2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
8 * Copyright (c) 2005-2010 Alan L. Cox <alc@cs.rice.edu>
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
46 * Copyright (c) 2003 Networks Associates Technology, Inc.
47 * All rights reserved.
49 * This software was developed for the FreeBSD Project by Jake Burkholder,
50 * Safeport Network Services, and Network Associates Laboratories, the
51 * Security Research Division of Network Associates, Inc. under
52 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA
53 * CHATS research program.
55 * Redistribution and use in source and binary forms, with or without
56 * modification, are permitted provided that the following conditions
58 * 1. Redistributions of source code must retain the above copyright
59 * notice, this list of conditions and the following disclaimer.
60 * 2. Redistributions in binary form must reproduce the above copyright
61 * notice, this list of conditions and the following disclaimer in the
62 * documentation and/or other materials provided with the distribution.
64 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
65 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
66 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
67 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
68 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
69 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
70 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
71 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
72 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
73 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
77 #include <sys/cdefs.h>
78 __FBSDID("$FreeBSD$");
81 * Manages physical address maps.
83 * In addition to hardware address maps, this
84 * module is called upon to provide software-use-only
85 * maps which may or may not be stored in the same
86 * form as hardware maps. These pseudo-maps are
87 * used to store intermediate results from copy
88 * operations to and from address spaces.
90 * Since the information managed by this module is
91 * also stored by the logical address mapping module,
92 * this module may throw away valid virtual-to-physical
93 * mappings at almost any time. However, invalidations
94 * of virtual-to-physical mappings must be done as
97 * In order to cope with hardware architectures which
98 * make virtual-to-physical map invalidates expensive,
99 * this module may delay invalidate or reduced protection
100 * operations until such time as they are actually
101 * necessary. This module is given full information as
102 * to which processors are currently using which maps,
103 * and to when physical maps must be made correct.
106 #include "opt_apic.h"
108 #include "opt_pmap.h"
110 #include "opt_xbox.h"
112 #include <sys/param.h>
113 #include <sys/systm.h>
114 #include <sys/kernel.h>
116 #include <sys/lock.h>
117 #include <sys/malloc.h>
118 #include <sys/mman.h>
119 #include <sys/msgbuf.h>
120 #include <sys/mutex.h>
121 #include <sys/proc.h>
122 #include <sys/rwlock.h>
123 #include <sys/sf_buf.h>
125 #include <sys/vmmeter.h>
126 #include <sys/sched.h>
127 #include <sys/sysctl.h>
131 #include <sys/cpuset.h>
135 #include <vm/vm_param.h>
136 #include <vm/vm_kern.h>
137 #include <vm/vm_page.h>
138 #include <vm/vm_map.h>
139 #include <vm/vm_object.h>
140 #include <vm/vm_extern.h>
141 #include <vm/vm_pageout.h>
142 #include <vm/vm_pager.h>
143 #include <vm/vm_reserv.h>
148 #include <machine/intr_machdep.h>
149 #include <machine/apicvar.h>
151 #include <machine/cpu.h>
152 #include <machine/cputypes.h>
153 #include <machine/md_var.h>
154 #include <machine/pcb.h>
155 #include <machine/specialreg.h>
157 #include <machine/smp.h>
161 #include <machine/xbox.h>
164 #if !defined(CPU_DISABLE_SSE) && defined(I686_CPU)
165 #define CPU_ENABLE_SSE
168 #ifndef PMAP_SHPGPERPROC
169 #define PMAP_SHPGPERPROC 200
172 #if !defined(DIAGNOSTIC)
173 #ifdef __GNUC_GNU_INLINE__
174 #define PMAP_INLINE __attribute__((__gnu_inline__)) inline
176 #define PMAP_INLINE extern inline
183 #define PV_STAT(x) do { x ; } while (0)
185 #define PV_STAT(x) do { } while (0)
188 #define pa_index(pa) ((pa) >> PDRSHIFT)
189 #define pa_to_pvh(pa) (&pv_table[pa_index(pa)])
192 * Get PDEs and PTEs for user/kernel address space
194 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
195 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
197 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
198 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
199 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
200 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
201 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
203 #define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \
204 atomic_clear_int((u_int *)(pte), PG_W))
205 #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v)))
207 struct pmap kernel_pmap_store;
208 LIST_HEAD(pmaplist, pmap);
209 static struct pmaplist allpmaps;
210 static struct mtx allpmaps_lock;
212 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
213 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
214 int pgeflag = 0; /* PG_G or-in */
215 int pseflag = 0; /* PG_PS or-in */
217 static int nkpt = NKPT;
218 vm_offset_t kernel_vm_end = KERNBASE + NKPT * NBPDR;
219 extern u_int32_t KERNend;
220 extern u_int32_t KPTphys;
224 static uma_zone_t pdptzone;
227 static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters");
229 static int pat_works = 1;
230 SYSCTL_INT(_vm_pmap, OID_AUTO, pat_works, CTLFLAG_RD, &pat_works, 1,
231 "Is page attribute table fully functional?");
233 static int pg_ps_enabled = 1;
234 SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RDTUN, &pg_ps_enabled, 0,
235 "Are large page mappings enabled?");
237 #define PAT_INDEX_SIZE 8
238 static int pat_index[PAT_INDEX_SIZE]; /* cache mode to PAT index conversion */
241 * Isolate the global pv list lock from data and other locks to prevent false
242 * sharing within the cache.
246 char padding[CACHE_LINE_SIZE - sizeof(struct rwlock)];
247 } pvh_global __aligned(CACHE_LINE_SIZE);
249 #define pvh_global_lock pvh_global.lock
252 * Data for the pv entry allocation mechanism
254 static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks);
255 static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0;
256 static struct md_page *pv_table;
257 static int shpgperproc = PMAP_SHPGPERPROC;
259 struct pv_chunk *pv_chunkbase; /* KVA block for pv_chunks */
260 int pv_maxchunks; /* How many chunks we have KVA for */
261 vm_offset_t pv_vafree; /* freelist stored in the PTE */
264 * All those kernel PT submaps that BSD is so fond of
273 static struct sysmaps sysmaps_pcpu[MAXCPU];
275 static pd_entry_t *KPTD;
278 struct msgbuf *msgbufp = 0;
283 static caddr_t crashdumpmap;
285 static pt_entry_t *PMAP1 = 0, *PMAP2;
286 static pt_entry_t *PADDR1 = 0, *PADDR2;
289 static int PMAP1changedcpu;
290 SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD,
292 "Number of times pmap_pte_quick changed CPU with same PMAP1");
294 static int PMAP1changed;
295 SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD,
297 "Number of times pmap_pte_quick changed PMAP1");
298 static int PMAP1unchanged;
299 SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD,
301 "Number of times pmap_pte_quick didn't change PMAP1");
302 static struct mtx PMAP2mutex;
304 static void free_pv_chunk(struct pv_chunk *pc);
305 static void free_pv_entry(pmap_t pmap, pv_entry_t pv);
306 static pv_entry_t get_pv_entry(pmap_t pmap, boolean_t try);
307 static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa);
308 static boolean_t pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa);
309 static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa);
310 static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va);
311 static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap,
313 static int pmap_pvh_wired_mappings(struct md_page *pvh, int count);
315 static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va);
316 static boolean_t pmap_enter_pde(pmap_t pmap, vm_offset_t va, vm_page_t m,
318 static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va,
319 vm_page_t m, vm_prot_t prot, vm_page_t mpte);
320 static void pmap_flush_page(vm_page_t m);
321 static void pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte);
322 static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte);
323 static boolean_t pmap_is_modified_pvh(struct md_page *pvh);
324 static boolean_t pmap_is_referenced_pvh(struct md_page *pvh);
325 static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode);
326 static void pmap_kenter_pde(vm_offset_t va, pd_entry_t newpde);
327 static vm_page_t pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va);
328 static void pmap_pde_attr(pd_entry_t *pde, int cache_bits);
329 static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va);
330 static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva,
332 static void pmap_pte_attr(pt_entry_t *pte, int cache_bits);
333 static void pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
335 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva,
337 static void pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte);
338 static void pmap_remove_page(struct pmap *pmap, vm_offset_t va,
340 static void pmap_remove_entry(struct pmap *pmap, vm_page_t m,
342 static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m);
343 static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va,
345 static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
347 static void pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde);
349 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags);
351 static vm_page_t _pmap_allocpte(pmap_t pmap, u_int ptepindex, int flags);
352 static void _pmap_unwire_ptp(pmap_t pmap, vm_page_t m, vm_page_t *free);
353 static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va);
354 static void pmap_pte_release(pt_entry_t *pte);
355 static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t *);
357 static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait);
359 static void pmap_set_pg(void);
361 static __inline void pagezero(void *page);
363 CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t));
364 CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t));
367 * If you get an error here, then you set KVA_PAGES wrong! See the
368 * description of KVA_PAGES in sys/i386/include/pmap.h. It must be
369 * multiple of 4 for a normal kernel, or a multiple of 8 for a PAE.
371 CTASSERT(KERNBASE % (1 << 24) == 0);
374 * Bootstrap the system enough to run with virtual memory.
376 * On the i386 this is called after mapping has already been enabled
377 * and just syncs the pmap module with what has already been done.
378 * [We can't call it easily with mapping off since the kernel is not
379 * mapped with PA == VA, hence we would have to relocate every address
380 * from the linked base (virtual) address "KERNBASE" to the actual
381 * (physical) address starting relative to 0]
384 pmap_bootstrap(vm_paddr_t firstaddr)
387 pt_entry_t *pte, *unused;
388 struct sysmaps *sysmaps;
392 * Initialize the first available kernel virtual address. However,
393 * using "firstaddr" may waste a few pages of the kernel virtual
394 * address space, because locore may not have mapped every physical
395 * page that it allocated. Preferably, locore would provide a first
396 * unused virtual address in addition to "firstaddr".
398 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
400 virtual_end = VM_MAX_KERNEL_ADDRESS;
403 * Initialize the kernel pmap (which is statically allocated).
405 PMAP_LOCK_INIT(kernel_pmap);
406 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD);
408 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT);
410 kernel_pmap->pm_root = NULL;
411 CPU_FILL(&kernel_pmap->pm_active); /* don't allow deactivation */
412 TAILQ_INIT(&kernel_pmap->pm_pvchunk);
415 * Initialize the global pv list lock.
417 rw_init(&pvh_global_lock, "pmap pv global");
419 LIST_INIT(&allpmaps);
422 * Request a spin mutex so that changes to allpmaps cannot be
423 * preempted by smp_rendezvous_cpus(). Otherwise,
424 * pmap_update_pde_kernel() could access allpmaps while it is
427 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN);
428 mtx_lock_spin(&allpmaps_lock);
429 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list);
430 mtx_unlock_spin(&allpmaps_lock);
433 * Reserve some special page table entries/VA space for temporary
436 #define SYSMAP(c, p, v, n) \
437 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
443 * CMAP1/CMAP2 are used for zeroing and copying pages.
444 * CMAP3 is used for the idle process page zeroing.
446 for (i = 0; i < MAXCPU; i++) {
447 sysmaps = &sysmaps_pcpu[i];
448 mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF);
449 SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1)
450 SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1)
452 SYSMAP(caddr_t, CMAP3, CADDR3, 1)
457 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS)
460 * ptvmmap is used for reading arbitrary physical pages via /dev/mem.
462 SYSMAP(caddr_t, unused, ptvmmap, 1)
465 * msgbufp is used to map the system message buffer.
467 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(msgbufsize)))
470 * KPTmap is used by pmap_kextract().
472 * KPTmap is first initialized by locore. However, that initial
473 * KPTmap can only support NKPT page table pages. Here, a larger
474 * KPTmap is created that can support KVA_PAGES page table pages.
476 SYSMAP(pt_entry_t *, KPTD, KPTmap, KVA_PAGES)
478 for (i = 0; i < NKPT; i++)
479 KPTD[i] = (KPTphys + (i << PAGE_SHIFT)) | pgeflag | PG_RW | PG_V;
482 * Adjust the start of the KPTD and KPTmap so that the implementation
483 * of pmap_kextract() and pmap_growkernel() can be made simpler.
486 KPTmap -= i386_btop(KPTDI << PDRSHIFT);
489 * PADDR1 and PADDR2 are used by pmap_pte_quick() and pmap_pte(),
492 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1)
493 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1)
495 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF);
500 * Leave in place an identity mapping (virt == phys) for the low 1 MB
501 * physical memory region that is used by the ACPI wakeup code. This
502 * mapping must not have PG_G set.
505 /* FIXME: This is gross, but needed for the XBOX. Since we are in such
506 * an early stadium, we cannot yet neatly map video memory ... :-(
507 * Better fixes are very welcome! */
508 if (!arch_i386_is_xbox)
510 for (i = 1; i < NKPT; i++)
513 /* Initialize the PAT MSR if present. */
516 /* Turn on PG_G on kernel page(s) */
526 int pat_table[PAT_INDEX_SIZE];
531 /* Set default PAT index table. */
532 for (i = 0; i < PAT_INDEX_SIZE; i++)
534 pat_table[PAT_WRITE_BACK] = 0;
535 pat_table[PAT_WRITE_THROUGH] = 1;
536 pat_table[PAT_UNCACHEABLE] = 3;
537 pat_table[PAT_WRITE_COMBINING] = 3;
538 pat_table[PAT_WRITE_PROTECTED] = 3;
539 pat_table[PAT_UNCACHED] = 3;
541 /* Bail if this CPU doesn't implement PAT. */
542 if ((cpu_feature & CPUID_PAT) == 0) {
543 for (i = 0; i < PAT_INDEX_SIZE; i++)
544 pat_index[i] = pat_table[i];
550 * Due to some Intel errata, we can only safely use the lower 4
553 * Intel Pentium III Processor Specification Update
554 * Errata E.27 (Upper Four PAT Entries Not Usable With Mode B
557 * Intel Pentium IV Processor Specification Update
558 * Errata N46 (PAT Index MSB May Be Calculated Incorrectly)
560 if (cpu_vendor_id == CPU_VENDOR_INTEL &&
561 !(CPUID_TO_FAMILY(cpu_id) == 6 && CPUID_TO_MODEL(cpu_id) >= 0xe))
564 /* Initialize default PAT entries. */
565 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |
566 PAT_VALUE(1, PAT_WRITE_THROUGH) |
567 PAT_VALUE(2, PAT_UNCACHED) |
568 PAT_VALUE(3, PAT_UNCACHEABLE) |
569 PAT_VALUE(4, PAT_WRITE_BACK) |
570 PAT_VALUE(5, PAT_WRITE_THROUGH) |
571 PAT_VALUE(6, PAT_UNCACHED) |
572 PAT_VALUE(7, PAT_UNCACHEABLE);
576 * Leave the indices 0-3 at the default of WB, WT, UC-, and UC.
577 * Program 5 and 6 as WP and WC.
578 * Leave 4 and 7 as WB and UC.
580 pat_msr &= ~(PAT_MASK(5) | PAT_MASK(6));
581 pat_msr |= PAT_VALUE(5, PAT_WRITE_PROTECTED) |
582 PAT_VALUE(6, PAT_WRITE_COMBINING);
583 pat_table[PAT_UNCACHED] = 2;
584 pat_table[PAT_WRITE_PROTECTED] = 5;
585 pat_table[PAT_WRITE_COMBINING] = 6;
588 * Just replace PAT Index 2 with WC instead of UC-.
590 pat_msr &= ~PAT_MASK(2);
591 pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING);
592 pat_table[PAT_WRITE_COMBINING] = 2;
597 load_cr4(cr4 & ~CR4_PGE);
599 /* Disable caches (CD = 1, NW = 0). */
601 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
603 /* Flushes caches and TLBs. */
607 /* Update PAT and index table. */
608 wrmsr(MSR_PAT, pat_msr);
609 for (i = 0; i < PAT_INDEX_SIZE; i++)
610 pat_index[i] = pat_table[i];
612 /* Flush caches and TLBs again. */
616 /* Restore caches and PGE. */
622 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on.
628 vm_offset_t va, endva;
633 endva = KERNBASE + KERNend;
636 va = KERNBASE + KERNLOAD;
638 pdir_pde(PTD, va) |= pgeflag;
639 invltlb(); /* Play it safe, invltlb() every time */
643 va = (vm_offset_t)btext;
648 invltlb(); /* Play it safe, invltlb() every time */
655 * Initialize a vm_page's machine-dependent fields.
658 pmap_page_init(vm_page_t m)
661 TAILQ_INIT(&m->md.pv_list);
662 m->md.pat_mode = PAT_WRITE_BACK;
667 pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
670 /* Inform UMA that this allocator uses kernel_map/object. */
671 *flags = UMA_SLAB_KERNEL;
672 return ((void *)kmem_alloc_contig(kernel_map, bytes, wait, 0x0ULL,
673 0xffffffffULL, 1, 0, VM_MEMATTR_DEFAULT));
678 * ABuse the pte nodes for unmapped kva to thread a kva freelist through.
