4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
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
44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
45 * $DragonFly: src/sys/platform/pc32/i386/pmap.c,v 1.87 2008/08/25 17:01:38 dillon Exp $
49 * Manages physical address maps.
51 * In most cases the vm_token must be held when manipulating a user pmap
52 * or elements within a vm_page, and the kvm_token must be held when
53 * manipulating the kernel pmap. Operations on user pmaps may require
54 * additional synchronization.
56 * In some cases the caller may hold the required tokens to prevent pmap
57 * functions from blocking on those same tokens. This typically only works
58 * for lookup-style operations.
61 * PMAP_DEBUG - see platform/pc32/include/pmap.h
64 #include "opt_disable_pse.h"
66 #include "opt_msgbuf.h"
68 #include <sys/param.h>
69 #include <sys/systm.h>
70 #include <sys/kernel.h>
72 #include <sys/msgbuf.h>
73 #include <sys/vmmeter.h>
77 #include <vm/vm_param.h>
78 #include <sys/sysctl.h>
80 #include <vm/vm_kern.h>
81 #include <vm/vm_page.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_extern.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_zone.h>
90 #include <sys/thread2.h>
91 #include <sys/sysref2.h>
93 #include <machine/cputypes.h>
94 #include <machine/md_var.h>
95 #include <machine/specialreg.h>
96 #include <machine/smp.h>
97 #include <machine_base/apic/apicreg.h>
98 #include <machine/globaldata.h>
99 #include <machine/pmap.h>
100 #include <machine/pmap_inval.h>
102 #define PMAP_KEEP_PDIRS
103 #ifndef PMAP_SHPGPERPROC
104 #define PMAP_SHPGPERPROC 200
107 #if defined(DIAGNOSTIC)
108 #define PMAP_DIAGNOSTIC
113 #if !defined(PMAP_DIAGNOSTIC)
114 #define PMAP_INLINE __inline
120 * Get PDEs and PTEs for user/kernel address space
122 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
123 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
125 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
126 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
127 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
128 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
129 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
132 * Given a map and a machine independent protection code,
133 * convert to a vax protection code.
135 #define pte_prot(m, p) \
136 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
137 static int protection_codes[8];
139 struct pmap kernel_pmap;
140 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
142 vm_paddr_t avail_start; /* PA of first available physical page */
143 vm_paddr_t avail_end; /* PA of last available physical page */
144 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
145 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
146 vm_offset_t virtual2_start;
147 vm_offset_t virtual2_end;
148 vm_offset_t KvaStart; /* VA start of KVA space */
149 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
150 vm_offset_t KvaSize; /* max size of kernel virtual address space */
151 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
152 static int pgeflag; /* PG_G or-in */
153 static int pseflag; /* PG_PS or-in */
155 static vm_object_t kptobj;
158 vm_offset_t kernel_vm_end;
161 * Data for the pv entry allocation mechanism
163 static vm_zone_t pvzone;
164 static struct vm_zone pvzone_store;
165 static struct vm_object pvzone_obj;
166 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
167 static int pmap_pagedaemon_waken = 0;
168 static struct pv_entry *pvinit;
171 * Considering all the issues I'm having with pmap caching, if breakage
172 * continues to occur, and for debugging, I've added a sysctl that will
173 * just do an unconditional invltlb.
175 static int dreadful_invltlb;
177 SYSCTL_INT(_vm, OID_AUTO, dreadful_invltlb,
178 CTLFLAG_RW, &dreadful_invltlb, 0, "");
181 * All those kernel PT submaps that BSD is so fond of
183 pt_entry_t *CMAP1 = 0, *ptmmap;
184 caddr_t CADDR1 = 0, ptvmmap = 0;
185 static pt_entry_t *msgbufmap;
186 struct msgbuf *msgbufp=0;
191 static pt_entry_t *pt_crashdumpmap;
192 static caddr_t crashdumpmap;
194 extern pt_entry_t *SMPpt;
196 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
197 static unsigned * get_ptbase (pmap_t pmap);
198 static pv_entry_t get_pv_entry (void);
199 static void i386_protection_init (void);
200 static __inline void pmap_clearbit (vm_page_t m, int bit);
202 static void pmap_remove_all (vm_page_t m);
203 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
204 vm_offset_t sva, pmap_inval_info_t info);
205 static void pmap_remove_page (struct pmap *pmap,
206 vm_offset_t va, pmap_inval_info_t info);
207 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
208 vm_offset_t va, pmap_inval_info_t info);
209 static boolean_t pmap_testbit (vm_page_t m, int bit);
210 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
211 vm_page_t mpte, vm_page_t m);
213 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
215 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
216 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
217 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
218 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
219 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
220 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
222 static unsigned pdir4mb;
225 * Move the kernel virtual free pointer to the next
226 * 4MB. This is used to help improve performance
227 * by using a large (4MB) page for much of the kernel
228 * (.text, .data, .bss)
232 pmap_kmem_choose(vm_offset_t addr)
234 vm_offset_t newaddr = addr;
236 if (cpu_feature & CPUID_PSE) {
237 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
244 * This function returns a pointer to the pte entry in the pmap and has
245 * the side effect of potentially retaining a cached mapping of the pmap.
247 * The caller must hold vm_token and the returned value is only valid
248 * until the caller blocks or releases the token.
252 pmap_pte(pmap_t pmap, vm_offset_t va)
256 ASSERT_LWKT_TOKEN_HELD(&vm_token);
258 pdeaddr = (unsigned *) pmap_pde(pmap, va);
259 if (*pdeaddr & PG_PS)
262 return get_ptbase(pmap) + i386_btop(va);
268 * pmap_pte using the kernel_pmap
270 * Used for debugging, no requirements.
273 pmap_kernel_pte(vm_offset_t va)
277 pdeaddr = (unsigned *) pmap_pde(&kernel_pmap, va);
278 if (*pdeaddr & PG_PS)
281 return (unsigned *)vtopte(va);
288 * Super fast pmap_pte routine best used when scanning the pv lists.
289 * This eliminates many course-grained invltlb calls. Note that many of
290 * the pv list scans are across different pmaps and it is very wasteful
291 * to do an entire invltlb when checking a single mapping.
293 * Should only be called while in a critical section.
295 * The caller must hold vm_token and the returned value is only valid
296 * until the caller blocks or releases the token.
300 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
302 struct mdglobaldata *gd = mdcpu;
305 ASSERT_LWKT_TOKEN_HELD(&vm_token);
306 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
307 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
308 unsigned index = i386_btop(va);
309 /* are we current address space or kernel? */
310 if ((pmap == &kernel_pmap) ||
311 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
312 return (unsigned *) PTmap + index;
314 newpf = pde & PG_FRAME;
315 if (((*(unsigned *)gd->gd_PMAP1) & PG_FRAME) != newpf) {
316 *(unsigned *)gd->gd_PMAP1 = newpf | PG_RW | PG_V;
317 cpu_invlpg(gd->gd_PADDR1);
319 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
326 * Bootstrap the system enough to run with virtual memory.
328 * On the i386 this is called after mapping has already been enabled
329 * and just syncs the pmap module with what has already been done.
330 * [We can't call it easily with mapping off since the kernel is not
331 * mapped with PA == VA, hence we would have to relocate every address
332 * from the linked base (virtual) address "KERNBASE" to the actual
333 * (physical) address starting relative to 0]
336 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
340 struct mdglobaldata *gd;
344 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
345 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
346 KvaEnd = KvaStart + KvaSize;
348 avail_start = firstaddr;
351 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
352 * too large. It should instead be correctly calculated in locore.s and
353 * not based on 'first' (which is a physical address, not a virtual
354 * address, for the start of unused physical memory). The kernel
355 * page tables are NOT double mapped and thus should not be included
356 * in this calculation.
358 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
359 virtual_start = pmap_kmem_choose(virtual_start);
360 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
363 * Initialize protection array.
365 i386_protection_init();
368 * The kernel's pmap is statically allocated so we don't have to use
369 * pmap_create, which is unlikely to work correctly at this part of
370 * the boot sequence (XXX and which no longer exists).
372 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
373 kernel_pmap.pm_count = 1;
374 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
375 TAILQ_INIT(&kernel_pmap.pm_pvlist);
379 * Reserve some special page table entries/VA space for temporary
382 #define SYSMAP(c, p, v, n) \
383 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
386 pte = (pt_entry_t *) pmap_kernel_pte(va);
389 * CMAP1/CMAP2 are used for zeroing and copying pages.
391 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
396 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
399 * ptvmmap is used for reading arbitrary physical pages via
402 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
405 * msgbufp is used to map the system message buffer.
406 * XXX msgbufmap is not used.
