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, "Debugging sysctl to force invltlb on pmap operations");
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)
594 KKASSERT(pv->pv_m == m);
596 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
603 panic("test_m_maps_pv: failed m %p pv %p\n", m, pv);
607 ptbase_assert(struct pmap *pmap)
609 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
611 /* are we current address space or kernel? */
612 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME))
614 KKASSERT(frame == (*mycpu->gd_GDMAP1 & PG_FRAME));
619 #define test_m_maps_pv(m, pv)
620 #define ptbase_assert(pmap)
624 #if defined(PMAP_DIAGNOSTIC)
627 * This code checks for non-writeable/modified pages.
628 * This should be an invalid condition.
631 pmap_nw_modified(pt_entry_t ptea)
637 if ((pte & (PG_M|PG_RW)) == PG_M)
646 * This routine defines the region(s) of memory that should not be tested
647 * for the modified bit.
651 static PMAP_INLINE int
652 pmap_track_modified(vm_offset_t va)
654 if ((va < clean_sva) || (va >= clean_eva))
661 * Retrieve the mapped page table base for a particular pmap. Use our self
662 * mapping for the kernel_pmap or our current pmap.
664 * For foreign pmaps we use the per-cpu page table map. Since this involves
665 * installing a ptd it's actually (per-process x per-cpu). However, we
666 * still cannot depend on our mapping to survive thread switches because
667 * the process might be threaded and switching to another thread for the
668 * same process on the same cpu will allow that other thread to make its
671 * This could be a bit confusing but the jist is for something like the
672 * vkernel which uses foreign pmaps all the time this represents a pretty
673 * good cache that avoids unnecessary invltlb()s.
675 * The caller must hold vm_token and the returned value is only valid
676 * until the caller blocks or releases the token.
679 get_ptbase(pmap_t pmap)
681 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
682 struct mdglobaldata *gd = mdcpu;
684 ASSERT_LWKT_TOKEN_HELD(&vm_token);
687 * We can use PTmap if the pmap is our current address space or
688 * the kernel address space.
690 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
691 return (unsigned *) PTmap;
695 * Otherwise we use the per-cpu alternative page table map. Each
696 * cpu gets its own map. Because of this we cannot use this map
697 * from interrupts or threads which can preempt.
699 * Even if we already have the map cached we may still have to
700 * invalidate the TLB if another cpu modified a PDE in the map.
702 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
703 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
705 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
706 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
707 pmap->pm_cached |= gd->mi.gd_cpumask;
709 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
710 pmap->pm_cached |= gd->mi.gd_cpumask;
712 } else if (dreadful_invltlb) {
715 return ((unsigned *)gd->gd_GDADDR1);
721 * Extract the physical page address associated with the map/VA pair.
723 * The caller may hold vm_token if it desires non-blocking operation.
726 pmap_extract(pmap_t pmap, vm_offset_t va)
729 vm_offset_t pdirindex;
731 lwkt_gettoken(&vm_token);
732 pdirindex = va >> PDRSHIFT;
733 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
735 if ((rtval & PG_PS) != 0) {
736 rtval &= ~(NBPDR - 1);
737 rtval |= va & (NBPDR - 1);
739 pte = get_ptbase(pmap) + i386_btop(va);
740 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
745 lwkt_reltoken(&vm_token);
749 /***************************************************
750 * Low level mapping routines.....
751 ***************************************************/
754 * Map a wired VM page to a KVA, fully SMP synchronized.
756 * No requirements, non blocking.
759 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
763 pmap_inval_info info;
765 pmap_inval_init(&info);
766 npte = pa | PG_RW | PG_V | pgeflag;
767 pte = (unsigned *)vtopte(va);
768 pmap_inval_interlock(&info, &kernel_pmap, va);
770 pmap_inval_deinterlock(&info, &kernel_pmap);
771 pmap_inval_done(&info);
775 * Map a wired VM page to a KVA, synchronized on current cpu only.
777 * No requirements, non blocking.
780 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
785 npte = pa | PG_RW | PG_V | pgeflag;
786 pte = (unsigned *)vtopte(va);
788 cpu_invlpg((void *)va);
792 * Synchronize a previously entered VA on all cpus.
794 * No requirements, non blocking.
797 pmap_kenter_sync(vm_offset_t va)
799 pmap_inval_info info;
801 pmap_inval_init(&info);
802 pmap_inval_interlock(&info, &kernel_pmap, va);
803 pmap_inval_deinterlock(&info, &kernel_pmap);
804 pmap_inval_done(&info);
808 * Synchronize a previously entered VA on the current cpu only.
810 * No requirements, non blocking.
813 pmap_kenter_sync_quick(vm_offset_t va)
815 cpu_invlpg((void *)va);
819 * Remove a page from the kernel pagetables, fully SMP synchronized.
821 * No requirements, non blocking.
824 pmap_kremove(vm_offset_t va)
827 pmap_inval_info info;
829 pmap_inval_init(&info);
830 pte = (unsigned *)vtopte(va);
831 pmap_inval_interlock(&info, &kernel_pmap, va);
833 pmap_inval_deinterlock(&info, &kernel_pmap);
834 pmap_inval_done(&info);
838 * Remove a page from the kernel pagetables, synchronized on current cpu only.
840 * No requirements, non blocking.
843 pmap_kremove_quick(vm_offset_t va)
846 pte = (unsigned *)vtopte(va);
848 cpu_invlpg((void *)va);
852 * Adjust the permissions of a page in the kernel page table,
853 * synchronized on the current cpu only.
855 * No requirements, non blocking.
858 pmap_kmodify_rw(vm_offset_t va)
860 atomic_set_int(vtopte(va), PG_RW);
861 cpu_invlpg((void *)va);
865 * Adjust the permissions of a page in the kernel page table,
866 * synchronized on the current cpu only.
868 * No requirements, non blocking.
871 pmap_kmodify_nc(vm_offset_t va)
873 atomic_set_int(vtopte(va), PG_N);
874 cpu_invlpg((void *)va);
878 * Map a range of physical addresses into kernel virtual address space.
880 * No requirements, non blocking.
883 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
885 vm_offset_t sva, virt;
888 while (start < end) {
889 pmap_kenter(virt, start);
898 * Add a list of wired pages to the kva, fully SMP synchronized.
900 * No requirements, non blocking.
903 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
907 end_va = va + count * PAGE_SIZE;
909 while (va < end_va) {
912 pte = (unsigned *)vtopte(va);
913 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
914 cpu_invlpg((void *)va);
919 smp_invltlb(); /* XXX */
924 * Remove pages from KVA, fully SMP synchronized.
926 * No requirements, non blocking.
929 pmap_qremove(vm_offset_t va, int count)
933 end_va = va + count*PAGE_SIZE;
935 while (va < end_va) {
938 pte = (unsigned *)vtopte(va);
940 cpu_invlpg((void *)va);
949 * This routine works like vm_page_lookup() but also blocks as long as the
950 * page is busy. This routine does not busy the page it returns.
952 * The caller must hold vm_token.
955 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
959 ASSERT_LWKT_TOKEN_HELD(&vm_token);
961 m = vm_page_lookup(object, pindex);
962 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
968 * Create a new thread and optionally associate it with a (new) process.