680 * - Must deal with pages in order to ensure that none of the PG_* bits
681 * are ever set, PG_V in particular.
682 * - Assumes we can write to ptes without pte_store() atomic ops, even
683 * on PAE systems. This should be ok.
684 * - Assumes nothing will ever test these addresses for 0 to indicate
685 * no mapping instead of correctly checking PG_V.
686 * - Assumes a vm_offset_t will fit in a pte (true for i386).
687 * Because PG_V is never set, there can be no mappings to invalidate.
690 pmap_ptelist_alloc(vm_offset_t *head)
697 return (va); /* Out of memory */
701 panic("pmap_ptelist_alloc: va with PG_V set!");
707 pmap_ptelist_free(vm_offset_t *head, vm_offset_t va)
712 panic("pmap_ptelist_free: freeing va with PG_V set!");
714 *pte = *head; /* virtual! PG_V is 0 though */
719 pmap_ptelist_init(vm_offset_t *head, void *base, int npages)
725 for (i = npages - 1; i >= 0; i--) {
726 va = (vm_offset_t)base + i * PAGE_SIZE;
727 pmap_ptelist_free(head, va);
733 * Initialize the pmap module.
734 * Called by vm_init, to initialize any structures that the pmap
735 * system needs to map virtual memory.
745 * Initialize the vm page array entries for the kernel pmap's
748 for (i = 0; i < NKPT; i++) {
749 mpte = PHYS_TO_VM_PAGE(KPTphys + (i << PAGE_SHIFT));
750 KASSERT(mpte >= vm_page_array &&
751 mpte < &vm_page_array[vm_page_array_size],
752 ("pmap_init: page table page is out of range"));
753 mpte->pindex = i + KPTDI;
754 mpte->phys_addr = KPTphys + (i << PAGE_SHIFT);
758 * Initialize the address space (zone) for the pv entries. Set a
759 * high water mark so that the system can recover from excessive
760 * numbers of pv entries.
762 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
763 pv_entry_max = shpgperproc * maxproc + cnt.v_page_count;
764 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
765 pv_entry_max = roundup(pv_entry_max, _NPCPV);
766 pv_entry_high_water = 9 * (pv_entry_max / 10);
769 * If the kernel is running on a virtual machine, then it must assume
770 * that MCA is enabled by the hypervisor. Moreover, the kernel must
771 * be prepared for the hypervisor changing the vendor and family that
772 * are reported by CPUID. Consequently, the workaround for AMD Family
773 * 10h Erratum 383 is enabled if the processor's feature set does not
774 * include at least one feature that is only supported by older Intel
775 * or newer AMD processors.
777 if (vm_guest == VM_GUEST_VM && (cpu_feature & CPUID_SS) == 0 &&
778 (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI |
779 CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP |
781 workaround_erratum383 = 1;
784 * Are large page mappings supported and enabled?
786 TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled);
789 else if (pg_ps_enabled) {
790 KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
791 ("pmap_init: can't assign to pagesizes[1]"));
792 pagesizes[1] = NBPDR;
796 * Calculate the size of the pv head table for superpages.
798 for (i = 0; phys_avail[i + 1]; i += 2);
799 pv_npg = round_4mpage(phys_avail[(i - 2) + 1]) / NBPDR;
802 * Allocate memory for the pv head table for superpages.
804 s = (vm_size_t)(pv_npg * sizeof(struct md_page));
806 pv_table = (struct md_page *)kmem_alloc(kernel_map, s);
807 for (i = 0; i < pv_npg; i++)
808 TAILQ_INIT(&pv_table[i].pv_list);
810 pv_maxchunks = MAX(pv_entry_max / _NPCPV, maxproc);
811 pv_chunkbase = (struct pv_chunk *)kmem_alloc_nofault(kernel_map,
812 PAGE_SIZE * pv_maxchunks);
813 if (pv_chunkbase == NULL)
814 panic("pmap_init: not enough kvm for pv chunks");
815 pmap_ptelist_init(&pv_vafree, pv_chunkbase, pv_maxchunks);
817 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL,
818 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1,
819 UMA_ZONE_VM | UMA_ZONE_NOFREE);
820 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf);
825 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_max, CTLFLAG_RD, &pv_entry_max, 0,
826 "Max number of PV entries");
827 SYSCTL_INT(_vm_pmap, OID_AUTO, shpgperproc, CTLFLAG_RD, &shpgperproc, 0,
828 "Page share factor per proc");
830 static SYSCTL_NODE(_vm_pmap, OID_AUTO, pde, CTLFLAG_RD, 0,
831 "2/4MB page mapping counters");
833 static u_long pmap_pde_demotions;
834 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, demotions, CTLFLAG_RD,
835 &pmap_pde_demotions, 0, "2/4MB page demotions");
837 static u_long pmap_pde_mappings;
838 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, mappings, CTLFLAG_RD,
839 &pmap_pde_mappings, 0, "2/4MB page mappings");
841 static u_long pmap_pde_p_failures;
842 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, p_failures, CTLFLAG_RD,
843 &pmap_pde_p_failures, 0, "2/4MB page promotion failures");
845 static u_long pmap_pde_promotions;
846 SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, promotions, CTLFLAG_RD,
847 &pmap_pde_promotions, 0, "2/4MB page promotions");
849 /***************************************************
850 * Low level helper routines.....
851 ***************************************************/
854 * Determine the appropriate bits to set in a PTE or PDE for a specified
858 pmap_cache_bits(int mode, boolean_t is_pde)
860 int cache_bits, pat_flag, pat_idx;
862 if (mode < 0 || mode >= PAT_INDEX_SIZE || pat_index[mode] < 0)
863 panic("Unknown caching mode %d\n", mode);
865 /* The PAT bit is different for PTE's and PDE's. */
866 pat_flag = is_pde ? PG_PDE_PAT : PG_PTE_PAT;
868 /* Map the caching mode to a PAT index. */
869 pat_idx = pat_index[mode];
871 /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */
874 cache_bits |= pat_flag;
876 cache_bits |= PG_NC_PCD;
878 cache_bits |= PG_NC_PWT;
883 * The caller is responsible for maintaining TLB consistency.
886 pmap_kenter_pde(vm_offset_t va, pd_entry_t newpde)
890 boolean_t PTD_updated;
893 mtx_lock_spin(&allpmaps_lock);
894 LIST_FOREACH(pmap, &allpmaps, pm_list) {
895 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] &
898 pde = pmap_pde(pmap, va);
899 pde_store(pde, newpde);
901 mtx_unlock_spin(&allpmaps_lock);
903 ("pmap_kenter_pde: current page table is not in allpmaps"));
907 * After changing the page size for the specified virtual address in the page
908 * table, flush the corresponding entries from the processor's TLB. Only the
909 * calling processor's TLB is affected.
911 * The calling thread must be pinned to a processor.
914 pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde)
918 if ((newpde & PG_PS) == 0)
919 /* Demotion: flush a specific 2MB page mapping. */
921 else if ((newpde & PG_G) == 0)
923 * Promotion: flush every 4KB page mapping from the TLB
924 * because there are too many to flush individually.
929 * Promotion: flush every 4KB page mapping from the TLB,
930 * including any global (PG_G) mappings.
933 load_cr4(cr4 & ~CR4_PGE);
935 * Although preemption at this point could be detrimental to
936 * performance, it would not lead to an error. PG_G is simply
937 * ignored if CR4.PGE is clear. Moreover, in case this block
938 * is re-entered, the load_cr4() either above or below will
939 * modify CR4.PGE flushing the TLB.
941 load_cr4(cr4 | CR4_PGE);
946 * For SMP, these functions have to use the IPI mechanism for coherence.
948 * N.B.: Before calling any of the following TLB invalidation functions,
949 * the calling processor must ensure that all stores updating a non-
950 * kernel page table are globally performed. Otherwise, another
951 * processor could cache an old, pre-update entry without being
952 * invalidated. This can happen one of two ways: (1) The pmap becomes
953 * active on another processor after its pm_active field is checked by
954 * one of the following functions but before a store updating the page
955 * table is globally performed. (2) The pmap becomes active on another
956 * processor before its pm_active field is checked but due to
957 * speculative loads one of the following functions stills reads the
958 * pmap as inactive on the other processor.
960 * The kernel page table is exempt because its pm_active field is
961 * immutable. The kernel page table is always active on every
965 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
971 if (pmap == kernel_pmap || !CPU_CMP(&pmap->pm_active, &all_cpus)) {
975 cpuid = PCPU_GET(cpuid);
976 other_cpus = all_cpus;
977 CPU_CLR(cpuid, &other_cpus);
978 if (CPU_ISSET(cpuid, &pmap->pm_active))
980 CPU_AND(&other_cpus, &pmap->pm_active);
981 if (!CPU_EMPTY(&other_cpus))
982 smp_masked_invlpg(other_cpus, va);
988 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
995 if (pmap == kernel_pmap || !CPU_CMP(&pmap->pm_active, &all_cpus)) {
996 for (addr = sva; addr < eva; addr += PAGE_SIZE)
998 smp_invlpg_range(sva, eva);
1000 cpuid = PCPU_GET(cpuid);
1001 other_cpus = all_cpus;
1002 CPU_CLR(cpuid, &other_cpus);
1003 if (CPU_ISSET(cpuid, &pmap->pm_active))
1004 for (addr = sva; addr < eva; addr += PAGE_SIZE)
1006 CPU_AND(&other_cpus, &pmap->pm_active);
1007 if (!CPU_EMPTY(&other_cpus))
1008 smp_masked_invlpg_range(other_cpus, sva, eva);
1014 pmap_invalidate_all(pmap_t pmap)
1016 cpuset_t other_cpus;
1020 if (pmap == kernel_pmap || !CPU_CMP(&pmap->pm_active, &all_cpus)) {
1024 cpuid = PCPU_GET(cpuid);
1025 other_cpus = all_cpus;
1026 CPU_CLR(cpuid, &other_cpus);
1027 if (CPU_ISSET(cpuid, &pmap->pm_active))
1029 CPU_AND(&other_cpus, &pmap->pm_active);
1030 if (!CPU_EMPTY(&other_cpus))
1031 smp_masked_invltlb(other_cpus);
1037 pmap_invalidate_cache(void)
1047 cpuset_t invalidate; /* processors that invalidate their TLB */
1051 u_int store; /* processor that updates the PDE */
1055 pmap_update_pde_kernel(void *arg)
1057 struct pde_action *act = arg;
1061 if (act->store == PCPU_GET(cpuid)) {
1064 * Elsewhere, this operation requires allpmaps_lock for
1065 * synchronization. Here, it does not because it is being
1066 * performed in the context of an all_cpus rendezvous.
1068 LIST_FOREACH(pmap, &allpmaps, pm_list) {
1069 pde = pmap_pde(pmap, act->va);
1070 pde_store(pde, act->newpde);
1076 pmap_update_pde_user(void *arg)
1078 struct pde_action *act = arg;
1080 if (act->store == PCPU_GET(cpuid))
1081 pde_store(act->pde, act->newpde);
1085 pmap_update_pde_teardown(void *arg)
1087 struct pde_action *act = arg;
1089 if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate))
1090 pmap_update_pde_invalidate(act->va, act->newpde);
1094 * Change the page size for the specified virtual address in a way that
1095 * prevents any possibility of the TLB ever having two entries that map the
1096 * same virtual address using different page sizes. This is the recommended
1097 * workaround for Erratum 383 on AMD Family 10h processors. It prevents a
1098 * machine check exception for a TLB state that is improperly diagnosed as a
1102 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
1104 struct pde_action act;
1105 cpuset_t active, other_cpus;
1109 cpuid = PCPU_GET(cpuid);
1110 other_cpus = all_cpus;
1111 CPU_CLR(cpuid, &other_cpus);
1112 if (pmap == kernel_pmap)
1115 active = pmap->pm_active;
1116 if (CPU_OVERLAP(&active, &other_cpus)) {
1118 act.invalidate = active;
1121 act.newpde = newpde;
1122 CPU_SET(cpuid, &active);
1123 smp_rendezvous_cpus(active,
1124 smp_no_rendevous_barrier, pmap == kernel_pmap ?
1125 pmap_update_pde_kernel : pmap_update_pde_user,
1126 pmap_update_pde_teardown, &act);
1128 if (pmap == kernel_pmap)
1129 pmap_kenter_pde(va, newpde);
1131 pde_store(pde, newpde);
1132 if (CPU_ISSET(cpuid, &active))
1133 pmap_update_pde_invalidate(va, newpde);
1139 * Normal, non-SMP, 486+ invalidation functions.
1140 * We inline these within pmap.c for speed.
1143 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
1146 if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active))
1151 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1155 if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active))
1156 for (addr = sva; addr < eva; addr += PAGE_SIZE)
1161 pmap_invalidate_all(pmap_t pmap)
1164 if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active))
1169 pmap_invalidate_cache(void)
1176 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
1179 if (pmap == kernel_pmap)
1180 pmap_kenter_pde(va, newpde);
1182 pde_store(pde, newpde);
1183 if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active))
1184 pmap_update_pde_invalidate(va, newpde);
1188 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1191 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1194 KASSERT((sva & PAGE_MASK) == 0,
1195 ("pmap_invalidate_cache_range: sva not page-aligned"));
1196 KASSERT((eva & PAGE_MASK) == 0,
1197 ("pmap_invalidate_cache_range: eva not page-aligned"));
1199 if (cpu_feature & CPUID_SS)
1200 ; /* If "Self Snoop" is supported, do nothing. */
1201 else if ((cpu_feature & CPUID_CLFSH) != 0 &&
1202 eva - sva < PMAP_CLFLUSH_THRESHOLD) {
1206 * XXX: Some CPUs fault, hang, or trash the local APIC
1207 * registers if we use CLFLUSH on the local APIC
1208 * range. The local APIC is always uncached, so we
1209 * don't need to flush for that range anyway.
1211 if (pmap_kextract(sva) == lapic_paddr)
1215 * Otherwise, do per-cache line flush. Use the mfence
1216 * instruction to insure that previous stores are
1217 * included in the write-back. The processor
1218 * propagates flush to other processors in the cache
1222 for (; sva < eva; sva += cpu_clflush_line_size)
1228 * No targeted cache flush methods are supported by CPU,
1229 * or the supplied range is bigger than 2MB.
1230 * Globally invalidate cache.
1232 pmap_invalidate_cache();
1237 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1241 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1242 (cpu_feature & CPUID_CLFSH) == 0) {
1243 pmap_invalidate_cache();
1245 for (i = 0; i < count; i++)
1246 pmap_flush_page(pages[i]);
1251 * Are we current address space or kernel? N.B. We return FALSE when
1252 * a pmap's page table is in use because a kernel thread is borrowing
1253 * it. The borrowed page table can change spontaneously, making any
1254 * dependence on its continued use subject to a race condition.
1257 pmap_is_current(pmap_t pmap)
1260 return (pmap == kernel_pmap ||
1261 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) &&
1262 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)));
1266 * If the given pmap is not the current or kernel pmap, the returned pte must
1267 * be released by passing it to pmap_pte_release().
1270 pmap_pte(pmap_t pmap, vm_offset_t va)
1275 pde = pmap_pde(pmap, va);
1279 /* are we current address space or kernel? */
1280 if (pmap_is_current(pmap))
1281 return (vtopte(va));
1282 mtx_lock(&PMAP2mutex);
1283 newpf = *pde & PG_FRAME;
1284 if ((*PMAP2 & PG_FRAME) != newpf) {
1285 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M;
1286 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2);
1288 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1)));
1294 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte
1297 static __inline void
1298 pmap_pte_release(pt_entry_t *pte)
1301 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2)
1302 mtx_unlock(&PMAP2mutex);
1306 * NB: The sequence of updating a page table followed by accesses to the
1307 * corresponding pages is subject to the situation described in the "AMD64
1308 * Architecture Programmer's Manual Volume 2: System Programming" rev. 3.23,
1309 * "7.3.1 Special Coherency Considerations". Therefore, issuing the INVLPG
1310 * right after modifying the PTE bits is crucial.
1312 static __inline void
1313 invlcaddr(void *caddr)
1316 invlpg((u_int)caddr);
1320 * Super fast pmap_pte routine best used when scanning
1321 * the pv lists. This eliminates many coarse-grained
1322 * invltlb calls. Note that many of the pv list
1323 * scans are across different pmaps. It is very wasteful
1324 * to do an entire invltlb for checking a single mapping.
1326 * If the given pmap is not the current pmap, pvh_global_lock
1327 * must be held and curthread pinned to a CPU.
