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) 2014-2018 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.
110 #include "opt_pmap.h"
113 #include <sys/param.h>
114 #include <sys/bitstring.h>
116 #include <sys/systm.h>
117 #include <sys/kernel.h>
119 #include <sys/lock.h>
120 #include <sys/malloc.h>
121 #include <sys/mman.h>
122 #include <sys/mutex.h>
123 #include <sys/proc.h>
124 #include <sys/rwlock.h>
126 #include <sys/turnstile.h>
127 #include <sys/vmem.h>
128 #include <sys/vmmeter.h>
129 #include <sys/sched.h>
130 #include <sys/sysctl.h>
134 #include <vm/vm_param.h>
135 #include <vm/vm_kern.h>
136 #include <vm/vm_page.h>
137 #include <vm/vm_map.h>
138 #include <vm/vm_object.h>
139 #include <vm/vm_extern.h>
140 #include <vm/vm_pageout.h>
141 #include <vm/vm_pager.h>
142 #include <vm/vm_phys.h>
143 #include <vm/vm_radix.h>
144 #include <vm/vm_reserv.h>
147 #include <machine/intr_machdep.h>
148 #include <x86/apicvar.h>
149 #include <x86/ifunc.h>
150 #include <machine/cpu.h>
151 #include <machine/cputypes.h>
152 #include <machine/md_var.h>
153 #include <machine/pcb.h>
154 #include <machine/specialreg.h>
156 #include <machine/smp.h>
158 #include <machine/tss.h>
160 static __inline boolean_t
161 pmap_type_guest(pmap_t pmap)
164 return ((pmap->pm_type == PT_EPT) || (pmap->pm_type == PT_RVI));
167 static __inline boolean_t
168 pmap_emulate_ad_bits(pmap_t pmap)
171 return ((pmap->pm_flags & PMAP_EMULATE_AD_BITS) != 0);
174 static __inline pt_entry_t
175 pmap_valid_bit(pmap_t pmap)
179 switch (pmap->pm_type) {
185 if (pmap_emulate_ad_bits(pmap))
186 mask = EPT_PG_EMUL_V;
191 panic("pmap_valid_bit: invalid pm_type %d", pmap->pm_type);
197 static __inline pt_entry_t
198 pmap_rw_bit(pmap_t pmap)
202 switch (pmap->pm_type) {
208 if (pmap_emulate_ad_bits(pmap))
209 mask = EPT_PG_EMUL_RW;
214 panic("pmap_rw_bit: invalid pm_type %d", pmap->pm_type);
220 static pt_entry_t pg_g;
222 static __inline pt_entry_t
223 pmap_global_bit(pmap_t pmap)
227 switch (pmap->pm_type) {
236 panic("pmap_global_bit: invalid pm_type %d", pmap->pm_type);
242 static __inline pt_entry_t
243 pmap_accessed_bit(pmap_t pmap)
247 switch (pmap->pm_type) {
253 if (pmap_emulate_ad_bits(pmap))
259 panic("pmap_accessed_bit: invalid pm_type %d", pmap->pm_type);
265 static __inline pt_entry_t
266 pmap_modified_bit(pmap_t pmap)
270 switch (pmap->pm_type) {
276 if (pmap_emulate_ad_bits(pmap))
282 panic("pmap_modified_bit: invalid pm_type %d", pmap->pm_type);
288 #if !defined(DIAGNOSTIC)
289 #ifdef __GNUC_GNU_INLINE__
290 #define PMAP_INLINE __attribute__((__gnu_inline__)) inline
292 #define PMAP_INLINE extern inline
299 #define PV_STAT(x) do { x ; } while (0)
301 #define PV_STAT(x) do { } while (0)
304 #define pa_index(pa) ((pa) >> PDRSHIFT)
305 #define pa_to_pvh(pa) (&pv_table[pa_index(pa)])
307 #define NPV_LIST_LOCKS MAXCPU
309 #define PHYS_TO_PV_LIST_LOCK(pa) \
310 (&pv_list_locks[pa_index(pa) % NPV_LIST_LOCKS])
312 #define CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa) do { \
313 struct rwlock **_lockp = (lockp); \
314 struct rwlock *_new_lock; \
316 _new_lock = PHYS_TO_PV_LIST_LOCK(pa); \
317 if (_new_lock != *_lockp) { \
318 if (*_lockp != NULL) \
319 rw_wunlock(*_lockp); \
320 *_lockp = _new_lock; \
325 #define CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m) \
326 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, VM_PAGE_TO_PHYS(m))
328 #define RELEASE_PV_LIST_LOCK(lockp) do { \
329 struct rwlock **_lockp = (lockp); \
331 if (*_lockp != NULL) { \
332 rw_wunlock(*_lockp); \
337 #define VM_PAGE_TO_PV_LIST_LOCK(m) \
338 PHYS_TO_PV_LIST_LOCK(VM_PAGE_TO_PHYS(m))
340 struct pmap kernel_pmap_store;
342 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
343 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
346 SYSCTL_INT(_machdep, OID_AUTO, nkpt, CTLFLAG_RD, &nkpt, 0,
347 "Number of kernel page table pages allocated on bootup");
350 vm_paddr_t dmaplimit;
351 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
354 static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters");
356 static int pat_works = 1;
357 SYSCTL_INT(_vm_pmap, OID_AUTO, pat_works, CTLFLAG_RD, &pat_works, 1,
358 "Is page attribute table fully functional?");
360 static int pg_ps_enabled = 1;
361 SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
362 &pg_ps_enabled, 0, "Are large page mappings enabled?");
364 #define PAT_INDEX_SIZE 8
365 static int pat_index[PAT_INDEX_SIZE]; /* cache mode to PAT index conversion */
367 static u_int64_t KPTphys; /* phys addr of kernel level 1 */
368 static u_int64_t KPDphys; /* phys addr of kernel level 2 */
369 u_int64_t KPDPphys; /* phys addr of kernel level 3 */
370 u_int64_t KPML4phys; /* phys addr of kernel level 4 */
372 static u_int64_t DMPDphys; /* phys addr of direct mapped level 2 */
373 static u_int64_t DMPDPphys; /* phys addr of direct mapped level 3 */
374 static int ndmpdpphys; /* number of DMPDPphys pages */
376 static vm_paddr_t KERNend; /* phys addr of end of bootstrap data */
379 * pmap_mapdev support pre initialization (i.e. console)
381 #define PMAP_PREINIT_MAPPING_COUNT 8
382 static struct pmap_preinit_mapping {
387 } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT];
388 static int pmap_initialized;
391 * Data for the pv entry allocation mechanism.
392 * Updates to pv_invl_gen are protected by the pv_list_locks[]
393 * elements, but reads are not.
395 static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks);
396 static struct mtx __exclusive_cache_line pv_chunks_mutex;
397 static struct rwlock __exclusive_cache_line pv_list_locks[NPV_LIST_LOCKS];
398 static u_long pv_invl_gen[NPV_LIST_LOCKS];
399 static struct md_page *pv_table;
400 static struct md_page pv_dummy;
403 * All those kernel PT submaps that BSD is so fond of
405 pt_entry_t *CMAP1 = NULL;
407 static vm_offset_t qframe = 0;
408 static struct mtx qframe_mtx;
410 static int pmap_flags = PMAP_PDE_SUPERPAGE; /* flags for x86 pmaps */
412 static vmem_t *large_vmem;
413 static u_int lm_ents;
415 int pmap_pcid_enabled = 1;
416 SYSCTL_INT(_vm_pmap, OID_AUTO, pcid_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
417 &pmap_pcid_enabled, 0, "Is TLB Context ID enabled ?");
418 int invpcid_works = 0;
419 SYSCTL_INT(_vm_pmap, OID_AUTO, invpcid_works, CTLFLAG_RD, &invpcid_works, 0,
420 "Is the invpcid instruction available ?");
422 int __read_frequently pti = 0;
423 SYSCTL_INT(_vm_pmap, OID_AUTO, pti, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
425 "Page Table Isolation enabled");
426 static vm_object_t pti_obj;
427 static pml4_entry_t *pti_pml4;
428 static vm_pindex_t pti_pg_idx;
429 static bool pti_finalized;
432 pmap_pcid_save_cnt_proc(SYSCTL_HANDLER_ARGS)
439 res += cpuid_to_pcpu[i]->pc_pm_save_cnt;
441 return (sysctl_handle_64(oidp, &res, 0, req));
443 SYSCTL_PROC(_vm_pmap, OID_AUTO, pcid_save_cnt, CTLTYPE_U64 | CTLFLAG_RW |
444 CTLFLAG_MPSAFE, NULL, 0, pmap_pcid_save_cnt_proc, "QU",
445 "Count of saved TLB context on switch");
447 static LIST_HEAD(, pmap_invl_gen) pmap_invl_gen_tracker =
448 LIST_HEAD_INITIALIZER(&pmap_invl_gen_tracker);
449 static struct mtx invl_gen_mtx;
450 static u_long pmap_invl_gen = 0;
451 /* Fake lock object to satisfy turnstiles interface. */
452 static struct lock_object invl_gen_ts = {
460 return (curthread->td_md.md_invl_gen.gen == 0);
463 #define PMAP_ASSERT_NOT_IN_DI() \
464 KASSERT(pmap_not_in_di(), ("DI already started"))
467 * Start a new Delayed Invalidation (DI) block of code, executed by
468 * the current thread. Within a DI block, the current thread may
469 * destroy both the page table and PV list entries for a mapping and
470 * then release the corresponding PV list lock before ensuring that
471 * the mapping is flushed from the TLBs of any processors with the
475 pmap_delayed_invl_started(void)
477 struct pmap_invl_gen *invl_gen;
480 invl_gen = &curthread->td_md.md_invl_gen;
481 PMAP_ASSERT_NOT_IN_DI();
482 mtx_lock(&invl_gen_mtx);
483 if (LIST_EMPTY(&pmap_invl_gen_tracker))
484 currgen = pmap_invl_gen;
486 currgen = LIST_FIRST(&pmap_invl_gen_tracker)->gen;
487 invl_gen->gen = currgen + 1;
488 LIST_INSERT_HEAD(&pmap_invl_gen_tracker, invl_gen, link);
489 mtx_unlock(&invl_gen_mtx);
493 * Finish the DI block, previously started by the current thread. All
494 * required TLB flushes for the pages marked by
495 * pmap_delayed_invl_page() must be finished before this function is
498 * This function works by bumping the global DI generation number to
499 * the generation number of the current thread's DI, unless there is a
500 * pending DI that started earlier. In the latter case, bumping the
501 * global DI generation number would incorrectly signal that the
502 * earlier DI had finished. Instead, this function bumps the earlier
503 * DI's generation number to match the generation number of the
504 * current thread's DI.
507 pmap_delayed_invl_finished(void)
509 struct pmap_invl_gen *invl_gen, *next;
510 struct turnstile *ts;
512 invl_gen = &curthread->td_md.md_invl_gen;
513 KASSERT(invl_gen->gen != 0, ("missed invl_started"));
514 mtx_lock(&invl_gen_mtx);
515 next = LIST_NEXT(invl_gen, link);
517 turnstile_chain_lock(&invl_gen_ts);
518 ts = turnstile_lookup(&invl_gen_ts);
519 pmap_invl_gen = invl_gen->gen;
521 turnstile_broadcast(ts, TS_SHARED_QUEUE);
522 turnstile_unpend(ts);
524 turnstile_chain_unlock(&invl_gen_ts);
526 next->gen = invl_gen->gen;
528 LIST_REMOVE(invl_gen, link);
529 mtx_unlock(&invl_gen_mtx);
534 static long invl_wait;
535 SYSCTL_LONG(_vm_pmap, OID_AUTO, invl_wait, CTLFLAG_RD, &invl_wait, 0,
536 "Number of times DI invalidation blocked pmap_remove_all/write");
540 pmap_delayed_invl_genp(vm_page_t m)
543 return (&pv_invl_gen[pa_index(VM_PAGE_TO_PHYS(m)) % NPV_LIST_LOCKS]);
547 * Ensure that all currently executing DI blocks, that need to flush
548 * TLB for the given page m, actually flushed the TLB at the time the
549 * function returned. If the page m has an empty PV list and we call
550 * pmap_delayed_invl_wait(), upon its return we know that no CPU has a
551 * valid mapping for the page m in either its page table or TLB.
553 * This function works by blocking until the global DI generation
554 * number catches up with the generation number associated with the
555 * given page m and its PV list. Since this function's callers
556 * typically own an object lock and sometimes own a page lock, it
557 * cannot sleep. Instead, it blocks on a turnstile to relinquish the
561 pmap_delayed_invl_wait(vm_page_t m)
563 struct turnstile *ts;
566 bool accounted = false;
569 m_gen = pmap_delayed_invl_genp(m);
570 while (*m_gen > pmap_invl_gen) {
573 atomic_add_long(&invl_wait, 1);
577 ts = turnstile_trywait(&invl_gen_ts);
578 if (*m_gen > pmap_invl_gen)
579 turnstile_wait(ts, NULL, TS_SHARED_QUEUE);
581 turnstile_cancel(ts);
586 * Mark the page m's PV list as participating in the current thread's
587 * DI block. Any threads concurrently using m's PV list to remove or
588 * restrict all mappings to m will wait for the current thread's DI
589 * block to complete before proceeding.
591 * The function works by setting the DI generation number for m's PV
592 * list to at least the DI generation number of the current thread.
593 * This forces a caller of pmap_delayed_invl_wait() to block until
594 * current thread calls pmap_delayed_invl_finished().
597 pmap_delayed_invl_page(vm_page_t m)
601 rw_assert(VM_PAGE_TO_PV_LIST_LOCK(m), RA_WLOCKED);
602 gen = curthread->td_md.md_invl_gen.gen;
605 m_gen = pmap_delayed_invl_genp(m);
613 static caddr_t crashdumpmap;
616 * Internal flags for pmap_enter()'s helper functions.
618 #define PMAP_ENTER_NORECLAIM 0x1000000 /* Don't reclaim PV entries. */
619 #define PMAP_ENTER_NOREPLACE 0x2000000 /* Don't replace mappings. */
621 static void free_pv_chunk(struct pv_chunk *pc);
622 static void free_pv_entry(pmap_t pmap, pv_entry_t pv);
623 static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp);
624 static int popcnt_pc_map_pq(uint64_t *map);
625 static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp);
626 static void reserve_pv_entries(pmap_t pmap, int needed,
627 struct rwlock **lockp);
628 static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
629 struct rwlock **lockp);
630 static bool pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde,
631 u_int flags, struct rwlock **lockp);
632 #if VM_NRESERVLEVEL > 0
633 static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
634 struct rwlock **lockp);
636 static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va);
637 static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap,
640 static int pmap_change_attr_locked(vm_offset_t va, vm_size_t size, int mode,
642 static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va);
643 static boolean_t pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde,
644 vm_offset_t va, struct rwlock **lockp);
645 static boolean_t pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe,
647 static bool pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m,
648 vm_prot_t prot, struct rwlock **lockp);
649 static int pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde,
650 u_int flags, vm_page_t m, struct rwlock **lockp);
651 static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va,
652 vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp);
653 static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte);
654 static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte);
655 static void pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva,
657 static void pmap_invalidate_cache_range_all(vm_offset_t sva,
659 static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va,
661 static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode);
662 static vm_page_t pmap_large_map_getptp_unlocked(void);
663 static void pmap_pde_attr(pd_entry_t *pde, int cache_bits, int mask);
664 #if VM_NRESERVLEVEL > 0
665 static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va,
666 struct rwlock **lockp);
668 static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva,
670 static void pmap_pte_attr(pt_entry_t *pte, int cache_bits, int mask);
671 static void pmap_pti_add_kva_locked(vm_offset_t sva, vm_offset_t eva,
673 static pdp_entry_t *pmap_pti_pdpe(vm_offset_t va);
674 static pd_entry_t *pmap_pti_pde(vm_offset_t va);
675 static void pmap_pti_wire_pte(void *pte);
676 static int pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
677 struct spglist *free, struct rwlock **lockp);
678 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva,
679 pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp);
680 static vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va);
681 static void pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
682 struct spglist *free);
683 static bool pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
684 pd_entry_t *pde, struct spglist *free,
685 struct rwlock **lockp);
686 static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va,
687 vm_page_t m, struct rwlock **lockp);
688 static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
690 static void pmap_update_pde_invalidate(pmap_t, vm_offset_t va, pd_entry_t pde);
692 static vm_page_t _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
693 struct rwlock **lockp);
694 static vm_page_t pmap_allocpde(pmap_t pmap, vm_offset_t va,
695 struct rwlock **lockp);
696 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va,
697 struct rwlock **lockp);
699 static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m,
700 struct spglist *free);
701 static int pmap_unuse_pt(pmap_t, vm_offset_t, pd_entry_t, struct spglist *);
703 /********************/
704 /* Inline functions */
705 /********************/
707 /* Return a non-clipped PD index for a given VA */
708 static __inline vm_pindex_t
709 pmap_pde_pindex(vm_offset_t va)
711 return (va >> PDRSHIFT);
715 /* Return a pointer to the PML4 slot that corresponds to a VA */
716 static __inline pml4_entry_t *
717 pmap_pml4e(pmap_t pmap, vm_offset_t va)
720 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
723 /* Return a pointer to the PDP slot that corresponds to a VA */
724 static __inline pdp_entry_t *
725 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
729 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
730 return (&pdpe[pmap_pdpe_index(va)]);
733 /* Return a pointer to the PDP slot that corresponds to a VA */
734 static __inline pdp_entry_t *
735 pmap_pdpe(pmap_t pmap, vm_offset_t va)
740 PG_V = pmap_valid_bit(pmap);
741 pml4e = pmap_pml4e(pmap, va);
742 if ((*pml4e & PG_V) == 0)
744 return (pmap_pml4e_to_pdpe(pml4e, va));
747 /* Return a pointer to the PD slot that corresponds to a VA */
748 static __inline pd_entry_t *
749 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
753 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
754 return (&pde[pmap_pde_index(va)]);
757 /* Return a pointer to the PD slot that corresponds to a VA */
758 static __inline pd_entry_t *
759 pmap_pde(pmap_t pmap, vm_offset_t va)
764 PG_V = pmap_valid_bit(pmap);
765 pdpe = pmap_pdpe(pmap, va);
766 if (pdpe == NULL || (*pdpe & PG_V) == 0)
768 return (pmap_pdpe_to_pde(pdpe, va));
771 /* Return a pointer to the PT slot that corresponds to a VA */
772 static __inline pt_entry_t *
773 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
777 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
778 return (&pte[pmap_pte_index(va)]);
781 /* Return a pointer to the PT slot that corresponds to a VA */
782 static __inline pt_entry_t *
783 pmap_pte(pmap_t pmap, vm_offset_t va)
788 PG_V = pmap_valid_bit(pmap);
789 pde = pmap_pde(pmap, va);
790 if (pde == NULL || (*pde & PG_V) == 0)
792 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
793 return ((pt_entry_t *)pde);
794 return (pmap_pde_to_pte(pde, va));
798 pmap_resident_count_inc(pmap_t pmap, int count)
801 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
802 pmap->pm_stats.resident_count += count;
806 pmap_resident_count_dec(pmap_t pmap, int count)
809 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
810 KASSERT(pmap->pm_stats.resident_count >= count,
811 ("pmap %p resident count underflow %ld %d", pmap,
812 pmap->pm_stats.resident_count, count));
813 pmap->pm_stats.resident_count -= count;
816 PMAP_INLINE pt_entry_t *
817 vtopte(vm_offset_t va)
819 u_int64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
821 KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopte on a uva/gpa 0x%0lx", va));
823 return (PTmap + ((va >> PAGE_SHIFT) & mask));
826 static __inline pd_entry_t *
827 vtopde(vm_offset_t va)
829 u_int64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
831 KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopde on a uva/gpa 0x%0lx", va));
833 return (PDmap + ((va >> PDRSHIFT) & mask));
837 allocpages(vm_paddr_t *firstaddr, int n)
842 bzero((void *)ret, n * PAGE_SIZE);
843 *firstaddr += n * PAGE_SIZE;
847 CTASSERT(powerof2(NDMPML4E));
849 /* number of kernel PDP slots */
850 #define NKPDPE(ptpgs) howmany(ptpgs, NPDEPG)
853 nkpt_init(vm_paddr_t addr)
860 pt_pages = howmany(addr, 1 << PDRSHIFT);
861 pt_pages += NKPDPE(pt_pages);
864 * Add some slop beyond the bare minimum required for bootstrapping
867 * This is quite important when allocating KVA for kernel modules.
868 * The modules are required to be linked in the negative 2GB of
869 * the address space. If we run out of KVA in this region then
870 * pmap_growkernel() will need to allocate page table pages to map
871 * the entire 512GB of KVA space which is an unnecessary tax on
874 * Secondly, device memory mapped as part of setting up the low-
875 * level console(s) is taken from KVA, starting at virtual_avail.
876 * This is because cninit() is called after pmap_bootstrap() but
877 * before vm_init() and pmap_init(). 20MB for a frame buffer is
880 pt_pages += 32; /* 64MB additional slop. */
886 * Returns the proper write/execute permission for a physical page that is
887 * part of the initial boot allocations.
889 * If the page has kernel text, it is marked as read-only. If the page has
890 * kernel read-only data, it is marked as read-only/not-executable. If the
891 * page has only read-write data, it is marked as read-write/not-executable.
892 * If the page is below/above the kernel range, it is marked as read-write.
894 * This function operates on 2M pages, since we map the kernel space that
897 * Note that this doesn't currently provide any protection for modules.
899 static inline pt_entry_t
900 bootaddr_rwx(vm_paddr_t pa)
904 * Everything in the same 2M page as the start of the kernel
905 * should be static. On the other hand, things in the same 2M
906 * page as the end of the kernel could be read-write/executable,
907 * as the kernel image is not guaranteed to end on a 2M boundary.
909 if (pa < trunc_2mpage(btext - KERNBASE) ||
910 pa >= trunc_2mpage(_end - KERNBASE))
913 * The linker should ensure that the read-only and read-write
914 * portions don't share the same 2M page, so this shouldn't
915 * impact read-only data. However, in any case, any page with
916 * read-write data needs to be read-write.
918 if (pa >= trunc_2mpage(brwsection - KERNBASE))
919 return (X86_PG_RW | pg_nx);
921 * Mark any 2M page containing kernel text as read-only. Mark
922 * other pages with read-only data as read-only and not executable.
923 * (It is likely a small portion of the read-only data section will
924 * be marked as read-only, but executable. This should be acceptable
925 * since the read-only protection will keep the data from changing.)
926 * Note that fixups to the .text section will still work until we
929 if (pa < round_2mpage(etext - KERNBASE))
935 create_pagetables(vm_paddr_t *firstaddr)
937 int i, j, ndm1g, nkpdpe, nkdmpde;
942 uint64_t DMPDkernphys;
944 /* Allocate page table pages for the direct map */
945 ndmpdp = howmany(ptoa(Maxmem), NBPDP);
946 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
948 ndmpdpphys = howmany(ndmpdp, NPDPEPG);
949 if (ndmpdpphys > NDMPML4E) {
951 * Each NDMPML4E allows 512 GB, so limit to that,
952 * and then readjust ndmpdp and ndmpdpphys.
954 printf("NDMPML4E limits system to %d GB\n", NDMPML4E * 512);
955 Maxmem = atop(NDMPML4E * NBPML4);
956 ndmpdpphys = NDMPML4E;
957 ndmpdp = NDMPML4E * NPDEPG;
959 DMPDPphys = allocpages(firstaddr, ndmpdpphys);
961 if ((amd_feature & AMDID_PAGE1GB) != 0) {
963 * Calculate the number of 1G pages that will fully fit in
966 ndm1g = ptoa(Maxmem) >> PDPSHIFT;
969 * Allocate 2M pages for the kernel. These will be used in
970 * place of the first one or more 1G pages from ndm1g.
972 nkdmpde = howmany((vm_offset_t)(brwsection - KERNBASE), NBPDP);
973 DMPDkernphys = allocpages(firstaddr, nkdmpde);
976 DMPDphys = allocpages(firstaddr, ndmpdp - ndm1g);
977 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
980 KPML4phys = allocpages(firstaddr, 1);
981 KPDPphys = allocpages(firstaddr, NKPML4E);
984 * Allocate the initial number of kernel page table pages required to
985 * bootstrap. We defer this until after all memory-size dependent
986 * allocations are done (e.g. direct map), so that we don't have to
987 * build in too much slop in our estimate.
989 * Note that when NKPML4E > 1, we have an empty page underneath
990 * all but the KPML4I'th one, so we need NKPML4E-1 extra (zeroed)
991 * pages. (pmap_enter requires a PD page to exist for each KPML4E.)
993 nkpt_init(*firstaddr);
994 nkpdpe = NKPDPE(nkpt);
996 KPTphys = allocpages(firstaddr, nkpt);
997 KPDphys = allocpages(firstaddr, nkpdpe);
999 /* Fill in the underlying page table pages */
1000 /* XXX not fully used, underneath 2M pages */
1001 pt_p = (pt_entry_t *)KPTphys;
1002 for (i = 0; ptoa(i) < *firstaddr; i++)
1003 pt_p[i] = ptoa(i) | X86_PG_V | pg_g | bootaddr_rwx(ptoa(i));
1005 /* Now map the page tables at their location within PTmap */
1006 pd_p = (pd_entry_t *)KPDphys;
1007 for (i = 0; i < nkpt; i++)
1008 pd_p[i] = (KPTphys + ptoa(i)) | X86_PG_RW | X86_PG_V;
1010 /* Map from zero to end of allocations under 2M pages */
1011 /* This replaces some of the KPTphys entries above */
1012 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++)
1013 /* Preset PG_M and PG_A because demotion expects it. */
1014 pd_p[i] = (i << PDRSHIFT) | X86_PG_V | PG_PS | pg_g |
1015 X86_PG_M | X86_PG_A | bootaddr_rwx(i << PDRSHIFT);
1018 * Because we map the physical blocks in 2M pages, adjust firstaddr
1019 * to record the physical blocks we've actually mapped into kernel
1020 * virtual address space.
1022 *firstaddr = round_2mpage(*firstaddr);
1024 /* And connect up the PD to the PDP (leaving room for L4 pages) */
1025 pdp_p = (pdp_entry_t *)(KPDPphys + ptoa(KPML4I - KPML4BASE));
1026 for (i = 0; i < nkpdpe; i++)
1027 pdp_p[i + KPDPI] = (KPDphys + ptoa(i)) | X86_PG_RW | X86_PG_V;
1030 * Now, set up the direct map region using 2MB and/or 1GB pages. If
1031 * the end of physical memory is not aligned to a 1GB page boundary,
1032 * then the residual physical memory is mapped with 2MB pages. Later,
1033 * if pmap_mapdev{_attr}() uses the direct map for non-write-back
1034 * memory, pmap_change_attr() will demote any 2MB or 1GB page mappings
1035 * that are partially used.
1037 pd_p = (pd_entry_t *)DMPDphys;
1038 for (i = NPDEPG * ndm1g, j = 0; i < NPDEPG * ndmpdp; i++, j++) {
1039 pd_p[j] = (vm_paddr_t)i << PDRSHIFT;
1040 /* Preset PG_M and PG_A because demotion expects it. */
1041 pd_p[j] |= X86_PG_RW | X86_PG_V | PG_PS | pg_g |
1042 X86_PG_M | X86_PG_A | pg_nx;
1044 pdp_p = (pdp_entry_t *)DMPDPphys;
1045 for (i = 0; i < ndm1g; i++) {
1046 pdp_p[i] = (vm_paddr_t)i << PDPSHIFT;
1047 /* Preset PG_M and PG_A because demotion expects it. */
1048 pdp_p[i] |= X86_PG_RW | X86_PG_V | PG_PS | pg_g |
1049 X86_PG_M | X86_PG_A | pg_nx;
1051 for (j = 0; i < ndmpdp; i++, j++) {
1052 pdp_p[i] = DMPDphys + ptoa(j);
1053 pdp_p[i] |= X86_PG_RW | X86_PG_V;
1057 * Instead of using a 1G page for the memory containing the kernel,
1058 * use 2M pages with appropriate permissions. (If using 1G pages,
1059 * this will partially overwrite the PDPEs above.)
1062 pd_p = (pd_entry_t *)DMPDkernphys;
1063 for (i = 0; i < (NPDEPG * nkdmpde); i++)
1064 pd_p[i] = (i << PDRSHIFT) | X86_PG_V | PG_PS | pg_g |
1065 X86_PG_M | X86_PG_A | pg_nx |
1066 bootaddr_rwx(i << PDRSHIFT);
1067 for (i = 0; i < nkdmpde; i++)
1068 pdp_p[i] = (DMPDkernphys + ptoa(i)) | X86_PG_RW |
1072 /* And recursively map PML4 to itself in order to get PTmap */
1073 p4_p = (pml4_entry_t *)KPML4phys;
1074 p4_p[PML4PML4I] = KPML4phys;
1075 p4_p[PML4PML4I] |= X86_PG_RW | X86_PG_V | pg_nx;
1077 /* Connect the Direct Map slot(s) up to the PML4. */
1078 for (i = 0; i < ndmpdpphys; i++) {
1079 p4_p[DMPML4I + i] = DMPDPphys + ptoa(i);
1080 p4_p[DMPML4I + i] |= X86_PG_RW | X86_PG_V;
1083 /* Connect the KVA slots up to the PML4 */
1084 for (i = 0; i < NKPML4E; i++) {
1085 p4_p[KPML4BASE + i] = KPDPphys + ptoa(i);
1086 p4_p[KPML4BASE + i] |= X86_PG_RW | X86_PG_V;
1091 * Bootstrap the system enough to run with virtual memory.
1093 * On amd64 this is called after mapping has already been enabled
1094 * and just syncs the pmap module with what has already been done.
1095 * [We can't call it easily with mapping off since the kernel is not
1096 * mapped with PA == VA, hence we would have to relocate every address
1097 * from the linked base (virtual) address "KERNBASE" to the actual
1098 * (physical) address starting relative to 0]
1101 pmap_bootstrap(vm_paddr_t *firstaddr)
1109 KERNend = *firstaddr;
1110 res = atop(KERNend - (vm_paddr_t)kernphys);
1116 * Create an initial set of page tables to run the kernel in.
1118 create_pagetables(firstaddr);
1121 * Add a physical memory segment (vm_phys_seg) corresponding to the
1122 * preallocated kernel page table pages so that vm_page structures
1123 * representing these pages will be created. The vm_page structures
1124 * are required for promotion of the corresponding kernel virtual
1125 * addresses to superpage mappings.
1127 vm_phys_add_seg(KPTphys, KPTphys + ptoa(nkpt));
1129 virtual_avail = (vm_offset_t) KERNBASE + *firstaddr;
1130 virtual_end = VM_MAX_KERNEL_ADDRESS;
1133 * Enable PG_G global pages, then switch to the kernel page
1134 * table from the bootstrap page table. After the switch, it
1135 * is possible to enable SMEP and SMAP since PG_U bits are
1141 load_cr3(KPML4phys);
1142 if (cpu_stdext_feature & CPUID_STDEXT_SMEP)
1144 if (cpu_stdext_feature & CPUID_STDEXT_SMAP)
1149 * Initialize the kernel pmap (which is statically allocated).