969 * NOTE! the new thread's cpu may not equal the current cpu.
972 pmap_init_thread(thread_t td)
974 /* enforce pcb placement */
975 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
976 td->td_savefpu = &td->td_pcb->pcb_save;
977 td->td_sp = (char *)td->td_pcb - 16;
981 * This routine directly affects the fork perf for a process.
984 pmap_init_proc(struct proc *p)
989 * Dispose the UPAGES for a process that has exited.
990 * This routine directly impacts the exit perf of a process.
993 pmap_dispose_proc(struct proc *p)
995 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
998 /***************************************************
999 * Page table page management routines.....
1000 ***************************************************/
1003 * This routine unholds page table pages, and if the hold count
1004 * drops to zero, then it decrements the wire count.
1006 * The caller must hold vm_token.
1007 * This function can block.
1010 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1013 * Wait until we can busy the page ourselves. We cannot have
1014 * any active flushes if we block.
1016 if (m->flags & PG_BUSY) {
1017 pmap_inval_flush(info);
1018 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1021 KASSERT(m->queue == PQ_NONE,
1022 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1024 if (m->hold_count == 1) {
1026 * Unmap the page table page.
1028 * NOTE: We must clear pm_cached for all cpus, including
1029 * the current one, when clearing a page directory
1033 pmap_inval_interlock(info, pmap, -1);
1034 KKASSERT(pmap->pm_pdir[m->pindex]);
1035 pmap->pm_pdir[m->pindex] = 0;
1036 pmap->pm_cached = 0;
1037 pmap_inval_deinterlock(info, pmap);
1039 KKASSERT(pmap->pm_stats.resident_count > 0);
1040 --pmap->pm_stats.resident_count;
1042 if (pmap->pm_ptphint == m)
1043 pmap->pm_ptphint = NULL;
1046 * This was our last hold, the page had better be unwired
1047 * after we decrement wire_count.
1049 * FUTURE NOTE: shared page directory page could result in
1050 * multiple wire counts.
1054 KKASSERT(m->wire_count == 0);
1055 --vmstats.v_wire_count;
1056 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1058 vm_page_free_zero(m);
1061 KKASSERT(m->hold_count > 1);
1068 * The caller must hold vm_token.
1069 * This function can block.
1071 static PMAP_INLINE int
1072 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1074 KKASSERT(m->hold_count > 0);
1075 if (m->hold_count > 1) {
1079 return _pmap_unwire_pte_hold(pmap, m, info);
1084 * After removing a (user) page table entry, this routine is used to
1085 * conditionally free the page, and manage the hold/wire counts.
1087 * The caller must hold vm_token.
1088 * This function can block regardless.
1091 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1092 pmap_inval_info_t info)
1096 if (va >= UPT_MIN_ADDRESS)
1100 ptepindex = (va >> PDRSHIFT);
1101 if (pmap->pm_ptphint &&
1102 (pmap->pm_ptphint->pindex == ptepindex)) {
1103 mpte = pmap->pm_ptphint;
1105 pmap_inval_flush(info);
1106 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1107 pmap->pm_ptphint = mpte;
1111 return pmap_unwire_pte_hold(pmap, mpte, info);
1115 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1116 * it, and IdlePTD, represents the template used to update all other pmaps.
1118 * On architectures where the kernel pmap is not integrated into the user
1119 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1120 * kernel_pmap should be used to directly access the kernel_pmap.
1125 pmap_pinit0(struct pmap *pmap)
1128 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1129 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1131 pmap->pm_active = 0;
1132 pmap->pm_cached = 0;
1133 pmap->pm_ptphint = NULL;
1134 TAILQ_INIT(&pmap->pm_pvlist);
1135 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1139 * Initialize a preallocated and zeroed pmap structure,
1140 * such as one in a vmspace structure.
1145 pmap_pinit(struct pmap *pmap)
1150 * No need to allocate page table space yet but we do need a valid
1151 * page directory table.
1153 if (pmap->pm_pdir == NULL) {
1155 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1159 * Allocate an object for the ptes
1161 if (pmap->pm_pteobj == NULL)
1162 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1165 * Allocate the page directory page, unless we already have
1166 * one cached. If we used the cached page the wire_count will
1167 * already be set appropriately.
1169 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1170 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1171 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1172 pmap->pm_pdirm = ptdpg;
1173 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1174 ptdpg->valid = VM_PAGE_BITS_ALL;
1175 ptdpg->wire_count = 1;
1176 ++vmstats.v_wire_count;
1177 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1179 if ((ptdpg->flags & PG_ZERO) == 0)
1180 bzero(pmap->pm_pdir, PAGE_SIZE);
1183 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1186 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1188 /* install self-referential address mapping entry */
1189 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1190 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1193 pmap->pm_active = 0;
1194 pmap->pm_cached = 0;
1195 pmap->pm_ptphint = NULL;
1196 TAILQ_INIT(&pmap->pm_pvlist);
1197 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1198 pmap->pm_stats.resident_count = 1;
1202 * Clean up a pmap structure so it can be physically freed. This routine
1203 * is called by the vmspace dtor function. A great deal of pmap data is
1204 * left passively mapped to improve vmspace management so we have a bit
1205 * of cleanup work to do here.
1210 pmap_puninit(pmap_t pmap)
1214 KKASSERT(pmap->pm_active == 0);
1215 lwkt_gettoken(&vm_token);
1216 if ((p = pmap->pm_pdirm) != NULL) {
1217 KKASSERT(pmap->pm_pdir != NULL);
1218 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1220 vmstats.v_wire_count--;
1221 KKASSERT((p->flags & PG_BUSY) == 0);
1223 vm_page_free_zero(p);
1224 pmap->pm_pdirm = NULL;
1226 lwkt_reltoken(&vm_token);
1227 if (pmap->pm_pdir) {
1228 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1229 pmap->pm_pdir = NULL;
1231 if (pmap->pm_pteobj) {
1232 vm_object_deallocate(pmap->pm_pteobj);
1233 pmap->pm_pteobj = NULL;
1238 * Wire in kernel global address entries. To avoid a race condition
1239 * between pmap initialization and pmap_growkernel, this procedure
1240 * adds the pmap to the master list (which growkernel scans to update),
1241 * then copies the template.
1246 pmap_pinit2(struct pmap *pmap)
1249 lwkt_gettoken(&vm_token);
1250 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1251 /* XXX copies current process, does not fill in MPPTDI */
1252 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1253 lwkt_reltoken(&vm_token);
1258 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1259 * 0 on failure (if the procedure had to sleep).
1261 * When asked to remove the page directory page itself, we actually just
1262 * leave it cached so we do not have to incur the SMP inval overhead of
1263 * removing the kernel mapping. pmap_puninit() will take care of it.
1265 * The caller must hold vm_token.
1266 * This function can block regardless.
1269 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1271 unsigned *pde = (unsigned *) pmap->pm_pdir;
1274 * This code optimizes the case of freeing non-busy
1275 * page-table pages. Those pages are zero now, and
1276 * might as well be placed directly into the zero queue.
1278 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1284 * Remove the page table page from the processes address space.
