu_id) == 0x37 ||
1925 CPUID_TO_MODEL(cpu_id) == 0x86 ||
1926 CPUID_TO_MODEL(cpu_id) == 0x1c ||
1927 CPUID_TO_MODEL(cpu_id) == 0x4a ||
1928 CPUID_TO_MODEL(cpu_id) == 0x4c ||
1929 CPUID_TO_MODEL(cpu_id) == 0x4d ||
1930 CPUID_TO_MODEL(cpu_id) == 0x5a ||
1931 CPUID_TO_MODEL(cpu_id) == 0x5c ||
1932 CPUID_TO_MODEL(cpu_id) == 0x5d ||
1933 CPUID_TO_MODEL(cpu_id) == 0x5f ||
1934 CPUID_TO_MODEL(cpu_id) == 0x6e ||
1935 CPUID_TO_MODEL(cpu_id) == 0x7a ||
1936 CPUID_TO_MODEL(cpu_id) == 0x57 || /* Knights */
1937 CPUID_TO_MODEL(cpu_id) == 0x85))))
1938 pmap_allow_2m_x_ept = 1;
1939 TUNABLE_INT_FETCH("hw.allow_2m_x_ept", &pmap_allow_2m_x_ept);
1943 pmap_allow_2m_x_page(pmap_t pmap, bool executable)
1946 return (pmap->pm_type != PT_EPT || !executable ||
1947 !pmap_allow_2m_x_ept);
1952 pmap_init_pv_table(void)
1954 struct pmap_large_md_page *pvd;
1956 long start, end, highest, pv_npg;
1957 int domain, i, j, pages;
1960 * We strongly depend on the size being a power of two, so the assert
1961 * is overzealous. However, should the struct be resized to a
1962 * different power of two, the code below needs to be revisited.
1964 CTASSERT((sizeof(*pvd) == 64));
1967 * Calculate the size of the array.
1969 pmap_last_pa = vm_phys_segs[vm_phys_nsegs - 1].end;
1970 pv_npg = howmany(pmap_last_pa, NBPDR);
1971 s = (vm_size_t)pv_npg * sizeof(struct pmap_large_md_page);
1973 pv_table = (struct pmap_large_md_page *)kva_alloc(s);
1974 if (pv_table == NULL)
1975 panic("%s: kva_alloc failed\n", __func__);
1978 * Iterate physical segments to allocate space for respective pages.
1982 for (i = 0; i < vm_phys_nsegs; i++) {
1983 end = vm_phys_segs[i].end / NBPDR;
1984 domain = vm_phys_segs[i].domain;
1989 start = highest + 1;
1990 pvd = &pv_table[start];
1992 pages = end - start + 1;
1993 s = round_page(pages * sizeof(*pvd));
1994 highest = start + (s / sizeof(*pvd)) - 1;
1996 for (j = 0; j < s; j += PAGE_SIZE) {
1997 vm_page_t m = vm_page_alloc_domain(NULL, 0,
1998 domain, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ);
2000 panic("vm_page_alloc_domain failed for %lx\n", (vm_offset_t)pvd + j);
2001 pmap_qenter((vm_offset_t)pvd + j, &m, 1);
2004 for (j = 0; j < s / sizeof(*pvd); j++) {
2005 rw_init_flags(&pvd->pv_lock, "pmap pv list", RW_NEW);
2006 TAILQ_INIT(&pvd->pv_page.pv_list);
2007 pvd->pv_page.pv_gen = 0;
2008 pvd->pv_page.pat_mode = 0;
2009 pvd->pv_invl_gen = 0;
2013 pvd = &pv_dummy_large;
2014 rw_init_flags(&pvd->pv_lock, "pmap pv list dummy", RW_NEW);
2015 TAILQ_INIT(&pvd->pv_page.pv_list);
2016 pvd->pv_page.pv_gen = 0;
2017 pvd->pv_page.pat_mode = 0;
2018 pvd->pv_invl_gen = 0;
2022 pmap_init_pv_table(void)
2028 * Initialize the pool of pv list locks.
2030 for (i = 0; i < NPV_LIST_LOCKS; i++)
2031 rw_init(&pv_list_locks[i], "pmap pv list");
2034 * Calculate the size of the pv head table for superpages.
2036 pv_npg = howmany(vm_phys_segs[vm_phys_nsegs - 1].end, NBPDR);
2039 * Allocate memory for the pv head table for superpages.
2041 s = (vm_size_t)pv_npg * sizeof(struct md_page);
2043 pv_table = (struct md_page *)kmem_malloc(s, M_WAITOK | M_ZERO);
2044 for (i = 0; i < pv_npg; i++)
2045 TAILQ_INIT(&pv_table[i].pv_list);
2046 TAILQ_INIT(&pv_dummy.pv_list);
2051 * Initialize the pmap module.
2052 * Called by vm_init, to initialize any structures that the pmap
2053 * system needs to map virtual memory.
2058 struct pmap_preinit_mapping *ppim;
2060 int error, i, ret, skz63;
2062 /* L1TF, reserve page @0 unconditionally */
2063 vm_page_blacklist_add(0, bootverbose);
2065 /* Detect bare-metal Skylake Server and Skylake-X. */
2066 if (vm_guest == VM_GUEST_NO && cpu_vendor_id == CPU_VENDOR_INTEL &&
2067 CPUID_TO_FAMILY(cpu_id) == 0x6 && CPUID_TO_MODEL(cpu_id) == 0x55) {
2069 * Skylake-X errata SKZ63. Processor May Hang When
2070 * Executing Code In an HLE Transaction Region between
2071 * 40000000H and 403FFFFFH.
2073 * Mark the pages in the range as preallocated. It
2074 * seems to be impossible to distinguish between
2075 * Skylake Server and Skylake X.
2078 TUNABLE_INT_FETCH("hw.skz63_enable", &skz63);
2081 printf("SKZ63: skipping 4M RAM starting "
2082 "at physical 1G\n");
2083 for (i = 0; i < atop(0x400000); i++) {
2084 ret = vm_page_blacklist_add(0x40000000 +
2086 if (!ret && bootverbose)
2087 printf("page at %#lx already used\n",
2088 0x40000000 + ptoa(i));
2094 pmap_allow_2m_x_ept_recalculate();
2097 * Initialize the vm page array entries for the kernel pmap's
2100 PMAP_LOCK(kernel_pmap);
2101 for (i = 0; i < nkpt; i++) {
2102 mpte = PHYS_TO_VM_PAGE(KPTphys + (i << PAGE_SHIFT));
2103 KASSERT(mpte >= vm_page_array &&
2104 mpte < &vm_page_array[vm_page_array_size],
2105 ("pmap_init: page table page is out of range"));
2106 mpte->pindex = pmap_pde_pindex(KERNBASE) + i;
2107 mpte->phys_addr = KPTphys + (i << PAGE_SHIFT);
2108 mpte->ref_count = 1;
2111 * Collect the page table pages that were replaced by a 2MB
2112 * page in create_pagetables(). They are zero filled.
2114 if (i << PDRSHIFT < KERNend &&
2115 pmap_insert_pt_page(kernel_pmap, mpte, false))
2116 panic("pmap_init: pmap_insert_pt_page failed");
2118 PMAP_UNLOCK(kernel_pmap);
2122 * If the kernel is running on a virtual machine, then it must assume
2123 * that MCA is enabled by the hypervisor. Moreover, the kernel must
2124 * be prepared for the hypervisor changing the vendor and family that
2125 * are reported by CPUID. Consequently, the workaround for AMD Family
2126 * 10h Erratum 383 is enabled if the processor's feature set does not
2127 * include at least one feature that is only supported by older Intel
2128 * or newer AMD processors.
2130 if (vm_guest != VM_GUEST_NO && (cpu_feature & CPUID_SS) == 0 &&
2131 (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI |
2132 CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP |
2134 workaround_erratum383 = 1;
2137 * Are large page mappings enabled?
2139 TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled);
2140 if (pg_ps_enabled) {
2141 KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
2142 ("pmap_init: can't assign to pagesizes[1]"));
2143 pagesizes[1] = NBPDR;
2147 * Initialize pv chunk lists.
2149 for (i = 0; i < PMAP_MEMDOM; i++) {
2150 mtx_init(&pv_chunks[i].pvc_lock, "pmap pv chunk list", NULL, MTX_DEF);
2151 TAILQ_INIT(&pv_chunks[i].pvc_list);
2153 pmap_init_pv_table();
2155 pmap_initialized = 1;
2156 for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
2157 ppim = pmap_preinit_mapping + i;
2160 /* Make the direct map consistent */
2161 if (ppim->pa < dmaplimit && ppim->pa + ppim->sz <= dmaplimit) {
2162 (void)pmap_change_attr(PHYS_TO_DMAP(ppim->pa),
2163 ppim->sz, ppim->mode);
2167 printf("PPIM %u: PA=%#lx, VA=%#lx, size=%#lx, mode=%#x\n", i,
2168 ppim->pa, ppim->va, ppim->sz, ppim->mode);
2171 mtx_init(&qframe_mtx, "qfrmlk", NULL, MTX_SPIN);
2172 error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK,
2173 (vmem_addr_t *)&qframe);
2175 panic("qframe allocation failed");
2178 TUNABLE_INT_FETCH("vm.pmap.large_map_pml4_entries", &lm_ents);
2179 if (lm_ents > LMEPML4I - LMSPML4I + 1)
2180 lm_ents = LMEPML4I - LMSPML4I + 1;
2182 printf("pmap: large map %u PML4 slots (%lu GB)\n",
2183 lm_ents, (u_long)lm_ents * (NBPML4 / 1024 / 1024 / 1024));
2185 large_vmem = vmem_create("large", LARGEMAP_MIN_ADDRESS,
2186 (vmem_size_t)lm_ents * NBPML4, PAGE_SIZE, 0, M_WAITOK);
2187 if (large_vmem == NULL) {
2188 printf("pmap: cannot create large map\n");
2191 for (i = 0; i < lm_ents; i++) {
2192 m = pmap_large_map_getptp_unlocked();
2193 kernel_pmap->pm_pml4[LMSPML4I + i] = X86_PG_V |
2194 X86_PG_RW | X86_PG_A | X86_PG_M | pg_nx |
2200 SYSCTL_UINT(_vm_pmap, OID_AUTO, large_map_pml4_entries,
2201 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &lm_ents, 0,
2202 "Maximum number of PML4 entries for use by large map (tunable). "
2203 "Each entry corresponds to 512GB of address space.");
2205 static SYSCTL_NODE(_vm_pmap, OID_AUTO, pde, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
2206 "2MB page mapping counters");
2208 static u_long pmap_pde_demotions;
2209 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, demotions, CTLFLAG_RD,
2210 &pmap_pde_demotions, 0, "2MB page demotions");
2212 static u_long pmap_pde_mappings;
2213 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, mappings, CTLFLAG_RD,
2214 &pmap_pde_mappings, 0, "2MB page mappings");
2216 static u_long pmap_pde_p_failures;
2217 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, p_failures, CTLFLAG_RD,
2218 &pmap_pde_p_failures, 0, "2MB page promotion failures");
2220 static u_long pmap_pde_promotions;
2221 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, promotions, CTLFLAG_RD,
2222 &pmap_pde_promotions, 0, "2MB page promotions");
2224 static SYSCTL_NODE(_vm_pmap, OID_AUTO, pdpe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
2225 "1GB page mapping counters");
2227 static u_long pmap_pdpe_demotions;
2228 SYSCTL_ULONG(_vm_pmap_pdpe, OID_AUTO, demotions, CTLFLAG_RD,
2229 &pmap_pdpe_demotions, 0, "1GB page demotions");
2231 /***************************************************
2232 * Low level helper routines.....
2233 ***************************************************/
2236 pmap_swap_pat(pmap_t pmap, pt_entry_t entry)
2238 int x86_pat_bits = X86_PG_PTE_PAT | X86_PG_PDE_PAT;
2240 switch (pmap->pm_type) {
2243 /* Verify that both PAT bits are not set at the same time */
2244 KASSERT((entry & x86_pat_bits) != x86_pat_bits,
2245 ("Invalid PAT bits in entry %#lx", entry));
2247 /* Swap the PAT bits if one of them is set */
2248 if ((entry & x86_pat_bits) != 0)
2249 entry ^= x86_pat_bits;
2253 * Nothing to do - the memory attributes are represented
2254 * the same way for regular pages and superpages.
2258 panic("pmap_switch_pat_bits: bad pm_type %d", pmap->pm_type);
2265 pmap_is_valid_memattr(pmap_t pmap __unused, vm_memattr_t mode)
2268 return (mode >= 0 && mode < PAT_INDEX_SIZE &&
2269 pat_index[(int)mode] >= 0);
2273 * Determine the appropriate bits to set in a PTE or PDE for a specified
2277 pmap_cache_bits(pmap_t pmap, int mode, boolean_t is_pde)
2279 int cache_bits, pat_flag, pat_idx;
2281 if (!pmap_is_valid_memattr(pmap, mode))
2282 panic("Unknown caching mode %d\n", mode);
2284 switch (pmap->pm_type) {
2287 /* The PAT bit is different for PTE's and PDE's. */
2288 pat_flag = is_pde ? X86_PG_PDE_PAT : X86_PG_PTE_PAT;
2290 /* Map the caching mode to a PAT index. */
2291 pat_idx = pat_index[mode];
2293 /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */
2296 cache_bits |= pat_flag;
2298 cache_bits |= PG_NC_PCD;
2300 cache_bits |= PG_NC_PWT;
2304 cache_bits = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(mode);
2308 panic("unsupported pmap type %d", pmap->pm_type);
2311 return (cache_bits);
2315 pmap_cache_mask(pmap_t pmap, boolean_t is_pde)
2319 switch (pmap->pm_type) {
2322 mask = is_pde ? X86_PG_PDE_CACHE : X86_PG_PTE_CACHE;
2325 mask = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(0x7);
2328 panic("pmap_cache_mask: invalid pm_type %d", pmap->pm_type);
2335 pmap_pat_index(pmap_t pmap, pt_entry_t pte, bool is_pde)
2337 int pat_flag, pat_idx;
2340 switch (pmap->pm_type) {
2343 /* The PAT bit is different for PTE's and PDE's. */
2344 pat_flag = is_pde ? X86_PG_PDE_PAT : X86_PG_PTE_PAT;
2346 if ((pte & pat_flag) != 0)
2348 if ((pte & PG_NC_PCD) != 0)
2350 if ((pte & PG_NC_PWT) != 0)
2354 if ((pte & EPT_PG_IGNORE_PAT) != 0)
2355 panic("EPT PTE %#lx has no PAT memory type", pte);
2356 pat_idx = (pte & EPT_PG_MEMORY_TYPE(0x7)) >> 3;
2360 /* See pmap_init_pat(). */
2370 pmap_ps_enabled(pmap_t pmap)
2373 return (pg_ps_enabled && (pmap->pm_flags & PMAP_PDE_SUPERPAGE) != 0);
2377 pmap_update_pde_store(pmap_t pmap, pd_entry_t *pde, pd_entry_t newpde)
2380 switch (pmap->pm_type) {
2387 * This is a little bogus since the generation number is
2388 * supposed to be bumped up when a region of the address
2389 * space is invalidated in the page tables.
2391 * In this case the old PDE entry is valid but yet we want
2392 * to make sure that any mappings using the old entry are
2393 * invalidated in the TLB.
2395 * The reason this works as expected is because we rendezvous
2396 * "all" host cpus and force any vcpu context to exit as a
2399 atomic_add_acq_long(&pmap->pm_eptgen, 1);
2402 panic("pmap_update_pde_store: bad pm_type %d", pmap->pm_type);
2404 pde_store(pde, newpde);
2408 * After changing the page size for the specified virtual address in the page
2409 * table, flush the corresponding entries from the processor's TLB. Only the
2410 * calling processor's TLB is affected.
2412 * The calling thread must be pinned to a processor.
2415 pmap_update_pde_invalidate(pmap_t pmap, vm_offset_t va, pd_entry_t newpde)
2419 if (pmap_type_guest(pmap))
2422 KASSERT(pmap->pm_type == PT_X86,
2423 ("pmap_update_pde_invalidate: invalid type %d", pmap->pm_type));
2425 PG_G = pmap_global_bit(pmap);
2427 if ((newpde & PG_PS) == 0)
2428 /* Demotion: flush a specific 2MB page mapping. */
2430 else if ((newpde & PG_G) == 0)
2432 * Promotion: flush every 4KB page mapping from the TLB
2433 * because there are too many to flush individually.
2438 * Promotion: flush every 4KB page mapping from the TLB,
2439 * including any global (PG_G) mappings.
2447 * For SMP, these functions have to use the IPI mechanism for coherence.
2449 * N.B.: Before calling any of the following TLB invalidation functions,
2450 * the calling processor must ensure that all stores updating a non-
2451 * kernel page table are globally performed. Otherwise, another
2452 * processor could cache an old, pre-update entry without being
2453 * invalidated. This can happen one of two ways: (1) The pmap becomes
2454 * active on another processor after its pm_active field is checked by
2455 * one of the following functions but before a store updating the page
2456 * table is globally performed. (2) The pmap becomes active on another
2457 * processor before its pm_active field is checked but due to
2458 * speculative loads one of the following functions stills reads the
2459 * pmap as inactive on the other processor.
2461 * The kernel page table is exempt because its pm_active field is
2462 * immutable. The kernel page table is always active on every
2467 * Interrupt the cpus that are executing in the guest context.
2468 * This will force the vcpu to exit and the cached EPT mappings
2469 * will be invalidated by the host before the next vmresume.
2471 static __inline void
2472 pmap_invalidate_ept(pmap_t pmap)
2477 KASSERT(!CPU_ISSET(curcpu, &pmap->pm_active),
2478 ("pmap_invalidate_ept: absurd pm_active"));
2481 * The TLB mappings associated with a vcpu context are not
2482 * flushed each time a different vcpu is chosen to execute.
2484 * This is in contrast with a process's vtop mappings that
2485 * are flushed from the TLB on each context switch.
2487 * Therefore we need to do more than just a TLB shootdown on
2488 * the active cpus in 'pmap->pm_active'. To do this we keep
2489 * track of the number of invalidations performed on this pmap.
2491 * Each vcpu keeps a cache of this counter and compares it
2492 * just before a vmresume. If the counter is out-of-date an
2493 * invept will be done to flush stale mappings from the TLB.
2495 atomic_add_acq_long(&pmap->pm_eptgen, 1);
2498 * Force the vcpu to exit and trap back into the hypervisor.
2500 ipinum = pmap->pm_flags & PMAP_NESTED_IPIMASK;
2501 ipi_selected(pmap->pm_active, ipinum);
2506 pmap_invalidate_cpu_mask(pmap_t pmap)
2509 return (pmap == kernel_pmap ? all_cpus : pmap->pm_active);
2513 pmap_invalidate_page_pcid(pmap_t pmap, vm_offset_t va,
2514 const bool invpcid_works1)
2516 struct invpcid_descr d;
2517 uint64_t kcr3, ucr3;
2521 cpuid = PCPU_GET(cpuid);
2522 if (pmap == PCPU_GET(curpmap)) {
2523 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
2525 * Because pm_pcid is recalculated on a
2526 * context switch, we must disable switching.
2527 * Otherwise, we might use a stale value
2531 pcid = pmap->pm_pcids[cpuid].pm_pcid;
2532 if (invpcid_works1) {
2533 d.pcid = pcid | PMAP_PCID_USER_PT;
2536 invpcid(&d, INVPCID_ADDR);
2538 kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
2539 ucr3 = pmap->pm_ucr3 | pcid |
2540 PMAP_PCID_USER_PT | CR3_PCID_SAVE;
2541 pmap_pti_pcid_invlpg(ucr3, kcr3, va);
2546 pmap->pm_pcids[cpuid].pm_gen = 0;
2550 pmap->pm_pcids[i].pm_gen = 0;
2554 * The fence is between stores to pm_gen and the read of the
2555 * pm_active mask. We need to ensure that it is impossible
2556 * for us to miss the bit update in pm_active and
2557 * simultaneously observe a non-zero pm_gen in
2558 * pmap_activate_sw(), otherwise TLB update is missed.
2559 * Without the fence, IA32 allows such an outcome. Note that
2560 * pm_active is updated by a locked operation, which provides
2561 * the reciprocal fence.
2563 atomic_thread_fence_seq_cst();
2567 pmap_invalidate_page_pcid_invpcid(pmap_t pmap, vm_offset_t va)
2570 pmap_invalidate_page_pcid(pmap, va, true);
2574 pmap_invalidate_page_pcid_noinvpcid(pmap_t pmap, vm_offset_t va)
2577 pmap_invalidate_page_pcid(pmap, va, false);
2581 pmap_invalidate_page_nopcid(pmap_t pmap, vm_offset_t va)
2585 DEFINE_IFUNC(static, void, pmap_invalidate_page_mode, (pmap_t, vm_offset_t))
2588 if (pmap_pcid_enabled)
2589 return (invpcid_works ? pmap_invalidate_page_pcid_invpcid :
2590 pmap_invalidate_page_pcid_noinvpcid);
2591 return (pmap_invalidate_page_nopcid);
2595 pmap_invalidate_page_curcpu_cb(pmap_t pmap, vm_offset_t va,
2596 vm_offset_t addr2 __unused)
2599 if (pmap == kernel_pmap) {
2602 if (pmap == PCPU_GET(curpmap))
2604 pmap_invalidate_page_mode(pmap, va);
2609 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
2612 if (pmap_type_guest(pmap)) {
2613 pmap_invalidate_ept(pmap);
2617 KASSERT(pmap->pm_type == PT_X86,
2618 ("pmap_invalidate_page: invalid type %d", pmap->pm_type));
2620 smp_masked_invlpg(pmap_invalidate_cpu_mask(pmap), va, pmap,
2621 pmap_invalidate_page_curcpu_cb);
2624 /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */
2625 #define PMAP_INVLPG_THRESHOLD (4 * 1024 * PAGE_SIZE)
2628 pmap_invalidate_range_pcid(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
2629 const bool invpcid_works1)
2631 struct invpcid_descr d;
2632 uint64_t kcr3, ucr3;
2636 cpuid = PCPU_GET(cpuid);
2637 if (pmap == PCPU_GET(curpmap)) {
2638 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
2640 pcid = pmap->pm_pcids[cpuid].pm_pcid;
2641 if (invpcid_works1) {
2642 d.pcid = pcid | PMAP_PCID_USER_PT;
2645 for (; d.addr < eva; d.addr += PAGE_SIZE)
2646 invpcid(&d, INVPCID_ADDR);
2648 kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
2649 ucr3 = pmap->pm_ucr3 | pcid |
2650 PMAP_PCID_USER_PT | CR3_PCID_SAVE;
2651 pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva);
2656 pmap->pm_pcids[cpuid].pm_gen = 0;
2660 pmap->pm_pcids[i].pm_gen = 0;
2662 /* See the comment in pmap_invalidate_page_pcid(). */
2663 atomic_thread_fence_seq_cst();
2667 pmap_invalidate_range_pcid_invpcid(pmap_t pmap, vm_offset_t sva,
2671 pmap_invalidate_range_pcid(pmap, sva, eva, true);
2675 pmap_invalidate_range_pcid_noinvpcid(pmap_t pmap, vm_offset_t sva,
2679 pmap_invalidate_range_pcid(pmap, sva, eva, false);
2683 pmap_invalidate_range_nopcid(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2687 DEFINE_IFUNC(static, void, pmap_invalidate_range_mode, (pmap_t, vm_offset_t,
2691 if (pmap_pcid_enabled)
2692 return (invpcid_works ? pmap_invalidate_range_pcid_invpcid :
2693 pmap_invalidate_range_pcid_noinvpcid);
2694 return (pmap_invalidate_range_nopcid);
2698 pmap_invalidate_range_curcpu_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2702 if (pmap == kernel_pmap) {
2703 for (addr = sva; addr < eva; addr += PAGE_SIZE)
2706 if (pmap == PCPU_GET(curpmap)) {
2707 for (addr = sva; addr < eva; addr += PAGE_SIZE)
2710 pmap_invalidate_range_mode(pmap, sva, eva);
2715 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2718 if (eva - sva >= PMAP_INVLPG_THRESHOLD) {
2719 pmap_invalidate_all(pmap);
2723 if (pmap_type_guest(pmap)) {
2724 pmap_invalidate_ept(pmap);
2728 KASSERT(pmap->pm_type == PT_X86,
2729 ("pmap_invalidate_range: invalid type %d", pmap->pm_type));
2731 smp_masked_invlpg_range(pmap_invalidate_cpu_mask(pmap), sva, eva, pmap,
2732 pmap_invalidate_range_curcpu_cb);
2736 pmap_invalidate_all_pcid(pmap_t pmap, bool invpcid_works1)
2738 struct invpcid_descr d;
2739 uint64_t kcr3, ucr3;
2743 if (pmap == kernel_pmap) {
2744 if (invpcid_works1) {
2745 bzero(&d, sizeof(d));
2746 invpcid(&d, INVPCID_CTXGLOB);
2751 cpuid = PCPU_GET(cpuid);
2752 if (pmap == PCPU_GET(curpmap)) {
2754 pcid = pmap->pm_pcids[cpuid].pm_pcid;
2755 if (invpcid_works1) {
2759 invpcid(&d, INVPCID_CTX);
2760 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
2761 d.pcid |= PMAP_PCID_USER_PT;
2762 invpcid(&d, INVPCID_CTX);
2765 kcr3 = pmap->pm_cr3 | pcid;
2766 ucr3 = pmap->pm_ucr3;
2767 if (ucr3 != PMAP_NO_CR3) {
2768 ucr3 |= pcid | PMAP_PCID_USER_PT;
2769 pmap_pti_pcid_invalidate(ucr3, kcr3);
2776 pmap->pm_pcids[cpuid].pm_gen = 0;
2779 pmap->pm_pcids[i].pm_gen = 0;
2782 /* See the comment in pmap_invalidate_page_pcid(). */
2783 atomic_thread_fence_seq_cst();
2787 pmap_invalidate_all_pcid_invpcid(pmap_t pmap)
2790 pmap_invalidate_all_pcid(pmap, true);
2794 pmap_invalidate_all_pcid_noinvpcid(pmap_t pmap)
2797 pmap_invalidate_all_pcid(pmap, false);
2801 pmap_invalidate_all_nopcid(pmap_t pmap)
2804 if (pmap == kernel_pmap)
2806 else if (pmap == PCPU_GET(curpmap))
2810 DEFINE_IFUNC(static, void, pmap_invalidate_all_mode, (pmap_t))
2813 if (pmap_pcid_enabled)
2814 return (invpcid_works ? pmap_invalidate_all_pcid_invpcid :
2815 pmap_invalidate_all_pcid_noinvpcid);
2816 return (pmap_invalidate_all_nopcid);
2820 pmap_invalidate_all_curcpu_cb(pmap_t pmap, vm_offset_t addr1 __unused,
2821 vm_offset_t addr2 __unused)
2824 pmap_invalidate_all_mode(pmap);
2828 pmap_invalidate_all(pmap_t pmap)
2831 if (pmap_type_guest(pmap)) {
2832 pmap_invalidate_ept(pmap);
2836 KASSERT(pmap->pm_type == PT_X86,
2837 ("pmap_invalidate_all: invalid type %d", pmap->pm_type));
2839 smp_masked_invltlb(pmap_invalidate_cpu_mask(pmap), pmap,
2840 pmap_invalidate_all_curcpu_cb);
2844 pmap_invalidate_cache_curcpu_cb(pmap_t pmap __unused, vm_offset_t va __unused,
2845 vm_offset_t addr2 __unused)
2852 pmap_invalidate_cache(void)
2855 smp_cache_flush(pmap_invalidate_cache_curcpu_cb);
2859 cpuset_t invalidate; /* processors that invalidate their TLB */
2864 u_int store; /* processor that updates the PDE */
2868 pmap_update_pde_action(void *arg)
2870 struct pde_action *act = arg;
2872 if (act->store == PCPU_GET(cpuid))
2873 pmap_update_pde_store(act->pmap, act->pde, act->newpde);
2877 pmap_update_pde_teardown(void *arg)
2879 struct pde_action *act = arg;
2881 if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate))
2882 pmap_update_pde_invalidate(act->pmap, act->va, act->newpde);
2886 * Change the page size for the specified virtual address in a way that
2887 * prevents any possibility of the TLB ever having two entries that map the
2888 * same virtual address using different page sizes. This is the recommended
2889 * workaround for Erratum 383 on AMD Family 10h processors. It prevents a
2890 * machine check exception for a TLB state that is improperly diagnosed as a
2894 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
2896 struct pde_action act;
2897 cpuset_t active, other_cpus;
2901 cpuid = PCPU_GET(cpuid);
2902 other_cpus = all_cpus;
2903 CPU_CLR(cpuid, &other_cpus);
2904 if (pmap == kernel_pmap || pmap_type_guest(pmap))
2907 active = pmap->pm_active;
2909 if (CPU_OVERLAP(&active, &other_cpus)) {
2911 act.invalidate = active;
2915 act.newpde = newpde;
2916 CPU_SET(cpuid, &active);
2917 smp_rendezvous_cpus(active,
2918 smp_no_rendezvous_barrier, pmap_update_pde_action,
2919 pmap_update_pde_teardown, &act);
2921 pmap_update_pde_store(pmap, pde, newpde);
2922 if (CPU_ISSET(cpuid, &active))
2923 pmap_update_pde_invalidate(pmap, va, newpde);
2929 * Normal, non-SMP, invalidation functions.
2932 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
2934 struct invpcid_descr d;
2935 uint64_t kcr3, ucr3;
2938 if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) {
2942 KASSERT(pmap->pm_type == PT_X86,
2943 ("pmap_invalidate_range: unknown type %d", pmap->pm_type));
2945 if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) {
2947 if (pmap == PCPU_GET(curpmap) && pmap_pcid_enabled &&
2948 pmap->pm_ucr3 != PMAP_NO_CR3) {
2950 pcid = pmap->pm_pcids[0].pm_pcid;
2951 if (invpcid_works) {
2952 d.pcid = pcid | PMAP_PCID_USER_PT;
2955 invpcid(&d, INVPCID_ADDR);
2957 kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
2958 ucr3 = pmap->pm_ucr3 | pcid |
2959 PMAP_PCID_USER_PT | CR3_PCID_SAVE;
2960 pmap_pti_pcid_invlpg(ucr3, kcr3, va);
2964 } else if (pmap_pcid_enabled)
2965 pmap->pm_pcids[0].pm_gen = 0;
2969 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2971 struct invpcid_descr d;
2973 uint64_t kcr3, ucr3;
2975 if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) {
2979 KASSERT(pmap->pm_type == PT_X86,
2980 ("pmap_invalidate_range: unknown type %d", pmap->pm_type));
2982 if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) {
2983 for (addr = sva; addr < eva; addr += PAGE_SIZE)
2985 if (pmap == PCPU_GET(curpmap) && pmap_pcid_enabled &&
2986 pmap->pm_ucr3 != PMAP_NO_CR3) {
2988 if (invpcid_works) {
2989 d.pcid = pmap->pm_pcids[0].pm_pcid |
2993 for (; d.addr < eva; d.addr += PAGE_SIZE)
2994 invpcid(&d, INVPCID_ADDR);
2996 kcr3 = pmap->pm_cr3 | pmap->pm_pcids[0].