1150 * Count bootstrap data as being resident in case any of this data is
1151 * later unmapped (using pmap_remove()) and freed.
1153 PMAP_LOCK_INIT(kernel_pmap);
1154 kernel_pmap->pm_pml4 = (pdp_entry_t *)PHYS_TO_DMAP(KPML4phys);
1155 kernel_pmap->pm_cr3 = KPML4phys;
1156 kernel_pmap->pm_ucr3 = PMAP_NO_CR3;
1157 CPU_FILL(&kernel_pmap->pm_active); /* don't allow deactivation */
1158 TAILQ_INIT(&kernel_pmap->pm_pvchunk);
1159 kernel_pmap->pm_stats.resident_count = res;
1160 kernel_pmap->pm_flags = pmap_flags;
1163 * Initialize the TLB invalidations generation number lock.
1165 mtx_init(&invl_gen_mtx, "invlgn", NULL, MTX_DEF);
1168 * Reserve some special page table entries/VA space for temporary
1171 #define SYSMAP(c, p, v, n) \
1172 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
1178 * Crashdump maps. The first page is reused as CMAP1 for the
1181 SYSMAP(caddr_t, CMAP1, crashdumpmap, MAXDUMPPGS)
1182 CADDR1 = crashdumpmap;
1187 * Initialize the PAT MSR.
1188 * pmap_init_pat() clears and sets CR4_PGE, which, as a
1189 * side-effect, invalidates stale PG_G TLB entries that might
1190 * have been created in our pre-boot environment.
1194 /* Initialize TLB Context Id. */
1195 if (pmap_pcid_enabled) {
1196 for (i = 0; i < MAXCPU; i++) {
1197 kernel_pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN;
1198 kernel_pmap->pm_pcids[i].pm_gen = 1;
1202 * PMAP_PCID_KERN + 1 is used for initialization of
1203 * proc0 pmap. The pmap' pcid state might be used by
1204 * EFIRT entry before first context switch, so it
1205 * needs to be valid.
1207 PCPU_SET(pcid_next, PMAP_PCID_KERN + 2);
1208 PCPU_SET(pcid_gen, 1);
1211 * pcpu area for APs is zeroed during AP startup.
1212 * pc_pcid_next and pc_pcid_gen are initialized by AP
1213 * during pcpu setup.
1215 load_cr4(rcr4() | CR4_PCIDE);
1220 * Setup the PAT MSR.
1225 int pat_table[PAT_INDEX_SIZE];
1230 /* Bail if this CPU doesn't implement PAT. */
1231 if ((cpu_feature & CPUID_PAT) == 0)
1234 /* Set default PAT index table. */
1235 for (i = 0; i < PAT_INDEX_SIZE; i++)
1237 pat_table[PAT_WRITE_BACK] = 0;
1238 pat_table[PAT_WRITE_THROUGH] = 1;
1239 pat_table[PAT_UNCACHEABLE] = 3;
1240 pat_table[PAT_WRITE_COMBINING] = 3;
1241 pat_table[PAT_WRITE_PROTECTED] = 3;
1242 pat_table[PAT_UNCACHED] = 3;
1244 /* Initialize default PAT entries. */
1245 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |
1246 PAT_VALUE(1, PAT_WRITE_THROUGH) |
1247 PAT_VALUE(2, PAT_UNCACHED) |
1248 PAT_VALUE(3, PAT_UNCACHEABLE) |
1249 PAT_VALUE(4, PAT_WRITE_BACK) |
1250 PAT_VALUE(5, PAT_WRITE_THROUGH) |
1251 PAT_VALUE(6, PAT_UNCACHED) |
1252 PAT_VALUE(7, PAT_UNCACHEABLE);
1256 * Leave the indices 0-3 at the default of WB, WT, UC-, and UC.
1257 * Program 5 and 6 as WP and WC.
1258 * Leave 4 and 7 as WB and UC.
1260 pat_msr &= ~(PAT_MASK(5) | PAT_MASK(6));
1261 pat_msr |= PAT_VALUE(5, PAT_WRITE_PROTECTED) |
1262 PAT_VALUE(6, PAT_WRITE_COMBINING);
1263 pat_table[PAT_UNCACHED] = 2;
1264 pat_table[PAT_WRITE_PROTECTED] = 5;
1265 pat_table[PAT_WRITE_COMBINING] = 6;
1268 * Just replace PAT Index 2 with WC instead of UC-.
1270 pat_msr &= ~PAT_MASK(2);
1271 pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING);
1272 pat_table[PAT_WRITE_COMBINING] = 2;
1277 load_cr4(cr4 & ~CR4_PGE);
1279 /* Disable caches (CD = 1, NW = 0). */
1281 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1283 /* Flushes caches and TLBs. */
1287 /* Update PAT and index table. */
1288 wrmsr(MSR_PAT, pat_msr);
1289 for (i = 0; i < PAT_INDEX_SIZE; i++)
1290 pat_index[i] = pat_table[i];
1292 /* Flush caches and TLBs again. */
1296 /* Restore caches and PGE. */
1302 * Initialize a vm_page's machine-dependent fields.
1305 pmap_page_init(vm_page_t m)
1308 TAILQ_INIT(&m->md.pv_list);
1309 m->md.pat_mode = PAT_WRITE_BACK;
1313 * Initialize the pmap module.
1314 * Called by vm_init, to initialize any structures that the pmap
1315 * system needs to map virtual memory.
1320 struct pmap_preinit_mapping *ppim;
1323 int error, i, pv_npg, ret, skz63;
1325 /* L1TF, reserve page @0 unconditionally */
1326 vm_page_blacklist_add(0, bootverbose);
1328 /* Detect bare-metal Skylake Server and Skylake-X. */
1329 if (vm_guest == VM_GUEST_NO && cpu_vendor_id == CPU_VENDOR_INTEL &&
1330 CPUID_TO_FAMILY(cpu_id) == 0x6 && CPUID_TO_MODEL(cpu_id) == 0x55) {
1332 * Skylake-X errata SKZ63. Processor May Hang When
1333 * Executing Code In an HLE Transaction Region between
1334 * 40000000H and 403FFFFFH.
1336 * Mark the pages in the range as preallocated. It
1337 * seems to be impossible to distinguish between
1338 * Skylake Server and Skylake X.
1341 TUNABLE_INT_FETCH("hw.skz63_enable", &skz63);
1344 printf("SKZ63: skipping 4M RAM starting "
1345 "at physical 1G\n");
1346 for (i = 0; i < atop(0x400000); i++) {
1347 ret = vm_page_blacklist_add(0x40000000 +
1349 if (!ret && bootverbose)
1350 printf("page at %#lx already used\n",
1351 0x40000000 + ptoa(i));
1357 * Initialize the vm page array entries for the kernel pmap's
1360 PMAP_LOCK(kernel_pmap);
1361 for (i = 0; i < nkpt; i++) {
1362 mpte = PHYS_TO_VM_PAGE(KPTphys + (i << PAGE_SHIFT));
1363 KASSERT(mpte >= vm_page_array &&
1364 mpte < &vm_page_array[vm_page_array_size],
1365 ("pmap_init: page table page is out of range"));
1366 mpte->pindex = pmap_pde_pindex(KERNBASE) + i;
1367 mpte->phys_addr = KPTphys + (i << PAGE_SHIFT);
1368 mpte->wire_count = 1;
1369 if (i << PDRSHIFT < KERNend &&
1370 pmap_insert_pt_page(kernel_pmap, mpte))
1371 panic("pmap_init: pmap_insert_pt_page failed");
1373 PMAP_UNLOCK(kernel_pmap);
1377 * If the kernel is running on a virtual machine, then it must assume
1378 * that MCA is enabled by the hypervisor. Moreover, the kernel must
1379 * be prepared for the hypervisor changing the vendor and family that
1380 * are reported by CPUID. Consequently, the workaround for AMD Family
1381 * 10h Erratum 383 is enabled if the processor's feature set does not
1382 * include at least one feature that is only supported by older Intel
1383 * or newer AMD processors.
1385 if (vm_guest != VM_GUEST_NO && (cpu_feature & CPUID_SS) == 0 &&
1386 (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI |
1387 CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP |
1389 workaround_erratum383 = 1;
1392 * Are large page mappings enabled?
1394 TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled);
1395 if (pg_ps_enabled) {
1396 KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
1397 ("pmap_init: can't assign to pagesizes[1]"));
1398 pagesizes[1] = NBPDR;
1402 * Initialize the pv chunk list mutex.
1404 mtx_init(&pv_chunks_mutex, "pmap pv chunk list", NULL, MTX_DEF);
1407 * Initialize the pool of pv list locks.
1409 for (i = 0; i < NPV_LIST_LOCKS; i++)
1410 rw_init(&pv_list_locks[i], "pmap pv list");
1413 * Calculate the size of the pv head table for superpages.
1415 pv_npg = howmany(vm_phys_segs[vm_phys_nsegs - 1].end, NBPDR);
1418 * Allocate memory for the pv head table for superpages.
1420 s = (vm_size_t)(pv_npg * sizeof(struct md_page));
1422 pv_table = (struct md_page *)kmem_malloc(s, M_WAITOK | M_ZERO);
1423 for (i = 0; i < pv_npg; i++)
1424 TAILQ_INIT(&pv_table[i].pv_list);
1425 TAILQ_INIT(&pv_dummy.pv_list);
1427 pmap_initialized = 1;
1428 for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
1429 ppim = pmap_preinit_mapping + i;
1432 /* Make the direct map consistent */
1433 if (ppim->pa < dmaplimit && ppim->pa + ppim->sz <= dmaplimit) {
1434 (void)pmap_change_attr(PHYS_TO_DMAP(ppim->pa),
1435 ppim->sz, ppim->mode);
1439 printf("PPIM %u: PA=%#lx, VA=%#lx, size=%#lx, mode=%#x\n", i,
1440 ppim->pa, ppim->va, ppim->sz, ppim->mode);
1443 mtx_init(&qframe_mtx, "qfrmlk", NULL, MTX_SPIN);
1444 error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK,
1445 (vmem_addr_t *)&qframe);
1447 panic("qframe allocation failed");
1450 TUNABLE_INT_FETCH("vm.pmap.large_map_pml4_entries", &lm_ents);
1451 if (lm_ents > LMEPML4I - LMSPML4I + 1)
1452 lm_ents = LMEPML4I - LMSPML4I + 1;
1454 printf("pmap: large map %u PML4 slots (%lu Gb)\n",
1455 lm_ents, (u_long)lm_ents * (NBPML4 / 1024 / 1024 / 1024));
1457 large_vmem = vmem_create("large", LARGEMAP_MIN_ADDRESS,
1458 (vmem_size_t)lm_ents * NBPML4, PAGE_SIZE, 0, M_WAITOK);
1459 if (large_vmem == NULL) {
1460 printf("pmap: cannot create large map\n");
1463 for (i = 0; i < lm_ents; i++) {
1464 m = pmap_large_map_getptp_unlocked();
1465 kernel_pmap->pm_pml4[LMSPML4I + i] = X86_PG_V |
1466 X86_PG_RW | X86_PG_A | X86_PG_M | pg_nx |
1472 static SYSCTL_NODE(_vm_pmap, OID_AUTO, pde, CTLFLAG_RD, 0,
1473 "2MB page mapping counters");
1475 static u_long pmap_pde_demotions;
1476 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, demotions, CTLFLAG_RD,
1477 &pmap_pde_demotions, 0, "2MB page demotions");
1479 static u_long pmap_pde_mappings;
1480 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, mappings, CTLFLAG_RD,
1481 &pmap_pde_mappings, 0, "2MB page mappings");
1483 static u_long pmap_pde_p_failures;
1484 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, p_failures, CTLFLAG_RD,
1485 &pmap_pde_p_failures, 0, "2MB page promotion failures");
1487 static u_long pmap_pde_promotions;
1488 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, promotions, CTLFLAG_RD,
1489 &pmap_pde_promotions, 0, "2MB page promotions");
1491 static SYSCTL_NODE(_vm_pmap, OID_AUTO, pdpe, CTLFLAG_RD, 0,
1492 "1GB page mapping counters");
1494 static u_long pmap_pdpe_demotions;
1495 SYSCTL_ULONG(_vm_pmap_pdpe, OID_AUTO, demotions, CTLFLAG_RD,
1496 &pmap_pdpe_demotions, 0, "1GB page demotions");
1498 /***************************************************
1499 * Low level helper routines.....
1500 ***************************************************/
1503 pmap_swap_pat(pmap_t pmap, pt_entry_t entry)
1505 int x86_pat_bits = X86_PG_PTE_PAT | X86_PG_PDE_PAT;
1507 switch (pmap->pm_type) {
1510 /* Verify that both PAT bits are not set at the same time */
1511 KASSERT((entry & x86_pat_bits) != x86_pat_bits,
1512 ("Invalid PAT bits in entry %#lx", entry));
1514 /* Swap the PAT bits if one of them is set */
1515 if ((entry & x86_pat_bits) != 0)
1516 entry ^= x86_pat_bits;
1520 * Nothing to do - the memory attributes are represented
1521 * the same way for regular pages and superpages.
1525 panic("pmap_switch_pat_bits: bad pm_type %d", pmap->pm_type);
1532 pmap_is_valid_memattr(pmap_t pmap __unused, vm_memattr_t mode)
1535 return (mode >= 0 && mode < PAT_INDEX_SIZE &&
1536 pat_index[(int)mode] >= 0);
1540 * Determine the appropriate bits to set in a PTE or PDE for a specified
1544 pmap_cache_bits(pmap_t pmap, int mode, boolean_t is_pde)
1546 int cache_bits, pat_flag, pat_idx;
1548 if (!pmap_is_valid_memattr(pmap, mode))
1549 panic("Unknown caching mode %d\n", mode);
1551 switch (pmap->pm_type) {
1554 /* The PAT bit is different for PTE's and PDE's. */
1555 pat_flag = is_pde ? X86_PG_PDE_PAT : X86_PG_PTE_PAT;
1557 /* Map the caching mode to a PAT index. */
1558 pat_idx = pat_index[mode];
1560 /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */
1563 cache_bits |= pat_flag;
1565 cache_bits |= PG_NC_PCD;
1567 cache_bits |= PG_NC_PWT;
1571 cache_bits = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(mode);
1575 panic("unsupported pmap type %d", pmap->pm_type);
1578 return (cache_bits);
1582 pmap_cache_mask(pmap_t pmap, boolean_t is_pde)
1586 switch (pmap->pm_type) {
1589 mask = is_pde ? X86_PG_PDE_CACHE : X86_PG_PTE_CACHE;
1592 mask = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(0x7);
1595 panic("pmap_cache_mask: invalid pm_type %d", pmap->pm_type);
1602 pmap_ps_enabled(pmap_t pmap)
1605 return (pg_ps_enabled && (pmap->pm_flags & PMAP_PDE_SUPERPAGE) != 0);
1609 pmap_update_pde_store(pmap_t pmap, pd_entry_t *pde, pd_entry_t newpde)
1612 switch (pmap->pm_type) {
1619 * This is a little bogus since the generation number is
1620 * supposed to be bumped up when a region of the address
1621 * space is invalidated in the page tables.
1623 * In this case the old PDE entry is valid but yet we want
1624 * to make sure that any mappings using the old entry are
1625 * invalidated in the TLB.
1627 * The reason this works as expected is because we rendezvous
1628 * "all" host cpus and force any vcpu context to exit as a
1631 atomic_add_acq_long(&pmap->pm_eptgen, 1);
1634 panic("pmap_update_pde_store: bad pm_type %d", pmap->pm_type);
1636 pde_store(pde, newpde);
1640 * After changing the page size for the specified virtual address in the page
1641 * table, flush the corresponding entries from the processor's TLB. Only the
1642 * calling processor's TLB is affected.
1644 * The calling thread must be pinned to a processor.
1647 pmap_update_pde_invalidate(pmap_t pmap, vm_offset_t va, pd_entry_t newpde)
1651 if (pmap_type_guest(pmap))
1654 KASSERT(pmap->pm_type == PT_X86,
1655 ("pmap_update_pde_invalidate: invalid type %d", pmap->pm_type));
1657 PG_G = pmap_global_bit(pmap);
1659 if ((newpde & PG_PS) == 0)
1660 /* Demotion: flush a specific 2MB page mapping. */
1662 else if ((newpde & PG_G) == 0)
1664 * Promotion: flush every 4KB page mapping from the TLB
1665 * because there are too many to flush individually.
1670 * Promotion: flush every 4KB page mapping from the TLB,
1671 * including any global (PG_G) mappings.
1679 * For SMP, these functions have to use the IPI mechanism for coherence.
1681 * N.B.: Before calling any of the following TLB invalidation functions,
1682 * the calling processor must ensure that all stores updating a non-
1683 * kernel page table are globally performed. Otherwise, another
1684 * processor could cache an old, pre-update entry without being
1685 * invalidated. This can happen one of two ways: (1) The pmap becomes
1686 * active on another processor after its pm_active field is checked by
1687 * one of the following functions but before a store updating the page
1688 * table is globally performed. (2) The pmap becomes active on another
1689 * processor before its pm_active field is checked but due to
1690 * speculative loads one of the following functions stills reads the
1691 * pmap as inactive on the other processor.
1693 * The kernel page table is exempt because its pm_active field is
1694 * immutable. The kernel page table is always active on every
1699 * Interrupt the cpus that are executing in the guest context.
1700 * This will force the vcpu to exit and the cached EPT mappings
1701 * will be invalidated by the host before the next vmresume.
1703 static __inline void
1704 pmap_invalidate_ept(pmap_t pmap)
1709 KASSERT(!CPU_ISSET(curcpu, &pmap->pm_active),
1710 ("pmap_invalidate_ept: absurd pm_active"));
1713 * The TLB mappings associated with a vcpu context are not
1714 * flushed each time a different vcpu is chosen to execute.
1716 * This is in contrast with a process's vtop mappings that
1717 * are flushed from the TLB on each context switch.
1719 * Therefore we need to do more than just a TLB shootdown on
1720 * the active cpus in 'pmap->pm_active'. To do this we keep
1721 * track of the number of invalidations performed on this pmap.
1723 * Each vcpu keeps a cache of this counter and compares it
1724 * just before a vmresume. If the counter is out-of-date an
1725 * invept will be done to flush stale mappings from the TLB.
1727 atomic_add_acq_long(&pmap->pm_eptgen, 1);
1730 * Force the vcpu to exit and trap back into the hypervisor.
1732 ipinum = pmap->pm_flags & PMAP_NESTED_IPIMASK;
1733 ipi_selected(pmap->pm_active, ipinum);
1738 pmap_invalidate_cpu_mask(pmap_t pmap)
1741 return (pmap == kernel_pmap ? all_cpus : pmap->pm_active);
1745 pmap_invalidate_page_pcid(pmap_t pmap, vm_offset_t va,
1746 const bool invpcid_works1)
1748 struct invpcid_descr d;
1749 uint64_t kcr3, ucr3;
1753 cpuid = PCPU_GET(cpuid);
1754 if (pmap == PCPU_GET(curpmap)) {
1755 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
1757 * Because pm_pcid is recalculated on a
1758 * context switch, we must disable switching.
1759 * Otherwise, we might use a stale value
1763 pcid = pmap->pm_pcids[cpuid].pm_pcid;
1764 if (invpcid_works1) {
1765 d.pcid = pcid | PMAP_PCID_USER_PT;
1768 invpcid(&d, INVPCID_ADDR);
1770 kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
1771 ucr3 = pmap->pm_ucr3 | pcid |
1772 PMAP_PCID_USER_PT | CR3_PCID_SAVE;
1773 pmap_pti_pcid_invlpg(ucr3, kcr3, va);
1778 pmap->pm_pcids[cpuid].pm_gen = 0;
1782 pmap->pm_pcids[i].pm_gen = 0;
1786 * The fence is between stores to pm_gen and the read of the
1787 * pm_active mask. We need to ensure that it is impossible
1788 * for us to miss the bit update in pm_active and
1789 * simultaneously observe a non-zero pm_gen in
1790 * pmap_activate_sw(), otherwise TLB update is missed.
1791 * Without the fence, IA32 allows such an outcome. Note that
1792 * pm_active is updated by a locked operation, which provides
1793 * the reciprocal fence.
1795 atomic_thread_fence_seq_cst();
1799 pmap_invalidate_page_pcid_invpcid(pmap_t pmap, vm_offset_t va)
1802 pmap_invalidate_page_pcid(pmap, va, true);
1806 pmap_invalidate_page_pcid_noinvpcid(pmap_t pmap, vm_offset_t va)
1809 pmap_invalidate_page_pcid(pmap, va, false);
1813 pmap_invalidate_page_nopcid(pmap_t pmap, vm_offset_t va)
1817 DEFINE_IFUNC(static, void, pmap_invalidate_page_mode, (pmap_t, vm_offset_t),
1821 if (pmap_pcid_enabled)
1822 return (invpcid_works ? pmap_invalidate_page_pcid_invpcid :
1823 pmap_invalidate_page_pcid_noinvpcid);
1824 return (pmap_invalidate_page_nopcid);
1828 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
1831 if (pmap_type_guest(pmap)) {
1832 pmap_invalidate_ept(pmap);
1836 KASSERT(pmap->pm_type == PT_X86,
1837 ("pmap_invalidate_page: invalid type %d", pmap->pm_type));
1840 if (pmap == kernel_pmap) {
1843 if (pmap == PCPU_GET(curpmap))
1845 pmap_invalidate_page_mode(pmap, va);
1847 smp_masked_invlpg(pmap_invalidate_cpu_mask(pmap), va, pmap);
1851 /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */
1852 #define PMAP_INVLPG_THRESHOLD (4 * 1024 * PAGE_SIZE)
1855 pmap_invalidate_range_pcid(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
1856 const bool invpcid_works1)
1858 struct invpcid_descr d;
1859 uint64_t kcr3, ucr3;
1863 cpuid = PCPU_GET(cpuid);
1864 if (pmap == PCPU_GET(curpmap)) {
1865 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
1867 pcid = pmap->pm_pcids[cpuid].pm_pcid;
1868 if (invpcid_works1) {
1869 d.pcid = pcid | PMAP_PCID_USER_PT;
1872 for (; d.addr < eva; d.addr += PAGE_SIZE)
1873 invpcid(&d, INVPCID_ADDR);
1875 kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
1876 ucr3 = pmap->pm_ucr3 | pcid |
1877 PMAP_PCID_USER_PT | CR3_PCID_SAVE;
1878 pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva);
1883 pmap->pm_pcids[cpuid].pm_gen = 0;
1887 pmap->pm_pcids[i].pm_gen = 0;
1889 /* See the comment in pmap_invalidate_page_pcid(). */
1890 atomic_thread_fence_seq_cst();
1894 pmap_invalidate_range_pcid_invpcid(pmap_t pmap, vm_offset_t sva,
1898 pmap_invalidate_range_pcid(pmap, sva, eva, true);
1902 pmap_invalidate_range_pcid_noinvpcid(pmap_t pmap, vm_offset_t sva,
1906 pmap_invalidate_range_pcid(pmap, sva, eva, false);
1910 pmap_invalidate_range_nopcid(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1914 DEFINE_IFUNC(static, void, pmap_invalidate_range_mode, (pmap_t, vm_offset_t,
1915 vm_offset_t), static)
1918 if (pmap_pcid_enabled)
1919 return (invpcid_works ? pmap_invalidate_range_pcid_invpcid :
1920 pmap_invalidate_range_pcid_noinvpcid);
1921 return (pmap_invalidate_range_nopcid);
1925 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1929 if (eva - sva >= PMAP_INVLPG_THRESHOLD) {
1930 pmap_invalidate_all(pmap);
1934 if (pmap_type_guest(pmap)) {
1935 pmap_invalidate_ept(pmap);
1939 KASSERT(pmap->pm_type == PT_X86,
1940 ("pmap_invalidate_range: invalid type %d", pmap->pm_type));
1943 if (pmap == kernel_pmap) {
1944 for (addr = sva; addr < eva; addr += PAGE_SIZE)
1947 if (pmap == PCPU_GET(curpmap)) {
1948 for (addr = sva; addr < eva; addr += PAGE_SIZE)
1951 pmap_invalidate_range_mode(pmap, sva, eva);
1953 smp_masked_invlpg_range(pmap_invalidate_cpu_mask(pmap), sva, eva, pmap);
1958 pmap_invalidate_all_pcid(pmap_t pmap, bool invpcid_works1)
1960 struct invpcid_descr d;
1961 uint64_t kcr3, ucr3;
1965 if (pmap == kernel_pmap) {
1966 if (invpcid_works1) {
1967 bzero(&d, sizeof(d));
1968 invpcid(&d, INVPCID_CTXGLOB);
1973 cpuid = PCPU_GET(cpuid);
1974 if (pmap == PCPU_GET(curpmap)) {
1976 pcid = pmap->pm_pcids[cpuid].pm_pcid;
1977 if (invpcid_works1) {
1981 invpcid(&d, INVPCID_CTX);
1982 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
1983 d.pcid |= PMAP_PCID_USER_PT;
1984 invpcid(&d, INVPCID_CTX);
1987 kcr3 = pmap->pm_cr3 | pcid;
1988 ucr3 = pmap->pm_ucr3;
1989 if (ucr3 != PMAP_NO_CR3) {
1990 ucr3 |= pcid | PMAP_PCID_USER_PT;
1991 pmap_pti_pcid_invalidate(ucr3, kcr3);
1998 pmap->pm_pcids[cpuid].pm_gen = 0;
2001 pmap->pm_pcids[i].pm_gen = 0;
2004 /* See the comment in pmap_invalidate_page_pcid(). */
2005 atomic_thread_fence_seq_cst();
2009 pmap_invalidate_all_pcid_invpcid(pmap_t pmap)
2012 pmap_invalidate_all_pcid(pmap, true);
2016 pmap_invalidate_all_pcid_noinvpcid(pmap_t pmap)
2019 pmap_invalidate_all_pcid(pmap, false);
2023 pmap_invalidate_all_nopcid(pmap_t pmap)
2026 if (pmap == kernel_pmap)
2028 else if (pmap == PCPU_GET(curpmap))
2032 DEFINE_IFUNC(static, void, pmap_invalidate_all_mode, (pmap_t), static)
2035 if (pmap_pcid_enabled)
2036 return (invpcid_works ? pmap_invalidate_all_pcid_invpcid :
2037 pmap_invalidate_all_pcid_noinvpcid);
2038 return (pmap_invalidate_all_nopcid);
2042 pmap_invalidate_all(pmap_t pmap)
2045 if (pmap_type_guest(pmap)) {
2046 pmap_invalidate_ept(pmap);
2050 KASSERT(pmap->pm_type == PT_X86,
2051 ("pmap_invalidate_all: invalid type %d", pmap->pm_type));
2054 pmap_invalidate_all_mode(pmap);
2055 smp_masked_invltlb(pmap_invalidate_cpu_mask(pmap), pmap);
2060 pmap_invalidate_cache(void)
2070 cpuset_t invalidate; /* processors that invalidate their TLB */
2075 u_int store; /* processor that updates the PDE */
2079 pmap_update_pde_action(void *arg)
2081 struct pde_action *act = arg;
2083 if (act->store == PCPU_GET(cpuid))
2084 pmap_update_pde_store(act->pmap, act->pde, act->newpde);
2088 pmap_update_pde_teardown(void *arg)
2090 struct pde_action *act = arg;
2092 if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate))
2093 pmap_update_pde_invalidate(act->pmap, act->va, act->newpde);
2097 * Change the page size for the specified virtual address in a way that
2098 * prevents any possibility of the TLB ever having two entries that map the
2099 * same virtual address using different page sizes. This is the recommended
2100 * workaround for Erratum 383 on AMD Family 10h processors. It prevents a
2101 * machine check exception for a TLB state that is improperly diagnosed as a
2105 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
2107 struct pde_action act;
2108 cpuset_t active, other_cpus;
2112 cpuid = PCPU_GET(cpuid);
2113 other_cpus = all_cpus;
2114 CPU_CLR(cpuid, &other_cpus);
2115 if (pmap == kernel_pmap || pmap_type_guest(pmap))
2118 active = pmap->pm_active;
2120 if (CPU_OVERLAP(&active, &other_cpus)) {
2122 act.invalidate = active;
2126 act.newpde = newpde;
2127 CPU_SET(cpuid, &active);
2128 smp_rendezvous_cpus(active,
2129 smp_no_rendezvous_barrier, pmap_update_pde_action,
2130 pmap_update_pde_teardown, &act);
2132 pmap_update_pde_store(pmap, pde, newpde);
2133 if (CPU_ISSET(cpuid, &active))
2134 pmap_update_pde_invalidate(pmap, va, newpde);
2140 * Normal, non-SMP, invalidation functions.
2143 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
2145 struct invpcid_descr d;
2146 uint64_t kcr3, ucr3;
2149 if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) {
2153 KASSERT(pmap->pm_type == PT_X86,
2154 ("pmap_invalidate_range: unknown type %d", pmap->pm_type));
2156 if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) {
2158 if (pmap == PCPU_GET(curpmap) && pmap_pcid_enabled &&
2159 pmap->pm_ucr3 != PMAP_NO_CR3) {
2161 pcid = pmap->pm_pcids[0].pm_pcid;
2162 if (invpcid_works) {
2163 d.pcid = pcid | PMAP_PCID_USER_PT;
2166 invpcid(&d, INVPCID_ADDR);
2168 kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
2169 ucr3 = pmap->pm_ucr3 | pcid |
2170 PMAP_PCID_USER_PT | CR3_PCID_SAVE;
2171 pmap_pti_pcid_invlpg(ucr3, kcr3, va);
2175 } else if (pmap_pcid_enabled)
2176 pmap->pm_pcids[0].pm_gen = 0;
2180 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2182 struct invpcid_descr d;
2184 uint64_t kcr3, ucr3;
2186 if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) {
2190 KASSERT(pmap->pm_type == PT_X86,
2191 ("pmap_invalidate_range: unknown type %d", pmap->pm_type));
2193 if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) {
2194 for (addr = sva; addr < eva; addr += PAGE_SIZE)
2196 if (pmap == PCPU_GET(curpmap) && pmap_pcid_enabled &&
2197 pmap->pm_ucr3 != PMAP_NO_CR3) {
2199 if (invpcid_works) {
2200 d.pcid = pmap->pm_pcids[0].pm_pcid |
2204 for (; d.addr < eva; d.addr += PAGE_SIZE)
2205 invpcid(&d, INVPCID_ADDR);
2207 kcr3 = pmap->pm_cr3 | pmap->pm_pcids[0].
2208 pm_pcid | CR3_PCID_SAVE;
2209 ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[0].