1286 KKASSERT(pmap->pm_stats.resident_count > 0);
1287 KKASSERT(pde[p->pindex]);
1289 --pmap->pm_stats.resident_count;
1290 pmap->pm_cached = 0;
1292 if (p->hold_count) {
1293 panic("pmap_release: freeing held page table page");
1295 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1296 pmap->pm_ptphint = NULL;
1299 * We leave the page directory page cached, wired, and mapped in
1300 * the pmap until the dtor function (pmap_puninit()) gets called.
1301 * However, still clean it up so we can set PG_ZERO.
1303 * The pmap has already been removed from the pmap_list in the
1306 if (p->pindex == PTDPTDI) {
1307 bzero(pde + KPTDI, nkpt * PTESIZE);
1308 bzero(pde + MPPTDI, (NPDEPG - MPPTDI) * PTESIZE);
1309 vm_page_flag_set(p, PG_ZERO);
1313 vmstats.v_wire_count--;
1314 vm_page_free_zero(p);
1320 * This routine is called if the page table page is not mapped correctly.
1322 * The caller must hold vm_token.
1325 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1327 vm_offset_t pteva, ptepa;
1331 * Find or fabricate a new pagetable page
1333 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1334 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1336 KASSERT(m->queue == PQ_NONE,
1337 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1340 * Increment the hold count for the page we will be returning to
1346 * It is possible that someone else got in and mapped by the page
1347 * directory page while we were blocked, if so just unbusy and
1348 * return the held page.
1350 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1351 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1356 if (m->wire_count == 0)
1357 vmstats.v_wire_count++;
1362 * Map the pagetable page into the process address space, if
1363 * it isn't already there.
1365 * NOTE: For safety clear pm_cached for all cpus including the
1366 * current one when adding a PDE to the map.
1368 ++pmap->pm_stats.resident_count;
1370 ptepa = VM_PAGE_TO_PHYS(m);
1371 pmap->pm_pdir[ptepindex] =
1372 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1373 pmap->pm_cached = 0;
1376 * Set the page table hint
1378 pmap->pm_ptphint = m;
1381 * Try to use the new mapping, but if we cannot, then
1382 * do it with the routine that maps the page explicitly.
1384 if ((m->flags & PG_ZERO) == 0) {
1385 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1386 (((unsigned) PTDpde) & PG_FRAME)) {
1387 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1388 bzero((caddr_t) pteva, PAGE_SIZE);
1390 pmap_zero_page(ptepa);
1395 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1399 m->valid = VM_PAGE_BITS_ALL;
1400 vm_page_flag_clear(m, PG_ZERO);
1401 vm_page_flag_set(m, PG_MAPPED);
1408 * Allocate a page table entry for a va.
1410 * The caller must hold vm_token.
1413 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1420 * Calculate pagetable page index
1422 ptepindex = va >> PDRSHIFT;
1425 * Get the page directory entry
1427 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1430 * This supports switching from a 4MB page to a
1433 if (ptepa & PG_PS) {
1434 pmap->pm_pdir[ptepindex] = 0;
1441 * If the page table page is mapped, we just increment the
1442 * hold count, and activate it.
1446 * In order to get the page table page, try the
1449 if (pmap->pm_ptphint &&
1450 (pmap->pm_ptphint->pindex == ptepindex)) {
1451 m = pmap->pm_ptphint;
1453 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1454 pmap->pm_ptphint = m;
1460 * Here if the pte page isn't mapped, or if it has been deallocated.
1462 return _pmap_allocpte(pmap, ptepindex);
1466 /***************************************************
1467 * Pmap allocation/deallocation routines.
1468 ***************************************************/
1471 * Release any resources held by the given physical map.
1472 * Called when a pmap initialized by pmap_pinit is being released.
1473 * Should only be called if the map contains no valid mappings.
1477 static int pmap_release_callback(struct vm_page *p, void *data);
1480 pmap_release(struct pmap *pmap)
1482 vm_object_t object = pmap->pm_pteobj;
1483 struct rb_vm_page_scan_info info;
1485 KASSERT(pmap->pm_active == 0,
1486 ("pmap still active! %08x", pmap->pm_active));
1487 #if defined(DIAGNOSTIC)
1488 if (object->ref_count != 1)
1489 panic("pmap_release: pteobj reference count != 1");
1493 info.object = object;
1495 lwkt_gettoken(&vm_token);
1496 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1503 info.limit = object->generation;
1505 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1506 pmap_release_callback, &info);
1507 if (info.error == 0 && info.mpte) {
1508 if (!pmap_release_free_page(pmap, info.mpte))
1512 } while (info.error);
1513 pmap->pm_cached = 0;
1514 lwkt_reltoken(&vm_token);
1518 * The caller must hold vm_token.
1521 pmap_release_callback(struct vm_page *p, void *data)
1523 struct rb_vm_page_scan_info *info = data;
1525 if (p->pindex == PTDPTDI) {
1529 if (!pmap_release_free_page(info->pmap, p)) {
1533 if (info->object->generation != info->limit) {
1541 * Grow the number of kernel page table entries, if needed.
1546 pmap_growkernel(vm_offset_t addr)
1549 vm_offset_t ptppaddr;
1554 lwkt_gettoken(&vm_token);
1555 if (kernel_vm_end == 0) {
1556 kernel_vm_end = KERNBASE;
1558 while (pdir_pde(PTD, kernel_vm_end)) {
1559 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1560 ~(PAGE_SIZE * NPTEPG - 1);
1564 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1565 while (kernel_vm_end < addr) {
1566 if (pdir_pde(PTD, kernel_vm_end)) {
1567 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1568 ~(PAGE_SIZE * NPTEPG - 1);
1573 * This index is bogus, but out of the way
1575 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_NORMAL |
1577 VM_ALLOC_INTERRUPT);
1579 panic("pmap_growkernel: no memory to grow kernel");
1582 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1583 pmap_zero_page(ptppaddr);
1584 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1585 pdir_pde(PTD, kernel_vm_end) = newpdir;
1586 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1590 * This update must be interlocked with pmap_pinit2.
1592 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1593 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1595 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1596 ~(PAGE_SIZE * NPTEPG - 1);
1598 lwkt_reltoken(&vm_token);
1603 * Retire the given physical map from service.
1605 * Should only be called if the map contains no valid mappings.
1610 pmap_destroy(pmap_t pmap)
1615 lwkt_gettoken(&vm_token);
1616 if (--pmap->pm_count == 0) {
1618 panic("destroying a pmap is not yet implemented");
1620 lwkt_reltoken(&vm_token);
1624 * Add a reference to the specified pmap.
1629 pmap_reference(pmap_t pmap)
1632 lwkt_gettoken(&vm_token);
1634 lwkt_reltoken(&vm_token);
1638 /***************************************************
1639 * page management routines.
1640 ***************************************************/
1643 * free the pv_entry back to the free list. This function may be
1644 * called from an interrupt.
1646 * The caller must hold vm_token.
1648 static PMAP_INLINE void
1649 free_pv_entry(pv_entry_t pv)
1652 KKASSERT(pv->pv_m != NULL);
1660 * get a new pv_entry, allocating a block from the system
1661 * when needed. This function may be called from an interrupt.
1663 * The caller must hold vm_token.