2997 pm_pcid | CR3_PCID_SAVE;
2998 ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[0].
2999 pm_pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE;
3000 pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva);
3004 } else if (pmap_pcid_enabled) {
3005 pmap->pm_pcids[0].pm_gen = 0;
3010 pmap_invalidate_all(pmap_t pmap)
3012 struct invpcid_descr d;
3013 uint64_t kcr3, ucr3;
3015 if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) {
3019 KASSERT(pmap->pm_type == PT_X86,
3020 ("pmap_invalidate_all: unknown type %d", pmap->pm_type));
3022 if (pmap == kernel_pmap) {
3023 if (pmap_pcid_enabled && invpcid_works) {
3024 bzero(&d, sizeof(d));
3025 invpcid(&d, INVPCID_CTXGLOB);
3029 } else if (pmap == PCPU_GET(curpmap)) {
3030 if (pmap_pcid_enabled) {
3032 if (invpcid_works) {
3033 d.pcid = pmap->pm_pcids[0].pm_pcid;
3036 invpcid(&d, INVPCID_CTX);
3037 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
3038 d.pcid |= PMAP_PCID_USER_PT;
3039 invpcid(&d, INVPCID_CTX);
3042 kcr3 = pmap->pm_cr3 | pmap->pm_pcids[0].pm_pcid;
3043 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
3044 ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[
3045 0].pm_pcid | PMAP_PCID_USER_PT;
3046 pmap_pti_pcid_invalidate(ucr3, kcr3);
3054 } else if (pmap_pcid_enabled) {
3055 pmap->pm_pcids[0].pm_gen = 0;
3060 pmap_invalidate_cache(void)
3067 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
3070 pmap_update_pde_store(pmap, pde, newpde);
3071 if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap))
3072 pmap_update_pde_invalidate(pmap, va, newpde);
3074 pmap->pm_pcids[0].pm_gen = 0;
3079 pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde)
3083 * When the PDE has PG_PROMOTED set, the 2MB page mapping was created
3084 * by a promotion that did not invalidate the 512 4KB page mappings
3085 * that might exist in the TLB. Consequently, at this point, the TLB
3086 * may hold both 4KB and 2MB page mappings for the address range [va,
3087 * va + NBPDR). Therefore, the entire range must be invalidated here.
3088 * In contrast, when PG_PROMOTED is clear, the TLB will not hold any
3089 * 4KB page mappings for the address range [va, va + NBPDR), and so a
3090 * single INVLPG suffices to invalidate the 2MB page mapping from the
3093 if ((pde & PG_PROMOTED) != 0)
3094 pmap_invalidate_range(pmap, va, va + NBPDR - 1);
3096 pmap_invalidate_page(pmap, va);
3099 DEFINE_IFUNC(, void, pmap_invalidate_cache_range,
3100 (vm_offset_t sva, vm_offset_t eva))
3103 if ((cpu_feature & CPUID_SS) != 0)
3104 return (pmap_invalidate_cache_range_selfsnoop);
3105 if ((cpu_feature & CPUID_CLFSH) != 0)
3106 return (pmap_force_invalidate_cache_range);
3107 return (pmap_invalidate_cache_range_all);
3110 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
3113 pmap_invalidate_cache_range_check_align(vm_offset_t sva, vm_offset_t eva)
3116 KASSERT((sva & PAGE_MASK) == 0,
3117 ("pmap_invalidate_cache_range: sva not page-aligned"));
3118 KASSERT((eva & PAGE_MASK) == 0,
3119 ("pmap_invalidate_cache_range: eva not page-aligned"));
3123 pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva, vm_offset_t eva)
3126 pmap_invalidate_cache_range_check_align(sva, eva);
3130 pmap_force_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
3133 sva &= ~(vm_offset_t)(cpu_clflush_line_size - 1);
3136 * XXX: Some CPUs fault, hang, or trash the local APIC
3137 * registers if we use CLFLUSH on the local APIC range. The
3138 * local APIC is always uncached, so we don't need to flush
3139 * for that range anyway.
3141 if (pmap_kextract(sva) == lapic_paddr)
3144 if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0) {
3146 * Do per-cache line flush. Use a locked
3147 * instruction to insure that previous stores are
3148 * included in the write-back. The processor
3149 * propagates flush to other processors in the cache
3152 atomic_thread_fence_seq_cst();
3153 for (; sva < eva; sva += cpu_clflush_line_size)
3155 atomic_thread_fence_seq_cst();
3158 * Writes are ordered by CLFLUSH on Intel CPUs.
3160 if (cpu_vendor_id != CPU_VENDOR_INTEL)
3162 for (; sva < eva; sva += cpu_clflush_line_size)
3164 if (cpu_vendor_id != CPU_VENDOR_INTEL)
3170 pmap_invalidate_cache_range_all(vm_offset_t sva, vm_offset_t eva)
3173 pmap_invalidate_cache_range_check_align(sva, eva);
3174 pmap_invalidate_cache();
3178 * Remove the specified set of pages from the data and instruction caches.
3180 * In contrast to pmap_invalidate_cache_range(), this function does not
3181 * rely on the CPU's self-snoop feature, because it is intended for use
3182 * when moving pages into a different cache domain.
3185 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
3187 vm_offset_t daddr, eva;
3191 useclflushopt = (cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0;
3192 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
3193 ((cpu_feature & CPUID_CLFSH) == 0 && !useclflushopt))
3194 pmap_invalidate_cache();
3197 atomic_thread_fence_seq_cst();
3198 else if (cpu_vendor_id != CPU_VENDOR_INTEL)
3200 for (i = 0; i < count; i++) {
3201 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
3202 eva = daddr + PAGE_SIZE;
3203 for (; daddr < eva; daddr += cpu_clflush_line_size) {
3211 atomic_thread_fence_seq_cst();
3212 else if (cpu_vendor_id != CPU_VENDOR_INTEL)
3218 pmap_flush_cache_range(vm_offset_t sva, vm_offset_t eva)
3221 pmap_invalidate_cache_range_check_align(sva, eva);
3223 if ((cpu_stdext_feature & CPUID_STDEXT_CLWB) == 0) {
3224 pmap_force_invalidate_cache_range(sva, eva);
3228 /* See comment in pmap_force_invalidate_cache_range(). */
3229 if (pmap_kextract(sva) == lapic_paddr)
3232 atomic_thread_fence_seq_cst();
3233 for (; sva < eva; sva += cpu_clflush_line_size)
3235 atomic_thread_fence_seq_cst();
3239 pmap_flush_cache_phys_range(vm_paddr_t spa, vm_paddr_t epa, vm_memattr_t mattr)
3243 int error, pte_bits;
3245 KASSERT((spa & PAGE_MASK) == 0,
3246 ("pmap_flush_cache_phys_range: spa not page-aligned"));
3247 KASSERT((epa & PAGE_MASK) == 0,
3248 ("pmap_flush_cache_phys_range: epa not page-aligned"));
3250 if (spa < dmaplimit) {
3251 pmap_flush_cache_range(PHYS_TO_DMAP(spa), PHYS_TO_DMAP(MIN(
3253 if (dmaplimit >= epa)
3258 pte_bits = pmap_cache_bits(kernel_pmap, mattr, 0) | X86_PG_RW |
3260 error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK,
3262 KASSERT(error == 0, ("vmem_alloc failed: %d", error));
3263 pte = vtopte(vaddr);
3264 for (; spa < epa; spa += PAGE_SIZE) {
3266 pte_store(pte, spa | pte_bits);
3268 /* XXXKIB atomic inside flush_cache_range are excessive */
3269 pmap_flush_cache_range(vaddr, vaddr + PAGE_SIZE);
3272 vmem_free(kernel_arena, vaddr, PAGE_SIZE);
3276 * Routine: pmap_extract
3278 * Extract the physical page address associated
3279 * with the given map/virtual_address pair.
3282 pmap_extract(pmap_t pmap, vm_offset_t va)
3286 pt_entry_t *pte, PG_V;
3290 PG_V = pmap_valid_bit(pmap);
3292 pdpe = pmap_pdpe(pmap, va);
3293 if (pdpe != NULL && (*pdpe & PG_V) != 0) {
3294 if ((*pdpe & PG_PS) != 0)
3295 pa = (*pdpe & PG_PS_FRAME) | (va & PDPMASK);
3297 pde = pmap_pdpe_to_pde(pdpe, va);
3298 if ((*pde & PG_V) != 0) {
3299 if ((*pde & PG_PS) != 0) {
3300 pa = (*pde & PG_PS_FRAME) |
3303 pte = pmap_pde_to_pte(pde, va);
3304 pa = (*pte & PG_FRAME) |
3315 * Routine: pmap_extract_and_hold
3317 * Atomically extract and hold the physical page
3318 * with the given pmap and virtual address pair
3319 * if that mapping permits the given protection.
3322 pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
3324 pd_entry_t pde, *pdep;
3325 pt_entry_t pte, PG_RW, PG_V;
3329 PG_RW = pmap_rw_bit(pmap);
3330 PG_V = pmap_valid_bit(pmap);
3333 pdep = pmap_pde(pmap, va);
3334 if (pdep != NULL && (pde = *pdep)) {
3336 if ((pde & PG_RW) != 0 || (prot & VM_PROT_WRITE) == 0)
3337 m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) |
3340 pte = *pmap_pde_to_pte(pdep, va);
3341 if ((pte & PG_V) != 0 &&
3342 ((pte & PG_RW) != 0 || (prot & VM_PROT_WRITE) == 0))
3343 m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
3345 if (m != NULL && !vm_page_wire_mapped(m))
3353 pmap_kextract(vm_offset_t va)
3358 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
3359 pa = DMAP_TO_PHYS(va);
3360 } else if (PMAP_ADDRESS_IN_LARGEMAP(va)) {
3361 pa = pmap_large_map_kextract(va);
3365 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
3368 * Beware of a concurrent promotion that changes the
3369 * PDE at this point! For example, vtopte() must not
3370 * be used to access the PTE because it would use the
3371 * new PDE. It is, however, safe to use the old PDE
3372 * because the page table page is preserved by the
3375 pa = *pmap_pde_to_pte(&pde, va);
3376 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
3382 /***************************************************
3383 * Low level mapping routines.....
3384 ***************************************************/
3387 * Add a wired page to the kva.
3388 * Note: not SMP coherent.
3391 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
3396 pte_store(pte, pa | X86_PG_RW | X86_PG_V | pg_g | pg_nx);
3399 static __inline void
3400 pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode)
3406 cache_bits = pmap_cache_bits(kernel_pmap, mode, 0);
3407 pte_store(pte, pa | X86_PG_RW | X86_PG_V | pg_g | pg_nx | cache_bits);
3411 * Remove a page from the kernel pagetables.
3412 * Note: not SMP coherent.
3415 pmap_kremove(vm_offset_t va)
3424 * Used to map a range of physical addresses into kernel
3425 * virtual address space.
3427 * The value passed in '*virt' is a suggested virtual address for
3428 * the mapping. Architectures which can support a direct-mapped
3429 * physical to virtual region can return the appropriate address
3430 * within that region, leaving '*virt' unchanged. Other
3431 * architectures should map the pages starting at '*virt' and
3432 * update '*virt' with the first usable address after the mapped
3436 pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot)
3438 return PHYS_TO_DMAP(start);
3443 * Add a list of wired pages to the kva
3444 * this routine is only used for temporary
3445 * kernel mappings that do not need to have
3446 * page modification or references recorded.
3447 * Note that old mappings are simply written
3448 * over. The page *must* be wired.
3449 * Note: SMP coherent. Uses a ranged shootdown IPI.
3452 pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count)
3454 pt_entry_t *endpte, oldpte, pa, *pte;
3460 endpte = pte + count;
3461 while (pte < endpte) {
3463 cache_bits = pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0);
3464 pa = VM_PAGE_TO_PHYS(m) | cache_bits;
3465 if ((*pte & (PG_FRAME | X86_PG_PTE_CACHE)) != pa) {
3467 pte_store(pte, pa | pg_g | pg_nx | X86_PG_RW | X86_PG_V);
3471 if (__predict_false((oldpte & X86_PG_V) != 0))
3472 pmap_invalidate_range(kernel_pmap, sva, sva + count *
3477 * This routine tears out page mappings from the
3478 * kernel -- it is meant only for temporary mappings.
3479 * Note: SMP coherent. Uses a ranged shootdown IPI.
3482 pmap_qremove(vm_offset_t sva, int count)
3487 while (count-- > 0) {
3488 KASSERT(va >= VM_MIN_KERNEL_ADDRESS, ("usermode va %lx", va));
3492 pmap_invalidate_range(kernel_pmap, sva, va);
3495 /***************************************************
3496 * Page table page management routines.....
3497 ***************************************************/
3499 * Schedule the specified unused page table page to be freed. Specifically,
3500 * add the page to the specified list of pages that will be released to the
3501 * physical memory manager after the TLB has been updated.
3503 static __inline void
3504 pmap_add_delayed_free_list(vm_page_t m, struct spglist *free,
3505 boolean_t set_PG_ZERO)
3509 m->flags |= PG_ZERO;
3511 m->flags &= ~PG_ZERO;
3512 SLIST_INSERT_HEAD(free, m, plinks.s.ss);
3516 * Inserts the specified page table page into the specified pmap's collection
3517 * of idle page table pages. Each of a pmap's page table pages is responsible
3518 * for mapping a distinct range of virtual addresses. The pmap's collection is
3519 * ordered by this virtual address range.
3521 * If "promoted" is false, then the page table page "mpte" must be zero filled.
3524 pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted)
3527 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3528 mpte->valid = promoted ? VM_PAGE_BITS_ALL : 0;
3529 return (vm_radix_insert(&pmap->pm_root, mpte));
3533 * Removes the page table page mapping the specified virtual address from the
3534 * specified pmap's collection of idle page table pages, and returns it.
3535 * Otherwise, returns NULL if there is no page table page corresponding to the
3536 * specified virtual address.
3538 static __inline vm_page_t
3539 pmap_remove_pt_page(pmap_t pmap, vm_offset_t va)
3542 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3543 return (vm_radix_remove(&pmap->pm_root, pmap_pde_pindex(va)));
3547 * Decrements a page table page's reference count, which is used to record the
3548 * number of valid page table entries within the page. If the reference count
3549 * drops to zero, then the page table page is unmapped. Returns TRUE if the
3550 * page table page was unmapped and FALSE otherwise.
3552 static inline boolean_t
3553 pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
3557 if (m->ref_count == 0) {
3558 _pmap_unwire_ptp(pmap, va, m, free);
3565 _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
3568 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3570 * unmap the page table page
3572 if (m->pindex >= NUPDE + NUPDPE) {
3575 pml4 = pmap_pml4e(pmap, va);
3577 if (pmap->pm_pml4u != NULL && va <= VM_MAXUSER_ADDRESS) {
3578 pml4 = &pmap->pm_pml4u[pmap_pml4e_index(va)];
3581 } else if (m->pindex >= NUPDE) {
3584 pdp = pmap_pdpe(pmap, va);
3589 pd = pmap_pde(pmap, va);
3592 pmap_resident_count_dec(pmap, 1);
3593 if (m->pindex < NUPDE) {
3594 /* We just released a PT, unhold the matching PD */
3597 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
3598 pmap_unwire_ptp(pmap, va, pdpg, free);
3599 } else if (m->pindex < NUPDE + NUPDPE) {
3600 /* We just released a PD, unhold the matching PDP */
3603 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
3604 pmap_unwire_ptp(pmap, va, pdppg, free);
3608 * Put page on a list so that it is released after
3609 * *ALL* TLB shootdown is done
3611 pmap_add_delayed_free_list(m, free, TRUE);
3615 * After removing a page table entry, this routine is used to
3616 * conditionally free the page, and manage the reference count.
3619 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, pd_entry_t ptepde,
3620 struct spglist *free)
3624 if (va >= VM_MAXUSER_ADDRESS)
3626 KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0"));
3627 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME);
3628 return (pmap_unwire_ptp(pmap, va, mpte, free));
3632 * Release a page table page reference after a failed attempt to create a
3636 pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
3638 struct spglist free;
3641 if (pmap_unwire_ptp(pmap, va, mpte, &free)) {
3643 * Although "va" was never mapped, paging-structure caches
3644 * could nonetheless have entries that refer to the freed
3645 * page table pages. Invalidate those entries.
3647 pmap_invalidate_page(pmap, va);
3648 vm_page_free_pages_toq(&free, true);
3653 pmap_pinit0(pmap_t pmap)
3659 PMAP_LOCK_INIT(pmap);
3660 pmap->pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
3661 pmap->pm_pml4u = NULL;
3662 pmap->pm_cr3 = KPML4phys;
3663 /* hack to keep pmap_pti_pcid_invalidate() alive */
3664 pmap->pm_ucr3 = PMAP_NO_CR3;
3665 pmap->pm_root.rt_root = 0;
3666 CPU_ZERO(&pmap->pm_active);
3667 TAILQ_INIT(&pmap->pm_pvchunk);
3668 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
3669 pmap->pm_flags = pmap_flags;
3671 pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN + 1;
3672 pmap->pm_pcids[i].pm_gen = 1;
3674 pmap_activate_boot(pmap);
3679 p->p_md.md_flags |= P_MD_KPTI;
3682 pmap_thread_init_invl_gen(td);
3684 if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) {
3685 pmap_pkru_ranges_zone = uma_zcreate("pkru ranges",
3686 sizeof(struct pmap_pkru_range), NULL, NULL, NULL, NULL,
3692 pmap_pinit_pml4(vm_page_t pml4pg)
3694 pml4_entry_t *pm_pml4;
3697 pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4pg));
3699 /* Wire in kernel global address entries. */
3700 for (i = 0; i < NKPML4E; i++) {
3701 pm_pml4[KPML4BASE + i] = (KPDPphys + ptoa(i)) | X86_PG_RW |
3704 for (i = 0; i < ndmpdpphys; i++) {
3705 pm_pml4[DMPML4I + i] = (DMPDPphys + ptoa(i)) | X86_PG_RW |
3709 /* install self-referential address mapping entry(s) */
3710 pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pml4pg) | X86_PG_V | X86_PG_RW |
3711 X86_PG_A | X86_PG_M;
3713 /* install large map entries if configured */
3714 for (i = 0; i < lm_ents; i++)
3715 pm_pml4[LMSPML4I + i] = kernel_pmap->pm_pml4[LMSPML4I + i];
3719 pmap_pinit_pml4_pti(vm_page_t pml4pg)
3721 pml4_entry_t *pm_pml4;
3724 pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4pg));
3725 for (i = 0; i < NPML4EPG; i++)
3726 pm_pml4[i] = pti_pml4[i];
3730 * Initialize a preallocated and zeroed pmap structure,
3731 * such as one in a vmspace structure.
3734 pmap_pinit_type(pmap_t pmap, enum pmap_type pm_type, int flags)
3736 vm_page_t pml4pg, pml4pgu;
3737 vm_paddr_t pml4phys;
3741 * allocate the page directory page
3743 pml4pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
3744 VM_ALLOC_WIRED | VM_ALLOC_ZERO | VM_ALLOC_WAITOK);
3746 pml4phys = VM_PAGE_TO_PHYS(pml4pg);
3747 pmap->pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(pml4phys);
3749 pmap->pm_pcids[i].pm_pcid = PMAP_PCID_NONE;
3750 pmap->pm_pcids[i].pm_gen = 0;
3752 pmap->pm_cr3 = PMAP_NO_CR3; /* initialize to an invalid value */
3753 pmap->pm_ucr3 = PMAP_NO_CR3;
3754 pmap->pm_pml4u = NULL;
3756 pmap->pm_type = pm_type;
3757 if ((pml4pg->flags & PG_ZERO) == 0)
3758 pagezero(pmap->pm_pml4);
3761 * Do not install the host kernel mappings in the nested page
3762 * tables. These mappings are meaningless in the guest physical
3764 * Install minimal kernel mappings in PTI case.
3766 if (pm_type == PT_X86) {
3767 pmap->pm_cr3 = pml4phys;
3768 pmap_pinit_pml4(pml4pg);
3769 if ((curproc->p_md.md_flags & P_MD_KPTI) != 0) {
3770 pml4pgu = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
3771 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
3772 pmap->pm_pml4u = (pml4_entry_t *)PHYS_TO_DMAP(
3773 VM_PAGE_TO_PHYS(pml4pgu));
3774 pmap_pinit_pml4_pti(pml4pgu);
3775 pmap->pm_ucr3 = VM_PAGE_TO_PHYS(pml4pgu);
3777 if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) {
3778 rangeset_init(&pmap->pm_pkru, pkru_dup_range,
3779 pkru_free_range, pmap, M_NOWAIT);
3783 pmap->pm_root.rt_root = 0;
3784 CPU_ZERO(&pmap->pm_active);
3785 TAILQ_INIT(&pmap->pm_pvchunk);
3786 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
3787 pmap->pm_flags = flags;
3788 pmap->pm_eptgen = 0;
3794 pmap_pinit(pmap_t pmap)
3797 return (pmap_pinit_type(pmap, PT_X86, pmap_flags));
3801 * This routine is called if the desired page table page does not exist.
3803 * If page table page allocation fails, this routine may sleep before
3804 * returning NULL. It sleeps only if a lock pointer was given.
3806 * Note: If a page allocation fails at page table level two or three,
3807 * one or two pages may be held during the wait, only to be released
3808 * afterwards. This conservative approach is easily argued to avoid
3812 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp)
3814 vm_page_t m, pdppg, pdpg;
3815 pt_entry_t PG_A, PG_M, PG_RW, PG_V;
3817 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3819 PG_A = pmap_accessed_bit(pmap);
3820 PG_M = pmap_modified_bit(pmap);
3821 PG_V = pmap_valid_bit(pmap);
3822 PG_RW = pmap_rw_bit(pmap);
3825 * Allocate a page table page.
3827 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ |
3828 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
3829 if (lockp != NULL) {
3830 RELEASE_PV_LIST_LOCK(lockp);
3832 PMAP_ASSERT_NOT_IN_DI();
3838 * Indicate the need to retry. While waiting, the page table
3839 * page may have been allocated.
3843 if ((m->flags & PG_ZERO) == 0)
3847 * Map the pagetable page into the process address space, if
3848 * it isn't already there.
3851 if (ptepindex >= (NUPDE + NUPDPE)) {
3852 pml4_entry_t *pml4, *pml4u;
3853 vm_pindex_t pml4index;
3855 /* Wire up a new PDPE page */
3856 pml4index = ptepindex - (NUPDE + NUPDPE);
3857 pml4 = &pmap->pm_pml4[pml4index];
3858 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
3859 if (pmap->pm_pml4u != NULL && pml4index < NUPML4E) {
3861 * PTI: Make all user-space mappings in the
3862 * kernel-mode page table no-execute so that
3863 * we detect any programming errors that leave
3864 * the kernel-mode page table active on return
3867 if (pmap->pm_ucr3 != PMAP_NO_CR3)
3870 pml4u = &pmap->pm_pml4u[pml4index];
3871 *pml4u = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V |
3875 } else if (ptepindex >= NUPDE) {
3876 vm_pindex_t pml4index;
3877 vm_pindex_t pdpindex;
3881 /* Wire up a new PDE page */
3882 pdpindex = ptepindex - NUPDE;
3883 pml4index = pdpindex >> NPML4EPGSHIFT;
3885 pml4 = &pmap->pm_pml4[pml4index];
3886 if ((*pml4 & PG_V) == 0) {
3887 /* Have to allocate a new pdp, recurse */
3888 if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml4index,
3890 vm_page_unwire_noq(m);
3891 vm_page_free_zero(m);
3895 /* Add reference to pdp page */
3896 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
3899 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
3901 /* Now find the pdp page */
3902 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
3903 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
3906 vm_pindex_t pml4index;
3907 vm_pindex_t pdpindex;
3912 /* Wire up a new PTE page */
3913 pdpindex = ptepindex >> NPDPEPGSHIFT;
3914 pml4index = pdpindex >> NPML4EPGSHIFT;
3916 /* First, find the pdp and check that its valid. */
3917 pml4 = &pmap->pm_pml4[pml4index];
3918 if ((*pml4 & PG_V) == 0) {
3919 /* Have to allocate a new pd, recurse */
3920 if (_pmap_allocpte(pmap, NUPDE + pdpindex,
3922 vm_page_unwire_noq(m);
3923 vm_page_free_zero(m);
3926 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
3927 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
3929 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
3930 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
3931 if ((*pdp & PG_V) == 0) {
3932 /* Have to allocate a new pd, recurse */
3933 if (_pmap_allocpte(pmap, NUPDE + pdpindex,
3935 vm_page_unwire_noq(m);
3936 vm_page_free_zero(m);
3940 /* Add reference to the pd page */
3941 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
3945 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
3947 /* Now we know where the page directory page is */
3948 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
3949 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
3952 pmap_resident_count_inc(pmap, 1);
3958 pmap_alloc_pde(pmap_t pmap, vm_offset_t va, vm_page_t *pdpgp,
3959 struct rwlock **lockp)
3961 pdp_entry_t *pdpe, PG_V;
3964 vm_pindex_t pdpindex;
3966 PG_V = pmap_valid_bit(pmap);
3969 pdpe = pmap_pdpe(pmap, va);
3970 if (pdpe != NULL && (*pdpe & PG_V) != 0) {
3971 pde = pmap_pdpe_to_pde(pdpe, va);
3972 if (va < VM_MAXUSER_ADDRESS) {
3973 /* Add a reference to the pd page. */
3974 pdpg = PHYS_TO_VM_PAGE(*pdpe & PG_FRAME);
3978 } else if (va < VM_MAXUSER_ADDRESS) {
3979 /* Allocate a pd page. */
3980 pdpindex = pmap_pde_pindex(va) >> NPDPEPGSHIFT;
3981 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex, lockp);
3988 pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pdpg));
3989 pde = &pde[pmap_pde_index(va)];
3991 panic("pmap_alloc_pde: missing page table page for va %#lx",
3998 pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp)
4000 vm_pindex_t ptepindex;
4001 pd_entry_t *pd, PG_V;
4004 PG_V = pmap_valid_bit(pmap);
4007 * Calculate pagetable page index
4009 ptepindex = pmap_pde_pindex(va);
4012 * Get the page directory entry
4014 pd = pmap_pde(pmap, va);
4017 * This supports switching from a 2MB page to a
4020 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
4021 if (!pmap_demote_pde_locked(pmap, pd, va, lockp)) {
4023 * Invalidation of the 2MB page mapping may have caused
4024 * the deallocation of the underlying PD page.
4031 * If the page table page is mapped, we just increment the
4032 * hold count, and activate it.
4034 if (pd != NULL && (*pd & PG_V) != 0) {
4035 m = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
4039 * Here if the pte page isn't mapped, or if it has been
4042 m = _pmap_allocpte(pmap, ptepindex, lockp);
4043 if (m == NULL && lockp != NULL)
4050 /***************************************************
4051 * Pmap allocation/deallocation routines.
4052 ***************************************************/
4055 * Release any resources held by the given physical map.
4056 * Called when a pmap initialized by pmap_pinit is being released.
4057 * Should only be called if the map contains no valid mappings.
4060 pmap_release(pmap_t pmap)
4065 KASSERT(pmap->pm_stats.resident_count == 0,
4066 ("pmap_release: pmap resident count %ld != 0",
4067 pmap->pm_stats.resident_count));
4068 KASSERT(vm_radix_is_empty(&pmap->pm_root),
4069 ("pmap_release: pmap has reserved page table page(s)"));
4070 KASSERT(CPU_EMPTY(&pmap->pm_active),
4071 ("releasing active pmap %p", pmap));
4073 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_pml4));
4075 for (i = 0; i < NKPML4E; i++) /* KVA */
4076 pmap->pm_pml4[KPML4BASE + i] = 0;
4077 for (i = 0; i < ndmpdpphys; i++)/* Direct Map */
4078 pmap->pm_pml4[DMPML4I + i] = 0;
4079 pmap->pm_pml4[PML4PML4I] = 0; /* Recursive Mapping */
4080 for (i = 0; i < lm_ents; i++) /* Large Map */
4081 pmap->pm_pml4[LMSPML4I + i] = 0;
4083 vm_page_unwire_noq(m);
4084 vm_page_free_zero(m);
4086 if (pmap->pm_pml4u != NULL) {
4087 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_pml4u));
4088 vm_page_unwire_noq(m);
4091 if (pmap->pm_type == PT_X86 &&
4092 (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0)
4093 rangeset_fini(&pmap->pm_pkru);
4097 kvm_size(SYSCTL_HANDLER_ARGS)
4099 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
4101 return sysctl_handle_long(oidp, &ksize, 0, req);
4103 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_MPSAFE,
4104 0, 0, kvm_size, "LU",
4108 kvm_free(SYSCTL_HANDLER_ARGS)
4110 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
4112 return sysctl_handle_long(oidp, &kfree, 0, req);
4114 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_MPSAFE,
4115 0, 0, kvm_free, "LU",
4116 "Amount of KVM free");
4119 * Allocate physical memory for the vm_page array and map it into KVA,
4120 * attempting to back the vm_pages with domain-local memory.
4123 pmap_page_array_startup(long pages)
4126 pd_entry_t *pde, newpdir;
4127 vm_offset_t va, start, end;
4132 vm_page_array_size = pages;
4134 start = VM_MIN_KERNEL_ADDRESS;
4135 end = start + pages * sizeof(struct vm_page);
4136 for (va = start; va < end; va += NBPDR) {
4137 pfn = first_page + (va - start) / sizeof(struct vm_page);
4138 domain = _vm_phys_domain(ptoa(pfn));
4139 pdpe = pmap_pdpe(kernel_pmap, va);
4140 if ((*pdpe & X86_PG_V) == 0) {
4141 pa = vm_phys_early_alloc(domain, PAGE_SIZE);
4143 pagezero((void *)PHYS_TO_DMAP(pa));
4144 *pdpe = (pdp_entry_t)(pa | X86_PG_V | X86_PG_RW |
4145 X86_PG_A | X86_PG_M);
4147 pde = pmap_pdpe_to_pde(pdpe, va);
4148 if ((*pde & X86_PG_V) != 0)
4149 panic("Unexpected pde");
4150 pa = vm_phys_early_alloc(domain, NBPDR);
4151 for (i = 0; i < NPDEPG; i++)
4152 dump_add_page(pa + i * PAGE_SIZE);
4153 newpdir = (pd_entry_t)(pa | X86_PG_V | X86_PG_RW | X86_PG_A |
4154 X86_PG_M | PG_PS | pg_g | pg_nx);
4155 pde_store(pde, newpdir);
4157 vm_page_array = (vm_page_t)start;
4161 * grow the number of kernel page table entries, if needed
4164 pmap_growkernel(vm_offset_t addr)
4168 pd_entry_t *pde, newpdir;
4171 mtx_assert(&kernel_map->system_mtx, MA_OWNED);
4174 * Return if "addr" is within the range of kernel page table pages
4175 * that were preallocated during pmap bootstrap. Moreover, leave
4176 * "kernel_vm_end" and the kernel page table as they were.