408 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
409 atop(round_page(MSGBUF_SIZE)))
414 for (i = 0; i < NKPT; i++)
418 * PG_G is terribly broken on SMP because we IPI invltlb's in some
419 * cases rather then invl1pg. Actually, I don't even know why it
420 * works under UP because self-referential page table mappings
425 if (cpu_feature & CPUID_PGE)
430 * Initialize the 4MB page size flag
434 * The 4MB page version of the initial
435 * kernel page mapping.
439 #if !defined(DISABLE_PSE)
440 if (cpu_feature & CPUID_PSE) {
443 * Note that we have enabled PSE mode
446 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
447 ptditmp &= ~(NBPDR - 1);
448 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
453 * Enable the PSE mode. If we are SMP we can't do this
454 * now because the APs will not be able to use it when
457 load_cr4(rcr4() | CR4_PSE);
460 * We can do the mapping here for the single processor
461 * case. We simply ignore the old page table page from
465 * For SMP, we still need 4K pages to bootstrap APs,
466 * PSE will be enabled as soon as all APs are up.
468 PTD[KPTDI] = (pd_entry_t)ptditmp;
469 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
476 * We need to finish setting up the globaldata page for the BSP.
477 * locore has already populated the page table for the mdglobaldata
480 pg = MDGLOBALDATA_BASEALLOC_PAGES;
481 gd = &CPU_prvspace[0].mdglobaldata;
482 gd->gd_CMAP1 = &SMPpt[pg + 0];
483 gd->gd_CMAP2 = &SMPpt[pg + 1];
484 gd->gd_CMAP3 = &SMPpt[pg + 2];
485 gd->gd_PMAP1 = &SMPpt[pg + 3];
486 gd->gd_GDMAP1 = &PTD[APTDPTDI];
487 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
488 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
489 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
490 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
491 gd->gd_GDADDR1= (unsigned *)VADDR(APTDPTDI, 0);
498 * Set 4mb pdir for mp startup
503 if (pseflag && (cpu_feature & CPUID_PSE)) {
504 load_cr4(rcr4() | CR4_PSE);
505 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
506 kernel_pmap.pm_pdir[KPTDI] =
507 PTD[KPTDI] = (pd_entry_t)pdir4mb;
515 * Initialize the pmap module, called by vm_init()
517 * Called from the low level boot code only.
526 * object for kernel page table pages
528 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
531 * Allocate memory for random pmap data structures. Includes the
535 for(i = 0; i < vm_page_array_size; i++) {
538 m = &vm_page_array[i];
539 TAILQ_INIT(&m->md.pv_list);
540 m->md.pv_list_count = 0;
544 * init the pv free list
546 initial_pvs = vm_page_array_size;
547 if (initial_pvs < MINPV)
549 pvzone = &pvzone_store;
550 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
551 initial_pvs * sizeof (struct pv_entry));
552 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
556 * Now it is safe to enable pv_table recording.
558 pmap_initialized = TRUE;
562 * Initialize the address space (zone) for the pv_entries. Set a
563 * high water mark so that the system can recover from excessive
564 * numbers of pv entries.
566 * Called from the low level boot code only.
571 int shpgperproc = PMAP_SHPGPERPROC;
573 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
574 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
575 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
576 pv_entry_high_water = 9 * (pv_entry_max / 10);
577 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
581 /***************************************************
582 * Low level helper routines.....
583 ***************************************************/
588 test_m_maps_pv(vm_page_t m, pv_entry_t pv)
593 KKASSERT(pv->pv_m == m);
595 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
599 panic("test_m_maps_pv: failed m %p pv %p\n", m, pv);
603 ptbase_assert(struct pmap *pmap)
605 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
607 /* are we current address space or kernel? */
608 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME))
610 KKASSERT(frame == (*mycpu->gd_GDMAP1 & PG_FRAME));
615 #define test_m_maps_pv(m, pv)
616 #define ptbase_assert(pmap)
620 #if defined(PMAP_DIAGNOSTIC)
623 * This code checks for non-writeable/modified pages.
624 * This should be an invalid condition.
627 pmap_nw_modified(pt_entry_t ptea)
633 if ((pte & (PG_M|PG_RW)) == PG_M)
642 * This routine defines the region(s) of memory that should not be tested
643 * for the modified bit.
647 static PMAP_INLINE int
648 pmap_track_modified(vm_offset_t va)
650 if ((va < clean_sva) || (va >= clean_eva))
657 * Retrieve the mapped page table base for a particular pmap. Use our self
658 * mapping for the kernel_pmap or our current pmap.
660 * For foreign pmaps we use the per-cpu page table map. Since this involves
661 * installing a ptd it's actually (per-process x per-cpu). However, we
662 * still cannot depend on our mapping to survive thread switches because
663 * the process might be threaded and switching to another thread for the
664 * same process on the same cpu will allow that other thread to make its
667 * This could be a bit confusing but the jist is for something like the
668 * vkernel which uses foreign pmaps all the time this represents a pretty
669 * good cache that avoids unnecessary invltlb()s.
671 * The caller must hold vm_token and the returned value is only valid
672 * until the caller blocks or releases the token.
675 get_ptbase(pmap_t pmap)
677 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
678 struct mdglobaldata *gd = mdcpu;
680 ASSERT_LWKT_TOKEN_HELD(&vm_token);
683 * We can use PTmap if the pmap is our current address space or
684 * the kernel address space.
686 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
687 return (unsigned *) PTmap;
691 * Otherwise we use the per-cpu alternative page table map. Each
692 * cpu gets its own map. Because of this we cannot use this map
693 * from interrupts or threads which can preempt.
695 * Even if we already have the map cached we may still have to
696 * invalidate the TLB if another cpu modified a PDE in the map.
698 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
699 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
701 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
702 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
703 pmap->pm_cached |= gd->mi.gd_cpumask;
705 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
706 pmap->pm_cached |= gd->mi.gd_cpumask;
708 } else if (dreadful_invltlb) {
711 return ((unsigned *)gd->gd_GDADDR1);
717 * Extract the physical page address associated with the map/VA pair.
719 * The caller may hold vm_token if it desires non-blocking operation.
722 pmap_extract(pmap_t pmap, vm_offset_t va)
725 vm_offset_t pdirindex;
727 lwkt_gettoken(&vm_token);
728 pdirindex = va >> PDRSHIFT;
729 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
731 if ((rtval & PG_PS) != 0) {
732 rtval &= ~(NBPDR - 1);
733 rtval |= va & (NBPDR - 1);
735 pte = get_ptbase(pmap) + i386_btop(va);
736 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
741 lwkt_reltoken(&vm_token);
745 /***************************************************
746 * Low level mapping routines.....
747 ***************************************************/
750 * Map a wired VM page to a KVA, fully SMP synchronized.
752 * No requirements, non blocking.
755 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
759 pmap_inval_info info;
761 pmap_inval_init(&info);
762 npte = pa | PG_RW | PG_V | pgeflag;
763 pte = (unsigned *)vtopte(va);
764 pmap_inval_interlock(&info, &kernel_pmap, va);
766 pmap_inval_deinterlock(&info, &kernel_pmap);
767 pmap_inval_done(&info);
771 * Map a wired VM page to a KVA, synchronized on current cpu only.
773 * No requirements, non blocking.
776 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
781 npte = pa | PG_RW | PG_V | pgeflag;
782 pte = (unsigned *)vtopte(va);
784 cpu_invlpg((void *)va);
788 * Synchronize a previously entered VA on all cpus.
790 * No requirements, non blocking.
793 pmap_kenter_sync(vm_offset_t va)
795 pmap_inval_info info;
797 pmap_inval_init(&info);
798 pmap_inval_interlock(&info, &kernel_pmap, va);
799 pmap_inval_deinterlock(&info, &kernel_pmap);
800 pmap_inval_done(&info);
804 * Synchronize a previously entered VA on the current cpu only.
806 * No requirements, non blocking.
809 pmap_kenter_sync_quick(vm_offset_t va)
811 cpu_invlpg((void *)va);
815 * Remove a page from the kernel pagetables, fully SMP synchronized.
817 * No requirements, non blocking.
820 pmap_kremove(vm_offset_t va)
823 pmap_inval_info info;
825 pmap_inval_init(&info);
826 pte = (unsigned *)vtopte(va);
827 pmap_inval_interlock(&info, &kernel_pmap, va);
829 pmap_inval_deinterlock(&info, &kernel_pmap);
830 pmap_inval_done(&info);
834 * Remove a page from the kernel pagetables, synchronized on current cpu only.
836 * No requirements, non blocking.