1669 if (pv_entry_high_water &&
1670 (pv_entry_count > pv_entry_high_water) &&
1671 (pmap_pagedaemon_waken == 0)) {
1672 pmap_pagedaemon_waken = 1;
1673 wakeup (&vm_pages_needed);
1675 return zalloc(pvzone);
1679 * This routine is very drastic, but can save the system
1689 static int warningdone=0;
1691 if (pmap_pagedaemon_waken == 0)
1693 lwkt_gettoken(&vm_token);
1694 pmap_pagedaemon_waken = 0;
1696 if (warningdone < 5) {
1697 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1701 for(i = 0; i < vm_page_array_size; i++) {
1702 m = &vm_page_array[i];
1703 if (m->wire_count || m->hold_count || m->busy ||
1704 (m->flags & PG_BUSY)) {
1709 lwkt_reltoken(&vm_token);
1714 * If it is the first entry on the list, it is actually
1715 * in the header and we must copy the following entry up
1716 * to the header. Otherwise we must search the list for
1717 * the entry. In either case we free the now unused entry.
1719 * The caller must hold vm_token.
1722 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1723 vm_offset_t va, pmap_inval_info_t info)
1729 ASSERT_LWKT_TOKEN_HELD(&vm_token);
1730 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1731 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1732 if (pmap == pv->pv_pmap && va == pv->pv_va)
1736 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1738 KKASSERT(pv->pv_pmap == pmap);
1740 if (va == pv->pv_va)
1747 test_m_maps_pv(m, pv);
1748 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1749 m->md.pv_list_count--;
1750 if (TAILQ_EMPTY(&m->md.pv_list))
1751 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1752 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1753 ++pmap->pm_generation;
1754 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1761 * Create a pv entry for page at pa for (pmap, va).
1763 * The caller must hold vm_token.
1766 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1771 pv = get_pv_entry();
1773 KKASSERT(pv->pv_m == NULL);
1780 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1781 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1782 ++pmap->pm_generation;
1783 m->md.pv_list_count++;
1789 * pmap_remove_pte: do the things to unmap a page in a process.
1791 * The caller must hold vm_token.
1793 * WARNING! As with most other pmap functions this one can block, so
1794 * callers using temporary page table mappings must reload
1798 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1799 pmap_inval_info_t info)
1804 ptbase_assert(pmap);
1805 pmap_inval_interlock(info, pmap, va);
1806 ptbase_assert(pmap);
1807 oldpte = loadandclear(ptq);
1809 pmap->pm_stats.wired_count -= 1;
1810 pmap_inval_deinterlock(info, pmap);
1813 * Machines that don't support invlpg, also don't support
1814 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1818 cpu_invlpg((void *)va);
1819 KKASSERT(pmap->pm_stats.resident_count > 0);
1820 --pmap->pm_stats.resident_count;
1821 if (oldpte & PG_MANAGED) {
1822 m = PHYS_TO_VM_PAGE(oldpte);
1823 if (oldpte & PG_M) {
1824 #if defined(PMAP_DIAGNOSTIC)
1825 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1826 kprintf("pmap_remove: modified page not "
1827 "writable: va: %p, pte: 0x%lx\n",
1828 (void *)va, (long)oldpte);
1831 if (pmap_track_modified(va))
1835 vm_page_flag_set(m, PG_REFERENCED);
1836 return pmap_remove_entry(pmap, m, va, info);
1838 return pmap_unuse_pt(pmap, va, NULL, info);
1845 * Remove a single page from a process address space.
1847 * The caller must hold vm_token.
1850 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1855 * if there is no pte for this address, just skip it!!! Otherwise
1856 * get a local va for mappings for this pmap and remove the entry.
1858 if (*pmap_pde(pmap, va) != 0) {
1859 ptq = get_ptbase(pmap) + i386_btop(va);
1861 pmap_remove_pte(pmap, ptq, va, info);
1868 * Remove the given range of addresses from the specified map.
1870 * It is assumed that the start and end are properly rounded to the page
1876 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1880 vm_offset_t ptpaddr;
1881 vm_offset_t sindex, eindex;
1882 struct pmap_inval_info info;
1887 lwkt_gettoken(&vm_token);
1888 if (pmap->pm_stats.resident_count == 0) {
1889 lwkt_reltoken(&vm_token);
1893 pmap_inval_init(&info);
1896 * special handling of removing one page. a very
1897 * common operation and easy to short circuit some
1900 if (((sva + PAGE_SIZE) == eva) &&
1901 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1902 pmap_remove_page(pmap, sva, &info);
1903 pmap_inval_done(&info);
1904 lwkt_reltoken(&vm_token);
1909 * Get a local virtual address for the mappings that are being
1912 sindex = i386_btop(sva);
1913 eindex = i386_btop(eva);
1915 for (; sindex < eindex; sindex = pdnxt) {
1919 * Calculate index for next page table.
1921 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1922 if (pmap->pm_stats.resident_count == 0)
1925 pdirindex = sindex / NPDEPG;
1926 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1927 pmap_inval_interlock(&info, pmap, -1);
1928 pmap->pm_pdir[pdirindex] = 0;
1929 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1930 pmap->pm_cached = 0;
1931 pmap_inval_deinterlock(&info, pmap);
1936 * Weed out invalid mappings. Note: we assume that the page
1937 * directory table is always allocated, and in kernel virtual.
1943 * Limit our scan to either the end of the va represented
1944 * by the current page table page, or to the end of the
1945 * range being removed.
1947 if (pdnxt > eindex) {
1952 * NOTE: pmap_remove_pte() can block and wipe the temporary
1955 for (; sindex != pdnxt; sindex++) {
1958 ptbase = get_ptbase(pmap);
1959 if (ptbase[sindex] == 0)
1961 va = i386_ptob(sindex);
1962 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1966 pmap_inval_done(&info);
1967 lwkt_reltoken(&vm_token);
1971 * Removes this physical page from all physical maps in which it resides.
1972 * Reflects back modify bits to the pager.
1977 pmap_remove_all(vm_page_t m)
1979 struct pmap_inval_info info;
1980 unsigned *pte, tpte;
1983 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1986 lwkt_gettoken(&vm_token);
1987 pmap_inval_init(&info);
1988 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1989 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1990 --pv->pv_pmap->pm_stats.resident_count;
1992 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1993 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
1994 tpte = loadandclear(pte);
1996 pv->pv_pmap->pm_stats.wired_count--;
1997 pmap_inval_deinterlock(&info, pv->pv_pmap);
1999 vm_page_flag_set(m, PG_REFERENCED);
2001 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
2005 * Update the vm_page_t clean and reference bits.