1330 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
1335 pde = pmap_pde(pmap, va);
1339 /* are we current address space or kernel? */
1340 if (pmap_is_current(pmap))
1341 return (vtopte(va));
1342 rw_assert(&pvh_global_lock, RA_WLOCKED);
1343 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
1344 newpf = *pde & PG_FRAME;
1345 if ((*PMAP1 & PG_FRAME) != newpf) {
1346 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M;
1348 PMAP1cpu = PCPU_GET(cpuid);
1354 if (PMAP1cpu != PCPU_GET(cpuid)) {
1355 PMAP1cpu = PCPU_GET(cpuid);
1361 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1)));
1367 * Routine: pmap_extract
1369 * Extract the physical page address associated
1370 * with the given map/virtual_address pair.
1373 pmap_extract(pmap_t pmap, vm_offset_t va)
1381 pde = pmap->pm_pdir[va >> PDRSHIFT];
1383 if ((pde & PG_PS) != 0)
1384 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
1386 pte = pmap_pte(pmap, va);
1387 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
1388 pmap_pte_release(pte);
1396 * Routine: pmap_extract_and_hold
1398 * Atomically extract and hold the physical page
1399 * with the given pmap and virtual address pair
1400 * if that mapping permits the given protection.
1403 pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1406 pt_entry_t pte, *ptep;
1414 pde = *pmap_pde(pmap, va);
1417 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) {
1418 if (vm_page_pa_tryrelock(pmap, (pde &
1419 PG_PS_FRAME) | (va & PDRMASK), &pa))
1421 m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) |
1426 ptep = pmap_pte(pmap, va);
1428 pmap_pte_release(ptep);
1430 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) {
1431 if (vm_page_pa_tryrelock(pmap, pte & PG_FRAME,
1434 m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
1444 /***************************************************
1445 * Low level mapping routines.....
1446 ***************************************************/
1449 * Add a wired page to the kva.
1450 * Note: not SMP coherent.
1452 * This function may be used before pmap_bootstrap() is called.
1455 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1460 pte_store(pte, pa | PG_RW | PG_V | pgeflag);
1463 static __inline void
1464 pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode)
1469 pte_store(pte, pa | PG_RW | PG_V | pgeflag | pmap_cache_bits(mode, 0));
1473 * Remove a page from the kernel pagetables.
1474 * Note: not SMP coherent.
1476 * This function may be used before pmap_bootstrap() is called.
1479 pmap_kremove(vm_offset_t va)
1488 * Used to map a range of physical addresses into kernel
1489 * virtual address space.
1491 * The value passed in '*virt' is a suggested virtual address for
1492 * the mapping. Architectures which can support a direct-mapped
1493 * physical to virtual region can return the appropriate address
1494 * within that region, leaving '*virt' unchanged. Other
1495 * architectures should map the pages starting at '*virt' and
1496 * update '*virt' with the first usable address after the mapped
1500 pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot)
1502 vm_offset_t va, sva;
1503 vm_paddr_t superpage_offset;
1508 * Does the physical address range's size and alignment permit at
1509 * least one superpage mapping to be created?
1511 superpage_offset = start & PDRMASK;
1512 if ((end - start) - ((NBPDR - superpage_offset) & PDRMASK) >= NBPDR) {
1514 * Increase the starting virtual address so that its alignment
1515 * does not preclude the use of superpage mappings.
1517 if ((va & PDRMASK) < superpage_offset)
1518 va = (va & ~PDRMASK) + superpage_offset;
1519 else if ((va & PDRMASK) > superpage_offset)
1520 va = ((va + PDRMASK) & ~PDRMASK) + superpage_offset;
1523 while (start < end) {
1524 if ((start & PDRMASK) == 0 && end - start >= NBPDR &&
1526 KASSERT((va & PDRMASK) == 0,
1527 ("pmap_map: misaligned va %#x", va));
1528 newpde = start | PG_PS | pgeflag | PG_RW | PG_V;
1529 pmap_kenter_pde(va, newpde);
1533 pmap_kenter(va, start);
1538 pmap_invalidate_range(kernel_pmap, sva, va);
1545 * Add a list of wired pages to the kva
1546 * this routine is only used for temporary
1547 * kernel mappings that do not need to have
1548 * page modification or references recorded.
1549 * Note that old mappings are simply written
1550 * over. The page *must* be wired.
1551 * Note: SMP coherent. Uses a ranged shootdown IPI.
1554 pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count)
1556 pt_entry_t *endpte, oldpte, pa, *pte;
1561 endpte = pte + count;
1562 while (pte < endpte) {
1564 pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.pat_mode, 0);
1565 if ((*pte & (PG_FRAME | PG_PTE_CACHE)) != pa) {
1567 pte_store(pte, pa | pgeflag | PG_RW | PG_V);
1571 if (__predict_false((oldpte & PG_V) != 0))
1572 pmap_invalidate_range(kernel_pmap, sva, sva + count *
1577 * This routine tears out page mappings from the
1578 * kernel -- it is meant only for temporary mappings.
1579 * Note: SMP coherent. Uses a ranged shootdown IPI.
1582 pmap_qremove(vm_offset_t sva, int count)
1587 while (count-- > 0) {
1591 pmap_invalidate_range(kernel_pmap, sva, va);
1594 /***************************************************
1595 * Page table page management routines.....
1596 ***************************************************/
1597 static __inline void
1598 pmap_free_zero_pages(vm_page_t free)
1602 while (free != NULL) {
1605 /* Preserve the page's PG_ZERO setting. */
1606 vm_page_free_toq(m);
1611 * Schedule the specified unused page table page to be freed. Specifically,
1612 * add the page to the specified list of pages that will be released to the
1613 * physical memory manager after the TLB has been updated.
1615 static __inline void
1616 pmap_add_delayed_free_list(vm_page_t m, vm_page_t *free, boolean_t set_PG_ZERO)
1620 m->flags |= PG_ZERO;
1622 m->flags &= ~PG_ZERO;
1628 * Inserts the specified page table page into the specified pmap's collection
1629 * of idle page table pages. Each of a pmap's page table pages is responsible
1630 * for mapping a distinct range of virtual addresses. The pmap's collection is
1631 * ordered by this virtual address range.
1634 pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte)
1638 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1639 root = pmap->pm_root;
1644 root = vm_page_splay(mpte->pindex, root);
1645 if (mpte->pindex < root->pindex) {
1646 mpte->left = root->left;
1649 } else if (mpte->pindex == root->pindex)
1650 panic("pmap_insert_pt_page: pindex already inserted");
1652 mpte->right = root->right;
1657 pmap->pm_root = mpte;
1661 * Looks for a page table page mapping the specified virtual address in the
1662 * specified pmap's collection of idle page table pages. Returns NULL if there
1663 * is no page table page corresponding to the specified virtual address.
1666 pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va)
1669 vm_pindex_t pindex = va >> PDRSHIFT;
1671 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1672 if ((mpte = pmap->pm_root) != NULL && mpte->pindex != pindex) {
1673 mpte = vm_page_splay(pindex, mpte);
1674 if ((pmap->pm_root = mpte)->pindex != pindex)
1681 * Removes the specified page table page from the specified pmap's collection
1682 * of idle page table pages. The specified page table page must be a member of
1683 * the pmap's collection.
1686 pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte)
1690 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1691 if (mpte != pmap->pm_root)
1692 vm_page_splay(mpte->pindex, pmap->pm_root);
1693 if (mpte->left == NULL)
1696 root = vm_page_splay(mpte->pindex, mpte->left);
1697 root->right = mpte->right;
1699 pmap->pm_root = root;
1703 * Decrements a page table page's wire count, which is used to record the
1704 * number of valid page table entries within the page. If the wire count
1705 * drops to zero, then the page table page is unmapped. Returns TRUE if the
1706 * page table page was unmapped and FALSE otherwise.
1708 static inline boolean_t
1709 pmap_unwire_ptp(pmap_t pmap, vm_page_t m, vm_page_t *free)
1713 if (m->wire_count == 0) {
1714 _pmap_unwire_ptp(pmap, m, free);
1721 _pmap_unwire_ptp(pmap_t pmap, vm_page_t m, vm_page_t *free)
1726 * unmap the page table page
1728 pmap->pm_pdir[m->pindex] = 0;
1729 --pmap->pm_stats.resident_count;
1732 * This is a release store so that the ordinary store unmapping
1733 * the page table page is globally performed before TLB shoot-
1736 atomic_subtract_rel_int(&cnt.v_wire_count, 1);
1739 * Do an invltlb to make the invalidated mapping
1740 * take effect immediately.
1742 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex);
1743 pmap_invalidate_page(pmap, pteva);
1746 * Put page on a list so that it is released after
1747 * *ALL* TLB shootdown is done
1749 pmap_add_delayed_free_list(m, free, TRUE);
1753 * After removing a page table entry, this routine is used to
1754 * conditionally free the page, and manage the hold/wire counts.
1757 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t *free)
1762 if (va >= VM_MAXUSER_ADDRESS)
1764 ptepde = *pmap_pde(pmap, va);
1765 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME);
1766 return (pmap_unwire_ptp(pmap, mpte, free));
1770 * Initialize the pmap for the swapper process.
1773 pmap_pinit0(pmap_t pmap)
1776 PMAP_LOCK_INIT(pmap);
1778 * Since the page table directory is shared with the kernel pmap,
1779 * which is already included in the list "allpmaps", this pmap does
1780 * not need to be inserted into that list.
1782 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD);
1784 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT);
1786 pmap->pm_root = NULL;
1787 CPU_ZERO(&pmap->pm_active);
1788 PCPU_SET(curpmap, pmap);
1789 TAILQ_INIT(&pmap->pm_pvchunk);
1790 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1794 * Initialize a preallocated and zeroed pmap structure,
1795 * such as one in a vmspace structure.
1798 pmap_pinit(pmap_t pmap)
1800 vm_page_t m, ptdpg[NPGPTD];
1804 PMAP_LOCK_INIT(pmap);
1807 * No need to allocate page table space yet but we do need a valid
1808 * page directory table.
1810 if (pmap->pm_pdir == NULL) {
1811 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map,
1813 if (pmap->pm_pdir == NULL) {
1814 PMAP_LOCK_DESTROY(pmap);
1818 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO);
1819 KASSERT(((vm_offset_t)pmap->pm_pdpt &
1820 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0,
1821 ("pmap_pinit: pdpt misaligned"));
1822 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30),
1823 ("pmap_pinit: pdpt above 4g"));
1825 pmap->pm_root = NULL;
1827 KASSERT(pmap->pm_root == NULL,
1828 ("pmap_pinit: pmap has reserved page table page(s)"));
1831 * allocate the page directory page(s)
1833 for (i = 0; i < NPGPTD;) {
1834 m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
1835 VM_ALLOC_WIRED | VM_ALLOC_ZERO);
1843 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD);
1845 for (i = 0; i < NPGPTD; i++)
1846 if ((ptdpg[i]->flags & PG_ZERO) == 0)
1847 pagezero(pmap->pm_pdir + (i * NPDEPG));
1849 mtx_lock_spin(&allpmaps_lock);
1850 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1851 /* Copy the kernel page table directory entries. */
1852 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t));
1853 mtx_unlock_spin(&allpmaps_lock);
1855 /* install self-referential address mapping entry(s) */
1856 for (i = 0; i < NPGPTD; i++) {
1857 pa = VM_PAGE_TO_PHYS(ptdpg[i]);
1858 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M;
1860 pmap->pm_pdpt[i] = pa | PG_V;
1864 CPU_ZERO(&pmap->pm_active);
1865 TAILQ_INIT(&pmap->pm_pvchunk);
1866 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1872 * this routine is called if the page table page is not
1876 _pmap_allocpte(pmap_t pmap, u_int ptepindex, int flags)
1881 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT ||
1882 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK,
1883 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK"));
1886 * Allocate a page table page.
1888 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ |
1889 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
1890 if (flags & M_WAITOK) {
1892 rw_wunlock(&pvh_global_lock);
1894 rw_wlock(&pvh_global_lock);
1899 * Indicate the need to retry. While waiting, the page table
1900 * page may have been allocated.
1904 if ((m->flags & PG_ZERO) == 0)
1908 * Map the pagetable page into the process address space, if
1909 * it isn't already there.
1912 pmap->pm_stats.resident_count++;
1914 ptepa = VM_PAGE_TO_PHYS(m);
1915 pmap->pm_pdir[ptepindex] =
1916 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1922 pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags)
1928 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT ||
1929 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK,
1930 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK"));
1933 * Calculate pagetable page index
1935 ptepindex = va >> PDRSHIFT;
1938 * Get the page directory entry
1940 ptepa = pmap->pm_pdir[ptepindex];
1943 * This supports switching from a 4MB page to a
1946 if (ptepa & PG_PS) {
1947 (void)pmap_demote_pde(pmap, &pmap->pm_pdir[ptepindex], va);
1948 ptepa = pmap->pm_pdir[ptepindex];
1952 * If the page table page is mapped, we just increment the
1953 * hold count, and activate it.
1956 m = PHYS_TO_VM_PAGE(ptepa & PG_FRAME);
1960 * Here if the pte page isn't mapped, or if it has
1963 m = _pmap_allocpte(pmap, ptepindex, flags);
1964 if (m == NULL && (flags & M_WAITOK))
1971 /***************************************************
1972 * Pmap allocation/deallocation routines.
1973 ***************************************************/
1977 * Deal with a SMP shootdown of other users of the pmap that we are
1978 * trying to dispose of. This can be a bit hairy.
1980 static cpuset_t *lazymask;
1981 static u_int lazyptd;
1982 static volatile u_int lazywait;
1984 void pmap_lazyfix_action(void);
1987 pmap_lazyfix_action(void)
1991 (*ipi_lazypmap_counts[PCPU_GET(cpuid)])++;
1993 if (rcr3() == lazyptd)
1994 load_cr3(curpcb->pcb_cr3);
1995 CPU_CLR_ATOMIC(PCPU_GET(cpuid), lazymask);
1996 atomic_store_rel_int(&lazywait, 1);
2000 pmap_lazyfix_self(u_int cpuid)
2003 if (rcr3() == lazyptd)
2004 load_cr3(curpcb->pcb_cr3);
2005 CPU_CLR_ATOMIC(cpuid, lazymask);
2010 pmap_lazyfix(pmap_t pmap)
2012 cpuset_t mymask, mask;
2016 mask = pmap->pm_active;
2017 while (!CPU_EMPTY(&mask)) {
2020 /* Find least significant set bit. */
2021 lsb = CPU_FFS(&mask);
2024 CPU_SETOF(lsb, &mask);
2025 mtx_lock_spin(&smp_ipi_mtx);
2027 lazyptd = vtophys(pmap->pm_pdpt);
2029 lazyptd = vtophys(pmap->pm_pdir);
2031 cpuid = PCPU_GET(cpuid);
2033 /* Use a cpuset just for having an easy check. */
2034 CPU_SETOF(cpuid, &mymask);
2035 if (!CPU_CMP(&mask, &mymask)) {
2036 lazymask = &pmap->pm_active;
2037 pmap_lazyfix_self(cpuid);
2039 atomic_store_rel_int((u_int *)&lazymask,
2040 (u_int)&pmap->pm_active);
2041 atomic_store_rel_int(&lazywait, 0);
2042 ipi_selected(mask, IPI_LAZYPMAP);
2043 while (lazywait == 0) {
2049 mtx_unlock_spin(&smp_ipi_mtx);
2051 printf("pmap_lazyfix: spun for 50000000\n");
2052 mask = pmap->pm_active;
2059 * Cleaning up on uniprocessor is easy. For various reasons, we're
2060 * unlikely to have to even execute this code, including the fact
2061 * that the cleanup is deferred until the parent does a wait(2), which
2062 * means that another userland process has run.
2065 pmap_lazyfix(pmap_t pmap)
2069 cr3 = vtophys(pmap->pm_pdir);
2070 if (cr3 == rcr3()) {
2071 load_cr3(curpcb->pcb_cr3);
2072 CPU_CLR(PCPU_GET(cpuid), &pmap->pm_active);
2078 * Release any resources held by the given physical map.
2079 * Called when a pmap initialized by pmap_pinit is being released.
2080 * Should only be called if the map contains no valid mappings.