2210 pm_pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE;
2211 pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva);
2215 } else if (pmap_pcid_enabled) {
2216 pmap->pm_pcids[0].pm_gen = 0;
2221 pmap_invalidate_all(pmap_t pmap)
2223 struct invpcid_descr d;
2224 uint64_t kcr3, ucr3;
2226 if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) {
2230 KASSERT(pmap->pm_type == PT_X86,
2231 ("pmap_invalidate_all: unknown type %d", pmap->pm_type));
2233 if (pmap == kernel_pmap) {
2234 if (pmap_pcid_enabled && invpcid_works) {
2235 bzero(&d, sizeof(d));
2236 invpcid(&d, INVPCID_CTXGLOB);
2240 } else if (pmap == PCPU_GET(curpmap)) {
2241 if (pmap_pcid_enabled) {
2243 if (invpcid_works) {
2244 d.pcid = pmap->pm_pcids[0].pm_pcid;
2247 invpcid(&d, INVPCID_CTX);
2248 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
2249 d.pcid |= PMAP_PCID_USER_PT;
2250 invpcid(&d, INVPCID_CTX);
2253 kcr3 = pmap->pm_cr3 | pmap->pm_pcids[0].pm_pcid;
2254 if (pmap->pm_ucr3 != PMAP_NO_CR3) {
2255 ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[
2256 0].pm_pcid | PMAP_PCID_USER_PT;
2257 pmap_pti_pcid_invalidate(ucr3, kcr3);
2265 } else if (pmap_pcid_enabled) {
2266 pmap->pm_pcids[0].pm_gen = 0;
2271 pmap_invalidate_cache(void)
2278 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
2281 pmap_update_pde_store(pmap, pde, newpde);
2282 if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap))
2283 pmap_update_pde_invalidate(pmap, va, newpde);
2285 pmap->pm_pcids[0].pm_gen = 0;
2290 pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde)
2294 * When the PDE has PG_PROMOTED set, the 2MB page mapping was created
2295 * by a promotion that did not invalidate the 512 4KB page mappings
2296 * that might exist in the TLB. Consequently, at this point, the TLB
2297 * may hold both 4KB and 2MB page mappings for the address range [va,
2298 * va + NBPDR). Therefore, the entire range must be invalidated here.
2299 * In contrast, when PG_PROMOTED is clear, the TLB will not hold any
2300 * 4KB page mappings for the address range [va, va + NBPDR), and so a
2301 * single INVLPG suffices to invalidate the 2MB page mapping from the
2304 if ((pde & PG_PROMOTED) != 0)
2305 pmap_invalidate_range(pmap, va, va + NBPDR - 1);
2307 pmap_invalidate_page(pmap, va);
2310 DEFINE_IFUNC(, void, pmap_invalidate_cache_range,
2311 (vm_offset_t sva, vm_offset_t eva), static)
2314 if ((cpu_feature & CPUID_SS) != 0)
2315 return (pmap_invalidate_cache_range_selfsnoop);
2316 if ((cpu_feature & CPUID_CLFSH) != 0)
2317 return (pmap_force_invalidate_cache_range);
2318 return (pmap_invalidate_cache_range_all);
2321 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
2324 pmap_invalidate_cache_range_check_align(vm_offset_t sva, vm_offset_t eva)
2327 KASSERT((sva & PAGE_MASK) == 0,
2328 ("pmap_invalidate_cache_range: sva not page-aligned"));
2329 KASSERT((eva & PAGE_MASK) == 0,
2330 ("pmap_invalidate_cache_range: eva not page-aligned"));
2334 pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva, vm_offset_t eva)
2337 pmap_invalidate_cache_range_check_align(sva, eva);
2341 pmap_force_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
2344 sva &= ~(vm_offset_t)(cpu_clflush_line_size - 1);
2347 * XXX: Some CPUs fault, hang, or trash the local APIC
2348 * registers if we use CLFLUSH on the local APIC range. The
2349 * local APIC is always uncached, so we don't need to flush
2350 * for that range anyway.
2352 if (pmap_kextract(sva) == lapic_paddr)
2355 if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0) {
2357 * Do per-cache line flush. Use the sfence
2358 * instruction to insure that previous stores are
2359 * included in the write-back. The processor
2360 * propagates flush to other processors in the cache
2364 for (; sva < eva; sva += cpu_clflush_line_size)
2369 * Writes are ordered by CLFLUSH on Intel CPUs.
2371 if (cpu_vendor_id != CPU_VENDOR_INTEL)
2373 for (; sva < eva; sva += cpu_clflush_line_size)
2375 if (cpu_vendor_id != CPU_VENDOR_INTEL)
2381 pmap_invalidate_cache_range_all(vm_offset_t sva, vm_offset_t eva)
2384 pmap_invalidate_cache_range_check_align(sva, eva);
2385 pmap_invalidate_cache();
2389 * Remove the specified set of pages from the data and instruction caches.
2391 * In contrast to pmap_invalidate_cache_range(), this function does not
2392 * rely on the CPU's self-snoop feature, because it is intended for use
2393 * when moving pages into a different cache domain.
2396 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
2398 vm_offset_t daddr, eva;
2402 useclflushopt = (cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0;
2403 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
2404 ((cpu_feature & CPUID_CLFSH) == 0 && !useclflushopt))
2405 pmap_invalidate_cache();
2409 else if (cpu_vendor_id != CPU_VENDOR_INTEL)
2411 for (i = 0; i < count; i++) {
2412 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
2413 eva = daddr + PAGE_SIZE;
2414 for (; daddr < eva; daddr += cpu_clflush_line_size) {
2423 else if (cpu_vendor_id != CPU_VENDOR_INTEL)
2429 pmap_flush_cache_range(vm_offset_t sva, vm_offset_t eva)
2432 pmap_invalidate_cache_range_check_align(sva, eva);
2434 if ((cpu_stdext_feature & CPUID_STDEXT_CLWB) == 0) {
2435 pmap_force_invalidate_cache_range(sva, eva);
2439 /* See comment in pmap_force_invalidate_cache_range(). */
2440 if (pmap_kextract(sva) == lapic_paddr)
2444 for (; sva < eva; sva += cpu_clflush_line_size)
2450 pmap_flush_cache_phys_range(vm_paddr_t spa, vm_paddr_t epa, vm_memattr_t mattr)
2454 int error, pte_bits;
2456 KASSERT((spa & PAGE_MASK) == 0,
2457 ("pmap_flush_cache_phys_range: spa not page-aligned"));
2458 KASSERT((epa & PAGE_MASK) == 0,
2459 ("pmap_flush_cache_phys_range: epa not page-aligned"));
2461 if (spa < dmaplimit) {
2462 pmap_flush_cache_range(PHYS_TO_DMAP(spa), PHYS_TO_DMAP(MIN(
2464 if (dmaplimit >= epa)
2469 pte_bits = pmap_cache_bits(kernel_pmap, mattr, 0) | X86_PG_RW |
2471 error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK,
2473 KASSERT(error == 0, ("vmem_alloc failed: %d", error));
2474 pte = vtopte(vaddr);
2475 for (; spa < epa; spa += PAGE_SIZE) {
2477 pte_store(pte, spa | pte_bits);
2479 /* XXXKIB sfences inside flush_cache_range are excessive */
2480 pmap_flush_cache_range(vaddr, vaddr + PAGE_SIZE);
2483 vmem_free(kernel_arena, vaddr, PAGE_SIZE);
2487 * Routine: pmap_extract
2489 * Extract the physical page address associated
2490 * with the given map/virtual_address pair.
2493 pmap_extract(pmap_t pmap, vm_offset_t va)
2497 pt_entry_t *pte, PG_V;
2501 PG_V = pmap_valid_bit(pmap);
2503 pdpe = pmap_pdpe(pmap, va);
2504 if (pdpe != NULL && (*pdpe & PG_V) != 0) {
2505 if ((*pdpe & PG_PS) != 0)
2506 pa = (*pdpe & PG_PS_FRAME) | (va & PDPMASK);
2508 pde = pmap_pdpe_to_pde(pdpe, va);
2509 if ((*pde & PG_V) != 0) {
2510 if ((*pde & PG_PS) != 0) {
2511 pa = (*pde & PG_PS_FRAME) |
2514 pte = pmap_pde_to_pte(pde, va);
2515 pa = (*pte & PG_FRAME) |
2526 * Routine: pmap_extract_and_hold
2528 * Atomically extract and hold the physical page
2529 * with the given pmap and virtual address pair
2530 * if that mapping permits the given protection.
2533 pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
2535 pd_entry_t pde, *pdep;
2536 pt_entry_t pte, PG_RW, PG_V;
2542 PG_RW = pmap_rw_bit(pmap);
2543 PG_V = pmap_valid_bit(pmap);
2546 pdep = pmap_pde(pmap, va);
2547 if (pdep != NULL && (pde = *pdep)) {
2549 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) {
2550 if (vm_page_pa_tryrelock(pmap, (pde &
2551 PG_PS_FRAME) | (va & PDRMASK), &pa))
2553 m = PHYS_TO_VM_PAGE(pa);
2556 pte = *pmap_pde_to_pte(pdep, va);
2558 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) {
2559 if (vm_page_pa_tryrelock(pmap, pte & PG_FRAME,
2562 m = PHYS_TO_VM_PAGE(pa);
2574 pmap_kextract(vm_offset_t va)
2579 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
2580 pa = DMAP_TO_PHYS(va);
2584 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
2587 * Beware of a concurrent promotion that changes the
2588 * PDE at this point! For example, vtopte() must not
2589 * be used to access the PTE because it would use the
2590 * new PDE. It is, however, safe to use the old PDE
2591 * because the page table page is preserved by the
2594 pa = *pmap_pde_to_pte(&pde, va);
2595 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
2601 /***************************************************
2602 * Low level mapping routines.....
2603 ***************************************************/
2606 * Add a wired page to the kva.
2607 * Note: not SMP coherent.
2610 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
2615 pte_store(pte, pa | X86_PG_RW | X86_PG_V | pg_g);
2618 static __inline void
2619 pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode)
2625 cache_bits = pmap_cache_bits(kernel_pmap, mode, 0);
2626 pte_store(pte, pa | X86_PG_RW | X86_PG_V | pg_g | cache_bits);
2630 * Remove a page from the kernel pagetables.
2631 * Note: not SMP coherent.
2634 pmap_kremove(vm_offset_t va)
2643 * Used to map a range of physical addresses into kernel
2644 * virtual address space.
2646 * The value passed in '*virt' is a suggested virtual address for
2647 * the mapping. Architectures which can support a direct-mapped
2648 * physical to virtual region can return the appropriate address
2649 * within that region, leaving '*virt' unchanged. Other
2650 * architectures should map the pages starting at '*virt' and
2651 * update '*virt' with the first usable address after the mapped
2655 pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot)
2657 return PHYS_TO_DMAP(start);
2662 * Add a list of wired pages to the kva
2663 * this routine is only used for temporary
2664 * kernel mappings that do not need to have
2665 * page modification or references recorded.
2666 * Note that old mappings are simply written
2667 * over. The page *must* be wired.
2668 * Note: SMP coherent. Uses a ranged shootdown IPI.
2671 pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count)
2673 pt_entry_t *endpte, oldpte, pa, *pte;
2679 endpte = pte + count;
2680 while (pte < endpte) {
2682 cache_bits = pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0);
2683 pa = VM_PAGE_TO_PHYS(m) | cache_bits;
2684 if ((*pte & (PG_FRAME | X86_PG_PTE_CACHE)) != pa) {
2686 pte_store(pte, pa | pg_g | pg_nx | X86_PG_RW | X86_PG_V);
2690 if (__predict_false((oldpte & X86_PG_V) != 0))
2691 pmap_invalidate_range(kernel_pmap, sva, sva + count *
2696 * This routine tears out page mappings from the
2697 * kernel -- it is meant only for temporary mappings.
2698 * Note: SMP coherent. Uses a ranged shootdown IPI.
2701 pmap_qremove(vm_offset_t sva, int count)
2706 while (count-- > 0) {
2707 KASSERT(va >= VM_MIN_KERNEL_ADDRESS, ("usermode va %lx", va));
2711 pmap_invalidate_range(kernel_pmap, sva, va);
2714 /***************************************************
2715 * Page table page management routines.....
2716 ***************************************************/
2718 * Schedule the specified unused page table page to be freed. Specifically,
2719 * add the page to the specified list of pages that will be released to the
2720 * physical memory manager after the TLB has been updated.
2722 static __inline void
2723 pmap_add_delayed_free_list(vm_page_t m, struct spglist *free,
2724 boolean_t set_PG_ZERO)
2728 m->flags |= PG_ZERO;
2730 m->flags &= ~PG_ZERO;
2731 SLIST_INSERT_HEAD(free, m, plinks.s.ss);
2735 * Inserts the specified page table page into the specified pmap's collection
2736 * of idle page table pages. Each of a pmap's page table pages is responsible
2737 * for mapping a distinct range of virtual addresses. The pmap's collection is
2738 * ordered by this virtual address range.
2741 pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte)
2744 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2745 return (vm_radix_insert(&pmap->pm_root, mpte));
2749 * Removes the page table page mapping the specified virtual address from the
2750 * specified pmap's collection of idle page table pages, and returns it.
2751 * Otherwise, returns NULL if there is no page table page corresponding to the
2752 * specified virtual address.
2754 static __inline vm_page_t
2755 pmap_remove_pt_page(pmap_t pmap, vm_offset_t va)
2758 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2759 return (vm_radix_remove(&pmap->pm_root, pmap_pde_pindex(va)));
2763 * Decrements a page table page's wire count, which is used to record the
2764 * number of valid page table entries within the page. If the wire count
2765 * drops to zero, then the page table page is unmapped. Returns TRUE if the
2766 * page table page was unmapped and FALSE otherwise.
2768 static inline boolean_t
2769 pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
2773 if (m->wire_count == 0) {
2774 _pmap_unwire_ptp(pmap, va, m, free);
2781 _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
2784 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2786 * unmap the page table page
2788 if (m->pindex >= (NUPDE + NUPDPE)) {
2791 pml4 = pmap_pml4e(pmap, va);
2793 if (pmap->pm_pml4u != NULL && va <= VM_MAXUSER_ADDRESS) {
2794 pml4 = &pmap->pm_pml4u[pmap_pml4e_index(va)];
2797 } else if (m->pindex >= NUPDE) {
2800 pdp = pmap_pdpe(pmap, va);
2805 pd = pmap_pde(pmap, va);
2808 pmap_resident_count_dec(pmap, 1);
2809 if (m->pindex < NUPDE) {
2810 /* We just released a PT, unhold the matching PD */
2813 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
2814 pmap_unwire_ptp(pmap, va, pdpg, free);
2816 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
2817 /* We just released a PD, unhold the matching PDP */
2820 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
2821 pmap_unwire_ptp(pmap, va, pdppg, free);
2825 * Put page on a list so that it is released after
2826 * *ALL* TLB shootdown is done
2828 pmap_add_delayed_free_list(m, free, TRUE);
2832 * After removing a page table entry, this routine is used to
2833 * conditionally free the page, and manage the hold/wire counts.
2836 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, pd_entry_t ptepde,
2837 struct spglist *free)
2841 if (va >= VM_MAXUSER_ADDRESS)
2843 KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0"));
2844 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME);
2845 return (pmap_unwire_ptp(pmap, va, mpte, free));
2849 pmap_pinit0(pmap_t pmap)
2853 PMAP_LOCK_INIT(pmap);
2854 pmap->pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
2855 pmap->pm_pml4u = NULL;
2856 pmap->pm_cr3 = KPML4phys;
2857 /* hack to keep pmap_pti_pcid_invalidate() alive */
2858 pmap->pm_ucr3 = PMAP_NO_CR3;
2859 pmap->pm_root.rt_root = 0;
2860 CPU_ZERO(&pmap->pm_active);
2861 TAILQ_INIT(&pmap->pm_pvchunk);
2862 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
2863 pmap->pm_flags = pmap_flags;
2865 pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN + 1;
2866 pmap->pm_pcids[i].pm_gen = 1;
2868 pmap_activate_boot(pmap);
2872 pmap_pinit_pml4(vm_page_t pml4pg)
2874 pml4_entry_t *pm_pml4;
2877 pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4pg));
2879 /* Wire in kernel global address entries. */
2880 for (i = 0; i < NKPML4E; i++) {
2881 pm_pml4[KPML4BASE + i] = (KPDPphys + ptoa(i)) | X86_PG_RW |
2884 for (i = 0; i < ndmpdpphys; i++) {
2885 pm_pml4[DMPML4I + i] = (DMPDPphys + ptoa(i)) | X86_PG_RW |
2889 /* install self-referential address mapping entry(s) */
2890 pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pml4pg) | X86_PG_V | X86_PG_RW |
2891 X86_PG_A | X86_PG_M;
2893 /* install large map entries if configured */
2894 for (i = 0; i < lm_ents; i++)
2895 pm_pml4[LMSPML4I + i] = kernel_pmap->pm_pml4[LMSPML4I + i];
2899 pmap_pinit_pml4_pti(vm_page_t pml4pg)
2901 pml4_entry_t *pm_pml4;
2904 pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4pg));
2905 for (i = 0; i < NPML4EPG; i++)
2906 pm_pml4[i] = pti_pml4[i];
2910 * Initialize a preallocated and zeroed pmap structure,
2911 * such as one in a vmspace structure.
2914 pmap_pinit_type(pmap_t pmap, enum pmap_type pm_type, int flags)
2916 vm_page_t pml4pg, pml4pgu;
2917 vm_paddr_t pml4phys;
2921 * allocate the page directory page
2923 pml4pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
2924 VM_ALLOC_WIRED | VM_ALLOC_ZERO | VM_ALLOC_WAITOK);
2926 pml4phys = VM_PAGE_TO_PHYS(pml4pg);
2927 pmap->pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(pml4phys);
2929 pmap->pm_pcids[i].pm_pcid = PMAP_PCID_NONE;
2930 pmap->pm_pcids[i].pm_gen = 0;
2932 pmap->pm_cr3 = PMAP_NO_CR3; /* initialize to an invalid value */
2933 pmap->pm_ucr3 = PMAP_NO_CR3;
2934 pmap->pm_pml4u = NULL;
2936 pmap->pm_type = pm_type;
2937 if ((pml4pg->flags & PG_ZERO) == 0)
2938 pagezero(pmap->pm_pml4);
2941 * Do not install the host kernel mappings in the nested page
2942 * tables. These mappings are meaningless in the guest physical
2944 * Install minimal kernel mappings in PTI case.
2946 if (pm_type == PT_X86) {
2947 pmap->pm_cr3 = pml4phys;
2948 pmap_pinit_pml4(pml4pg);
2950 pml4pgu = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
2951 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
2952 pmap->pm_pml4u = (pml4_entry_t *)PHYS_TO_DMAP(
2953 VM_PAGE_TO_PHYS(pml4pgu));
2954 pmap_pinit_pml4_pti(pml4pgu);
2955 pmap->pm_ucr3 = VM_PAGE_TO_PHYS(pml4pgu);
2959 pmap->pm_root.rt_root = 0;
2960 CPU_ZERO(&pmap->pm_active);
2961 TAILQ_INIT(&pmap->pm_pvchunk);
2962 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
2963 pmap->pm_flags = flags;
2964 pmap->pm_eptgen = 0;
2970 pmap_pinit(pmap_t pmap)
2973 return (pmap_pinit_type(pmap, PT_X86, pmap_flags));
2977 * This routine is called if the desired page table page does not exist.
2979 * If page table page allocation fails, this routine may sleep before
2980 * returning NULL. It sleeps only if a lock pointer was given.
2982 * Note: If a page allocation fails at page table level two or three,
2983 * one or two pages may be held during the wait, only to be released
2984 * afterwards. This conservative approach is easily argued to avoid
2988 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp)
2990 vm_page_t m, pdppg, pdpg;
2991 pt_entry_t PG_A, PG_M, PG_RW, PG_V;
2993 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2995 PG_A = pmap_accessed_bit(pmap);
2996 PG_M = pmap_modified_bit(pmap);
2997 PG_V = pmap_valid_bit(pmap);
2998 PG_RW = pmap_rw_bit(pmap);
3001 * Allocate a page table page.
3003 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ |
3004 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
3005 if (lockp != NULL) {
3006 RELEASE_PV_LIST_LOCK(lockp);
3008 PMAP_ASSERT_NOT_IN_DI();
3014 * Indicate the need to retry. While waiting, the page table
3015 * page may have been allocated.
3019 if ((m->flags & PG_ZERO) == 0)
3023 * Map the pagetable page into the process address space, if
3024 * it isn't already there.
3027 if (ptepindex >= (NUPDE + NUPDPE)) {
3028 pml4_entry_t *pml4, *pml4u;
3029 vm_pindex_t pml4index;
3031 /* Wire up a new PDPE page */
3032 pml4index = ptepindex - (NUPDE + NUPDPE);
3033 pml4 = &pmap->pm_pml4[pml4index];
3034 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
3035 if (pmap->pm_pml4u != NULL && pml4index < NUPML4E) {
3037 * PTI: Make all user-space mappings in the
3038 * kernel-mode page table no-execute so that
3039 * we detect any programming errors that leave
3040 * the kernel-mode page table active on return
3043 if (pmap->pm_ucr3 != PMAP_NO_CR3)
3046 pml4u = &pmap->pm_pml4u[pml4index];
3047 *pml4u = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V |
3051 } else if (ptepindex >= NUPDE) {
3052 vm_pindex_t pml4index;
3053 vm_pindex_t pdpindex;
3057 /* Wire up a new PDE page */
3058 pdpindex = ptepindex - NUPDE;
3059 pml4index = pdpindex >> NPML4EPGSHIFT;
3061 pml4 = &pmap->pm_pml4[pml4index];
3062 if ((*pml4 & PG_V) == 0) {
3063 /* Have to allocate a new pdp, recurse */
3064 if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml4index,
3066 vm_page_unwire_noq(m);
3067 vm_page_free_zero(m);
3071 /* Add reference to pdp page */
3072 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
3073 pdppg->wire_count++;
3075 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
3077 /* Now find the pdp page */
3078 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
3079 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
3082 vm_pindex_t pml4index;
3083 vm_pindex_t pdpindex;
3088 /* Wire up a new PTE page */
3089 pdpindex = ptepindex >> NPDPEPGSHIFT;
3090 pml4index = pdpindex >> NPML4EPGSHIFT;
3092 /* First, find the pdp and check that its valid. */
3093 pml4 = &pmap->pm_pml4[pml4index];
3094 if ((*pml4 & PG_V) == 0) {
3095 /* Have to allocate a new pd, recurse */
3096 if (_pmap_allocpte(pmap, NUPDE + pdpindex,
3098 vm_page_unwire_noq(m);
3099 vm_page_free_zero(m);
3102 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
3103 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
3105 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
3106 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
3107 if ((*pdp & PG_V) == 0) {
3108 /* Have to allocate a new pd, recurse */
3109 if (_pmap_allocpte(pmap, NUPDE + pdpindex,
3111 vm_page_unwire_noq(m);
3112 vm_page_free_zero(m);
3116 /* Add reference to the pd page */
3117 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
3121 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
3123 /* Now we know where the page directory page is */
3124 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
3125 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
3128 pmap_resident_count_inc(pmap, 1);
3134 pmap_allocpde(pmap_t pmap, vm_offset_t va, struct rwlock **lockp)
3136 vm_pindex_t pdpindex, ptepindex;
3137 pdp_entry_t *pdpe, PG_V;
3140 PG_V = pmap_valid_bit(pmap);
3143 pdpe = pmap_pdpe(pmap, va);
3144 if (pdpe != NULL && (*pdpe & PG_V) != 0) {
3145 /* Add a reference to the pd page. */
3146 pdpg = PHYS_TO_VM_PAGE(*pdpe & PG_FRAME);
3149 /* Allocate a pd page. */
3150 ptepindex = pmap_pde_pindex(va);
3151 pdpindex = ptepindex >> NPDPEPGSHIFT;
3152 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex, lockp);
3153 if (pdpg == NULL && lockp != NULL)
3160 pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp)
3162 vm_pindex_t ptepindex;
3163 pd_entry_t *pd, PG_V;
3166 PG_V = pmap_valid_bit(pmap);
3169 * Calculate pagetable page index
3171 ptepindex = pmap_pde_pindex(va);
3174 * Get the page directory entry
3176 pd = pmap_pde(pmap, va);
3179 * This supports switching from a 2MB page to a
3182 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
3183 if (!pmap_demote_pde_locked(pmap, pd, va, lockp)) {
3185 * Invalidation of the 2MB page mapping may have caused
3186 * the deallocation of the underlying PD page.
3193 * If the page table page is mapped, we just increment the
3194 * hold count, and activate it.
3196 if (pd != NULL && (*pd & PG_V) != 0) {
3197 m = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
3201 * Here if the pte page isn't mapped, or if it has been
3204 m = _pmap_allocpte(pmap, ptepindex, lockp);
3205 if (m == NULL && lockp != NULL)
3212 /***************************************************
3213 * Pmap allocation/deallocation routines.
3214 ***************************************************/
3217 * Release any resources held by the given physical map.
3218 * Called when a pmap initialized by pmap_pinit is being released.
3219 * Should only be called if the map contains no valid mappings.
3222 pmap_release(pmap_t pmap)
3227 KASSERT(pmap->pm_stats.resident_count == 0,
3228 ("pmap_release: pmap resident count %ld != 0",
3229 pmap->pm_stats.resident_count));
3230 KASSERT(vm_radix_is_empty(&pmap->pm_root),
3231 ("pmap_release: pmap has reserved page table page(s)"));
3232 KASSERT(CPU_EMPTY(&pmap->pm_active),
3233 ("releasing active pmap %p", pmap));
3235 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_pml4));
3237 for (i = 0; i < NKPML4E; i++) /* KVA */
3238 pmap->pm_pml4[KPML4BASE + i] = 0;
3239 for (i = 0; i < ndmpdpphys; i++)/* Direct Map */
3240 pmap->pm_pml4[DMPML4I + i] = 0;
3241 pmap->pm_pml4[PML4PML4I] = 0; /* Recursive Mapping */
3242 for (i = 0; i < lm_ents; i++) /* Large Map */
3243 pmap->pm_pml4[LMSPML4I + i] = 0;
3245 vm_page_unwire_noq(m);
3246 vm_page_free_zero(m);
3248 if (pmap->pm_pml4u != NULL) {
3249 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_pml4u));
3250 vm_page_unwire_noq(m);
3256 kvm_size(SYSCTL_HANDLER_ARGS)
3258 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS;
3260 return sysctl_handle_long(oidp, &ksize, 0, req);
3262 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
3263 0, 0, kvm_size, "LU", "Size of KVM");
3266 kvm_free(SYSCTL_HANDLER_ARGS)
3268 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
3270 return sysctl_handle_long(oidp, &kfree, 0, req);
3272 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
3273 0, 0, kvm_free, "LU", "Amount of KVM free");
3276 * grow the number of kernel page table entries, if needed
3279 pmap_growkernel(vm_offset_t addr)
3283 pd_entry_t *pde, newpdir;
3286 mtx_assert(&kernel_map->system_mtx, MA_OWNED);
3289 * Return if "addr" is within the range of kernel page table pages
3290 * that were preallocated during pmap bootstrap. Moreover, leave
3291 * "kernel_vm_end" and the kernel page table as they were.
3293 * The correctness of this action is based on the following
3294 * argument: vm_map_insert() allocates contiguous ranges of the
3295 * kernel virtual address space. It calls this function if a range
3296 * ends after "kernel_vm_end". If the kernel is mapped between
3297 * "kernel_vm_end" and "addr", then the range cannot begin at
3298 * "kernel_vm_end". In fact, its beginning address cannot be less
3299 * than the kernel. Thus, there is no immediate need to allocate
3300 * any new kernel page table pages between "kernel_vm_end" and
3303 if (KERNBASE < addr && addr <= KERNBASE + nkpt * NBPDR)
3306 addr = roundup2(addr, NBPDR);
3307 if (addr - 1 >= vm_map_max(kernel_map))
3308 addr = vm_map_max(kernel_map);
3309 while (kernel_vm_end < addr) {
3310 pdpe = pmap_pdpe(kernel_pmap, kernel_vm_end);
3311 if ((*pdpe & X86_PG_V) == 0) {
3312 /* We need a new PDP entry */
3313 nkpg = vm_page_alloc(NULL, kernel_vm_end >> PDPSHIFT,
3314 VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ |
3315 VM_ALLOC_WIRED | VM_ALLOC_ZERO);
3317 panic("pmap_growkernel: no memory to grow kernel");
3318 if ((nkpg->flags & PG_ZERO) == 0)
3319 pmap_zero_page(nkpg);
3320 paddr = VM_PAGE_TO_PHYS(nkpg);
3321 *pdpe = (pdp_entry_t)(paddr | X86_PG_V | X86_PG_RW |
3322 X86_PG_A | X86_PG_M);
3323 continue; /* try again */
3325 pde = pmap_pdpe_to_pde(pdpe, kernel_vm_end);
3326 if ((*pde & X86_PG_V) != 0) {
3327 kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK;
3328 if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
3329 kernel_vm_end = vm_map_max(kernel_map);
3335 nkpg = vm_page_alloc(NULL, pmap_pde_pindex(kernel_vm_end),
3336 VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
3339 panic("pmap_growkernel: no memory to grow kernel");
3340 if ((nkpg->flags & PG_ZERO) == 0)
3341 pmap_zero_page(nkpg);
3342 paddr = VM_PAGE_TO_PHYS(nkpg);
3343 newpdir = paddr | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M;
3344 pde_store(pde, newpdir);
3346 kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK;
3347 if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
3348 kernel_vm_end = vm_map_max(kernel_map);
3355 /***************************************************
3356 * page management routines.
3357 ***************************************************/
3359 CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE);
3360 CTASSERT(_NPCM == 3);
3361 CTASSERT(_NPCPV == 168);
3363 static __inline struct pv_chunk *
3364 pv_to_chunk(pv_entry_t pv)
3367 return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK));
3370 #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap)
3372 #define PC_FREE0 0xfffffffffffffffful
3373 #define PC_FREE1 0xfffffffffffffffful
3374 #define PC_FREE2 0x000000fffffffffful
3376 static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1, PC_FREE2 };
3379 static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail;
3381 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0,
3382 "Current number of pv entry chunks");
3383 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0,
3384 "Current number of pv entry chunks allocated");
3385 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0,
3386 "Current number of pv entry chunks frees");
3387 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0,
3388 "Number of times tried to get a chunk page but failed.");
3390 static long pv_entry_frees, pv_entry_allocs, pv_entry_count;
3391 static int pv_entry_spare;
3393 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0,
3394 "Current number of pv entry frees");
3395 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0,
3396 "Current number of pv entry allocs");
3397 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0,
3398 "Current number of pv entries");
3399 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0,
3400 "Current number of spare pv entries");
3404 reclaim_pv_chunk_leave_pmap(pmap_t pmap, pmap_t locked_pmap, bool start_di)
3409 pmap_invalidate_all(pmap);
3410 if (pmap != locked_pmap)
3413 pmap_delayed_invl_finished();
3417 * We are in a serious low memory condition. Resort to
3418 * drastic measures to free some pages so we can allocate
3419 * another pv entry chunk.
3421 * Returns NULL if PV entries were reclaimed from the specified pmap.
3423 * We do not, however, unmap 2mpages because subsequent accesses will
3424 * allocate per-page pv entries until repromotion occurs, thereby
3425 * exacerbating the shortage of free pv entries.