839 pmap_kremove_quick(vm_offset_t va)
842 pte = (unsigned *)vtopte(va);
844 cpu_invlpg((void *)va);
848 * Adjust the permissions of a page in the kernel page table,
849 * synchronized on the current cpu only.
851 * No requirements, non blocking.
854 pmap_kmodify_rw(vm_offset_t va)
856 atomic_set_int(vtopte(va), PG_RW);
857 cpu_invlpg((void *)va);
861 * Adjust the permissions of a page in the kernel page table,
862 * synchronized on the current cpu only.
864 * No requirements, non blocking.
867 pmap_kmodify_nc(vm_offset_t va)
869 atomic_set_int(vtopte(va), PG_N);
870 cpu_invlpg((void *)va);
874 * Map a range of physical addresses into kernel virtual address space.
876 * No requirements, non blocking.
879 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
881 vm_offset_t sva, virt;
884 while (start < end) {
885 pmap_kenter(virt, start);
894 * Add a list of wired pages to the kva, fully SMP synchronized.
896 * No requirements, non blocking.
899 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
903 end_va = va + count * PAGE_SIZE;
905 while (va < end_va) {
908 pte = (unsigned *)vtopte(va);
909 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
910 cpu_invlpg((void *)va);
915 smp_invltlb(); /* XXX */
920 * Remove pages from KVA, fully SMP synchronized.
922 * No requirements, non blocking.
925 pmap_qremove(vm_offset_t va, int count)
929 end_va = va + count*PAGE_SIZE;
931 while (va < end_va) {
934 pte = (unsigned *)vtopte(va);
936 cpu_invlpg((void *)va);
945 * This routine works like vm_page_lookup() but also blocks as long as the
946 * page is busy. This routine does not busy the page it returns.
948 * The caller must hold vm_token.
951 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
955 ASSERT_LWKT_TOKEN_HELD(&vm_token);
957 m = vm_page_lookup(object, pindex);
958 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
964 * Create a new thread and optionally associate it with a (new) process.
965 * NOTE! the new thread's cpu may not equal the current cpu.
968 pmap_init_thread(thread_t td)
970 /* enforce pcb placement */
971 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
972 td->td_savefpu = &td->td_pcb->pcb_save;
973 td->td_sp = (char *)td->td_pcb - 16;
977 * This routine directly affects the fork perf for a process.
980 pmap_init_proc(struct proc *p)
985 * Dispose the UPAGES for a process that has exited.
986 * This routine directly impacts the exit perf of a process.
989 pmap_dispose_proc(struct proc *p)
991 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
994 /***************************************************
995 * Page table page management routines.....
996 ***************************************************/
999 * This routine unholds page table pages, and if the hold count
1000 * drops to zero, then it decrements the wire count.
1002 * The caller must hold vm_token.
1003 * This function can block.
1006 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1009 * Wait until we can busy the page ourselves. We cannot have
1010 * any active flushes if we block.
1012 if (m->flags & PG_BUSY) {
1013 pmap_inval_flush(info);
1014 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1017 KASSERT(m->queue == PQ_NONE,
1018 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1020 if (m->hold_count == 1) {
1022 * Unmap the page table page.
1024 * NOTE: We must clear pm_cached for all cpus, including
1025 * the current one, when clearing a page directory
1029 pmap_inval_interlock(info, pmap, -1);
1030 KKASSERT(pmap->pm_pdir[m->pindex]);
1031 pmap->pm_pdir[m->pindex] = 0;
1032 pmap->pm_cached = 0;
1033 pmap_inval_deinterlock(info, pmap);
1035 KKASSERT(pmap->pm_stats.resident_count > 0);
1036 --pmap->pm_stats.resident_count;
1038 if (pmap->pm_ptphint == m)
1039 pmap->pm_ptphint = NULL;
1042 * This was our last hold, the page had better be unwired
1043 * after we decrement wire_count.
1045 * FUTURE NOTE: shared page directory page could result in
1046 * multiple wire counts.
1050 KKASSERT(m->wire_count == 0);
1051 --vmstats.v_wire_count;
1052 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1054 vm_page_free_zero(m);
1057 KKASSERT(m->hold_count > 1);
1064 * The caller must hold vm_token.
1065 * This function can block.
1067 static PMAP_INLINE int
1068 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1070 KKASSERT(m->hold_count > 0);
1071 if (m->hold_count > 1) {
1075 return _pmap_unwire_pte_hold(pmap, m, info);
1080 * After removing a (user) page table entry, this routine is used to
1081 * conditionally free the page, and manage the hold/wire counts.
1083 * The caller must hold vm_token.
1084 * This function can block regardless.
1087 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1088 pmap_inval_info_t info)
1092 if (va >= UPT_MIN_ADDRESS)
1096 ptepindex = (va >> PDRSHIFT);
1097 if (pmap->pm_ptphint &&
1098 (pmap->pm_ptphint->pindex == ptepindex)) {
1099 mpte = pmap->pm_ptphint;
1101 pmap_inval_flush(info);
1102 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1103 pmap->pm_ptphint = mpte;
1107 return pmap_unwire_pte_hold(pmap, mpte, info);
1111 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1112 * it, and IdlePTD, represents the template used to update all other pmaps.
1114 * On architectures where the kernel pmap is not integrated into the user
1115 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1116 * kernel_pmap should be used to directly access the kernel_pmap.
1121 pmap_pinit0(struct pmap *pmap)
1124 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1125 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1127 pmap->pm_active = 0;
1128 pmap->pm_cached = 0;
1129 pmap->pm_ptphint = NULL;
1130 TAILQ_INIT(&pmap->pm_pvlist);
1131 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1135 * Initialize a preallocated and zeroed pmap structure,
1136 * such as one in a vmspace structure.
1141 pmap_pinit(struct pmap *pmap)
1146 * No need to allocate page table space yet but we do need a valid
1147 * page directory table.
1149 if (pmap->pm_pdir == NULL) {
1151 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1155 * Allocate an object for the ptes
1157 if (pmap->pm_pteobj == NULL)
1158 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1161 * Allocate the page directory page, unless we already have
1162 * one cached. If we used the cached page the wire_count will
1163 * already be set appropriately.
1165 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1166 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1167 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1168 pmap->pm_pdirm = ptdpg;
1169 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1170 ptdpg->valid = VM_PAGE_BITS_ALL;
1171 ptdpg->wire_count = 1;
1172 ++vmstats.v_wire_count;
1173 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1175 if ((ptdpg->flags & PG_ZERO) == 0)
1176 bzero(pmap->pm_pdir, PAGE_SIZE);
1179 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1182 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1184 /* install self-referential address mapping entry */
1185 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1186 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1189 pmap->pm_active = 0;
1190 pmap->pm_cached = 0;
1191 pmap->pm_ptphint = NULL;
1192 TAILQ_INIT(&pmap->pm_pvlist);
1193 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1194 pmap->pm_stats.resident_count = 1;
1198 * Clean up a pmap structure so it can be physically freed. This routine
1199 * is called by the vmspace dtor function. A great deal of pmap data is
1200 * left passively mapped to improve vmspace management so we have a bit
1201 * of cleanup work to do here.
1206 pmap_puninit(pmap_t pmap)
1210 KKASSERT(pmap->pm_active == 0);
1211 if ((p = pmap->pm_pdirm) != NULL) {
1212 KKASSERT(pmap->pm_pdir != NULL);
1213 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1215 vmstats.v_wire_count--;
1216 KKASSERT((p->flags & PG_BUSY) == 0);
1218 vm_page_free_zero(p);
1219 pmap->pm_pdirm = NULL;
1221 if (pmap->pm_pdir) {
1222 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1223 pmap->pm_pdir = NULL;
1225 if (pmap->pm_pteobj) {
1226 vm_object_deallocate(pmap->pm_pteobj);
1227 pmap->pm_pteobj = NULL;
1232 * Wire in kernel global address entries. To avoid a race condition
1233 * between pmap initialization and pmap_growkernel, this procedure
1234 * adds the pmap to the master list (which growkernel scans to update),
1235 * then copies the template.
1240 pmap_pinit2(struct pmap *pmap)
1242 lwkt_gettoken(&vm_token);
1243 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1244 /* XXX copies current process, does not fill in MPPTDI */
1245 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1246 lwkt_reltoken(&vm_token);
1250 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1251 * 0 on failure (if the procedure had to sleep).
1253 * When asked to remove the page directory page itself, we actually just
1254 * leave it cached so we do not have to incur the SMP inval overhead of
1255 * removing the kernel mapping. pmap_puninit() will take care of it.
1257 * The caller must hold vm_token.
1258 * This function can block regardless.