2083 pmap_release(pmap_t pmap)
2085 vm_page_t m, ptdpg[NPGPTD];
2088 KASSERT(pmap->pm_stats.resident_count == 0,
2089 ("pmap_release: pmap resident count %ld != 0",
2090 pmap->pm_stats.resident_count));
2091 KASSERT(pmap->pm_root == NULL,
2092 ("pmap_release: pmap has reserved page table page(s)"));
2095 mtx_lock_spin(&allpmaps_lock);
2096 LIST_REMOVE(pmap, pm_list);
2097 mtx_unlock_spin(&allpmaps_lock);
2099 for (i = 0; i < NPGPTD; i++)
2100 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i] &
2103 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) *
2104 sizeof(*pmap->pm_pdir));
2106 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD);
2108 for (i = 0; i < NPGPTD; i++) {
2111 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME),
2112 ("pmap_release: got wrong ptd page"));
2115 atomic_subtract_int(&cnt.v_wire_count, 1);
2116 vm_page_free_zero(m);
2118 PMAP_LOCK_DESTROY(pmap);
2122 kvm_size(SYSCTL_HANDLER_ARGS)
2124 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
2126 return (sysctl_handle_long(oidp, &ksize, 0, req));
2128 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
2129 0, 0, kvm_size, "IU", "Size of KVM");
2132 kvm_free(SYSCTL_HANDLER_ARGS)
2134 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
2136 return (sysctl_handle_long(oidp, &kfree, 0, req));
2138 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
2139 0, 0, kvm_free, "IU", "Amount of KVM free");
2142 * grow the number of kernel page table entries, if needed
2145 pmap_growkernel(vm_offset_t addr)
2147 vm_paddr_t ptppaddr;
2151 mtx_assert(&kernel_map->system_mtx, MA_OWNED);
2152 addr = roundup2(addr, NBPDR);
2153 if (addr - 1 >= kernel_map->max_offset)
2154 addr = kernel_map->max_offset;
2155 while (kernel_vm_end < addr) {
2156 if (pdir_pde(PTD, kernel_vm_end)) {
2157 kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK;
2158 if (kernel_vm_end - 1 >= kernel_map->max_offset) {
2159 kernel_vm_end = kernel_map->max_offset;
2165 nkpg = vm_page_alloc(NULL, kernel_vm_end >> PDRSHIFT,
2166 VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
2169 panic("pmap_growkernel: no memory to grow kernel");
2173 if ((nkpg->flags & PG_ZERO) == 0)
2174 pmap_zero_page(nkpg);
2175 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2176 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2177 pdir_pde(KPTD, kernel_vm_end) = pgeflag | newpdir;
2179 pmap_kenter_pde(kernel_vm_end, newpdir);
2180 kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK;
2181 if (kernel_vm_end - 1 >= kernel_map->max_offset) {
2182 kernel_vm_end = kernel_map->max_offset;
2189 /***************************************************
2190 * page management routines.
2191 ***************************************************/
2193 CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE);
2194 CTASSERT(_NPCM == 11);
2195 CTASSERT(_NPCPV == 336);
2197 static __inline struct pv_chunk *
2198 pv_to_chunk(pv_entry_t pv)
2201 return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK));
2204 #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap)
2206 #define PC_FREE0_9 0xfffffffful /* Free values for index 0 through 9 */
2207 #define PC_FREE10 0x0000fffful /* Free values for index 10 */
2209 static const uint32_t pc_freemask[_NPCM] = {
2210 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9,
2211 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9,
2212 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9,
2213 PC_FREE0_9, PC_FREE10
2216 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0,
2217 "Current number of pv entries");
2220 static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail;
2222 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0,
2223 "Current number of pv entry chunks");
2224 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0,
2225 "Current number of pv entry chunks allocated");
2226 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0,
2227 "Current number of pv entry chunks frees");
2228 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0,
2229 "Number of times tried to get a chunk page but failed.");
2231 static long pv_entry_frees, pv_entry_allocs;
2232 static int pv_entry_spare;
2234 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0,
2235 "Current number of pv entry frees");
2236 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0,
2237 "Current number of pv entry allocs");
2238 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0,
2239 "Current number of spare pv entries");
2243 * We are in a serious low memory condition. Resort to
2244 * drastic measures to free some pages so we can allocate
2245 * another pv entry chunk.
2248 pmap_pv_reclaim(pmap_t locked_pmap)
2251 struct pv_chunk *pc;
2252 struct md_page *pvh;
2255 pt_entry_t *pte, tpte;
2258 vm_page_t free, m, m_pc;
2260 int bit, field, freed;
2262 PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED);
2265 TAILQ_INIT(&newtail);
2266 while ((pc = TAILQ_FIRST(&pv_chunks)) != NULL && (pv_vafree == 0 ||
2268 TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
2269 if (pmap != pc->pc_pmap) {
2271 pmap_invalidate_all(pmap);
2272 if (pmap != locked_pmap)
2276 /* Avoid deadlock and lock recursion. */
2277 if (pmap > locked_pmap)
2279 else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) {
2281 TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2287 * Destroy every non-wired, 4 KB page mapping in the chunk.
2290 for (field = 0; field < _NPCM; field++) {
2291 for (inuse = ~pc->pc_map[field] & pc_freemask[field];
2292 inuse != 0; inuse &= ~(1UL << bit)) {
2294 pv = &pc->pc_pventry[field * 32 + bit];
2296 pde = pmap_pde(pmap, va);
2297 if ((*pde & PG_PS) != 0)
2299 pte = pmap_pte(pmap, va);
2301 if ((tpte & PG_W) == 0)
2302 tpte = pte_load_clear(pte);
2303 pmap_pte_release(pte);
2304 if ((tpte & PG_W) != 0)
2307 ("pmap_pv_reclaim: pmap %p va %x zero pte",
2309 if ((tpte & PG_G) != 0)
2310 pmap_invalidate_page(pmap, va);
2311 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
2312 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
2314 if ((tpte & PG_A) != 0)
2315 vm_page_aflag_set(m, PGA_REFERENCED);
2316 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2317 if (TAILQ_EMPTY(&m->md.pv_list) &&
2318 (m->flags & PG_FICTITIOUS) == 0) {
2319 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
2320 if (TAILQ_EMPTY(&pvh->pv_list)) {
2321 vm_page_aflag_clear(m,
2325 pc->pc_map[field] |= 1UL << bit;
2326 pmap_unuse_pt(pmap, va, &free);
2331 TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2334 /* Every freed mapping is for a 4 KB page. */
2335 pmap->pm_stats.resident_count -= freed;
2336 PV_STAT(pv_entry_frees += freed);
2337 PV_STAT(pv_entry_spare += freed);
2338 pv_entry_count -= freed;
2339 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2340 for (field = 0; field < _NPCM; field++)
2341 if (pc->pc_map[field] != pc_freemask[field]) {
2342 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc,
2344 TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2347 * One freed pv entry in locked_pmap is
2350 if (pmap == locked_pmap)
2354 if (field == _NPCM) {
2355 PV_STAT(pv_entry_spare -= _NPCPV);
2356 PV_STAT(pc_chunk_count--);
2357 PV_STAT(pc_chunk_frees++);
2358 /* Entire chunk is free; return it. */
2359 m_pc = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc));
2360 pmap_qremove((vm_offset_t)pc, 1);
2361 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc);
2366 TAILQ_CONCAT(&pv_chunks, &newtail, pc_lru);
2368 pmap_invalidate_all(pmap);
2369 if (pmap != locked_pmap)
2372 if (m_pc == NULL && pv_vafree != 0 && free != NULL) {
2375 /* Recycle a freed page table page. */
2376 m_pc->wire_count = 1;
2377 atomic_add_int(&cnt.v_wire_count, 1);
2379 pmap_free_zero_pages(free);
2384 * free the pv_entry back to the free list
2387 free_pv_entry(pmap_t pmap, pv_entry_t pv)
2389 struct pv_chunk *pc;
2390 int idx, field, bit;
2392 rw_assert(&pvh_global_lock, RA_WLOCKED);
2393 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2394 PV_STAT(pv_entry_frees++);
2395 PV_STAT(pv_entry_spare++);
2397 pc = pv_to_chunk(pv);
2398 idx = pv - &pc->pc_pventry[0];
2401 pc->pc_map[field] |= 1ul << bit;
2402 for (idx = 0; idx < _NPCM; idx++)
2403 if (pc->pc_map[idx] != pc_freemask[idx]) {
2405 * 98% of the time, pc is already at the head of the
2406 * list. If it isn't already, move it to the head.
2408 if (__predict_false(TAILQ_FIRST(&pmap->pm_pvchunk) !=
2410 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2411 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc,
2416 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2421 free_pv_chunk(struct pv_chunk *pc)
2425 TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
2426 PV_STAT(pv_entry_spare -= _NPCPV);
2427 PV_STAT(pc_chunk_count--);
2428 PV_STAT(pc_chunk_frees++);
2429 /* entire chunk is free, return it */
2430 m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc));
2431 pmap_qremove((vm_offset_t)pc, 1);
2432 vm_page_unwire(m, 0);
2434 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc);
2438 * get a new pv_entry, allocating a block from the system
2442 get_pv_entry(pmap_t pmap, boolean_t try)
2444 static const struct timeval printinterval = { 60, 0 };
2445 static struct timeval lastprint;
2448 struct pv_chunk *pc;
2451 rw_assert(&pvh_global_lock, RA_WLOCKED);
2452 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2453 PV_STAT(pv_entry_allocs++);
2455 if (pv_entry_count > pv_entry_high_water)
2456 if (ratecheck(&lastprint, &printinterval))
2457 printf("Approaching the limit on PV entries, consider "
2458 "increasing either the vm.pmap.shpgperproc or the "
2459 "vm.pmap.pv_entry_max tunable.\n");
2461 pc = TAILQ_FIRST(&pmap->pm_pvchunk);
2463 for (field = 0; field < _NPCM; field++) {
2464 if (pc->pc_map[field]) {
2465 bit = bsfl(pc->pc_map[field]);
2469 if (field < _NPCM) {
2470 pv = &pc->pc_pventry[field * 32 + bit];
2471 pc->pc_map[field] &= ~(1ul << bit);
2472 /* If this was the last item, move it to tail */
2473 for (field = 0; field < _NPCM; field++)
2474 if (pc->pc_map[field] != 0) {
2475 PV_STAT(pv_entry_spare--);
2476 return (pv); /* not full, return */
2478 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2479 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
2480 PV_STAT(pv_entry_spare--);
2485 * Access to the ptelist "pv_vafree" is synchronized by the pvh
2486 * global lock. If "pv_vafree" is currently non-empty, it will
2487 * remain non-empty until pmap_ptelist_alloc() completes.
2489 if (pv_vafree == 0 || (m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
2490 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) {
2493 PV_STAT(pc_chunk_tryfail++);
2496 m = pmap_pv_reclaim(pmap);
2500 PV_STAT(pc_chunk_count++);
2501 PV_STAT(pc_chunk_allocs++);
2502 pc = (struct pv_chunk *)pmap_ptelist_alloc(&pv_vafree);
2503 pmap_qenter((vm_offset_t)pc, &m, 1);
2505 pc->pc_map[0] = pc_freemask[0] & ~1ul; /* preallocated bit 0 */
2506 for (field = 1; field < _NPCM; field++)
2507 pc->pc_map[field] = pc_freemask[field];
2508 TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru);
2509 pv = &pc->pc_pventry[0];
2510 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
2511 PV_STAT(pv_entry_spare += _NPCPV - 1);
2515 static __inline pv_entry_t
2516 pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
2520 rw_assert(&pvh_global_lock, RA_WLOCKED);
2521 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) {
2522 if (pmap == PV_PMAP(pv) && va == pv->pv_va) {
2523 TAILQ_REMOVE(&pvh->pv_list, pv, pv_list);
2531 pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa)
2533 struct md_page *pvh;
2535 vm_offset_t va_last;
2538 rw_assert(&pvh_global_lock, RA_WLOCKED);
2539 KASSERT((pa & PDRMASK) == 0,
2540 ("pmap_pv_demote_pde: pa is not 4mpage aligned"));
2543 * Transfer the 4mpage's pv entry for this mapping to the first
2546 pvh = pa_to_pvh(pa);
2547 va = trunc_4mpage(va);
2548 pv = pmap_pvh_remove(pvh, pmap, va);
2549 KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found"));
2550 m = PHYS_TO_VM_PAGE(pa);
2551 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2552 /* Instantiate the remaining NPTEPG - 1 pv entries. */
2553 va_last = va + NBPDR - PAGE_SIZE;
2556 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2557 ("pmap_pv_demote_pde: page %p is not managed", m));
2559 pmap_insert_entry(pmap, va, m);
2560 } while (va < va_last);
2564 pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa)
2566 struct md_page *pvh;
2568 vm_offset_t va_last;
2571 rw_assert(&pvh_global_lock, RA_WLOCKED);
2572 KASSERT((pa & PDRMASK) == 0,
2573 ("pmap_pv_promote_pde: pa is not 4mpage aligned"));
2576 * Transfer the first page's pv entry for this mapping to the
2577 * 4mpage's pv list. Aside from avoiding the cost of a call
2578 * to get_pv_entry(), a transfer avoids the possibility that
2579 * get_pv_entry() calls pmap_collect() and that pmap_collect()
2580 * removes one of the mappings that is being promoted.
2582 m = PHYS_TO_VM_PAGE(pa);
2583 va = trunc_4mpage(va);
2584 pv = pmap_pvh_remove(&m->md, pmap, va);
2585 KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found"));
2586 pvh = pa_to_pvh(pa);
2587 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_list);
2588 /* Free the remaining NPTEPG - 1 pv entries. */
2589 va_last = va + NBPDR - PAGE_SIZE;
2593 pmap_pvh_free(&m->md, pmap, va);
2594 } while (va < va_last);
2598 pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
2602 pv = pmap_pvh_remove(pvh, pmap, va);
2603 KASSERT(pv != NULL, ("pmap_pvh_free: pv not found"));
2604 free_pv_entry(pmap, pv);
2608 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
2610 struct md_page *pvh;
2612 rw_assert(&pvh_global_lock, RA_WLOCKED);
2613 pmap_pvh_free(&m->md, pmap, va);
2614 if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) {
2615 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
2616 if (TAILQ_EMPTY(&pvh->pv_list))
2617 vm_page_aflag_clear(m, PGA_WRITEABLE);
2622 * Create a pv entry for page at pa for
2626 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
2630 rw_assert(&pvh_global_lock, RA_WLOCKED);
2631 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2632 pv = get_pv_entry(pmap, FALSE);
2634 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2638 * Conditionally create a pv entry.
2641 pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
2645 rw_assert(&pvh_global_lock, RA_WLOCKED);
2646 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2647 if (pv_entry_count < pv_entry_high_water &&
2648 (pv = get_pv_entry(pmap, TRUE)) != NULL) {
2650 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2657 * Create the pv entries for each of the pages within a superpage.
2660 pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa)
2662 struct md_page *pvh;
2665 rw_assert(&pvh_global_lock, RA_WLOCKED);
2666 if (pv_entry_count < pv_entry_high_water &&
2667 (pv = get_pv_entry(pmap, TRUE)) != NULL) {
2669 pvh = pa_to_pvh(pa);
2670 TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_list);
2677 * Fills a page table page with mappings to consecutive physical pages.
2680 pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte)
2684 for (pte = firstpte; pte < firstpte + NPTEPG; pte++) {
2686 newpte += PAGE_SIZE;
2691 * Tries to demote a 2- or 4MB page mapping. If demotion fails, the
2692 * 2- or 4MB page mapping is invalidated.
2695 pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
2697 pd_entry_t newpde, oldpde;
2698 pt_entry_t *firstpte, newpte;
2700 vm_page_t free, mpte;
2702 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2704 KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V),
2705 ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V"));
2706 mpte = pmap_lookup_pt_page(pmap, va);
2708 pmap_remove_pt_page(pmap, mpte);
2710 KASSERT((oldpde & PG_W) == 0,
2711 ("pmap_demote_pde: page table page for a wired mapping"
2715 * Invalidate the 2- or 4MB page mapping and return
2716 * "failure" if the mapping was never accessed or the
2717 * allocation of the new page table page fails.
2719 if ((oldpde & PG_A) == 0 || (mpte = vm_page_alloc(NULL,
2720 va >> PDRSHIFT, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL |
2721 VM_ALLOC_WIRED)) == NULL) {
2723 pmap_remove_pde(pmap, pde, trunc_4mpage(va), &free);
2724 pmap_invalidate_page(pmap, trunc_4mpage(va));
2725 pmap_free_zero_pages(free);
2726 CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#x"
2727 " in pmap %p", va, pmap);
2730 if (va < VM_MAXUSER_ADDRESS)
2731 pmap->pm_stats.resident_count++;
2733 mptepa = VM_PAGE_TO_PHYS(mpte);
2736 * If the page mapping is in the kernel's address space, then the
2737 * KPTmap can provide access to the page table page. Otherwise,
2738 * temporarily map the page table page (mpte) into the kernel's
2739 * address space at either PADDR1 or PADDR2.