3428 reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp)
3430 struct pv_chunk *pc, *pc_marker, *pc_marker_end;
3431 struct pv_chunk_header pc_marker_b, pc_marker_end_b;
3432 struct md_page *pvh;
3434 pmap_t next_pmap, pmap;
3435 pt_entry_t *pte, tpte;
3436 pt_entry_t PG_G, PG_A, PG_M, PG_RW;
3440 struct spglist free;
3442 int bit, field, freed;
3444 static int active_reclaims = 0;
3446 PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED);
3447 KASSERT(lockp != NULL, ("reclaim_pv_chunk: lockp is NULL"));
3450 PG_G = PG_A = PG_M = PG_RW = 0;
3452 bzero(&pc_marker_b, sizeof(pc_marker_b));
3453 bzero(&pc_marker_end_b, sizeof(pc_marker_end_b));
3454 pc_marker = (struct pv_chunk *)&pc_marker_b;
3455 pc_marker_end = (struct pv_chunk *)&pc_marker_end_b;
3458 * A delayed invalidation block should already be active if
3459 * pmap_advise() or pmap_remove() called this function by way
3460 * of pmap_demote_pde_locked().
3462 start_di = pmap_not_in_di();
3464 mtx_lock(&pv_chunks_mutex);
3466 TAILQ_INSERT_HEAD(&pv_chunks, pc_marker, pc_lru);
3467 TAILQ_INSERT_TAIL(&pv_chunks, pc_marker_end, pc_lru);
3468 while ((pc = TAILQ_NEXT(pc_marker, pc_lru)) != pc_marker_end &&
3469 SLIST_EMPTY(&free)) {
3470 next_pmap = pc->pc_pmap;
3471 if (next_pmap == NULL) {
3473 * The next chunk is a marker. However, it is
3474 * not our marker, so active_reclaims must be
3475 * > 1. Consequently, the next_chunk code
3476 * will not rotate the pv_chunks list.
3480 mtx_unlock(&pv_chunks_mutex);
3483 * A pv_chunk can only be removed from the pc_lru list
3484 * when both pc_chunks_mutex is owned and the
3485 * corresponding pmap is locked.
3487 if (pmap != next_pmap) {
3488 reclaim_pv_chunk_leave_pmap(pmap, locked_pmap,
3491 /* Avoid deadlock and lock recursion. */
3492 if (pmap > locked_pmap) {
3493 RELEASE_PV_LIST_LOCK(lockp);
3496 pmap_delayed_invl_started();
3497 mtx_lock(&pv_chunks_mutex);
3499 } else if (pmap != locked_pmap) {
3500 if (PMAP_TRYLOCK(pmap)) {
3502 pmap_delayed_invl_started();
3503 mtx_lock(&pv_chunks_mutex);
3506 pmap = NULL; /* pmap is not locked */
3507 mtx_lock(&pv_chunks_mutex);
3508 pc = TAILQ_NEXT(pc_marker, pc_lru);
3510 pc->pc_pmap != next_pmap)
3514 } else if (start_di)
3515 pmap_delayed_invl_started();
3516 PG_G = pmap_global_bit(pmap);
3517 PG_A = pmap_accessed_bit(pmap);
3518 PG_M = pmap_modified_bit(pmap);
3519 PG_RW = pmap_rw_bit(pmap);
3523 * Destroy every non-wired, 4 KB page mapping in the chunk.
3526 for (field = 0; field < _NPCM; field++) {
3527 for (inuse = ~pc->pc_map[field] & pc_freemask[field];
3528 inuse != 0; inuse &= ~(1UL << bit)) {
3530 pv = &pc->pc_pventry[field * 64 + bit];
3532 pde = pmap_pde(pmap, va);
3533 if ((*pde & PG_PS) != 0)
3535 pte = pmap_pde_to_pte(pde, va);
3536 if ((*pte & PG_W) != 0)
3538 tpte = pte_load_clear(pte);
3539 if ((tpte & PG_G) != 0)
3540 pmap_invalidate_page(pmap, va);
3541 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3542 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
3544 if ((tpte & PG_A) != 0)
3545 vm_page_aflag_set(m, PGA_REFERENCED);
3546 CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
3547 TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
3549 if (TAILQ_EMPTY(&m->md.pv_list) &&
3550 (m->flags & PG_FICTITIOUS) == 0) {
3551 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
3552 if (TAILQ_EMPTY(&pvh->pv_list)) {
3553 vm_page_aflag_clear(m,
3557 pmap_delayed_invl_page(m);
3558 pc->pc_map[field] |= 1UL << bit;
3559 pmap_unuse_pt(pmap, va, *pde, &free);
3564 mtx_lock(&pv_chunks_mutex);
3567 /* Every freed mapping is for a 4 KB page. */
3568 pmap_resident_count_dec(pmap, freed);
3569 PV_STAT(atomic_add_long(&pv_entry_frees, freed));
3570 PV_STAT(atomic_add_int(&pv_entry_spare, freed));
3571 PV_STAT(atomic_subtract_long(&pv_entry_count, freed));
3572 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
3573 if (pc->pc_map[0] == PC_FREE0 && pc->pc_map[1] == PC_FREE1 &&
3574 pc->pc_map[2] == PC_FREE2) {
3575 PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV));
3576 PV_STAT(atomic_subtract_int(&pc_chunk_count, 1));
3577 PV_STAT(atomic_add_int(&pc_chunk_frees, 1));
3578 /* Entire chunk is free; return it. */
3579 m_pc = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc));
3580 dump_drop_page(m_pc->phys_addr);
3581 mtx_lock(&pv_chunks_mutex);
3582 TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
3585 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
3586 mtx_lock(&pv_chunks_mutex);
3587 /* One freed pv entry in locked_pmap is sufficient. */
3588 if (pmap == locked_pmap)
3591 TAILQ_REMOVE(&pv_chunks, pc_marker, pc_lru);
3592 TAILQ_INSERT_AFTER(&pv_chunks, pc, pc_marker, pc_lru);
3593 if (active_reclaims == 1 && pmap != NULL) {
3595 * Rotate the pv chunks list so that we do not
3596 * scan the same pv chunks that could not be
3597 * freed (because they contained a wired
3598 * and/or superpage mapping) on every
3599 * invocation of reclaim_pv_chunk().
3601 while ((pc = TAILQ_FIRST(&pv_chunks)) != pc_marker) {
3602 MPASS(pc->pc_pmap != NULL);
3603 TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
3604 TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru);
3608 TAILQ_REMOVE(&pv_chunks, pc_marker, pc_lru);
3609 TAILQ_REMOVE(&pv_chunks, pc_marker_end, pc_lru);
3611 mtx_unlock(&pv_chunks_mutex);
3612 reclaim_pv_chunk_leave_pmap(pmap, locked_pmap, start_di);
3613 if (m_pc == NULL && !SLIST_EMPTY(&free)) {
3614 m_pc = SLIST_FIRST(&free);
3615 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
3616 /* Recycle a freed page table page. */
3617 m_pc->wire_count = 1;
3619 vm_page_free_pages_toq(&free, true);
3624 * free the pv_entry back to the free list
3627 free_pv_entry(pmap_t pmap, pv_entry_t pv)
3629 struct pv_chunk *pc;
3630 int idx, field, bit;
3632 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3633 PV_STAT(atomic_add_long(&pv_entry_frees, 1));
3634 PV_STAT(atomic_add_int(&pv_entry_spare, 1));
3635 PV_STAT(atomic_subtract_long(&pv_entry_count, 1));
3636 pc = pv_to_chunk(pv);
3637 idx = pv - &pc->pc_pventry[0];
3640 pc->pc_map[field] |= 1ul << bit;
3641 if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1 ||
3642 pc->pc_map[2] != PC_FREE2) {
3643 /* 98% of the time, pc is already at the head of the list. */
3644 if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) {
3645 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
3646 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
3650 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
3655 free_pv_chunk(struct pv_chunk *pc)
3659 mtx_lock(&pv_chunks_mutex);
3660 TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
3661 mtx_unlock(&pv_chunks_mutex);
3662 PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV));
3663 PV_STAT(atomic_subtract_int(&pc_chunk_count, 1));
3664 PV_STAT(atomic_add_int(&pc_chunk_frees, 1));
3665 /* entire chunk is free, return it */
3666 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc));
3667 dump_drop_page(m->phys_addr);
3668 vm_page_unwire(m, PQ_NONE);
3673 * Returns a new PV entry, allocating a new PV chunk from the system when
3674 * needed. If this PV chunk allocation fails and a PV list lock pointer was
3675 * given, a PV chunk is reclaimed from an arbitrary pmap. Otherwise, NULL is
3678 * The given PV list lock may be released.
3681 get_pv_entry(pmap_t pmap, struct rwlock **lockp)
3685 struct pv_chunk *pc;
3688 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3689 PV_STAT(atomic_add_long(&pv_entry_allocs, 1));
3691 pc = TAILQ_FIRST(&pmap->pm_pvchunk);
3693 for (field = 0; field < _NPCM; field++) {
3694 if (pc->pc_map[field]) {
3695 bit = bsfq(pc->pc_map[field]);
3699 if (field < _NPCM) {
3700 pv = &pc->pc_pventry[field * 64 + bit];
3701 pc->pc_map[field] &= ~(1ul << bit);
3702 /* If this was the last item, move it to tail */
3703 if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 &&
3704 pc->pc_map[2] == 0) {
3705 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
3706 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc,
3709 PV_STAT(atomic_add_long(&pv_entry_count, 1));
3710 PV_STAT(atomic_subtract_int(&pv_entry_spare, 1));
3714 /* No free items, allocate another chunk */
3715 m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
3718 if (lockp == NULL) {
3719 PV_STAT(pc_chunk_tryfail++);
3722 m = reclaim_pv_chunk(pmap, lockp);
3726 PV_STAT(atomic_add_int(&pc_chunk_count, 1));
3727 PV_STAT(atomic_add_int(&pc_chunk_allocs, 1));
3728 dump_add_page(m->phys_addr);
3729 pc = (void *)PHYS_TO_DMAP(m->phys_addr);
3731 pc->pc_map[0] = PC_FREE0 & ~1ul; /* preallocated bit 0 */
3732 pc->pc_map[1] = PC_FREE1;
3733 pc->pc_map[2] = PC_FREE2;
3734 mtx_lock(&pv_chunks_mutex);
3735 TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru);
3736 mtx_unlock(&pv_chunks_mutex);
3737 pv = &pc->pc_pventry[0];
3738 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
3739 PV_STAT(atomic_add_long(&pv_entry_count, 1));
3740 PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV - 1));
3745 * Returns the number of one bits within the given PV chunk map.
3747 * The erratas for Intel processors state that "POPCNT Instruction May
3748 * Take Longer to Execute Than Expected". It is believed that the
3749 * issue is the spurious dependency on the destination register.
3750 * Provide a hint to the register rename logic that the destination
3751 * value is overwritten, by clearing it, as suggested in the
3752 * optimization manual. It should be cheap for unaffected processors
3755 * Reference numbers for erratas are
3756 * 4th Gen Core: HSD146
3757 * 5th Gen Core: BDM85
3758 * 6th Gen Core: SKL029
3761 popcnt_pc_map_pq(uint64_t *map)
3765 __asm __volatile("xorl %k0,%k0;popcntq %2,%0;"
3766 "xorl %k1,%k1;popcntq %3,%1;addl %k1,%k0;"
3767 "xorl %k1,%k1;popcntq %4,%1;addl %k1,%k0"
3768 : "=&r" (result), "=&r" (tmp)
3769 : "m" (map[0]), "m" (map[1]), "m" (map[2]));
3774 * Ensure that the number of spare PV entries in the specified pmap meets or
3775 * exceeds the given count, "needed".
3777 * The given PV list lock may be released.
3780 reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp)
3782 struct pch new_tail;
3783 struct pv_chunk *pc;
3788 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3789 KASSERT(lockp != NULL, ("reserve_pv_entries: lockp is NULL"));
3792 * Newly allocated PV chunks must be stored in a private list until
3793 * the required number of PV chunks have been allocated. Otherwise,
3794 * reclaim_pv_chunk() could recycle one of these chunks. In
3795 * contrast, these chunks must be added to the pmap upon allocation.
3797 TAILQ_INIT(&new_tail);
3800 TAILQ_FOREACH(pc, &pmap->pm_pvchunk, pc_list) {
3802 if ((cpu_feature2 & CPUID2_POPCNT) == 0)
3803 bit_count((bitstr_t *)pc->pc_map, 0,
3804 sizeof(pc->pc_map) * NBBY, &free);
3807 free = popcnt_pc_map_pq(pc->pc_map);
3811 if (avail >= needed)
3814 for (reclaimed = false; avail < needed; avail += _NPCPV) {
3815 m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
3818 m = reclaim_pv_chunk(pmap, lockp);
3823 PV_STAT(atomic_add_int(&pc_chunk_count, 1));
3824 PV_STAT(atomic_add_int(&pc_chunk_allocs, 1));
3825 dump_add_page(m->phys_addr);
3826 pc = (void *)PHYS_TO_DMAP(m->phys_addr);
3828 pc->pc_map[0] = PC_FREE0;
3829 pc->pc_map[1] = PC_FREE1;
3830 pc->pc_map[2] = PC_FREE2;
3831 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
3832 TAILQ_INSERT_TAIL(&new_tail, pc, pc_lru);
3833 PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV));
3836 * The reclaim might have freed a chunk from the current pmap.
3837 * If that chunk contained available entries, we need to
3838 * re-count the number of available entries.
3843 if (!TAILQ_EMPTY(&new_tail)) {
3844 mtx_lock(&pv_chunks_mutex);
3845 TAILQ_CONCAT(&pv_chunks, &new_tail, pc_lru);
3846 mtx_unlock(&pv_chunks_mutex);
3851 * First find and then remove the pv entry for the specified pmap and virtual
3852 * address from the specified pv list. Returns the pv entry if found and NULL
3853 * otherwise. This operation can be performed on pv lists for either 4KB or
3854 * 2MB page mappings.
3856 static __inline pv_entry_t
3857 pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
3861 TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
3862 if (pmap == PV_PMAP(pv) && va == pv->pv_va) {
3863 TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
3872 * After demotion from a 2MB page mapping to 512 4KB page mappings,
3873 * destroy the pv entry for the 2MB page mapping and reinstantiate the pv
3874 * entries for each of the 4KB page mappings.
3877 pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
3878 struct rwlock **lockp)
3880 struct md_page *pvh;
3881 struct pv_chunk *pc;
3883 vm_offset_t va_last;
3887 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3888 KASSERT((pa & PDRMASK) == 0,
3889 ("pmap_pv_demote_pde: pa is not 2mpage aligned"));
3890 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
3893 * Transfer the 2mpage's pv entry for this mapping to the first
3894 * page's pv list. Once this transfer begins, the pv list lock
3895 * must not be released until the last pv entry is reinstantiated.
3897 pvh = pa_to_pvh(pa);
3898 va = trunc_2mpage(va);
3899 pv = pmap_pvh_remove(pvh, pmap, va);
3900 KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found"));
3901 m = PHYS_TO_VM_PAGE(pa);
3902 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
3904 /* Instantiate the remaining NPTEPG - 1 pv entries. */
3905 PV_STAT(atomic_add_long(&pv_entry_allocs, NPTEPG - 1));
3906 va_last = va + NBPDR - PAGE_SIZE;
3908 pc = TAILQ_FIRST(&pmap->pm_pvchunk);
3909 KASSERT(pc->pc_map[0] != 0 || pc->pc_map[1] != 0 ||
3910 pc->pc_map[2] != 0, ("pmap_pv_demote_pde: missing spare"));
3911 for (field = 0; field < _NPCM; field++) {
3912 while (pc->pc_map[field]) {
3913 bit = bsfq(pc->pc_map[field]);
3914 pc->pc_map[field] &= ~(1ul << bit);
3915 pv = &pc->pc_pventry[field * 64 + bit];
3919 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3920 ("pmap_pv_demote_pde: page %p is not managed", m));
3921 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
3927 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
3928 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
3931 if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) {
3932 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
3933 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
3935 PV_STAT(atomic_add_long(&pv_entry_count, NPTEPG - 1));
3936 PV_STAT(atomic_subtract_int(&pv_entry_spare, NPTEPG - 1));
3939 #if VM_NRESERVLEVEL > 0
3941 * After promotion from 512 4KB page mappings to a single 2MB page mapping,
3942 * replace the many pv entries for the 4KB page mappings by a single pv entry
3943 * for the 2MB page mapping.
3946 pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
3947 struct rwlock **lockp)
3949 struct md_page *pvh;
3951 vm_offset_t va_last;
3954 KASSERT((pa & PDRMASK) == 0,
3955 ("pmap_pv_promote_pde: pa is not 2mpage aligned"));
3956 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
3959 * Transfer the first page's pv entry for this mapping to the 2mpage's
3960 * pv list. Aside from avoiding the cost of a call to get_pv_entry(),
3961 * a transfer avoids the possibility that get_pv_entry() calls
3962 * reclaim_pv_chunk() and that reclaim_pv_chunk() removes one of the
3963 * mappings that is being promoted.
3965 m = PHYS_TO_VM_PAGE(pa);
3966 va = trunc_2mpage(va);
3967 pv = pmap_pvh_remove(&m->md, pmap, va);
3968 KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found"));
3969 pvh = pa_to_pvh(pa);
3970 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
3972 /* Free the remaining NPTEPG - 1 pv entries. */
3973 va_last = va + NBPDR - PAGE_SIZE;
3977 pmap_pvh_free(&m->md, pmap, va);
3978 } while (va < va_last);
3980 #endif /* VM_NRESERVLEVEL > 0 */
3983 * First find and then destroy the pv entry for the specified pmap and virtual
3984 * address. This operation can be performed on pv lists for either 4KB or 2MB
3988 pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
3992 pv = pmap_pvh_remove(pvh, pmap, va);
3993 KASSERT(pv != NULL, ("pmap_pvh_free: pv not found"));
3994 free_pv_entry(pmap, pv);
3998 * Conditionally create the PV entry for a 4KB page mapping if the required
3999 * memory can be allocated without resorting to reclamation.
4002 pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m,
4003 struct rwlock **lockp)
4007 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4008 /* Pass NULL instead of the lock pointer to disable reclamation. */
4009 if ((pv = get_pv_entry(pmap, NULL)) != NULL) {
4011 CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
4012 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
4020 * Create the PV entry for a 2MB page mapping. Always returns true unless the
4021 * flag PMAP_ENTER_NORECLAIM is specified. If that flag is specified, returns
4022 * false if the PV entry cannot be allocated without resorting to reclamation.
4025 pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags,
4026 struct rwlock **lockp)
4028 struct md_page *pvh;
4032 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4033 /* Pass NULL instead of the lock pointer to disable reclamation. */
4034 if ((pv = get_pv_entry(pmap, (flags & PMAP_ENTER_NORECLAIM) != 0 ?
4035 NULL : lockp)) == NULL)
4038 pa = pde & PG_PS_FRAME;
4039 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
4040 pvh = pa_to_pvh(pa);
4041 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
4047 * Fills a page table page with mappings to consecutive physical pages.
4050 pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte)
4054 for (pte = firstpte; pte < firstpte + NPTEPG; pte++) {
4056 newpte += PAGE_SIZE;
4061 * Tries to demote a 2MB page mapping. If demotion fails, the 2MB page
4062 * mapping is invalidated.
4065 pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
4067 struct rwlock *lock;
4071 rv = pmap_demote_pde_locked(pmap, pde, va, &lock);
4078 pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va,
4079 struct rwlock **lockp)
4081 pd_entry_t newpde, oldpde;
4082 pt_entry_t *firstpte, newpte;
4083 pt_entry_t PG_A, PG_G, PG_M, PG_RW, PG_V;
4086 struct spglist free;
4090 PG_G = pmap_global_bit(pmap);
4091 PG_A = pmap_accessed_bit(pmap);
4092 PG_M = pmap_modified_bit(pmap);
4093 PG_RW = pmap_rw_bit(pmap);
4094 PG_V = pmap_valid_bit(pmap);
4095 PG_PTE_CACHE = pmap_cache_mask(pmap, 0);
4097 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4099 KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V),
4100 ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V"));
4101 if ((oldpde & PG_A) == 0 || (mpte = pmap_remove_pt_page(pmap, va)) ==
4103 KASSERT((oldpde & PG_W) == 0,
4104 ("pmap_demote_pde: page table page for a wired mapping"
4108 * Invalidate the 2MB page mapping and return "failure" if the
4109 * mapping was never accessed or the allocation of the new
4110 * page table page fails. If the 2MB page mapping belongs to
4111 * the direct map region of the kernel's address space, then
4112 * the page allocation request specifies the highest possible
4113 * priority (VM_ALLOC_INTERRUPT). Otherwise, the priority is
4114 * normal. Page table pages are preallocated for every other
4115 * part of the kernel address space, so the direct map region
4116 * is the only part of the kernel address space that must be
4119 if ((oldpde & PG_A) == 0 || (mpte = vm_page_alloc(NULL,
4120 pmap_pde_pindex(va), (va >= DMAP_MIN_ADDRESS && va <
4121 DMAP_MAX_ADDRESS ? VM_ALLOC_INTERRUPT : VM_ALLOC_NORMAL) |
4122 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) {
4124 sva = trunc_2mpage(va);
4125 pmap_remove_pde(pmap, pde, sva, &free, lockp);
4126 if ((oldpde & PG_G) == 0)
4127 pmap_invalidate_pde_page(pmap, sva, oldpde);
4128 vm_page_free_pages_toq(&free, true);
4129 CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#lx"
4130 " in pmap %p", va, pmap);
4133 if (va < VM_MAXUSER_ADDRESS)
4134 pmap_resident_count_inc(pmap, 1);
4136 mptepa = VM_PAGE_TO_PHYS(mpte);
4137 firstpte = (pt_entry_t *)PHYS_TO_DMAP(mptepa);
4138 newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V;
4139 KASSERT((oldpde & PG_A) != 0,
4140 ("pmap_demote_pde: oldpde is missing PG_A"));
4141 KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW,
4142 ("pmap_demote_pde: oldpde is missing PG_M"));
4143 newpte = oldpde & ~PG_PS;
4144 newpte = pmap_swap_pat(pmap, newpte);
4147 * If the page table page is new, initialize it.
4149 if (mpte->wire_count == 1) {
4150 mpte->wire_count = NPTEPG;
4151 pmap_fill_ptp(firstpte, newpte);
4153 KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME),
4154 ("pmap_demote_pde: firstpte and newpte map different physical"
4158 * If the mapping has changed attributes, update the page table
4161 if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE))
4162 pmap_fill_ptp(firstpte, newpte);
4165 * The spare PV entries must be reserved prior to demoting the
4166 * mapping, that is, prior to changing the PDE. Otherwise, the state
4167 * of the PDE and the PV lists will be inconsistent, which can result
4168 * in reclaim_pv_chunk() attempting to remove a PV entry from the
4169 * wrong PV list and pmap_pv_demote_pde() failing to find the expected
4170 * PV entry for the 2MB page mapping that is being demoted.
4172 if ((oldpde & PG_MANAGED) != 0)
4173 reserve_pv_entries(pmap, NPTEPG - 1, lockp);
4176 * Demote the mapping. This pmap is locked. The old PDE has
4177 * PG_A set. If the old PDE has PG_RW set, it also has PG_M
4178 * set. Thus, there is no danger of a race with another
4179 * processor changing the setting of PG_A and/or PG_M between
4180 * the read above and the store below.
4182 if (workaround_erratum383)
4183 pmap_update_pde(pmap, va, pde, newpde);
4185 pde_store(pde, newpde);
4188 * Invalidate a stale recursive mapping of the page table page.
4190 if (va >= VM_MAXUSER_ADDRESS)
4191 pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va));
4194 * Demote the PV entry.
4196 if ((oldpde & PG_MANAGED) != 0)
4197 pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME, lockp);
4199 atomic_add_long(&pmap_pde_demotions, 1);
4200 CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#lx"
4201 " in pmap %p", va, pmap);
4206 * pmap_remove_kernel_pde: Remove a kernel superpage mapping.
4209 pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
4215 KASSERT(pmap == kernel_pmap, ("pmap %p is not kernel_pmap", pmap));
4216 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4217 mpte = pmap_remove_pt_page(pmap, va);
4219 panic("pmap_remove_kernel_pde: Missing pt page.");
4221 mptepa = VM_PAGE_TO_PHYS(mpte);
4222 newpde = mptepa | X86_PG_M | X86_PG_A | X86_PG_RW | X86_PG_V;
4225 * Initialize the page table page.
4227 pagezero((void *)PHYS_TO_DMAP(mptepa));
4230 * Demote the mapping.
4232 if (workaround_erratum383)
4233 pmap_update_pde(pmap, va, pde, newpde);
4235 pde_store(pde, newpde);
4238 * Invalidate a stale recursive mapping of the page table page.
4240 pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va));
4244 * pmap_remove_pde: do the things to unmap a superpage in a process
4247 pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
4248 struct spglist *free, struct rwlock **lockp)
4250 struct md_page *pvh;
4252 vm_offset_t eva, va;
4254 pt_entry_t PG_G, PG_A, PG_M, PG_RW;
4256 PG_G = pmap_global_bit(pmap);
4257 PG_A = pmap_accessed_bit(pmap);
4258 PG_M = pmap_modified_bit(pmap);
4259 PG_RW = pmap_rw_bit(pmap);
4261 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4262 KASSERT((sva & PDRMASK) == 0,
4263 ("pmap_remove_pde: sva is not 2mpage aligned"));
4264 oldpde = pte_load_clear(pdq);
4266 pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE;
4267 if ((oldpde & PG_G) != 0)
4268 pmap_invalidate_pde_page(kernel_pmap, sva, oldpde);
4269 pmap_resident_count_dec(pmap, NBPDR / PAGE_SIZE);
4270 if (oldpde & PG_MANAGED) {
4271 CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, oldpde & PG_PS_FRAME);
4272 pvh = pa_to_pvh(oldpde & PG_PS_FRAME);
4273 pmap_pvh_free(pvh, pmap, sva);
4275 for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
4276 va < eva; va += PAGE_SIZE, m++) {
4277 if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW))
4280 vm_page_aflag_set(m, PGA_REFERENCED);
4281 if (TAILQ_EMPTY(&m->md.pv_list) &&
4282 TAILQ_EMPTY(&pvh->pv_list))
4283 vm_page_aflag_clear(m, PGA_WRITEABLE);
4284 pmap_delayed_invl_page(m);
4287 if (pmap == kernel_pmap) {
4288 pmap_remove_kernel_pde(pmap, pdq, sva);
4290 mpte = pmap_remove_pt_page(pmap, sva);
4292 pmap_resident_count_dec(pmap, 1);
4293 KASSERT(mpte->wire_count == NPTEPG,
4294 ("pmap_remove_pde: pte page wire count error"));
4295 mpte->wire_count = 0;
4296 pmap_add_delayed_free_list(mpte, free, FALSE);
4299 return (pmap_unuse_pt(pmap, sva, *pmap_pdpe(pmap, sva), free));
4303 * pmap_remove_pte: do the things to unmap a page in a process
4306 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va,
4307 pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp)
4309 struct md_page *pvh;
4310 pt_entry_t oldpte, PG_A, PG_M, PG_RW;
4313 PG_A = pmap_accessed_bit(pmap);
4314 PG_M = pmap_modified_bit(pmap);
4315 PG_RW = pmap_rw_bit(pmap);
4317 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4318 oldpte = pte_load_clear(ptq);
4320 pmap->pm_stats.wired_count -= 1;
4321 pmap_resident_count_dec(pmap, 1);
4322 if (oldpte & PG_MANAGED) {
4323 m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME);
4324 if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
4327 vm_page_aflag_set(m, PGA_REFERENCED);
4328 CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
4329 pmap_pvh_free(&m->md, pmap, va);
4330 if (TAILQ_EMPTY(&m->md.pv_list) &&
4331 (m->flags & PG_FICTITIOUS) == 0) {
4332 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4333 if (TAILQ_EMPTY(&pvh->pv_list))
4334 vm_page_aflag_clear(m, PGA_WRITEABLE);
4336 pmap_delayed_invl_page(m);
4338 return (pmap_unuse_pt(pmap, va, ptepde, free));
4342 * Remove a single page from a process address space
4345 pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
4346 struct spglist *free)
4348 struct rwlock *lock;
4349 pt_entry_t *pte, PG_V;
4351 PG_V = pmap_valid_bit(pmap);
4352 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4353 if ((*pde & PG_V) == 0)
4355 pte = pmap_pde_to_pte(pde, va);
4356 if ((*pte & PG_V) == 0)
4359 pmap_remove_pte(pmap, pte, va, *pde, free, &lock);
4362 pmap_invalidate_page(pmap, va);
4366 * Removes the specified range of addresses from the page table page.
4369 pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
4370 pd_entry_t *pde, struct spglist *free, struct rwlock **lockp)
4372 pt_entry_t PG_G, *pte;
4376 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4377 PG_G = pmap_global_bit(pmap);
4380 for (pte = pmap_pde_to_pte(pde, sva); sva != eva; pte++,
4384 pmap_invalidate_range(pmap, va, sva);
4389 if ((*pte & PG_G) == 0)
4393 if (pmap_remove_pte(pmap, pte, sva, *pde, free, lockp)) {
4399 pmap_invalidate_range(pmap, va, sva);
4404 * Remove the given range of addresses from the specified map.
4406 * It is assumed that the start and end are properly
4407 * rounded to the page size.
4410 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4412 struct rwlock *lock;
4413 vm_offset_t va_next;
4414 pml4_entry_t *pml4e;
4416 pd_entry_t ptpaddr, *pde;
4417 pt_entry_t PG_G, PG_V;
4418 struct spglist free;
4421 PG_G = pmap_global_bit(pmap);
4422 PG_V = pmap_valid_bit(pmap);
4425 * Perform an unsynchronized read. This is, however, safe.
4427 if (pmap->pm_stats.resident_count == 0)
4433 pmap_delayed_invl_started();
4437 * special handling of removing one page. a very
4438 * common operation and easy to short circuit some
4441 if (sva + PAGE_SIZE == eva) {
4442 pde = pmap_pde(pmap, sva);
4443 if (pde && (*pde & PG_PS) == 0) {
4444 pmap_remove_page(pmap, sva, pde, &free);
4450 for (; sva < eva; sva = va_next) {
4452 if (pmap->pm_stats.resident_count == 0)
4455 pml4e = pmap_pml4e(pmap, sva);
4456 if ((*pml4e & PG_V) == 0) {
4457 va_next = (sva + NBPML4) & ~PML4MASK;
4463 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
4464 if ((*pdpe & PG_V) == 0) {
4465 va_next = (sva + NBPDP) & ~PDPMASK;
4472 * Calculate index for next page table.
4474 va_next = (sva + NBPDR) & ~PDRMASK;
4478 pde = pmap_pdpe_to_pde(pdpe, sva);
4482 * Weed out invalid mappings.
4488 * Check for large page.
4490 if ((ptpaddr & PG_PS) != 0) {
4492 * Are we removing the entire large page? If not,
4493 * demote the mapping and fall through.
4495 if (sva + NBPDR == va_next && eva >= va_next) {
4497 * The TLB entry for a PG_G mapping is
4498 * invalidated by pmap_remove_pde().