2008 #if defined(PMAP_DIAGNOSTIC)
2009 if (pmap_nw_modified((pt_entry_t) tpte)) {
2010 kprintf("pmap_remove_all: modified page "
2011 "not writable: va: %p, pte: 0x%lx\n",
2012 (void *)pv->pv_va, (long)tpte);
2015 if (pmap_track_modified(pv->pv_va))
2019 KKASSERT(pv->pv_m == m);
2021 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2022 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2023 ++pv->pv_pmap->pm_generation;
2024 m->md.pv_list_count--;
2025 if (TAILQ_EMPTY(&m->md.pv_list))
2026 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2027 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2030 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2031 pmap_inval_done(&info);
2032 lwkt_reltoken(&vm_token);
2036 * Set the physical protection on the specified range of this map
2042 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2045 vm_offset_t pdnxt, ptpaddr;
2046 vm_pindex_t sindex, eindex;
2047 pmap_inval_info info;
2052 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2053 pmap_remove(pmap, sva, eva);
2057 if (prot & VM_PROT_WRITE)
2060 lwkt_gettoken(&vm_token);
2061 pmap_inval_init(&info);
2063 ptbase = get_ptbase(pmap);
2065 sindex = i386_btop(sva);
2066 eindex = i386_btop(eva);
2068 for (; sindex < eindex; sindex = pdnxt) {
2071 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2073 pdirindex = sindex / NPDEPG;
2074 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
2075 pmap_inval_interlock(&info, pmap, -1);
2076 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2077 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2078 pmap_inval_deinterlock(&info, pmap);
2083 * Weed out invalid mappings. Note: we assume that the page
2084 * directory table is always allocated, and in kernel virtual.
2089 if (pdnxt > eindex) {
2093 for (; sindex != pdnxt; sindex++) {
2099 * XXX non-optimal. Note also that there can be
2100 * no pmap_inval_flush() calls until after we modify
2101 * ptbase[sindex] (or otherwise we have to do another
2102 * pmap_inval_interlock() call).
2104 pmap_inval_interlock(&info, pmap, i386_ptob(sindex));
2106 pbits = ptbase[sindex];
2109 if (pbits & PG_MANAGED) {
2112 m = PHYS_TO_VM_PAGE(pbits);
2113 vm_page_flag_set(m, PG_REFERENCED);
2117 if (pmap_track_modified(i386_ptob(sindex))) {
2119 m = PHYS_TO_VM_PAGE(pbits);
2126 if (pbits != cbits &&
2127 !atomic_cmpset_int(ptbase + sindex, pbits, cbits)) {
2130 pmap_inval_deinterlock(&info, pmap);
2133 pmap_inval_done(&info);
2134 lwkt_reltoken(&vm_token);
2138 * Insert the given physical page (p) at the specified virtual address (v)
2139 * in the target physical map with the protection requested.
2141 * If specified, the page will be wired down, meaning that the related pte
2142 * cannot be reclaimed.
2147 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2153 vm_offset_t origpte, newpte;
2155 pmap_inval_info info;
2161 #ifdef PMAP_DIAGNOSTIC
2163 panic("pmap_enter: toobig");
2164 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2165 panic("pmap_enter: invalid to pmap_enter page "
2166 "table pages (va: %p)", (void *)va);
2169 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2170 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2171 print_backtrace(-1);
2173 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2174 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2175 print_backtrace(-1);
2178 lwkt_gettoken(&vm_token);
2181 * In the case that a page table page is not
2182 * resident, we are creating it here.
2184 if (va < UPT_MIN_ADDRESS)
2185 mpte = pmap_allocpte(pmap, va);
2189 pmap_inval_init(&info);
2190 pte = pmap_pte(pmap, va);
2193 * Page Directory table entry not valid, we need a new PT page
2196 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p\n",
2197 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
2200 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2201 origpte = *(vm_offset_t *)pte;
2202 opa = origpte & PG_FRAME;
2204 if (origpte & PG_PS)
2205 panic("pmap_enter: attempted pmap_enter on 4MB page");
2208 * Mapping has not changed, must be protection or wiring change.
2210 if (origpte && (opa == pa)) {
2212 * Wiring change, just update stats. We don't worry about
2213 * wiring PT pages as they remain resident as long as there
2214 * are valid mappings in them. Hence, if a user page is wired,
2215 * the PT page will be also.
2217 if (wired && ((origpte & PG_W) == 0))
2218 pmap->pm_stats.wired_count++;
2219 else if (!wired && (origpte & PG_W))
2220 pmap->pm_stats.wired_count--;
2222 #if defined(PMAP_DIAGNOSTIC)
2223 if (pmap_nw_modified((pt_entry_t) origpte)) {
2224 kprintf("pmap_enter: modified page not "
2225 "writable: va: %p, pte: 0x%lx\n",
2226 (void *)va, (long )origpte);
2231 * Remove the extra pte reference. Note that we cannot
2232 * optimize the RO->RW case because we have adjusted the
2233 * wiring count above and may need to adjust the wiring
2240 * We might be turning off write access to the page,
2241 * so we go ahead and sense modify status.
2243 if (origpte & PG_MANAGED) {
2244 if ((origpte & PG_M) && pmap_track_modified(va)) {
2246 om = PHYS_TO_VM_PAGE(opa);
2250 KKASSERT(m->flags & PG_MAPPED);
2255 * Mapping has changed, invalidate old range and fall through to
2256 * handle validating new mapping.
2258 * Since we have a ref on the page directory page pmap_pte()
2259 * will always return non-NULL.
2261 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2262 * to get wiped. reload the ptbase. I'm not sure if it is
2263 * also possible to race another pmap_enter() but check for
2269 KKASSERT((origpte & PG_FRAME) ==
2270 (*(vm_offset_t *)pte & PG_FRAME));
2271 err = pmap_remove_pte(pmap, pte, va, &info);
2273 panic("pmap_enter: pte vanished, va: %p", (void *)va);
2274 pte = pmap_pte(pmap, va);
2275 origpte = *(vm_offset_t *)pte;
2276 opa = origpte & PG_FRAME;
2278 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2284 * Enter on the PV list if part of our managed memory. Note that we
2285 * raise IPL while manipulating pv_table since pmap_enter can be
2286 * called at interrupt time.
2288 if (pmap_initialized &&
2289 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2290 pmap_insert_entry(pmap, va, mpte, m);
2291 ptbase_assert(pmap);
2293 vm_page_flag_set(m, PG_MAPPED);
2297 * Increment counters
2299 ++pmap->pm_stats.resident_count;
2301 pmap->pm_stats.wired_count++;
2302 KKASSERT(*pte == 0);
2306 * Now validate mapping with desired protection/wiring.
2308 ptbase_assert(pmap);
2309 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2313 if (va < UPT_MIN_ADDRESS)
2315 if (pmap == &kernel_pmap)
2319 * if the mapping or permission bits are different, we need
2320 * to update the pte.
2322 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2323 pmap_inval_interlock(&info, pmap, va);
2324 ptbase_assert(pmap);
2325 KKASSERT(*pte == 0 ||
2326 (*pte & PG_FRAME) == (newpte & PG_FRAME));
2327 *pte = newpte | PG_A;
2328 pmap_inval_deinterlock(&info, pmap);
2330 vm_page_flag_set(m, PG_WRITEABLE);
2332 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2333 pmap_inval_done(&info);
2334 lwkt_reltoken(&vm_token);
2338 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2339 * This code also assumes that the pmap has no pre-existing entry for this
2342 * This code currently may only be used on user pmaps, not kernel_pmap.
2347 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2354 pmap_inval_info info;
2356 lwkt_gettoken(&vm_token);
2357 pmap_inval_init(&info);
2359 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2360 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2361 print_backtrace(-1);
2363 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2364 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2365 print_backtrace(-1);
2368 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2371 * Calculate the page table page (mpte), allocating it if necessary.