2742 firstpte = &KPTmap[i386_btop(trunc_4mpage(va))];
2743 else if (curthread->td_pinned > 0 && rw_wowned(&pvh_global_lock)) {
2744 if ((*PMAP1 & PG_FRAME) != mptepa) {
2745 *PMAP1 = mptepa | PG_RW | PG_V | PG_A | PG_M;
2747 PMAP1cpu = PCPU_GET(cpuid);
2753 if (PMAP1cpu != PCPU_GET(cpuid)) {
2754 PMAP1cpu = PCPU_GET(cpuid);
2762 mtx_lock(&PMAP2mutex);
2763 if ((*PMAP2 & PG_FRAME) != mptepa) {
2764 *PMAP2 = mptepa | PG_RW | PG_V | PG_A | PG_M;
2765 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2);
2769 newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V;
2770 KASSERT((oldpde & PG_A) != 0,
2771 ("pmap_demote_pde: oldpde is missing PG_A"));
2772 KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW,
2773 ("pmap_demote_pde: oldpde is missing PG_M"));
2774 newpte = oldpde & ~PG_PS;
2775 if ((newpte & PG_PDE_PAT) != 0)
2776 newpte ^= PG_PDE_PAT | PG_PTE_PAT;
2779 * If the page table page is new, initialize it.
2781 if (mpte->wire_count == 1) {
2782 mpte->wire_count = NPTEPG;
2783 pmap_fill_ptp(firstpte, newpte);
2785 KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME),
2786 ("pmap_demote_pde: firstpte and newpte map different physical"
2790 * If the mapping has changed attributes, update the page table
2793 if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE))
2794 pmap_fill_ptp(firstpte, newpte);
2797 * Demote the mapping. This pmap is locked. The old PDE has
2798 * PG_A set. If the old PDE has PG_RW set, it also has PG_M
2799 * set. Thus, there is no danger of a race with another
2800 * processor changing the setting of PG_A and/or PG_M between
2801 * the read above and the store below.
2803 if (workaround_erratum383)
2804 pmap_update_pde(pmap, va, pde, newpde);
2805 else if (pmap == kernel_pmap)
2806 pmap_kenter_pde(va, newpde);
2808 pde_store(pde, newpde);
2809 if (firstpte == PADDR2)
2810 mtx_unlock(&PMAP2mutex);
2813 * Invalidate the recursive mapping of the page table page.
2815 pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va));
2818 * Demote the pv entry. This depends on the earlier demotion
2819 * of the mapping. Specifically, the (re)creation of a per-
2820 * page pv entry might trigger the execution of pmap_collect(),
2821 * which might reclaim a newly (re)created per-page pv entry
2822 * and destroy the associated mapping. In order to destroy
2823 * the mapping, the PDE must have already changed from mapping
2824 * the 2mpage to referencing the page table page.
2826 if ((oldpde & PG_MANAGED) != 0)
2827 pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME);
2829 pmap_pde_demotions++;
2830 CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#x"
2831 " in pmap %p", va, pmap);
2836 * pmap_remove_pde: do the things to unmap a superpage in a process
2839 pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
2842 struct md_page *pvh;
2844 vm_offset_t eva, va;
2847 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2848 KASSERT((sva & PDRMASK) == 0,
2849 ("pmap_remove_pde: sva is not 4mpage aligned"));
2850 oldpde = pte_load_clear(pdq);
2852 pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE;
2855 * Machines that don't support invlpg, also don't support
2859 pmap_invalidate_page(kernel_pmap, sva);
2860 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2861 if (oldpde & PG_MANAGED) {
2862 pvh = pa_to_pvh(oldpde & PG_PS_FRAME);
2863 pmap_pvh_free(pvh, pmap, sva);
2865 for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
2866 va < eva; va += PAGE_SIZE, m++) {
2867 if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW))
2870 vm_page_aflag_set(m, PGA_REFERENCED);
2871 if (TAILQ_EMPTY(&m->md.pv_list) &&
2872 TAILQ_EMPTY(&pvh->pv_list))
2873 vm_page_aflag_clear(m, PGA_WRITEABLE);
2876 if (pmap == kernel_pmap) {
2877 if (!pmap_demote_pde(pmap, pdq, sva))
2878 panic("pmap_remove_pde: failed demotion");
2880 mpte = pmap_lookup_pt_page(pmap, sva);
2882 pmap_remove_pt_page(pmap, mpte);
2883 pmap->pm_stats.resident_count--;
2884 KASSERT(mpte->wire_count == NPTEPG,
2885 ("pmap_remove_pde: pte page wire count error"));
2886 mpte->wire_count = 0;
2887 pmap_add_delayed_free_list(mpte, free, FALSE);
2888 atomic_subtract_int(&cnt.v_wire_count, 1);
2894 * pmap_remove_pte: do the things to unmap a page in a process
2897 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va, vm_page_t *free)
2902 rw_assert(&pvh_global_lock, RA_WLOCKED);
2903 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2904 oldpte = pte_load_clear(ptq);
2905 KASSERT(oldpte != 0,
2906 ("pmap_remove_pte: pmap %p va %x zero pte", pmap, va));
2908 pmap->pm_stats.wired_count -= 1;
2910 * Machines that don't support invlpg, also don't support
2914 pmap_invalidate_page(kernel_pmap, va);
2915 pmap->pm_stats.resident_count -= 1;
2916 if (oldpte & PG_MANAGED) {
2917 m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME);
2918 if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
2921 vm_page_aflag_set(m, PGA_REFERENCED);
2922 pmap_remove_entry(pmap, m, va);
2924 return (pmap_unuse_pt(pmap, va, free));
2928 * Remove a single page from a process address space
2931 pmap_remove_page(pmap_t pmap, vm_offset_t va, vm_page_t *free)
2935 rw_assert(&pvh_global_lock, RA_WLOCKED);
2936 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
2937 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2938 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0)
2940 pmap_remove_pte(pmap, pte, va, free);
2941 pmap_invalidate_page(pmap, va);
2945 * Remove the given range of addresses from the specified map.
2947 * It is assumed that the start and end are properly
2948 * rounded to the page size.
2951 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2956 vm_page_t free = NULL;
2960 * Perform an unsynchronized read. This is, however, safe.
2962 if (pmap->pm_stats.resident_count == 0)
2967 rw_wlock(&pvh_global_lock);
2972 * special handling of removing one page. a very
2973 * common operation and easy to short circuit some
2976 if ((sva + PAGE_SIZE == eva) &&
2977 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
2978 pmap_remove_page(pmap, sva, &free);
2982 for (; sva < eva; sva = pdnxt) {
2986 * Calculate index for next page table.
2988 pdnxt = (sva + NBPDR) & ~PDRMASK;
2991 if (pmap->pm_stats.resident_count == 0)
2994 pdirindex = sva >> PDRSHIFT;
2995 ptpaddr = pmap->pm_pdir[pdirindex];
2998 * Weed out invalid mappings. Note: we assume that the page
2999 * directory table is always allocated, and in kernel virtual.
3005 * Check for large page.
3007 if ((ptpaddr & PG_PS) != 0) {
3009 * Are we removing the entire large page? If not,
3010 * demote the mapping and fall through.
3012 if (sva + NBPDR == pdnxt && eva >= pdnxt) {
3014 * The TLB entry for a PG_G mapping is
3015 * invalidated by pmap_remove_pde().
3017 if ((ptpaddr & PG_G) == 0)
3019 pmap_remove_pde(pmap,
3020 &pmap->pm_pdir[pdirindex], sva, &free);
3022 } else if (!pmap_demote_pde(pmap,
3023 &pmap->pm_pdir[pdirindex], sva)) {
3024 /* The large page mapping was destroyed. */
3030 * Limit our scan to either the end of the va represented
3031 * by the current page table page, or to the end of the
3032 * range being removed.
3037 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
3043 * The TLB entry for a PG_G mapping is invalidated
3044 * by pmap_remove_pte().
3046 if ((*pte & PG_G) == 0)
3048 if (pmap_remove_pte(pmap, pte, sva, &free))
3055 pmap_invalidate_all(pmap);
3056 rw_wunlock(&pvh_global_lock);
3058 pmap_free_zero_pages(free);
3062 * Routine: pmap_remove_all
3064 * Removes this physical page from
3065 * all physical maps in which it resides.
3066 * Reflects back modify bits to the pager.
3069 * Original versions of this routine were very
3070 * inefficient because they iteratively called
3071 * pmap_remove (slow...)
3075 pmap_remove_all(vm_page_t m)
3077 struct md_page *pvh;
3080 pt_entry_t *pte, tpte;
3085 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3086 ("pmap_remove_all: page %p is not managed", m));
3088 rw_wlock(&pvh_global_lock);
3090 if ((m->flags & PG_FICTITIOUS) != 0)
3091 goto small_mappings;
3092 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
3093 while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) {
3097 pde = pmap_pde(pmap, va);
3098 (void)pmap_demote_pde(pmap, pde, va);
3102 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3105 pmap->pm_stats.resident_count--;
3106 pde = pmap_pde(pmap, pv->pv_va);
3107 KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found"
3108 " a 4mpage in page %p's pv list", m));
3109 pte = pmap_pte_quick(pmap, pv->pv_va);
3110 tpte = pte_load_clear(pte);
3111 KASSERT(tpte != 0, ("pmap_remove_all: pmap %p va %x zero pte",
3114 pmap->pm_stats.wired_count--;
3116 vm_page_aflag_set(m, PGA_REFERENCED);
3119 * Update the vm_page_t clean and reference bits.
3121 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
3123 pmap_unuse_pt(pmap, pv->pv_va, &free);
3124 pmap_invalidate_page(pmap, pv->pv_va);
3125 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3126 free_pv_entry(pmap, pv);
3129 vm_page_aflag_clear(m, PGA_WRITEABLE);
3131 rw_wunlock(&pvh_global_lock);
3132 pmap_free_zero_pages(free);
3136 * pmap_protect_pde: do the things to protect a 4mpage in a process
3139 pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot)
3141 pd_entry_t newpde, oldpde;
3142 vm_offset_t eva, va;
3144 boolean_t anychanged;
3146 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3147 KASSERT((sva & PDRMASK) == 0,
3148 ("pmap_protect_pde: sva is not 4mpage aligned"));
3151 oldpde = newpde = *pde;
3152 if (oldpde & PG_MANAGED) {
3154 for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
3155 va < eva; va += PAGE_SIZE, m++)
3156 if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW))
3159 if ((prot & VM_PROT_WRITE) == 0)
3160 newpde &= ~(PG_RW | PG_M);
3162 if ((prot & VM_PROT_EXECUTE) == 0)
3165 if (newpde != oldpde) {
3166 if (!pde_cmpset(pde, oldpde, newpde))
3169 pmap_invalidate_page(pmap, sva);
3173 return (anychanged);
3177 * Set the physical protection on the
3178 * specified range of this map as requested.
3181 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3186 boolean_t anychanged, pv_lists_locked;
3188 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3189 pmap_remove(pmap, sva, eva);
3194 if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) ==
3195 (VM_PROT_WRITE|VM_PROT_EXECUTE))
3198 if (prot & VM_PROT_WRITE)
3202 if (pmap_is_current(pmap))
3203 pv_lists_locked = FALSE;
3205 pv_lists_locked = TRUE;
3207 rw_wlock(&pvh_global_lock);
3213 for (; sva < eva; sva = pdnxt) {
3214 pt_entry_t obits, pbits;
3217 pdnxt = (sva + NBPDR) & ~PDRMASK;
3221 pdirindex = sva >> PDRSHIFT;
3222 ptpaddr = pmap->pm_pdir[pdirindex];
3225 * Weed out invalid mappings. Note: we assume that the page
3226 * directory table is always allocated, and in kernel virtual.
3232 * Check for large page.
3234 if ((ptpaddr & PG_PS) != 0) {
3236 * Are we protecting the entire large page? If not,
3237 * demote the mapping and fall through.
3239 if (sva + NBPDR == pdnxt && eva >= pdnxt) {
3241 * The TLB entry for a PG_G mapping is
3242 * invalidated by pmap_protect_pde().
3244 if (pmap_protect_pde(pmap,
3245 &pmap->pm_pdir[pdirindex], sva, prot))
3249 if (!pv_lists_locked) {
3250 pv_lists_locked = TRUE;
3251 if (!rw_try_wlock(&pvh_global_lock)) {
3253 pmap_invalidate_all(
3260 if (!pmap_demote_pde(pmap,
3261 &pmap->pm_pdir[pdirindex], sva)) {
3263 * The large page mapping was
3274 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
3280 * Regardless of whether a pte is 32 or 64 bits in
3281 * size, PG_RW, PG_A, and PG_M are among the least
3282 * significant 32 bits.
3284 obits = pbits = *pte;
3285 if ((pbits & PG_V) == 0)
3288 if ((prot & VM_PROT_WRITE) == 0) {
3289 if ((pbits & (PG_MANAGED | PG_M | PG_RW)) ==
3290 (PG_MANAGED | PG_M | PG_RW)) {
3291 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3294 pbits &= ~(PG_RW | PG_M);
3297 if ((prot & VM_PROT_EXECUTE) == 0)
3301 if (pbits != obits) {
3303 if (!atomic_cmpset_64(pte, obits, pbits))
3306 if (!atomic_cmpset_int((u_int *)pte, obits,
3311 pmap_invalidate_page(pmap, sva);
3318 pmap_invalidate_all(pmap);
3319 if (pv_lists_locked) {
3321 rw_wunlock(&pvh_global_lock);
3327 * Tries to promote the 512 or 1024, contiguous 4KB page mappings that are
3328 * within a single page table page (PTP) to a single 2- or 4MB page mapping.
3329 * For promotion to occur, two conditions must be met: (1) the 4KB page
3330 * mappings must map aligned, contiguous physical memory and (2) the 4KB page
3331 * mappings must have identical characteristics.
3333 * Managed (PG_MANAGED) mappings within the kernel address space are not
3334 * promoted. The reason is that kernel PDEs are replicated in each pmap but
3335 * pmap_clear_ptes() and pmap_ts_referenced() only read the PDE from the kernel
3339 pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
3342 pt_entry_t *firstpte, oldpte, pa, *pte;
3343 vm_offset_t oldpteva;
3346 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3349 * Examine the first PTE in the specified PTP. Abort if this PTE is
3350 * either invalid, unused, or does not map the first 4KB physical page
3351 * within a 2- or 4MB page.
3353 firstpte = pmap_pte_quick(pmap, trunc_4mpage(va));
3356 if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V)) {
3357 pmap_pde_p_failures++;
3358 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x"
3359 " in pmap %p", va, pmap);
3362 if ((*firstpte & PG_MANAGED) != 0 && pmap == kernel_pmap) {
3363 pmap_pde_p_failures++;
3364 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x"
3365 " in pmap %p", va, pmap);
3368 if ((newpde & (PG_M | PG_RW)) == PG_RW) {
3370 * When PG_M is already clear, PG_RW can be cleared without
3371 * a TLB invalidation.
3373 if (!atomic_cmpset_int((u_int *)firstpte, newpde, newpde &
3380 * Examine each of the other PTEs in the specified PTP. Abort if this
3381 * PTE maps an unexpected 4KB physical page or does not have identical
3382 * characteristics to the first PTE.
3384 pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE;
3385 for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) {
3388 if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) {
3389 pmap_pde_p_failures++;
3390 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x"
3391 " in pmap %p", va, pmap);
3394 if ((oldpte & (PG_M | PG_RW)) == PG_RW) {
3396 * When PG_M is already clear, PG_RW can be cleared
3397 * without a TLB invalidation.
3399 if (!atomic_cmpset_int((u_int *)pte, oldpte,
3403 oldpteva = (oldpte & PG_FRAME & PDRMASK) |
3405 CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#x"
3406 " in pmap %p", oldpteva, pmap);
3408 if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) {
3409 pmap_pde_p_failures++;
3410 CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x"
3411 " in pmap %p", va, pmap);
3418 * Save the page table page in its current state until the PDE
3419 * mapping the superpage is demoted by pmap_demote_pde() or
3420 * destroyed by pmap_remove_pde().
3422 mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
3423 KASSERT(mpte >= vm_page_array &&
3424 mpte < &vm_page_array[vm_page_array_size],
3425 ("pmap_promote_pde: page table page is out of range"));
3426 KASSERT(mpte->pindex == va >> PDRSHIFT,
3427 ("pmap_promote_pde: page table page's pindex is wrong"));
3428 pmap_insert_pt_page(pmap, mpte);
3431 * Promote the pv entries.
3433 if ((newpde & PG_MANAGED) != 0)
3434 pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME);
3437 * Propagate the PAT index to its proper position.
3439 if ((newpde & PG_PTE_PAT) != 0)
3440 newpde ^= PG_PDE_PAT | PG_PTE_PAT;
3443 * Map the superpage.