1261 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1263 unsigned *pde = (unsigned *) pmap->pm_pdir;
1266 * This code optimizes the case of freeing non-busy
1267 * page-table pages. Those pages are zero now, and
1268 * might as well be placed directly into the zero queue.
1270 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1276 * Remove the page table page from the processes address space.
1278 KKASSERT(pmap->pm_stats.resident_count > 0);
1279 KKASSERT(pde[p->pindex]);
1281 --pmap->pm_stats.resident_count;
1282 pmap->pm_cached = 0;
1284 if (p->hold_count) {
1285 panic("pmap_release: freeing held page table page");
1287 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1288 pmap->pm_ptphint = NULL;
1291 * We leave the page directory page cached, wired, and mapped in
1292 * the pmap until the dtor function (pmap_puninit()) gets called.
1293 * However, still clean it up so we can set PG_ZERO.
1295 * The pmap has already been removed from the pmap_list in the
1298 if (p->pindex == PTDPTDI) {
1299 bzero(pde + KPTDI, nkpt * PTESIZE);
1300 bzero(pde + MPPTDI, (NPDEPG - MPPTDI) * PTESIZE);
1301 vm_page_flag_set(p, PG_ZERO);
1305 vmstats.v_wire_count--;
1306 vm_page_free_zero(p);
1312 * This routine is called if the page table page is not mapped correctly.
1314 * The caller must hold vm_token.
1317 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1319 vm_offset_t pteva, ptepa;
1323 * Find or fabricate a new pagetable page
1325 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1326 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1328 KASSERT(m->queue == PQ_NONE,
1329 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1332 * Increment the hold count for the page we will be returning to
1338 * It is possible that someone else got in and mapped by the page
1339 * directory page while we were blocked, if so just unbusy and
1340 * return the held page.
1342 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1343 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1348 if (m->wire_count == 0)
1349 vmstats.v_wire_count++;
1354 * Map the pagetable page into the process address space, if
1355 * it isn't already there.
1357 * NOTE: For safety clear pm_cached for all cpus including the
1358 * current one when adding a PDE to the map.
1360 ++pmap->pm_stats.resident_count;
1362 ptepa = VM_PAGE_TO_PHYS(m);
1363 pmap->pm_pdir[ptepindex] =
1364 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1365 pmap->pm_cached = 0;
1368 * Set the page table hint
1370 pmap->pm_ptphint = m;
1373 * Try to use the new mapping, but if we cannot, then
1374 * do it with the routine that maps the page explicitly.
1376 if ((m->flags & PG_ZERO) == 0) {
1377 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1378 (((unsigned) PTDpde) & PG_FRAME)) {
1379 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1380 bzero((caddr_t) pteva, PAGE_SIZE);
1382 pmap_zero_page(ptepa);
1387 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1391 m->valid = VM_PAGE_BITS_ALL;
1392 vm_page_flag_clear(m, PG_ZERO);
1393 vm_page_flag_set(m, PG_MAPPED);
1400 * Allocate a page table entry for a va.
1402 * The caller must hold vm_token.
1405 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1412 * Calculate pagetable page index
1414 ptepindex = va >> PDRSHIFT;
1417 * Get the page directory entry
1419 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1422 * This supports switching from a 4MB page to a
1425 if (ptepa & PG_PS) {
1426 pmap->pm_pdir[ptepindex] = 0;
1433 * If the page table page is mapped, we just increment the
1434 * hold count, and activate it.
1438 * In order to get the page table page, try the
1441 if (pmap->pm_ptphint &&
1442 (pmap->pm_ptphint->pindex == ptepindex)) {
1443 m = pmap->pm_ptphint;
1445 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1446 pmap->pm_ptphint = m;
1452 * Here if the pte page isn't mapped, or if it has been deallocated.
1454 return _pmap_allocpte(pmap, ptepindex);
1458 /***************************************************
1459 * Pmap allocation/deallocation routines.
1460 ***************************************************/
1463 * Release any resources held by the given physical map.
1464 * Called when a pmap initialized by pmap_pinit is being released.
1465 * Should only be called if the map contains no valid mappings.
1469 static int pmap_release_callback(struct vm_page *p, void *data);
1472 pmap_release(struct pmap *pmap)
1474 vm_object_t object = pmap->pm_pteobj;
1475 struct rb_vm_page_scan_info info;
1477 KASSERT(pmap->pm_active == 0,
1478 ("pmap still active! %08x", pmap->pm_active));
1479 #if defined(DIAGNOSTIC)
1480 if (object->ref_count != 1)
1481 panic("pmap_release: pteobj reference count != 1");
1485 info.object = object;
1486 lwkt_gettoken(&vm_token);
1487 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1492 info.limit = object->generation;
1494 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1495 pmap_release_callback, &info);
1496 if (info.error == 0 && info.mpte) {
1497 if (!pmap_release_free_page(pmap, info.mpte))
1500 } while (info.error);
1501 pmap->pm_cached = 0;
1502 lwkt_reltoken(&vm_token);
1506 * The caller must hold vm_token.
1509 pmap_release_callback(struct vm_page *p, void *data)
1511 struct rb_vm_page_scan_info *info = data;
1513 if (p->pindex == PTDPTDI) {
1517 if (!pmap_release_free_page(info->pmap, p)) {
1521 if (info->object->generation != info->limit) {
1529 * Grow the number of kernel page table entries, if needed.
1534 pmap_growkernel(vm_offset_t addr)
1537 vm_offset_t ptppaddr;
1541 lwkt_gettoken(&vm_token);
1542 if (kernel_vm_end == 0) {
1543 kernel_vm_end = KERNBASE;
1545 while (pdir_pde(PTD, kernel_vm_end)) {
1546 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1547 ~(PAGE_SIZE * NPTEPG - 1);
1551 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1552 while (kernel_vm_end < addr) {
1553 if (pdir_pde(PTD, kernel_vm_end)) {
1554 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1555 ~(PAGE_SIZE * NPTEPG - 1);
1560 * This index is bogus, but out of the way
1562 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_NORMAL |
1564 VM_ALLOC_INTERRUPT);
1566 panic("pmap_growkernel: no memory to grow kernel");
1569 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1570 pmap_zero_page(ptppaddr);
1571 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1572 pdir_pde(PTD, kernel_vm_end) = newpdir;
1573 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1577 * This update must be interlocked with pmap_pinit2.
1579 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1580 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1582 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1583 ~(PAGE_SIZE * NPTEPG - 1);
1585 lwkt_reltoken(&vm_token);
1589 * Retire the given physical map from service.
1591 * Should only be called if the map contains no valid mappings.
1596 pmap_destroy(pmap_t pmap)
1601 lwkt_gettoken(&vm_token);
1602 if (--pmap->pm_count == 0) {
1604 panic("destroying a pmap is not yet implemented");
1606 lwkt_reltoken(&vm_token);
1610 * Add a reference to the specified pmap.
1615 pmap_reference(pmap_t pmap)
1618 lwkt_gettoken(&vm_token);
1620 lwkt_reltoken(&vm_token);
1624 /***************************************************
1625 * page management routines.
1626 ***************************************************/
1629 * free the pv_entry back to the free list. This function may be
1630 * called from an interrupt.
1632 * The caller must hold vm_token.
1634 static PMAP_INLINE void
1635 free_pv_entry(pv_entry_t pv)
1638 KKASSERT(pv->pv_m != NULL);
1646 * get a new pv_entry, allocating a block from the system
1647 * when needed. This function may be called from an interrupt.
1649 * The caller must hold vm_token.
1655 if (pv_entry_high_water &&
1656 (pv_entry_count > pv_entry_high_water) &&
1657 (pmap_pagedaemon_waken == 0)) {
1658 pmap_pagedaemon_waken = 1;
1659 wakeup (&vm_pages_needed);
1661 return zalloc(pvzone);
1665 * This routine is very drastic, but can save the system
1675 static int warningdone=0;
1677 if (pmap_pagedaemon_waken == 0)
1679 lwkt_gettoken(&vm_token);
1680 pmap_pagedaemon_waken = 0;
1682 if (warningdone < 5) {
1683 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1687 for(i = 0; i < vm_page_array_size; i++) {
1688 m = &vm_page_array[i];
1689 if (m->wire_count || m->hold_count || m->busy ||
1690 (m->flags & PG_BUSY)) {
1695 lwkt_reltoken(&vm_token);
1700 * If it is the first entry on the list, it is actually
1701 * in the header and we must copy the following entry up
1702 * to the header. Otherwise we must search the list for
1703 * the entry. In either case we free the now unused entry.
1705 * The caller must hold vm_token.