4500 if ((ptpaddr & PG_G) == 0)
4502 pmap_remove_pde(pmap, pde, sva, &free, &lock);
4504 } else if (!pmap_demote_pde_locked(pmap, pde, sva,
4506 /* The large page mapping was destroyed. */
4513 * Limit our scan to either the end of the va represented
4514 * by the current page table page, or to the end of the
4515 * range being removed.
4520 if (pmap_remove_ptes(pmap, sva, va_next, pde, &free, &lock))
4527 pmap_invalidate_all(pmap);
4529 pmap_delayed_invl_finished();
4530 vm_page_free_pages_toq(&free, true);
4534 * Routine: pmap_remove_all
4536 * Removes this physical page from
4537 * all physical maps in which it resides.
4538 * Reflects back modify bits to the pager.
4541 * Original versions of this routine were very
4542 * inefficient because they iteratively called
4543 * pmap_remove (slow...)
4547 pmap_remove_all(vm_page_t m)
4549 struct md_page *pvh;
4552 struct rwlock *lock;
4553 pt_entry_t *pte, tpte, PG_A, PG_M, PG_RW;
4556 struct spglist free;
4557 int pvh_gen, md_gen;
4559 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4560 ("pmap_remove_all: page %p is not managed", m));
4562 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
4563 pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
4564 pa_to_pvh(VM_PAGE_TO_PHYS(m));
4567 while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) {
4569 if (!PMAP_TRYLOCK(pmap)) {
4570 pvh_gen = pvh->pv_gen;
4574 if (pvh_gen != pvh->pv_gen) {
4581 pde = pmap_pde(pmap, va);
4582 (void)pmap_demote_pde_locked(pmap, pde, va, &lock);
4585 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4587 if (!PMAP_TRYLOCK(pmap)) {
4588 pvh_gen = pvh->pv_gen;
4589 md_gen = m->md.pv_gen;
4593 if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
4599 PG_A = pmap_accessed_bit(pmap);
4600 PG_M = pmap_modified_bit(pmap);
4601 PG_RW = pmap_rw_bit(pmap);
4602 pmap_resident_count_dec(pmap, 1);
4603 pde = pmap_pde(pmap, pv->pv_va);
4604 KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found"
4605 " a 2mpage in page %p's pv list", m));
4606 pte = pmap_pde_to_pte(pde, pv->pv_va);
4607 tpte = pte_load_clear(pte);
4609 pmap->pm_stats.wired_count--;
4611 vm_page_aflag_set(m, PGA_REFERENCED);
4614 * Update the vm_page_t clean and reference bits.
4616 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
4618 pmap_unuse_pt(pmap, pv->pv_va, *pde, &free);
4619 pmap_invalidate_page(pmap, pv->pv_va);
4620 TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
4622 free_pv_entry(pmap, pv);
4625 vm_page_aflag_clear(m, PGA_WRITEABLE);
4627 pmap_delayed_invl_wait(m);
4628 vm_page_free_pages_toq(&free, true);
4632 * pmap_protect_pde: do the things to protect a 2mpage in a process
4635 pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot)
4637 pd_entry_t newpde, oldpde;
4638 vm_offset_t eva, va;
4640 boolean_t anychanged;
4641 pt_entry_t PG_G, PG_M, PG_RW;
4643 PG_G = pmap_global_bit(pmap);
4644 PG_M = pmap_modified_bit(pmap);
4645 PG_RW = pmap_rw_bit(pmap);
4647 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4648 KASSERT((sva & PDRMASK) == 0,
4649 ("pmap_protect_pde: sva is not 2mpage aligned"));
4652 oldpde = newpde = *pde;
4653 if ((oldpde & (PG_MANAGED | PG_M | PG_RW)) ==
4654 (PG_MANAGED | PG_M | PG_RW)) {
4656 for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
4657 va < eva; va += PAGE_SIZE, m++)
4660 if ((prot & VM_PROT_WRITE) == 0)
4661 newpde &= ~(PG_RW | PG_M);
4662 if ((prot & VM_PROT_EXECUTE) == 0)
4664 if (newpde != oldpde) {
4666 * As an optimization to future operations on this PDE, clear
4667 * PG_PROMOTED. The impending invalidation will remove any
4668 * lingering 4KB page mappings from the TLB.
4670 if (!atomic_cmpset_long(pde, oldpde, newpde & ~PG_PROMOTED))
4672 if ((oldpde & PG_G) != 0)
4673 pmap_invalidate_pde_page(kernel_pmap, sva, oldpde);
4677 return (anychanged);
4681 * Set the physical protection on the
4682 * specified range of this map as requested.
4685 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4687 vm_offset_t va_next;
4688 pml4_entry_t *pml4e;
4690 pd_entry_t ptpaddr, *pde;
4691 pt_entry_t *pte, PG_G, PG_M, PG_RW, PG_V;
4692 boolean_t anychanged;
4694 KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot));
4695 if (prot == VM_PROT_NONE) {
4696 pmap_remove(pmap, sva, eva);
4700 if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) ==
4701 (VM_PROT_WRITE|VM_PROT_EXECUTE))
4704 PG_G = pmap_global_bit(pmap);
4705 PG_M = pmap_modified_bit(pmap);
4706 PG_V = pmap_valid_bit(pmap);
4707 PG_RW = pmap_rw_bit(pmap);
4711 * Although this function delays and batches the invalidation
4712 * of stale TLB entries, it does not need to call
4713 * pmap_delayed_invl_started() and
4714 * pmap_delayed_invl_finished(), because it does not
4715 * ordinarily destroy mappings. Stale TLB entries from
4716 * protection-only changes need only be invalidated before the
4717 * pmap lock is released, because protection-only changes do
4718 * not destroy PV entries. Even operations that iterate over
4719 * a physical page's PV list of mappings, like
4720 * pmap_remove_write(), acquire the pmap lock for each
4721 * mapping. Consequently, for protection-only changes, the
4722 * pmap lock suffices to synchronize both page table and TLB
4725 * This function only destroys a mapping if pmap_demote_pde()
4726 * fails. In that case, stale TLB entries are immediately
4731 for (; sva < eva; sva = va_next) {
4733 pml4e = pmap_pml4e(pmap, sva);
4734 if ((*pml4e & PG_V) == 0) {
4735 va_next = (sva + NBPML4) & ~PML4MASK;
4741 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
4742 if ((*pdpe & PG_V) == 0) {
4743 va_next = (sva + NBPDP) & ~PDPMASK;
4749 va_next = (sva + NBPDR) & ~PDRMASK;
4753 pde = pmap_pdpe_to_pde(pdpe, sva);
4757 * Weed out invalid mappings.
4763 * Check for large page.
4765 if ((ptpaddr & PG_PS) != 0) {
4767 * Are we protecting the entire large page? If not,
4768 * demote the mapping and fall through.
4770 if (sva + NBPDR == va_next && eva >= va_next) {
4772 * The TLB entry for a PG_G mapping is
4773 * invalidated by pmap_protect_pde().
4775 if (pmap_protect_pde(pmap, pde, sva, prot))
4778 } else if (!pmap_demote_pde(pmap, pde, sva)) {
4780 * The large page mapping was destroyed.
4789 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
4791 pt_entry_t obits, pbits;
4795 obits = pbits = *pte;
4796 if ((pbits & PG_V) == 0)
4799 if ((prot & VM_PROT_WRITE) == 0) {
4800 if ((pbits & (PG_MANAGED | PG_M | PG_RW)) ==
4801 (PG_MANAGED | PG_M | PG_RW)) {
4802 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4805 pbits &= ~(PG_RW | PG_M);
4807 if ((prot & VM_PROT_EXECUTE) == 0)
4810 if (pbits != obits) {
4811 if (!atomic_cmpset_long(pte, obits, pbits))
4814 pmap_invalidate_page(pmap, sva);
4821 pmap_invalidate_all(pmap);
4825 #if VM_NRESERVLEVEL > 0
4827 * Tries to promote the 512, contiguous 4KB page mappings that are within a
4828 * single page table page (PTP) to a single 2MB page mapping. For promotion
4829 * to occur, two conditions must be met: (1) the 4KB page mappings must map
4830 * aligned, contiguous physical memory and (2) the 4KB page mappings must have
4831 * identical characteristics.
4834 pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va,
4835 struct rwlock **lockp)
4838 pt_entry_t *firstpte, oldpte, pa, *pte;
4839 pt_entry_t PG_G, PG_A, PG_M, PG_RW, PG_V;
4843 PG_A = pmap_accessed_bit(pmap);
4844 PG_G = pmap_global_bit(pmap);
4845 PG_M = pmap_modified_bit(pmap);
4846 PG_V = pmap_valid_bit(pmap);
4847 PG_RW = pmap_rw_bit(pmap);
4848 PG_PTE_CACHE = pmap_cache_mask(pmap, 0);
4850 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4853 * Examine the first PTE in the specified PTP. Abort if this PTE is
4854 * either invalid, unused, or does not map the first 4KB physical page
4855 * within a 2MB page.
4857 firstpte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
4860 if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V)) {
4861 atomic_add_long(&pmap_pde_p_failures, 1);
4862 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx"
4863 " in pmap %p", va, pmap);
4866 if ((newpde & (PG_M | PG_RW)) == PG_RW) {
4868 * When PG_M is already clear, PG_RW can be cleared without
4869 * a TLB invalidation.
4871 if (!atomic_cmpset_long(firstpte, newpde, newpde & ~PG_RW))
4877 * Examine each of the other PTEs in the specified PTP. Abort if this
4878 * PTE maps an unexpected 4KB physical page or does not have identical
4879 * characteristics to the first PTE.
4881 pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE;
4882 for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) {
4885 if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) {
4886 atomic_add_long(&pmap_pde_p_failures, 1);
4887 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx"
4888 " in pmap %p", va, pmap);
4891 if ((oldpte & (PG_M | PG_RW)) == PG_RW) {
4893 * When PG_M is already clear, PG_RW can be cleared
4894 * without a TLB invalidation.
4896 if (!atomic_cmpset_long(pte, oldpte, oldpte & ~PG_RW))
4899 CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#lx"
4900 " in pmap %p", (oldpte & PG_FRAME & PDRMASK) |
4901 (va & ~PDRMASK), pmap);
4903 if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) {
4904 atomic_add_long(&pmap_pde_p_failures, 1);
4905 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx"
4906 " in pmap %p", va, pmap);
4913 * Save the page table page in its current state until the PDE
4914 * mapping the superpage is demoted by pmap_demote_pde() or
4915 * destroyed by pmap_remove_pde().
4917 mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
4918 KASSERT(mpte >= vm_page_array &&
4919 mpte < &vm_page_array[vm_page_array_size],
4920 ("pmap_promote_pde: page table page is out of range"));
4921 KASSERT(mpte->pindex == pmap_pde_pindex(va),
4922 ("pmap_promote_pde: page table page's pindex is wrong"));
4923 if (pmap_insert_pt_page(pmap, mpte)) {
4924 atomic_add_long(&pmap_pde_p_failures, 1);
4926 "pmap_promote_pde: failure for va %#lx in pmap %p", va,
4932 * Promote the pv entries.
4934 if ((newpde & PG_MANAGED) != 0)
4935 pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME, lockp);
4938 * Propagate the PAT index to its proper position.
4940 newpde = pmap_swap_pat(pmap, newpde);
4943 * Map the superpage.
4945 if (workaround_erratum383)
4946 pmap_update_pde(pmap, va, pde, PG_PS | newpde);
4948 pde_store(pde, PG_PROMOTED | PG_PS | newpde);
4950 atomic_add_long(&pmap_pde_promotions, 1);
4951 CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#lx"
4952 " in pmap %p", va, pmap);
4954 #endif /* VM_NRESERVLEVEL > 0 */
4957 * Insert the given physical page (p) at
4958 * the specified virtual address (v) in the
4959 * target physical map with the protection requested.
4961 * If specified, the page will be wired down, meaning
4962 * that the related pte can not be reclaimed.
4964 * NB: This is the only routine which MAY NOT lazy-evaluate
4965 * or lose information. That is, this routine must actually
4966 * insert this page into the given map NOW.
4968 * When destroying both a page table and PV entry, this function
4969 * performs the TLB invalidation before releasing the PV list
4970 * lock, so we do not need pmap_delayed_invl_page() calls here.
4973 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4974 u_int flags, int8_t psind)
4976 struct rwlock *lock;
4978 pt_entry_t *pte, PG_G, PG_A, PG_M, PG_RW, PG_V;
4979 pt_entry_t newpte, origpte;
4986 PG_A = pmap_accessed_bit(pmap);
4987 PG_G = pmap_global_bit(pmap);
4988 PG_M = pmap_modified_bit(pmap);
4989 PG_V = pmap_valid_bit(pmap);
4990 PG_RW = pmap_rw_bit(pmap);
4992 va = trunc_page(va);
4993 KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig"));
4994 KASSERT(va < UPT_MIN_ADDRESS || va >= UPT_MAX_ADDRESS,
4995 ("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)",
4997 KASSERT((m->oflags & VPO_UNMANAGED) != 0 || va < kmi.clean_sva ||
4998 va >= kmi.clean_eva,
4999 ("pmap_enter: managed mapping within the clean submap"));
5000 if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
5001 VM_OBJECT_ASSERT_LOCKED(m->object);
5002 KASSERT((flags & PMAP_ENTER_RESERVED) == 0,
5003 ("pmap_enter: flags %u has reserved bits set", flags));
5004 pa = VM_PAGE_TO_PHYS(m);
5005 newpte = (pt_entry_t)(pa | PG_A | PG_V);
5006 if ((flags & VM_PROT_WRITE) != 0)
5008 if ((prot & VM_PROT_WRITE) != 0)
5010 KASSERT((newpte & (PG_M | PG_RW)) != PG_M,
5011 ("pmap_enter: flags includes VM_PROT_WRITE but prot doesn't"));
5012 if ((prot & VM_PROT_EXECUTE) == 0)
5014 if ((flags & PMAP_ENTER_WIRED) != 0)
5016 if (va < VM_MAXUSER_ADDRESS)
5018 if (pmap == kernel_pmap)
5020 newpte |= pmap_cache_bits(pmap, m->md.pat_mode, psind > 0);
5023 * Set modified bit gratuitously for writeable mappings if
5024 * the page is unmanaged. We do not want to take a fault
5025 * to do the dirty bit accounting for these mappings.
5027 if ((m->oflags & VPO_UNMANAGED) != 0) {
5028 if ((newpte & PG_RW) != 0)
5031 newpte |= PG_MANAGED;
5036 /* Assert the required virtual and physical alignment. */
5037 KASSERT((va & PDRMASK) == 0, ("pmap_enter: va unaligned"));
5038 KASSERT(m->psind > 0, ("pmap_enter: m->psind < psind"));
5039 rv = pmap_enter_pde(pmap, va, newpte | PG_PS, flags, m, &lock);
5045 * In the case that a page table page is not
5046 * resident, we are creating it here.
5049 pde = pmap_pde(pmap, va);
5050 if (pde != NULL && (*pde & PG_V) != 0 && ((*pde & PG_PS) == 0 ||
5051 pmap_demote_pde_locked(pmap, pde, va, &lock))) {
5052 pte = pmap_pde_to_pte(pde, va);
5053 if (va < VM_MAXUSER_ADDRESS && mpte == NULL) {
5054 mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
5057 } else if (va < VM_MAXUSER_ADDRESS) {
5059 * Here if the pte page isn't mapped, or if it has been
5062 nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0;
5063 mpte = _pmap_allocpte(pmap, pmap_pde_pindex(va),
5064 nosleep ? NULL : &lock);
5065 if (mpte == NULL && nosleep) {
5066 rv = KERN_RESOURCE_SHORTAGE;
5071 panic("pmap_enter: invalid page directory va=%#lx", va);
5077 * Is the specified virtual address already mapped?
5079 if ((origpte & PG_V) != 0) {
5081 * Wiring change, just update stats. We don't worry about
5082 * wiring PT pages as they remain resident as long as there
5083 * are valid mappings in them. Hence, if a user page is wired,
5084 * the PT page will be also.
5086 if ((newpte & PG_W) != 0 && (origpte & PG_W) == 0)
5087 pmap->pm_stats.wired_count++;
5088 else if ((newpte & PG_W) == 0 && (origpte & PG_W) != 0)
5089 pmap->pm_stats.wired_count--;
5092 * Remove the extra PT page reference.
5096 KASSERT(mpte->wire_count > 0,
5097 ("pmap_enter: missing reference to page table page,"
5102 * Has the physical page changed?
5104 opa = origpte & PG_FRAME;
5107 * No, might be a protection or wiring change.
5109 if ((origpte & PG_MANAGED) != 0 &&
5110 (newpte & PG_RW) != 0)
5111 vm_page_aflag_set(m, PGA_WRITEABLE);
5112 if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0)
5118 * The physical page has changed. Temporarily invalidate
5119 * the mapping. This ensures that all threads sharing the
5120 * pmap keep a consistent view of the mapping, which is
5121 * necessary for the correct handling of COW faults. It
5122 * also permits reuse of the old mapping's PV entry,
5123 * avoiding an allocation.
5125 * For consistency, handle unmanaged mappings the same way.
5127 origpte = pte_load_clear(pte);
5128 KASSERT((origpte & PG_FRAME) == opa,
5129 ("pmap_enter: unexpected pa update for %#lx", va));
5130 if ((origpte & PG_MANAGED) != 0) {
5131 om = PHYS_TO_VM_PAGE(opa);
5134 * The pmap lock is sufficient to synchronize with
5135 * concurrent calls to pmap_page_test_mappings() and
5136 * pmap_ts_referenced().
5138 if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5140 if ((origpte & PG_A) != 0)
5141 vm_page_aflag_set(om, PGA_REFERENCED);
5142 CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, opa);
5143 pv = pmap_pvh_remove(&om->md, pmap, va);
5144 if ((newpte & PG_MANAGED) == 0)
5145 free_pv_entry(pmap, pv);
5146 if ((om->aflags & PGA_WRITEABLE) != 0 &&
5147 TAILQ_EMPTY(&om->md.pv_list) &&
5148 ((om->flags & PG_FICTITIOUS) != 0 ||
5149 TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list)))
5150 vm_page_aflag_clear(om, PGA_WRITEABLE);
5152 if ((origpte & PG_A) != 0)
5153 pmap_invalidate_page(pmap, va);
5157 * Increment the counters.
5159 if ((newpte & PG_W) != 0)
5160 pmap->pm_stats.wired_count++;
5161 pmap_resident_count_inc(pmap, 1);
5165 * Enter on the PV list if part of our managed memory.
5167 if ((newpte & PG_MANAGED) != 0) {
5169 pv = get_pv_entry(pmap, &lock);
5172 CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, pa);
5173 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
5175 if ((newpte & PG_RW) != 0)
5176 vm_page_aflag_set(m, PGA_WRITEABLE);
5182 if ((origpte & PG_V) != 0) {
5184 origpte = pte_load_store(pte, newpte);
5185 KASSERT((origpte & PG_FRAME) == pa,
5186 ("pmap_enter: unexpected pa update for %#lx", va));
5187 if ((newpte & PG_M) == 0 && (origpte & (PG_M | PG_RW)) ==
5189 if ((origpte & PG_MANAGED) != 0)
5193 * Although the PTE may still have PG_RW set, TLB
5194 * invalidation may nonetheless be required because
5195 * the PTE no longer has PG_M set.
5197 } else if ((origpte & PG_NX) != 0 || (newpte & PG_NX) == 0) {
5199 * This PTE change does not require TLB invalidation.
5203 if ((origpte & PG_A) != 0)
5204 pmap_invalidate_page(pmap, va);
5206 pte_store(pte, newpte);
5210 #if VM_NRESERVLEVEL > 0
5212 * If both the page table page and the reservation are fully
5213 * populated, then attempt promotion.
5215 if ((mpte == NULL || mpte->wire_count == NPTEPG) &&
5216 pmap_ps_enabled(pmap) &&
5217 (m->flags & PG_FICTITIOUS) == 0 &&
5218 vm_reserv_level_iffullpop(m) == 0)
5219 pmap_promote_pde(pmap, pde, va, &lock);
5231 * Tries to create a read- and/or execute-only 2MB page mapping. Returns true
5232 * if successful. Returns false if (1) a page table page cannot be allocated
5233 * without sleeping, (2) a mapping already exists at the specified virtual
5234 * address, or (3) a PV entry cannot be allocated without reclaiming another
5238 pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
5239 struct rwlock **lockp)
5244 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5245 PG_V = pmap_valid_bit(pmap);
5246 newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(pmap, m->md.pat_mode, 1) |
5248 if ((m->oflags & VPO_UNMANAGED) == 0)
5249 newpde |= PG_MANAGED;
5250 if ((prot & VM_PROT_EXECUTE) == 0)
5252 if (va < VM_MAXUSER_ADDRESS)
5254 return (pmap_enter_pde(pmap, va, newpde, PMAP_ENTER_NOSLEEP |
5255 PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, NULL, lockp) ==
5260 * Tries to create the specified 2MB page mapping. Returns KERN_SUCCESS if
5261 * the mapping was created, and either KERN_FAILURE or KERN_RESOURCE_SHORTAGE
5262 * otherwise. Returns KERN_FAILURE if PMAP_ENTER_NOREPLACE was specified and
5263 * a mapping already exists at the specified virtual address. Returns
5264 * KERN_RESOURCE_SHORTAGE if PMAP_ENTER_NOSLEEP was specified and a page table
5265 * page allocation failed. Returns KERN_RESOURCE_SHORTAGE if
5266 * PMAP_ENTER_NORECLAIM was specified and a PV entry allocation failed.
5268 * The parameter "m" is only used when creating a managed, writeable mapping.
5271 pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags,
5272 vm_page_t m, struct rwlock **lockp)
5274 struct spglist free;
5275 pd_entry_t oldpde, *pde;
5276 pt_entry_t PG_G, PG_RW, PG_V;
5279 PG_G = pmap_global_bit(pmap);
5280 PG_RW = pmap_rw_bit(pmap);
5281 KASSERT((newpde & (pmap_modified_bit(pmap) | PG_RW)) != PG_RW,
5282 ("pmap_enter_pde: newpde is missing PG_M"));
5283 PG_V = pmap_valid_bit(pmap);
5284 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5286 if ((pdpg = pmap_allocpde(pmap, va, (flags & PMAP_ENTER_NOSLEEP) != 0 ?
5287 NULL : lockp)) == NULL) {
5288 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
5289 " in pmap %p", va, pmap);
5290 return (KERN_RESOURCE_SHORTAGE);
5292 pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pdpg));
5293 pde = &pde[pmap_pde_index(va)];
5295 if ((oldpde & PG_V) != 0) {
5296 KASSERT(pdpg->wire_count > 1,
5297 ("pmap_enter_pde: pdpg's wire count is too low"));
5298 if ((flags & PMAP_ENTER_NOREPLACE) != 0) {
5300 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
5301 " in pmap %p", va, pmap);
5302 return (KERN_FAILURE);
5304 /* Break the existing mapping(s). */
5306 if ((oldpde & PG_PS) != 0) {
5308 * The reference to the PD page that was acquired by
5309 * pmap_allocpde() ensures that it won't be freed.
5310 * However, if the PDE resulted from a promotion, then
5311 * a reserved PT page could be freed.
5313 (void)pmap_remove_pde(pmap, pde, va, &free, lockp);
5314 if ((oldpde & PG_G) == 0)
5315 pmap_invalidate_pde_page(pmap, va, oldpde);
5317 pmap_delayed_invl_started();
5318 if (pmap_remove_ptes(pmap, va, va + NBPDR, pde, &free,
5320 pmap_invalidate_all(pmap);
5321 pmap_delayed_invl_finished();
5323 vm_page_free_pages_toq(&free, true);
5324 if (va >= VM_MAXUSER_ADDRESS) {
5325 mt = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
5326 if (pmap_insert_pt_page(pmap, mt)) {
5328 * XXX Currently, this can't happen because
5329 * we do not perform pmap_enter(psind == 1)
5330 * on the kernel pmap.
5332 panic("pmap_enter_pde: trie insert failed");
5335 KASSERT(*pde == 0, ("pmap_enter_pde: non-zero pde %p",
5338 if ((newpde & PG_MANAGED) != 0) {
5340 * Abort this mapping if its PV entry could not be created.
5342 if (!pmap_pv_insert_pde(pmap, va, newpde, flags, lockp)) {
5344 if (pmap_unwire_ptp(pmap, va, pdpg, &free)) {
5346 * Although "va" is not mapped, paging-
5347 * structure caches could nonetheless have
5348 * entries that refer to the freed page table
5349 * pages. Invalidate those entries.
5351 pmap_invalidate_page(pmap, va);
5352 vm_page_free_pages_toq(&free, true);
5354 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
5355 " in pmap %p", va, pmap);
5356 return (KERN_RESOURCE_SHORTAGE);
5358 if ((newpde & PG_RW) != 0) {
5359 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
5360 vm_page_aflag_set(mt, PGA_WRITEABLE);
5365 * Increment counters.
5367 if ((newpde & PG_W) != 0)
5368 pmap->pm_stats.wired_count += NBPDR / PAGE_SIZE;
5369 pmap_resident_count_inc(pmap, NBPDR / PAGE_SIZE);
5372 * Map the superpage. (This is not a promoted mapping; there will not
5373 * be any lingering 4KB page mappings in the TLB.)
5375 pde_store(pde, newpde);
5377 atomic_add_long(&pmap_pde_mappings, 1);
5378 CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx"
5379 " in pmap %p", va, pmap);
5380 return (KERN_SUCCESS);
5384 * Maps a sequence of resident pages belonging to the same object.
5385 * The sequence begins with the given page m_start. This page is
5386 * mapped at the given virtual address start. Each subsequent page is
5387 * mapped at a virtual address that is offset from start by the same
5388 * amount as the page is offset from m_start within the object. The
5389 * last page in the sequence is the page with the largest offset from
5390 * m_start that can be mapped at a virtual address less than the given
5391 * virtual address end. Not every virtual page between start and end
5392 * is mapped; only those for which a resident page exists with the
5393 * corresponding offset from m_start are mapped.
5396 pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end,
5397 vm_page_t m_start, vm_prot_t prot)
5399 struct rwlock *lock;
5402 vm_pindex_t diff, psize;
5404 VM_OBJECT_ASSERT_LOCKED(m_start->object);
5406 psize = atop(end - start);
5411 while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
5412 va = start + ptoa(diff);
5413 if ((va & PDRMASK) == 0 && va + NBPDR <= end &&
5414 m->psind == 1 && pmap_ps_enabled(pmap) &&
5415 pmap_enter_2mpage(pmap, va, m, prot, &lock))
5416 m = &m[NBPDR / PAGE_SIZE - 1];
5418 mpte = pmap_enter_quick_locked(pmap, va, m, prot,
5420 m = TAILQ_NEXT(m, listq);
5428 * this code makes some *MAJOR* assumptions:
5429 * 1. Current pmap & pmap exists.
5432 * 4. No page table pages.
5433 * but is *MUCH* faster than pmap_enter...
5437 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot)
5439 struct rwlock *lock;
5443 (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL, &lock);
5450 pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m,
5451 vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp)
5453 struct spglist free;
5454 pt_entry_t *pte, PG_V;
5457 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva ||
5458 (m->oflags & VPO_UNMANAGED) != 0,
5459 ("pmap_enter_quick_locked: managed mapping within the clean submap"));
5460 PG_V = pmap_valid_bit(pmap);
5461 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5464 * In the case that a page table page is not
5465 * resident, we are creating it here.
5467 if (va < VM_MAXUSER_ADDRESS) {
5468 vm_pindex_t ptepindex;
5472 * Calculate pagetable page index
5474 ptepindex = pmap_pde_pindex(va);
5475 if (mpte && (mpte->pindex == ptepindex)) {
5479 * Get the page directory entry
5481 ptepa = pmap_pde(pmap, va);
5484 * If the page table page is mapped, we just increment
5485 * the hold count, and activate it. Otherwise, we
5486 * attempt to allocate a page table page. If this
5487 * attempt fails, we don't retry. Instead, we give up.
5489 if (ptepa && (*ptepa & PG_V) != 0) {
5492 mpte = PHYS_TO_VM_PAGE(*ptepa & PG_FRAME);
5496 * Pass NULL instead of the PV list lock
5497 * pointer, because we don't intend to sleep.
5499 mpte = _pmap_allocpte(pmap, ptepindex, NULL);
5504 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte));
5505 pte = &pte[pmap_pte_index(va)];
5519 * Enter on the PV list if part of our managed memory.
5521 if ((m->oflags & VPO_UNMANAGED) == 0 &&
5522 !pmap_try_insert_pv_entry(pmap, va, m, lockp)) {
5525 if (pmap_unwire_ptp(pmap, va, mpte, &free)) {
5527 * Although "va" is not mapped, paging-
5528 * structure caches could nonetheless have
5529 * entries that refer to the freed page table
5530 * pages. Invalidate those entries.
5532 pmap_invalidate_page(pmap, va);
5533 vm_page_free_pages_toq(&free, true);
5541 * Increment counters
5543 pmap_resident_count_inc(pmap, 1);
5545 pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(pmap, m->md.pat_mode, 0);
5546 if ((prot & VM_PROT_EXECUTE) == 0)
5550 * Now validate mapping with RO protection
5552 if ((m->oflags & VPO_UNMANAGED) != 0)
5553 pte_store(pte, pa | PG_V | PG_U);
5555 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED);
5560 * Make a temporary mapping for a physical address. This is only intended
5561 * to be used for panic dumps.
5564 pmap_kenter_temporary(vm_paddr_t pa, int i)
5568 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
5569 pmap_kenter(va, pa);
5571 return ((void *)crashdumpmap);
5575 * This code maps large physical mmap regions into the
5576 * processor address space. Note that some shortcuts
5577 * are taken, but the code works.
5580 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object,
5581 vm_pindex_t pindex, vm_size_t size)
5584 pt_entry_t PG_A, PG_M, PG_RW, PG_V;
5585 vm_paddr_t pa, ptepa;
5589 PG_A = pmap_accessed_bit(pmap);
5590 PG_M = pmap_modified_bit(pmap);
5591 PG_V = pmap_valid_bit(pmap);
5592 PG_RW = pmap_rw_bit(pmap);
5594 VM_OBJECT_ASSERT_WLOCKED(object);
5595 KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG,
5596 ("pmap_object_init_pt: non-device object"));
5597 if ((addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) {
5598 if (!pmap_ps_enabled(pmap))
5600 if (!vm_object_populate(object, pindex, pindex + atop(size)))
5602 p = vm_page_lookup(object, pindex);
5603 KASSERT(p->valid == VM_PAGE_BITS_ALL,
5604 ("pmap_object_init_pt: invalid page %p", p));
5605 pat_mode = p->md.pat_mode;
5608 * Abort the mapping if the first page is not physically
5609 * aligned to a 2MB page boundary.
5611 ptepa = VM_PAGE_TO_PHYS(p);
5612 if (ptepa & (NBPDR - 1))
5616 * Skip the first page. Abort the mapping if the rest of
5617 * the pages are not physically contiguous or have differing
5618 * memory attributes.