4178 * The correctness of this action is based on the following
4179 * argument: vm_map_insert() allocates contiguous ranges of the
4180 * kernel virtual address space. It calls this function if a range
4181 * ends after "kernel_vm_end". If the kernel is mapped between
4182 * "kernel_vm_end" and "addr", then the range cannot begin at
4183 * "kernel_vm_end". In fact, its beginning address cannot be less
4184 * than the kernel. Thus, there is no immediate need to allocate
4185 * any new kernel page table pages between "kernel_vm_end" and
4188 if (KERNBASE < addr && addr <= KERNBASE + nkpt * NBPDR)
4191 addr = roundup2(addr, NBPDR);
4192 if (addr - 1 >= vm_map_max(kernel_map))
4193 addr = vm_map_max(kernel_map);
4194 while (kernel_vm_end < addr) {
4195 pdpe = pmap_pdpe(kernel_pmap, kernel_vm_end);
4196 if ((*pdpe & X86_PG_V) == 0) {
4197 /* We need a new PDP entry */
4198 nkpg = vm_page_alloc(NULL, kernel_vm_end >> PDPSHIFT,
4199 VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ |
4200 VM_ALLOC_WIRED | VM_ALLOC_ZERO);
4202 panic("pmap_growkernel: no memory to grow kernel");
4203 if ((nkpg->flags & PG_ZERO) == 0)
4204 pmap_zero_page(nkpg);
4205 paddr = VM_PAGE_TO_PHYS(nkpg);
4206 *pdpe = (pdp_entry_t)(paddr | X86_PG_V | X86_PG_RW |
4207 X86_PG_A | X86_PG_M);
4208 continue; /* try again */
4210 pde = pmap_pdpe_to_pde(pdpe, kernel_vm_end);
4211 if ((*pde & X86_PG_V) != 0) {
4212 kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK;
4213 if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
4214 kernel_vm_end = vm_map_max(kernel_map);
4220 nkpg = vm_page_alloc(NULL, pmap_pde_pindex(kernel_vm_end),
4221 VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
4224 panic("pmap_growkernel: no memory to grow kernel");
4225 if ((nkpg->flags & PG_ZERO) == 0)
4226 pmap_zero_page(nkpg);
4227 paddr = VM_PAGE_TO_PHYS(nkpg);
4228 newpdir = paddr | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M;
4229 pde_store(pde, newpdir);
4231 kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK;
4232 if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
4233 kernel_vm_end = vm_map_max(kernel_map);
4240 /***************************************************
4241 * page management routines.
4242 ***************************************************/
4244 CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE);
4245 CTASSERT(_NPCM == 3);
4246 CTASSERT(_NPCPV == 168);
4248 static __inline struct pv_chunk *
4249 pv_to_chunk(pv_entry_t pv)
4252 return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK));
4255 #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap)
4257 #define PC_FREE0 0xfffffffffffffffful
4258 #define PC_FREE1 0xfffffffffffffffful
4259 #define PC_FREE2 0x000000fffffffffful
4261 static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1, PC_FREE2 };
4264 static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail;
4266 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0,
4267 "Current number of pv entry chunks");
4268 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0,
4269 "Current number of pv entry chunks allocated");
4270 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0,
4271 "Current number of pv entry chunks frees");
4272 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0,
4273 "Number of times tried to get a chunk page but failed.");
4275 static long pv_entry_frees, pv_entry_allocs, pv_entry_count;
4276 static int pv_entry_spare;
4278 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0,
4279 "Current number of pv entry frees");
4280 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0,
4281 "Current number of pv entry allocs");
4282 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0,
4283 "Current number of pv entries");
4284 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0,
4285 "Current number of spare pv entries");
4289 reclaim_pv_chunk_leave_pmap(pmap_t pmap, pmap_t locked_pmap, bool start_di)
4294 pmap_invalidate_all(pmap);
4295 if (pmap != locked_pmap)
4298 pmap_delayed_invl_finish();
4302 * We are in a serious low memory condition. Resort to
4303 * drastic measures to free some pages so we can allocate
4304 * another pv entry chunk.
4306 * Returns NULL if PV entries were reclaimed from the specified pmap.
4308 * We do not, however, unmap 2mpages because subsequent accesses will
4309 * allocate per-page pv entries until repromotion occurs, thereby
4310 * exacerbating the shortage of free pv entries.
4313 reclaim_pv_chunk_domain(pmap_t locked_pmap, struct rwlock **lockp, int domain)
4315 struct pv_chunks_list *pvc;
4316 struct pv_chunk *pc, *pc_marker, *pc_marker_end;
4317 struct pv_chunk_header pc_marker_b, pc_marker_end_b;
4318 struct md_page *pvh;
4320 pmap_t next_pmap, pmap;
4321 pt_entry_t *pte, tpte;
4322 pt_entry_t PG_G, PG_A, PG_M, PG_RW;
4326 struct spglist free;
4328 int bit, field, freed;
4329 bool start_di, restart;
4331 PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED);
4332 KASSERT(lockp != NULL, ("reclaim_pv_chunk: lockp is NULL"));
4335 PG_G = PG_A = PG_M = PG_RW = 0;
4337 bzero(&pc_marker_b, sizeof(pc_marker_b));
4338 bzero(&pc_marker_end_b, sizeof(pc_marker_end_b));
4339 pc_marker = (struct pv_chunk *)&pc_marker_b;
4340 pc_marker_end = (struct pv_chunk *)&pc_marker_end_b;
4343 * A delayed invalidation block should already be active if
4344 * pmap_advise() or pmap_remove() called this function by way
4345 * of pmap_demote_pde_locked().
4347 start_di = pmap_not_in_di();
4349 pvc = &pv_chunks[domain];
4350 mtx_lock(&pvc->pvc_lock);
4351 pvc->active_reclaims++;
4352 TAILQ_INSERT_HEAD(&pvc->pvc_list, pc_marker, pc_lru);
4353 TAILQ_INSERT_TAIL(&pvc->pvc_list, pc_marker_end, pc_lru);
4354 while ((pc = TAILQ_NEXT(pc_marker, pc_lru)) != pc_marker_end &&
4355 SLIST_EMPTY(&free)) {
4356 next_pmap = pc->pc_pmap;
4357 if (next_pmap == NULL) {
4359 * The next chunk is a marker. However, it is
4360 * not our marker, so active_reclaims must be
4361 * > 1. Consequently, the next_chunk code
4362 * will not rotate the pv_chunks list.
4366 mtx_unlock(&pvc->pvc_lock);
4369 * A pv_chunk can only be removed from the pc_lru list
4370 * when both pc_chunks_mutex is owned and the
4371 * corresponding pmap is locked.
4373 if (pmap != next_pmap) {
4375 reclaim_pv_chunk_leave_pmap(pmap, locked_pmap,
4378 /* Avoid deadlock and lock recursion. */
4379 if (pmap > locked_pmap) {
4380 RELEASE_PV_LIST_LOCK(lockp);
4383 pmap_delayed_invl_start();
4384 mtx_lock(&pvc->pvc_lock);
4386 } else if (pmap != locked_pmap) {
4387 if (PMAP_TRYLOCK(pmap)) {
4389 pmap_delayed_invl_start();
4390 mtx_lock(&pvc->pvc_lock);
4393 pmap = NULL; /* pmap is not locked */
4394 mtx_lock(&pvc->pvc_lock);
4395 pc = TAILQ_NEXT(pc_marker, pc_lru);
4397 pc->pc_pmap != next_pmap)
4401 } else if (start_di)
4402 pmap_delayed_invl_start();
4403 PG_G = pmap_global_bit(pmap);
4404 PG_A = pmap_accessed_bit(pmap);
4405 PG_M = pmap_modified_bit(pmap);
4406 PG_RW = pmap_rw_bit(pmap);
4412 * Destroy every non-wired, 4 KB page mapping in the chunk.
4415 for (field = 0; field < _NPCM; field++) {
4416 for (inuse = ~pc->pc_map[field] & pc_freemask[field];
4417 inuse != 0; inuse &= ~(1UL << bit)) {
4419 pv = &pc->pc_pventry[field * 64 + bit];
4421 pde = pmap_pde(pmap, va);
4422 if ((*pde & PG_PS) != 0)
4424 pte = pmap_pde_to_pte(pde, va);
4425 if ((*pte & PG_W) != 0)
4427 tpte = pte_load_clear(pte);
4428 if ((tpte & PG_G) != 0)
4429 pmap_invalidate_page(pmap, va);
4430 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
4431 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
4433 if ((tpte & PG_A) != 0)
4434 vm_page_aflag_set(m, PGA_REFERENCED);
4435 CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
4436 TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
4438 if (TAILQ_EMPTY(&m->md.pv_list) &&
4439 (m->flags & PG_FICTITIOUS) == 0) {
4440 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4441 if (TAILQ_EMPTY(&pvh->pv_list)) {
4442 vm_page_aflag_clear(m,
4446 pmap_delayed_invl_page(m);
4447 pc->pc_map[field] |= 1UL << bit;
4448 pmap_unuse_pt(pmap, va, *pde, &free);
4453 mtx_lock(&pvc->pvc_lock);
4456 /* Every freed mapping is for a 4 KB page. */
4457 pmap_resident_count_dec(pmap, freed);
4458 PV_STAT(atomic_add_long(&pv_entry_frees, freed));
4459 PV_STAT(atomic_add_int(&pv_entry_spare, freed));
4460 PV_STAT(atomic_subtract_long(&pv_entry_count, freed));
4461 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4462 if (pc->pc_map[0] == PC_FREE0 && pc->pc_map[1] == PC_FREE1 &&
4463 pc->pc_map[2] == PC_FREE2) {
4464 PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV));
4465 PV_STAT(atomic_subtract_int(&pc_chunk_count, 1));
4466 PV_STAT(atomic_add_int(&pc_chunk_frees, 1));
4467 /* Entire chunk is free; return it. */
4468 m_pc = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc));
4469 dump_drop_page(m_pc->phys_addr);
4470 mtx_lock(&pvc->pvc_lock);
4471 TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru);
4474 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
4475 mtx_lock(&pvc->pvc_lock);
4476 /* One freed pv entry in locked_pmap is sufficient. */
4477 if (pmap == locked_pmap)
4480 TAILQ_REMOVE(&pvc->pvc_list, pc_marker, pc_lru);
4481 TAILQ_INSERT_AFTER(&pvc->pvc_list, pc, pc_marker, pc_lru);
4482 if (pvc->active_reclaims == 1 && pmap != NULL) {
4484 * Rotate the pv chunks list so that we do not
4485 * scan the same pv chunks that could not be
4486 * freed (because they contained a wired
4487 * and/or superpage mapping) on every
4488 * invocation of reclaim_pv_chunk().
4490 while ((pc = TAILQ_FIRST(&pvc->pvc_list)) != pc_marker) {
4491 MPASS(pc->pc_pmap != NULL);
4492 TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru);
4493 TAILQ_INSERT_TAIL(&pvc->pvc_list, pc, pc_lru);
4497 TAILQ_REMOVE(&pvc->pvc_list, pc_marker, pc_lru);
4498 TAILQ_REMOVE(&pvc->pvc_list, pc_marker_end, pc_lru);
4499 pvc->active_reclaims--;
4500 mtx_unlock(&pvc->pvc_lock);
4501 reclaim_pv_chunk_leave_pmap(pmap, locked_pmap, start_di);
4502 if (m_pc == NULL && !SLIST_EMPTY(&free)) {
4503 m_pc = SLIST_FIRST(&free);
4504 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
4505 /* Recycle a freed page table page. */
4506 m_pc->ref_count = 1;
4508 vm_page_free_pages_toq(&free, true);
4513 reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp)
4518 domain = PCPU_GET(domain);
4519 for (i = 0; i < vm_ndomains; i++) {
4520 m = reclaim_pv_chunk_domain(locked_pmap, lockp, domain);
4523 domain = (domain + 1) % vm_ndomains;
4530 * free the pv_entry back to the free list
4533 free_pv_entry(pmap_t pmap, pv_entry_t pv)
4535 struct pv_chunk *pc;
4536 int idx, field, bit;
4538 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4539 PV_STAT(atomic_add_long(&pv_entry_frees, 1));
4540 PV_STAT(atomic_add_int(&pv_entry_spare, 1));
4541 PV_STAT(atomic_subtract_long(&pv_entry_count, 1));
4542 pc = pv_to_chunk(pv);
4543 idx = pv - &pc->pc_pventry[0];
4546 pc->pc_map[field] |= 1ul << bit;
4547 if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1 ||
4548 pc->pc_map[2] != PC_FREE2) {
4549 /* 98% of the time, pc is already at the head of the list. */
4550 if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) {
4551 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4552 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
4556 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4561 free_pv_chunk_dequeued(struct pv_chunk *pc)
4565 PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV));
4566 PV_STAT(atomic_subtract_int(&pc_chunk_count, 1));
4567 PV_STAT(atomic_add_int(&pc_chunk_frees, 1));
4568 /* entire chunk is free, return it */
4569 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc));
4570 dump_drop_page(m->phys_addr);
4571 vm_page_unwire_noq(m);
4576 free_pv_chunk(struct pv_chunk *pc)
4578 struct pv_chunks_list *pvc;
4580 pvc = &pv_chunks[pc_to_domain(pc)];
4581 mtx_lock(&pvc->pvc_lock);
4582 TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru);
4583 mtx_unlock(&pvc->pvc_lock);
4584 free_pv_chunk_dequeued(pc);
4588 free_pv_chunk_batch(struct pv_chunklist *batch)
4590 struct pv_chunks_list *pvc;
4591 struct pv_chunk *pc, *npc;
4594 for (i = 0; i < vm_ndomains; i++) {
4595 if (TAILQ_EMPTY(&batch[i]))
4597 pvc = &pv_chunks[i];
4598 mtx_lock(&pvc->pvc_lock);
4599 TAILQ_FOREACH(pc, &batch[i], pc_list) {
4600 TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru);
4602 mtx_unlock(&pvc->pvc_lock);
4605 for (i = 0; i < vm_ndomains; i++) {
4606 TAILQ_FOREACH_SAFE(pc, &batch[i], pc_list, npc) {
4607 free_pv_chunk_dequeued(pc);
4613 * Returns a new PV entry, allocating a new PV chunk from the system when
4614 * needed. If this PV chunk allocation fails and a PV list lock pointer was
4615 * given, a PV chunk is reclaimed from an arbitrary pmap. Otherwise, NULL is
4618 * The given PV list lock may be released.
4621 get_pv_entry(pmap_t pmap, struct rwlock **lockp)
4623 struct pv_chunks_list *pvc;
4626 struct pv_chunk *pc;
4629 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4630 PV_STAT(atomic_add_long(&pv_entry_allocs, 1));
4632 pc = TAILQ_FIRST(&pmap->pm_pvchunk);
4634 for (field = 0; field < _NPCM; field++) {
4635 if (pc->pc_map[field]) {
4636 bit = bsfq(pc->pc_map[field]);
4640 if (field < _NPCM) {
4641 pv = &pc->pc_pventry[field * 64 + bit];
4642 pc->pc_map[field] &= ~(1ul << bit);
4643 /* If this was the last item, move it to tail */
4644 if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 &&
4645 pc->pc_map[2] == 0) {
4646 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4647 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc,
4650 PV_STAT(atomic_add_long(&pv_entry_count, 1));
4651 PV_STAT(atomic_subtract_int(&pv_entry_spare, 1));
4655 /* No free items, allocate another chunk */
4656 m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
4659 if (lockp == NULL) {
4660 PV_STAT(pc_chunk_tryfail++);
4663 m = reclaim_pv_chunk(pmap, lockp);
4667 PV_STAT(atomic_add_int(&pc_chunk_count, 1));
4668 PV_STAT(atomic_add_int(&pc_chunk_allocs, 1));
4669 dump_add_page(m->phys_addr);
4670 pc = (void *)PHYS_TO_DMAP(m->phys_addr);
4672 pc->pc_map[0] = PC_FREE0 & ~1ul; /* preallocated bit 0 */
4673 pc->pc_map[1] = PC_FREE1;
4674 pc->pc_map[2] = PC_FREE2;
4675 pvc = &pv_chunks[_vm_phys_domain(m->phys_addr)];
4676 mtx_lock(&pvc->pvc_lock);
4677 TAILQ_INSERT_TAIL(&pvc->pvc_list, pc, pc_lru);
4678 mtx_unlock(&pvc->pvc_lock);
4679 pv = &pc->pc_pventry[0];
4680 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
4681 PV_STAT(atomic_add_long(&pv_entry_count, 1));
4682 PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV - 1));
4687 * Returns the number of one bits within the given PV chunk map.
4689 * The erratas for Intel processors state that "POPCNT Instruction May
4690 * Take Longer to Execute Than Expected". It is believed that the
4691 * issue is the spurious dependency on the destination register.
4692 * Provide a hint to the register rename logic that the destination
4693 * value is overwritten, by clearing it, as suggested in the
4694 * optimization manual. It should be cheap for unaffected processors
4697 * Reference numbers for erratas are
4698 * 4th Gen Core: HSD146
4699 * 5th Gen Core: BDM85
4700 * 6th Gen Core: SKL029
4703 popcnt_pc_map_pq(uint64_t *map)
4707 __asm __volatile("xorl %k0,%k0;popcntq %2,%0;"
4708 "xorl %k1,%k1;popcntq %3,%1;addl %k1,%k0;"
4709 "xorl %k1,%k1;popcntq %4,%1;addl %k1,%k0"
4710 : "=&r" (result), "=&r" (tmp)
4711 : "m" (map[0]), "m" (map[1]), "m" (map[2]));
4716 * Ensure that the number of spare PV entries in the specified pmap meets or
4717 * exceeds the given count, "needed".
4719 * The given PV list lock may be released.
4722 reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp)
4724 struct pv_chunks_list *pvc;
4725 struct pch new_tail[PMAP_MEMDOM];
4726 struct pv_chunk *pc;
4731 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4732 KASSERT(lockp != NULL, ("reserve_pv_entries: lockp is NULL"));
4735 * Newly allocated PV chunks must be stored in a private list until
4736 * the required number of PV chunks have been allocated. Otherwise,
4737 * reclaim_pv_chunk() could recycle one of these chunks. In
4738 * contrast, these chunks must be added to the pmap upon allocation.
4740 for (i = 0; i < PMAP_MEMDOM; i++)
4741 TAILQ_INIT(&new_tail[i]);
4744 TAILQ_FOREACH(pc, &pmap->pm_pvchunk, pc_list) {
4746 if ((cpu_feature2 & CPUID2_POPCNT) == 0)
4747 bit_count((bitstr_t *)pc->pc_map, 0,
4748 sizeof(pc->pc_map) * NBBY, &free);
4751 free = popcnt_pc_map_pq(pc->pc_map);
4755 if (avail >= needed)
4758 for (reclaimed = false; avail < needed; avail += _NPCPV) {
4759 m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
4762 m = reclaim_pv_chunk(pmap, lockp);
4767 PV_STAT(atomic_add_int(&pc_chunk_count, 1));
4768 PV_STAT(atomic_add_int(&pc_chunk_allocs, 1));
4769 dump_add_page(m->phys_addr);
4770 pc = (void *)PHYS_TO_DMAP(m->phys_addr);
4772 pc->pc_map[0] = PC_FREE0;
4773 pc->pc_map[1] = PC_FREE1;
4774 pc->pc_map[2] = PC_FREE2;
4775 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
4776 TAILQ_INSERT_TAIL(&new_tail[pc_to_domain(pc)], pc, pc_lru);
4777 PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV));
4780 * The reclaim might have freed a chunk from the current pmap.
4781 * If that chunk contained available entries, we need to
4782 * re-count the number of available entries.
4787 for (i = 0; i < vm_ndomains; i++) {
4788 if (TAILQ_EMPTY(&new_tail[i]))
4790 pvc = &pv_chunks[i];
4791 mtx_lock(&pvc->pvc_lock);
4792 TAILQ_CONCAT(&pvc->pvc_list, &new_tail[i], pc_lru);
4793 mtx_unlock(&pvc->pvc_lock);
4798 * First find and then remove the pv entry for the specified pmap and virtual
4799 * address from the specified pv list. Returns the pv entry if found and NULL
4800 * otherwise. This operation can be performed on pv lists for either 4KB or
4801 * 2MB page mappings.
4803 static __inline pv_entry_t
4804 pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
4808 TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
4809 if (pmap == PV_PMAP(pv) && va == pv->pv_va) {
4810 TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
4819 * After demotion from a 2MB page mapping to 512 4KB page mappings,
4820 * destroy the pv entry for the 2MB page mapping and reinstantiate the pv
4821 * entries for each of the 4KB page mappings.
4824 pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
4825 struct rwlock **lockp)
4827 struct md_page *pvh;
4828 struct pv_chunk *pc;
4830 vm_offset_t va_last;
4834 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4835 KASSERT((pa & PDRMASK) == 0,
4836 ("pmap_pv_demote_pde: pa is not 2mpage aligned"));
4837 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
4840 * Transfer the 2mpage's pv entry for this mapping to the first
4841 * page's pv list. Once this transfer begins, the pv list lock
4842 * must not be released until the last pv entry is reinstantiated.
4844 pvh = pa_to_pvh(pa);
4845 va = trunc_2mpage(va);
4846 pv = pmap_pvh_remove(pvh, pmap, va);
4847 KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found"));
4848 m = PHYS_TO_VM_PAGE(pa);
4849 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
4851 /* Instantiate the remaining NPTEPG - 1 pv entries. */
4852 PV_STAT(atomic_add_long(&pv_entry_allocs, NPTEPG - 1));
4853 va_last = va + NBPDR - PAGE_SIZE;
4855 pc = TAILQ_FIRST(&pmap->pm_pvchunk);
4856 KASSERT(pc->pc_map[0] != 0 || pc->pc_map[1] != 0 ||
4857 pc->pc_map[2] != 0, ("pmap_pv_demote_pde: missing spare"));
4858 for (field = 0; field < _NPCM; field++) {
4859 while (pc->pc_map[field]) {
4860 bit = bsfq(pc->pc_map[field]);
4861 pc->pc_map[field] &= ~(1ul << bit);
4862 pv = &pc->pc_pventry[field * 64 + bit];
4866 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4867 ("pmap_pv_demote_pde: page %p is not managed", m));
4868 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
4874 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4875 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
4878 if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) {
4879 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4880 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
4882 PV_STAT(atomic_add_long(&pv_entry_count, NPTEPG - 1));
4883 PV_STAT(atomic_subtract_int(&pv_entry_spare, NPTEPG - 1));
4886 #if VM_NRESERVLEVEL > 0
4888 * After promotion from 512 4KB page mappings to a single 2MB page mapping,
4889 * replace the many pv entries for the 4KB page mappings by a single pv entry
4890 * for the 2MB page mapping.
4893 pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
4894 struct rwlock **lockp)
4896 struct md_page *pvh;
4898 vm_offset_t va_last;
4901 KASSERT((pa & PDRMASK) == 0,
4902 ("pmap_pv_promote_pde: pa is not 2mpage aligned"));
4903 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
4906 * Transfer the first page's pv entry for this mapping to the 2mpage's
4907 * pv list. Aside from avoiding the cost of a call to get_pv_entry(),
4908 * a transfer avoids the possibility that get_pv_entry() calls
4909 * reclaim_pv_chunk() and that reclaim_pv_chunk() removes one of the
4910 * mappings that is being promoted.
4912 m = PHYS_TO_VM_PAGE(pa);
4913 va = trunc_2mpage(va);
4914 pv = pmap_pvh_remove(&m->md, pmap, va);
4915 KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found"));
4916 pvh = pa_to_pvh(pa);
4917 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
4919 /* Free the remaining NPTEPG - 1 pv entries. */
4920 va_last = va + NBPDR - PAGE_SIZE;
4924 pmap_pvh_free(&m->md, pmap, va);
4925 } while (va < va_last);
4927 #endif /* VM_NRESERVLEVEL > 0 */
4930 * First find and then destroy the pv entry for the specified pmap and virtual
4931 * address. This operation can be performed on pv lists for either 4KB or 2MB
4935 pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
4939 pv = pmap_pvh_remove(pvh, pmap, va);
4940 KASSERT(pv != NULL, ("pmap_pvh_free: pv not found"));
4941 free_pv_entry(pmap, pv);
4945 * Conditionally create the PV entry for a 4KB page mapping if the required
4946 * memory can be allocated without resorting to reclamation.
4949 pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m,
4950 struct rwlock **lockp)
4954 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4955 /* Pass NULL instead of the lock pointer to disable reclamation. */
4956 if ((pv = get_pv_entry(pmap, NULL)) != NULL) {
4958 CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
4959 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
4967 * Create the PV entry for a 2MB page mapping. Always returns true unless the
4968 * flag PMAP_ENTER_NORECLAIM is specified. If that flag is specified, returns
4969 * false if the PV entry cannot be allocated without resorting to reclamation.
4972 pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags,
4973 struct rwlock **lockp)
4975 struct md_page *pvh;
4979 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4980 /* Pass NULL instead of the lock pointer to disable reclamation. */
4981 if ((pv = get_pv_entry(pmap, (flags & PMAP_ENTER_NORECLAIM) != 0 ?
4982 NULL : lockp)) == NULL)
4985 pa = pde & PG_PS_FRAME;
4986 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
4987 pvh = pa_to_pvh(pa);
4988 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
4994 * Fills a page table page with mappings to consecutive physical pages.
4997 pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte)
5001 for (pte = firstpte; pte < firstpte + NPTEPG; pte++) {
5003 newpte += PAGE_SIZE;
5008 * Tries to demote a 2MB page mapping. If demotion fails, the 2MB page
5009 * mapping is invalidated.
5012 pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
5014 struct rwlock *lock;
5018 rv = pmap_demote_pde_locked(pmap, pde, va, &lock);
5025 pmap_demote_pde_check(pt_entry_t *firstpte __unused, pt_entry_t newpte __unused)
5029 pt_entry_t *xpte, *ypte;
5031 for (xpte = firstpte; xpte < firstpte + NPTEPG;
5032 xpte++, newpte += PAGE_SIZE) {
5033 if ((*xpte & PG_FRAME) != (newpte & PG_FRAME)) {
5034 printf("pmap_demote_pde: xpte %zd and newpte map "
5035 "different pages: found %#lx, expected %#lx\n",
5036 xpte - firstpte, *xpte, newpte);
5037 printf("page table dump\n");
5038 for (ypte = firstpte; ypte < firstpte + NPTEPG; ypte++)
5039 printf("%zd %#lx\n", ypte - firstpte, *ypte);
5044 KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME),
5045 ("pmap_demote_pde: firstpte and newpte map different physical"
5052 pmap_demote_pde_abort(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
5053 pd_entry_t oldpde, struct rwlock **lockp)
5055 struct spglist free;
5059 sva = trunc_2mpage(va);
5060 pmap_remove_pde(pmap, pde, sva, &free, lockp);
5061 if ((oldpde & pmap_global_bit(pmap)) == 0)
5062 pmap_invalidate_pde_page(pmap, sva, oldpde);
5063 vm_page_free_pages_toq(&free, true);
5064 CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#lx in pmap %p",
5069 pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va,
5070 struct rwlock **lockp)
5072 pd_entry_t newpde, oldpde;
5073 pt_entry_t *firstpte, newpte;
5074 pt_entry_t PG_A, PG_G, PG_M, PG_PKU_MASK, PG_RW, PG_V;
5080 PG_A = pmap_accessed_bit(pmap);
5081 PG_G = pmap_global_bit(pmap);
5082 PG_M = pmap_modified_bit(pmap);
5083 PG_RW = pmap_rw_bit(pmap);
5084 PG_V = pmap_valid_bit(pmap);
5085 PG_PTE_CACHE = pmap_cache_mask(pmap, 0);
5086 PG_PKU_MASK = pmap_pku_mask_bit(pmap);
5088 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5089 in_kernel = va >= VM_MAXUSER_ADDRESS;
5091 KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V),
5092 ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V"));
5095 * Invalidate the 2MB page mapping and return "failure" if the
5096 * mapping was never accessed.
5098 if ((oldpde & PG_A) == 0) {
5099 KASSERT((oldpde & PG_W) == 0,
5100 ("pmap_demote_pde: a wired mapping is missing PG_A"));
5101 pmap_demote_pde_abort(pmap, va, pde, oldpde, lockp);
5105 mpte = pmap_remove_pt_page(pmap, va);
5107 KASSERT((oldpde & PG_W) == 0,
5108 ("pmap_demote_pde: page table page for a wired mapping"
5112 * If the page table page is missing and the mapping
5113 * is for a kernel address, the mapping must belong to
5114 * the direct map. Page table pages are preallocated
5115 * for every other part of the kernel address space,
5116 * so the direct map region is the only part of the
5117 * kernel address space that must be handled here.
5119 KASSERT(!in_kernel || (va >= DMAP_MIN_ADDRESS &&
5120 va < DMAP_MAX_ADDRESS),
5121 ("pmap_demote_pde: No saved mpte for va %#lx", va));
5124 * If the 2MB page mapping belongs to the direct map
5125 * region of the kernel's address space, then the page
5126 * allocation request specifies the highest possible
5127 * priority (VM_ALLOC_INTERRUPT). Otherwise, the
5128 * priority is normal.
5130 mpte = vm_page_alloc(NULL, pmap_pde_pindex(va),
5131 (in_kernel ? VM_ALLOC_INTERRUPT : VM_ALLOC_NORMAL) |
5132 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED);
5135 * If the allocation of the new page table page fails,
5136 * invalidate the 2MB page mapping and return "failure".
5139 pmap_demote_pde_abort(pmap, va, pde, oldpde, lockp);
5144 mpte->ref_count = NPTEPG;
5145 pmap_resident_count_inc(pmap, 1);
5148 mptepa = VM_PAGE_TO_PHYS(mpte);
5149 firstpte = (pt_entry_t *)PHYS_TO_DMAP(mptepa);
5150 newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V;
5151 KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW,
5152 ("pmap_demote_pde: oldpde is missing PG_M"));
5153 newpte = oldpde & ~PG_PS;
5154 newpte = pmap_swap_pat(pmap, newpte);
5157 * If the page table page is not leftover from an earlier promotion,
5160 if (mpte->valid == 0)
5161 pmap_fill_ptp(firstpte, newpte);
5163 pmap_demote_pde_check(firstpte, newpte);
5166 * If the mapping has changed attributes, update the page table
5169 if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE))
5170 pmap_fill_ptp(firstpte, newpte);
5173 * The spare PV entries must be reserved prior to demoting the
5174 * mapping, that is, prior to changing the PDE. Otherwise, the state
5175 * of the PDE and the PV lists will be inconsistent, which can result
5176 * in reclaim_pv_chunk() attempting to remove a PV entry from the
5177 * wrong PV list and pmap_pv_demote_pde() failing to find the expected
5178 * PV entry for the 2MB page mapping that is being demoted.