1708 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1709 vm_offset_t va, pmap_inval_info_t info)
1714 ASSERT_LWKT_TOKEN_HELD(&vm_token);
1715 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1716 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1717 if (pmap == pv->pv_pmap && va == pv->pv_va)
1721 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1723 KKASSERT(pv->pv_pmap == pmap);
1725 if (va == pv->pv_va)
1732 test_m_maps_pv(m, pv);
1733 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1734 m->md.pv_list_count--;
1735 if (TAILQ_EMPTY(&m->md.pv_list))
1736 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1737 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1738 ++pmap->pm_generation;
1739 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1745 * Create a pv entry for page at pa for (pmap, va).
1747 * The caller must hold vm_token.
1750 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1754 pv = get_pv_entry();
1756 KKASSERT(pv->pv_m == NULL);
1763 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1764 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1765 ++pmap->pm_generation;
1766 m->md.pv_list_count++;
1770 * pmap_remove_pte: do the things to unmap a page in a process.
1772 * The caller must hold vm_token.
1774 * WARNING! As with most other pmap functions this one can block, so
1775 * callers using temporary page table mappings must reload
1779 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1780 pmap_inval_info_t info)
1785 ptbase_assert(pmap);
1786 pmap_inval_interlock(info, pmap, va);
1787 ptbase_assert(pmap);
1788 oldpte = loadandclear(ptq);
1790 pmap->pm_stats.wired_count -= 1;
1791 pmap_inval_deinterlock(info, pmap);
1794 * Machines that don't support invlpg, also don't support
1795 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1799 cpu_invlpg((void *)va);
1800 KKASSERT(pmap->pm_stats.resident_count > 0);
1801 --pmap->pm_stats.resident_count;
1802 if (oldpte & PG_MANAGED) {
1803 m = PHYS_TO_VM_PAGE(oldpte);
1804 if (oldpte & PG_M) {
1805 #if defined(PMAP_DIAGNOSTIC)
1806 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1807 kprintf("pmap_remove: modified page not "
1808 "writable: va: %p, pte: 0x%lx\n",
1809 (void *)va, (long)oldpte);
1812 if (pmap_track_modified(va))
1816 vm_page_flag_set(m, PG_REFERENCED);
1817 return pmap_remove_entry(pmap, m, va, info);
1819 return pmap_unuse_pt(pmap, va, NULL, info);
1826 * Remove a single page from a process address space.
1828 * The caller must hold vm_token.
1831 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1836 * if there is no pte for this address, just skip it!!! Otherwise
1837 * get a local va for mappings for this pmap and remove the entry.
1839 if (*pmap_pde(pmap, va) != 0) {
1840 ptq = get_ptbase(pmap) + i386_btop(va);
1842 pmap_remove_pte(pmap, ptq, va, info);
1849 * Remove the given range of addresses from the specified map.
1851 * It is assumed that the start and end are properly rounded to the page
1857 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1861 vm_offset_t ptpaddr;
1862 vm_offset_t sindex, eindex;
1863 struct pmap_inval_info info;
1868 lwkt_gettoken(&vm_token);
1869 if (pmap->pm_stats.resident_count == 0) {
1870 lwkt_reltoken(&vm_token);
1874 pmap_inval_init(&info);
1877 * special handling of removing one page. a very
1878 * common operation and easy to short circuit some
1881 if (((sva + PAGE_SIZE) == eva) &&
1882 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1883 pmap_remove_page(pmap, sva, &info);
1884 pmap_inval_done(&info);
1885 lwkt_reltoken(&vm_token);
1890 * Get a local virtual address for the mappings that are being
1893 sindex = i386_btop(sva);
1894 eindex = i386_btop(eva);
1896 for (; sindex < eindex; sindex = pdnxt) {
1900 * Calculate index for next page table.
1902 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1903 if (pmap->pm_stats.resident_count == 0)
1906 pdirindex = sindex / NPDEPG;
1907 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1908 pmap_inval_interlock(&info, pmap, -1);
1909 pmap->pm_pdir[pdirindex] = 0;
1910 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1911 pmap->pm_cached = 0;
1912 pmap_inval_deinterlock(&info, pmap);
1917 * Weed out invalid mappings. Note: we assume that the page
1918 * directory table is always allocated, and in kernel virtual.
1924 * Limit our scan to either the end of the va represented
1925 * by the current page table page, or to the end of the
1926 * range being removed.
1928 if (pdnxt > eindex) {
1933 * NOTE: pmap_remove_pte() can block and wipe the temporary
1936 for (; sindex != pdnxt; sindex++) {
1939 ptbase = get_ptbase(pmap);
1940 if (ptbase[sindex] == 0)
1942 va = i386_ptob(sindex);
1943 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1947 pmap_inval_done(&info);
1948 lwkt_reltoken(&vm_token);
1952 * Removes this physical page from all physical maps in which it resides.
1953 * Reflects back modify bits to the pager.
1958 pmap_remove_all(vm_page_t m)
1960 struct pmap_inval_info info;
1961 unsigned *pte, tpte;
1964 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1967 lwkt_gettoken(&vm_token);
1968 pmap_inval_init(&info);
1969 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1970 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1971 --pv->pv_pmap->pm_stats.resident_count;
1973 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1974 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
1975 tpte = loadandclear(pte);
1977 pv->pv_pmap->pm_stats.wired_count--;
1978 pmap_inval_deinterlock(&info, pv->pv_pmap);
1980 vm_page_flag_set(m, PG_REFERENCED);
1982 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
1986 * Update the vm_page_t clean and reference bits.
1989 #if defined(PMAP_DIAGNOSTIC)
1990 if (pmap_nw_modified((pt_entry_t) tpte)) {
1991 kprintf("pmap_remove_all: modified page "
1992 "not writable: va: %p, pte: 0x%lx\n",
1993 (void *)pv->pv_va, (long)tpte);
1996 if (pmap_track_modified(pv->pv_va))
2000 KKASSERT(pv->pv_m == m);
2002 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2003 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2004 ++pv->pv_pmap->pm_generation;
2005 m->md.pv_list_count--;
2006 if (TAILQ_EMPTY(&m->md.pv_list))
2007 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2008 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2011 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2012 pmap_inval_done(&info);
2013 lwkt_reltoken(&vm_token);
2017 * Set the physical protection on the specified range of this map
2023 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2026 vm_offset_t pdnxt, ptpaddr;
2027 vm_pindex_t sindex, eindex;
2028 pmap_inval_info info;
2033 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2034 lwkt_gettoken(&vm_token);
2035 pmap_remove(pmap, sva, eva);
2036 lwkt_reltoken(&vm_token);
2040 if (prot & VM_PROT_WRITE)
2043 lwkt_gettoken(&vm_token);
2044 pmap_inval_init(&info);
2046 ptbase = get_ptbase(pmap);
2048 sindex = i386_btop(sva);
2049 eindex = i386_btop(eva);
2051 for (; sindex < eindex; sindex = pdnxt) {
2054 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2056 pdirindex = sindex / NPDEPG;
2057 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
2058 pmap_inval_interlock(&info, pmap, -1);
2059 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2060 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2061 pmap_inval_deinterlock(&info, pmap);
2066 * Weed out invalid mappings. Note: we assume that the page
2067 * directory table is always allocated, and in kernel virtual.
2072 if (pdnxt > eindex) {
2076 for (; sindex != pdnxt; sindex++) {
2082 * XXX non-optimal. Note also that there can be
2083 * no pmap_inval_flush() calls until after we modify
2084 * ptbase[sindex] (or otherwise we have to do another
2085 * pmap_inval_interlock() call).
2087 pmap_inval_interlock(&info, pmap, i386_ptob(sindex));
2089 pbits = ptbase[sindex];
2092 if (pbits & PG_MANAGED) {
2095 m = PHYS_TO_VM_PAGE(pbits);
2096 vm_page_flag_set(m, PG_REFERENCED);
2100 if (pmap_track_modified(i386_ptob(sindex))) {
2102 m = PHYS_TO_VM_PAGE(pbits);
2109 if (pbits != cbits &&
2110 !atomic_cmpset_int(ptbase + sindex, pbits, cbits)) {
2113 pmap_inval_deinterlock(&info, pmap);
2116 pmap_inval_done(&info);
2117 lwkt_reltoken(&vm_token);
2121 * Insert the given physical page (p) at the specified virtual address (v)
2122 * in the target physical map with the protection requested.
2124 * If specified, the page will be wired down, meaning that the related pte
2125 * cannot be reclaimed.