2373 * A held page table page (mpte), or NULL, is passed onto the
2374 * section following.
2376 if (va < UPT_MIN_ADDRESS) {
2378 * Calculate pagetable page index
2380 ptepindex = va >> PDRSHIFT;
2384 * Get the page directory entry
2386 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2389 * If the page table page is mapped, we just increment
2390 * the hold count, and activate it.
2394 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2395 if (pmap->pm_ptphint &&
2396 (pmap->pm_ptphint->pindex == ptepindex)) {
2397 mpte = pmap->pm_ptphint;
2399 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2400 pmap->pm_ptphint = mpte;
2405 mpte = _pmap_allocpte(pmap, ptepindex);
2407 } while (mpte == NULL);
2410 /* this code path is not yet used */
2414 * With a valid (and held) page directory page, we can just use
2415 * vtopte() to get to the pte. If the pte is already present
2416 * we do not disturb it.
2418 pte = (unsigned *)vtopte(va);
2421 pmap_unwire_pte_hold(pmap, mpte, &info);
2422 pa = VM_PAGE_TO_PHYS(m);
2423 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2424 pmap_inval_done(&info);
2425 lwkt_reltoken(&vm_token);
2430 * Enter on the PV list if part of our managed memory
2432 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2433 pmap_insert_entry(pmap, va, mpte, m);
2434 vm_page_flag_set(m, PG_MAPPED);
2438 * Increment counters
2440 ++pmap->pm_stats.resident_count;
2442 pa = VM_PAGE_TO_PHYS(m);
2445 * Now validate mapping with RO protection
2447 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2448 *pte = pa | PG_V | PG_U;
2450 *pte = pa | PG_V | PG_U | PG_MANAGED;
2451 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2452 pmap_inval_done(&info);
2453 lwkt_reltoken(&vm_token);
2457 * Make a temporary mapping for a physical address. This is only intended
2458 * to be used for panic dumps.
2463 pmap_kenter_temporary(vm_paddr_t pa, int i)
2465 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2466 return ((void *)crashdumpmap);
2469 #define MAX_INIT_PT (96)
2472 * This routine preloads the ptes for a given object into the specified pmap.
2473 * This eliminates the blast of soft faults on process startup and
2474 * immediately after an mmap.
2478 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2481 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2482 vm_object_t object, vm_pindex_t pindex,
2483 vm_size_t size, int limit)
2485 struct rb_vm_page_scan_info info;
2490 * We can't preinit if read access isn't set or there is no pmap
2493 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2497 * We can't preinit if the pmap is not the current pmap
2499 lp = curthread->td_lwp;
2500 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2503 psize = i386_btop(size);
2505 if ((object->type != OBJT_VNODE) ||
2506 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2507 (object->resident_page_count > MAX_INIT_PT))) {
2511 if (psize + pindex > object->size) {
2512 if (object->size < pindex)
2514 psize = object->size - pindex;
2521 * Use a red-black scan to traverse the requested range and load
2522 * any valid pages found into the pmap.
2524 * We cannot safely scan the object's memq unless we are in a
2525 * critical section since interrupts can remove pages from objects.
2527 info.start_pindex = pindex;
2528 info.end_pindex = pindex + psize - 1;
2535 lwkt_gettoken(&vm_token);
2536 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2537 pmap_object_init_pt_callback, &info);
2538 lwkt_reltoken(&vm_token);
2543 * The caller must hold vm_token.
2547 pmap_object_init_pt_callback(vm_page_t p, void *data)
2549 struct rb_vm_page_scan_info *info = data;
2550 vm_pindex_t rel_index;
2552 * don't allow an madvise to blow away our really
2553 * free pages allocating pv entries.
2555 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2556 vmstats.v_free_count < vmstats.v_free_reserved) {
2559 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2560 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2561 if ((p->queue - p->pc) == PQ_CACHE)
2562 vm_page_deactivate(p);
2564 rel_index = p->pindex - info->start_pindex;
2565 pmap_enter_quick(info->pmap,
2566 info->addr + i386_ptob(rel_index), p);
2573 * Return TRUE if the pmap is in shape to trivially
2574 * pre-fault the specified address.
2576 * Returns FALSE if it would be non-trivial or if a
2577 * pte is already loaded into the slot.
2582 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2587 lwkt_gettoken(&vm_token);
2588 if ((*pmap_pde(pmap, addr)) == 0) {
2591 pte = (unsigned *) vtopte(addr);
2592 ret = (*pte) ? 0 : 1;
2594 lwkt_reltoken(&vm_token);
2599 * Change the wiring attribute for a map/virtual-adderss pair. The mapping
2600 * must already exist.
2605 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2612 lwkt_gettoken(&vm_token);
2613 pte = pmap_pte(pmap, va);
2615 if (wired && !pmap_pte_w(pte))
2616 pmap->pm_stats.wired_count++;
2617 else if (!wired && pmap_pte_w(pte))
2618 pmap->pm_stats.wired_count--;
2621 * Wiring is not a hardware characteristic so there is no need to
2622 * invalidate TLB. However, in an SMP environment we must use
2623 * a locked bus cycle to update the pte (if we are not using
2624 * the pmap_inval_*() API that is)... it's ok to do this for simple
2629 atomic_set_int(pte, PG_W);
2631 atomic_clear_int(pte, PG_W);
2634 atomic_set_int_nonlocked(pte, PG_W);
2636 atomic_clear_int_nonlocked(pte, PG_W);
2638 lwkt_reltoken(&vm_token);
2642 * Copy the range specified by src_addr/len from the source map to the
2643 * range dst_addr/len in the destination map.
2645 * This routine is only advisory and need not do anything.
2650 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2651 vm_size_t len, vm_offset_t src_addr)
2657 * Zero the specified PA by mapping the page into KVM and clearing its
2663 pmap_zero_page(vm_paddr_t phys)
2665 struct mdglobaldata *gd = mdcpu;
2668 if (*(int *)gd->gd_CMAP3)
2669 panic("pmap_zero_page: CMAP3 busy");
2670 *(int *)gd->gd_CMAP3 =
2671 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2672 cpu_invlpg(gd->gd_CADDR3);
2674 #if defined(I686_CPU)
2675 if (cpu_class == CPUCLASS_686)
2676 i686_pagezero(gd->gd_CADDR3);
2679 bzero(gd->gd_CADDR3, PAGE_SIZE);
2680 *(int *) gd->gd_CMAP3 = 0;
2685 * Assert that a page is empty, panic if it isn't.
2690 pmap_page_assertzero(vm_paddr_t phys)
2692 struct mdglobaldata *gd = mdcpu;
2696 if (*(int *)gd->gd_CMAP3)
2697 panic("pmap_zero_page: CMAP3 busy");
2698 *(int *)gd->gd_CMAP3 =
2699 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2700 cpu_invlpg(gd->gd_CADDR3);
2701 for (i = 0; i < PAGE_SIZE; i += 4) {
2702 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2703 panic("pmap_page_assertzero() @ %p not zero!\n",
2704 (void *)gd->gd_CADDR3);
2707 *(int *) gd->gd_CMAP3 = 0;
2712 * Zero part of a physical page by mapping it into memory and clearing
2713 * its contents with bzero.