5180 if ((oldpde & PG_MANAGED) != 0)
5181 reserve_pv_entries(pmap, NPTEPG - 1, lockp);
5184 * Demote the mapping. This pmap is locked. The old PDE has
5185 * PG_A set. If the old PDE has PG_RW set, it also has PG_M
5186 * set. Thus, there is no danger of a race with another
5187 * processor changing the setting of PG_A and/or PG_M between
5188 * the read above and the store below.
5190 if (workaround_erratum383)
5191 pmap_update_pde(pmap, va, pde, newpde);
5193 pde_store(pde, newpde);
5196 * Invalidate a stale recursive mapping of the page table page.
5199 pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va));
5202 * Demote the PV entry.
5204 if ((oldpde & PG_MANAGED) != 0)
5205 pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME, lockp);
5207 atomic_add_long(&pmap_pde_demotions, 1);
5208 CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#lx in pmap %p",
5214 * pmap_remove_kernel_pde: Remove a kernel superpage mapping.
5217 pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
5223 KASSERT(pmap == kernel_pmap, ("pmap %p is not kernel_pmap", pmap));
5224 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5225 mpte = pmap_remove_pt_page(pmap, va);
5227 panic("pmap_remove_kernel_pde: Missing pt page.");
5229 mptepa = VM_PAGE_TO_PHYS(mpte);
5230 newpde = mptepa | X86_PG_M | X86_PG_A | X86_PG_RW | X86_PG_V;
5233 * If this page table page was unmapped by a promotion, then it
5234 * contains valid mappings. Zero it to invalidate those mappings.
5236 if (mpte->valid != 0)
5237 pagezero((void *)PHYS_TO_DMAP(mptepa));
5240 * Demote the mapping.
5242 if (workaround_erratum383)
5243 pmap_update_pde(pmap, va, pde, newpde);
5245 pde_store(pde, newpde);
5248 * Invalidate a stale recursive mapping of the page table page.
5250 pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va));
5254 * pmap_remove_pde: do the things to unmap a superpage in a process
5257 pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
5258 struct spglist *free, struct rwlock **lockp)
5260 struct md_page *pvh;
5262 vm_offset_t eva, va;
5264 pt_entry_t PG_G, PG_A, PG_M, PG_RW;
5266 PG_G = pmap_global_bit(pmap);
5267 PG_A = pmap_accessed_bit(pmap);
5268 PG_M = pmap_modified_bit(pmap);
5269 PG_RW = pmap_rw_bit(pmap);
5271 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5272 KASSERT((sva & PDRMASK) == 0,
5273 ("pmap_remove_pde: sva is not 2mpage aligned"));
5274 oldpde = pte_load_clear(pdq);
5276 pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE;
5277 if ((oldpde & PG_G) != 0)
5278 pmap_invalidate_pde_page(kernel_pmap, sva, oldpde);
5279 pmap_resident_count_dec(pmap, NBPDR / PAGE_SIZE);
5280 if (oldpde & PG_MANAGED) {
5281 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, oldpde & PG_PS_FRAME);
5282 pvh = pa_to_pvh(oldpde & PG_PS_FRAME);
5283 pmap_pvh_free(pvh, pmap, sva);
5285 for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
5286 va < eva; va += PAGE_SIZE, m++) {
5287 if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW))
5290 vm_page_aflag_set(m, PGA_REFERENCED);
5291 if (TAILQ_EMPTY(&m->md.pv_list) &&
5292 TAILQ_EMPTY(&pvh->pv_list))
5293 vm_page_aflag_clear(m, PGA_WRITEABLE);
5294 pmap_delayed_invl_page(m);
5297 if (pmap == kernel_pmap) {
5298 pmap_remove_kernel_pde(pmap, pdq, sva);
5300 mpte = pmap_remove_pt_page(pmap, sva);
5302 KASSERT(mpte->valid == VM_PAGE_BITS_ALL,
5303 ("pmap_remove_pde: pte page not promoted"));
5304 pmap_resident_count_dec(pmap, 1);
5305 KASSERT(mpte->ref_count == NPTEPG,
5306 ("pmap_remove_pde: pte page ref count error"));
5307 mpte->ref_count = 0;
5308 pmap_add_delayed_free_list(mpte, free, FALSE);
5311 return (pmap_unuse_pt(pmap, sva, *pmap_pdpe(pmap, sva), free));
5315 * pmap_remove_pte: do the things to unmap a page in a process
5318 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va,
5319 pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp)
5321 struct md_page *pvh;
5322 pt_entry_t oldpte, PG_A, PG_M, PG_RW;
5325 PG_A = pmap_accessed_bit(pmap);
5326 PG_M = pmap_modified_bit(pmap);
5327 PG_RW = pmap_rw_bit(pmap);
5329 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5330 oldpte = pte_load_clear(ptq);
5332 pmap->pm_stats.wired_count -= 1;
5333 pmap_resident_count_dec(pmap, 1);
5334 if (oldpte & PG_MANAGED) {
5335 m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME);
5336 if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5339 vm_page_aflag_set(m, PGA_REFERENCED);
5340 CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
5341 pmap_pvh_free(&m->md, pmap, va);
5342 if (TAILQ_EMPTY(&m->md.pv_list) &&
5343 (m->flags & PG_FICTITIOUS) == 0) {
5344 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
5345 if (TAILQ_EMPTY(&pvh->pv_list))
5346 vm_page_aflag_clear(m, PGA_WRITEABLE);
5348 pmap_delayed_invl_page(m);
5350 return (pmap_unuse_pt(pmap, va, ptepde, free));
5354 * Remove a single page from a process address space
5357 pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
5358 struct spglist *free)
5360 struct rwlock *lock;
5361 pt_entry_t *pte, PG_V;
5363 PG_V = pmap_valid_bit(pmap);
5364 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5365 if ((*pde & PG_V) == 0)
5367 pte = pmap_pde_to_pte(pde, va);
5368 if ((*pte & PG_V) == 0)
5371 pmap_remove_pte(pmap, pte, va, *pde, free, &lock);
5374 pmap_invalidate_page(pmap, va);
5378 * Removes the specified range of addresses from the page table page.
5381 pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
5382 pd_entry_t *pde, struct spglist *free, struct rwlock **lockp)
5384 pt_entry_t PG_G, *pte;
5388 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5389 PG_G = pmap_global_bit(pmap);
5392 for (pte = pmap_pde_to_pte(pde, sva); sva != eva; pte++,
5396 pmap_invalidate_range(pmap, va, sva);
5401 if ((*pte & PG_G) == 0)
5405 if (pmap_remove_pte(pmap, pte, sva, *pde, free, lockp)) {
5411 pmap_invalidate_range(pmap, va, sva);
5416 * Remove the given range of addresses from the specified map.
5418 * It is assumed that the start and end are properly
5419 * rounded to the page size.
5422 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5424 struct rwlock *lock;
5425 vm_offset_t va_next;
5426 pml4_entry_t *pml4e;
5428 pd_entry_t ptpaddr, *pde;
5429 pt_entry_t PG_G, PG_V;
5430 struct spglist free;
5433 PG_G = pmap_global_bit(pmap);
5434 PG_V = pmap_valid_bit(pmap);
5437 * Perform an unsynchronized read. This is, however, safe.
5439 if (pmap->pm_stats.resident_count == 0)
5445 pmap_delayed_invl_start();
5447 pmap_pkru_on_remove(pmap, sva, eva);
5450 * special handling of removing one page. a very
5451 * common operation and easy to short circuit some
5454 if (sva + PAGE_SIZE == eva) {
5455 pde = pmap_pde(pmap, sva);
5456 if (pde && (*pde & PG_PS) == 0) {
5457 pmap_remove_page(pmap, sva, pde, &free);
5463 for (; sva < eva; sva = va_next) {
5465 if (pmap->pm_stats.resident_count == 0)
5468 pml4e = pmap_pml4e(pmap, sva);
5469 if ((*pml4e & PG_V) == 0) {
5470 va_next = (sva + NBPML4) & ~PML4MASK;
5476 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
5477 if ((*pdpe & PG_V) == 0) {
5478 va_next = (sva + NBPDP) & ~PDPMASK;
5485 * Calculate index for next page table.
5487 va_next = (sva + NBPDR) & ~PDRMASK;
5491 pde = pmap_pdpe_to_pde(pdpe, sva);
5495 * Weed out invalid mappings.
5501 * Check for large page.
5503 if ((ptpaddr & PG_PS) != 0) {
5505 * Are we removing the entire large page? If not,
5506 * demote the mapping and fall through.
5508 if (sva + NBPDR == va_next && eva >= va_next) {
5510 * The TLB entry for a PG_G mapping is
5511 * invalidated by pmap_remove_pde().
5513 if ((ptpaddr & PG_G) == 0)
5515 pmap_remove_pde(pmap, pde, sva, &free, &lock);
5517 } else if (!pmap_demote_pde_locked(pmap, pde, sva,
5519 /* The large page mapping was destroyed. */
5526 * Limit our scan to either the end of the va represented
5527 * by the current page table page, or to the end of the
5528 * range being removed.
5533 if (pmap_remove_ptes(pmap, sva, va_next, pde, &free, &lock))
5540 pmap_invalidate_all(pmap);
5542 pmap_delayed_invl_finish();
5543 vm_page_free_pages_toq(&free, true);
5547 * Routine: pmap_remove_all
5549 * Removes this physical page from
5550 * all physical maps in which it resides.
5551 * Reflects back modify bits to the pager.
5554 * Original versions of this routine were very
5555 * inefficient because they iteratively called
5556 * pmap_remove (slow...)
5560 pmap_remove_all(vm_page_t m)
5562 struct md_page *pvh;
5565 struct rwlock *lock;
5566 pt_entry_t *pte, tpte, PG_A, PG_M, PG_RW;
5569 struct spglist free;
5570 int pvh_gen, md_gen;
5572 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5573 ("pmap_remove_all: page %p is not managed", m));
5575 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
5576 pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
5577 pa_to_pvh(VM_PAGE_TO_PHYS(m));
5580 while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) {
5582 if (!PMAP_TRYLOCK(pmap)) {
5583 pvh_gen = pvh->pv_gen;
5587 if (pvh_gen != pvh->pv_gen) {
5594 pde = pmap_pde(pmap, va);
5595 (void)pmap_demote_pde_locked(pmap, pde, va, &lock);
5598 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
5600 if (!PMAP_TRYLOCK(pmap)) {
5601 pvh_gen = pvh->pv_gen;
5602 md_gen = m->md.pv_gen;
5606 if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
5612 PG_A = pmap_accessed_bit(pmap);
5613 PG_M = pmap_modified_bit(pmap);
5614 PG_RW = pmap_rw_bit(pmap);
5615 pmap_resident_count_dec(pmap, 1);
5616 pde = pmap_pde(pmap, pv->pv_va);
5617 KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found"
5618 " a 2mpage in page %p's pv list", m));
5619 pte = pmap_pde_to_pte(pde, pv->pv_va);
5620 tpte = pte_load_clear(pte);
5622 pmap->pm_stats.wired_count--;
5624 vm_page_aflag_set(m, PGA_REFERENCED);
5627 * Update the vm_page_t clean and reference bits.
5629 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5631 pmap_unuse_pt(pmap, pv->pv_va, *pde, &free);
5632 pmap_invalidate_page(pmap, pv->pv_va);
5633 TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
5635 free_pv_entry(pmap, pv);
5638 vm_page_aflag_clear(m, PGA_WRITEABLE);
5640 pmap_delayed_invl_wait(m);
5641 vm_page_free_pages_toq(&free, true);
5645 * pmap_protect_pde: do the things to protect a 2mpage in a process
5648 pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot)
5650 pd_entry_t newpde, oldpde;
5652 boolean_t anychanged;
5653 pt_entry_t PG_G, PG_M, PG_RW;
5655 PG_G = pmap_global_bit(pmap);
5656 PG_M = pmap_modified_bit(pmap);
5657 PG_RW = pmap_rw_bit(pmap);
5659 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5660 KASSERT((sva & PDRMASK) == 0,
5661 ("pmap_protect_pde: sva is not 2mpage aligned"));
5664 oldpde = newpde = *pde;
5665 if ((prot & VM_PROT_WRITE) == 0) {
5666 if ((oldpde & (PG_MANAGED | PG_M | PG_RW)) ==
5667 (PG_MANAGED | PG_M | PG_RW)) {
5668 m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
5669 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
5672 newpde &= ~(PG_RW | PG_M);
5674 if ((prot & VM_PROT_EXECUTE) == 0)
5676 if (newpde != oldpde) {
5678 * As an optimization to future operations on this PDE, clear
5679 * PG_PROMOTED. The impending invalidation will remove any
5680 * lingering 4KB page mappings from the TLB.
5682 if (!atomic_cmpset_long(pde, oldpde, newpde & ~PG_PROMOTED))
5684 if ((oldpde & PG_G) != 0)
5685 pmap_invalidate_pde_page(kernel_pmap, sva, oldpde);
5689 return (anychanged);
5693 * Set the physical protection on the
5694 * specified range of this map as requested.
5697 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
5699 vm_offset_t va_next;
5700 pml4_entry_t *pml4e;
5702 pd_entry_t ptpaddr, *pde;
5703 pt_entry_t *pte, PG_G, PG_M, PG_RW, PG_V;
5704 boolean_t anychanged;
5706 KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot));
5707 if (prot == VM_PROT_NONE) {
5708 pmap_remove(pmap, sva, eva);
5712 if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) ==
5713 (VM_PROT_WRITE|VM_PROT_EXECUTE))
5716 PG_G = pmap_global_bit(pmap);
5717 PG_M = pmap_modified_bit(pmap);
5718 PG_V = pmap_valid_bit(pmap);
5719 PG_RW = pmap_rw_bit(pmap);
5723 * Although this function delays and batches the invalidation
5724 * of stale TLB entries, it does not need to call
5725 * pmap_delayed_invl_start() and
5726 * pmap_delayed_invl_finish(), because it does not
5727 * ordinarily destroy mappings. Stale TLB entries from
5728 * protection-only changes need only be invalidated before the
5729 * pmap lock is released, because protection-only changes do
5730 * not destroy PV entries. Even operations that iterate over
5731 * a physical page's PV list of mappings, like
5732 * pmap_remove_write(), acquire the pmap lock for each
5733 * mapping. Consequently, for protection-only changes, the
5734 * pmap lock suffices to synchronize both page table and TLB
5737 * This function only destroys a mapping if pmap_demote_pde()
5738 * fails. In that case, stale TLB entries are immediately
5743 for (; sva < eva; sva = va_next) {
5745 pml4e = pmap_pml4e(pmap, sva);
5746 if ((*pml4e & PG_V) == 0) {
5747 va_next = (sva + NBPML4) & ~PML4MASK;
5753 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
5754 if ((*pdpe & PG_V) == 0) {
5755 va_next = (sva + NBPDP) & ~PDPMASK;
5761 va_next = (sva + NBPDR) & ~PDRMASK;
5765 pde = pmap_pdpe_to_pde(pdpe, sva);
5769 * Weed out invalid mappings.
5775 * Check for large page.
5777 if ((ptpaddr & PG_PS) != 0) {
5779 * Are we protecting the entire large page? If not,
5780 * demote the mapping and fall through.
5782 if (sva + NBPDR == va_next && eva >= va_next) {
5784 * The TLB entry for a PG_G mapping is
5785 * invalidated by pmap_protect_pde().
5787 if (pmap_protect_pde(pmap, pde, sva, prot))
5790 } else if (!pmap_demote_pde(pmap, pde, sva)) {
5792 * The large page mapping was destroyed.
5801 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
5803 pt_entry_t obits, pbits;
5807 obits = pbits = *pte;
5808 if ((pbits & PG_V) == 0)
5811 if ((prot & VM_PROT_WRITE) == 0) {
5812 if ((pbits & (PG_MANAGED | PG_M | PG_RW)) ==
5813 (PG_MANAGED | PG_M | PG_RW)) {
5814 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
5817 pbits &= ~(PG_RW | PG_M);
5819 if ((prot & VM_PROT_EXECUTE) == 0)
5822 if (pbits != obits) {
5823 if (!atomic_cmpset_long(pte, obits, pbits))
5826 pmap_invalidate_page(pmap, sva);
5833 pmap_invalidate_all(pmap);
5837 #if VM_NRESERVLEVEL > 0
5839 pmap_pde_ept_executable(pmap_t pmap, pd_entry_t pde)
5842 if (pmap->pm_type != PT_EPT)
5844 return ((pde & EPT_PG_EXECUTE) != 0);
5848 * Tries to promote the 512, contiguous 4KB page mappings that are within a
5849 * single page table page (PTP) to a single 2MB page mapping. For promotion
5850 * to occur, two conditions must be met: (1) the 4KB page mappings must map
5851 * aligned, contiguous physical memory and (2) the 4KB page mappings must have
5852 * identical characteristics.
5855 pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va,
5856 struct rwlock **lockp)
5859 pt_entry_t *firstpte, oldpte, pa, *pte;
5860 pt_entry_t PG_G, PG_A, PG_M, PG_RW, PG_V, PG_PKU_MASK;
5864 PG_A = pmap_accessed_bit(pmap);
5865 PG_G = pmap_global_bit(pmap);
5866 PG_M = pmap_modified_bit(pmap);
5867 PG_V = pmap_valid_bit(pmap);
5868 PG_RW = pmap_rw_bit(pmap);
5869 PG_PKU_MASK = pmap_pku_mask_bit(pmap);
5870 PG_PTE_CACHE = pmap_cache_mask(pmap, 0);
5872 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5875 * Examine the first PTE in the specified PTP. Abort if this PTE is
5876 * either invalid, unused, or does not map the first 4KB physical page
5877 * within a 2MB page.
5879 firstpte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
5882 if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V) ||
5883 !pmap_allow_2m_x_page(pmap, pmap_pde_ept_executable(pmap,
5885 atomic_add_long(&pmap_pde_p_failures, 1);
5886 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx"
5887 " in pmap %p", va, pmap);
5890 if ((newpde & (PG_M | PG_RW)) == PG_RW) {
5892 * When PG_M is already clear, PG_RW can be cleared without
5893 * a TLB invalidation.
5895 if (!atomic_cmpset_long(firstpte, newpde, newpde & ~PG_RW))
5901 * Examine each of the other PTEs in the specified PTP. Abort if this
5902 * PTE maps an unexpected 4KB physical page or does not have identical
5903 * characteristics to the first PTE.
5905 pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE;
5906 for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) {
5909 if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) {
5910 atomic_add_long(&pmap_pde_p_failures, 1);
5911 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx"
5912 " in pmap %p", va, pmap);
5915 if ((oldpte & (PG_M | PG_RW)) == PG_RW) {
5917 * When PG_M is already clear, PG_RW can be cleared
5918 * without a TLB invalidation.
5920 if (!atomic_cmpset_long(pte, oldpte, oldpte & ~PG_RW))
5923 CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#lx"
5924 " in pmap %p", (oldpte & PG_FRAME & PDRMASK) |
5925 (va & ~PDRMASK), pmap);
5927 if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) {
5928 atomic_add_long(&pmap_pde_p_failures, 1);
5929 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx"
5930 " in pmap %p", va, pmap);
5937 * Save the page table page in its current state until the PDE
5938 * mapping the superpage is demoted by pmap_demote_pde() or
5939 * destroyed by pmap_remove_pde().
5941 mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
5942 KASSERT(mpte >= vm_page_array &&
5943 mpte < &vm_page_array[vm_page_array_size],
5944 ("pmap_promote_pde: page table page is out of range"));
5945 KASSERT(mpte->pindex == pmap_pde_pindex(va),
5946 ("pmap_promote_pde: page table page's pindex is wrong"));
5947 if (pmap_insert_pt_page(pmap, mpte, true)) {
5948 atomic_add_long(&pmap_pde_p_failures, 1);
5950 "pmap_promote_pde: failure for va %#lx in pmap %p", va,
5956 * Promote the pv entries.
5958 if ((newpde & PG_MANAGED) != 0)
5959 pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME, lockp);
5962 * Propagate the PAT index to its proper position.
5964 newpde = pmap_swap_pat(pmap, newpde);
5967 * Map the superpage.
5969 if (workaround_erratum383)
5970 pmap_update_pde(pmap, va, pde, PG_PS | newpde);
5972 pde_store(pde, PG_PROMOTED | PG_PS | newpde);
5974 atomic_add_long(&pmap_pde_promotions, 1);
5975 CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#lx"
5976 " in pmap %p", va, pmap);
5978 #endif /* VM_NRESERVLEVEL > 0 */
5981 * Insert the given physical page (p) at
5982 * the specified virtual address (v) in the
5983 * target physical map with the protection requested.
5985 * If specified, the page will be wired down, meaning
5986 * that the related pte can not be reclaimed.
5988 * NB: This is the only routine which MAY NOT lazy-evaluate
5989 * or lose information. That is, this routine must actually
5990 * insert this page into the given map NOW.
5992 * When destroying both a page table and PV entry, this function
5993 * performs the TLB invalidation before releasing the PV list
5994 * lock, so we do not need pmap_delayed_invl_page() calls here.
5997 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
5998 u_int flags, int8_t psind)
6000 struct rwlock *lock;
6002 pt_entry_t *pte, PG_G, PG_A, PG_M, PG_RW, PG_V;
6003 pt_entry_t newpte, origpte;
6010 PG_A = pmap_accessed_bit(pmap);
6011 PG_G = pmap_global_bit(pmap);
6012 PG_M = pmap_modified_bit(pmap);
6013 PG_V = pmap_valid_bit(pmap);
6014 PG_RW = pmap_rw_bit(pmap);
6016 va = trunc_page(va);
6017 KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig"));
6018 KASSERT(va < UPT_MIN_ADDRESS || va >= UPT_MAX_ADDRESS,
6019 ("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)",
6021 KASSERT((m->oflags & VPO_UNMANAGED) != 0 || va < kmi.clean_sva ||
6022 va >= kmi.clean_eva,
6023 ("pmap_enter: managed mapping within the clean submap"));
6024 if ((m->oflags & VPO_UNMANAGED) == 0)
6025 VM_PAGE_OBJECT_BUSY_ASSERT(m);
6026 KASSERT((flags & PMAP_ENTER_RESERVED) == 0,
6027 ("pmap_enter: flags %u has reserved bits set", flags));
6028 pa = VM_PAGE_TO_PHYS(m);
6029 newpte = (pt_entry_t)(pa | PG_A | PG_V);
6030 if ((flags & VM_PROT_WRITE) != 0)
6032 if ((prot & VM_PROT_WRITE) != 0)
6034 KASSERT((newpte & (PG_M | PG_RW)) != PG_M,
6035 ("pmap_enter: flags includes VM_PROT_WRITE but prot doesn't"));
6036 if ((prot & VM_PROT_EXECUTE) == 0)
6038 if ((flags & PMAP_ENTER_WIRED) != 0)
6040 if (va < VM_MAXUSER_ADDRESS)
6042 if (pmap == kernel_pmap)
6044 newpte |= pmap_cache_bits(pmap, m->md.pat_mode, psind > 0);
6047 * Set modified bit gratuitously for writeable mappings if
6048 * the page is unmanaged. We do not want to take a fault
6049 * to do the dirty bit accounting for these mappings.
6051 if ((m->oflags & VPO_UNMANAGED) != 0) {
6052 if ((newpte & PG_RW) != 0)
6055 newpte |= PG_MANAGED;
6060 /* Assert the required virtual and physical alignment. */
6061 KASSERT((va & PDRMASK) == 0, ("pmap_enter: va unaligned"));
6062 KASSERT(m->psind > 0, ("pmap_enter: m->psind < psind"));
6063 rv = pmap_enter_pde(pmap, va, newpte | PG_PS, flags, m, &lock);
6069 * In the case that a page table page is not
6070 * resident, we are creating it here.
6073 pde = pmap_pde(pmap, va);
6074 if (pde != NULL && (*pde & PG_V) != 0 && ((*pde & PG_PS) == 0 ||
6075 pmap_demote_pde_locked(pmap, pde, va, &lock))) {
6076 pte = pmap_pde_to_pte(pde, va);
6077 if (va < VM_MAXUSER_ADDRESS && mpte == NULL) {
6078 mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
6081 } else if (va < VM_MAXUSER_ADDRESS) {
6083 * Here if the pte page isn't mapped, or if it has been
6086 nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0;
6087 mpte = _pmap_allocpte(pmap, pmap_pde_pindex(va),
6088 nosleep ? NULL : &lock);
6089 if (mpte == NULL && nosleep) {
6090 rv = KERN_RESOURCE_SHORTAGE;
6095 panic("pmap_enter: invalid page directory va=%#lx", va);
6099 if (va < VM_MAXUSER_ADDRESS && pmap->pm_type == PT_X86)
6100 newpte |= pmap_pkru_get(pmap, va);
6103 * Is the specified virtual address already mapped?
6105 if ((origpte & PG_V) != 0) {
6107 * Wiring change, just update stats. We don't worry about
6108 * wiring PT pages as they remain resident as long as there
6109 * are valid mappings in them. Hence, if a user page is wired,
6110 * the PT page will be also.
6112 if ((newpte & PG_W) != 0 && (origpte & PG_W) == 0)
6113 pmap->pm_stats.wired_count++;
6114 else if ((newpte & PG_W) == 0 && (origpte & PG_W) != 0)
6115 pmap->pm_stats.wired_count--;
6118 * Remove the extra PT page reference.
6122 KASSERT(mpte->ref_count > 0,
6123 ("pmap_enter: missing reference to page table page,"
6128 * Has the physical page changed?
6130 opa = origpte & PG_FRAME;
6133 * No, might be a protection or wiring change.
6135 if ((origpte & PG_MANAGED) != 0 &&
6136 (newpte & PG_RW) != 0)
6137 vm_page_aflag_set(m, PGA_WRITEABLE);
6138 if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0)
6144 * The physical page has changed. Temporarily invalidate
6145 * the mapping. This ensures that all threads sharing the
6146 * pmap keep a consistent view of the mapping, which is
6147 * necessary for the correct handling of COW faults. It
6148 * also permits reuse of the old mapping's PV entry,
6149 * avoiding an allocation.
6151 * For consistency, handle unmanaged mappings the same way.
6153 origpte = pte_load_clear(pte);
6154 KASSERT((origpte & PG_FRAME) == opa,
6155 ("pmap_enter: unexpected pa update for %#lx", va));
6156 if ((origpte & PG_MANAGED) != 0) {
6157 om = PHYS_TO_VM_PAGE(opa);
6160 * The pmap lock is sufficient to synchronize with
6161 * concurrent calls to pmap_page_test_mappings() and
6162 * pmap_ts_referenced().
6164 if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
6166 if ((origpte & PG_A) != 0) {
6167 pmap_invalidate_page(pmap, va);
6168 vm_page_aflag_set(om, PGA_REFERENCED);
6170 CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, opa);
6171 pv = pmap_pvh_remove(&om->md, pmap, va);
6173 ("pmap_enter: no PV entry for %#lx", va));
6174 if ((newpte & PG_MANAGED) == 0)
6175 free_pv_entry(pmap, pv);
6176 if ((om->a.flags & PGA_WRITEABLE) != 0 &&
6177 TAILQ_EMPTY(&om->md.pv_list) &&
6178 ((om->flags & PG_FICTITIOUS) != 0 ||
6179 TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list)))
6180 vm_page_aflag_clear(om, PGA_WRITEABLE);
6183 * Since this mapping is unmanaged, assume that PG_A
6186 pmap_invalidate_page(pmap, va);
6191 * Increment the counters.
6193 if ((newpte & PG_W) != 0)
6194 pmap->pm_stats.wired_count++;
6195 pmap_resident_count_inc(pmap, 1);
6199 * Enter on the PV list if part of our managed memory.
6201 if ((newpte & PG_MANAGED) != 0) {
6203 pv = get_pv_entry(pmap, &lock);
6206 CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, pa);
6207 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
6209 if ((newpte & PG_RW) != 0)
6210 vm_page_aflag_set(m, PGA_WRITEABLE);
6216 if ((origpte & PG_V) != 0) {
6218 origpte = pte_load_store(pte, newpte);
6219 KASSERT((origpte & PG_FRAME) == pa,
6220 ("pmap_enter: unexpected pa update for %#lx", va));
6221 if ((newpte & PG_M) == 0 && (origpte & (PG_M | PG_RW)) ==
6223 if ((origpte & PG_MANAGED) != 0)
6227 * Although the PTE may still have PG_RW set, TLB
6228 * invalidation may nonetheless be required because
6229 * the PTE no longer has PG_M set.
6231 } else if ((origpte & PG_NX) != 0 || (newpte & PG_NX) == 0) {
6233 * This PTE change does not require TLB invalidation.
6237 if ((origpte & PG_A) != 0)
6238 pmap_invalidate_page(pmap, va);
6240 pte_store(pte, newpte);
6244 #if VM_NRESERVLEVEL > 0
6246 * If both the page table page and the reservation are fully
6247 * populated, then attempt promotion.
6249 if ((mpte == NULL || mpte->ref_count == NPTEPG) &&
6250 pmap_ps_enabled(pmap) &&
6251 (m->flags & PG_FICTITIOUS) == 0 &&
6252 vm_reserv_level_iffullpop(m) == 0)
6253 pmap_promote_pde(pmap, pde, va, &lock);
6265 * Tries to create a read- and/or execute-only 2MB page mapping. Returns true
6266 * if successful. Returns false if (1) a page table page cannot be allocated
6267 * without sleeping, (2) a mapping already exists at the specified virtual
6268 * address, or (3) a PV entry cannot be allocated without reclaiming another
6272 pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
6273 struct rwlock **lockp)
6278 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
6279 PG_V = pmap_valid_bit(pmap);
6280 newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(pmap, m->md.pat_mode, 1) |
6282 if ((m->oflags & VPO_UNMANAGED) == 0)
6283 newpde |= PG_MANAGED;
6284 if ((prot & VM_PROT_EXECUTE) == 0)
6286 if (va < VM_MAXUSER_ADDRESS)
6288 return (pmap_enter_pde(pmap, va, newpde, PMAP_ENTER_NOSLEEP |
6289 PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, NULL, lockp) ==
6294 * Returns true if every page table entry in the specified page table page is
6298 pmap_every_pte_zero(vm_paddr_t pa)
6300 pt_entry_t *pt_end, *pte;
6302 KASSERT((pa & PAGE_MASK) == 0, ("pa is misaligned"));
6303 pte = (pt_entry_t *)PHYS_TO_DMAP(pa);
6304 for (pt_end = pte + NPTEPG; pte < pt_end; pte++) {
6312 * Tries to create the specified 2MB page mapping. Returns KERN_SUCCESS if
6313 * the mapping was created, and either KERN_FAILURE or KERN_RESOURCE_SHORTAGE
6314 * otherwise. Returns KERN_FAILURE if PMAP_ENTER_NOREPLACE was specified and
6315 * a mapping already exists at the specified virtual address. Returns
6316 * KERN_RESOURCE_SHORTAGE if PMAP_ENTER_NOSLEEP was specified and a page table
6317 * page allocation failed. Returns KERN_RESOURCE_SHORTAGE if
6318 * PMAP_ENTER_NORECLAIM was specified and a PV entry allocation failed.