2130 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2136 vm_offset_t origpte, newpte;
2138 pmap_inval_info info;
2144 #ifdef PMAP_DIAGNOSTIC
2146 panic("pmap_enter: toobig");
2147 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2148 panic("pmap_enter: invalid to pmap_enter page "
2149 "table pages (va: %p)", (void *)va);
2152 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2153 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2154 print_backtrace(-1);
2156 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2157 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2158 print_backtrace(-1);
2161 lwkt_gettoken(&vm_token);
2164 * In the case that a page table page is not
2165 * resident, we are creating it here.
2167 if (va < UPT_MIN_ADDRESS)
2168 mpte = pmap_allocpte(pmap, va);
2172 pmap_inval_init(&info);
2173 pte = pmap_pte(pmap, va);
2176 * Page Directory table entry not valid, we need a new PT page
2179 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p\n",
2180 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
2183 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2184 origpte = *(vm_offset_t *)pte;
2185 opa = origpte & PG_FRAME;
2187 if (origpte & PG_PS)
2188 panic("pmap_enter: attempted pmap_enter on 4MB page");
2191 * Mapping has not changed, must be protection or wiring change.
2193 if (origpte && (opa == pa)) {
2195 * Wiring change, just update stats. We don't worry about
2196 * wiring PT pages as they remain resident as long as there
2197 * are valid mappings in them. Hence, if a user page is wired,
2198 * the PT page will be also.
2200 if (wired && ((origpte & PG_W) == 0))
2201 pmap->pm_stats.wired_count++;
2202 else if (!wired && (origpte & PG_W))
2203 pmap->pm_stats.wired_count--;
2205 #if defined(PMAP_DIAGNOSTIC)
2206 if (pmap_nw_modified((pt_entry_t) origpte)) {
2207 kprintf("pmap_enter: modified page not "
2208 "writable: va: %p, pte: 0x%lx\n",
2209 (void *)va, (long )origpte);
2214 * Remove the extra pte reference. Note that we cannot
2215 * optimize the RO->RW case because we have adjusted the
2216 * wiring count above and may need to adjust the wiring
2223 * We might be turning off write access to the page,
2224 * so we go ahead and sense modify status.
2226 if (origpte & PG_MANAGED) {
2227 if ((origpte & PG_M) && pmap_track_modified(va)) {
2229 om = PHYS_TO_VM_PAGE(opa);
2233 KKASSERT(m->flags & PG_MAPPED);
2238 * Mapping has changed, invalidate old range and fall through to
2239 * handle validating new mapping.
2241 * Since we have a ref on the page directory page pmap_pte()
2242 * will always return non-NULL.
2244 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2245 * to get wiped. reload the ptbase. I'm not sure if it is
2246 * also possible to race another pmap_enter() but check for
2252 KKASSERT((origpte & PG_FRAME) ==
2253 (*(vm_offset_t *)pte & PG_FRAME));
2254 err = pmap_remove_pte(pmap, pte, va, &info);
2256 panic("pmap_enter: pte vanished, va: %p", (void *)va);
2257 pte = pmap_pte(pmap, va);
2258 origpte = *(vm_offset_t *)pte;
2259 opa = origpte & PG_FRAME;
2261 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2267 * Enter on the PV list if part of our managed memory. Note that we
2268 * raise IPL while manipulating pv_table since pmap_enter can be
2269 * called at interrupt time.
2271 if (pmap_initialized &&
2272 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2273 pmap_insert_entry(pmap, va, mpte, m);
2274 ptbase_assert(pmap);
2276 vm_page_flag_set(m, PG_MAPPED);
2280 * Increment counters
2282 ++pmap->pm_stats.resident_count;
2284 pmap->pm_stats.wired_count++;
2285 KKASSERT(*pte == 0);
2289 * Now validate mapping with desired protection/wiring.
2291 ptbase_assert(pmap);
2292 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2296 if (va < UPT_MIN_ADDRESS)
2298 if (pmap == &kernel_pmap)
2302 * if the mapping or permission bits are different, we need
2303 * to update the pte.
2305 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2306 pmap_inval_interlock(&info, pmap, va);
2307 ptbase_assert(pmap);
2308 KKASSERT(*pte == 0 ||
2309 (*pte & PG_FRAME) == (newpte & PG_FRAME));
2310 *pte = newpte | PG_A;
2311 pmap_inval_deinterlock(&info, pmap);
2313 vm_page_flag_set(m, PG_WRITEABLE);
2315 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2316 pmap_inval_done(&info);
2317 lwkt_reltoken(&vm_token);
2321 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2322 * This code also assumes that the pmap has no pre-existing entry for this
2325 * This code currently may only be used on user pmaps, not kernel_pmap.
2330 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2337 pmap_inval_info info;
2339 lwkt_gettoken(&vm_token);
2340 pmap_inval_init(&info);
2342 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2343 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2344 print_backtrace(-1);
2346 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2347 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2348 print_backtrace(-1);
2351 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2354 * Calculate the page table page (mpte), allocating it if necessary.
2356 * A held page table page (mpte), or NULL, is passed onto the
2357 * section following.
2359 if (va < UPT_MIN_ADDRESS) {
2361 * Calculate pagetable page index
2363 ptepindex = va >> PDRSHIFT;
2367 * Get the page directory entry
2369 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2372 * If the page table page is mapped, we just increment
2373 * the hold count, and activate it.
2377 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2378 if (pmap->pm_ptphint &&
2379 (pmap->pm_ptphint->pindex == ptepindex)) {
2380 mpte = pmap->pm_ptphint;
2382 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2383 pmap->pm_ptphint = mpte;
2388 mpte = _pmap_allocpte(pmap, ptepindex);
2390 } while (mpte == NULL);
2393 /* this code path is not yet used */
2397 * With a valid (and held) page directory page, we can just use
2398 * vtopte() to get to the pte. If the pte is already present
2399 * we do not disturb it.
2401 pte = (unsigned *)vtopte(va);
2404 pmap_unwire_pte_hold(pmap, mpte, &info);
2405 pa = VM_PAGE_TO_PHYS(m);
2406 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2407 pmap_inval_done(&info);
2408 lwkt_reltoken(&vm_token);
2413 * Enter on the PV list if part of our managed memory
2415 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2416 pmap_insert_entry(pmap, va, mpte, m);
2417 vm_page_flag_set(m, PG_MAPPED);
2421 * Increment counters
2423 ++pmap->pm_stats.resident_count;
2425 pa = VM_PAGE_TO_PHYS(m);
2428 * Now validate mapping with RO protection
2430 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2431 *pte = pa | PG_V | PG_U;
2433 *pte = pa | PG_V | PG_U | PG_MANAGED;
2434 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2435 pmap_inval_done(&info);
2436 lwkt_reltoken(&vm_token);
2440 * Make a temporary mapping for a physical address. This is only intended
2441 * to be used for panic dumps.
2446 pmap_kenter_temporary(vm_paddr_t pa, int i)
2448 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2449 return ((void *)crashdumpmap);
2452 #define MAX_INIT_PT (96)
2455 * This routine preloads the ptes for a given object into the specified pmap.
2456 * This eliminates the blast of soft faults on process startup and
2457 * immediately after an mmap.
2461 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2464 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2465 vm_object_t object, vm_pindex_t pindex,
2466 vm_size_t size, int limit)
2468 struct rb_vm_page_scan_info info;
2473 * We can't preinit if read access isn't set or there is no pmap
2476 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2480 * We can't preinit if the pmap is not the current pmap
2482 lp = curthread->td_lwp;
2483 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2486 psize = i386_btop(size);
2488 if ((object->type != OBJT_VNODE) ||
2489 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2490 (object->resident_page_count > MAX_INIT_PT))) {
2494 if (psize + pindex > object->size) {
2495 if (object->size < pindex)
2497 psize = object->size - pindex;
2504 * Use a red-black scan to traverse the requested range and load
2505 * any valid pages found into the pmap.
2507 * We cannot safely scan the object's memq unless we are in a
2508 * critical section since interrupts can remove pages from objects.
2510 info.start_pindex = pindex;
2511 info.end_pindex = pindex + psize - 1;
2517 lwkt_gettoken(&vm_token);
2518 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2519 pmap_object_init_pt_callback, &info);
2520 lwkt_reltoken(&vm_token);
2524 * The caller must hold vm_token.
2528 pmap_object_init_pt_callback(vm_page_t p, void *data)
2530 struct rb_vm_page_scan_info *info = data;
2531 vm_pindex_t rel_index;
2533 * don't allow an madvise to blow away our really
2534 * free pages allocating pv entries.
2536 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2537 vmstats.v_free_count < vmstats.v_free_reserved) {
2540 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2541 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2542 if ((p->queue - p->pc) == PQ_CACHE)
2543 vm_page_deactivate(p);
2545 rel_index = p->pindex - info->start_pindex;
2546 pmap_enter_quick(info->pmap,
2547 info->addr + i386_ptob(rel_index), p);
2554 * Return TRUE if the pmap is in shape to trivially
2555 * pre-fault the specified address.