3445 if (workaround_erratum383)
3446 pmap_update_pde(pmap, va, pde, PG_PS | newpde);
3447 else if (pmap == kernel_pmap)
3448 pmap_kenter_pde(va, PG_PS | newpde);
3450 pde_store(pde, PG_PS | newpde);
3452 pmap_pde_promotions++;
3453 CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#x"
3454 " in pmap %p", va, pmap);
3458 * Insert the given physical page (p) at
3459 * the specified virtual address (v) in the
3460 * target physical map with the protection requested.
3462 * If specified, the page will be wired down, meaning
3463 * that the related pte can not be reclaimed.
3465 * NB: This is the only routine which MAY NOT lazy-evaluate
3466 * or lose information. That is, this routine must actually
3467 * insert this page into the given map NOW.
3470 pmap_enter(pmap_t pmap, vm_offset_t va, vm_prot_t access, vm_page_t m,
3471 vm_prot_t prot, boolean_t wired)
3475 pt_entry_t newpte, origpte;
3481 va = trunc_page(va);
3482 KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig"));
3483 KASSERT(va < UPT_MIN_ADDRESS || va >= UPT_MAX_ADDRESS,
3484 ("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)",
3486 KASSERT((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) != 0 ||
3487 VM_OBJECT_LOCKED(m->object),
3488 ("pmap_enter: page %p is not busy", m));
3492 rw_wlock(&pvh_global_lock);
3497 * In the case that a page table page is not
3498 * resident, we are creating it here.
3500 if (va < VM_MAXUSER_ADDRESS) {
3501 mpte = pmap_allocpte(pmap, va, M_WAITOK);
3504 pde = pmap_pde(pmap, va);
3505 if ((*pde & PG_PS) != 0)
3506 panic("pmap_enter: attempted pmap_enter on 4MB page");
3507 pte = pmap_pte_quick(pmap, va);
3510 * Page Directory table entry not valid, we need a new PT page
3513 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x",
3514 (uintmax_t)pmap->pm_pdir[PTDPTDI], va);
3517 pa = VM_PAGE_TO_PHYS(m);
3520 opa = origpte & PG_FRAME;
3523 * Mapping has not changed, must be protection or wiring change.
3525 if (origpte && (opa == pa)) {
3527 * Wiring change, just update stats. We don't worry about
3528 * wiring PT pages as they remain resident as long as there
3529 * are valid mappings in them. Hence, if a user page is wired,
3530 * the PT page will be also.
3532 if (wired && ((origpte & PG_W) == 0))
3533 pmap->pm_stats.wired_count++;
3534 else if (!wired && (origpte & PG_W))
3535 pmap->pm_stats.wired_count--;
3538 * Remove extra pte reference
3543 if (origpte & PG_MANAGED) {
3553 * Mapping has changed, invalidate old range and fall through to
3554 * handle validating new mapping.
3558 pmap->pm_stats.wired_count--;
3559 if (origpte & PG_MANAGED) {
3560 om = PHYS_TO_VM_PAGE(opa);
3561 pv = pmap_pvh_remove(&om->md, pmap, va);
3565 KASSERT(mpte->wire_count > 0,
3566 ("pmap_enter: missing reference to page table page,"
3570 pmap->pm_stats.resident_count++;
3573 * Enter on the PV list if part of our managed memory.
3575 if ((m->oflags & VPO_UNMANAGED) == 0) {
3576 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva,
3577 ("pmap_enter: managed mapping within the clean submap"));
3579 pv = get_pv_entry(pmap, FALSE);
3581 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3583 } else if (pv != NULL)
3584 free_pv_entry(pmap, pv);
3587 * Increment counters
3590 pmap->pm_stats.wired_count++;
3594 * Now validate mapping with desired protection/wiring.
3596 newpte = (pt_entry_t)(pa | pmap_cache_bits(m->md.pat_mode, 0) | PG_V);
3597 if ((prot & VM_PROT_WRITE) != 0) {
3599 if ((newpte & PG_MANAGED) != 0)
3600 vm_page_aflag_set(m, PGA_WRITEABLE);
3603 if ((prot & VM_PROT_EXECUTE) == 0)
3608 if (va < VM_MAXUSER_ADDRESS)
3610 if (pmap == kernel_pmap)
3614 * if the mapping or permission bits are different, we need
3615 * to update the pte.
3617 if ((origpte & ~(PG_M|PG_A)) != newpte) {
3619 if ((access & VM_PROT_WRITE) != 0)
3621 if (origpte & PG_V) {
3623 origpte = pte_load_store(pte, newpte);
3624 if (origpte & PG_A) {
3625 if (origpte & PG_MANAGED)
3626 vm_page_aflag_set(om, PGA_REFERENCED);
3627 if (opa != VM_PAGE_TO_PHYS(m))
3630 if ((origpte & PG_NX) == 0 &&
3631 (newpte & PG_NX) != 0)
3635 if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
3636 if ((origpte & PG_MANAGED) != 0)
3638 if ((prot & VM_PROT_WRITE) == 0)
3641 if ((origpte & PG_MANAGED) != 0 &&
3642 TAILQ_EMPTY(&om->md.pv_list) &&
3643 ((om->flags & PG_FICTITIOUS) != 0 ||
3644 TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list)))
3645 vm_page_aflag_clear(om, PGA_WRITEABLE);
3647 pmap_invalidate_page(pmap, va);
3649 pte_store(pte, newpte);
3653 * If both the page table page and the reservation are fully
3654 * populated, then attempt promotion.
3656 if ((mpte == NULL || mpte->wire_count == NPTEPG) &&
3657 pg_ps_enabled && (m->flags & PG_FICTITIOUS) == 0 &&
3658 vm_reserv_level_iffullpop(m) == 0)
3659 pmap_promote_pde(pmap, pde, va);
3662 rw_wunlock(&pvh_global_lock);
3667 * Tries to create a 2- or 4MB page mapping. Returns TRUE if successful and
3668 * FALSE otherwise. Fails if (1) a page table page cannot be allocated without
3669 * blocking, (2) a mapping already exists at the specified virtual address, or
3670 * (3) a pv entry cannot be allocated without reclaiming another pv entry.
3673 pmap_enter_pde(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot)
3675 pd_entry_t *pde, newpde;
3677 rw_assert(&pvh_global_lock, RA_WLOCKED);
3678 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3679 pde = pmap_pde(pmap, va);
3681 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
3682 " in pmap %p", va, pmap);
3685 newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.pat_mode, 1) |
3687 if ((m->oflags & VPO_UNMANAGED) == 0) {
3688 newpde |= PG_MANAGED;
3691 * Abort this mapping if its PV entry could not be created.
3693 if (!pmap_pv_insert_pde(pmap, va, VM_PAGE_TO_PHYS(m))) {
3694 CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
3695 " in pmap %p", va, pmap);
3700 if ((prot & VM_PROT_EXECUTE) == 0)
3703 if (va < VM_MAXUSER_ADDRESS)
3707 * Increment counters.
3709 pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3712 * Map the superpage.
3714 pde_store(pde, newpde);
3716 pmap_pde_mappings++;
3717 CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx"
3718 " in pmap %p", va, pmap);
3723 * Maps a sequence of resident pages belonging to the same object.
3724 * The sequence begins with the given page m_start. This page is
3725 * mapped at the given virtual address start. Each subsequent page is
3726 * mapped at a virtual address that is offset from start by the same
3727 * amount as the page is offset from m_start within the object. The
3728 * last page in the sequence is the page with the largest offset from
3729 * m_start that can be mapped at a virtual address less than the given
3730 * virtual address end. Not every virtual page between start and end
3731 * is mapped; only those for which a resident page exists with the
3732 * corresponding offset from m_start are mapped.
3735 pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end,
3736 vm_page_t m_start, vm_prot_t prot)
3740 vm_pindex_t diff, psize;
3742 VM_OBJECT_LOCK_ASSERT(m_start->object, MA_OWNED);
3743 psize = atop(end - start);
3746 rw_wlock(&pvh_global_lock);
3748 while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
3749 va = start + ptoa(diff);
3750 if ((va & PDRMASK) == 0 && va + NBPDR <= end &&
3751 (VM_PAGE_TO_PHYS(m) & PDRMASK) == 0 &&
3752 pg_ps_enabled && vm_reserv_level_iffullpop(m) == 0 &&
3753 pmap_enter_pde(pmap, va, m, prot))
3754 m = &m[NBPDR / PAGE_SIZE - 1];
3756 mpte = pmap_enter_quick_locked(pmap, va, m, prot,
3758 m = TAILQ_NEXT(m, listq);
3760 rw_wunlock(&pvh_global_lock);
3765 * this code makes some *MAJOR* assumptions:
3766 * 1. Current pmap & pmap exists.
3769 * 4. No page table pages.
3770 * but is *MUCH* faster than pmap_enter...
3774 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot)
3777 rw_wlock(&pvh_global_lock);
3779 (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL);
3780 rw_wunlock(&pvh_global_lock);
3785 pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m,
3786 vm_prot_t prot, vm_page_t mpte)
3792 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva ||
3793 (m->oflags & VPO_UNMANAGED) != 0,
3794 ("pmap_enter_quick_locked: managed mapping within the clean submap"));
3795 rw_assert(&pvh_global_lock, RA_WLOCKED);
3796 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3799 * In the case that a page table page is not
3800 * resident, we are creating it here.
3802 if (va < VM_MAXUSER_ADDRESS) {
3807 * Calculate pagetable page index
3809 ptepindex = va >> PDRSHIFT;
3810 if (mpte && (mpte->pindex == ptepindex)) {
3814 * Get the page directory entry
3816 ptepa = pmap->pm_pdir[ptepindex];
3819 * If the page table page is mapped, we just increment
3820 * the hold count, and activate it.
3825 mpte = PHYS_TO_VM_PAGE(ptepa & PG_FRAME);
3828 mpte = _pmap_allocpte(pmap, ptepindex,
3839 * This call to vtopte makes the assumption that we are
3840 * entering the page into the current pmap. In order to support
3841 * quick entry into any pmap, one would likely use pmap_pte_quick.
3842 * But that isn't as quick as vtopte.
3854 * Enter on the PV list if part of our managed memory.
3856 if ((m->oflags & VPO_UNMANAGED) == 0 &&
3857 !pmap_try_insert_pv_entry(pmap, va, m)) {
3860 if (pmap_unwire_ptp(pmap, mpte, &free)) {
3861 pmap_invalidate_page(pmap, va);
3862 pmap_free_zero_pages(free);
3871 * Increment counters
3873 pmap->pm_stats.resident_count++;
3875 pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.pat_mode, 0);
3877 if ((prot & VM_PROT_EXECUTE) == 0)
3882 * Now validate mapping with RO protection
3884 if ((m->oflags & VPO_UNMANAGED) != 0)
3885 pte_store(pte, pa | PG_V | PG_U);
3887 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED);
3892 * Make a temporary mapping for a physical address. This is only intended
3893 * to be used for panic dumps.
3896 pmap_kenter_temporary(vm_paddr_t pa, int i)
3900 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
3901 pmap_kenter(va, pa);
3903 return ((void *)crashdumpmap);
3907 * This code maps large physical mmap regions into the
3908 * processor address space. Note that some shortcuts
3909 * are taken, but the code works.
3912 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object,
3913 vm_pindex_t pindex, vm_size_t size)
3916 vm_paddr_t pa, ptepa;
3920 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
3921 KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG,
3922 ("pmap_object_init_pt: non-device object"));
3924 (addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) {
3925 if (!vm_object_populate(object, pindex, pindex + atop(size)))
3927 p = vm_page_lookup(object, pindex);
3928 KASSERT(p->valid == VM_PAGE_BITS_ALL,
3929 ("pmap_object_init_pt: invalid page %p", p));
3930 pat_mode = p->md.pat_mode;
3933 * Abort the mapping if the first page is not physically
3934 * aligned to a 2/4MB page boundary.
3936 ptepa = VM_PAGE_TO_PHYS(p);
3937 if (ptepa & (NBPDR - 1))
3941 * Skip the first page. Abort the mapping if the rest of
3942 * the pages are not physically contiguous or have differing
3943 * memory attributes.
3945 p = TAILQ_NEXT(p, listq);
3946 for (pa = ptepa + PAGE_SIZE; pa < ptepa + size;
3948 KASSERT(p->valid == VM_PAGE_BITS_ALL,
3949 ("pmap_object_init_pt: invalid page %p", p));
3950 if (pa != VM_PAGE_TO_PHYS(p) ||
3951 pat_mode != p->md.pat_mode)
3953 p = TAILQ_NEXT(p, listq);
3957 * Map using 2/4MB pages. Since "ptepa" is 2/4M aligned and
3958 * "size" is a multiple of 2/4M, adding the PAT setting to
3959 * "pa" will not affect the termination of this loop.
3962 for (pa = ptepa | pmap_cache_bits(pat_mode, 1); pa < ptepa +
3963 size; pa += NBPDR) {
3964 pde = pmap_pde(pmap, addr);
3966 pde_store(pde, pa | PG_PS | PG_M | PG_A |
3967 PG_U | PG_RW | PG_V);
3968 pmap->pm_stats.resident_count += NBPDR /
3970 pmap_pde_mappings++;
3972 /* Else continue on if the PDE is already valid. */
3980 * Routine: pmap_change_wiring
3981 * Function: Change the wiring attribute for a map/virtual-address
3983 * In/out conditions:
3984 * The mapping must already exist in the pmap.
3987 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
3991 boolean_t are_queues_locked;
3993 are_queues_locked = FALSE;
3996 pde = pmap_pde(pmap, va);
3997 if ((*pde & PG_PS) != 0) {
3998 if (!wired != ((*pde & PG_W) == 0)) {
3999 if (!are_queues_locked) {
4000 are_queues_locked = TRUE;
4001 if (!rw_try_wlock(&pvh_global_lock)) {
4003 rw_wlock(&pvh_global_lock);
4007 if (!pmap_demote_pde(pmap, pde, va))
4008 panic("pmap_change_wiring: demotion failed");
4012 pte = pmap_pte(pmap, va);
4014 if (wired && !pmap_pte_w(pte))
4015 pmap->pm_stats.wired_count++;
4016 else if (!wired && pmap_pte_w(pte))
4017 pmap->pm_stats.wired_count--;
4020 * Wiring is not a hardware characteristic so there is no need to
4023 pmap_pte_set_w(pte, wired);
4024 pmap_pte_release(pte);
4026 if (are_queues_locked)
4027 rw_wunlock(&pvh_global_lock);
4034 * Copy the range specified by src_addr/len
4035 * from the source map to the range dst_addr/len
4036 * in the destination map.
4038 * This routine is only advisory and need not do anything.
4042 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
4043 vm_offset_t src_addr)
4047 vm_offset_t end_addr = src_addr + len;
4050 if (dst_addr != src_addr)
4053 if (!pmap_is_current(src_pmap))
4056 rw_wlock(&pvh_global_lock);
4057 if (dst_pmap < src_pmap) {
4058 PMAP_LOCK(dst_pmap);
4059 PMAP_LOCK(src_pmap);
4061 PMAP_LOCK(src_pmap);
4062 PMAP_LOCK(dst_pmap);
4065 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
4066 pt_entry_t *src_pte, *dst_pte;
4067 vm_page_t dstmpte, srcmpte;
4068 pd_entry_t srcptepaddr;
4071 KASSERT(addr < UPT_MIN_ADDRESS,
4072 ("pmap_copy: invalid to pmap_copy page tables"));
4074 pdnxt = (addr + NBPDR) & ~PDRMASK;
4077 ptepindex = addr >> PDRSHIFT;
4079 srcptepaddr = src_pmap->pm_pdir[ptepindex];
4080 if (srcptepaddr == 0)
4083 if (srcptepaddr & PG_PS) {
4084 if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr)
4086 if (dst_pmap->pm_pdir[ptepindex] == 0 &&
4087 ((srcptepaddr & PG_MANAGED) == 0 ||
4088 pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr &
4090 dst_pmap->pm_pdir[ptepindex] = srcptepaddr &
4092 dst_pmap->pm_stats.resident_count +=
4098 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr & PG_FRAME);
4099 KASSERT(srcmpte->wire_count > 0,
4100 ("pmap_copy: source page table page is unused"));
4102 if (pdnxt > end_addr)
4105 src_pte = vtopte(addr);
4106 while (addr < pdnxt) {
4110 * we only virtual copy managed pages
4112 if ((ptetemp & PG_MANAGED) != 0) {
4113 dstmpte = pmap_allocpte(dst_pmap, addr,
4115 if (dstmpte == NULL)
4117 dst_pte = pmap_pte_quick(dst_pmap, addr);
4118 if (*dst_pte == 0 &&
4119 pmap_try_insert_pv_entry(dst_pmap, addr,
4120 PHYS_TO_VM_PAGE(ptetemp & PG_FRAME))) {
4122 * Clear the wired, modified, and
4123 * accessed (referenced) bits
4126 *dst_pte = ptetemp & ~(PG_W | PG_M |
4128 dst_pmap->pm_stats.resident_count++;
4131 if (pmap_unwire_ptp(dst_pmap, dstmpte,
4133 pmap_invalidate_page(dst_pmap,
4135 pmap_free_zero_pages(free);
4139 if (dstmpte->wire_count >= srcmpte->wire_count)
4148 rw_wunlock(&pvh_global_lock);
4149 PMAP_UNLOCK(src_pmap);
4150 PMAP_UNLOCK(dst_pmap);
4153 static __inline void
4154 pagezero(void *page)
4156 #if defined(I686_CPU)
4157 if (cpu_class == CPUCLASS_686) {
4158 #if defined(CPU_ENABLE_SSE)
4159 if (cpu_feature & CPUID_SSE2)
4160 sse2_pagezero(page);
4163 i686_pagezero(page);
4166 bzero(page, PAGE_SIZE);
4170 * pmap_zero_page zeros the specified hardware page by mapping
4171 * the page into KVM and using bzero to clear its contents.