6320 * The parameter "m" is only used when creating a managed, writeable mapping.
6323 pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags,
6324 vm_page_t m, struct rwlock **lockp)
6326 struct spglist free;
6327 pd_entry_t oldpde, *pde;
6328 pt_entry_t PG_G, PG_RW, PG_V;
6331 KASSERT(pmap == kernel_pmap || (newpde & PG_W) == 0,
6332 ("pmap_enter_pde: cannot create wired user mapping"));
6333 PG_G = pmap_global_bit(pmap);
6334 PG_RW = pmap_rw_bit(pmap);
6335 KASSERT((newpde & (pmap_modified_bit(pmap) | PG_RW)) != PG_RW,
6336 ("pmap_enter_pde: newpde is missing PG_M"));
6337 PG_V = pmap_valid_bit(pmap);
6338 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
6340 if (!pmap_allow_2m_x_page(pmap, pmap_pde_ept_executable(pmap,
6342 CTR2(KTR_PMAP, "pmap_enter_pde: 2m x blocked for va %#lx"
6343 " in pmap %p", va, pmap);
6344 return (KERN_FAILURE);
6346 if ((pde = pmap_alloc_pde(pmap, va, &pdpg, (flags &
6347 PMAP_ENTER_NOSLEEP) != 0 ? NULL : lockp)) == NULL) {
6348 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
6349 " in pmap %p", va, pmap);
6350 return (KERN_RESOURCE_SHORTAGE);
6354 * If pkru is not same for the whole pde range, return failure
6355 * and let vm_fault() cope. Check after pde allocation, since
6358 if (!pmap_pkru_same(pmap, va, va + NBPDR)) {
6359 pmap_abort_ptp(pmap, va, pdpg);
6360 return (KERN_FAILURE);
6362 if (va < VM_MAXUSER_ADDRESS && pmap->pm_type == PT_X86) {
6363 newpde &= ~X86_PG_PKU_MASK;
6364 newpde |= pmap_pkru_get(pmap, va);
6368 * If there are existing mappings, either abort or remove them.
6371 if ((oldpde & PG_V) != 0) {
6372 KASSERT(pdpg == NULL || pdpg->ref_count > 1,
6373 ("pmap_enter_pde: pdpg's reference count is too low"));
6374 if ((flags & PMAP_ENTER_NOREPLACE) != 0 && (va <
6375 VM_MAXUSER_ADDRESS || (oldpde & PG_PS) != 0 ||
6376 !pmap_every_pte_zero(oldpde & PG_FRAME))) {
6379 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
6380 " in pmap %p", va, pmap);
6381 return (KERN_FAILURE);
6383 /* Break the existing mapping(s). */
6385 if ((oldpde & PG_PS) != 0) {
6387 * The reference to the PD page that was acquired by
6388 * pmap_alloc_pde() ensures that it won't be freed.
6389 * However, if the PDE resulted from a promotion, then
6390 * a reserved PT page could be freed.
6392 (void)pmap_remove_pde(pmap, pde, va, &free, lockp);
6393 if ((oldpde & PG_G) == 0)
6394 pmap_invalidate_pde_page(pmap, va, oldpde);
6396 pmap_delayed_invl_start();
6397 if (pmap_remove_ptes(pmap, va, va + NBPDR, pde, &free,
6399 pmap_invalidate_all(pmap);
6400 pmap_delayed_invl_finish();
6402 if (va < VM_MAXUSER_ADDRESS) {
6403 vm_page_free_pages_toq(&free, true);
6404 KASSERT(*pde == 0, ("pmap_enter_pde: non-zero pde %p",
6407 KASSERT(SLIST_EMPTY(&free),
6408 ("pmap_enter_pde: freed kernel page table page"));
6411 * Both pmap_remove_pde() and pmap_remove_ptes() will
6412 * leave the kernel page table page zero filled.
6414 mt = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
6415 if (pmap_insert_pt_page(pmap, mt, false))
6416 panic("pmap_enter_pde: trie insert failed");
6420 if ((newpde & PG_MANAGED) != 0) {
6422 * Abort this mapping if its PV entry could not be created.
6424 if (!pmap_pv_insert_pde(pmap, va, newpde, flags, lockp)) {
6426 pmap_abort_ptp(pmap, va, pdpg);
6427 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
6428 " in pmap %p", va, pmap);
6429 return (KERN_RESOURCE_SHORTAGE);
6431 if ((newpde & PG_RW) != 0) {
6432 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
6433 vm_page_aflag_set(mt, PGA_WRITEABLE);
6438 * Increment counters.
6440 if ((newpde & PG_W) != 0)
6441 pmap->pm_stats.wired_count += NBPDR / PAGE_SIZE;
6442 pmap_resident_count_inc(pmap, NBPDR / PAGE_SIZE);
6445 * Map the superpage. (This is not a promoted mapping; there will not
6446 * be any lingering 4KB page mappings in the TLB.)
6448 pde_store(pde, newpde);
6450 atomic_add_long(&pmap_pde_mappings, 1);
6451 CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx in pmap %p",
6453 return (KERN_SUCCESS);
6457 * Maps a sequence of resident pages belonging to the same object.
6458 * The sequence begins with the given page m_start. This page is
6459 * mapped at the given virtual address start. Each subsequent page is
6460 * mapped at a virtual address that is offset from start by the same
6461 * amount as the page is offset from m_start within the object. The
6462 * last page in the sequence is the page with the largest offset from
6463 * m_start that can be mapped at a virtual address less than the given
6464 * virtual address end. Not every virtual page between start and end
6465 * is mapped; only those for which a resident page exists with the
6466 * corresponding offset from m_start are mapped.
6469 pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end,
6470 vm_page_t m_start, vm_prot_t prot)
6472 struct rwlock *lock;
6475 vm_pindex_t diff, psize;
6477 VM_OBJECT_ASSERT_LOCKED(m_start->object);
6479 psize = atop(end - start);
6484 while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
6485 va = start + ptoa(diff);
6486 if ((va & PDRMASK) == 0 && va + NBPDR <= end &&
6487 m->psind == 1 && pmap_ps_enabled(pmap) &&
6488 pmap_allow_2m_x_page(pmap, (prot & VM_PROT_EXECUTE) != 0) &&
6489 pmap_enter_2mpage(pmap, va, m, prot, &lock))
6490 m = &m[NBPDR / PAGE_SIZE - 1];
6492 mpte = pmap_enter_quick_locked(pmap, va, m, prot,
6494 m = TAILQ_NEXT(m, listq);
6502 * this code makes some *MAJOR* assumptions:
6503 * 1. Current pmap & pmap exists.
6506 * 4. No page table pages.
6507 * but is *MUCH* faster than pmap_enter...
6511 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot)
6513 struct rwlock *lock;
6517 (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL, &lock);
6524 pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m,
6525 vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp)
6527 pt_entry_t newpte, *pte, PG_V;
6529 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva ||
6530 (m->oflags & VPO_UNMANAGED) != 0,
6531 ("pmap_enter_quick_locked: managed mapping within the clean submap"));
6532 PG_V = pmap_valid_bit(pmap);
6533 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
6536 * In the case that a page table page is not
6537 * resident, we are creating it here.
6539 if (va < VM_MAXUSER_ADDRESS) {
6540 vm_pindex_t ptepindex;
6544 * Calculate pagetable page index
6546 ptepindex = pmap_pde_pindex(va);
6547 if (mpte && (mpte->pindex == ptepindex)) {
6551 * Get the page directory entry
6553 ptepa = pmap_pde(pmap, va);
6556 * If the page table page is mapped, we just increment
6557 * the hold count, and activate it. Otherwise, we
6558 * attempt to allocate a page table page. If this
6559 * attempt fails, we don't retry. Instead, we give up.
6561 if (ptepa && (*ptepa & PG_V) != 0) {
6564 mpte = PHYS_TO_VM_PAGE(*ptepa & PG_FRAME);
6568 * Pass NULL instead of the PV list lock
6569 * pointer, because we don't intend to sleep.
6571 mpte = _pmap_allocpte(pmap, ptepindex, NULL);
6576 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte));
6577 pte = &pte[pmap_pte_index(va)];
6589 * Enter on the PV list if part of our managed memory.
6591 if ((m->oflags & VPO_UNMANAGED) == 0 &&
6592 !pmap_try_insert_pv_entry(pmap, va, m, lockp)) {
6594 pmap_abort_ptp(pmap, va, mpte);
6599 * Increment counters
6601 pmap_resident_count_inc(pmap, 1);
6603 newpte = VM_PAGE_TO_PHYS(m) | PG_V |
6604 pmap_cache_bits(pmap, m->md.pat_mode, 0);
6605 if ((m->oflags & VPO_UNMANAGED) == 0)
6606 newpte |= PG_MANAGED;
6607 if ((prot & VM_PROT_EXECUTE) == 0)
6609 if (va < VM_MAXUSER_ADDRESS)
6610 newpte |= PG_U | pmap_pkru_get(pmap, va);
6611 pte_store(pte, newpte);
6616 * Make a temporary mapping for a physical address. This is only intended
6617 * to be used for panic dumps.
6620 pmap_kenter_temporary(vm_paddr_t pa, int i)
6624 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
6625 pmap_kenter(va, pa);
6627 return ((void *)crashdumpmap);
6631 * This code maps large physical mmap regions into the
6632 * processor address space. Note that some shortcuts
6633 * are taken, but the code works.
6636 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object,
6637 vm_pindex_t pindex, vm_size_t size)
6640 pt_entry_t PG_A, PG_M, PG_RW, PG_V;
6641 vm_paddr_t pa, ptepa;
6645 PG_A = pmap_accessed_bit(pmap);
6646 PG_M = pmap_modified_bit(pmap);
6647 PG_V = pmap_valid_bit(pmap);
6648 PG_RW = pmap_rw_bit(pmap);
6650 VM_OBJECT_ASSERT_WLOCKED(object);
6651 KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG,
6652 ("pmap_object_init_pt: non-device object"));
6653 if ((addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) {
6654 if (!pmap_ps_enabled(pmap))
6656 if (!vm_object_populate(object, pindex, pindex + atop(size)))
6658 p = vm_page_lookup(object, pindex);
6659 KASSERT(p->valid == VM_PAGE_BITS_ALL,
6660 ("pmap_object_init_pt: invalid page %p", p));
6661 pat_mode = p->md.pat_mode;
6664 * Abort the mapping if the first page is not physically
6665 * aligned to a 2MB page boundary.
6667 ptepa = VM_PAGE_TO_PHYS(p);
6668 if (ptepa & (NBPDR - 1))
6672 * Skip the first page. Abort the mapping if the rest of
6673 * the pages are not physically contiguous or have differing
6674 * memory attributes.
6676 p = TAILQ_NEXT(p, listq);
6677 for (pa = ptepa + PAGE_SIZE; pa < ptepa + size;
6679 KASSERT(p->valid == VM_PAGE_BITS_ALL,
6680 ("pmap_object_init_pt: invalid page %p", p));
6681 if (pa != VM_PAGE_TO_PHYS(p) ||
6682 pat_mode != p->md.pat_mode)
6684 p = TAILQ_NEXT(p, listq);
6688 * Map using 2MB pages. Since "ptepa" is 2M aligned and
6689 * "size" is a multiple of 2M, adding the PAT setting to "pa"
6690 * will not affect the termination of this loop.
6693 for (pa = ptepa | pmap_cache_bits(pmap, pat_mode, 1);
6694 pa < ptepa + size; pa += NBPDR) {
6695 pde = pmap_alloc_pde(pmap, addr, &pdpg, NULL);
6698 * The creation of mappings below is only an
6699 * optimization. If a page directory page
6700 * cannot be allocated without blocking,
6701 * continue on to the next mapping rather than
6707 if ((*pde & PG_V) == 0) {
6708 pde_store(pde, pa | PG_PS | PG_M | PG_A |
6709 PG_U | PG_RW | PG_V);
6710 pmap_resident_count_inc(pmap, NBPDR / PAGE_SIZE);
6711 atomic_add_long(&pmap_pde_mappings, 1);
6713 /* Continue on if the PDE is already valid. */
6715 KASSERT(pdpg->ref_count > 0,
6716 ("pmap_object_init_pt: missing reference "
6717 "to page directory page, va: 0x%lx", addr));
6726 * Clear the wired attribute from the mappings for the specified range of
6727 * addresses in the given pmap. Every valid mapping within that range
6728 * must have the wired attribute set. In contrast, invalid mappings
6729 * cannot have the wired attribute set, so they are ignored.
6731 * The wired attribute of the page table entry is not a hardware
6732 * feature, so there is no need to invalidate any TLB entries.
6733 * Since pmap_demote_pde() for the wired entry must never fail,
6734 * pmap_delayed_invl_start()/finish() calls around the
6735 * function are not needed.
6738 pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
6740 vm_offset_t va_next;
6741 pml4_entry_t *pml4e;
6744 pt_entry_t *pte, PG_V;
6746 PG_V = pmap_valid_bit(pmap);
6748 for (; sva < eva; sva = va_next) {
6749 pml4e = pmap_pml4e(pmap, sva);
6750 if ((*pml4e & PG_V) == 0) {
6751 va_next = (sva + NBPML4) & ~PML4MASK;
6756 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
6757 if ((*pdpe & PG_V) == 0) {
6758 va_next = (sva + NBPDP) & ~PDPMASK;
6763 va_next = (sva + NBPDR) & ~PDRMASK;
6766 pde = pmap_pdpe_to_pde(pdpe, sva);
6767 if ((*pde & PG_V) == 0)
6769 if ((*pde & PG_PS) != 0) {
6770 if ((*pde & PG_W) == 0)
6771 panic("pmap_unwire: pde %#jx is missing PG_W",
6775 * Are we unwiring the entire large page? If not,
6776 * demote the mapping and fall through.
6778 if (sva + NBPDR == va_next && eva >= va_next) {
6779 atomic_clear_long(pde, PG_W);
6780 pmap->pm_stats.wired_count -= NBPDR /
6783 } else if (!pmap_demote_pde(pmap, pde, sva))
6784 panic("pmap_unwire: demotion failed");
6788 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
6790 if ((*pte & PG_V) == 0)
6792 if ((*pte & PG_W) == 0)
6793 panic("pmap_unwire: pte %#jx is missing PG_W",
6797 * PG_W must be cleared atomically. Although the pmap
6798 * lock synchronizes access to PG_W, another processor
6799 * could be setting PG_M and/or PG_A concurrently.
6801 atomic_clear_long(pte, PG_W);
6802 pmap->pm_stats.wired_count--;
6809 * Copy the range specified by src_addr/len
6810 * from the source map to the range dst_addr/len
6811 * in the destination map.
6813 * This routine is only advisory and need not do anything.
6816 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
6817 vm_offset_t src_addr)
6819 struct rwlock *lock;
6820 pml4_entry_t *pml4e;
6822 pd_entry_t *pde, srcptepaddr;
6823 pt_entry_t *dst_pte, PG_A, PG_M, PG_V, ptetemp, *src_pte;
6824 vm_offset_t addr, end_addr, va_next;
6825 vm_page_t dst_pdpg, dstmpte, srcmpte;
6827 if (dst_addr != src_addr)
6830 if (dst_pmap->pm_type != src_pmap->pm_type)
6834 * EPT page table entries that require emulation of A/D bits are
6835 * sensitive to clearing the PG_A bit (aka EPT_PG_READ). Although
6836 * we clear PG_M (aka EPT_PG_WRITE) concomitantly, the PG_U bit
6837 * (aka EPT_PG_EXECUTE) could still be set. Since some EPT
6838 * implementations flag an EPT misconfiguration for exec-only
6839 * mappings we skip this function entirely for emulated pmaps.
6841 if (pmap_emulate_ad_bits(dst_pmap))
6844 end_addr = src_addr + len;
6846 if (dst_pmap < src_pmap) {
6847 PMAP_LOCK(dst_pmap);
6848 PMAP_LOCK(src_pmap);
6850 PMAP_LOCK(src_pmap);
6851 PMAP_LOCK(dst_pmap);
6854 PG_A = pmap_accessed_bit(dst_pmap);
6855 PG_M = pmap_modified_bit(dst_pmap);
6856 PG_V = pmap_valid_bit(dst_pmap);
6858 for (addr = src_addr; addr < end_addr; addr = va_next) {
6859 KASSERT(addr < UPT_MIN_ADDRESS,
6860 ("pmap_copy: invalid to pmap_copy page tables"));
6862 pml4e = pmap_pml4e(src_pmap, addr);
6863 if ((*pml4e & PG_V) == 0) {
6864 va_next = (addr + NBPML4) & ~PML4MASK;
6870 pdpe = pmap_pml4e_to_pdpe(pml4e, addr);
6871 if ((*pdpe & PG_V) == 0) {
6872 va_next = (addr + NBPDP) & ~PDPMASK;
6878 va_next = (addr + NBPDR) & ~PDRMASK;
6882 pde = pmap_pdpe_to_pde(pdpe, addr);
6884 if (srcptepaddr == 0)
6887 if (srcptepaddr & PG_PS) {
6888 if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr)
6890 pde = pmap_alloc_pde(dst_pmap, addr, &dst_pdpg, NULL);
6893 if (*pde == 0 && ((srcptepaddr & PG_MANAGED) == 0 ||
6894 pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr,
6895 PMAP_ENTER_NORECLAIM, &lock))) {
6896 *pde = srcptepaddr & ~PG_W;
6897 pmap_resident_count_inc(dst_pmap, NBPDR /
6899 atomic_add_long(&pmap_pde_mappings, 1);
6901 pmap_abort_ptp(dst_pmap, addr, dst_pdpg);
6905 srcptepaddr &= PG_FRAME;
6906 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr);
6907 KASSERT(srcmpte->ref_count > 0,
6908 ("pmap_copy: source page table page is unused"));
6910 if (va_next > end_addr)
6913 src_pte = (pt_entry_t *)PHYS_TO_DMAP(srcptepaddr);
6914 src_pte = &src_pte[pmap_pte_index(addr)];
6916 for (; addr < va_next; addr += PAGE_SIZE, src_pte++) {
6920 * We only virtual copy managed pages.
6922 if ((ptetemp & PG_MANAGED) == 0)
6925 if (dstmpte != NULL) {
6926 KASSERT(dstmpte->pindex ==
6927 pmap_pde_pindex(addr),
6928 ("dstmpte pindex/addr mismatch"));
6929 dstmpte->ref_count++;
6930 } else if ((dstmpte = pmap_allocpte(dst_pmap, addr,
6933 dst_pte = (pt_entry_t *)
6934 PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dstmpte));
6935 dst_pte = &dst_pte[pmap_pte_index(addr)];
6936 if (*dst_pte == 0 &&
6937 pmap_try_insert_pv_entry(dst_pmap, addr,
6938 PHYS_TO_VM_PAGE(ptetemp & PG_FRAME), &lock)) {
6940 * Clear the wired, modified, and accessed
6941 * (referenced) bits during the copy.
6943 *dst_pte = ptetemp & ~(PG_W | PG_M | PG_A);
6944 pmap_resident_count_inc(dst_pmap, 1);
6946 pmap_abort_ptp(dst_pmap, addr, dstmpte);
6949 /* Have we copied all of the valid mappings? */
6950 if (dstmpte->ref_count >= srcmpte->ref_count)
6957 PMAP_UNLOCK(src_pmap);
6958 PMAP_UNLOCK(dst_pmap);
6962 pmap_vmspace_copy(pmap_t dst_pmap, pmap_t src_pmap)
6966 if (dst_pmap->pm_type != src_pmap->pm_type ||
6967 dst_pmap->pm_type != PT_X86 ||
6968 (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0)
6971 if (dst_pmap < src_pmap) {
6972 PMAP_LOCK(dst_pmap);
6973 PMAP_LOCK(src_pmap);
6975 PMAP_LOCK(src_pmap);
6976 PMAP_LOCK(dst_pmap);
6978 error = pmap_pkru_copy(dst_pmap, src_pmap);
6979 /* Clean up partial copy on failure due to no memory. */
6980 if (error == ENOMEM)
6981 pmap_pkru_deassign_all(dst_pmap);
6982 PMAP_UNLOCK(src_pmap);
6983 PMAP_UNLOCK(dst_pmap);
6984 if (error != ENOMEM)
6992 * Zero the specified hardware page.
6995 pmap_zero_page(vm_page_t m)
6997 vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
6999 pagezero((void *)va);
7003 * Zero an an area within a single hardware page. off and size must not
7004 * cover an area beyond a single hardware page.
7007 pmap_zero_page_area(vm_page_t m, int off, int size)
7009 vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
7011 if (off == 0 && size == PAGE_SIZE)
7012 pagezero((void *)va);
7014 bzero((char *)va + off, size);
7018 * Copy 1 specified hardware page to another.
7021 pmap_copy_page(vm_page_t msrc, vm_page_t mdst)
7023 vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc));
7024 vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst));
7026 pagecopy((void *)src, (void *)dst);
7029 int unmapped_buf_allowed = 1;
7032 pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[],
7033 vm_offset_t b_offset, int xfersize)
7037 vm_offset_t vaddr[2], a_pg_offset, b_pg_offset;
7041 while (xfersize > 0) {
7042 a_pg_offset = a_offset & PAGE_MASK;
7043 pages[0] = ma[a_offset >> PAGE_SHIFT];
7044 b_pg_offset = b_offset & PAGE_MASK;
7045 pages[1] = mb[b_offset >> PAGE_SHIFT];
7046 cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
7047 cnt = min(cnt, PAGE_SIZE - b_pg_offset);
7048 mapped = pmap_map_io_transient(pages, vaddr, 2, FALSE);
7049 a_cp = (char *)vaddr[0] + a_pg_offset;
7050 b_cp = (char *)vaddr[1] + b_pg_offset;
7051 bcopy(a_cp, b_cp, cnt);
7052 if (__predict_false(mapped))
7053 pmap_unmap_io_transient(pages, vaddr, 2, FALSE);
7061 * Returns true if the pmap's pv is one of the first
7062 * 16 pvs linked to from this page. This count may
7063 * be changed upwards or downwards in the future; it
7064 * is only necessary that true be returned for a small
7065 * subset of pmaps for proper page aging.
7068 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
7070 struct md_page *pvh;
7071 struct rwlock *lock;
7076 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
7077 ("pmap_page_exists_quick: page %p is not managed", m));
7079 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
7081 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
7082 if (PV_PMAP(pv) == pmap) {
7090 if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) {
7091 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
7092 TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
7093 if (PV_PMAP(pv) == pmap) {
7107 * pmap_page_wired_mappings:
7109 * Return the number of managed mappings to the given physical page
7113 pmap_page_wired_mappings(vm_page_t m)
7115 struct rwlock *lock;
7116 struct md_page *pvh;
7120 int count, md_gen, pvh_gen;
7122 if ((m->oflags & VPO_UNMANAGED) != 0)
7124 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
7128 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
7130 if (!PMAP_TRYLOCK(pmap)) {
7131 md_gen = m->md.pv_gen;
7135 if (md_gen != m->md.pv_gen) {
7140 pte = pmap_pte(pmap, pv->pv_va);
7141 if ((*pte & PG_W) != 0)
7145 if ((m->flags & PG_FICTITIOUS) == 0) {
7146 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
7147 TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
7149 if (!PMAP_TRYLOCK(pmap)) {
7150 md_gen = m->md.pv_gen;
7151 pvh_gen = pvh->pv_gen;
7155 if (md_gen != m->md.pv_gen ||
7156 pvh_gen != pvh->pv_gen) {
7161 pte = pmap_pde(pmap, pv->pv_va);
7162 if ((*pte & PG_W) != 0)
7172 * Returns TRUE if the given page is mapped individually or as part of
7173 * a 2mpage. Otherwise, returns FALSE.
7176 pmap_page_is_mapped(vm_page_t m)
7178 struct rwlock *lock;
7181 if ((m->oflags & VPO_UNMANAGED) != 0)
7183 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
7185 rv = !TAILQ_EMPTY(&m->md.pv_list) ||
7186 ((m->flags & PG_FICTITIOUS) == 0 &&
7187 !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list));
7193 * Destroy all managed, non-wired mappings in the given user-space
7194 * pmap. This pmap cannot be active on any processor besides the
7197 * This function cannot be applied to the kernel pmap. Moreover, it
7198 * is not intended for general use. It is only to be used during
7199 * process termination. Consequently, it can be implemented in ways
7200 * that make it faster than pmap_remove(). First, it can more quickly
7201 * destroy mappings by iterating over the pmap's collection of PV
7202 * entries, rather than searching the page table. Second, it doesn't
7203 * have to test and clear the page table entries atomically, because
7204 * no processor is currently accessing the user address space. In
7205 * particular, a page table entry's dirty bit won't change state once
7206 * this function starts.
7208 * Although this function destroys all of the pmap's managed,
7209 * non-wired mappings, it can delay and batch the invalidation of TLB
7210 * entries without calling pmap_delayed_invl_start() and
7211 * pmap_delayed_invl_finish(). Because the pmap is not active on
7212 * any other processor, none of these TLB entries will ever be used
7213 * before their eventual invalidation. Consequently, there is no need
7214 * for either pmap_remove_all() or pmap_remove_write() to wait for
7215 * that eventual TLB invalidation.
7218 pmap_remove_pages(pmap_t pmap)
7221 pt_entry_t *pte, tpte;
7222 pt_entry_t PG_M, PG_RW, PG_V;
7223 struct spglist free;
7224 struct pv_chunklist free_chunks[PMAP_MEMDOM];
7225 vm_page_t m, mpte, mt;
7227 struct md_page *pvh;
7228 struct pv_chunk *pc, *npc;
7229 struct rwlock *lock;
7231 uint64_t inuse, bitmask;
7232 int allfree, field, freed, i, idx;
7233 boolean_t superpage;
7237 * Assert that the given pmap is only active on the current
7238 * CPU. Unfortunately, we cannot block another CPU from
7239 * activating the pmap while this function is executing.
7241 KASSERT(pmap == PCPU_GET(curpmap), ("non-current pmap %p", pmap));
7244 cpuset_t other_cpus;
7246 other_cpus = all_cpus;
7248 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
7249 CPU_AND(&other_cpus, &pmap->pm_active);
7251 KASSERT(CPU_EMPTY(&other_cpus), ("pmap active %p", pmap));
7256 PG_M = pmap_modified_bit(pmap);
7257 PG_V = pmap_valid_bit(pmap);
7258 PG_RW = pmap_rw_bit(pmap);
7260 for (i = 0; i < PMAP_MEMDOM; i++)
7261 TAILQ_INIT(&free_chunks[i]);
7264 TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) {
7267 for (field = 0; field < _NPCM; field++) {
7268 inuse = ~pc->pc_map[field] & pc_freemask[field];
7269 while (inuse != 0) {
7271 bitmask = 1UL << bit;
7272 idx = field * 64 + bit;
7273 pv = &pc->pc_pventry[idx];
7276 pte = pmap_pdpe(pmap, pv->pv_va);
7278 pte = pmap_pdpe_to_pde(pte, pv->pv_va);
7280 if ((tpte & (PG_PS | PG_V)) == PG_V) {
7283 pte = (pt_entry_t *)PHYS_TO_DMAP(tpte &
7285 pte = &pte[pmap_pte_index(pv->pv_va)];
7289 * Keep track whether 'tpte' is a
7290 * superpage explicitly instead of
7291 * relying on PG_PS being set.
7293 * This is because PG_PS is numerically
7294 * identical to PG_PTE_PAT and thus a
7295 * regular page could be mistaken for
7301 if ((tpte & PG_V) == 0) {
7302 panic("bad pte va %lx pte %lx",
7307 * We cannot remove wired pages from a process' mapping at this time
7315 pa = tpte & PG_PS_FRAME;
7317 pa = tpte & PG_FRAME;
7319 m = PHYS_TO_VM_PAGE(pa);
7320 KASSERT(m->phys_addr == pa,
7321 ("vm_page_t %p phys_addr mismatch %016jx %016jx",
7322 m, (uintmax_t)m->phys_addr,
7325 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
7326 m < &vm_page_array[vm_page_array_size],
7327 ("pmap_remove_pages: bad tpte %#jx",
7333 * Update the vm_page_t clean/reference bits.