2557 * Returns FALSE if it would be non-trivial or if a
2558 * pte is already loaded into the slot.
2563 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2568 lwkt_gettoken(&vm_token);
2569 if ((*pmap_pde(pmap, addr)) == 0) {
2572 pte = (unsigned *) vtopte(addr);
2573 ret = (*pte) ? 0 : 1;
2575 lwkt_reltoken(&vm_token);
2580 * Change the wiring attribute for a map/virtual-adderss pair. The mapping
2581 * must already exist.
2586 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2593 lwkt_gettoken(&vm_token);
2594 pte = pmap_pte(pmap, va);
2596 if (wired && !pmap_pte_w(pte))
2597 pmap->pm_stats.wired_count++;
2598 else if (!wired && pmap_pte_w(pte))
2599 pmap->pm_stats.wired_count--;
2602 * Wiring is not a hardware characteristic so there is no need to
2603 * invalidate TLB. However, in an SMP environment we must use
2604 * a locked bus cycle to update the pte (if we are not using
2605 * the pmap_inval_*() API that is)... it's ok to do this for simple
2610 atomic_set_int(pte, PG_W);
2612 atomic_clear_int(pte, PG_W);
2615 atomic_set_int_nonlocked(pte, PG_W);
2617 atomic_clear_int_nonlocked(pte, PG_W);
2619 lwkt_reltoken(&vm_token);
2623 * Copy the range specified by src_addr/len from the source map to the
2624 * range dst_addr/len in the destination map.
2626 * This routine is only advisory and need not do anything.
2631 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2632 vm_size_t len, vm_offset_t src_addr)
2638 * Zero the specified PA by mapping the page into KVM and clearing its
2644 pmap_zero_page(vm_paddr_t phys)
2646 struct mdglobaldata *gd = mdcpu;
2649 if (*(int *)gd->gd_CMAP3)
2650 panic("pmap_zero_page: CMAP3 busy");
2651 *(int *)gd->gd_CMAP3 =
2652 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2653 cpu_invlpg(gd->gd_CADDR3);
2655 #if defined(I686_CPU)
2656 if (cpu_class == CPUCLASS_686)
2657 i686_pagezero(gd->gd_CADDR3);
2660 bzero(gd->gd_CADDR3, PAGE_SIZE);
2661 *(int *) gd->gd_CMAP3 = 0;
2666 * Assert that a page is empty, panic if it isn't.
2671 pmap_page_assertzero(vm_paddr_t phys)
2673 struct mdglobaldata *gd = mdcpu;
2677 if (*(int *)gd->gd_CMAP3)
2678 panic("pmap_zero_page: CMAP3 busy");
2679 *(int *)gd->gd_CMAP3 =
2680 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2681 cpu_invlpg(gd->gd_CADDR3);
2682 for (i = 0; i < PAGE_SIZE; i += 4) {
2683 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2684 panic("pmap_page_assertzero() @ %p not zero!\n",
2685 (void *)gd->gd_CADDR3);
2688 *(int *) gd->gd_CMAP3 = 0;
2693 * Zero part of a physical page by mapping it into memory and clearing
2694 * its contents with bzero.
2696 * off and size may not cover an area beyond a single hardware page.
2701 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2703 struct mdglobaldata *gd = mdcpu;
2706 if (*(int *) gd->gd_CMAP3)
2707 panic("pmap_zero_page: CMAP3 busy");
2708 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2709 cpu_invlpg(gd->gd_CADDR3);
2711 #if defined(I686_CPU)
2712 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2713 i686_pagezero(gd->gd_CADDR3);
2716 bzero((char *)gd->gd_CADDR3 + off, size);
2717 *(int *) gd->gd_CMAP3 = 0;
2722 * Copy the physical page from the source PA to the target PA.
2723 * This function may be called from an interrupt. No locking
2729 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2731 struct mdglobaldata *gd = mdcpu;
2734 if (*(int *) gd->gd_CMAP1)
2735 panic("pmap_copy_page: CMAP1 busy");
2736 if (*(int *) gd->gd_CMAP2)
2737 panic("pmap_copy_page: CMAP2 busy");
2739 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2740 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2742 cpu_invlpg(gd->gd_CADDR1);
2743 cpu_invlpg(gd->gd_CADDR2);
2745 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2747 *(int *) gd->gd_CMAP1 = 0;
2748 *(int *) gd->gd_CMAP2 = 0;
2753 * Copy the physical page from the source PA to the target PA.
2754 * This function may be called from an interrupt. No locking
2760 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2762 struct mdglobaldata *gd = mdcpu;
2765 if (*(int *) gd->gd_CMAP1)
2766 panic("pmap_copy_page: CMAP1 busy");
2767 if (*(int *) gd->gd_CMAP2)
2768 panic("pmap_copy_page: CMAP2 busy");
2770 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2771 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2773 cpu_invlpg(gd->gd_CADDR1);
2774 cpu_invlpg(gd->gd_CADDR2);
2776 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2777 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2780 *(int *) gd->gd_CMAP1 = 0;
2781 *(int *) gd->gd_CMAP2 = 0;
2786 * Returns true if the pmap's pv is one of the first
2787 * 16 pvs linked to from this page. This count may
2788 * be changed upwards or downwards in the future; it
2789 * is only necessary that true be returned for a small
2790 * subset of pmaps for proper page aging.
2795 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2800 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2803 lwkt_gettoken(&vm_token);
2804 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2805 if (pv->pv_pmap == pmap) {
2813 lwkt_reltoken(&vm_token);
2818 * Remove all pages from specified address space
2819 * this aids process exit speeds. Also, this code
2820 * is special cased for current process only, but
2821 * can have the more generic (and slightly slower)
2822 * mode enabled. This is much faster than pmap_remove
2823 * in the case of running down an entire address space.
2828 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2831 unsigned *pte, tpte;
2834 pmap_inval_info info;
2836 int32_t save_generation;
2838 lp = curthread->td_lwp;
2839 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2844 lwkt_gettoken(&vm_token);
2845 pmap_inval_init(&info);
2846 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2847 if (pv->pv_va >= eva || pv->pv_va < sva) {
2848 npv = TAILQ_NEXT(pv, pv_plist);
2852 KKASSERT(pmap == pv->pv_pmap);
2855 pte = (unsigned *)vtopte(pv->pv_va);
2857 pte = pmap_pte_quick(pmap, pv->pv_va);
2859 pmap_inval_interlock(&info, pmap, pv->pv_va);
2862 * We cannot remove wired pages from a process' mapping
2866 pmap_inval_deinterlock(&info, pmap);
2867 npv = TAILQ_NEXT(pv, pv_plist);
2871 tpte = loadandclear(pte);
2872 pmap_inval_deinterlock(&info, pmap);
2874 m = PHYS_TO_VM_PAGE(tpte);
2875 test_m_maps_pv(m, pv);
2877 KASSERT(m < &vm_page_array[vm_page_array_size],
2878 ("pmap_remove_pages: bad tpte %x", tpte));
2880 KKASSERT(pmap->pm_stats.resident_count > 0);
2881 --pmap->pm_stats.resident_count;
2884 * Update the vm_page_t clean and reference bits.
2890 npv = TAILQ_NEXT(pv, pv_plist);
2892 KKASSERT(pv->pv_m == m);
2893 KKASSERT(pv->pv_pmap == pmap);
2895 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2896 save_generation = ++pmap->pm_generation;
2898 m->md.pv_list_count--;
2899 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2900 if (TAILQ_EMPTY(&m->md.pv_list))
2901 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2903 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2907 * Restart the scan if we blocked during the unuse or free
2908 * calls and other removals were made.
2910 if (save_generation != pmap->pm_generation) {
2911 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2912 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2915 pmap_inval_done(&info);
2916 lwkt_reltoken(&vm_token);
2920 * pmap_testbit tests bits in pte's
2921 * note that the testbit/clearbit routines are inline,
2922 * and a lot of things compile-time evaluate.
2924 * The caller must hold vm_token.
2927 pmap_testbit(vm_page_t m, int bit)
2932 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2935 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2938 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2940 * if the bit being tested is the modified bit, then
2941 * mark clean_map and ptes as never
2944 if (bit & (PG_A|PG_M)) {
2945 if (!pmap_track_modified(pv->pv_va))
2949 #if defined(PMAP_DIAGNOSTIC)
2951 kprintf("Null pmap (tb) at va: %p\n",
2956 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2964 * This routine is used to modify bits in ptes
2966 * The caller must hold vm_token.