2715 * off and size may not cover an area beyond a single hardware page.
2720 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2722 struct mdglobaldata *gd = mdcpu;
2725 if (*(int *) gd->gd_CMAP3)
2726 panic("pmap_zero_page: CMAP3 busy");
2727 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2728 cpu_invlpg(gd->gd_CADDR3);
2730 #if defined(I686_CPU)
2731 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2732 i686_pagezero(gd->gd_CADDR3);
2735 bzero((char *)gd->gd_CADDR3 + off, size);
2736 *(int *) gd->gd_CMAP3 = 0;
2741 * Copy the physical page from the source PA to the target PA.
2742 * This function may be called from an interrupt. No locking
2748 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2750 struct mdglobaldata *gd = mdcpu;
2753 if (*(int *) gd->gd_CMAP1)
2754 panic("pmap_copy_page: CMAP1 busy");
2755 if (*(int *) gd->gd_CMAP2)
2756 panic("pmap_copy_page: CMAP2 busy");
2758 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2759 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2761 cpu_invlpg(gd->gd_CADDR1);
2762 cpu_invlpg(gd->gd_CADDR2);
2764 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2766 *(int *) gd->gd_CMAP1 = 0;
2767 *(int *) gd->gd_CMAP2 = 0;
2772 * Copy the physical page from the source PA to the target PA.
2773 * This function may be called from an interrupt. No locking
2779 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2781 struct mdglobaldata *gd = mdcpu;
2784 if (*(int *) gd->gd_CMAP1)
2785 panic("pmap_copy_page: CMAP1 busy");
2786 if (*(int *) gd->gd_CMAP2)
2787 panic("pmap_copy_page: CMAP2 busy");
2789 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2790 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2792 cpu_invlpg(gd->gd_CADDR1);
2793 cpu_invlpg(gd->gd_CADDR2);
2795 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2796 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2799 *(int *) gd->gd_CMAP1 = 0;
2800 *(int *) gd->gd_CMAP2 = 0;
2805 * Returns true if the pmap's pv is one of the first
2806 * 16 pvs linked to from this page. This count may
2807 * be changed upwards or downwards in the future; it
2808 * is only necessary that true be returned for a small
2809 * subset of pmaps for proper page aging.
2814 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2819 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2823 lwkt_gettoken(&vm_token);
2824 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2825 if (pv->pv_pmap == pmap) {
2826 lwkt_reltoken(&vm_token);
2834 lwkt_reltoken(&vm_token);
2840 * Remove all pages from specified address space
2841 * this aids process exit speeds. Also, this code
2842 * is special cased for current process only, but
2843 * can have the more generic (and slightly slower)
2844 * mode enabled. This is much faster than pmap_remove
2845 * in the case of running down an entire address space.
2850 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2853 unsigned *pte, tpte;
2856 pmap_inval_info info;
2858 int32_t save_generation;
2860 lp = curthread->td_lwp;
2861 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2866 lwkt_gettoken(&vm_token);
2867 pmap_inval_init(&info);
2868 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2869 if (pv->pv_va >= eva || pv->pv_va < sva) {
2870 npv = TAILQ_NEXT(pv, pv_plist);
2874 KKASSERT(pmap == pv->pv_pmap);
2877 pte = (unsigned *)vtopte(pv->pv_va);
2879 pte = pmap_pte_quick(pmap, pv->pv_va);
2881 pmap_inval_interlock(&info, pmap, pv->pv_va);
2884 * We cannot remove wired pages from a process' mapping
2888 pmap_inval_deinterlock(&info, pmap);
2889 npv = TAILQ_NEXT(pv, pv_plist);
2893 tpte = loadandclear(pte);
2894 pmap_inval_deinterlock(&info, pmap);
2896 m = PHYS_TO_VM_PAGE(tpte);
2897 test_m_maps_pv(m, pv);
2899 KASSERT(m < &vm_page_array[vm_page_array_size],
2900 ("pmap_remove_pages: bad tpte %x", tpte));
2902 KKASSERT(pmap->pm_stats.resident_count > 0);
2903 --pmap->pm_stats.resident_count;
2906 * Update the vm_page_t clean and reference bits.
2912 npv = TAILQ_NEXT(pv, pv_plist);
2914 KKASSERT(pv->pv_m == m);
2915 KKASSERT(pv->pv_pmap == pmap);
2917 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2918 save_generation = ++pmap->pm_generation;
2920 m->md.pv_list_count--;
2921 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2922 if (TAILQ_EMPTY(&m->md.pv_list))
2923 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2925 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2929 * Restart the scan if we blocked during the unuse or free
2930 * calls and other removals were made.
2932 if (save_generation != pmap->pm_generation) {
2933 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2934 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2937 pmap_inval_done(&info);
2938 lwkt_reltoken(&vm_token);
2942 * pmap_testbit tests bits in pte's
2943 * note that the testbit/clearbit routines are inline,
2944 * and a lot of things compile-time evaluate.
2946 * The caller must hold vm_token.
2949 pmap_testbit(vm_page_t m, int bit)
2954 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2957 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2961 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2963 * if the bit being tested is the modified bit, then
2964 * mark clean_map and ptes as never
2967 if (bit & (PG_A|PG_M)) {
2968 if (!pmap_track_modified(pv->pv_va))
2972 #if defined(PMAP_DIAGNOSTIC)
2974 kprintf("Null pmap (tb) at va: %p\n",
2979 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2990 * This routine is used to modify bits in ptes
2992 * The caller must hold vm_token.
2994 static __inline void
2995 pmap_clearbit(vm_page_t m, int bit)
2997 struct pmap_inval_info info;
3002 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3005 pmap_inval_init(&info);
3008 * Loop over all current mappings setting/clearing as appropos If
3009 * setting RO do we need to clear the VAC?
3011 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3013 * don't write protect pager mappings
3016 if (!pmap_track_modified(pv->pv_va))
3020 #if defined(PMAP_DIAGNOSTIC)
3022 kprintf("Null pmap (cb) at va: %p\n",
3029 * Careful here. We can use a locked bus instruction to
3030 * clear PG_A or PG_M safely but we need to synchronize
3031 * with the target cpus when we mess with PG_RW.
3033 * We do not have to force synchronization when clearing
3034 * PG_M even for PTEs generated via virtual memory maps,
3035 * because the virtual kernel will invalidate the pmap
3036 * entry when/if it needs to resynchronize the Modify bit.
3039 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3040 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3047 atomic_clear_int(pte, PG_M|PG_RW);
3050 * The cpu may be trying to set PG_M
3051 * simultaniously with our clearing
3054 if (!atomic_cmpset_int(pte, pbits,
3058 } else if (bit == PG_M) {
3060 * We could also clear PG_RW here to force
3061 * a fault on write to redetect PG_M for
3062 * virtual kernels, but it isn't necessary
3063 * since virtual kernels invalidate the pte
3064 * when they clear the VPTE_M bit in their
3065 * virtual page tables.
3067 atomic_clear_int(pte, PG_M);
3069 atomic_clear_int(pte, bit);
3073 pmap_inval_deinterlock(&info, pv->pv_pmap);
3075 pmap_inval_done(&info);
3079 * Lower the permission for all mappings to a given page.