4174 pmap_zero_page(vm_page_t m)
4176 struct sysmaps *sysmaps;
4178 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
4179 mtx_lock(&sysmaps->lock);
4180 if (*sysmaps->CMAP2)
4181 panic("pmap_zero_page: CMAP2 busy");
4183 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M |
4184 pmap_cache_bits(m->md.pat_mode, 0);
4185 invlcaddr(sysmaps->CADDR2);
4186 pagezero(sysmaps->CADDR2);
4187 *sysmaps->CMAP2 = 0;
4189 mtx_unlock(&sysmaps->lock);
4193 * pmap_zero_page_area zeros the specified hardware page by mapping
4194 * the page into KVM and using bzero to clear its contents.
4196 * off and size may not cover an area beyond a single hardware page.
4199 pmap_zero_page_area(vm_page_t m, int off, int size)
4201 struct sysmaps *sysmaps;
4203 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
4204 mtx_lock(&sysmaps->lock);
4205 if (*sysmaps->CMAP2)
4206 panic("pmap_zero_page_area: CMAP2 busy");
4208 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M |
4209 pmap_cache_bits(m->md.pat_mode, 0);
4210 invlcaddr(sysmaps->CADDR2);
4211 if (off == 0 && size == PAGE_SIZE)
4212 pagezero(sysmaps->CADDR2);
4214 bzero((char *)sysmaps->CADDR2 + off, size);
4215 *sysmaps->CMAP2 = 0;
4217 mtx_unlock(&sysmaps->lock);
4221 * pmap_zero_page_idle zeros the specified hardware page by mapping
4222 * the page into KVM and using bzero to clear its contents. This
4223 * is intended to be called from the vm_pagezero process only and
4227 pmap_zero_page_idle(vm_page_t m)
4231 panic("pmap_zero_page_idle: CMAP3 busy");
4233 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M |
4234 pmap_cache_bits(m->md.pat_mode, 0);
4242 * pmap_copy_page copies the specified (machine independent)
4243 * page by mapping the page into virtual memory and using
4244 * bcopy to copy the page, one machine dependent page at a
4248 pmap_copy_page(vm_page_t src, vm_page_t dst)
4250 struct sysmaps *sysmaps;
4252 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
4253 mtx_lock(&sysmaps->lock);
4254 if (*sysmaps->CMAP1)
4255 panic("pmap_copy_page: CMAP1 busy");
4256 if (*sysmaps->CMAP2)
4257 panic("pmap_copy_page: CMAP2 busy");
4259 *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A |
4260 pmap_cache_bits(src->md.pat_mode, 0);
4261 invlcaddr(sysmaps->CADDR1);
4262 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M |
4263 pmap_cache_bits(dst->md.pat_mode, 0);
4264 invlcaddr(sysmaps->CADDR2);
4265 bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE);
4266 *sysmaps->CMAP1 = 0;
4267 *sysmaps->CMAP2 = 0;
4269 mtx_unlock(&sysmaps->lock);
4272 int unmapped_buf_allowed = 1;
4275 pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[],
4276 vm_offset_t b_offset, int xfersize)
4278 struct sysmaps *sysmaps;
4279 vm_page_t a_pg, b_pg;
4281 vm_offset_t a_pg_offset, b_pg_offset;
4284 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
4285 mtx_lock(&sysmaps->lock);
4286 if (*sysmaps->CMAP1 != 0)
4287 panic("pmap_copy_pages: CMAP1 busy");
4288 if (*sysmaps->CMAP2 != 0)
4289 panic("pmap_copy_pages: CMAP2 busy");
4291 while (xfersize > 0) {
4292 a_pg = ma[a_offset >> PAGE_SHIFT];
4293 a_pg_offset = a_offset & PAGE_MASK;
4294 cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
4295 b_pg = mb[b_offset >> PAGE_SHIFT];
4296 b_pg_offset = b_offset & PAGE_MASK;
4297 cnt = min(cnt, PAGE_SIZE - b_pg_offset);
4298 *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(a_pg) | PG_A |
4299 pmap_cache_bits(a_pg->md.pat_mode, 0);
4300 invlcaddr(sysmaps->CADDR1);
4301 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(b_pg) | PG_A |
4302 PG_M | pmap_cache_bits(b_pg->md.pat_mode, 0);
4303 invlcaddr(sysmaps->CADDR2);
4304 a_cp = sysmaps->CADDR1 + a_pg_offset;
4305 b_cp = sysmaps->CADDR2 + b_pg_offset;
4306 bcopy(a_cp, b_cp, cnt);
4311 *sysmaps->CMAP1 = 0;
4312 *sysmaps->CMAP2 = 0;
4314 mtx_unlock(&sysmaps->lock);
4318 * Returns true if the pmap's pv is one of the first
4319 * 16 pvs linked to from this page. This count may
4320 * be changed upwards or downwards in the future; it
4321 * is only necessary that true be returned for a small
4322 * subset of pmaps for proper page aging.
4325 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4327 struct md_page *pvh;
4332 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4333 ("pmap_page_exists_quick: page %p is not managed", m));
4335 rw_wlock(&pvh_global_lock);
4336 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4337 if (PV_PMAP(pv) == pmap) {
4345 if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) {
4346 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4347 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) {
4348 if (PV_PMAP(pv) == pmap) {
4357 rw_wunlock(&pvh_global_lock);
4362 * pmap_page_wired_mappings:
4364 * Return the number of managed mappings to the given physical page
4368 pmap_page_wired_mappings(vm_page_t m)
4373 if ((m->oflags & VPO_UNMANAGED) != 0)
4375 rw_wlock(&pvh_global_lock);
4376 count = pmap_pvh_wired_mappings(&m->md, count);
4377 if ((m->flags & PG_FICTITIOUS) == 0) {
4378 count = pmap_pvh_wired_mappings(pa_to_pvh(VM_PAGE_TO_PHYS(m)),
4381 rw_wunlock(&pvh_global_lock);
4386 * pmap_pvh_wired_mappings:
4388 * Return the updated number "count" of managed mappings that are wired.
4391 pmap_pvh_wired_mappings(struct md_page *pvh, int count)
4397 rw_assert(&pvh_global_lock, RA_WLOCKED);
4399 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) {
4402 pte = pmap_pte_quick(pmap, pv->pv_va);
4403 if ((*pte & PG_W) != 0)
4412 * Returns TRUE if the given page is mapped individually or as part of
4413 * a 4mpage. Otherwise, returns FALSE.
4416 pmap_page_is_mapped(vm_page_t m)
4420 if ((m->oflags & VPO_UNMANAGED) != 0)
4422 rw_wlock(&pvh_global_lock);
4423 rv = !TAILQ_EMPTY(&m->md.pv_list) ||
4424 ((m->flags & PG_FICTITIOUS) == 0 &&
4425 !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list));
4426 rw_wunlock(&pvh_global_lock);
4431 * Remove all pages from specified address space
4432 * this aids process exit speeds. Also, this code
4433 * is special cased for current process only, but
4434 * can have the more generic (and slightly slower)
4435 * mode enabled. This is much faster than pmap_remove
4436 * in the case of running down an entire address space.
4439 pmap_remove_pages(pmap_t pmap)
4441 pt_entry_t *pte, tpte;
4442 vm_page_t free = NULL;
4443 vm_page_t m, mpte, mt;
4445 struct md_page *pvh;
4446 struct pv_chunk *pc, *npc;
4449 uint32_t inuse, bitmask;
4452 if (pmap != PCPU_GET(curpmap)) {
4453 printf("warning: pmap_remove_pages called with non-current pmap\n");
4456 rw_wlock(&pvh_global_lock);
4459 TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) {
4460 KASSERT(pc->pc_pmap == pmap, ("Wrong pmap %p %p", pmap,
4463 for (field = 0; field < _NPCM; field++) {
4464 inuse = ~pc->pc_map[field] & pc_freemask[field];
4465 while (inuse != 0) {
4467 bitmask = 1UL << bit;
4468 idx = field * 32 + bit;
4469 pv = &pc->pc_pventry[idx];
4472 pte = pmap_pde(pmap, pv->pv_va);
4474 if ((tpte & PG_PS) == 0) {
4475 pte = vtopte(pv->pv_va);
4476 tpte = *pte & ~PG_PTE_PAT;
4481 "TPTE at %p IS ZERO @ VA %08x\n",
4487 * We cannot remove wired pages from a process' mapping at this time
4494 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
4495 KASSERT(m->phys_addr == (tpte & PG_FRAME),
4496 ("vm_page_t %p phys_addr mismatch %016jx %016jx",
4497 m, (uintmax_t)m->phys_addr,
4500 KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
4501 m < &vm_page_array[vm_page_array_size],
4502 ("pmap_remove_pages: bad tpte %#jx",
4508 * Update the vm_page_t clean/reference bits.
4510 if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
4511 if ((tpte & PG_PS) != 0) {
4512 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
4519 PV_STAT(pv_entry_frees++);
4520 PV_STAT(pv_entry_spare++);
4522 pc->pc_map[field] |= bitmask;
4523 if ((tpte & PG_PS) != 0) {
4524 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
4525 pvh = pa_to_pvh(tpte & PG_PS_FRAME);
4526 TAILQ_REMOVE(&pvh->pv_list, pv, pv_list);
4527 if (TAILQ_EMPTY(&pvh->pv_list)) {
4528 for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
4529 if (TAILQ_EMPTY(&mt->md.pv_list))
4530 vm_page_aflag_clear(mt, PGA_WRITEABLE);
4532 mpte = pmap_lookup_pt_page(pmap, pv->pv_va);
4534 pmap_remove_pt_page(pmap, mpte);
4535 pmap->pm_stats.resident_count--;
4536 KASSERT(mpte->wire_count == NPTEPG,
4537 ("pmap_remove_pages: pte page wire count error"));
4538 mpte->wire_count = 0;
4539 pmap_add_delayed_free_list(mpte, &free, FALSE);
4540 atomic_subtract_int(&cnt.v_wire_count, 1);
4543 pmap->pm_stats.resident_count--;
4544 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
4545 if (TAILQ_EMPTY(&m->md.pv_list) &&
4546 (m->flags & PG_FICTITIOUS) == 0) {
4547 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4548 if (TAILQ_EMPTY(&pvh->pv_list))
4549 vm_page_aflag_clear(m, PGA_WRITEABLE);
4551 pmap_unuse_pt(pmap, pv->pv_va, &free);
4556 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4561 pmap_invalidate_all(pmap);
4562 rw_wunlock(&pvh_global_lock);
4564 pmap_free_zero_pages(free);
4570 * Return whether or not the specified physical page was modified
4571 * in any physical maps.
4574 pmap_is_modified(vm_page_t m)
4578 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4579 ("pmap_is_modified: page %p is not managed", m));
4582 * If the page is not VPO_BUSY, then PGA_WRITEABLE cannot be
4583 * concurrently set while the object is locked. Thus, if PGA_WRITEABLE
4584 * is clear, no PTEs can have PG_M set.
4586 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
4587 if ((m->oflags & VPO_BUSY) == 0 &&
4588 (m->aflags & PGA_WRITEABLE) == 0)
4590 rw_wlock(&pvh_global_lock);
4591 rv = pmap_is_modified_pvh(&m->md) ||
4592 ((m->flags & PG_FICTITIOUS) == 0 &&
4593 pmap_is_modified_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m))));
4594 rw_wunlock(&pvh_global_lock);
4599 * Returns TRUE if any of the given mappings were used to modify
4600 * physical memory. Otherwise, returns FALSE. Both page and 2mpage
4601 * mappings are supported.
4604 pmap_is_modified_pvh(struct md_page *pvh)
4611 rw_assert(&pvh_global_lock, RA_WLOCKED);
4614 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) {
4617 pte = pmap_pte_quick(pmap, pv->pv_va);
4618 rv = (*pte & (PG_M | PG_RW)) == (PG_M | PG_RW);
4628 * pmap_is_prefaultable:
4630 * Return whether or not the specified virtual address is elgible
4634 pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr)
4642 pde = pmap_pde(pmap, addr);
4643 if (*pde != 0 && (*pde & PG_PS) == 0) {
4652 * pmap_is_referenced:
4654 * Return whether or not the specified physical page was referenced
4655 * in any physical maps.
4658 pmap_is_referenced(vm_page_t m)
4662 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4663 ("pmap_is_referenced: page %p is not managed", m));
4664 rw_wlock(&pvh_global_lock);
4665 rv = pmap_is_referenced_pvh(&m->md) ||
4666 ((m->flags & PG_FICTITIOUS) == 0 &&
4667 pmap_is_referenced_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m))));
4668 rw_wunlock(&pvh_global_lock);
4673 * Returns TRUE if any of the given mappings were referenced and FALSE
4674 * otherwise. Both page and 4mpage mappings are supported.
4677 pmap_is_referenced_pvh(struct md_page *pvh)
4684 rw_assert(&pvh_global_lock, RA_WLOCKED);
4687 TAILQ_FOREACH(pv, &pvh->pv_list, pv_list) {
4690 pte = pmap_pte_quick(pmap, pv->pv_va);
4691 rv = (*pte & (PG_A | PG_V)) == (PG_A | PG_V);
4701 * Clear the write and modified bits in each of the given page's mappings.
4704 pmap_remove_write(vm_page_t m)
4706 struct md_page *pvh;
4707 pv_entry_t next_pv, pv;
4710 pt_entry_t oldpte, *pte;
4713 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4714 ("pmap_remove_write: page %p is not managed", m));
4717 * If the page is not VPO_BUSY, then PGA_WRITEABLE cannot be set by
4718 * another thread while the object is locked. Thus, if PGA_WRITEABLE
4719 * is clear, no page table entries need updating.
4721 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
4722 if ((m->oflags & VPO_BUSY) == 0 &&
4723 (m->aflags & PGA_WRITEABLE) == 0)
4725 rw_wlock(&pvh_global_lock);
4727 if ((m->flags & PG_FICTITIOUS) != 0)
4728 goto small_mappings;
4729 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4730 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_list, next_pv) {
4734 pde = pmap_pde(pmap, va);
4735 if ((*pde & PG_RW) != 0)
4736 (void)pmap_demote_pde(pmap, pde, va);
4740 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4743 pde = pmap_pde(pmap, pv->pv_va);
4744 KASSERT((*pde & PG_PS) == 0, ("pmap_clear_write: found"
4745 " a 4mpage in page %p's pv list", m));
4746 pte = pmap_pte_quick(pmap, pv->pv_va);
4749 if ((oldpte & PG_RW) != 0) {
4751 * Regardless of whether a pte is 32 or 64 bits
4752 * in size, PG_RW and PG_M are among the least
4753 * significant 32 bits.
4755 if (!atomic_cmpset_int((u_int *)pte, oldpte,
4756 oldpte & ~(PG_RW | PG_M)))
4758 if ((oldpte & PG_M) != 0)
4760 pmap_invalidate_page(pmap, pv->pv_va);
4764 vm_page_aflag_clear(m, PGA_WRITEABLE);
4766 rw_wunlock(&pvh_global_lock);
4770 * pmap_ts_referenced:
4772 * Return a count of reference bits for a page, clearing those bits.
4773 * It is not necessary for every reference bit to be cleared, but it
4774 * is necessary that 0 only be returned when there are truly no
4775 * reference bits set.
4777 * XXX: The exact number of bits to check and clear is a matter that
4778 * should be tested and standardized at some point in the future for
4779 * optimal aging of shared pages.