2968 static __inline void
2969 pmap_clearbit(vm_page_t m, int bit)
2971 struct pmap_inval_info info;
2976 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2979 pmap_inval_init(&info);
2982 * Loop over all current mappings setting/clearing as appropos If
2983 * setting RO do we need to clear the VAC?
2985 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2987 * don't write protect pager mappings
2990 if (!pmap_track_modified(pv->pv_va))
2994 #if defined(PMAP_DIAGNOSTIC)
2996 kprintf("Null pmap (cb) at va: %p\n",
3003 * Careful here. We can use a locked bus instruction to
3004 * clear PG_A or PG_M safely but we need to synchronize
3005 * with the target cpus when we mess with PG_RW.
3007 * We do not have to force synchronization when clearing
3008 * PG_M even for PTEs generated via virtual memory maps,
3009 * because the virtual kernel will invalidate the pmap
3010 * entry when/if it needs to resynchronize the Modify bit.
3013 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3014 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3021 atomic_clear_int(pte, PG_M|PG_RW);
3024 * The cpu may be trying to set PG_M
3025 * simultaniously with our clearing
3028 if (!atomic_cmpset_int(pte, pbits,
3032 } else if (bit == PG_M) {
3034 * We could also clear PG_RW here to force
3035 * a fault on write to redetect PG_M for
3036 * virtual kernels, but it isn't necessary
3037 * since virtual kernels invalidate the pte
3038 * when they clear the VPTE_M bit in their
3039 * virtual page tables.
3041 atomic_clear_int(pte, PG_M);
3043 atomic_clear_int(pte, bit);
3047 pmap_inval_deinterlock(&info, pv->pv_pmap);
3049 pmap_inval_done(&info);
3053 * Lower the permission for all mappings to a given page.
3058 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3060 if ((prot & VM_PROT_WRITE) == 0) {
3061 lwkt_gettoken(&vm_token);
3062 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3063 pmap_clearbit(m, PG_RW);
3064 vm_page_flag_clear(m, PG_WRITEABLE);
3068 lwkt_reltoken(&vm_token);
3073 * Return the physical address given a physical page index.
3078 pmap_phys_address(vm_pindex_t ppn)
3080 return (i386_ptob(ppn));
3084 * Return a count of reference bits for a page, clearing those bits.
3085 * It is not necessary for every reference bit to be cleared, but it
3086 * is necessary that 0 only be returned when there are truly no
3087 * reference bits set.
3089 * XXX: The exact number of bits to check and clear is a matter that
3090 * should be tested and standardized at some point in the future for
3091 * optimal aging of shared pages.
3096 pmap_ts_referenced(vm_page_t m)
3098 pv_entry_t pv, pvf, pvn;
3102 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3105 lwkt_gettoken(&vm_token);
3107 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3112 pvn = TAILQ_NEXT(pv, pv_list);
3115 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3116 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3119 if (!pmap_track_modified(pv->pv_va))
3122 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3124 if (pte && (*pte & PG_A)) {
3126 atomic_clear_int(pte, PG_A);
3128 atomic_clear_int_nonlocked(pte, PG_A);
3135 } while ((pv = pvn) != NULL && pv != pvf);
3138 lwkt_reltoken(&vm_token);
3144 * Return whether or not the specified physical page was modified
3145 * in any physical maps.
3150 pmap_is_modified(vm_page_t m)
3154 lwkt_gettoken(&vm_token);
3155 res = pmap_testbit(m, PG_M);
3156 lwkt_reltoken(&vm_token);
3161 * Clear the modify bits on the specified physical page.
3166 pmap_clear_modify(vm_page_t m)
3168 lwkt_gettoken(&vm_token);
3169 pmap_clearbit(m, PG_M);
3170 lwkt_reltoken(&vm_token);
3174 * Clear the reference bit on the specified physical page.
3179 pmap_clear_reference(vm_page_t m)
3181 lwkt_gettoken(&vm_token);
3182 pmap_clearbit(m, PG_A);
3183 lwkt_reltoken(&vm_token);
3187 * Miscellaneous support routines follow
3189 * Called from the low level boot code only.
3192 i386_protection_init(void)
3196 kp = protection_codes;
3197 for (prot = 0; prot < 8; prot++) {
3199 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3201 * Read access is also 0. There isn't any execute bit,
3202 * so just make it readable.
3204 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3205 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3206 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3209 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3210 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3211 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3212 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3220 * Map a set of physical memory pages into the kernel virtual
3221 * address space. Return a pointer to where it is mapped. This
3222 * routine is intended to be used for mapping device memory,
3225 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3231 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3233 vm_offset_t va, tmpva, offset;
3236 offset = pa & PAGE_MASK;
3237 size = roundup(offset + size, PAGE_SIZE);
3239 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3241 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3244 for (tmpva = va; size > 0;) {
3245 pte = (unsigned *)vtopte(tmpva);
3246 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3254 return ((void *)(va + offset));
3261 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3263 vm_offset_t base, offset;
3265 base = va & PG_FRAME;
3266 offset = va & PAGE_MASK;
3267 size = roundup(offset + size, PAGE_SIZE);
3268 pmap_qremove(va, size >> PAGE_SHIFT);
3269 kmem_free(&kernel_map, base, size);
3273 * perform the pmap work for mincore
3276 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3278 unsigned *ptep, pte;
3282 ptep = pmap_pte(pmap, addr);
3287 if ((pte = *ptep) != 0) {
3290 val = MINCORE_INCORE;
3291 if ((pte & PG_MANAGED) == 0)
3294 pa = pte & PG_FRAME;
3296 m = PHYS_TO_VM_PAGE(pa);
3302 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3304 * Modified by someone
3306 else if (m->dirty || pmap_is_modified(m))
3307 val |= MINCORE_MODIFIED_OTHER;
3312 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3315 * Referenced by someone
3317 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3318 val |= MINCORE_REFERENCED_OTHER;
3319 vm_page_flag_set(m, PG_REFERENCED);
3326 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3327 * vmspace will be ref'd and the old one will be deref'd.
3329 * The vmspace for all lwps associated with the process will be adjusted
3330 * and cr3 will be reloaded if any lwp is the current lwp.
3332 * Only called with new VM spaces.
3337 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3339 struct vmspace *oldvm;
3343 oldvm = p->p_vmspace;
3344 if (oldvm != newvm) {
3345 p->p_vmspace = newvm;
3346 KKASSERT(p->p_nthreads == 1);
3347 lp = RB_ROOT(&p->p_lwp_tree);
3348 pmap_setlwpvm(lp, newvm);
3350 sysref_get(&newvm->vm_sysref);
3351 sysref_put(&oldvm->vm_sysref);
3358 * Set the vmspace for a LWP. The vmspace is almost universally set the
3359 * same as the process vmspace, but virtual kernels need to swap out contexts
3360 * on a per-lwp basis.
3362 * Always called with a lp under the caller's direct control, either
3363 * unscheduled or the current lwp.
3368 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3370 struct vmspace *oldvm;
3374 oldvm = lp->lwp_vmspace;
3376 if (oldvm != newvm) {
3377 lp->lwp_vmspace = newvm;
3378 if (curthread->td_lwp == lp) {
3379 pmap = vmspace_pmap(newvm);
3381 atomic_set_int(&pmap->pm_active, mycpu->gd_cpumask);
3382 if (pmap->pm_active & CPUMASK_LOCK)
3383 pmap_interlock_wait(newvm);
3385 pmap->pm_active |= 1;
3387 #if defined(SWTCH_OPTIM_STATS)
3390 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3391 load_cr3(curthread->td_pcb->pcb_cr3);
3392 pmap = vmspace_pmap(oldvm);
3394 atomic_clear_int(&pmap->pm_active, mycpu->gd_cpumask);
3396 pmap->pm_active &= ~1;
3405 * Called when switching to a locked pmap, used to interlock against pmaps
3406 * undergoing modifications to prevent us from activating the MMU for the
3407 * target pmap until all such modifications have completed. We have to do
3408 * this because the thread making the modifications has already set up its
3409 * SMP synchronization mask.
3414 pmap_interlock_wait(struct vmspace *vm)
3416 struct pmap *pmap = &vm->vm_pmap;
3418 if (pmap->pm_active & CPUMASK_LOCK) {
3419 kprintf("Debug: pmap_interlock %08x\n", pmap->pm_active);
3420 while (pmap->pm_active & CPUMASK_LOCK) {
3423 lwkt_process_ipiq();
3431 * Return a page-directory alignment hint for device mappings which will
3432 * allow the use of super-pages for the mapping.
3437 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3440 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3444 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3449 * Return whether the PGE flag is supported globally.
3454 pmap_get_pgeflag(void)