7335 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
7337 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
7343 CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m);
7346 pc->pc_map[field] |= bitmask;
7348 pmap_resident_count_dec(pmap, NBPDR / PAGE_SIZE);
7349 pvh = pa_to_pvh(tpte & PG_PS_FRAME);
7350 TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
7352 if (TAILQ_EMPTY(&pvh->pv_list)) {
7353 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
7354 if ((mt->a.flags & PGA_WRITEABLE) != 0 &&
7355 TAILQ_EMPTY(&mt->md.pv_list))
7356 vm_page_aflag_clear(mt, PGA_WRITEABLE);
7358 mpte = pmap_remove_pt_page(pmap, pv->pv_va);
7360 KASSERT(mpte->valid == VM_PAGE_BITS_ALL,
7361 ("pmap_remove_pages: pte page not promoted"));
7362 pmap_resident_count_dec(pmap, 1);
7363 KASSERT(mpte->ref_count == NPTEPG,
7364 ("pmap_remove_pages: pte page reference count error"));
7365 mpte->ref_count = 0;
7366 pmap_add_delayed_free_list(mpte, &free, FALSE);
7369 pmap_resident_count_dec(pmap, 1);
7370 TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
7372 if ((m->a.flags & PGA_WRITEABLE) != 0 &&
7373 TAILQ_EMPTY(&m->md.pv_list) &&
7374 (m->flags & PG_FICTITIOUS) == 0) {
7375 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
7376 if (TAILQ_EMPTY(&pvh->pv_list))
7377 vm_page_aflag_clear(m, PGA_WRITEABLE);
7380 pmap_unuse_pt(pmap, pv->pv_va, ptepde, &free);
7384 PV_STAT(atomic_add_long(&pv_entry_frees, freed));
7385 PV_STAT(atomic_add_int(&pv_entry_spare, freed));
7386 PV_STAT(atomic_subtract_long(&pv_entry_count, freed));
7388 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
7389 TAILQ_INSERT_TAIL(&free_chunks[pc_to_domain(pc)], pc, pc_list);
7394 pmap_invalidate_all(pmap);
7395 pmap_pkru_deassign_all(pmap);
7396 free_pv_chunk_batch((struct pv_chunklist *)&free_chunks);
7398 vm_page_free_pages_toq(&free, true);
7402 pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified)
7404 struct rwlock *lock;
7406 struct md_page *pvh;
7407 pt_entry_t *pte, mask;
7408 pt_entry_t PG_A, PG_M, PG_RW, PG_V;
7410 int md_gen, pvh_gen;
7414 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
7417 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
7419 if (!PMAP_TRYLOCK(pmap)) {
7420 md_gen = m->md.pv_gen;
7424 if (md_gen != m->md.pv_gen) {
7429 pte = pmap_pte(pmap, pv->pv_va);
7432 PG_M = pmap_modified_bit(pmap);
7433 PG_RW = pmap_rw_bit(pmap);
7434 mask |= PG_RW | PG_M;
7437 PG_A = pmap_accessed_bit(pmap);
7438 PG_V = pmap_valid_bit(pmap);
7439 mask |= PG_V | PG_A;
7441 rv = (*pte & mask) == mask;
7446 if ((m->flags & PG_FICTITIOUS) == 0) {
7447 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
7448 TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
7450 if (!PMAP_TRYLOCK(pmap)) {
7451 md_gen = m->md.pv_gen;
7452 pvh_gen = pvh->pv_gen;
7456 if (md_gen != m->md.pv_gen ||
7457 pvh_gen != pvh->pv_gen) {
7462 pte = pmap_pde(pmap, pv->pv_va);
7465 PG_M = pmap_modified_bit(pmap);
7466 PG_RW = pmap_rw_bit(pmap);
7467 mask |= PG_RW | PG_M;
7470 PG_A = pmap_accessed_bit(pmap);
7471 PG_V = pmap_valid_bit(pmap);
7472 mask |= PG_V | PG_A;
7474 rv = (*pte & mask) == mask;
7488 * Return whether or not the specified physical page was modified
7489 * in any physical maps.
7492 pmap_is_modified(vm_page_t m)
7495 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
7496 ("pmap_is_modified: page %p is not managed", m));
7499 * If the page is not busied then this check is racy.
7501 if (!pmap_page_is_write_mapped(m))
7503 return (pmap_page_test_mappings(m, FALSE, TRUE));
7507 * pmap_is_prefaultable:
7509 * Return whether or not the specified virtual address is eligible
7513 pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr)
7516 pt_entry_t *pte, PG_V;
7519 PG_V = pmap_valid_bit(pmap);
7522 pde = pmap_pde(pmap, addr);
7523 if (pde != NULL && (*pde & (PG_PS | PG_V)) == PG_V) {
7524 pte = pmap_pde_to_pte(pde, addr);
7525 rv = (*pte & PG_V) == 0;
7532 * pmap_is_referenced:
7534 * Return whether or not the specified physical page was referenced
7535 * in any physical maps.
7538 pmap_is_referenced(vm_page_t m)
7541 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
7542 ("pmap_is_referenced: page %p is not managed", m));
7543 return (pmap_page_test_mappings(m, TRUE, FALSE));
7547 * Clear the write and modified bits in each of the given page's mappings.
7550 pmap_remove_write(vm_page_t m)
7552 struct md_page *pvh;
7554 struct rwlock *lock;
7555 pv_entry_t next_pv, pv;
7557 pt_entry_t oldpte, *pte, PG_M, PG_RW;
7559 int pvh_gen, md_gen;
7561 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
7562 ("pmap_remove_write: page %p is not managed", m));
7564 vm_page_assert_busied(m);
7565 if (!pmap_page_is_write_mapped(m))
7568 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
7569 pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
7570 pa_to_pvh(VM_PAGE_TO_PHYS(m));
7573 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) {
7575 if (!PMAP_TRYLOCK(pmap)) {
7576 pvh_gen = pvh->pv_gen;
7580 if (pvh_gen != pvh->pv_gen) {
7586 PG_RW = pmap_rw_bit(pmap);
7588 pde = pmap_pde(pmap, va);
7589 if ((*pde & PG_RW) != 0)
7590 (void)pmap_demote_pde_locked(pmap, pde, va, &lock);
7591 KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m),
7592 ("inconsistent pv lock %p %p for page %p",
7593 lock, VM_PAGE_TO_PV_LIST_LOCK(m), m));
7596 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
7598 if (!PMAP_TRYLOCK(pmap)) {
7599 pvh_gen = pvh->pv_gen;
7600 md_gen = m->md.pv_gen;
7604 if (pvh_gen != pvh->pv_gen ||
7605 md_gen != m->md.pv_gen) {
7611 PG_M = pmap_modified_bit(pmap);
7612 PG_RW = pmap_rw_bit(pmap);
7613 pde = pmap_pde(pmap, pv->pv_va);
7614 KASSERT((*pde & PG_PS) == 0,
7615 ("pmap_remove_write: found a 2mpage in page %p's pv list",
7617 pte = pmap_pde_to_pte(pde, pv->pv_va);
7620 if (oldpte & PG_RW) {
7621 if (!atomic_cmpset_long(pte, oldpte, oldpte &
7624 if ((oldpte & PG_M) != 0)
7626 pmap_invalidate_page(pmap, pv->pv_va);
7631 vm_page_aflag_clear(m, PGA_WRITEABLE);
7632 pmap_delayed_invl_wait(m);
7635 static __inline boolean_t
7636 safe_to_clear_referenced(pmap_t pmap, pt_entry_t pte)
7639 if (!pmap_emulate_ad_bits(pmap))
7642 KASSERT(pmap->pm_type == PT_EPT, ("invalid pm_type %d", pmap->pm_type));
7645 * XWR = 010 or 110 will cause an unconditional EPT misconfiguration
7646 * so we don't let the referenced (aka EPT_PG_READ) bit to be cleared
7647 * if the EPT_PG_WRITE bit is set.
7649 if ((pte & EPT_PG_WRITE) != 0)
7653 * XWR = 100 is allowed only if the PMAP_SUPPORTS_EXEC_ONLY is set.
7655 if ((pte & EPT_PG_EXECUTE) == 0 ||
7656 ((pmap->pm_flags & PMAP_SUPPORTS_EXEC_ONLY) != 0))
7663 * pmap_ts_referenced:
7665 * Return a count of reference bits for a page, clearing those bits.
7666 * It is not necessary for every reference bit to be cleared, but it
7667 * is necessary that 0 only be returned when there are truly no
7668 * reference bits set.
7670 * As an optimization, update the page's dirty field if a modified bit is
7671 * found while counting reference bits. This opportunistic update can be
7672 * performed at low cost and can eliminate the need for some future calls
7673 * to pmap_is_modified(). However, since this function stops after
7674 * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some
7675 * dirty pages. Those dirty pages will only be detected by a future call
7676 * to pmap_is_modified().
7678 * A DI block is not needed within this function, because
7679 * invalidations are performed before the PV list lock is
7683 pmap_ts_referenced(vm_page_t m)
7685 struct md_page *pvh;
7688 struct rwlock *lock;
7689 pd_entry_t oldpde, *pde;
7690 pt_entry_t *pte, PG_A, PG_M, PG_RW;
7693 int cleared, md_gen, not_cleared, pvh_gen;
7694 struct spglist free;
7697 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
7698 ("pmap_ts_referenced: page %p is not managed", m));
7701 pa = VM_PAGE_TO_PHYS(m);
7702 lock = PHYS_TO_PV_LIST_LOCK(pa);
7703 pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(pa);
7707 if ((pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL)
7708 goto small_mappings;
7714 if (!PMAP_TRYLOCK(pmap)) {
7715 pvh_gen = pvh->pv_gen;
7719 if (pvh_gen != pvh->pv_gen) {
7724 PG_A = pmap_accessed_bit(pmap);
7725 PG_M = pmap_modified_bit(pmap);
7726 PG_RW = pmap_rw_bit(pmap);
7728 pde = pmap_pde(pmap, pv->pv_va);
7730 if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
7732 * Although "oldpde" is mapping a 2MB page, because
7733 * this function is called at a 4KB page granularity,
7734 * we only update the 4KB page under test.
7738 if ((oldpde & PG_A) != 0) {
7740 * Since this reference bit is shared by 512 4KB
7741 * pages, it should not be cleared every time it is
7742 * tested. Apply a simple "hash" function on the
7743 * physical page number, the virtual superpage number,
7744 * and the pmap address to select one 4KB page out of
7745 * the 512 on which testing the reference bit will
7746 * result in clearing that reference bit. This
7747 * function is designed to avoid the selection of the
7748 * same 4KB page for every 2MB page mapping.
7750 * On demotion, a mapping that hasn't been referenced
7751 * is simply destroyed. To avoid the possibility of a
7752 * subsequent page fault on a demoted wired mapping,
7753 * always leave its reference bit set. Moreover,
7754 * since the superpage is wired, the current state of
7755 * its reference bit won't affect page replacement.
7757 if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PDRSHIFT) ^
7758 (uintptr_t)pmap) & (NPTEPG - 1)) == 0 &&
7759 (oldpde & PG_W) == 0) {
7760 if (safe_to_clear_referenced(pmap, oldpde)) {
7761 atomic_clear_long(pde, PG_A);
7762 pmap_invalidate_page(pmap, pv->pv_va);
7764 } else if (pmap_demote_pde_locked(pmap, pde,
7765 pv->pv_va, &lock)) {
7767 * Remove the mapping to a single page
7768 * so that a subsequent access may
7769 * repromote. Since the underlying
7770 * page table page is fully populated,
7771 * this removal never frees a page
7775 va += VM_PAGE_TO_PHYS(m) - (oldpde &
7777 pte = pmap_pde_to_pte(pde, va);
7778 pmap_remove_pte(pmap, pte, va, *pde,
7780 pmap_invalidate_page(pmap, va);
7786 * The superpage mapping was removed
7787 * entirely and therefore 'pv' is no
7795 KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m),
7796 ("inconsistent pv lock %p %p for page %p",
7797 lock, VM_PAGE_TO_PV_LIST_LOCK(m), m));
7802 /* Rotate the PV list if it has more than one entry. */
7803 if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) {
7804 TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
7805 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
7808 if (cleared + not_cleared >= PMAP_TS_REFERENCED_MAX)
7810 } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf);
7812 if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL)
7819 if (!PMAP_TRYLOCK(pmap)) {
7820 pvh_gen = pvh->pv_gen;
7821 md_gen = m->md.pv_gen;
7825 if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
7830 PG_A = pmap_accessed_bit(pmap);
7831 PG_M = pmap_modified_bit(pmap);
7832 PG_RW = pmap_rw_bit(pmap);
7833 pde = pmap_pde(pmap, pv->pv_va);
7834 KASSERT((*pde & PG_PS) == 0,
7835 ("pmap_ts_referenced: found a 2mpage in page %p's pv list",
7837 pte = pmap_pde_to_pte(pde, pv->pv_va);
7838 if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW))
7840 if ((*pte & PG_A) != 0) {
7841 if (safe_to_clear_referenced(pmap, *pte)) {
7842 atomic_clear_long(pte, PG_A);
7843 pmap_invalidate_page(pmap, pv->pv_va);
7845 } else if ((*pte & PG_W) == 0) {
7847 * Wired pages cannot be paged out so
7848 * doing accessed bit emulation for
7849 * them is wasted effort. We do the
7850 * hard work for unwired pages only.
7852 pmap_remove_pte(pmap, pte, pv->pv_va,
7853 *pde, &free, &lock);
7854 pmap_invalidate_page(pmap, pv->pv_va);
7859 KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m),
7860 ("inconsistent pv lock %p %p for page %p",
7861 lock, VM_PAGE_TO_PV_LIST_LOCK(m), m));
7866 /* Rotate the PV list if it has more than one entry. */
7867 if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) {
7868 TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
7869 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
7872 } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared +
7873 not_cleared < PMAP_TS_REFERENCED_MAX);
7876 vm_page_free_pages_toq(&free, true);
7877 return (cleared + not_cleared);
7881 * Apply the given advice to the specified range of addresses within the
7882 * given pmap. Depending on the advice, clear the referenced and/or
7883 * modified flags in each mapping and set the mapped page's dirty field.
7886 pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice)
7888 struct rwlock *lock;
7889 pml4_entry_t *pml4e;
7891 pd_entry_t oldpde, *pde;
7892 pt_entry_t *pte, PG_A, PG_G, PG_M, PG_RW, PG_V;
7893 vm_offset_t va, va_next;
7897 if (advice != MADV_DONTNEED && advice != MADV_FREE)
7901 * A/D bit emulation requires an alternate code path when clearing
7902 * the modified and accessed bits below. Since this function is
7903 * advisory in nature we skip it entirely for pmaps that require
7904 * A/D bit emulation.
7906 if (pmap_emulate_ad_bits(pmap))
7909 PG_A = pmap_accessed_bit(pmap);
7910 PG_G = pmap_global_bit(pmap);
7911 PG_M = pmap_modified_bit(pmap);
7912 PG_V = pmap_valid_bit(pmap);
7913 PG_RW = pmap_rw_bit(pmap);
7915 pmap_delayed_invl_start();
7917 for (; sva < eva; sva = va_next) {
7918 pml4e = pmap_pml4e(pmap, sva);
7919 if ((*pml4e & PG_V) == 0) {
7920 va_next = (sva + NBPML4) & ~PML4MASK;
7925 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
7926 if ((*pdpe & PG_V) == 0) {
7927 va_next = (sva + NBPDP) & ~PDPMASK;
7932 va_next = (sva + NBPDR) & ~PDRMASK;
7935 pde = pmap_pdpe_to_pde(pdpe, sva);
7937 if ((oldpde & PG_V) == 0)
7939 else if ((oldpde & PG_PS) != 0) {
7940 if ((oldpde & PG_MANAGED) == 0)
7943 if (!pmap_demote_pde_locked(pmap, pde, sva, &lock)) {
7948 * The large page mapping was destroyed.
7954 * Unless the page mappings are wired, remove the
7955 * mapping to a single page so that a subsequent
7956 * access may repromote. Choosing the last page
7957 * within the address range [sva, min(va_next, eva))
7958 * generally results in more repromotions. Since the
7959 * underlying page table page is fully populated, this
7960 * removal never frees a page table page.
7962 if ((oldpde & PG_W) == 0) {
7968 ("pmap_advise: no address gap"));
7969 pte = pmap_pde_to_pte(pde, va);
7970 KASSERT((*pte & PG_V) != 0,
7971 ("pmap_advise: invalid PTE"));
7972 pmap_remove_pte(pmap, pte, va, *pde, NULL,
7982 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
7984 if ((*pte & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V))
7986 else if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
7987 if (advice == MADV_DONTNEED) {
7989 * Future calls to pmap_is_modified()
7990 * can be avoided by making the page
7993 m = PHYS_TO_VM_PAGE(*pte & PG_FRAME);
7996 atomic_clear_long(pte, PG_M | PG_A);
7997 } else if ((*pte & PG_A) != 0)
7998 atomic_clear_long(pte, PG_A);
8002 if ((*pte & PG_G) != 0) {
8009 if (va != va_next) {
8010 pmap_invalidate_range(pmap, va, sva);
8015 pmap_invalidate_range(pmap, va, sva);
8018 pmap_invalidate_all(pmap);
8020 pmap_delayed_invl_finish();
8024 * Clear the modify bits on the specified physical page.
8027 pmap_clear_modify(vm_page_t m)
8029 struct md_page *pvh;
8031 pv_entry_t next_pv, pv;
8032 pd_entry_t oldpde, *pde;
8033 pt_entry_t *pte, PG_M, PG_RW;
8034 struct rwlock *lock;
8036 int md_gen, pvh_gen;
8038 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
8039 ("pmap_clear_modify: page %p is not managed", m));
8040 vm_page_assert_busied(m);
8042 if (!pmap_page_is_write_mapped(m))
8044 pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
8045 pa_to_pvh(VM_PAGE_TO_PHYS(m));
8046 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
8049 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) {
8051 if (!PMAP_TRYLOCK(pmap)) {
8052 pvh_gen = pvh->pv_gen;
8056 if (pvh_gen != pvh->pv_gen) {
8061 PG_M = pmap_modified_bit(pmap);
8062 PG_RW = pmap_rw_bit(pmap);
8064 pde = pmap_pde(pmap, va);
8066 /* If oldpde has PG_RW set, then it also has PG_M set. */
8067 if ((oldpde & PG_RW) != 0 &&
8068 pmap_demote_pde_locked(pmap, pde, va, &lock) &&
8069 (oldpde & PG_W) == 0) {
8071 * Write protect the mapping to a single page so that
8072 * a subsequent write access may repromote.
8074 va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME);
8075 pte = pmap_pde_to_pte(pde, va);
8076 atomic_clear_long(pte, PG_M | PG_RW);
8078 pmap_invalidate_page(pmap, va);
8082 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
8084 if (!PMAP_TRYLOCK(pmap)) {
8085 md_gen = m->md.pv_gen;
8086 pvh_gen = pvh->pv_gen;
8090 if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
8095 PG_M = pmap_modified_bit(pmap);
8096 PG_RW = pmap_rw_bit(pmap);
8097 pde = pmap_pde(pmap, pv->pv_va);
8098 KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found"
8099 " a 2mpage in page %p's pv list", m));
8100 pte = pmap_pde_to_pte(pde, pv->pv_va);
8101 if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
8102 atomic_clear_long(pte, PG_M);
8103 pmap_invalidate_page(pmap, pv->pv_va);
8111 * Miscellaneous support routines follow
8114 /* Adjust the properties for a leaf page table entry. */
8115 static __inline void
8116 pmap_pte_props(pt_entry_t *pte, u_long bits, u_long mask)
8120 opte = *(u_long *)pte;
8122 npte = opte & ~mask;
8124 } while (npte != opte && !atomic_fcmpset_long((u_long *)pte, &opte,
8129 * Map a set of physical memory pages into the kernel virtual
8130 * address space. Return a pointer to where it is mapped. This
8131 * routine is intended to be used for mapping device memory,
8135 pmap_mapdev_internal(vm_paddr_t pa, vm_size_t size, int mode, int flags)
8137 struct pmap_preinit_mapping *ppim;
8138 vm_offset_t va, offset;
8142 offset = pa & PAGE_MASK;
8143 size = round_page(offset + size);
8144 pa = trunc_page(pa);
8146 if (!pmap_initialized) {
8148 for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
8149 ppim = pmap_preinit_mapping + i;
8150 if (ppim->va == 0) {
8154 ppim->va = virtual_avail;
8155 virtual_avail += size;
8161 panic("%s: too many preinit mappings", __func__);
8164 * If we have a preinit mapping, re-use it.
8166 for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
8167 ppim = pmap_preinit_mapping + i;
8168 if (ppim->pa == pa && ppim->sz == size &&
8169 (ppim->mode == mode ||
8170 (flags & MAPDEV_SETATTR) == 0))
8171 return ((void *)(ppim->va + offset));
8174 * If the specified range of physical addresses fits within
8175 * the direct map window, use the direct map.
8177 if (pa < dmaplimit && pa + size <= dmaplimit) {
8178 va = PHYS_TO_DMAP(pa);
8179 if ((flags & MAPDEV_SETATTR) != 0) {
8180 PMAP_LOCK(kernel_pmap);
8181 i = pmap_change_props_locked(va, size,
8182 PROT_NONE, mode, flags);
8183 PMAP_UNLOCK(kernel_pmap);
8187 return ((void *)(va + offset));
8189 va = kva_alloc(size);
8191 panic("%s: Couldn't allocate KVA", __func__);
8193 for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE)
8194 pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode);
8195 pmap_invalidate_range(kernel_pmap, va, va + tmpsize);
8196 if ((flags & MAPDEV_FLUSHCACHE) != 0)
8197 pmap_invalidate_cache_range(va, va + tmpsize);
8198 return ((void *)(va + offset));
8202 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
8205 return (pmap_mapdev_internal(pa, size, mode, MAPDEV_FLUSHCACHE |
8210 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
8213 return (pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
8217 pmap_mapdev_pciecfg(vm_paddr_t pa, vm_size_t size)
8220 return (pmap_mapdev_internal(pa, size, PAT_UNCACHEABLE,
8225 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
8228 return (pmap_mapdev_internal(pa, size, PAT_WRITE_BACK,
8229 MAPDEV_FLUSHCACHE));
8233 pmap_unmapdev(vm_offset_t va, vm_size_t size)
8235 struct pmap_preinit_mapping *ppim;
8239 /* If we gave a direct map region in pmap_mapdev, do nothing */
8240 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS)
8242 offset = va & PAGE_MASK;
8243 size = round_page(offset + size);
8244 va = trunc_page(va);
8245 for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
8246 ppim = pmap_preinit_mapping + i;
8247 if (ppim->va == va && ppim->sz == size) {
8248 if (pmap_initialized)
8254 if (va + size == virtual_avail)
8259 if (pmap_initialized)
8264 * Tries to demote a 1GB page mapping.
8267 pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe, vm_offset_t va)
8269 pdp_entry_t newpdpe, oldpdpe;
8270 pd_entry_t *firstpde, newpde, *pde;
8271 pt_entry_t PG_A, PG_M, PG_RW, PG_V;
8275 PG_A = pmap_accessed_bit(pmap);
8276 PG_M = pmap_modified_bit(pmap);
8277 PG_V = pmap_valid_bit(pmap);
8278 PG_RW = pmap_rw_bit(pmap);
8280 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
8282 KASSERT((oldpdpe & (PG_PS | PG_V)) == (PG_PS | PG_V),
8283 ("pmap_demote_pdpe: oldpdpe is missing PG_PS and/or PG_V"));
8284 if ((pdpg = vm_page_alloc(NULL, va >> PDPSHIFT, VM_ALLOC_INTERRUPT |
8285 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) {
8286 CTR2(KTR_PMAP, "pmap_demote_pdpe: failure for va %#lx"
8287 " in pmap %p", va, pmap);
8290 pdpgpa = VM_PAGE_TO_PHYS(pdpg);
8291 firstpde = (pd_entry_t *)PHYS_TO_DMAP(pdpgpa);
8292 newpdpe = pdpgpa | PG_M | PG_A | (oldpdpe & PG_U) | PG_RW | PG_V;
8293 KASSERT((oldpdpe & PG_A) != 0,
8294 ("pmap_demote_pdpe: oldpdpe is missing PG_A"));
8295 KASSERT((oldpdpe & (PG_M | PG_RW)) != PG_RW,
8296 ("pmap_demote_pdpe: oldpdpe is missing PG_M"));
8300 * Initialize the page directory page.
8302 for (pde = firstpde; pde < firstpde + NPDEPG; pde++) {
8308 * Demote the mapping.
8313 * Invalidate a stale recursive mapping of the page directory page.
8315 pmap_invalidate_page(pmap, (vm_offset_t)vtopde(va));
8317 pmap_pdpe_demotions++;
8318 CTR2(KTR_PMAP, "pmap_demote_pdpe: success for va %#lx"
8319 " in pmap %p", va, pmap);
8324 * Sets the memory attribute for the specified page.
8327 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
8330 m->md.pat_mode = ma;
8333 * If "m" is a normal page, update its direct mapping. This update
8334 * can be relied upon to perform any cache operations that are
8335 * required for data coherence.
8337 if ((m->flags & PG_FICTITIOUS) == 0 &&
8338 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE,
8340 panic("memory attribute change on the direct map failed");
8344 * Changes the specified virtual address range's memory type to that given by
8345 * the parameter "mode". The specified virtual address range must be
8346 * completely contained within either the direct map or the kernel map. If
8347 * the virtual address range is contained within the kernel map, then the
8348 * memory type for each of the corresponding ranges of the direct map is also
8349 * changed. (The corresponding ranges of the direct map are those ranges that
8350 * map the same physical pages as the specified virtual address range.) These
8351 * changes to the direct map are necessary because Intel describes the
8352 * behavior of their processors as "undefined" if two or more mappings to the
8353 * same physical page have different memory types.
8355 * Returns zero if the change completed successfully, and either EINVAL or
8356 * ENOMEM if the change failed. Specifically, EINVAL is returned if some part
8357 * of the virtual address range was not mapped, and ENOMEM is returned if
8358 * there was insufficient memory available to complete the change. In the
8359 * latter case, the memory type may have been changed on some part of the
8360 * virtual address range or the direct map.
8363 pmap_change_attr(vm_offset_t va, vm_size_t size, int mode)
8367 PMAP_LOCK(kernel_pmap);
8368 error = pmap_change_props_locked(va, size, PROT_NONE, mode,
8370 PMAP_UNLOCK(kernel_pmap);
8375 * Changes the specified virtual address range's protections to those
8376 * specified by "prot". Like pmap_change_attr(), protections for aliases
8377 * in the direct map are updated as well. Protections on aliasing mappings may
8378 * be a subset of the requested protections; for example, mappings in the direct
8379 * map are never executable.
8382 pmap_change_prot(vm_offset_t va, vm_size_t size, vm_prot_t prot)
8386 /* Only supported within the kernel map. */
8387 if (va < VM_MIN_KERNEL_ADDRESS)
8390 PMAP_LOCK(kernel_pmap);
8391 error = pmap_change_props_locked(va, size, prot, -1,
8392 MAPDEV_ASSERTVALID);
8393 PMAP_UNLOCK(kernel_pmap);
8398 pmap_change_props_locked(vm_offset_t va, vm_size_t size, vm_prot_t prot,
8399 int mode, int flags)
8401 vm_offset_t base, offset, tmpva;
8402 vm_paddr_t pa_start, pa_end, pa_end1;
8404 pd_entry_t *pde, pde_bits, pde_mask;
8405 pt_entry_t *pte, pte_bits, pte_mask;
8409 PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED);
8410 base = trunc_page(va);
8411 offset = va & PAGE_MASK;
8412 size = round_page(offset + size);
8415 * Only supported on kernel virtual addresses, including the direct
8416 * map but excluding the recursive map.
8418 if (base < DMAP_MIN_ADDRESS)
8422 * Construct our flag sets and masks. "bits" is the subset of
8423 * "mask" that will be set in each modified PTE.
8425 * Mappings in the direct map are never allowed to be executable.
8427 pde_bits = pte_bits = 0;
8428 pde_mask = pte_mask = 0;
8430 pde_bits |= pmap_cache_bits(kernel_pmap, mode, true);
8431 pde_mask |= X86_PG_PDE_CACHE;
8432 pte_bits |= pmap_cache_bits(kernel_pmap, mode, false);
8433 pte_mask |= X86_PG_PTE_CACHE;
8435 if (prot != VM_PROT_NONE) {
8436 if ((prot & VM_PROT_WRITE) != 0) {
8437 pde_bits |= X86_PG_RW;
8438 pte_bits |= X86_PG_RW;
8440 if ((prot & VM_PROT_EXECUTE) == 0 ||
8441 va < VM_MIN_KERNEL_ADDRESS) {
8445 pde_mask |= X86_PG_RW | pg_nx;
8446 pte_mask |= X86_PG_RW | pg_nx;
8450 * Pages that aren't mapped aren't supported. Also break down 2MB pages
8451 * into 4KB pages if required.
8453 for (tmpva = base; tmpva < base + size; ) {
8454 pdpe = pmap_pdpe(kernel_pmap, tmpva);
8455 if (pdpe == NULL || *pdpe == 0) {
8456 KASSERT((flags & MAPDEV_ASSERTVALID) == 0,
8457 ("%s: addr %#lx is not mapped", __func__, tmpva));
8460 if (*pdpe & PG_PS) {
8462 * If the current 1GB page already has the required
8463 * properties, then we need not demote this page. Just
8464 * increment tmpva to the next 1GB page frame.
8466 if ((*pdpe & pde_mask) == pde_bits) {
8467 tmpva = trunc_1gpage(tmpva) + NBPDP;
8472 * If the current offset aligns with a 1GB page frame
8473 * and there is at least 1GB left within the range, then
8474 * we need not break down this page into 2MB pages.
8476 if ((tmpva & PDPMASK) == 0 &&
8477 tmpva + PDPMASK < base + size) {
8481 if (!pmap_demote_pdpe(kernel_pmap, pdpe, tmpva))
8484 pde = pmap_pdpe_to_pde(pdpe, tmpva);
8486 KASSERT((flags & MAPDEV_ASSERTVALID) == 0,
8487 ("%s: addr %#lx is not mapped", __func__, tmpva));
8492 * If the current 2MB page already has the required
8493 * properties, then we need not demote this page. Just
8494 * increment tmpva to the next 2MB page frame.
8496 if ((*pde & pde_mask) == pde_bits) {
8497 tmpva = trunc_2mpage(tmpva) + NBPDR;
8502 * If the current offset aligns with a 2MB page frame
8503 * and there is at least 2MB left within the range, then
8504 * we need not break down this page into 4KB pages.
8506 if ((tmpva & PDRMASK) == 0 &&
8507 tmpva + PDRMASK < base + size) {
8511 if (!pmap_demote_pde(kernel_pmap, pde, tmpva))
8514 pte = pmap_pde_to_pte(pde, tmpva);
8516 KASSERT((flags & MAPDEV_ASSERTVALID) == 0,
8517 ("%s: addr %#lx is not mapped", __func__, tmpva));
8525 * Ok, all the pages exist, so run through them updating their
8526 * properties if required.