4782 pmap_ts_referenced(vm_page_t m)
4784 struct md_page *pvh;
4785 pv_entry_t pv, pvf, pvn;
4787 pd_entry_t oldpde, *pde;
4792 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4793 ("pmap_ts_referenced: page %p is not managed", m));
4794 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4795 rw_wlock(&pvh_global_lock);
4797 if ((m->flags & PG_FICTITIOUS) != 0)
4798 goto small_mappings;
4799 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_list, pvn) {
4803 pde = pmap_pde(pmap, va);
4805 if ((oldpde & PG_A) != 0) {
4806 if (pmap_demote_pde(pmap, pde, va)) {
4807 if ((oldpde & PG_W) == 0) {
4809 * Remove the mapping to a single page
4810 * so that a subsequent access may
4811 * repromote. Since the underlying
4812 * page table page is fully populated,
4813 * this removal never frees a page
4816 va += VM_PAGE_TO_PHYS(m) - (oldpde &
4818 pmap_remove_page(pmap, va, NULL);
4830 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4833 pvn = TAILQ_NEXT(pv, pv_list);
4834 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
4835 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
4838 pde = pmap_pde(pmap, pv->pv_va);
4839 KASSERT((*pde & PG_PS) == 0, ("pmap_ts_referenced:"
4840 " found a 4mpage in page %p's pv list", m));
4841 pte = pmap_pte_quick(pmap, pv->pv_va);
4842 if ((*pte & PG_A) != 0) {
4843 atomic_clear_int((u_int *)pte, PG_A);
4844 pmap_invalidate_page(pmap, pv->pv_va);
4850 } while ((pv = pvn) != NULL && pv != pvf);
4854 rw_wunlock(&pvh_global_lock);
4859 * Clear the modify bits on the specified physical page.
4862 pmap_clear_modify(vm_page_t m)
4864 struct md_page *pvh;
4865 pv_entry_t next_pv, pv;
4867 pd_entry_t oldpde, *pde;
4868 pt_entry_t oldpte, *pte;
4871 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4872 ("pmap_clear_modify: page %p is not managed", m));
4873 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
4874 KASSERT((m->oflags & VPO_BUSY) == 0,
4875 ("pmap_clear_modify: page %p is busy", m));
4878 * If the page is not PGA_WRITEABLE, then no PTEs can have PG_M set.
4879 * If the object containing the page is locked and the page is not
4880 * VPO_BUSY, then PGA_WRITEABLE cannot be concurrently set.
4882 if ((m->aflags & PGA_WRITEABLE) == 0)
4884 rw_wlock(&pvh_global_lock);
4886 if ((m->flags & PG_FICTITIOUS) != 0)
4887 goto small_mappings;
4888 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4889 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_list, next_pv) {
4893 pde = pmap_pde(pmap, va);
4895 if ((oldpde & PG_RW) != 0) {
4896 if (pmap_demote_pde(pmap, pde, va)) {
4897 if ((oldpde & PG_W) == 0) {
4899 * Write protect the mapping to a
4900 * single page so that a subsequent
4901 * write access may repromote.
4903 va += VM_PAGE_TO_PHYS(m) - (oldpde &
4905 pte = pmap_pte_quick(pmap, va);
4907 if ((oldpte & PG_V) != 0) {
4909 * Regardless of whether a pte is 32 or 64 bits
4910 * in size, PG_RW and PG_M are among the least
4911 * significant 32 bits.
4913 while (!atomic_cmpset_int((u_int *)pte,
4915 oldpte & ~(PG_M | PG_RW)))
4918 pmap_invalidate_page(pmap, va);
4926 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4929 pde = pmap_pde(pmap, pv->pv_va);
4930 KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found"
4931 " a 4mpage in page %p's pv list", m));
4932 pte = pmap_pte_quick(pmap, pv->pv_va);
4933 if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
4935 * Regardless of whether a pte is 32 or 64 bits
4936 * in size, PG_M is among the least significant
4939 atomic_clear_int((u_int *)pte, PG_M);
4940 pmap_invalidate_page(pmap, pv->pv_va);
4945 rw_wunlock(&pvh_global_lock);
4949 * pmap_clear_reference:
4951 * Clear the reference bit on the specified physical page.
4954 pmap_clear_reference(vm_page_t m)
4956 struct md_page *pvh;
4957 pv_entry_t next_pv, pv;
4959 pd_entry_t oldpde, *pde;
4963 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4964 ("pmap_clear_reference: page %p is not managed", m));
4965 rw_wlock(&pvh_global_lock);
4967 if ((m->flags & PG_FICTITIOUS) != 0)
4968 goto small_mappings;
4969 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4970 TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_list, next_pv) {
4974 pde = pmap_pde(pmap, va);
4976 if ((oldpde & PG_A) != 0) {
4977 if (pmap_demote_pde(pmap, pde, va)) {
4979 * Remove the mapping to a single page so
4980 * that a subsequent access may repromote.
4981 * Since the underlying page table page is
4982 * fully populated, this removal never frees
4983 * a page table page.
4985 va += VM_PAGE_TO_PHYS(m) - (oldpde &
4987 pmap_remove_page(pmap, va, NULL);
4993 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4996 pde = pmap_pde(pmap, pv->pv_va);
4997 KASSERT((*pde & PG_PS) == 0, ("pmap_clear_reference: found"
4998 " a 4mpage in page %p's pv list", m));
4999 pte = pmap_pte_quick(pmap, pv->pv_va);
5000 if ((*pte & PG_A) != 0) {
5002 * Regardless of whether a pte is 32 or 64 bits
5003 * in size, PG_A is among the least significant
5006 atomic_clear_int((u_int *)pte, PG_A);
5007 pmap_invalidate_page(pmap, pv->pv_va);
5012 rw_wunlock(&pvh_global_lock);
5016 * Miscellaneous support routines follow
5019 /* Adjust the cache mode for a 4KB page mapped via a PTE. */
5020 static __inline void
5021 pmap_pte_attr(pt_entry_t *pte, int cache_bits)
5026 * The cache mode bits are all in the low 32-bits of the
5027 * PTE, so we can just spin on updating the low 32-bits.
5030 opte = *(u_int *)pte;
5031 npte = opte & ~PG_PTE_CACHE;
5033 } while (npte != opte && !atomic_cmpset_int((u_int *)pte, opte, npte));
5036 /* Adjust the cache mode for a 2/4MB page mapped via a PDE. */
5037 static __inline void
5038 pmap_pde_attr(pd_entry_t *pde, int cache_bits)
5043 * The cache mode bits are all in the low 32-bits of the
5044 * PDE, so we can just spin on updating the low 32-bits.
5047 opde = *(u_int *)pde;
5048 npde = opde & ~PG_PDE_CACHE;
5050 } while (npde != opde && !atomic_cmpset_int((u_int *)pde, opde, npde));
5054 * Map a set of physical memory pages into the kernel virtual
5055 * address space. Return a pointer to where it is mapped. This
5056 * routine is intended to be used for mapping device memory,
5060 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5062 vm_offset_t va, offset;
5065 offset = pa & PAGE_MASK;
5066 size = roundup(offset + size, PAGE_SIZE);
5069 if (pa < KERNLOAD && pa + size <= KERNLOAD)
5072 va = kmem_alloc_nofault(kernel_map, size);
5074 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5076 for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE)
5077 pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode);
5078 pmap_invalidate_range(kernel_pmap, va, va + tmpsize);
5079 pmap_invalidate_cache_range(va, va + size);
5080 return ((void *)(va + offset));
5084 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5087 return (pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5091 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5094 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5098 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5100 vm_offset_t base, offset;
5102 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD)
5104 base = trunc_page(va);
5105 offset = va & PAGE_MASK;
5106 size = roundup(offset + size, PAGE_SIZE);
5107 kmem_free(kernel_map, base, size);
5111 * Sets the memory attribute for the specified page.
5114 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5117 m->md.pat_mode = ma;
5118 if ((m->flags & PG_FICTITIOUS) != 0)
5122 * If "m" is a normal page, flush it from the cache.
5123 * See pmap_invalidate_cache_range().
5125 * First, try to find an existing mapping of the page by sf
5126 * buffer. sf_buf_invalidate_cache() modifies mapping and
5127 * flushes the cache.
5129 if (sf_buf_invalidate_cache(m))
5133 * If page is not mapped by sf buffer, but CPU does not
5134 * support self snoop, map the page transient and do
5135 * invalidation. In the worst case, whole cache is flushed by
5136 * pmap_invalidate_cache_range().
5138 if ((cpu_feature & CPUID_SS) == 0)
5143 pmap_flush_page(vm_page_t m)
5145 struct sysmaps *sysmaps;
5146 vm_offset_t sva, eva;
5148 if ((cpu_feature & CPUID_CLFSH) != 0) {
5149 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
5150 mtx_lock(&sysmaps->lock);
5151 if (*sysmaps->CMAP2)
5152 panic("pmap_flush_page: CMAP2 busy");
5154 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) |
5155 PG_A | PG_M | pmap_cache_bits(m->md.pat_mode, 0);
5156 invlcaddr(sysmaps->CADDR2);
5157 sva = (vm_offset_t)sysmaps->CADDR2;
5158 eva = sva + PAGE_SIZE;
5161 * Use mfence despite the ordering implied by
5162 * mtx_{un,}lock() because clflush is not guaranteed
5163 * to be ordered by any other instruction.
5166 for (; sva < eva; sva += cpu_clflush_line_size)
5169 *sysmaps->CMAP2 = 0;
5171 mtx_unlock(&sysmaps->lock);
5173 pmap_invalidate_cache();
5177 * Changes the specified virtual address range's memory type to that given by
5178 * the parameter "mode". The specified virtual address range must be
5179 * completely contained within either the kernel map.
5181 * Returns zero if the change completed successfully, and either EINVAL or
5182 * ENOMEM if the change failed. Specifically, EINVAL is returned if some part
5183 * of the virtual address range was not mapped, and ENOMEM is returned if
5184 * there was insufficient memory available to complete the change.
5187 pmap_change_attr(vm_offset_t va, vm_size_t size, int mode)
5189 vm_offset_t base, offset, tmpva;
5192 int cache_bits_pte, cache_bits_pde;
5195 base = trunc_page(va);
5196 offset = va & PAGE_MASK;
5197 size = roundup(offset + size, PAGE_SIZE);
5200 * Only supported on kernel virtual addresses above the recursive map.
5202 if (base < VM_MIN_KERNEL_ADDRESS)
5205 cache_bits_pde = pmap_cache_bits(mode, 1);
5206 cache_bits_pte = pmap_cache_bits(mode, 0);
5210 * Pages that aren't mapped aren't supported. Also break down
5211 * 2/4MB pages into 4KB pages if required.
5213 PMAP_LOCK(kernel_pmap);
5214 for (tmpva = base; tmpva < base + size; ) {
5215 pde = pmap_pde(kernel_pmap, tmpva);
5217 PMAP_UNLOCK(kernel_pmap);
5222 * If the current 2/4MB page already has
5223 * the required memory type, then we need not
5224 * demote this page. Just increment tmpva to
5225 * the next 2/4MB page frame.
5227 if ((*pde & PG_PDE_CACHE) == cache_bits_pde) {
5228 tmpva = trunc_4mpage(tmpva) + NBPDR;
5233 * If the current offset aligns with a 2/4MB
5234 * page frame and there is at least 2/4MB left
5235 * within the range, then we need not break
5236 * down this page into 4KB pages.
5238 if ((tmpva & PDRMASK) == 0 &&
5239 tmpva + PDRMASK < base + size) {
5243 if (!pmap_demote_pde(kernel_pmap, pde, tmpva)) {
5244 PMAP_UNLOCK(kernel_pmap);
5248 pte = vtopte(tmpva);
5250 PMAP_UNLOCK(kernel_pmap);
5255 PMAP_UNLOCK(kernel_pmap);
5258 * Ok, all the pages exist, so run through them updating their
5259 * cache mode if required.
5261 for (tmpva = base; tmpva < base + size; ) {
5262 pde = pmap_pde(kernel_pmap, tmpva);
5264 if ((*pde & PG_PDE_CACHE) != cache_bits_pde) {
5265 pmap_pde_attr(pde, cache_bits_pde);
5268 tmpva = trunc_4mpage(tmpva) + NBPDR;
5270 pte = vtopte(tmpva);
5271 if ((*pte & PG_PTE_CACHE) != cache_bits_pte) {
5272 pmap_pte_attr(pte, cache_bits_pte);
5280 * Flush CPU caches to make sure any data isn't cached that
5281 * shouldn't be, etc.
5284 pmap_invalidate_range(kernel_pmap, base, tmpva);
5285 pmap_invalidate_cache_range(base, tmpva);
5291 * perform the pmap work for mincore
5294 pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa)
5297 pt_entry_t *ptep, pte;
5303 pdep = pmap_pde(pmap, addr);
5305 if (*pdep & PG_PS) {
5307 /* Compute the physical address of the 4KB page. */
5308 pa = ((*pdep & PG_PS_FRAME) | (addr & PDRMASK)) &
5310 val = MINCORE_SUPER;
5312 ptep = pmap_pte(pmap, addr);
5314 pmap_pte_release(ptep);
5315 pa = pte & PG_FRAME;
5323 if ((pte & PG_V) != 0) {
5324 val |= MINCORE_INCORE;
5325 if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5326 val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
5327 if ((pte & PG_A) != 0)
5328 val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
5330 if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
5331 (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
5332 (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) {
5333 /* Ensure that "PHYS_TO_VM_PAGE(pa)->object" doesn't change. */
5334 if (vm_page_pa_tryrelock(pmap, pa, locked_pa))
5337 PA_UNLOCK_COND(*locked_pa);
5343 pmap_activate(struct thread *td)
5345 pmap_t pmap, oldpmap;
5350 pmap = vmspace_pmap(td->td_proc->p_vmspace);
5351 oldpmap = PCPU_GET(curpmap);
5352 cpuid = PCPU_GET(cpuid);
5354 CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active);
5355 CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
5357 CPU_CLR(cpuid, &oldpmap->pm_active);
5358 CPU_SET(cpuid, &pmap->pm_active);
5361 cr3 = vtophys(pmap->pm_pdpt);
5363 cr3 = vtophys(pmap->pm_pdir);
5366 * pmap_activate is for the current thread on the current cpu
5368 td->td_pcb->pcb_cr3 = cr3;
5370 PCPU_SET(curpmap, pmap);
5375 pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
5380 * Increase the starting virtual address of the given mapping if a
5381 * different alignment might result in more superpage mappings.
5384 pmap_align_superpage(vm_object_t object, vm_ooffset_t offset,
5385 vm_offset_t *addr, vm_size_t size)
5387 vm_offset_t superpage_offset;
5391 if (object != NULL && (object->flags & OBJ_COLORED) != 0)
5392 offset += ptoa(object->pg_color);
5393 superpage_offset = offset & PDRMASK;
5394 if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR ||
5395 (*addr & PDRMASK) == superpage_offset)
5397 if ((*addr & PDRMASK) < superpage_offset)
5398 *addr = (*addr & ~PDRMASK) + superpage_offset;
5400 *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset;
5404 #if defined(PMAP_DEBUG)
5405 pmap_pid_dump(int pid)
5412 sx_slock(&allproc_lock);
5413 FOREACH_PROC_IN_SYSTEM(p) {
5414 if (p->p_pid != pid)
5420 pmap = vmspace_pmap(p->p_vmspace);
5421 for (i = 0; i < NPDEPTD; i++) {
5424 vm_offset_t base = i << PDRSHIFT;
5426 pde = &pmap->pm_pdir[i];
5427 if (pde && pmap_pde_v(pde)) {
5428 for (j = 0; j < NPTEPG; j++) {
5429 vm_offset_t va = base + (j << PAGE_SHIFT);
5430 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) {
5435 sx_sunlock(&allproc_lock);
5438 pte = pmap_pte(pmap, va);
5439 if (pte && pmap_pte_v(pte)) {
5443 m = PHYS_TO_VM_PAGE(pa & PG_FRAME);
5444 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x",
5445 va, pa, m->hold_count, m->wire_count, m->flags);
5460 sx_sunlock(&allproc_lock);
5467 static void pads(pmap_t pm);
5468 void pmap_pvdump(vm_paddr_t pa);
5470 /* print address space of pmap*/
5478 if (pm == kernel_pmap)
5480 for (i = 0; i < NPDEPTD; i++)
5482 for (j = 0; j < NPTEPG; j++) {
5483 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
5484 if (pm == kernel_pmap && va < KERNBASE)
5486 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
5488 ptep = pmap_pte(pm, va);
5489 if (pmap_pte_v(ptep))
5490 printf("%x:%x ", va, *ptep);
5496 pmap_pvdump(vm_paddr_t pa)
5502 printf("pa %x", pa);
5503 m = PHYS_TO_VM_PAGE(pa);
5504 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5506 printf(" -> pmap %p, va %x", (void *)pmap, pv->pv_va);