5620 p = TAILQ_NEXT(p, listq);
5621 for (pa = ptepa + PAGE_SIZE; pa < ptepa + size;
5623 KASSERT(p->valid == VM_PAGE_BITS_ALL,
5624 ("pmap_object_init_pt: invalid page %p", p));
5625 if (pa != VM_PAGE_TO_PHYS(p) ||
5626 pat_mode != p->md.pat_mode)
5628 p = TAILQ_NEXT(p, listq);
5632 * Map using 2MB pages. Since "ptepa" is 2M aligned and
5633 * "size" is a multiple of 2M, adding the PAT setting to "pa"
5634 * will not affect the termination of this loop.
5637 for (pa = ptepa | pmap_cache_bits(pmap, pat_mode, 1);
5638 pa < ptepa + size; pa += NBPDR) {
5639 pdpg = pmap_allocpde(pmap, addr, NULL);
5642 * The creation of mappings below is only an
5643 * optimization. If a page directory page
5644 * cannot be allocated without blocking,
5645 * continue on to the next mapping rather than
5651 pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pdpg));
5652 pde = &pde[pmap_pde_index(addr)];
5653 if ((*pde & PG_V) == 0) {
5654 pde_store(pde, pa | PG_PS | PG_M | PG_A |
5655 PG_U | PG_RW | PG_V);
5656 pmap_resident_count_inc(pmap, NBPDR / PAGE_SIZE);
5657 atomic_add_long(&pmap_pde_mappings, 1);
5659 /* Continue on if the PDE is already valid. */
5661 KASSERT(pdpg->wire_count > 0,
5662 ("pmap_object_init_pt: missing reference "
5663 "to page directory page, va: 0x%lx", addr));
5672 * Clear the wired attribute from the mappings for the specified range of
5673 * addresses in the given pmap. Every valid mapping within that range
5674 * must have the wired attribute set. In contrast, invalid mappings
5675 * cannot have the wired attribute set, so they are ignored.
5677 * The wired attribute of the page table entry is not a hardware
5678 * feature, so there is no need to invalidate any TLB entries.
5679 * Since pmap_demote_pde() for the wired entry must never fail,
5680 * pmap_delayed_invl_started()/finished() calls around the
5681 * function are not needed.
5684 pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5686 vm_offset_t va_next;
5687 pml4_entry_t *pml4e;
5690 pt_entry_t *pte, PG_V;
5692 PG_V = pmap_valid_bit(pmap);
5694 for (; sva < eva; sva = va_next) {
5695 pml4e = pmap_pml4e(pmap, sva);
5696 if ((*pml4e & PG_V) == 0) {
5697 va_next = (sva + NBPML4) & ~PML4MASK;
5702 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
5703 if ((*pdpe & PG_V) == 0) {
5704 va_next = (sva + NBPDP) & ~PDPMASK;
5709 va_next = (sva + NBPDR) & ~PDRMASK;
5712 pde = pmap_pdpe_to_pde(pdpe, sva);
5713 if ((*pde & PG_V) == 0)
5715 if ((*pde & PG_PS) != 0) {
5716 if ((*pde & PG_W) == 0)
5717 panic("pmap_unwire: pde %#jx is missing PG_W",
5721 * Are we unwiring the entire large page? If not,
5722 * demote the mapping and fall through.
5724 if (sva + NBPDR == va_next && eva >= va_next) {
5725 atomic_clear_long(pde, PG_W);
5726 pmap->pm_stats.wired_count -= NBPDR /
5729 } else if (!pmap_demote_pde(pmap, pde, sva))
5730 panic("pmap_unwire: demotion failed");
5734 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
5736 if ((*pte & PG_V) == 0)
5738 if ((*pte & PG_W) == 0)
5739 panic("pmap_unwire: pte %#jx is missing PG_W",
5743 * PG_W must be cleared atomically. Although the pmap
5744 * lock synchronizes access to PG_W, another processor
5745 * could be setting PG_M and/or PG_A concurrently.
5747 atomic_clear_long(pte, PG_W);
5748 pmap->pm_stats.wired_count--;
5755 * Copy the range specified by src_addr/len
5756 * from the source map to the range dst_addr/len
5757 * in the destination map.
5759 * This routine is only advisory and need not do anything.
5763 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
5764 vm_offset_t src_addr)
5766 struct rwlock *lock;
5767 struct spglist free;
5769 vm_offset_t end_addr = src_addr + len;
5770 vm_offset_t va_next;
5771 vm_page_t dst_pdpg, dstmpte, srcmpte;
5772 pt_entry_t PG_A, PG_M, PG_V;
5774 if (dst_addr != src_addr)
5777 if (dst_pmap->pm_type != src_pmap->pm_type)
5781 * EPT page table entries that require emulation of A/D bits are
5782 * sensitive to clearing the PG_A bit (aka EPT_PG_READ). Although
5783 * we clear PG_M (aka EPT_PG_WRITE) concomitantly, the PG_U bit
5784 * (aka EPT_PG_EXECUTE) could still be set. Since some EPT
5785 * implementations flag an EPT misconfiguration for exec-only
5786 * mappings we skip this function entirely for emulated pmaps.
5788 if (pmap_emulate_ad_bits(dst_pmap))
5792 if (dst_pmap < src_pmap) {
5793 PMAP_LOCK(dst_pmap);
5794 PMAP_LOCK(src_pmap);
5796 PMAP_LOCK(src_pmap);
5797 PMAP_LOCK(dst_pmap);
5800 PG_A = pmap_accessed_bit(dst_pmap);
5801 PG_M = pmap_modified_bit(dst_pmap);
5802 PG_V = pmap_valid_bit(dst_pmap);
5804 for (addr = src_addr; addr < end_addr; addr = va_next) {
5805 pt_entry_t *src_pte, *dst_pte;
5806 pml4_entry_t *pml4e;
5808 pd_entry_t srcptepaddr, *pde;
5810 KASSERT(addr < UPT_MIN_ADDRESS,
5811 ("pmap_copy: invalid to pmap_copy page tables"));
5813 pml4e = pmap_pml4e(src_pmap, addr);
5814 if ((*pml4e & PG_V) == 0) {
5815 va_next = (addr + NBPML4) & ~PML4MASK;
5821 pdpe = pmap_pml4e_to_pdpe(pml4e, addr);
5822 if ((*pdpe & PG_V) == 0) {
5823 va_next = (addr + NBPDP) & ~PDPMASK;
5829 va_next = (addr + NBPDR) & ~PDRMASK;
5833 pde = pmap_pdpe_to_pde(pdpe, addr);
5835 if (srcptepaddr == 0)
5838 if (srcptepaddr & PG_PS) {
5839 if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr)
5841 dst_pdpg = pmap_allocpde(dst_pmap, addr, NULL);
5842 if (dst_pdpg == NULL)
5844 pde = (pd_entry_t *)
5845 PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dst_pdpg));
5846 pde = &pde[pmap_pde_index(addr)];
5847 if (*pde == 0 && ((srcptepaddr & PG_MANAGED) == 0 ||
5848 pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr,
5849 PMAP_ENTER_NORECLAIM, &lock))) {
5850 *pde = srcptepaddr & ~PG_W;
5851 pmap_resident_count_inc(dst_pmap, NBPDR / PAGE_SIZE);
5852 atomic_add_long(&pmap_pde_mappings, 1);
5854 dst_pdpg->wire_count--;
5858 srcptepaddr &= PG_FRAME;
5859 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr);
5860 KASSERT(srcmpte->wire_count > 0,
5861 ("pmap_copy: source page table page is unused"));
5863 if (va_next > end_addr)
5866 src_pte = (pt_entry_t *)PHYS_TO_DMAP(srcptepaddr);
5867 src_pte = &src_pte[pmap_pte_index(addr)];
5869 while (addr < va_next) {
5873 * we only virtual copy managed pages
5875 if ((ptetemp & PG_MANAGED) != 0) {
5876 if (dstmpte != NULL &&
5877 dstmpte->pindex == pmap_pde_pindex(addr))
5878 dstmpte->wire_count++;
5879 else if ((dstmpte = pmap_allocpte(dst_pmap,
5880 addr, NULL)) == NULL)
5882 dst_pte = (pt_entry_t *)
5883 PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dstmpte));
5884 dst_pte = &dst_pte[pmap_pte_index(addr)];
5885 if (*dst_pte == 0 &&
5886 pmap_try_insert_pv_entry(dst_pmap, addr,
5887 PHYS_TO_VM_PAGE(ptetemp & PG_FRAME),
5890 * Clear the wired, modified, and
5891 * accessed (referenced) bits
5894 *dst_pte = ptetemp & ~(PG_W | PG_M |
5896 pmap_resident_count_inc(dst_pmap, 1);
5899 if (pmap_unwire_ptp(dst_pmap, addr,
5902 * Although "addr" is not
5903 * mapped, paging-structure
5904 * caches could nonetheless
5905 * have entries that refer to
5906 * the freed page table pages.
5907 * Invalidate those entries.
5909 pmap_invalidate_page(dst_pmap,
5911 vm_page_free_pages_toq(&free,
5916 if (dstmpte->wire_count >= srcmpte->wire_count)
5926 PMAP_UNLOCK(src_pmap);
5927 PMAP_UNLOCK(dst_pmap);
5931 * Zero the specified hardware page.
5934 pmap_zero_page(vm_page_t m)
5936 vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
5938 pagezero((void *)va);
5942 * Zero an an area within a single hardware page. off and size must not
5943 * cover an area beyond a single hardware page.
5946 pmap_zero_page_area(vm_page_t m, int off, int size)
5948 vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
5950 if (off == 0 && size == PAGE_SIZE)
5951 pagezero((void *)va);
5953 bzero((char *)va + off, size);
5957 * Copy 1 specified hardware page to another.
5960 pmap_copy_page(vm_page_t msrc, vm_page_t mdst)
5962 vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc));
5963 vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst));
5965 pagecopy((void *)src, (void *)dst);
5968 int unmapped_buf_allowed = 1;
5971 pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[],
5972 vm_offset_t b_offset, int xfersize)
5976 vm_offset_t vaddr[2], a_pg_offset, b_pg_offset;
5980 while (xfersize > 0) {
5981 a_pg_offset = a_offset & PAGE_MASK;
5982 pages[0] = ma[a_offset >> PAGE_SHIFT];
5983 b_pg_offset = b_offset & PAGE_MASK;
5984 pages[1] = mb[b_offset >> PAGE_SHIFT];
5985 cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
5986 cnt = min(cnt, PAGE_SIZE - b_pg_offset);
5987 mapped = pmap_map_io_transient(pages, vaddr, 2, FALSE);
5988 a_cp = (char *)vaddr[0] + a_pg_offset;
5989 b_cp = (char *)vaddr[1] + b_pg_offset;
5990 bcopy(a_cp, b_cp, cnt);
5991 if (__predict_false(mapped))
5992 pmap_unmap_io_transient(pages, vaddr, 2, FALSE);
6000 * Returns true if the pmap's pv is one of the first
6001 * 16 pvs linked to from this page. This count may
6002 * be changed upwards or downwards in the future; it
6003 * is only necessary that true be returned for a small
6004 * subset of pmaps for proper page aging.
6007 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
6009 struct md_page *pvh;
6010 struct rwlock *lock;
6015 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
6016 ("pmap_page_exists_quick: page %p is not managed", m));
6018 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
6020 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
6021 if (PV_PMAP(pv) == pmap) {
6029 if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) {
6030 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
6031 TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
6032 if (PV_PMAP(pv) == pmap) {
6046 * pmap_page_wired_mappings:
6048 * Return the number of managed mappings to the given physical page
6052 pmap_page_wired_mappings(vm_page_t m)
6054 struct rwlock *lock;
6055 struct md_page *pvh;
6059 int count, md_gen, pvh_gen;
6061 if ((m->oflags & VPO_UNMANAGED) != 0)
6063 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
6067 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
6069 if (!PMAP_TRYLOCK(pmap)) {
6070 md_gen = m->md.pv_gen;
6074 if (md_gen != m->md.pv_gen) {
6079 pte = pmap_pte(pmap, pv->pv_va);
6080 if ((*pte & PG_W) != 0)
6084 if ((m->flags & PG_FICTITIOUS) == 0) {
6085 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
6086 TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
6088 if (!PMAP_TRYLOCK(pmap)) {
6089 md_gen = m->md.pv_gen;
6090 pvh_gen = pvh->pv_gen;
6094 if (md_gen != m->md.pv_gen ||
6095 pvh_gen != pvh->pv_gen) {
6100 pte = pmap_pde(pmap, pv->pv_va);
6101 if ((*pte & PG_W) != 0)
6111 * Returns TRUE if the given page is mapped individually or as part of
6112 * a 2mpage. Otherwise, returns FALSE.
6115 pmap_page_is_mapped(vm_page_t m)
6117 struct rwlock *lock;
6120 if ((m->oflags & VPO_UNMANAGED) != 0)
6122 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
6124 rv = !TAILQ_EMPTY(&m->md.pv_list) ||
6125 ((m->flags & PG_FICTITIOUS) == 0 &&
6126 !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list));
6132 * Destroy all managed, non-wired mappings in the given user-space
6133 * pmap. This pmap cannot be active on any processor besides the
6136 * This function cannot be applied to the kernel pmap. Moreover, it
6137 * is not intended for general use. It is only to be used during
6138 * process termination. Consequently, it can be implemented in ways
6139 * that make it faster than pmap_remove(). First, it can more quickly
6140 * destroy mappings by iterating over the pmap's collection of PV
6141 * entries, rather than searching the page table. Second, it doesn't
6142 * have to test and clear the page table entries atomically, because
6143 * no processor is currently accessing the user address space. In
6144 * particular, a page table entry's dirty bit won't change state once
6145 * this function starts.
6147 * Although this function destroys all of the pmap's managed,
6148 * non-wired mappings, it can delay and batch the invalidation of TLB
6149 * entries without calling pmap_delayed_invl_started() and
6150 * pmap_delayed_invl_finished(). Because the pmap is not active on
6151 * any other processor, none of these TLB entries will ever be used
6152 * before their eventual invalidation. Consequently, there is no need
6153 * for either pmap_remove_all() or pmap_remove_write() to wait for
6154 * that eventual TLB invalidation.
6157 pmap_remove_pages(pmap_t pmap)
6160 pt_entry_t *pte, tpte;
6161 pt_entry_t PG_M, PG_RW, PG_V;
6162 struct spglist free;
6163 vm_page_t m, mpte, mt;
6165 struct md_page *pvh;
6166 struct pv_chunk *pc, *npc;
6167 struct rwlock *lock;
6169 uint64_t inuse, bitmask;
6170 int allfree, field, freed, idx;
6171 boolean_t superpage;
6175 * Assert that the given pmap is only active on the current
6176 * CPU. Unfortunately, we cannot block another CPU from
6177 * activating the pmap while this function is executing.
6179 KASSERT(pmap == PCPU_GET(curpmap), ("non-current pmap %p", pmap));
6182 cpuset_t other_cpus;
6184 other_cpus = all_cpus;
6186 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
6187 CPU_AND(&other_cpus, &pmap->pm_active);
6189 KASSERT(CPU_EMPTY(&other_cpus), ("pmap active %p", pmap));
6194 PG_M = pmap_modified_bit(pmap);
6195 PG_V = pmap_valid_bit(pmap);
6196 PG_RW = pmap_rw_bit(pmap);
6200 TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) {
6203 for (field = 0; field < _NPCM; field++) {
6204 inuse = ~pc->pc_map[field] & pc_freemask[field];
6205 while (inuse != 0) {
6207 bitmask = 1UL << bit;
6208 idx = field * 64 + bit;
6209 pv = &pc->pc_pventry[idx];
6212 pte = pmap_pdpe(pmap, pv->pv_va);
6214 pte = pmap_pdpe_to_pde(pte, pv->pv_va);
6216 if ((tpte & (PG_PS | PG_V)) == PG_V) {
6219 pte = (pt_entry_t *)PHYS_TO_DMAP(tpte &
6221 pte = &pte[pmap_pte_index(pv->pv_va)];
6225 * Keep track whether 'tpte' is a
6226 * superpage explicitly instead of
6227 * relying on PG_PS being set.
6229 * This is because PG_PS is numerically
6230 * identical to PG_PTE_PAT and thus a
6231 * regular page could be mistaken for
6237 if ((tpte & PG_V) == 0) {
6238 panic("bad pte va %lx pte %lx",
6243 * We cannot remove wired pages from a process' mapping at this time
6251 pa = tpte & PG_PS_FRAME;
6253 pa = tpte & PG_FRAME;
6255 m = PHYS_TO_VM_PAGE(pa);
6256 KASSERT(m->phys_addr == pa,
6257 ("vm_page_t %p phys_addr mismatch %016jx %016jx",
6258 m, (uintmax_t)m->phys_addr,
6261 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
6262 m < &vm_page_array[vm_page_array_size],
6263 ("pmap_remove_pages: bad tpte %#jx",
6269 * Update the vm_page_t clean/reference bits.
6271 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
6273 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
6279 CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m);
6282 pc->pc_map[field] |= bitmask;
6284 pmap_resident_count_dec(pmap, NBPDR / PAGE_SIZE);
6285 pvh = pa_to_pvh(tpte & PG_PS_FRAME);
6286 TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
6288 if (TAILQ_EMPTY(&pvh->pv_list)) {
6289 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
6290 if ((mt->aflags & PGA_WRITEABLE) != 0 &&
6291 TAILQ_EMPTY(&mt->md.pv_list))
6292 vm_page_aflag_clear(mt, PGA_WRITEABLE);
6294 mpte = pmap_remove_pt_page(pmap, pv->pv_va);
6296 pmap_resident_count_dec(pmap, 1);
6297 KASSERT(mpte->wire_count == NPTEPG,
6298 ("pmap_remove_pages: pte page wire count error"));
6299 mpte->wire_count = 0;
6300 pmap_add_delayed_free_list(mpte, &free, FALSE);
6303 pmap_resident_count_dec(pmap, 1);
6304 TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
6306 if ((m->aflags & PGA_WRITEABLE) != 0 &&
6307 TAILQ_EMPTY(&m->md.pv_list) &&
6308 (m->flags & PG_FICTITIOUS) == 0) {
6309 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
6310 if (TAILQ_EMPTY(&pvh->pv_list))
6311 vm_page_aflag_clear(m, PGA_WRITEABLE);
6314 pmap_unuse_pt(pmap, pv->pv_va, ptepde, &free);
6318 PV_STAT(atomic_add_long(&pv_entry_frees, freed));
6319 PV_STAT(atomic_add_int(&pv_entry_spare, freed));
6320 PV_STAT(atomic_subtract_long(&pv_entry_count, freed));
6322 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
6328 pmap_invalidate_all(pmap);
6330 vm_page_free_pages_toq(&free, true);
6334 pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified)
6336 struct rwlock *lock;
6338 struct md_page *pvh;
6339 pt_entry_t *pte, mask;
6340 pt_entry_t PG_A, PG_M, PG_RW, PG_V;
6342 int md_gen, pvh_gen;
6346 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
6349 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
6351 if (!PMAP_TRYLOCK(pmap)) {
6352 md_gen = m->md.pv_gen;
6356 if (md_gen != m->md.pv_gen) {
6361 pte = pmap_pte(pmap, pv->pv_va);
6364 PG_M = pmap_modified_bit(pmap);
6365 PG_RW = pmap_rw_bit(pmap);
6366 mask |= PG_RW | PG_M;
6369 PG_A = pmap_accessed_bit(pmap);
6370 PG_V = pmap_valid_bit(pmap);
6371 mask |= PG_V | PG_A;
6373 rv = (*pte & mask) == mask;
6378 if ((m->flags & PG_FICTITIOUS) == 0) {
6379 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
6380 TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
6382 if (!PMAP_TRYLOCK(pmap)) {
6383 md_gen = m->md.pv_gen;
6384 pvh_gen = pvh->pv_gen;
6388 if (md_gen != m->md.pv_gen ||
6389 pvh_gen != pvh->pv_gen) {
6394 pte = pmap_pde(pmap, pv->pv_va);
6397 PG_M = pmap_modified_bit(pmap);
6398 PG_RW = pmap_rw_bit(pmap);
6399 mask |= PG_RW | PG_M;
6402 PG_A = pmap_accessed_bit(pmap);
6403 PG_V = pmap_valid_bit(pmap);
6404 mask |= PG_V | PG_A;
6406 rv = (*pte & mask) == mask;
6420 * Return whether or not the specified physical page was modified
6421 * in any physical maps.
6424 pmap_is_modified(vm_page_t m)
6427 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
6428 ("pmap_is_modified: page %p is not managed", m));
6431 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
6432 * concurrently set while the object is locked. Thus, if PGA_WRITEABLE
6433 * is clear, no PTEs can have PG_M set.
6435 VM_OBJECT_ASSERT_WLOCKED(m->object);
6436 if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
6438 return (pmap_page_test_mappings(m, FALSE, TRUE));
6442 * pmap_is_prefaultable:
6444 * Return whether or not the specified virtual address is eligible
6448 pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr)
6451 pt_entry_t *pte, PG_V;
6454 PG_V = pmap_valid_bit(pmap);
6457 pde = pmap_pde(pmap, addr);
6458 if (pde != NULL && (*pde & (PG_PS | PG_V)) == PG_V) {
6459 pte = pmap_pde_to_pte(pde, addr);
6460 rv = (*pte & PG_V) == 0;
6467 * pmap_is_referenced:
6469 * Return whether or not the specified physical page was referenced
6470 * in any physical maps.
6473 pmap_is_referenced(vm_page_t m)
6476 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
6477 ("pmap_is_referenced: page %p is not managed", m));
6478 return (pmap_page_test_mappings(m, TRUE, FALSE));
6482 * Clear the write and modified bits in each of the given page's mappings.
6485 pmap_remove_write(vm_page_t m)
6487 struct md_page *pvh;
6489 struct rwlock *lock;
6490 pv_entry_t next_pv, pv;
6492 pt_entry_t oldpte, *pte, PG_M, PG_RW;
6494 int pvh_gen, md_gen;
6496 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
6497 ("pmap_remove_write: page %p is not managed", m));
6500 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
6501 * set by another thread while the object is locked. Thus,
6502 * if PGA_WRITEABLE is clear, no page table entries need updating.
6504 VM_OBJECT_ASSERT_WLOCKED(m->object);
6505 if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
6507 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
6508 pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
6509 pa_to_pvh(VM_PAGE_TO_PHYS(m));
6512 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) {
6514 if (!PMAP_TRYLOCK(pmap)) {
6515 pvh_gen = pvh->pv_gen;
6519 if (pvh_gen != pvh->pv_gen) {
6525 PG_RW = pmap_rw_bit(pmap);
6527 pde = pmap_pde(pmap, va);
6528 if ((*pde & PG_RW) != 0)
6529 (void)pmap_demote_pde_locked(pmap, pde, va, &lock);
6530 KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m),
6531 ("inconsistent pv lock %p %p for page %p",
6532 lock, VM_PAGE_TO_PV_LIST_LOCK(m), m));
6535 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
6537 if (!PMAP_TRYLOCK(pmap)) {
6538 pvh_gen = pvh->pv_gen;
6539 md_gen = m->md.pv_gen;
6543 if (pvh_gen != pvh->pv_gen ||
6544 md_gen != m->md.pv_gen) {
6550 PG_M = pmap_modified_bit(pmap);
6551 PG_RW = pmap_rw_bit(pmap);
6552 pde = pmap_pde(pmap, pv->pv_va);
6553 KASSERT((*pde & PG_PS) == 0,
6554 ("pmap_remove_write: found a 2mpage in page %p's pv list",
6556 pte = pmap_pde_to_pte(pde, pv->pv_va);
6559 if (oldpte & PG_RW) {
6560 if (!atomic_cmpset_long(pte, oldpte, oldpte &
6563 if ((oldpte & PG_M) != 0)
6565 pmap_invalidate_page(pmap, pv->pv_va);
6570 vm_page_aflag_clear(m, PGA_WRITEABLE);
6571 pmap_delayed_invl_wait(m);
6574 static __inline boolean_t
6575 safe_to_clear_referenced(pmap_t pmap, pt_entry_t pte)
6578 if (!pmap_emulate_ad_bits(pmap))
6581 KASSERT(pmap->pm_type == PT_EPT, ("invalid pm_type %d", pmap->pm_type));
6584 * XWR = 010 or 110 will cause an unconditional EPT misconfiguration
6585 * so we don't let the referenced (aka EPT_PG_READ) bit to be cleared
6586 * if the EPT_PG_WRITE bit is set.
6588 if ((pte & EPT_PG_WRITE) != 0)
6592 * XWR = 100 is allowed only if the PMAP_SUPPORTS_EXEC_ONLY is set.
6594 if ((pte & EPT_PG_EXECUTE) == 0 ||
6595 ((pmap->pm_flags & PMAP_SUPPORTS_EXEC_ONLY) != 0))
6602 * pmap_ts_referenced:
6604 * Return a count of reference bits for a page, clearing those bits.
6605 * It is not necessary for every reference bit to be cleared, but it
6606 * is necessary that 0 only be returned when there are truly no
6607 * reference bits set.
6609 * As an optimization, update the page's dirty field if a modified bit is
6610 * found while counting reference bits. This opportunistic update can be
6611 * performed at low cost and can eliminate the need for some future calls
6612 * to pmap_is_modified(). However, since this function stops after
6613 * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some
6614 * dirty pages. Those dirty pages will only be detected by a future call
6615 * to pmap_is_modified().
6617 * A DI block is not needed within this function, because
6618 * invalidations are performed before the PV list lock is
6622 pmap_ts_referenced(vm_page_t m)
6624 struct md_page *pvh;
6627 struct rwlock *lock;
6628 pd_entry_t oldpde, *pde;
6629 pt_entry_t *pte, PG_A, PG_M, PG_RW;
6632 int cleared, md_gen, not_cleared, pvh_gen;
6633 struct spglist free;
6636 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
6637 ("pmap_ts_referenced: page %p is not managed", m));
6640 pa = VM_PAGE_TO_PHYS(m);
6641 lock = PHYS_TO_PV_LIST_LOCK(pa);
6642 pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(pa);
6646 if ((pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL)
6647 goto small_mappings;
6653 if (!PMAP_TRYLOCK(pmap)) {
6654 pvh_gen = pvh->pv_gen;
6658 if (pvh_gen != pvh->pv_gen) {
6663 PG_A = pmap_accessed_bit(pmap);
6664 PG_M = pmap_modified_bit(pmap);
6665 PG_RW = pmap_rw_bit(pmap);
6667 pde = pmap_pde(pmap, pv->pv_va);
6669 if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
6671 * Although "oldpde" is mapping a 2MB page, because
6672 * this function is called at a 4KB page granularity,
6673 * we only update the 4KB page under test.
6677 if ((oldpde & PG_A) != 0) {
6679 * Since this reference bit is shared by 512 4KB
6680 * pages, it should not be cleared every time it is
6681 * tested. Apply a simple "hash" function on the
6682 * physical page number, the virtual superpage number,
6683 * and the pmap address to select one 4KB page out of
6684 * the 512 on which testing the reference bit will
6685 * result in clearing that reference bit. This
6686 * function is designed to avoid the selection of the
6687 * same 4KB page for every 2MB page mapping.
6689 * On demotion, a mapping that hasn't been referenced
6690 * is simply destroyed. To avoid the possibility of a
6691 * subsequent page fault on a demoted wired mapping,
6692 * always leave its reference bit set. Moreover,
6693 * since the superpage is wired, the current state of
6694 * its reference bit won't affect page replacement.
6696 if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PDRSHIFT) ^
6697 (uintptr_t)pmap) & (NPTEPG - 1)) == 0 &&
6698 (oldpde & PG_W) == 0) {
6699 if (safe_to_clear_referenced(pmap, oldpde)) {
6700 atomic_clear_long(pde, PG_A);
6701 pmap_invalidate_page(pmap, pv->pv_va);
6703 } else if (pmap_demote_pde_locked(pmap, pde,
6704 pv->pv_va, &lock)) {
6706 * Remove the mapping to a single page
6707 * so that a subsequent access may
6708 * repromote. Since the underlying
6709 * page table page is fully populated,
6710 * this removal never frees a page
6714 va += VM_PAGE_TO_PHYS(m) - (oldpde &
6716 pte = pmap_pde_to_pte(pde, va);
6717 pmap_remove_pte(pmap, pte, va, *pde,
6719 pmap_invalidate_page(pmap, va);
6725 * The superpage mapping was removed
6726 * entirely and therefore 'pv' is no
6734 KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m),
6735 ("inconsistent pv lock %p %p for page %p",
6736 lock, VM_PAGE_TO_PV_LIST_LOCK(m), m));
6741 /* Rotate the PV list if it has more than one entry. */
6742 if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) {
6743 TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
6744 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
6747 if (cleared + not_cleared >= PMAP_TS_REFERENCED_MAX)
6749 } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf);
6751 if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL)
6758 if (!PMAP_TRYLOCK(pmap)) {
6759 pvh_gen = pvh->pv_gen;
6760 md_gen = m->md.pv_gen;
6764 if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
6769 PG_A = pmap_accessed_bit(pmap);
6770 PG_M = pmap_modified_bit(pmap);
6771 PG_RW = pmap_rw_bit(pmap);
6772 pde = pmap_pde(pmap, pv->pv_va);
6773 KASSERT((*pde & PG_PS) == 0,
6774 ("pmap_ts_referenced: found a 2mpage in page %p's pv list",
6776 pte = pmap_pde_to_pte(pde, pv->pv_va);
6777 if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW))
6779 if ((*pte & PG_A) != 0) {
6780 if (safe_to_clear_referenced(pmap, *pte)) {
6781 atomic_clear_long(pte, PG_A);
6782 pmap_invalidate_page(pmap, pv->pv_va);
6784 } else if ((*pte & PG_W) == 0) {
6786 * Wired pages cannot be paged out so
6787 * doing accessed bit emulation for
6788 * them is wasted effort. We do the
6789 * hard work for unwired pages only.
6791 pmap_remove_pte(pmap, pte, pv->pv_va,
6792 *pde, &free, &lock);
6793 pmap_invalidate_page(pmap, pv->pv_va);
6798 KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m),
6799 ("inconsistent pv lock %p %p for page %p",
6800 lock, VM_PAGE_TO_PV_LIST_LOCK(m), m));
6805 /* Rotate the PV list if it has more than one entry. */
6806 if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) {
6807 TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
6808 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
6811 } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared +
6812 not_cleared < PMAP_TS_REFERENCED_MAX);
6815 vm_page_free_pages_toq(&free, true);
6816 return (cleared + not_cleared);
6820 * Apply the given advice to the specified range of addresses within the
6821 * given pmap. Depending on the advice, clear the referenced and/or
6822 * modified flags in each mapping and set the mapped page's dirty field.
6825 pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice)
6827 struct rwlock *lock;
6828 pml4_entry_t *pml4e;
6830 pd_entry_t oldpde, *pde;
6831 pt_entry_t *pte, PG_A, PG_G, PG_M, PG_RW, PG_V;
6832 vm_offset_t va, va_next;
6834 boolean_t anychanged;
6836 if (advice != MADV_DONTNEED && advice != MADV_FREE)
6840 * A/D bit emulation requires an alternate code path when clearing
6841 * the modified and accessed bits below. Since this function is
6842 * advisory in nature we skip it entirely for pmaps that require
6843 * A/D bit emulation.