8529 pa_start = pa_end = 0;
8530 for (tmpva = base; tmpva < base + size; ) {
8531 pdpe = pmap_pdpe(kernel_pmap, tmpva);
8532 if (*pdpe & PG_PS) {
8533 if ((*pdpe & pde_mask) != pde_bits) {
8534 pmap_pte_props(pdpe, pde_bits, pde_mask);
8537 if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
8538 (*pdpe & PG_PS_FRAME) < dmaplimit) {
8539 if (pa_start == pa_end) {
8540 /* Start physical address run. */
8541 pa_start = *pdpe & PG_PS_FRAME;
8542 pa_end = pa_start + NBPDP;
8543 } else if (pa_end == (*pdpe & PG_PS_FRAME))
8546 /* Run ended, update direct map. */
8547 error = pmap_change_props_locked(
8548 PHYS_TO_DMAP(pa_start),
8549 pa_end - pa_start, prot, mode,
8553 /* Start physical address run. */
8554 pa_start = *pdpe & PG_PS_FRAME;
8555 pa_end = pa_start + NBPDP;
8558 tmpva = trunc_1gpage(tmpva) + NBPDP;
8561 pde = pmap_pdpe_to_pde(pdpe, tmpva);
8563 if ((*pde & pde_mask) != pde_bits) {
8564 pmap_pte_props(pde, pde_bits, pde_mask);
8567 if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
8568 (*pde & PG_PS_FRAME) < dmaplimit) {
8569 if (pa_start == pa_end) {
8570 /* Start physical address run. */
8571 pa_start = *pde & PG_PS_FRAME;
8572 pa_end = pa_start + NBPDR;
8573 } else if (pa_end == (*pde & PG_PS_FRAME))
8576 /* Run ended, update direct map. */
8577 error = pmap_change_props_locked(
8578 PHYS_TO_DMAP(pa_start),
8579 pa_end - pa_start, prot, mode,
8583 /* Start physical address run. */
8584 pa_start = *pde & PG_PS_FRAME;
8585 pa_end = pa_start + NBPDR;
8588 tmpva = trunc_2mpage(tmpva) + NBPDR;
8590 pte = pmap_pde_to_pte(pde, tmpva);
8591 if ((*pte & pte_mask) != pte_bits) {
8592 pmap_pte_props(pte, pte_bits, pte_mask);
8595 if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
8596 (*pte & PG_FRAME) < dmaplimit) {
8597 if (pa_start == pa_end) {
8598 /* Start physical address run. */
8599 pa_start = *pte & PG_FRAME;
8600 pa_end = pa_start + PAGE_SIZE;
8601 } else if (pa_end == (*pte & PG_FRAME))
8602 pa_end += PAGE_SIZE;
8604 /* Run ended, update direct map. */
8605 error = pmap_change_props_locked(
8606 PHYS_TO_DMAP(pa_start),
8607 pa_end - pa_start, prot, mode,
8611 /* Start physical address run. */
8612 pa_start = *pte & PG_FRAME;
8613 pa_end = pa_start + PAGE_SIZE;
8619 if (error == 0 && pa_start != pa_end && pa_start < dmaplimit) {
8620 pa_end1 = MIN(pa_end, dmaplimit);
8621 if (pa_start != pa_end1)
8622 error = pmap_change_props_locked(PHYS_TO_DMAP(pa_start),
8623 pa_end1 - pa_start, prot, mode, flags);
8627 * Flush CPU caches if required to make sure any data isn't cached that
8628 * shouldn't be, etc.
8631 pmap_invalidate_range(kernel_pmap, base, tmpva);
8632 if ((flags & MAPDEV_FLUSHCACHE) != 0)
8633 pmap_invalidate_cache_range(base, tmpva);
8639 * Demotes any mapping within the direct map region that covers more than the
8640 * specified range of physical addresses. This range's size must be a power
8641 * of two and its starting address must be a multiple of its size. Since the
8642 * demotion does not change any attributes of the mapping, a TLB invalidation
8643 * is not mandatory. The caller may, however, request a TLB invalidation.
8646 pmap_demote_DMAP(vm_paddr_t base, vm_size_t len, boolean_t invalidate)
8655 KASSERT(powerof2(len), ("pmap_demote_DMAP: len is not a power of 2"));
8656 KASSERT((base & (len - 1)) == 0,
8657 ("pmap_demote_DMAP: base is not a multiple of len"));
8658 if (len < NBPDP && base < dmaplimit) {
8659 va = PHYS_TO_DMAP(base);
8661 PMAP_LOCK(kernel_pmap);
8662 pdpe = pmap_pdpe(kernel_pmap, va);
8663 if ((*pdpe & X86_PG_V) == 0)
8664 panic("pmap_demote_DMAP: invalid PDPE");
8665 if ((*pdpe & PG_PS) != 0) {
8666 if (!pmap_demote_pdpe(kernel_pmap, pdpe, va))
8667 panic("pmap_demote_DMAP: PDPE failed");
8671 pde = pmap_pdpe_to_pde(pdpe, va);
8672 if ((*pde & X86_PG_V) == 0)
8673 panic("pmap_demote_DMAP: invalid PDE");
8674 if ((*pde & PG_PS) != 0) {
8675 if (!pmap_demote_pde(kernel_pmap, pde, va))
8676 panic("pmap_demote_DMAP: PDE failed");
8680 if (changed && invalidate)
8681 pmap_invalidate_page(kernel_pmap, va);
8682 PMAP_UNLOCK(kernel_pmap);
8687 * Perform the pmap work for mincore(2). If the page is not both referenced and
8688 * modified by this pmap, returns its physical address so that the caller can
8689 * find other mappings.
8692 pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap)
8695 pt_entry_t pte, PG_A, PG_M, PG_RW, PG_V;
8699 PG_A = pmap_accessed_bit(pmap);
8700 PG_M = pmap_modified_bit(pmap);
8701 PG_V = pmap_valid_bit(pmap);
8702 PG_RW = pmap_rw_bit(pmap);
8705 pdep = pmap_pde(pmap, addr);
8706 if (pdep != NULL && (*pdep & PG_V)) {
8707 if (*pdep & PG_PS) {
8709 /* Compute the physical address of the 4KB page. */
8710 pa = ((*pdep & PG_PS_FRAME) | (addr & PDRMASK)) &
8712 val = MINCORE_SUPER;
8714 pte = *pmap_pde_to_pte(pdep, addr);
8715 pa = pte & PG_FRAME;
8723 if ((pte & PG_V) != 0) {
8724 val |= MINCORE_INCORE;
8725 if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW))
8726 val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
8727 if ((pte & PG_A) != 0)
8728 val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
8730 if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
8731 (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
8732 (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) {
8740 pmap_pcid_alloc(pmap_t pmap, u_int cpuid)
8742 uint32_t gen, new_gen, pcid_next;
8744 CRITICAL_ASSERT(curthread);
8745 gen = PCPU_GET(pcid_gen);
8746 if (pmap->pm_pcids[cpuid].pm_pcid == PMAP_PCID_KERN)
8747 return (pti ? 0 : CR3_PCID_SAVE);
8748 if (pmap->pm_pcids[cpuid].pm_gen == gen)
8749 return (CR3_PCID_SAVE);
8750 pcid_next = PCPU_GET(pcid_next);
8751 KASSERT((!pti && pcid_next <= PMAP_PCID_OVERMAX) ||
8752 (pti && pcid_next <= PMAP_PCID_OVERMAX_KERN),
8753 ("cpu %d pcid_next %#x", cpuid, pcid_next));
8754 if ((!pti && pcid_next == PMAP_PCID_OVERMAX) ||
8755 (pti && pcid_next == PMAP_PCID_OVERMAX_KERN)) {
8759 PCPU_SET(pcid_gen, new_gen);
8760 pcid_next = PMAP_PCID_KERN + 1;
8764 pmap->pm_pcids[cpuid].pm_pcid = pcid_next;
8765 pmap->pm_pcids[cpuid].pm_gen = new_gen;
8766 PCPU_SET(pcid_next, pcid_next + 1);
8771 pmap_pcid_alloc_checked(pmap_t pmap, u_int cpuid)
8775 cached = pmap_pcid_alloc(pmap, cpuid);
8776 KASSERT(pmap->pm_pcids[cpuid].pm_pcid < PMAP_PCID_OVERMAX,
8777 ("pmap %p cpu %d pcid %#x", pmap, cpuid,
8778 pmap->pm_pcids[cpuid].pm_pcid));
8779 KASSERT(pmap->pm_pcids[cpuid].pm_pcid != PMAP_PCID_KERN ||
8780 pmap == kernel_pmap,
8781 ("non-kernel pmap pmap %p cpu %d pcid %#x",
8782 pmap, cpuid, pmap->pm_pcids[cpuid].pm_pcid));
8787 pmap_activate_sw_pti_post(struct thread *td, pmap_t pmap)
8790 PCPU_GET(tssp)->tss_rsp0 = pmap->pm_ucr3 != PMAP_NO_CR3 ?
8791 PCPU_GET(pti_rsp0) : (uintptr_t)td->td_md.md_stack_base;
8795 pmap_activate_sw_pcid_pti(pmap_t pmap, u_int cpuid, const bool invpcid_works1)
8797 struct invpcid_descr d;
8798 uint64_t cached, cr3, kcr3, ucr3;
8800 cached = pmap_pcid_alloc_checked(pmap, cpuid);
8802 if ((cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3)
8803 load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid);
8804 PCPU_SET(curpmap, pmap);
8805 kcr3 = pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid;
8806 ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[cpuid].pm_pcid |
8809 if (!cached && pmap->pm_ucr3 != PMAP_NO_CR3) {
8811 * Explicitly invalidate translations cached from the
8812 * user page table. They are not automatically
8813 * flushed by reload of cr3 with the kernel page table
8816 * Note that the if() condition is resolved statically
8817 * by using the function argument instead of
8818 * runtime-evaluated invpcid_works value.
8820 if (invpcid_works1) {
8821 d.pcid = PMAP_PCID_USER_PT |
8822 pmap->pm_pcids[cpuid].pm_pcid;
8825 invpcid(&d, INVPCID_CTX);
8827 pmap_pti_pcid_invalidate(ucr3, kcr3);
8831 PCPU_SET(kcr3, kcr3 | CR3_PCID_SAVE);
8832 PCPU_SET(ucr3, ucr3 | CR3_PCID_SAVE);
8834 PCPU_INC(pm_save_cnt);
8838 pmap_activate_sw_pcid_invpcid_pti(struct thread *td, pmap_t pmap, u_int cpuid)
8841 pmap_activate_sw_pcid_pti(pmap, cpuid, true);
8842 pmap_activate_sw_pti_post(td, pmap);
8846 pmap_activate_sw_pcid_noinvpcid_pti(struct thread *td, pmap_t pmap,
8852 * If the INVPCID instruction is not available,
8853 * invltlb_pcid_handler() is used to handle an invalidate_all
8854 * IPI, which checks for curpmap == smp_tlb_pmap. The below
8855 * sequence of operations has a window where %CR3 is loaded
8856 * with the new pmap's PML4 address, but the curpmap value has
8857 * not yet been updated. This causes the invltlb IPI handler,
8858 * which is called between the updates, to execute as a NOP,
8859 * which leaves stale TLB entries.
8861 * Note that the most typical use of pmap_activate_sw(), from
8862 * the context switch, is immune to this race, because
8863 * interrupts are disabled (while the thread lock is owned),
8864 * and the IPI happens after curpmap is updated. Protect
8865 * other callers in a similar way, by disabling interrupts
8866 * around the %cr3 register reload and curpmap assignment.
8868 rflags = intr_disable();
8869 pmap_activate_sw_pcid_pti(pmap, cpuid, false);
8870 intr_restore(rflags);
8871 pmap_activate_sw_pti_post(td, pmap);
8875 pmap_activate_sw_pcid_nopti(struct thread *td __unused, pmap_t pmap,
8878 uint64_t cached, cr3;
8880 cached = pmap_pcid_alloc_checked(pmap, cpuid);
8882 if (!cached || (cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3)
8883 load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid |
8885 PCPU_SET(curpmap, pmap);
8887 PCPU_INC(pm_save_cnt);
8891 pmap_activate_sw_pcid_noinvpcid_nopti(struct thread *td __unused, pmap_t pmap,
8896 rflags = intr_disable();
8897 pmap_activate_sw_pcid_nopti(td, pmap, cpuid);
8898 intr_restore(rflags);
8902 pmap_activate_sw_nopcid_nopti(struct thread *td __unused, pmap_t pmap,
8903 u_int cpuid __unused)
8906 load_cr3(pmap->pm_cr3);
8907 PCPU_SET(curpmap, pmap);
8911 pmap_activate_sw_nopcid_pti(struct thread *td, pmap_t pmap,
8912 u_int cpuid __unused)
8915 pmap_activate_sw_nopcid_nopti(td, pmap, cpuid);
8916 PCPU_SET(kcr3, pmap->pm_cr3);
8917 PCPU_SET(ucr3, pmap->pm_ucr3);
8918 pmap_activate_sw_pti_post(td, pmap);
8921 DEFINE_IFUNC(static, void, pmap_activate_sw_mode, (struct thread *, pmap_t,
8925 if (pmap_pcid_enabled && pti && invpcid_works)
8926 return (pmap_activate_sw_pcid_invpcid_pti);
8927 else if (pmap_pcid_enabled && pti && !invpcid_works)
8928 return (pmap_activate_sw_pcid_noinvpcid_pti);
8929 else if (pmap_pcid_enabled && !pti && invpcid_works)
8930 return (pmap_activate_sw_pcid_nopti);
8931 else if (pmap_pcid_enabled && !pti && !invpcid_works)
8932 return (pmap_activate_sw_pcid_noinvpcid_nopti);
8933 else if (!pmap_pcid_enabled && pti)
8934 return (pmap_activate_sw_nopcid_pti);
8935 else /* if (!pmap_pcid_enabled && !pti) */
8936 return (pmap_activate_sw_nopcid_nopti);
8940 pmap_activate_sw(struct thread *td)
8942 pmap_t oldpmap, pmap;
8945 oldpmap = PCPU_GET(curpmap);
8946 pmap = vmspace_pmap(td->td_proc->p_vmspace);
8947 if (oldpmap == pmap) {
8948 if (cpu_vendor_id != CPU_VENDOR_INTEL)
8952 cpuid = PCPU_GET(cpuid);
8954 CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
8956 CPU_SET(cpuid, &pmap->pm_active);
8958 pmap_activate_sw_mode(td, pmap, cpuid);
8960 CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active);
8962 CPU_CLR(cpuid, &oldpmap->pm_active);
8967 pmap_activate(struct thread *td)
8971 pmap_activate_sw(td);
8976 pmap_activate_boot(pmap_t pmap)
8982 * kernel_pmap must be never deactivated, and we ensure that
8983 * by never activating it at all.
8985 MPASS(pmap != kernel_pmap);
8987 cpuid = PCPU_GET(cpuid);
8989 CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
8991 CPU_SET(cpuid, &pmap->pm_active);
8993 PCPU_SET(curpmap, pmap);
8995 kcr3 = pmap->pm_cr3;
8996 if (pmap_pcid_enabled)
8997 kcr3 |= pmap->pm_pcids[cpuid].pm_pcid | CR3_PCID_SAVE;
9001 PCPU_SET(kcr3, kcr3);
9002 PCPU_SET(ucr3, PMAP_NO_CR3);
9006 pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
9011 * Increase the starting virtual address of the given mapping if a
9012 * different alignment might result in more superpage mappings.
9015 pmap_align_superpage(vm_object_t object, vm_ooffset_t offset,
9016 vm_offset_t *addr, vm_size_t size)
9018 vm_offset_t superpage_offset;
9022 if (object != NULL && (object->flags & OBJ_COLORED) != 0)
9023 offset += ptoa(object->pg_color);
9024 superpage_offset = offset & PDRMASK;
9025 if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR ||
9026 (*addr & PDRMASK) == superpage_offset)
9028 if ((*addr & PDRMASK) < superpage_offset)
9029 *addr = (*addr & ~PDRMASK) + superpage_offset;
9031 *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset;
9035 static unsigned long num_dirty_emulations;
9036 SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_dirty_emulations, CTLFLAG_RW,
9037 &num_dirty_emulations, 0, NULL);
9039 static unsigned long num_accessed_emulations;
9040 SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_accessed_emulations, CTLFLAG_RW,
9041 &num_accessed_emulations, 0, NULL);
9043 static unsigned long num_superpage_accessed_emulations;
9044 SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_superpage_accessed_emulations, CTLFLAG_RW,
9045 &num_superpage_accessed_emulations, 0, NULL);
9047 static unsigned long ad_emulation_superpage_promotions;
9048 SYSCTL_ULONG(_vm_pmap, OID_AUTO, ad_emulation_superpage_promotions, CTLFLAG_RW,
9049 &ad_emulation_superpage_promotions, 0, NULL);
9050 #endif /* INVARIANTS */
9053 pmap_emulate_accessed_dirty(pmap_t pmap, vm_offset_t va, int ftype)
9056 struct rwlock *lock;
9057 #if VM_NRESERVLEVEL > 0
9061 pt_entry_t *pte, PG_A, PG_M, PG_RW, PG_V;
9063 KASSERT(ftype == VM_PROT_READ || ftype == VM_PROT_WRITE,
9064 ("pmap_emulate_accessed_dirty: invalid fault type %d", ftype));
9066 if (!pmap_emulate_ad_bits(pmap))
9069 PG_A = pmap_accessed_bit(pmap);
9070 PG_M = pmap_modified_bit(pmap);
9071 PG_V = pmap_valid_bit(pmap);
9072 PG_RW = pmap_rw_bit(pmap);
9078 pde = pmap_pde(pmap, va);
9079 if (pde == NULL || (*pde & PG_V) == 0)
9082 if ((*pde & PG_PS) != 0) {
9083 if (ftype == VM_PROT_READ) {
9085 atomic_add_long(&num_superpage_accessed_emulations, 1);
9093 pte = pmap_pde_to_pte(pde, va);
9094 if ((*pte & PG_V) == 0)
9097 if (ftype == VM_PROT_WRITE) {
9098 if ((*pte & PG_RW) == 0)
9101 * Set the modified and accessed bits simultaneously.
9103 * Intel EPT PTEs that do software emulation of A/D bits map
9104 * PG_A and PG_M to EPT_PG_READ and EPT_PG_WRITE respectively.
9105 * An EPT misconfiguration is triggered if the PTE is writable
9106 * but not readable (WR=10). This is avoided by setting PG_A
9107 * and PG_M simultaneously.
9109 *pte |= PG_M | PG_A;
9114 #if VM_NRESERVLEVEL > 0
9115 /* try to promote the mapping */
9116 if (va < VM_MAXUSER_ADDRESS)
9117 mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
9121 m = PHYS_TO_VM_PAGE(*pte & PG_FRAME);
9123 if ((mpte == NULL || mpte->ref_count == NPTEPG) &&
9124 pmap_ps_enabled(pmap) &&
9125 (m->flags & PG_FICTITIOUS) == 0 &&
9126 vm_reserv_level_iffullpop(m) == 0) {
9127 pmap_promote_pde(pmap, pde, va, &lock);
9129 atomic_add_long(&ad_emulation_superpage_promotions, 1);
9135 if (ftype == VM_PROT_WRITE)
9136 atomic_add_long(&num_dirty_emulations, 1);
9138 atomic_add_long(&num_accessed_emulations, 1);
9140 rv = 0; /* success */
9149 pmap_get_mapping(pmap_t pmap, vm_offset_t va, uint64_t *ptr, int *num)
9154 pt_entry_t *pte, PG_V;
9158 PG_V = pmap_valid_bit(pmap);
9161 pml4 = pmap_pml4e(pmap, va);
9163 if ((*pml4 & PG_V) == 0)
9166 pdp = pmap_pml4e_to_pdpe(pml4, va);
9168 if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0)
9171 pde = pmap_pdpe_to_pde(pdp, va);
9173 if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0)
9176 pte = pmap_pde_to_pte(pde, va);
9185 * Get the kernel virtual address of a set of physical pages. If there are
9186 * physical addresses not covered by the DMAP perform a transient mapping
9187 * that will be removed when calling pmap_unmap_io_transient.
9189 * \param page The pages the caller wishes to obtain the virtual
9190 * address on the kernel memory map.
9191 * \param vaddr On return contains the kernel virtual memory address
9192 * of the pages passed in the page parameter.
9193 * \param count Number of pages passed in.
9194 * \param can_fault TRUE if the thread using the mapped pages can take
9195 * page faults, FALSE otherwise.
9197 * \returns TRUE if the caller must call pmap_unmap_io_transient when
9198 * finished or FALSE otherwise.
9202 pmap_map_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count,
9203 boolean_t can_fault)
9206 boolean_t needs_mapping;
9208 int cache_bits, error __unused, i;
9211 * Allocate any KVA space that we need, this is done in a separate
9212 * loop to prevent calling vmem_alloc while pinned.
9214 needs_mapping = FALSE;
9215 for (i = 0; i < count; i++) {
9216 paddr = VM_PAGE_TO_PHYS(page[i]);
9217 if (__predict_false(paddr >= dmaplimit)) {
9218 error = vmem_alloc(kernel_arena, PAGE_SIZE,
9219 M_BESTFIT | M_WAITOK, &vaddr[i]);
9220 KASSERT(error == 0, ("vmem_alloc failed: %d", error));
9221 needs_mapping = TRUE;
9223 vaddr[i] = PHYS_TO_DMAP(paddr);
9227 /* Exit early if everything is covered by the DMAP */
9232 * NB: The sequence of updating a page table followed by accesses
9233 * to the corresponding pages used in the !DMAP case is subject to
9234 * the situation described in the "AMD64 Architecture Programmer's
9235 * Manual Volume 2: System Programming" rev. 3.23, "7.3.1 Special
9236 * Coherency Considerations". Therefore, issuing the INVLPG right
9237 * after modifying the PTE bits is crucial.
9241 for (i = 0; i < count; i++) {
9242 paddr = VM_PAGE_TO_PHYS(page[i]);
9243 if (paddr >= dmaplimit) {
9246 * Slow path, since we can get page faults
9247 * while mappings are active don't pin the
9248 * thread to the CPU and instead add a global
9249 * mapping visible to all CPUs.
9251 pmap_qenter(vaddr[i], &page[i], 1);
9253 pte = vtopte(vaddr[i]);
9254 cache_bits = pmap_cache_bits(kernel_pmap,
9255 page[i]->md.pat_mode, 0);
9256 pte_store(pte, paddr | X86_PG_RW | X86_PG_V |
9263 return (needs_mapping);
9267 pmap_unmap_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count,
9268 boolean_t can_fault)
9275 for (i = 0; i < count; i++) {
9276 paddr = VM_PAGE_TO_PHYS(page[i]);
9277 if (paddr >= dmaplimit) {
9279 pmap_qremove(vaddr[i], 1);
9280 vmem_free(kernel_arena, vaddr[i], PAGE_SIZE);
9286 pmap_quick_enter_page(vm_page_t m)
9290 paddr = VM_PAGE_TO_PHYS(m);
9291 if (paddr < dmaplimit)
9292 return (PHYS_TO_DMAP(paddr));
9293 mtx_lock_spin(&qframe_mtx);
9294 KASSERT(*vtopte(qframe) == 0, ("qframe busy"));
9295 pte_store(vtopte(qframe), paddr | X86_PG_RW | X86_PG_V | X86_PG_A |
9296 X86_PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0));
9301 pmap_quick_remove_page(vm_offset_t addr)
9306 pte_store(vtopte(qframe), 0);
9308 mtx_unlock_spin(&qframe_mtx);
9312 * Pdp pages from the large map are managed differently from either
9313 * kernel or user page table pages. They are permanently allocated at
9314 * initialization time, and their reference count is permanently set to
9315 * zero. The pml4 entries pointing to those pages are copied into
9316 * each allocated pmap.
9318 * In contrast, pd and pt pages are managed like user page table
9319 * pages. They are dynamically allocated, and their reference count
9320 * represents the number of valid entries within the page.
9323 pmap_large_map_getptp_unlocked(void)
9327 m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
9329 if (m != NULL && (m->flags & PG_ZERO) == 0)
9335 pmap_large_map_getptp(void)
9339 PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED);
9340 m = pmap_large_map_getptp_unlocked();
9342 PMAP_UNLOCK(kernel_pmap);
9344 PMAP_LOCK(kernel_pmap);
9345 /* Callers retry. */
9350 static pdp_entry_t *
9351 pmap_large_map_pdpe(vm_offset_t va)
9353 vm_pindex_t pml4_idx;
9356 pml4_idx = pmap_pml4e_index(va);
9357 KASSERT(LMSPML4I <= pml4_idx && pml4_idx < LMSPML4I + lm_ents,
9358 ("pmap_large_map_pdpe: va %#jx out of range idx %#jx LMSPML4I "
9360 (uintmax_t)va, (uintmax_t)pml4_idx, LMSPML4I, lm_ents));
9361 KASSERT((kernel_pmap->pm_pml4[pml4_idx] & X86_PG_V) != 0,
9362 ("pmap_large_map_pdpe: invalid pml4 for va %#jx idx %#jx "
9363 "LMSPML4I %#jx lm_ents %d",
9364 (uintmax_t)va, (uintmax_t)pml4_idx, LMSPML4I, lm_ents));
9365 mphys = kernel_pmap->pm_pml4[pml4_idx] & PG_FRAME;
9366 return ((pdp_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pdpe_index(va));
9370 pmap_large_map_pde(vm_offset_t va)
9377 pdpe = pmap_large_map_pdpe(va);
9379 m = pmap_large_map_getptp();
9382 mphys = VM_PAGE_TO_PHYS(m);
9383 *pdpe = mphys | X86_PG_A | X86_PG_RW | X86_PG_V | pg_nx;
9385 MPASS((*pdpe & X86_PG_PS) == 0);
9386 mphys = *pdpe & PG_FRAME;
9388 return ((pd_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pde_index(va));
9392 pmap_large_map_pte(vm_offset_t va)
9399 pde = pmap_large_map_pde(va);
9401 m = pmap_large_map_getptp();
9404 mphys = VM_PAGE_TO_PHYS(m);
9405 *pde = mphys | X86_PG_A | X86_PG_RW | X86_PG_V | pg_nx;
9406 PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde))->ref_count++;
9408 MPASS((*pde & X86_PG_PS) == 0);
9409 mphys = *pde & PG_FRAME;
9411 return ((pt_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pte_index(va));
9415 pmap_large_map_kextract(vm_offset_t va)
9417 pdp_entry_t *pdpe, pdp;
9418 pd_entry_t *pde, pd;
9419 pt_entry_t *pte, pt;
9421 KASSERT(PMAP_ADDRESS_IN_LARGEMAP(va),
9422 ("not largemap range %#lx", (u_long)va));
9423 pdpe = pmap_large_map_pdpe(va);
9425 KASSERT((pdp & X86_PG_V) != 0,
9426 ("invalid pdp va %#lx pdpe %#lx pdp %#lx", va,
9427 (u_long)pdpe, pdp));
9428 if ((pdp & X86_PG_PS) != 0) {
9429 KASSERT((amd_feature & AMDID_PAGE1GB) != 0,
9430 ("no 1G pages, va %#lx pdpe %#lx pdp %#lx", va,
9431 (u_long)pdpe, pdp));
9432 return ((pdp & PG_PS_PDP_FRAME) | (va & PDPMASK));
9434 pde = pmap_pdpe_to_pde(pdpe, va);
9436 KASSERT((pd & X86_PG_V) != 0,
9437 ("invalid pd va %#lx pde %#lx pd %#lx", va, (u_long)pde, pd));
9438 if ((pd & X86_PG_PS) != 0)
9439 return ((pd & PG_PS_FRAME) | (va & PDRMASK));
9440 pte = pmap_pde_to_pte(pde, va);
9442 KASSERT((pt & X86_PG_V) != 0,
9443 ("invalid pte va %#lx pte %#lx pt %#lx", va, (u_long)pte, pt));
9444 return ((pt & PG_FRAME) | (va & PAGE_MASK));
9448 pmap_large_map_getva(vm_size_t len, vm_offset_t align, vm_offset_t phase,
9449 vmem_addr_t *vmem_res)
9453 * Large mappings are all but static. Consequently, there
9454 * is no point in waiting for an earlier allocation to be
9457 return (vmem_xalloc(large_vmem, len, align, phase, 0, VMEM_ADDR_MIN,
9458 VMEM_ADDR_MAX, M_NOWAIT | M_BESTFIT, vmem_res));
9462 pmap_large_map(vm_paddr_t spa, vm_size_t len, void **addr,
9468 vm_offset_t va, inc;
9469 vmem_addr_t vmem_res;
9473 if (len == 0 || spa + len < spa)
9476 /* See if DMAP can serve. */
9477 if (spa + len <= dmaplimit) {
9478 va = PHYS_TO_DMAP(spa);
9480 return (pmap_change_attr(va, len, mattr));
9484 * No, allocate KVA. Fit the address with best possible
9485 * alignment for superpages. Fall back to worse align if
9489 if ((amd_feature & AMDID_PAGE1GB) != 0 && rounddown2(spa + len,
9490 NBPDP) >= roundup2(spa, NBPDP) + NBPDP)
9491 error = pmap_large_map_getva(len, NBPDP, spa & PDPMASK,
9493 if (error != 0 && rounddown2(spa + len, NBPDR) >= roundup2(spa,
9495 error = pmap_large_map_getva(len, NBPDR, spa & PDRMASK,
9498 error = pmap_large_map_getva(len, PAGE_SIZE, 0, &vmem_res);
9503 * Fill pagetable. PG_M is not pre-set, we scan modified bits
9504 * in the pagetable to minimize flushing. No need to
9505 * invalidate TLB, since we only update invalid entries.