3084 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3086 if ((prot & VM_PROT_WRITE) == 0) {
3087 lwkt_gettoken(&vm_token);
3088 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3089 pmap_clearbit(m, PG_RW);
3090 vm_page_flag_clear(m, PG_WRITEABLE);
3094 lwkt_reltoken(&vm_token);
3099 * Return the physical address given a physical page index.
3104 pmap_phys_address(vm_pindex_t ppn)
3106 return (i386_ptob(ppn));
3110 * Return a count of reference bits for a page, clearing those bits.
3111 * It is not necessary for every reference bit to be cleared, but it
3112 * is necessary that 0 only be returned when there are truly no
3113 * reference bits set.
3115 * XXX: The exact number of bits to check and clear is a matter that
3116 * should be tested and standardized at some point in the future for
3117 * optimal aging of shared pages.
3122 pmap_ts_referenced(vm_page_t m)
3124 pv_entry_t pv, pvf, pvn;
3128 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3132 lwkt_gettoken(&vm_token);
3134 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3139 pvn = TAILQ_NEXT(pv, pv_list);
3142 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3143 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3146 if (!pmap_track_modified(pv->pv_va))
3149 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3151 if (pte && (*pte & PG_A)) {
3153 atomic_clear_int(pte, PG_A);
3155 atomic_clear_int_nonlocked(pte, PG_A);
3162 } while ((pv = pvn) != NULL && pv != pvf);
3165 lwkt_reltoken(&vm_token);
3172 * Return whether or not the specified physical page was modified
3173 * in any physical maps.
3178 pmap_is_modified(vm_page_t m)
3182 lwkt_gettoken(&vm_token);
3183 res = pmap_testbit(m, PG_M);
3184 lwkt_reltoken(&vm_token);
3189 * Clear the modify bits on the specified physical page.
3194 pmap_clear_modify(vm_page_t m)
3196 lwkt_gettoken(&vm_token);
3197 pmap_clearbit(m, PG_M);
3198 lwkt_reltoken(&vm_token);
3202 * Clear the reference bit on the specified physical page.
3207 pmap_clear_reference(vm_page_t m)
3209 lwkt_gettoken(&vm_token);
3210 pmap_clearbit(m, PG_A);
3211 lwkt_reltoken(&vm_token);
3215 * Miscellaneous support routines follow
3217 * Called from the low level boot code only.
3220 i386_protection_init(void)
3224 kp = protection_codes;
3225 for (prot = 0; prot < 8; prot++) {
3227 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3229 * Read access is also 0. There isn't any execute bit,
3230 * so just make it readable.
3232 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3233 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3234 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3237 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3238 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3239 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3240 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3248 * Map a set of physical memory pages into the kernel virtual
3249 * address space. Return a pointer to where it is mapped. This
3250 * routine is intended to be used for mapping device memory,
3253 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3259 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3261 vm_offset_t va, tmpva, offset;
3264 offset = pa & PAGE_MASK;
3265 size = roundup(offset + size, PAGE_SIZE);
3267 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3269 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3272 for (tmpva = va; size > 0;) {
3273 pte = (unsigned *)vtopte(tmpva);
3274 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3282 return ((void *)(va + offset));
3289 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3291 vm_offset_t base, offset;
3293 base = va & PG_FRAME;
3294 offset = va & PAGE_MASK;
3295 size = roundup(offset + size, PAGE_SIZE);
3296 pmap_qremove(va, size >> PAGE_SHIFT);
3297 kmem_free(&kernel_map, base, size);
3301 * Perform the pmap work for mincore
3303 * The caller must hold vm_token if the caller wishes a stable result,
3304 * and even in that case some bits can change due to third party accesses
3310 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3312 unsigned *ptep, pte;
3316 lwkt_gettoken(&vm_token);
3317 ptep = pmap_pte(pmap, addr);
3319 if (ptep && (pte = *ptep) != 0) {
3322 val = MINCORE_INCORE;
3323 if ((pte & PG_MANAGED) == 0)
3326 pa = pte & PG_FRAME;
3328 m = PHYS_TO_VM_PAGE(pa);
3334 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3335 } else if (m->dirty || pmap_is_modified(m)) {
3337 * Modified by someone else
3339 val |= MINCORE_MODIFIED_OTHER;
3346 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3347 } else if ((m->flags & PG_REFERENCED) ||
3348 pmap_ts_referenced(m)) {
3350 * Referenced by someone else
3352 val |= MINCORE_REFERENCED_OTHER;
3353 vm_page_flag_set(m, PG_REFERENCED);
3357 lwkt_reltoken(&vm_token);
3362 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3363 * vmspace will be ref'd and the old one will be deref'd.
3365 * cr3 will be reloaded if any lwp is the current lwp.
3367 * Only called with new VM spaces.
3368 * The process must have only a single thread.
3369 * No other requirements.
3372 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3374 struct vmspace *oldvm;
3378 oldvm = p->p_vmspace;
3379 if (oldvm != newvm) {
3380 p->p_vmspace = newvm;
3381 KKASSERT(p->p_nthreads == 1);
3382 lp = RB_ROOT(&p->p_lwp_tree);
3383 pmap_setlwpvm(lp, newvm);
3385 sysref_get(&newvm->vm_sysref);
3386 sysref_put(&oldvm->vm_sysref);
3393 * Set the vmspace for a LWP. The vmspace is almost universally set the
3394 * same as the process vmspace, but virtual kernels need to swap out contexts
3395 * on a per-lwp basis.
3397 * Always called with a lp under the caller's direct control, either
3398 * unscheduled or the current lwp.
3403 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3405 struct vmspace *oldvm;
3409 oldvm = lp->lwp_vmspace;
3411 if (oldvm != newvm) {
3412 lp->lwp_vmspace = newvm;
3413 if (curthread->td_lwp == lp) {
3414 pmap = vmspace_pmap(newvm);
3416 atomic_set_int(&pmap->pm_active, mycpu->gd_cpumask);
3417 if (pmap->pm_active & CPUMASK_LOCK)
3418 pmap_interlock_wait(newvm);
3420 pmap->pm_active |= 1;
3422 #if defined(SWTCH_OPTIM_STATS)
3425 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3426 load_cr3(curthread->td_pcb->pcb_cr3);
3427 pmap = vmspace_pmap(oldvm);
3429 atomic_clear_int(&pmap->pm_active, mycpu->gd_cpumask);
3431 pmap->pm_active &= ~1;
3440 * Called when switching to a locked pmap, used to interlock against pmaps
3441 * undergoing modifications to prevent us from activating the MMU for the
3442 * target pmap until all such modifications have completed. We have to do
3443 * this because the thread making the modifications has already set up its
3444 * SMP synchronization mask.
3449 pmap_interlock_wait(struct vmspace *vm)
3451 struct pmap *pmap = &vm->vm_pmap;
3453 if (pmap->pm_active & CPUMASK_LOCK) {
3454 while (pmap->pm_active & CPUMASK_LOCK) {
3457 lwkt_process_ipiq();
3465 * Return a page-directory alignment hint for device mappings which will
3466 * allow the use of super-pages for the mapping.
3471 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3474 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3478 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3483 * Return whether the PGE flag is supported globally.
3488 pmap_get_pgeflag(void)