6845 if (pmap_emulate_ad_bits(pmap))
6848 PG_A = pmap_accessed_bit(pmap);
6849 PG_G = pmap_global_bit(pmap);
6850 PG_M = pmap_modified_bit(pmap);
6851 PG_V = pmap_valid_bit(pmap);
6852 PG_RW = pmap_rw_bit(pmap);
6854 pmap_delayed_invl_started();
6856 for (; sva < eva; sva = va_next) {
6857 pml4e = pmap_pml4e(pmap, sva);
6858 if ((*pml4e & PG_V) == 0) {
6859 va_next = (sva + NBPML4) & ~PML4MASK;
6864 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
6865 if ((*pdpe & PG_V) == 0) {
6866 va_next = (sva + NBPDP) & ~PDPMASK;
6871 va_next = (sva + NBPDR) & ~PDRMASK;
6874 pde = pmap_pdpe_to_pde(pdpe, sva);
6876 if ((oldpde & PG_V) == 0)
6878 else if ((oldpde & PG_PS) != 0) {
6879 if ((oldpde & PG_MANAGED) == 0)
6882 if (!pmap_demote_pde_locked(pmap, pde, sva, &lock)) {
6887 * The large page mapping was destroyed.
6893 * Unless the page mappings are wired, remove the
6894 * mapping to a single page so that a subsequent
6895 * access may repromote. Since the underlying page
6896 * table page is fully populated, this removal never
6897 * frees a page table page.
6899 if ((oldpde & PG_W) == 0) {
6900 pte = pmap_pde_to_pte(pde, sva);
6901 KASSERT((*pte & PG_V) != 0,
6902 ("pmap_advise: invalid PTE"));
6903 pmap_remove_pte(pmap, pte, sva, *pde, NULL,
6913 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
6915 if ((*pte & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V))
6917 else if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
6918 if (advice == MADV_DONTNEED) {
6920 * Future calls to pmap_is_modified()
6921 * can be avoided by making the page
6924 m = PHYS_TO_VM_PAGE(*pte & PG_FRAME);
6927 atomic_clear_long(pte, PG_M | PG_A);
6928 } else if ((*pte & PG_A) != 0)
6929 atomic_clear_long(pte, PG_A);
6933 if ((*pte & PG_G) != 0) {
6940 if (va != va_next) {
6941 pmap_invalidate_range(pmap, va, sva);
6946 pmap_invalidate_range(pmap, va, sva);
6949 pmap_invalidate_all(pmap);
6951 pmap_delayed_invl_finished();
6955 * Clear the modify bits on the specified physical page.
6958 pmap_clear_modify(vm_page_t m)
6960 struct md_page *pvh;
6962 pv_entry_t next_pv, pv;
6963 pd_entry_t oldpde, *pde;
6964 pt_entry_t oldpte, *pte, PG_M, PG_RW, PG_V;
6965 struct rwlock *lock;
6967 int md_gen, pvh_gen;
6969 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
6970 ("pmap_clear_modify: page %p is not managed", m));
6971 VM_OBJECT_ASSERT_WLOCKED(m->object);
6972 KASSERT(!vm_page_xbusied(m),
6973 ("pmap_clear_modify: page %p is exclusive busied", m));
6976 * If the page is not PGA_WRITEABLE, then no PTEs can have PG_M set.
6977 * If the object containing the page is locked and the page is not
6978 * exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
6980 if ((m->aflags & PGA_WRITEABLE) == 0)
6982 pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
6983 pa_to_pvh(VM_PAGE_TO_PHYS(m));
6984 lock = VM_PAGE_TO_PV_LIST_LOCK(m);
6987 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) {
6989 if (!PMAP_TRYLOCK(pmap)) {
6990 pvh_gen = pvh->pv_gen;
6994 if (pvh_gen != pvh->pv_gen) {
6999 PG_M = pmap_modified_bit(pmap);
7000 PG_V = pmap_valid_bit(pmap);
7001 PG_RW = pmap_rw_bit(pmap);
7003 pde = pmap_pde(pmap, va);
7005 if ((oldpde & PG_RW) != 0) {
7006 if (pmap_demote_pde_locked(pmap, pde, va, &lock)) {
7007 if ((oldpde & PG_W) == 0) {
7009 * Write protect the mapping to a
7010 * single page so that a subsequent
7011 * write access may repromote.
7013 va += VM_PAGE_TO_PHYS(m) - (oldpde &
7015 pte = pmap_pde_to_pte(pde, va);
7017 if ((oldpte & PG_V) != 0) {
7018 while (!atomic_cmpset_long(pte,
7020 oldpte & ~(PG_M | PG_RW)))
7023 pmap_invalidate_page(pmap, va);
7030 TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
7032 if (!PMAP_TRYLOCK(pmap)) {
7033 md_gen = m->md.pv_gen;
7034 pvh_gen = pvh->pv_gen;
7038 if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
7043 PG_M = pmap_modified_bit(pmap);
7044 PG_RW = pmap_rw_bit(pmap);
7045 pde = pmap_pde(pmap, pv->pv_va);
7046 KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found"
7047 " a 2mpage in page %p's pv list", m));
7048 pte = pmap_pde_to_pte(pde, pv->pv_va);
7049 if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
7050 atomic_clear_long(pte, PG_M);
7051 pmap_invalidate_page(pmap, pv->pv_va);
7059 * Miscellaneous support routines follow
7062 /* Adjust the cache mode for a 4KB page mapped via a PTE. */
7063 static __inline void
7064 pmap_pte_attr(pt_entry_t *pte, int cache_bits, int mask)
7069 * The cache mode bits are all in the low 32-bits of the
7070 * PTE, so we can just spin on updating the low 32-bits.
7073 opte = *(u_int *)pte;
7074 npte = opte & ~mask;
7076 } while (npte != opte && !atomic_cmpset_int((u_int *)pte, opte, npte));
7079 /* Adjust the cache mode for a 2MB page mapped via a PDE. */
7080 static __inline void
7081 pmap_pde_attr(pd_entry_t *pde, int cache_bits, int mask)
7086 * The cache mode bits are all in the low 32-bits of the
7087 * PDE, so we can just spin on updating the low 32-bits.
7090 opde = *(u_int *)pde;
7091 npde = opde & ~mask;
7093 } while (npde != opde && !atomic_cmpset_int((u_int *)pde, opde, npde));
7097 * Map a set of physical memory pages into the kernel virtual
7098 * address space. Return a pointer to where it is mapped. This
7099 * routine is intended to be used for mapping device memory,
7103 pmap_mapdev_internal(vm_paddr_t pa, vm_size_t size, int mode, bool noflush)
7105 struct pmap_preinit_mapping *ppim;
7106 vm_offset_t va, offset;
7110 offset = pa & PAGE_MASK;
7111 size = round_page(offset + size);
7112 pa = trunc_page(pa);
7114 if (!pmap_initialized) {
7116 for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
7117 ppim = pmap_preinit_mapping + i;
7118 if (ppim->va == 0) {
7122 ppim->va = virtual_avail;
7123 virtual_avail += size;
7129 panic("%s: too many preinit mappings", __func__);
7132 * If we have a preinit mapping, re-use it.
7134 for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
7135 ppim = pmap_preinit_mapping + i;
7136 if (ppim->pa == pa && ppim->sz == size &&
7138 return ((void *)(ppim->va + offset));
7141 * If the specified range of physical addresses fits within
7142 * the direct map window, use the direct map.
7144 if (pa < dmaplimit && pa + size <= dmaplimit) {
7145 va = PHYS_TO_DMAP(pa);
7146 PMAP_LOCK(kernel_pmap);
7147 i = pmap_change_attr_locked(va, size, mode, noflush);
7148 PMAP_UNLOCK(kernel_pmap);
7150 return ((void *)(va + offset));
7152 va = kva_alloc(size);
7154 panic("%s: Couldn't allocate KVA", __func__);
7156 for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE)
7157 pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode);
7158 pmap_invalidate_range(kernel_pmap, va, va + tmpsize);
7160 pmap_invalidate_cache_range(va, va + tmpsize);
7161 return ((void *)(va + offset));
7165 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
7168 return (pmap_mapdev_internal(pa, size, mode, false));
7172 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
7175 return (pmap_mapdev_internal(pa, size, PAT_UNCACHEABLE, false));
7179 pmap_mapdev_pciecfg(vm_paddr_t pa, vm_size_t size)
7182 return (pmap_mapdev_internal(pa, size, PAT_UNCACHEABLE, true));
7186 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
7189 return (pmap_mapdev_internal(pa, size, PAT_WRITE_BACK, false));
7193 pmap_unmapdev(vm_offset_t va, vm_size_t size)
7195 struct pmap_preinit_mapping *ppim;
7199 /* If we gave a direct map region in pmap_mapdev, do nothing */
7200 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS)
7202 offset = va & PAGE_MASK;
7203 size = round_page(offset + size);
7204 va = trunc_page(va);
7205 for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
7206 ppim = pmap_preinit_mapping + i;
7207 if (ppim->va == va && ppim->sz == size) {
7208 if (pmap_initialized)
7214 if (va + size == virtual_avail)
7219 if (pmap_initialized)
7224 * Tries to demote a 1GB page mapping.
7227 pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe, vm_offset_t va)
7229 pdp_entry_t newpdpe, oldpdpe;
7230 pd_entry_t *firstpde, newpde, *pde;
7231 pt_entry_t PG_A, PG_M, PG_RW, PG_V;
7235 PG_A = pmap_accessed_bit(pmap);
7236 PG_M = pmap_modified_bit(pmap);
7237 PG_V = pmap_valid_bit(pmap);
7238 PG_RW = pmap_rw_bit(pmap);
7240 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
7242 KASSERT((oldpdpe & (PG_PS | PG_V)) == (PG_PS | PG_V),
7243 ("pmap_demote_pdpe: oldpdpe is missing PG_PS and/or PG_V"));
7244 if ((pdpg = vm_page_alloc(NULL, va >> PDPSHIFT, VM_ALLOC_INTERRUPT |
7245 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) {
7246 CTR2(KTR_PMAP, "pmap_demote_pdpe: failure for va %#lx"
7247 " in pmap %p", va, pmap);
7250 pdpgpa = VM_PAGE_TO_PHYS(pdpg);
7251 firstpde = (pd_entry_t *)PHYS_TO_DMAP(pdpgpa);
7252 newpdpe = pdpgpa | PG_M | PG_A | (oldpdpe & PG_U) | PG_RW | PG_V;
7253 KASSERT((oldpdpe & PG_A) != 0,
7254 ("pmap_demote_pdpe: oldpdpe is missing PG_A"));
7255 KASSERT((oldpdpe & (PG_M | PG_RW)) != PG_RW,
7256 ("pmap_demote_pdpe: oldpdpe is missing PG_M"));
7260 * Initialize the page directory page.
7262 for (pde = firstpde; pde < firstpde + NPDEPG; pde++) {
7268 * Demote the mapping.
7273 * Invalidate a stale recursive mapping of the page directory page.
7275 pmap_invalidate_page(pmap, (vm_offset_t)vtopde(va));
7277 pmap_pdpe_demotions++;
7278 CTR2(KTR_PMAP, "pmap_demote_pdpe: success for va %#lx"
7279 " in pmap %p", va, pmap);
7284 * Sets the memory attribute for the specified page.
7287 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
7290 m->md.pat_mode = ma;
7293 * If "m" is a normal page, update its direct mapping. This update
7294 * can be relied upon to perform any cache operations that are
7295 * required for data coherence.
7297 if ((m->flags & PG_FICTITIOUS) == 0 &&
7298 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE,
7300 panic("memory attribute change on the direct map failed");
7304 * Changes the specified virtual address range's memory type to that given by
7305 * the parameter "mode". The specified virtual address range must be
7306 * completely contained within either the direct map or the kernel map. If
7307 * the virtual address range is contained within the kernel map, then the
7308 * memory type for each of the corresponding ranges of the direct map is also
7309 * changed. (The corresponding ranges of the direct map are those ranges that
7310 * map the same physical pages as the specified virtual address range.) These
7311 * changes to the direct map are necessary because Intel describes the
7312 * behavior of their processors as "undefined" if two or more mappings to the
7313 * same physical page have different memory types.
7315 * Returns zero if the change completed successfully, and either EINVAL or
7316 * ENOMEM if the change failed. Specifically, EINVAL is returned if some part
7317 * of the virtual address range was not mapped, and ENOMEM is returned if
7318 * there was insufficient memory available to complete the change. In the
7319 * latter case, the memory type may have been changed on some part of the
7320 * virtual address range or the direct map.
7323 pmap_change_attr(vm_offset_t va, vm_size_t size, int mode)
7327 PMAP_LOCK(kernel_pmap);
7328 error = pmap_change_attr_locked(va, size, mode, false);
7329 PMAP_UNLOCK(kernel_pmap);
7334 pmap_change_attr_locked(vm_offset_t va, vm_size_t size, int mode, bool noflush)
7336 vm_offset_t base, offset, tmpva;
7337 vm_paddr_t pa_start, pa_end, pa_end1;
7341 int cache_bits_pte, cache_bits_pde, error;
7344 PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED);
7345 base = trunc_page(va);
7346 offset = va & PAGE_MASK;
7347 size = round_page(offset + size);
7350 * Only supported on kernel virtual addresses, including the direct
7351 * map but excluding the recursive map.
7353 if (base < DMAP_MIN_ADDRESS)
7356 cache_bits_pde = pmap_cache_bits(kernel_pmap, mode, 1);
7357 cache_bits_pte = pmap_cache_bits(kernel_pmap, mode, 0);
7361 * Pages that aren't mapped aren't supported. Also break down 2MB pages
7362 * into 4KB pages if required.
7364 for (tmpva = base; tmpva < base + size; ) {
7365 pdpe = pmap_pdpe(kernel_pmap, tmpva);
7366 if (pdpe == NULL || *pdpe == 0)
7368 if (*pdpe & PG_PS) {
7370 * If the current 1GB page already has the required
7371 * memory type, then we need not demote this page. Just
7372 * increment tmpva to the next 1GB page frame.
7374 if ((*pdpe & X86_PG_PDE_CACHE) == cache_bits_pde) {
7375 tmpva = trunc_1gpage(tmpva) + NBPDP;
7380 * If the current offset aligns with a 1GB page frame
7381 * and there is at least 1GB left within the range, then
7382 * we need not break down this page into 2MB pages.
7384 if ((tmpva & PDPMASK) == 0 &&
7385 tmpva + PDPMASK < base + size) {
7389 if (!pmap_demote_pdpe(kernel_pmap, pdpe, tmpva))
7392 pde = pmap_pdpe_to_pde(pdpe, tmpva);
7397 * If the current 2MB page already has the required
7398 * memory type, then we need not demote this page. Just
7399 * increment tmpva to the next 2MB page frame.
7401 if ((*pde & X86_PG_PDE_CACHE) == cache_bits_pde) {
7402 tmpva = trunc_2mpage(tmpva) + NBPDR;
7407 * If the current offset aligns with a 2MB page frame
7408 * and there is at least 2MB left within the range, then
7409 * we need not break down this page into 4KB pages.
7411 if ((tmpva & PDRMASK) == 0 &&
7412 tmpva + PDRMASK < base + size) {
7416 if (!pmap_demote_pde(kernel_pmap, pde, tmpva))
7419 pte = pmap_pde_to_pte(pde, tmpva);
7427 * Ok, all the pages exist, so run through them updating their
7428 * cache mode if required.
7430 pa_start = pa_end = 0;
7431 for (tmpva = base; tmpva < base + size; ) {
7432 pdpe = pmap_pdpe(kernel_pmap, tmpva);
7433 if (*pdpe & PG_PS) {
7434 if ((*pdpe & X86_PG_PDE_CACHE) != cache_bits_pde) {
7435 pmap_pde_attr(pdpe, cache_bits_pde,
7439 if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
7440 (*pdpe & PG_PS_FRAME) < dmaplimit) {
7441 if (pa_start == pa_end) {
7442 /* Start physical address run. */
7443 pa_start = *pdpe & PG_PS_FRAME;
7444 pa_end = pa_start + NBPDP;
7445 } else if (pa_end == (*pdpe & PG_PS_FRAME))
7448 /* Run ended, update direct map. */
7449 error = pmap_change_attr_locked(
7450 PHYS_TO_DMAP(pa_start),
7451 pa_end - pa_start, mode, noflush);
7454 /* Start physical address run. */
7455 pa_start = *pdpe & PG_PS_FRAME;
7456 pa_end = pa_start + NBPDP;
7459 tmpva = trunc_1gpage(tmpva) + NBPDP;
7462 pde = pmap_pdpe_to_pde(pdpe, tmpva);
7464 if ((*pde & X86_PG_PDE_CACHE) != cache_bits_pde) {
7465 pmap_pde_attr(pde, cache_bits_pde,
7469 if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
7470 (*pde & PG_PS_FRAME) < dmaplimit) {
7471 if (pa_start == pa_end) {
7472 /* Start physical address run. */
7473 pa_start = *pde & PG_PS_FRAME;
7474 pa_end = pa_start + NBPDR;
7475 } else if (pa_end == (*pde & PG_PS_FRAME))
7478 /* Run ended, update direct map. */
7479 error = pmap_change_attr_locked(
7480 PHYS_TO_DMAP(pa_start),
7481 pa_end - pa_start, mode, noflush);
7484 /* Start physical address run. */
7485 pa_start = *pde & PG_PS_FRAME;
7486 pa_end = pa_start + NBPDR;
7489 tmpva = trunc_2mpage(tmpva) + NBPDR;
7491 pte = pmap_pde_to_pte(pde, tmpva);
7492 if ((*pte & X86_PG_PTE_CACHE) != cache_bits_pte) {
7493 pmap_pte_attr(pte, cache_bits_pte,
7497 if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
7498 (*pte & PG_FRAME) < dmaplimit) {
7499 if (pa_start == pa_end) {
7500 /* Start physical address run. */
7501 pa_start = *pte & PG_FRAME;
7502 pa_end = pa_start + PAGE_SIZE;
7503 } else if (pa_end == (*pte & PG_FRAME))
7504 pa_end += PAGE_SIZE;
7506 /* Run ended, update direct map. */
7507 error = pmap_change_attr_locked(
7508 PHYS_TO_DMAP(pa_start),
7509 pa_end - pa_start, mode, noflush);
7512 /* Start physical address run. */
7513 pa_start = *pte & PG_FRAME;
7514 pa_end = pa_start + PAGE_SIZE;
7520 if (error == 0 && pa_start != pa_end && pa_start < dmaplimit) {
7521 pa_end1 = MIN(pa_end, dmaplimit);
7522 if (pa_start != pa_end1)
7523 error = pmap_change_attr_locked(PHYS_TO_DMAP(pa_start),
7524 pa_end1 - pa_start, mode, noflush);
7528 * Flush CPU caches if required to make sure any data isn't cached that
7529 * shouldn't be, etc.
7532 pmap_invalidate_range(kernel_pmap, base, tmpva);
7534 pmap_invalidate_cache_range(base, tmpva);
7540 * Demotes any mapping within the direct map region that covers more than the
7541 * specified range of physical addresses. This range's size must be a power
7542 * of two and its starting address must be a multiple of its size. Since the
7543 * demotion does not change any attributes of the mapping, a TLB invalidation
7544 * is not mandatory. The caller may, however, request a TLB invalidation.
7547 pmap_demote_DMAP(vm_paddr_t base, vm_size_t len, boolean_t invalidate)
7556 KASSERT(powerof2(len), ("pmap_demote_DMAP: len is not a power of 2"));
7557 KASSERT((base & (len - 1)) == 0,
7558 ("pmap_demote_DMAP: base is not a multiple of len"));
7559 if (len < NBPDP && base < dmaplimit) {
7560 va = PHYS_TO_DMAP(base);
7562 PMAP_LOCK(kernel_pmap);
7563 pdpe = pmap_pdpe(kernel_pmap, va);
7564 if ((*pdpe & X86_PG_V) == 0)
7565 panic("pmap_demote_DMAP: invalid PDPE");
7566 if ((*pdpe & PG_PS) != 0) {
7567 if (!pmap_demote_pdpe(kernel_pmap, pdpe, va))
7568 panic("pmap_demote_DMAP: PDPE failed");
7572 pde = pmap_pdpe_to_pde(pdpe, va);
7573 if ((*pde & X86_PG_V) == 0)
7574 panic("pmap_demote_DMAP: invalid PDE");
7575 if ((*pde & PG_PS) != 0) {
7576 if (!pmap_demote_pde(kernel_pmap, pde, va))
7577 panic("pmap_demote_DMAP: PDE failed");
7581 if (changed && invalidate)
7582 pmap_invalidate_page(kernel_pmap, va);
7583 PMAP_UNLOCK(kernel_pmap);
7588 * perform the pmap work for mincore
7591 pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa)
7594 pt_entry_t pte, PG_A, PG_M, PG_RW, PG_V;
7598 PG_A = pmap_accessed_bit(pmap);
7599 PG_M = pmap_modified_bit(pmap);
7600 PG_V = pmap_valid_bit(pmap);
7601 PG_RW = pmap_rw_bit(pmap);
7605 pdep = pmap_pde(pmap, addr);
7606 if (pdep != NULL && (*pdep & PG_V)) {
7607 if (*pdep & PG_PS) {
7609 /* Compute the physical address of the 4KB page. */
7610 pa = ((*pdep & PG_PS_FRAME) | (addr & PDRMASK)) &
7612 val = MINCORE_SUPER;
7614 pte = *pmap_pde_to_pte(pdep, addr);
7615 pa = pte & PG_FRAME;
7623 if ((pte & PG_V) != 0) {
7624 val |= MINCORE_INCORE;
7625 if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW))
7626 val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
7627 if ((pte & PG_A) != 0)
7628 val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
7630 if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
7631 (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
7632 (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) {
7633 /* Ensure that "PHYS_TO_VM_PAGE(pa)->object" doesn't change. */
7634 if (vm_page_pa_tryrelock(pmap, pa, locked_pa))
7637 PA_UNLOCK_COND(*locked_pa);
7643 pmap_pcid_alloc(pmap_t pmap, u_int cpuid)
7645 uint32_t gen, new_gen, pcid_next;
7647 CRITICAL_ASSERT(curthread);
7648 gen = PCPU_GET(pcid_gen);
7649 if (pmap->pm_pcids[cpuid].pm_pcid == PMAP_PCID_KERN)
7650 return (pti ? 0 : CR3_PCID_SAVE);
7651 if (pmap->pm_pcids[cpuid].pm_gen == gen)
7652 return (CR3_PCID_SAVE);
7653 pcid_next = PCPU_GET(pcid_next);
7654 KASSERT((!pti && pcid_next <= PMAP_PCID_OVERMAX) ||
7655 (pti && pcid_next <= PMAP_PCID_OVERMAX_KERN),
7656 ("cpu %d pcid_next %#x", cpuid, pcid_next));
7657 if ((!pti && pcid_next == PMAP_PCID_OVERMAX) ||
7658 (pti && pcid_next == PMAP_PCID_OVERMAX_KERN)) {
7662 PCPU_SET(pcid_gen, new_gen);
7663 pcid_next = PMAP_PCID_KERN + 1;
7667 pmap->pm_pcids[cpuid].pm_pcid = pcid_next;
7668 pmap->pm_pcids[cpuid].pm_gen = new_gen;
7669 PCPU_SET(pcid_next, pcid_next + 1);
7674 pmap_pcid_alloc_checked(pmap_t pmap, u_int cpuid)
7678 cached = pmap_pcid_alloc(pmap, cpuid);
7679 KASSERT(pmap->pm_pcids[cpuid].pm_pcid < PMAP_PCID_OVERMAX,
7680 ("pmap %p cpu %d pcid %#x", pmap, cpuid,
7681 pmap->pm_pcids[cpuid].pm_pcid));
7682 KASSERT(pmap->pm_pcids[cpuid].pm_pcid != PMAP_PCID_KERN ||
7683 pmap == kernel_pmap,
7684 ("non-kernel pmap pmap %p cpu %d pcid %#x",
7685 pmap, cpuid, pmap->pm_pcids[cpuid].pm_pcid));
7690 pmap_activate_sw_pti_post(pmap_t pmap)
7693 if (pmap->pm_ucr3 != PMAP_NO_CR3)
7694 PCPU_GET(tssp)->tss_rsp0 = ((vm_offset_t)PCPU_PTR(pti_stack) +
7695 PC_PTI_STACK_SZ * sizeof(uint64_t)) & ~0xful;
7699 pmap_activate_sw_pcid_pti(pmap_t pmap, u_int cpuid, const bool invpcid_works1)
7701 struct invpcid_descr d;
7702 uint64_t cached, cr3, kcr3, ucr3;
7704 cached = pmap_pcid_alloc_checked(pmap, cpuid);
7706 if ((cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3)
7707 load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid);
7708 PCPU_SET(curpmap, pmap);
7709 kcr3 = pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid;
7710 ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[cpuid].pm_pcid |
7713 if (!cached && pmap->pm_ucr3 != PMAP_NO_CR3) {
7715 * Explicitly invalidate translations cached from the
7716 * user page table. They are not automatically
7717 * flushed by reload of cr3 with the kernel page table
7720 * Note that the if() condition is resolved statically
7721 * by using the function argument instead of
7722 * runtime-evaluated invpcid_works value.
7724 if (invpcid_works1) {
7725 d.pcid = PMAP_PCID_USER_PT |
7726 pmap->pm_pcids[cpuid].pm_pcid;
7729 invpcid(&d, INVPCID_CTX);
7731 pmap_pti_pcid_invalidate(ucr3, kcr3);
7735 PCPU_SET(kcr3, kcr3 | CR3_PCID_SAVE);
7736 PCPU_SET(ucr3, ucr3 | CR3_PCID_SAVE);
7738 PCPU_INC(pm_save_cnt);
7742 pmap_activate_sw_pcid_invpcid_pti(pmap_t pmap, u_int cpuid)
7745 pmap_activate_sw_pcid_pti(pmap, cpuid, true);
7746 pmap_activate_sw_pti_post(pmap);
7750 pmap_activate_sw_pcid_noinvpcid_pti(pmap_t pmap, u_int cpuid)
7755 * If the INVPCID instruction is not available,
7756 * invltlb_pcid_handler() is used to handle an invalidate_all
7757 * IPI, which checks for curpmap == smp_tlb_pmap. The below
7758 * sequence of operations has a window where %CR3 is loaded
7759 * with the new pmap's PML4 address, but the curpmap value has
7760 * not yet been updated. This causes the invltlb IPI handler,
7761 * which is called between the updates, to execute as a NOP,
7762 * which leaves stale TLB entries.
7764 * Note that the most typical use of pmap_activate_sw(), from
7765 * the context switch, is immune to this race, because
7766 * interrupts are disabled (while the thread lock is owned),
7767 * and the IPI happens after curpmap is updated. Protect
7768 * other callers in a similar way, by disabling interrupts
7769 * around the %cr3 register reload and curpmap assignment.
7771 rflags = intr_disable();
7772 pmap_activate_sw_pcid_pti(pmap, cpuid, false);
7773 intr_restore(rflags);
7774 pmap_activate_sw_pti_post(pmap);
7778 pmap_activate_sw_pcid_nopti(pmap_t pmap, u_int cpuid)
7780 uint64_t cached, cr3;
7782 cached = pmap_pcid_alloc_checked(pmap, cpuid);
7784 if (!cached || (cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3)
7785 load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid |
7787 PCPU_SET(curpmap, pmap);
7789 PCPU_INC(pm_save_cnt);
7793 pmap_activate_sw_pcid_noinvpcid_nopti(pmap_t pmap, u_int cpuid)
7797 rflags = intr_disable();
7798 pmap_activate_sw_pcid_nopti(pmap, cpuid);
7799 intr_restore(rflags);
7803 pmap_activate_sw_nopcid_nopti(pmap_t pmap, u_int cpuid __unused)
7806 load_cr3(pmap->pm_cr3);
7807 PCPU_SET(curpmap, pmap);
7811 pmap_activate_sw_nopcid_pti(pmap_t pmap, u_int cpuid __unused)
7814 pmap_activate_sw_nopcid_nopti(pmap, cpuid);
7815 PCPU_SET(kcr3, pmap->pm_cr3);
7816 PCPU_SET(ucr3, pmap->pm_ucr3);
7817 pmap_activate_sw_pti_post(pmap);
7820 DEFINE_IFUNC(static, void, pmap_activate_sw_mode, (pmap_t, u_int), static)
7823 if (pmap_pcid_enabled && pti && invpcid_works)
7824 return (pmap_activate_sw_pcid_invpcid_pti);
7825 else if (pmap_pcid_enabled && pti && !invpcid_works)
7826 return (pmap_activate_sw_pcid_noinvpcid_pti);
7827 else if (pmap_pcid_enabled && !pti && invpcid_works)
7828 return (pmap_activate_sw_pcid_nopti);
7829 else if (pmap_pcid_enabled && !pti && !invpcid_works)
7830 return (pmap_activate_sw_pcid_noinvpcid_nopti);
7831 else if (!pmap_pcid_enabled && pti)
7832 return (pmap_activate_sw_nopcid_pti);
7833 else /* if (!pmap_pcid_enabled && !pti) */
7834 return (pmap_activate_sw_nopcid_nopti);
7838 pmap_activate_sw(struct thread *td)
7840 pmap_t oldpmap, pmap;
7843 oldpmap = PCPU_GET(curpmap);
7844 pmap = vmspace_pmap(td->td_proc->p_vmspace);
7845 if (oldpmap == pmap)
7847 cpuid = PCPU_GET(cpuid);
7849 CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
7851 CPU_SET(cpuid, &pmap->pm_active);
7853 pmap_activate_sw_mode(pmap, cpuid);
7855 CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active);
7857 CPU_CLR(cpuid, &oldpmap->pm_active);
7862 pmap_activate(struct thread *td)
7866 pmap_activate_sw(td);
7871 pmap_activate_boot(pmap_t pmap)
7877 * kernel_pmap must be never deactivated, and we ensure that
7878 * by never activating it at all.
7880 MPASS(pmap != kernel_pmap);
7882 cpuid = PCPU_GET(cpuid);
7884 CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
7886 CPU_SET(cpuid, &pmap->pm_active);
7888 PCPU_SET(curpmap, pmap);
7890 kcr3 = pmap->pm_cr3;
7891 if (pmap_pcid_enabled)
7892 kcr3 |= pmap->pm_pcids[cpuid].pm_pcid | CR3_PCID_SAVE;
7896 PCPU_SET(kcr3, kcr3);
7897 PCPU_SET(ucr3, PMAP_NO_CR3);
7901 pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
7906 * Increase the starting virtual address of the given mapping if a
7907 * different alignment might result in more superpage mappings.