9507 PMAP_LOCK(kernel_pmap);
9508 for (pa = spa, va = vmem_res; len > 0; pa += inc, va += inc,
9510 if ((amd_feature & AMDID_PAGE1GB) != 0 && len >= NBPDP &&
9511 (pa & PDPMASK) == 0 && (va & PDPMASK) == 0) {
9512 pdpe = pmap_large_map_pdpe(va);
9514 *pdpe = pa | pg_g | X86_PG_PS | X86_PG_RW |
9515 X86_PG_V | X86_PG_A | pg_nx |
9516 pmap_cache_bits(kernel_pmap, mattr, TRUE);
9518 } else if (len >= NBPDR && (pa & PDRMASK) == 0 &&
9519 (va & PDRMASK) == 0) {
9520 pde = pmap_large_map_pde(va);
9522 *pde = pa | pg_g | X86_PG_PS | X86_PG_RW |
9523 X86_PG_V | X86_PG_A | pg_nx |
9524 pmap_cache_bits(kernel_pmap, mattr, TRUE);
9525 PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde))->
9529 pte = pmap_large_map_pte(va);
9531 *pte = pa | pg_g | X86_PG_RW | X86_PG_V |
9532 X86_PG_A | pg_nx | pmap_cache_bits(kernel_pmap,
9534 PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte))->
9539 PMAP_UNLOCK(kernel_pmap);
9542 *addr = (void *)vmem_res;
9547 pmap_large_unmap(void *svaa, vm_size_t len)
9549 vm_offset_t sva, va;
9551 pdp_entry_t *pdpe, pdp;
9552 pd_entry_t *pde, pd;
9555 struct spglist spgf;
9557 sva = (vm_offset_t)svaa;
9558 if (len == 0 || sva + len < sva || (sva >= DMAP_MIN_ADDRESS &&
9559 sva + len <= DMAP_MIN_ADDRESS + dmaplimit))
9563 KASSERT(PMAP_ADDRESS_IN_LARGEMAP(sva) &&
9564 PMAP_ADDRESS_IN_LARGEMAP(sva + len - 1),
9565 ("not largemap range %#lx %#lx", (u_long)svaa, (u_long)svaa + len));
9566 PMAP_LOCK(kernel_pmap);
9567 for (va = sva; va < sva + len; va += inc) {
9568 pdpe = pmap_large_map_pdpe(va);
9570 KASSERT((pdp & X86_PG_V) != 0,
9571 ("invalid pdp va %#lx pdpe %#lx pdp %#lx", va,
9572 (u_long)pdpe, pdp));
9573 if ((pdp & X86_PG_PS) != 0) {
9574 KASSERT((amd_feature & AMDID_PAGE1GB) != 0,
9575 ("no 1G pages, va %#lx pdpe %#lx pdp %#lx", va,
9576 (u_long)pdpe, pdp));
9577 KASSERT((va & PDPMASK) == 0,
9578 ("PDPMASK bit set, va %#lx pdpe %#lx pdp %#lx", va,
9579 (u_long)pdpe, pdp));
9580 KASSERT(va + NBPDP <= sva + len,
9581 ("unmap covers partial 1GB page, sva %#lx va %#lx "
9582 "pdpe %#lx pdp %#lx len %#lx", sva, va,
9583 (u_long)pdpe, pdp, len));
9588 pde = pmap_pdpe_to_pde(pdpe, va);
9590 KASSERT((pd & X86_PG_V) != 0,
9591 ("invalid pd va %#lx pde %#lx pd %#lx", va,
9593 if ((pd & X86_PG_PS) != 0) {
9594 KASSERT((va & PDRMASK) == 0,
9595 ("PDRMASK bit set, va %#lx pde %#lx pd %#lx", va,
9597 KASSERT(va + NBPDR <= sva + len,
9598 ("unmap covers partial 2MB page, sva %#lx va %#lx "
9599 "pde %#lx pd %#lx len %#lx", sva, va, (u_long)pde,
9603 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pde));
9605 if (m->ref_count == 0) {
9607 SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss);
9611 pte = pmap_pde_to_pte(pde, va);
9612 KASSERT((*pte & X86_PG_V) != 0,
9613 ("invalid pte va %#lx pte %#lx pt %#lx", va,
9614 (u_long)pte, *pte));
9617 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pte));
9619 if (m->ref_count == 0) {
9621 SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss);
9622 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pde));
9624 if (m->ref_count == 0) {
9626 SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss);
9630 pmap_invalidate_range(kernel_pmap, sva, sva + len);
9631 PMAP_UNLOCK(kernel_pmap);
9632 vm_page_free_pages_toq(&spgf, false);
9633 vmem_free(large_vmem, sva, len);
9637 pmap_large_map_wb_fence_mfence(void)
9644 pmap_large_map_wb_fence_atomic(void)
9647 atomic_thread_fence_seq_cst();
9651 pmap_large_map_wb_fence_nop(void)
9655 DEFINE_IFUNC(static, void, pmap_large_map_wb_fence, (void))
9658 if (cpu_vendor_id != CPU_VENDOR_INTEL)
9659 return (pmap_large_map_wb_fence_mfence);
9660 else if ((cpu_stdext_feature & (CPUID_STDEXT_CLWB |
9661 CPUID_STDEXT_CLFLUSHOPT)) == 0)
9662 return (pmap_large_map_wb_fence_atomic);
9664 /* clflush is strongly enough ordered */
9665 return (pmap_large_map_wb_fence_nop);
9669 pmap_large_map_flush_range_clwb(vm_offset_t va, vm_size_t len)
9672 for (; len > 0; len -= cpu_clflush_line_size,
9673 va += cpu_clflush_line_size)
9678 pmap_large_map_flush_range_clflushopt(vm_offset_t va, vm_size_t len)
9681 for (; len > 0; len -= cpu_clflush_line_size,
9682 va += cpu_clflush_line_size)
9687 pmap_large_map_flush_range_clflush(vm_offset_t va, vm_size_t len)
9690 for (; len > 0; len -= cpu_clflush_line_size,
9691 va += cpu_clflush_line_size)
9696 pmap_large_map_flush_range_nop(vm_offset_t sva __unused, vm_size_t len __unused)
9700 DEFINE_IFUNC(static, void, pmap_large_map_flush_range, (vm_offset_t, vm_size_t))
9703 if ((cpu_stdext_feature & CPUID_STDEXT_CLWB) != 0)
9704 return (pmap_large_map_flush_range_clwb);
9705 else if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0)
9706 return (pmap_large_map_flush_range_clflushopt);
9707 else if ((cpu_feature & CPUID_CLFSH) != 0)
9708 return (pmap_large_map_flush_range_clflush);
9710 return (pmap_large_map_flush_range_nop);
9714 pmap_large_map_wb_large(vm_offset_t sva, vm_offset_t eva)
9716 volatile u_long *pe;
9722 for (va = sva; va < eva; va += inc) {
9724 if ((amd_feature & AMDID_PAGE1GB) != 0) {
9725 pe = (volatile u_long *)pmap_large_map_pdpe(va);
9727 if ((p & X86_PG_PS) != 0)
9731 pe = (volatile u_long *)pmap_large_map_pde(va);
9733 if ((p & X86_PG_PS) != 0)
9737 pe = (volatile u_long *)pmap_large_map_pte(va);
9743 if ((p & X86_PG_AVAIL1) != 0) {
9745 * Spin-wait for the end of a parallel
9752 * If we saw other write-back
9753 * occuring, we cannot rely on PG_M to
9754 * indicate state of the cache. The
9755 * PG_M bit is cleared before the
9756 * flush to avoid ignoring new writes,
9757 * and writes which are relevant for
9758 * us might happen after.
9764 if ((p & X86_PG_M) != 0 || seen_other) {
9765 if (!atomic_fcmpset_long(pe, &p,
9766 (p & ~X86_PG_M) | X86_PG_AVAIL1))
9768 * If we saw PG_M without
9769 * PG_AVAIL1, and then on the
9770 * next attempt we do not
9771 * observe either PG_M or
9772 * PG_AVAIL1, the other
9773 * write-back started after us
9774 * and finished before us. We
9775 * can rely on it doing our
9779 pmap_large_map_flush_range(va, inc);
9780 atomic_clear_long(pe, X86_PG_AVAIL1);
9789 * Write-back cache lines for the given address range.
9791 * Must be called only on the range or sub-range returned from
9792 * pmap_large_map(). Must not be called on the coalesced ranges.
9794 * Does nothing on CPUs without CLWB, CLFLUSHOPT, or CLFLUSH
9795 * instructions support.
9798 pmap_large_map_wb(void *svap, vm_size_t len)
9800 vm_offset_t eva, sva;
9802 sva = (vm_offset_t)svap;
9804 pmap_large_map_wb_fence();
9805 if (sva >= DMAP_MIN_ADDRESS && eva <= DMAP_MIN_ADDRESS + dmaplimit) {
9806 pmap_large_map_flush_range(sva, len);
9808 KASSERT(sva >= LARGEMAP_MIN_ADDRESS &&
9809 eva <= LARGEMAP_MIN_ADDRESS + lm_ents * NBPML4,
9810 ("pmap_large_map_wb: not largemap %#lx %#lx", sva, len));
9811 pmap_large_map_wb_large(sva, eva);
9813 pmap_large_map_wb_fence();
9817 pmap_pti_alloc_page(void)
9821 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
9822 m = vm_page_grab(pti_obj, pti_pg_idx++, VM_ALLOC_NOBUSY |
9823 VM_ALLOC_WIRED | VM_ALLOC_ZERO);
9828 pmap_pti_free_page(vm_page_t m)
9831 KASSERT(m->ref_count > 0, ("page %p not referenced", m));
9832 if (!vm_page_unwire_noq(m))
9834 vm_page_free_zero(m);
9848 pti_obj = vm_pager_allocate(OBJT_PHYS, NULL, 0, VM_PROT_ALL, 0, NULL);
9849 VM_OBJECT_WLOCK(pti_obj);
9850 pml4_pg = pmap_pti_alloc_page();
9851 pti_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4_pg));
9852 for (va = VM_MIN_KERNEL_ADDRESS; va <= VM_MAX_KERNEL_ADDRESS &&
9853 va >= VM_MIN_KERNEL_ADDRESS && va > NBPML4; va += NBPML4) {
9854 pdpe = pmap_pti_pdpe(va);
9855 pmap_pti_wire_pte(pdpe);
9857 pmap_pti_add_kva_locked((vm_offset_t)&__pcpu[0],
9858 (vm_offset_t)&__pcpu[0] + sizeof(__pcpu[0]) * MAXCPU, false);
9859 pmap_pti_add_kva_locked((vm_offset_t)idt, (vm_offset_t)idt +
9860 sizeof(struct gate_descriptor) * NIDT, false);
9862 /* Doublefault stack IST 1 */
9863 va = __pcpu[i].pc_common_tss.tss_ist1;
9864 pmap_pti_add_kva_locked(va - PAGE_SIZE, va, false);
9865 /* NMI stack IST 2 */
9866 va = __pcpu[i].pc_common_tss.tss_ist2 + sizeof(struct nmi_pcpu);
9867 pmap_pti_add_kva_locked(va - PAGE_SIZE, va, false);
9868 /* MC# stack IST 3 */
9869 va = __pcpu[i].pc_common_tss.tss_ist3 +
9870 sizeof(struct nmi_pcpu);
9871 pmap_pti_add_kva_locked(va - PAGE_SIZE, va, false);
9872 /* DB# stack IST 4 */
9873 va = __pcpu[i].pc_common_tss.tss_ist4 + sizeof(struct nmi_pcpu);
9874 pmap_pti_add_kva_locked(va - PAGE_SIZE, va, false);
9876 pmap_pti_add_kva_locked((vm_offset_t)kernphys + KERNBASE,
9877 (vm_offset_t)etext, true);
9878 pti_finalized = true;
9879 VM_OBJECT_WUNLOCK(pti_obj);
9881 SYSINIT(pmap_pti, SI_SUB_CPU + 1, SI_ORDER_ANY, pmap_pti_init, NULL);
9883 static pdp_entry_t *
9884 pmap_pti_pdpe(vm_offset_t va)
9886 pml4_entry_t *pml4e;
9889 vm_pindex_t pml4_idx;
9892 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
9894 pml4_idx = pmap_pml4e_index(va);
9895 pml4e = &pti_pml4[pml4_idx];
9899 panic("pml4 alloc after finalization\n");
9900 m = pmap_pti_alloc_page();
9902 pmap_pti_free_page(m);
9903 mphys = *pml4e & ~PAGE_MASK;
9905 mphys = VM_PAGE_TO_PHYS(m);
9906 *pml4e = mphys | X86_PG_RW | X86_PG_V;
9909 mphys = *pml4e & ~PAGE_MASK;
9911 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pdpe_index(va);
9916 pmap_pti_wire_pte(void *pte)
9920 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
9921 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte));
9926 pmap_pti_unwire_pde(void *pde, bool only_ref)
9930 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
9931 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde));
9932 MPASS(m->ref_count > 0);
9933 MPASS(only_ref || m->ref_count > 1);
9934 pmap_pti_free_page(m);
9938 pmap_pti_unwire_pte(void *pte, vm_offset_t va)
9943 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
9944 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte));
9945 MPASS(m->ref_count > 0);
9946 if (pmap_pti_free_page(m)) {
9947 pde = pmap_pti_pde(va);
9948 MPASS((*pde & (X86_PG_PS | X86_PG_V)) == X86_PG_V);
9950 pmap_pti_unwire_pde(pde, false);
9955 pmap_pti_pde(vm_offset_t va)
9963 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
9965 pdpe = pmap_pti_pdpe(va);
9967 m = pmap_pti_alloc_page();
9969 pmap_pti_free_page(m);
9970 MPASS((*pdpe & X86_PG_PS) == 0);
9971 mphys = *pdpe & ~PAGE_MASK;
9973 mphys = VM_PAGE_TO_PHYS(m);
9974 *pdpe = mphys | X86_PG_RW | X86_PG_V;
9977 MPASS((*pdpe & X86_PG_PS) == 0);
9978 mphys = *pdpe & ~PAGE_MASK;
9981 pde = (pd_entry_t *)PHYS_TO_DMAP(mphys);
9982 pd_idx = pmap_pde_index(va);
9988 pmap_pti_pte(vm_offset_t va, bool *unwire_pde)
9995 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
9997 pde = pmap_pti_pde(va);
9998 if (unwire_pde != NULL) {
10000 pmap_pti_wire_pte(pde);
10003 m = pmap_pti_alloc_page();
10005 pmap_pti_free_page(m);
10006 MPASS((*pde & X86_PG_PS) == 0);
10007 mphys = *pde & ~(PAGE_MASK | pg_nx);
10009 mphys = VM_PAGE_TO_PHYS(m);
10010 *pde = mphys | X86_PG_RW | X86_PG_V;
10011 if (unwire_pde != NULL)
10012 *unwire_pde = false;
10015 MPASS((*pde & X86_PG_PS) == 0);
10016 mphys = *pde & ~(PAGE_MASK | pg_nx);
10019 pte = (pt_entry_t *)PHYS_TO_DMAP(mphys);
10020 pte += pmap_pte_index(va);
10026 pmap_pti_add_kva_locked(vm_offset_t sva, vm_offset_t eva, bool exec)
10030 pt_entry_t *pte, ptev;
10033 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
10035 sva = trunc_page(sva);
10036 MPASS(sva > VM_MAXUSER_ADDRESS);
10037 eva = round_page(eva);
10039 for (; sva < eva; sva += PAGE_SIZE) {
10040 pte = pmap_pti_pte(sva, &unwire_pde);
10041 pa = pmap_kextract(sva);
10042 ptev = pa | X86_PG_RW | X86_PG_V | X86_PG_A | X86_PG_G |
10043 (exec ? 0 : pg_nx) | pmap_cache_bits(kernel_pmap,
10044 VM_MEMATTR_DEFAULT, FALSE);
10046 pte_store(pte, ptev);
10047 pmap_pti_wire_pte(pte);
10049 KASSERT(!pti_finalized,
10050 ("pti overlap after fin %#lx %#lx %#lx",
10052 KASSERT(*pte == ptev,
10053 ("pti non-identical pte after fin %#lx %#lx %#lx",
10057 pde = pmap_pti_pde(sva);
10058 pmap_pti_unwire_pde(pde, true);
10064 pmap_pti_add_kva(vm_offset_t sva, vm_offset_t eva, bool exec)
10069 VM_OBJECT_WLOCK(pti_obj);
10070 pmap_pti_add_kva_locked(sva, eva, exec);
10071 VM_OBJECT_WUNLOCK(pti_obj);
10075 pmap_pti_remove_kva(vm_offset_t sva, vm_offset_t eva)
10082 sva = rounddown2(sva, PAGE_SIZE);
10083 MPASS(sva > VM_MAXUSER_ADDRESS);
10084 eva = roundup2(eva, PAGE_SIZE);
10086 VM_OBJECT_WLOCK(pti_obj);
10087 for (va = sva; va < eva; va += PAGE_SIZE) {
10088 pte = pmap_pti_pte(va, NULL);
10089 KASSERT((*pte & X86_PG_V) != 0,
10090 ("invalid pte va %#lx pte %#lx pt %#lx", va,
10091 (u_long)pte, *pte));
10093 pmap_pti_unwire_pte(pte, va);
10095 pmap_invalidate_range(kernel_pmap, sva, eva);
10096 VM_OBJECT_WUNLOCK(pti_obj);
10100 pkru_dup_range(void *ctx __unused, void *data)
10102 struct pmap_pkru_range *node, *new_node;
10104 new_node = uma_zalloc(pmap_pkru_ranges_zone, M_NOWAIT);
10105 if (new_node == NULL)
10108 memcpy(new_node, node, sizeof(*node));
10113 pkru_free_range(void *ctx __unused, void *node)
10116 uma_zfree(pmap_pkru_ranges_zone, node);
10120 pmap_pkru_assign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx,
10123 struct pmap_pkru_range *ppr;
10126 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
10127 MPASS(pmap->pm_type == PT_X86);
10128 MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0);
10129 if ((flags & AMD64_PKRU_EXCL) != 0 &&
10130 !rangeset_check_empty(&pmap->pm_pkru, sva, eva))
10132 ppr = uma_zalloc(pmap_pkru_ranges_zone, M_NOWAIT);
10135 ppr->pkru_keyidx = keyidx;
10136 ppr->pkru_flags = flags & AMD64_PKRU_PERSIST;
10137 error = rangeset_insert(&pmap->pm_pkru, sva, eva, ppr);
10139 uma_zfree(pmap_pkru_ranges_zone, ppr);
10144 pmap_pkru_deassign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
10147 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
10148 MPASS(pmap->pm_type == PT_X86);
10149 MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0);
10150 return (rangeset_remove(&pmap->pm_pkru, sva, eva));
10154 pmap_pkru_deassign_all(pmap_t pmap)
10157 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
10158 if (pmap->pm_type == PT_X86 &&
10159 (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0)
10160 rangeset_remove_all(&pmap->pm_pkru);
10164 pmap_pkru_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
10166 struct pmap_pkru_range *ppr, *prev_ppr;
10169 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
10170 if (pmap->pm_type != PT_X86 ||
10171 (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0 ||
10172 sva >= VM_MAXUSER_ADDRESS)
10174 MPASS(eva <= VM_MAXUSER_ADDRESS);
10175 for (va = sva, prev_ppr = NULL; va < eva;) {
10176 ppr = rangeset_lookup(&pmap->pm_pkru, va);
10177 if ((ppr == NULL) ^ (prev_ppr == NULL))
10183 if (prev_ppr->pkru_keyidx != ppr->pkru_keyidx)
10185 va = ppr->pkru_rs_el.re_end;
10191 pmap_pkru_get(pmap_t pmap, vm_offset_t va)
10193 struct pmap_pkru_range *ppr;
10195 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
10196 if (pmap->pm_type != PT_X86 ||
10197 (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0 ||
10198 va >= VM_MAXUSER_ADDRESS)
10200 ppr = rangeset_lookup(&pmap->pm_pkru, va);
10202 return (X86_PG_PKU(ppr->pkru_keyidx));
10207 pred_pkru_on_remove(void *ctx __unused, void *r)
10209 struct pmap_pkru_range *ppr;
10212 return ((ppr->pkru_flags & AMD64_PKRU_PERSIST) == 0);
10216 pmap_pkru_on_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
10219 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
10220 if (pmap->pm_type == PT_X86 &&
10221 (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) {
10222 rangeset_remove_pred(&pmap->pm_pkru, sva, eva,
10223 pred_pkru_on_remove);
10228 pmap_pkru_copy(pmap_t dst_pmap, pmap_t src_pmap)
10231 PMAP_LOCK_ASSERT(dst_pmap, MA_OWNED);
10232 PMAP_LOCK_ASSERT(src_pmap, MA_OWNED);
10233 MPASS(dst_pmap->pm_type == PT_X86);
10234 MPASS(src_pmap->pm_type == PT_X86);
10235 MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0);
10236 if (src_pmap->pm_pkru.rs_data_ctx == NULL)
10238 return (rangeset_copy(&dst_pmap->pm_pkru, &src_pmap->pm_pkru));
10242 pmap_pkru_update_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
10245 pml4_entry_t *pml4e;
10247 pd_entry_t newpde, ptpaddr, *pde;
10248 pt_entry_t newpte, *ptep, pte;
10249 vm_offset_t va, va_next;
10252 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
10253 MPASS(pmap->pm_type == PT_X86);
10254 MPASS(keyidx <= PMAP_MAX_PKRU_IDX);
10256 for (changed = false, va = sva; va < eva; va = va_next) {
10257 pml4e = pmap_pml4e(pmap, va);
10258 if ((*pml4e & X86_PG_V) == 0) {
10259 va_next = (va + NBPML4) & ~PML4MASK;
10265 pdpe = pmap_pml4e_to_pdpe(pml4e, va);
10266 if ((*pdpe & X86_PG_V) == 0) {
10267 va_next = (va + NBPDP) & ~PDPMASK;
10273 va_next = (va + NBPDR) & ~PDRMASK;
10277 pde = pmap_pdpe_to_pde(pdpe, va);
10282 MPASS((ptpaddr & X86_PG_V) != 0);
10283 if ((ptpaddr & PG_PS) != 0) {
10284 if (va + NBPDR == va_next && eva >= va_next) {
10285 newpde = (ptpaddr & ~X86_PG_PKU_MASK) |
10286 X86_PG_PKU(keyidx);
10287 if (newpde != ptpaddr) {
10292 } else if (!pmap_demote_pde(pmap, pde, va)) {
10300 for (ptep = pmap_pde_to_pte(pde, va); va != va_next;
10301 ptep++, va += PAGE_SIZE) {
10303 if ((pte & X86_PG_V) == 0)
10305 newpte = (pte & ~X86_PG_PKU_MASK) | X86_PG_PKU(keyidx);
10306 if (newpte != pte) {
10313 pmap_invalidate_range(pmap, sva, eva);
10317 pmap_pkru_check_uargs(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
10318 u_int keyidx, int flags)
10321 if (pmap->pm_type != PT_X86 || keyidx > PMAP_MAX_PKRU_IDX ||
10322 (flags & ~(AMD64_PKRU_PERSIST | AMD64_PKRU_EXCL)) != 0)
10324 if (eva <= sva || eva > VM_MAXUSER_ADDRESS)
10326 if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0)
10332 pmap_pkru_set(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx,
10337 sva = trunc_page(sva);
10338 eva = round_page(eva);
10339 error = pmap_pkru_check_uargs(pmap, sva, eva, keyidx, flags);
10344 error = pmap_pkru_assign(pmap, sva, eva, keyidx, flags);
10346 pmap_pkru_update_range(pmap, sva, eva, keyidx);
10348 if (error != ENOMEM)
10356 pmap_pkru_clear(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
10360 sva = trunc_page(sva);
10361 eva = round_page(eva);
10362 error = pmap_pkru_check_uargs(pmap, sva, eva, 0, 0);
10367 error = pmap_pkru_deassign(pmap, sva, eva);
10369 pmap_pkru_update_range(pmap, sva, eva, 0);
10371 if (error != ENOMEM)
10379 * Track a range of the kernel's virtual address space that is contiguous
10380 * in various mapping attributes.
10382 struct pmap_kernel_map_range {
10391 sysctl_kmaps_dump(struct sbuf *sb, struct pmap_kernel_map_range *range,
10397 if (eva <= range->sva)
10400 pat_idx = pmap_pat_index(kernel_pmap, range->attrs, true);
10401 for (i = 0; i < PAT_INDEX_SIZE; i++)
10402 if (pat_index[i] == pat_idx)
10406 case PAT_WRITE_BACK:
10409 case PAT_WRITE_THROUGH:
10412 case PAT_UNCACHEABLE:
10418 case PAT_WRITE_PROTECTED:
10421 case PAT_WRITE_COMBINING:
10425 printf("%s: unknown PAT mode %#x for range 0x%016lx-0x%016lx\n",
10426 __func__, pat_idx, range->sva, eva);
10431 sbuf_printf(sb, "0x%016lx-0x%016lx r%c%c%c%c %s %d %d %d\n",
10433 (range->attrs & X86_PG_RW) != 0 ? 'w' : '-',
10434 (range->attrs & pg_nx) != 0 ? '-' : 'x',
10435 (range->attrs & X86_PG_U) != 0 ? 'u' : 's',
10436 (range->attrs & X86_PG_G) != 0 ? 'g' : '-',
10437 mode, range->pdpes, range->pdes, range->ptes);
10439 /* Reset to sentinel value. */
10440 range->sva = KVADDR(NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1);
10444 * Determine whether the attributes specified by a page table entry match those
10445 * being tracked by the current range. This is not quite as simple as a direct
10446 * flag comparison since some PAT modes have multiple representations.
10449 sysctl_kmaps_match(struct pmap_kernel_map_range *range, pt_entry_t attrs)
10451 pt_entry_t diff, mask;
10453 mask = X86_PG_G | X86_PG_RW | X86_PG_U | X86_PG_PDE_CACHE | pg_nx;
10454 diff = (range->attrs ^ attrs) & mask;
10457 if ((diff & ~X86_PG_PDE_PAT) == 0 &&
10458 pmap_pat_index(kernel_pmap, range->attrs, true) ==
10459 pmap_pat_index(kernel_pmap, attrs, true))
10465 sysctl_kmaps_reinit(struct pmap_kernel_map_range *range, vm_offset_t va,
10469 memset(range, 0, sizeof(*range));
10471 range->attrs = attrs;
10475 * Given a leaf PTE, derive the mapping's attributes. If they do not match
10476 * those of the current run, dump the address range and its attributes, and
10480 sysctl_kmaps_check(struct sbuf *sb, struct pmap_kernel_map_range *range,
10481 vm_offset_t va, pml4_entry_t pml4e, pdp_entry_t pdpe, pd_entry_t pde,
10486 attrs = pml4e & (X86_PG_RW | X86_PG_U | pg_nx);
10488 attrs |= pdpe & pg_nx;
10489 attrs &= pg_nx | (pdpe & (X86_PG_RW | X86_PG_U));
10490 if ((pdpe & PG_PS) != 0) {
10491 attrs |= pdpe & (X86_PG_G | X86_PG_PDE_CACHE);
10492 } else if (pde != 0) {
10493 attrs |= pde & pg_nx;
10494 attrs &= pg_nx | (pde & (X86_PG_RW | X86_PG_U));
10496 if ((pde & PG_PS) != 0) {
10497 attrs |= pde & (X86_PG_G | X86_PG_PDE_CACHE);
10498 } else if (pte != 0) {
10499 attrs |= pte & pg_nx;
10500 attrs &= pg_nx | (pte & (X86_PG_RW | X86_PG_U));
10501 attrs |= pte & (X86_PG_G | X86_PG_PTE_CACHE);
10503 /* Canonicalize by always using the PDE PAT bit. */
10504 if ((attrs & X86_PG_PTE_PAT) != 0)
10505 attrs ^= X86_PG_PDE_PAT | X86_PG_PTE_PAT;
10508 if (range->sva > va || !sysctl_kmaps_match(range, attrs)) {
10509 sysctl_kmaps_dump(sb, range, va);
10510 sysctl_kmaps_reinit(range, va, attrs);
10515 sysctl_kmaps(SYSCTL_HANDLER_ARGS)
10517 struct pmap_kernel_map_range range;
10518 struct sbuf sbuf, *sb;
10519 pml4_entry_t pml4e;
10520 pdp_entry_t *pdp, pdpe;
10521 pd_entry_t *pd, pde;
10522 pt_entry_t *pt, pte;
10525 int error, i, j, k, l;
10527 error = sysctl_wire_old_buffer(req, 0);
10531 sbuf_new_for_sysctl(sb, NULL, PAGE_SIZE, req);
10533 /* Sentinel value. */
10534 range.sva = KVADDR(NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1);
10537 * Iterate over the kernel page tables without holding the kernel pmap
10538 * lock. Outside of the large map, kernel page table pages are never
10539 * freed, so at worst we will observe inconsistencies in the output.
10540 * Within the large map, ensure that PDP and PD page addresses are
10541 * valid before descending.
10543 for (sva = 0, i = pmap_pml4e_index(sva); i < NPML4EPG; i++) {
10546 sbuf_printf(sb, "\nRecursive map:\n");
10549 sbuf_printf(sb, "\nDirect map:\n");
10552 sbuf_printf(sb, "\nKernel map:\n");
10555 sbuf_printf(sb, "\nLarge map:\n");
10559 /* Convert to canonical form. */
10560 if (sva == 1ul << 47)
10564 pml4e = kernel_pmap->pm_pml4[i];
10565 if ((pml4e & X86_PG_V) == 0) {
10566 sva = rounddown2(sva, NBPML4);
10567 sysctl_kmaps_dump(sb, &range, sva);
10571 pa = pml4e & PG_FRAME;
10572 pdp = (pdp_entry_t *)PHYS_TO_DMAP(pa);
10574 for (j = pmap_pdpe_index(sva); j < NPDPEPG; j++) {
10576 if ((pdpe & X86_PG_V) == 0) {
10577 sva = rounddown2(sva, NBPDP);
10578 sysctl_kmaps_dump(sb, &range, sva);
10582 pa = pdpe & PG_FRAME;
10583 if (PMAP_ADDRESS_IN_LARGEMAP(sva) &&
10584 vm_phys_paddr_to_vm_page(pa) == NULL)
10586 if ((pdpe & PG_PS) != 0) {
10587 sva = rounddown2(sva, NBPDP);
10588 sysctl_kmaps_check(sb, &range, sva, pml4e, pdpe,
10594 pd = (pd_entry_t *)PHYS_TO_DMAP(pa);
10596 for (k = pmap_pde_index(sva); k < NPDEPG; k++) {
10598 if ((pde & X86_PG_V) == 0) {
10599 sva = rounddown2(sva, NBPDR);
10600 sysctl_kmaps_dump(sb, &range, sva);
10604 pa = pde & PG_FRAME;
10605 if (PMAP_ADDRESS_IN_LARGEMAP(sva) &&
10606 vm_phys_paddr_to_vm_page(pa) == NULL)
10608 if ((pde & PG_PS) != 0) {
10609 sva = rounddown2(sva, NBPDR);
10610 sysctl_kmaps_check(sb, &range, sva,
10611 pml4e, pdpe, pde, 0);
10616 pt = (pt_entry_t *)PHYS_TO_DMAP(pa);
10618 for (l = pmap_pte_index(sva); l < NPTEPG; l++,
10619 sva += PAGE_SIZE) {
10621 if ((pte & X86_PG_V) == 0) {
10622 sysctl_kmaps_dump(sb, &range,
10626 sysctl_kmaps_check(sb, &range, sva,
10627 pml4e, pdpe, pde, pte);
10634 error = sbuf_finish(sb);
10638 SYSCTL_OID(_vm_pmap, OID_AUTO, kernel_maps,
10639 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
10640 NULL, 0, sysctl_kmaps, "A",
10641 "Dump kernel address layout");
10644 DB_SHOW_COMMAND(pte, pmap_print_pte)
10647 pml4_entry_t *pml4;
10650 pt_entry_t *pte, PG_V;
10654 db_printf("show pte addr\n");
10657 va = (vm_offset_t)addr;
10659 if (kdb_thread != NULL)
10660 pmap = vmspace_pmap(kdb_thread->td_proc->p_vmspace);
10662 pmap = PCPU_GET(curpmap);
10664 PG_V = pmap_valid_bit(pmap);
10665 pml4 = pmap_pml4e(pmap, va);
10666 db_printf("VA 0x%016lx pml4e 0x%016lx", va, *pml4);
10667 if ((*pml4 & PG_V) == 0) {
10671 pdp = pmap_pml4e_to_pdpe(pml4, va);
10672 db_printf(" pdpe 0x%016lx", *pdp);
10673 if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0) {
10677 pde = pmap_pdpe_to_pde(pdp, va);
10678 db_printf(" pde 0x%016lx", *pde);
10679 if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0) {
10683 pte = pmap_pde_to_pte(pde, va);
10684 db_printf(" pte 0x%016lx\n", *pte);
10687 DB_SHOW_COMMAND(phys2dmap, pmap_phys2dmap)
10692 a = (vm_paddr_t)addr;
10693 db_printf("0x%jx\n", (uintmax_t)PHYS_TO_DMAP(a));
10695 db_printf("show phys2dmap addr\n");