7910 pmap_align_superpage(vm_object_t object, vm_ooffset_t offset,
7911 vm_offset_t *addr, vm_size_t size)
7913 vm_offset_t superpage_offset;
7917 if (object != NULL && (object->flags & OBJ_COLORED) != 0)
7918 offset += ptoa(object->pg_color);
7919 superpage_offset = offset & PDRMASK;
7920 if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR ||
7921 (*addr & PDRMASK) == superpage_offset)
7923 if ((*addr & PDRMASK) < superpage_offset)
7924 *addr = (*addr & ~PDRMASK) + superpage_offset;
7926 *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset;
7930 static unsigned long num_dirty_emulations;
7931 SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_dirty_emulations, CTLFLAG_RW,
7932 &num_dirty_emulations, 0, NULL);
7934 static unsigned long num_accessed_emulations;
7935 SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_accessed_emulations, CTLFLAG_RW,
7936 &num_accessed_emulations, 0, NULL);
7938 static unsigned long num_superpage_accessed_emulations;
7939 SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_superpage_accessed_emulations, CTLFLAG_RW,
7940 &num_superpage_accessed_emulations, 0, NULL);
7942 static unsigned long ad_emulation_superpage_promotions;
7943 SYSCTL_ULONG(_vm_pmap, OID_AUTO, ad_emulation_superpage_promotions, CTLFLAG_RW,
7944 &ad_emulation_superpage_promotions, 0, NULL);
7945 #endif /* INVARIANTS */
7948 pmap_emulate_accessed_dirty(pmap_t pmap, vm_offset_t va, int ftype)
7951 struct rwlock *lock;
7952 #if VM_NRESERVLEVEL > 0
7956 pt_entry_t *pte, PG_A, PG_M, PG_RW, PG_V;
7958 KASSERT(ftype == VM_PROT_READ || ftype == VM_PROT_WRITE,
7959 ("pmap_emulate_accessed_dirty: invalid fault type %d", ftype));
7961 if (!pmap_emulate_ad_bits(pmap))
7964 PG_A = pmap_accessed_bit(pmap);
7965 PG_M = pmap_modified_bit(pmap);
7966 PG_V = pmap_valid_bit(pmap);
7967 PG_RW = pmap_rw_bit(pmap);
7973 pde = pmap_pde(pmap, va);
7974 if (pde == NULL || (*pde & PG_V) == 0)
7977 if ((*pde & PG_PS) != 0) {
7978 if (ftype == VM_PROT_READ) {
7980 atomic_add_long(&num_superpage_accessed_emulations, 1);
7988 pte = pmap_pde_to_pte(pde, va);
7989 if ((*pte & PG_V) == 0)
7992 if (ftype == VM_PROT_WRITE) {
7993 if ((*pte & PG_RW) == 0)
7996 * Set the modified and accessed bits simultaneously.
7998 * Intel EPT PTEs that do software emulation of A/D bits map
7999 * PG_A and PG_M to EPT_PG_READ and EPT_PG_WRITE respectively.
8000 * An EPT misconfiguration is triggered if the PTE is writable
8001 * but not readable (WR=10). This is avoided by setting PG_A
8002 * and PG_M simultaneously.
8004 *pte |= PG_M | PG_A;
8009 #if VM_NRESERVLEVEL > 0
8010 /* try to promote the mapping */
8011 if (va < VM_MAXUSER_ADDRESS)
8012 mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
8016 m = PHYS_TO_VM_PAGE(*pte & PG_FRAME);
8018 if ((mpte == NULL || mpte->wire_count == NPTEPG) &&
8019 pmap_ps_enabled(pmap) &&
8020 (m->flags & PG_FICTITIOUS) == 0 &&
8021 vm_reserv_level_iffullpop(m) == 0) {
8022 pmap_promote_pde(pmap, pde, va, &lock);
8024 atomic_add_long(&ad_emulation_superpage_promotions, 1);
8030 if (ftype == VM_PROT_WRITE)
8031 atomic_add_long(&num_dirty_emulations, 1);
8033 atomic_add_long(&num_accessed_emulations, 1);
8035 rv = 0; /* success */
8044 pmap_get_mapping(pmap_t pmap, vm_offset_t va, uint64_t *ptr, int *num)
8049 pt_entry_t *pte, PG_V;
8053 PG_V = pmap_valid_bit(pmap);
8056 pml4 = pmap_pml4e(pmap, va);
8058 if ((*pml4 & PG_V) == 0)
8061 pdp = pmap_pml4e_to_pdpe(pml4, va);
8063 if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0)
8066 pde = pmap_pdpe_to_pde(pdp, va);
8068 if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0)
8071 pte = pmap_pde_to_pte(pde, va);
8080 * Get the kernel virtual address of a set of physical pages. If there are
8081 * physical addresses not covered by the DMAP perform a transient mapping
8082 * that will be removed when calling pmap_unmap_io_transient.
8084 * \param page The pages the caller wishes to obtain the virtual
8085 * address on the kernel memory map.
8086 * \param vaddr On return contains the kernel virtual memory address
8087 * of the pages passed in the page parameter.
8088 * \param count Number of pages passed in.
8089 * \param can_fault TRUE if the thread using the mapped pages can take
8090 * page faults, FALSE otherwise.
8092 * \returns TRUE if the caller must call pmap_unmap_io_transient when
8093 * finished or FALSE otherwise.
8097 pmap_map_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count,
8098 boolean_t can_fault)
8101 boolean_t needs_mapping;
8103 int cache_bits, error __unused, i;
8106 * Allocate any KVA space that we need, this is done in a separate
8107 * loop to prevent calling vmem_alloc while pinned.
8109 needs_mapping = FALSE;
8110 for (i = 0; i < count; i++) {
8111 paddr = VM_PAGE_TO_PHYS(page[i]);
8112 if (__predict_false(paddr >= dmaplimit)) {
8113 error = vmem_alloc(kernel_arena, PAGE_SIZE,
8114 M_BESTFIT | M_WAITOK, &vaddr[i]);
8115 KASSERT(error == 0, ("vmem_alloc failed: %d", error));
8116 needs_mapping = TRUE;
8118 vaddr[i] = PHYS_TO_DMAP(paddr);
8122 /* Exit early if everything is covered by the DMAP */
8127 * NB: The sequence of updating a page table followed by accesses
8128 * to the corresponding pages used in the !DMAP case is subject to
8129 * the situation described in the "AMD64 Architecture Programmer's
8130 * Manual Volume 2: System Programming" rev. 3.23, "7.3.1 Special
8131 * Coherency Considerations". Therefore, issuing the INVLPG right
8132 * after modifying the PTE bits is crucial.
8136 for (i = 0; i < count; i++) {
8137 paddr = VM_PAGE_TO_PHYS(page[i]);
8138 if (paddr >= dmaplimit) {
8141 * Slow path, since we can get page faults
8142 * while mappings are active don't pin the
8143 * thread to the CPU and instead add a global
8144 * mapping visible to all CPUs.
8146 pmap_qenter(vaddr[i], &page[i], 1);
8148 pte = vtopte(vaddr[i]);
8149 cache_bits = pmap_cache_bits(kernel_pmap,
8150 page[i]->md.pat_mode, 0);
8151 pte_store(pte, paddr | X86_PG_RW | X86_PG_V |
8158 return (needs_mapping);
8162 pmap_unmap_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count,
8163 boolean_t can_fault)
8170 for (i = 0; i < count; i++) {
8171 paddr = VM_PAGE_TO_PHYS(page[i]);
8172 if (paddr >= dmaplimit) {
8174 pmap_qremove(vaddr[i], 1);
8175 vmem_free(kernel_arena, vaddr[i], PAGE_SIZE);
8181 pmap_quick_enter_page(vm_page_t m)
8185 paddr = VM_PAGE_TO_PHYS(m);
8186 if (paddr < dmaplimit)
8187 return (PHYS_TO_DMAP(paddr));
8188 mtx_lock_spin(&qframe_mtx);
8189 KASSERT(*vtopte(qframe) == 0, ("qframe busy"));
8190 pte_store(vtopte(qframe), paddr | X86_PG_RW | X86_PG_V | X86_PG_A |
8191 X86_PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0));
8196 pmap_quick_remove_page(vm_offset_t addr)
8201 pte_store(vtopte(qframe), 0);
8203 mtx_unlock_spin(&qframe_mtx);
8207 * Pdp pages from the large map are managed differently from either
8208 * kernel or user page table pages. They are permanently allocated at
8209 * initialization time, and their wire count is permanently set to
8210 * zero. The pml4 entries pointing to those pages are copied into
8211 * each allocated pmap.
8213 * In contrast, pd and pt pages are managed like user page table
8214 * pages. They are dynamically allocated, and their wire count
8215 * represents the number of valid entries within the page.
8218 pmap_large_map_getptp_unlocked(void)
8222 m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
8224 if (m != NULL && (m->flags & PG_ZERO) == 0)
8230 pmap_large_map_getptp(void)
8234 PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED);
8235 m = pmap_large_map_getptp_unlocked();
8237 PMAP_UNLOCK(kernel_pmap);
8239 PMAP_LOCK(kernel_pmap);
8240 /* Callers retry. */
8245 static pdp_entry_t *
8246 pmap_large_map_pdpe(vm_offset_t va)
8248 vm_pindex_t pml4_idx;
8251 pml4_idx = pmap_pml4e_index(va);
8252 KASSERT(LMSPML4I <= pml4_idx && pml4_idx < LMSPML4I + lm_ents,
8253 ("pmap_large_map_pdpe: va %#jx out of range idx %#jx LMSPML4I "
8255 (uintmax_t)va, (uintmax_t)pml4_idx, LMSPML4I, lm_ents));
8256 KASSERT((kernel_pmap->pm_pml4[pml4_idx] & X86_PG_V) != 0,
8257 ("pmap_large_map_pdpe: invalid pml4 for va %#jx idx %#jx "
8258 "LMSPML4I %#jx lm_ents %d",
8259 (uintmax_t)va, (uintmax_t)pml4_idx, LMSPML4I, lm_ents));
8260 mphys = kernel_pmap->pm_pml4[pml4_idx] & PG_FRAME;
8261 return ((pdp_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pdpe_index(va));
8265 pmap_large_map_pde(vm_offset_t va)
8272 pdpe = pmap_large_map_pdpe(va);
8274 m = pmap_large_map_getptp();
8277 mphys = VM_PAGE_TO_PHYS(m);
8278 *pdpe = mphys | X86_PG_A | X86_PG_RW | X86_PG_V | pg_nx;
8280 MPASS((*pdpe & X86_PG_PS) == 0);
8281 mphys = *pdpe & PG_FRAME;
8283 return ((pd_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pde_index(va));
8287 pmap_large_map_pte(vm_offset_t va)
8294 pde = pmap_large_map_pde(va);
8296 m = pmap_large_map_getptp();
8299 mphys = VM_PAGE_TO_PHYS(m);
8300 *pde = mphys | X86_PG_A | X86_PG_RW | X86_PG_V | pg_nx;
8301 PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde))->wire_count++;
8303 MPASS((*pde & X86_PG_PS) == 0);
8304 mphys = *pde & PG_FRAME;
8306 return ((pt_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pte_index(va));
8310 pmap_large_map_getva(vm_size_t len, vm_offset_t align, vm_offset_t phase,
8311 vmem_addr_t *vmem_res)
8315 * Large mappings are all but static. Consequently, there
8316 * is no point in waiting for an earlier allocation to be
8319 return (vmem_xalloc(large_vmem, len, align, phase, 0, VMEM_ADDR_MIN,
8320 VMEM_ADDR_MAX, M_NOWAIT | M_BESTFIT, vmem_res));
8324 pmap_large_map(vm_paddr_t spa, vm_size_t len, void **addr,
8330 vm_offset_t va, inc;
8331 vmem_addr_t vmem_res;
8335 if (len == 0 || spa + len < spa)
8338 /* See if DMAP can serve. */
8339 if (spa + len <= dmaplimit) {
8340 va = PHYS_TO_DMAP(spa);
8342 return (pmap_change_attr(va, len, mattr));
8346 * No, allocate KVA. Fit the address with best possible
8347 * alignment for superpages. Fall back to worse align if
8351 if ((amd_feature & AMDID_PAGE1GB) != 0 && rounddown2(spa + len,
8352 NBPDP) >= roundup2(spa, NBPDP) + NBPDP)
8353 error = pmap_large_map_getva(len, NBPDP, spa & PDPMASK,
8355 if (error != 0 && rounddown2(spa + len, NBPDR) >= roundup2(spa,
8357 error = pmap_large_map_getva(len, NBPDR, spa & PDRMASK,
8360 error = pmap_large_map_getva(len, PAGE_SIZE, 0, &vmem_res);
8365 * Fill pagetable. PG_M is not pre-set, we scan modified bits
8366 * in the pagetable to minimize flushing. No need to
8367 * invalidate TLB, since we only update invalid entries.
8369 PMAP_LOCK(kernel_pmap);
8370 for (pa = spa, va = vmem_res; len > 0; pa += inc, va += inc,
8372 if ((amd_feature & AMDID_PAGE1GB) != 0 && len >= NBPDP &&
8373 (pa & PDPMASK) == 0 && (va & PDPMASK) == 0) {
8374 pdpe = pmap_large_map_pdpe(va);
8376 *pdpe = pa | pg_g | X86_PG_PS | X86_PG_RW |
8377 X86_PG_V | X86_PG_A | pg_nx |
8378 pmap_cache_bits(kernel_pmap, mattr, TRUE);
8380 } else if (len >= NBPDR && (pa & PDRMASK) == 0 &&
8381 (va & PDRMASK) == 0) {
8382 pde = pmap_large_map_pde(va);
8384 *pde = pa | pg_g | X86_PG_PS | X86_PG_RW |
8385 X86_PG_V | X86_PG_A | pg_nx |
8386 pmap_cache_bits(kernel_pmap, mattr, TRUE);
8387 PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde))->
8391 pte = pmap_large_map_pte(va);
8393 *pte = pa | pg_g | X86_PG_RW | X86_PG_V |
8394 X86_PG_A | pg_nx | pmap_cache_bits(kernel_pmap,
8396 PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte))->
8401 PMAP_UNLOCK(kernel_pmap);
8404 *addr = (void *)vmem_res;
8409 pmap_large_unmap(void *svaa, vm_size_t len)
8411 vm_offset_t sva, va;
8413 pdp_entry_t *pdpe, pdp;
8414 pd_entry_t *pde, pd;
8417 struct spglist spgf;
8419 sva = (vm_offset_t)svaa;
8420 if (len == 0 || sva + len < sva || (sva >= DMAP_MIN_ADDRESS &&
8421 sva + len <= DMAP_MIN_ADDRESS + dmaplimit))
8425 KASSERT(LARGEMAP_MIN_ADDRESS <= sva && sva + len <=
8426 LARGEMAP_MAX_ADDRESS + NBPML4 * (u_long)lm_ents,
8427 ("not largemap range %#lx %#lx", (u_long)svaa, (u_long)svaa + len));
8428 PMAP_LOCK(kernel_pmap);
8429 for (va = sva; va < sva + len; va += inc) {
8430 pdpe = pmap_large_map_pdpe(va);
8432 KASSERT((pdp & X86_PG_V) != 0,
8433 ("invalid pdp va %#lx pdpe %#lx pdp %#lx", va,
8434 (u_long)pdpe, pdp));
8435 if ((pdp & X86_PG_PS) != 0) {
8436 KASSERT((amd_feature & AMDID_PAGE1GB) != 0,
8437 ("no 1G pages, va %#lx pdpe %#lx pdp %#lx", va,
8438 (u_long)pdpe, pdp));
8439 KASSERT((va & PDPMASK) == 0,
8440 ("PDPMASK bit set, va %#lx pdpe %#lx pdp %#lx", va,
8441 (u_long)pdpe, pdp));
8442 KASSERT(len <= NBPDP,
8443 ("len < NBPDP, sva %#lx va %#lx pdpe %#lx pdp %#lx "
8444 "len %#lx", sva, va, (u_long)pdpe, pdp, len));
8449 pde = pmap_pdpe_to_pde(pdpe, va);
8451 KASSERT((pd & X86_PG_V) != 0,
8452 ("invalid pd va %#lx pde %#lx pd %#lx", va,
8454 if ((pd & X86_PG_PS) != 0) {
8455 KASSERT((va & PDRMASK) == 0,
8456 ("PDRMASK bit set, va %#lx pde %#lx pd %#lx", va,
8458 KASSERT(len <= NBPDR,
8459 ("len < NBPDR, sva %#lx va %#lx pde %#lx pd %#lx "
8460 "len %#lx", sva, va, (u_long)pde, pd, len));
8463 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pde));
8465 if (m->wire_count == 0) {
8467 SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss);
8471 pte = pmap_pde_to_pte(pde, va);
8472 KASSERT((*pte & X86_PG_V) != 0,
8473 ("invalid pte va %#lx pte %#lx pt %#lx", va,
8474 (u_long)pte, *pte));
8477 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pte));
8479 if (m->wire_count == 0) {
8481 SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss);
8482 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pde));
8484 if (m->wire_count == 0) {
8486 SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss);
8490 pmap_invalidate_range(kernel_pmap, sva, sva + len);
8491 PMAP_UNLOCK(kernel_pmap);
8492 vm_page_free_pages_toq(&spgf, false);
8493 vmem_free(large_vmem, sva, len);
8497 pmap_large_map_wb_fence_mfence(void)
8504 pmap_large_map_wb_fence_sfence(void)
8511 pmap_large_map_wb_fence_nop(void)
8515 DEFINE_IFUNC(static, void, pmap_large_map_wb_fence, (void), static)
8518 if (cpu_vendor_id != CPU_VENDOR_INTEL)
8519 return (pmap_large_map_wb_fence_mfence);
8520 else if ((cpu_stdext_feature & (CPUID_STDEXT_CLWB |
8521 CPUID_STDEXT_CLFLUSHOPT)) == 0)
8522 return (pmap_large_map_wb_fence_sfence);
8524 /* clflush is strongly enough ordered */
8525 return (pmap_large_map_wb_fence_nop);
8529 pmap_large_map_flush_range_clwb(vm_offset_t va, vm_size_t len)
8532 for (; len > 0; len -= cpu_clflush_line_size,
8533 va += cpu_clflush_line_size)
8538 pmap_large_map_flush_range_clflushopt(vm_offset_t va, vm_size_t len)
8541 for (; len > 0; len -= cpu_clflush_line_size,
8542 va += cpu_clflush_line_size)
8547 pmap_large_map_flush_range_clflush(vm_offset_t va, vm_size_t len)
8550 for (; len > 0; len -= cpu_clflush_line_size,
8551 va += cpu_clflush_line_size)
8556 pmap_large_map_flush_range_nop(vm_offset_t sva __unused, vm_size_t len __unused)
8560 DEFINE_IFUNC(static, void, pmap_large_map_flush_range, (vm_offset_t, vm_size_t),
8564 if ((cpu_stdext_feature & CPUID_STDEXT_CLWB) != 0)
8565 return (pmap_large_map_flush_range_clwb);
8566 else if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0)
8567 return (pmap_large_map_flush_range_clflushopt);
8568 else if ((cpu_feature & CPUID_CLFSH) != 0)
8569 return (pmap_large_map_flush_range_clflush);
8571 return (pmap_large_map_flush_range_nop);
8575 pmap_large_map_wb_large(vm_offset_t sva, vm_offset_t eva)
8577 volatile u_long *pe;
8583 for (va = sva; va < eva; va += inc) {
8585 if ((amd_feature & AMDID_PAGE1GB) != 0) {
8586 pe = (volatile u_long *)pmap_large_map_pdpe(va);
8588 if ((p & X86_PG_PS) != 0)
8592 pe = (volatile u_long *)pmap_large_map_pde(va);
8594 if ((p & X86_PG_PS) != 0)
8598 pe = (volatile u_long *)pmap_large_map_pte(va);
8604 if ((p & X86_PG_AVAIL1) != 0) {
8606 * Spin-wait for the end of a parallel
8613 * If we saw other write-back
8614 * occuring, we cannot rely on PG_M to
8615 * indicate state of the cache. The
8616 * PG_M bit is cleared before the
8617 * flush to avoid ignoring new writes,
8618 * and writes which are relevant for
8619 * us might happen after.
8625 if ((p & X86_PG_M) != 0 || seen_other) {
8626 if (!atomic_fcmpset_long(pe, &p,
8627 (p & ~X86_PG_M) | X86_PG_AVAIL1))
8629 * If we saw PG_M without
8630 * PG_AVAIL1, and then on the
8631 * next attempt we do not
8632 * observe either PG_M or
8633 * PG_AVAIL1, the other
8634 * write-back started after us
8635 * and finished before us. We
8636 * can rely on it doing our
8640 pmap_large_map_flush_range(va, inc);
8641 atomic_clear_long(pe, X86_PG_AVAIL1);
8650 * Write-back cache lines for the given address range.
8652 * Must be called only on the range or sub-range returned from
8653 * pmap_large_map(). Must not be called on the coalesced ranges.
8655 * Does nothing on CPUs without CLWB, CLFLUSHOPT, or CLFLUSH
8656 * instructions support.
8659 pmap_large_map_wb(void *svap, vm_size_t len)
8661 vm_offset_t eva, sva;
8663 sva = (vm_offset_t)svap;
8665 pmap_large_map_wb_fence();
8666 if (sva >= DMAP_MIN_ADDRESS && eva <= DMAP_MIN_ADDRESS + dmaplimit) {
8667 pmap_large_map_flush_range(sva, len);
8669 KASSERT(sva >= LARGEMAP_MIN_ADDRESS &&
8670 eva <= LARGEMAP_MIN_ADDRESS + lm_ents * NBPML4,
8671 ("pmap_large_map_wb: not largemap %#lx %#lx", sva, len));
8672 pmap_large_map_wb_large(sva, eva);
8674 pmap_large_map_wb_fence();
8678 pmap_pti_alloc_page(void)
8682 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
8683 m = vm_page_grab(pti_obj, pti_pg_idx++, VM_ALLOC_NOBUSY |
8684 VM_ALLOC_WIRED | VM_ALLOC_ZERO);
8689 pmap_pti_free_page(vm_page_t m)
8692 KASSERT(m->wire_count > 0, ("page %p not wired", m));
8693 if (!vm_page_unwire_noq(m))
8695 vm_page_free_zero(m);
8709 pti_obj = vm_pager_allocate(OBJT_PHYS, NULL, 0, VM_PROT_ALL, 0, NULL);
8710 VM_OBJECT_WLOCK(pti_obj);
8711 pml4_pg = pmap_pti_alloc_page();
8712 pti_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4_pg));
8713 for (va = VM_MIN_KERNEL_ADDRESS; va <= VM_MAX_KERNEL_ADDRESS &&
8714 va >= VM_MIN_KERNEL_ADDRESS && va > NBPML4; va += NBPML4) {
8715 pdpe = pmap_pti_pdpe(va);
8716 pmap_pti_wire_pte(pdpe);
8718 pmap_pti_add_kva_locked((vm_offset_t)&__pcpu[0],
8719 (vm_offset_t)&__pcpu[0] + sizeof(__pcpu[0]) * MAXCPU, false);
8720 pmap_pti_add_kva_locked((vm_offset_t)gdt, (vm_offset_t)gdt +
8721 sizeof(struct user_segment_descriptor) * NGDT * MAXCPU, false);
8722 pmap_pti_add_kva_locked((vm_offset_t)idt, (vm_offset_t)idt +
8723 sizeof(struct gate_descriptor) * NIDT, false);
8724 pmap_pti_add_kva_locked((vm_offset_t)common_tss,
8725 (vm_offset_t)common_tss + sizeof(struct amd64tss) * MAXCPU, false);
8727 /* Doublefault stack IST 1 */
8728 va = common_tss[i].tss_ist1;
8729 pmap_pti_add_kva_locked(va - PAGE_SIZE, va, false);
8730 /* NMI stack IST 2 */
8731 va = common_tss[i].tss_ist2 + sizeof(struct nmi_pcpu);
8732 pmap_pti_add_kva_locked(va - PAGE_SIZE, va, false);
8733 /* MC# stack IST 3 */
8734 va = common_tss[i].tss_ist3 + sizeof(struct nmi_pcpu);
8735 pmap_pti_add_kva_locked(va - PAGE_SIZE, va, false);
8736 /* DB# stack IST 4 */
8737 va = common_tss[i].tss_ist4 + sizeof(struct nmi_pcpu);
8738 pmap_pti_add_kva_locked(va - PAGE_SIZE, va, false);
8740 pmap_pti_add_kva_locked((vm_offset_t)kernphys + KERNBASE,
8741 (vm_offset_t)etext, true);
8742 pti_finalized = true;
8743 VM_OBJECT_WUNLOCK(pti_obj);
8745 SYSINIT(pmap_pti, SI_SUB_CPU + 1, SI_ORDER_ANY, pmap_pti_init, NULL);
8747 static pdp_entry_t *
8748 pmap_pti_pdpe(vm_offset_t va)
8750 pml4_entry_t *pml4e;
8753 vm_pindex_t pml4_idx;
8756 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
8758 pml4_idx = pmap_pml4e_index(va);
8759 pml4e = &pti_pml4[pml4_idx];
8763 panic("pml4 alloc after finalization\n");
8764 m = pmap_pti_alloc_page();
8766 pmap_pti_free_page(m);
8767 mphys = *pml4e & ~PAGE_MASK;
8769 mphys = VM_PAGE_TO_PHYS(m);
8770 *pml4e = mphys | X86_PG_RW | X86_PG_V;
8773 mphys = *pml4e & ~PAGE_MASK;
8775 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pdpe_index(va);
8780 pmap_pti_wire_pte(void *pte)
8784 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
8785 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte));
8790 pmap_pti_unwire_pde(void *pde, bool only_ref)
8794 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
8795 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde));
8796 MPASS(m->wire_count > 0);
8797 MPASS(only_ref || m->wire_count > 1);
8798 pmap_pti_free_page(m);
8802 pmap_pti_unwire_pte(void *pte, vm_offset_t va)
8807 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
8808 m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte));
8809 MPASS(m->wire_count > 0);
8810 if (pmap_pti_free_page(m)) {
8811 pde = pmap_pti_pde(va);
8812 MPASS((*pde & (X86_PG_PS | X86_PG_V)) == X86_PG_V);
8814 pmap_pti_unwire_pde(pde, false);
8819 pmap_pti_pde(vm_offset_t va)
8827 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
8829 pdpe = pmap_pti_pdpe(va);
8831 m = pmap_pti_alloc_page();
8833 pmap_pti_free_page(m);
8834 MPASS((*pdpe & X86_PG_PS) == 0);
8835 mphys = *pdpe & ~PAGE_MASK;
8837 mphys = VM_PAGE_TO_PHYS(m);
8838 *pdpe = mphys | X86_PG_RW | X86_PG_V;
8841 MPASS((*pdpe & X86_PG_PS) == 0);
8842 mphys = *pdpe & ~PAGE_MASK;
8845 pde = (pd_entry_t *)PHYS_TO_DMAP(mphys);
8846 pd_idx = pmap_pde_index(va);
8852 pmap_pti_pte(vm_offset_t va, bool *unwire_pde)
8859 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
8861 pde = pmap_pti_pde(va);
8862 if (unwire_pde != NULL) {
8864 pmap_pti_wire_pte(pde);
8867 m = pmap_pti_alloc_page();
8869 pmap_pti_free_page(m);
8870 MPASS((*pde & X86_PG_PS) == 0);
8871 mphys = *pde & ~(PAGE_MASK | pg_nx);
8873 mphys = VM_PAGE_TO_PHYS(m);
8874 *pde = mphys | X86_PG_RW | X86_PG_V;
8875 if (unwire_pde != NULL)
8876 *unwire_pde = false;
8879 MPASS((*pde & X86_PG_PS) == 0);
8880 mphys = *pde & ~(PAGE_MASK | pg_nx);
8883 pte = (pt_entry_t *)PHYS_TO_DMAP(mphys);
8884 pte += pmap_pte_index(va);
8890 pmap_pti_add_kva_locked(vm_offset_t sva, vm_offset_t eva, bool exec)
8894 pt_entry_t *pte, ptev;
8897 VM_OBJECT_ASSERT_WLOCKED(pti_obj);
8899 sva = trunc_page(sva);
8900 MPASS(sva > VM_MAXUSER_ADDRESS);
8901 eva = round_page(eva);
8903 for (; sva < eva; sva += PAGE_SIZE) {
8904 pte = pmap_pti_pte(sva, &unwire_pde);
8905 pa = pmap_kextract(sva);
8906 ptev = pa | X86_PG_RW | X86_PG_V | X86_PG_A | X86_PG_G |
8907 (exec ? 0 : pg_nx) | pmap_cache_bits(kernel_pmap,
8908 VM_MEMATTR_DEFAULT, FALSE);
8910 pte_store(pte, ptev);
8911 pmap_pti_wire_pte(pte);
8913 KASSERT(!pti_finalized,
8914 ("pti overlap after fin %#lx %#lx %#lx",
8916 KASSERT(*pte == ptev,
8917 ("pti non-identical pte after fin %#lx %#lx %#lx",
8921 pde = pmap_pti_pde(sva);
8922 pmap_pti_unwire_pde(pde, true);
8928 pmap_pti_add_kva(vm_offset_t sva, vm_offset_t eva, bool exec)
8933 VM_OBJECT_WLOCK(pti_obj);
8934 pmap_pti_add_kva_locked(sva, eva, exec);
8935 VM_OBJECT_WUNLOCK(pti_obj);
8939 pmap_pti_remove_kva(vm_offset_t sva, vm_offset_t eva)
8946 sva = rounddown2(sva, PAGE_SIZE);
8947 MPASS(sva > VM_MAXUSER_ADDRESS);
8948 eva = roundup2(eva, PAGE_SIZE);
8950 VM_OBJECT_WLOCK(pti_obj);
8951 for (va = sva; va < eva; va += PAGE_SIZE) {
8952 pte = pmap_pti_pte(va, NULL);
8953 KASSERT((*pte & X86_PG_V) != 0,
8954 ("invalid pte va %#lx pte %#lx pt %#lx", va,
8955 (u_long)pte, *pte));
8957 pmap_pti_unwire_pte(pte, va);
8959 pmap_invalidate_range(kernel_pmap, sva, eva);
8960 VM_OBJECT_WUNLOCK(pti_obj);
8963 #include "opt_ddb.h"
8965 #include <sys/kdb.h>
8966 #include <ddb/ddb.h>
8968 DB_SHOW_COMMAND(pte, pmap_print_pte)
8974 pt_entry_t *pte, PG_V;
8978 db_printf("show pte addr\n");
8981 va = (vm_offset_t)addr;
8983 if (kdb_thread != NULL)
8984 pmap = vmspace_pmap(kdb_thread->td_proc->p_vmspace);
8986 pmap = PCPU_GET(curpmap);
8988 PG_V = pmap_valid_bit(pmap);
8989 pml4 = pmap_pml4e(pmap, va);
8990 db_printf("VA %#016lx pml4e %#016lx", va, *pml4);
8991 if ((*pml4 & PG_V) == 0) {
8995 pdp = pmap_pml4e_to_pdpe(pml4, va);
8996 db_printf(" pdpe %#016lx", *pdp);
8997 if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0) {
9001 pde = pmap_pdpe_to_pde(pdp, va);
9002 db_printf(" pde %#016lx", *pde);
9003 if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0) {
9007 pte = pmap_pde_to_pte(pde, va);
9008 db_printf(" pte %#016lx\n", *pte);
9011 DB_SHOW_COMMAND(phys2dmap, pmap_phys2dmap)
9016 a = (vm_paddr_t)addr;
9017 db_printf("0x%jx\n", (uintmax_t)PHYS_TO_DMAP(a));
9019 db_printf("show phys2dmap addr\n");