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
105 #define PMAP_PVLIMIT 1400000 /* i386 kvm problems */
108 #if defined(DIAGNOSTIC)
109 #define PMAP_DIAGNOSTIC
114 #if !defined(PMAP_DIAGNOSTIC)
115 #define PMAP_INLINE __inline
121 * Get PDEs and PTEs for user/kernel address space
123 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
124 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
126 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
127 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
128 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
129 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
130 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
133 * Given a map and a machine independent protection code,
134 * convert to a vax protection code.
136 #define pte_prot(m, p) \
137 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
138 static int protection_codes[8];
140 struct pmap kernel_pmap;
141 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
143 vm_paddr_t avail_start; /* PA of first available physical page */
144 vm_paddr_t avail_end; /* PA of last available physical page */
145 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
146 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
147 vm_offset_t virtual2_start;
148 vm_offset_t virtual2_end;
149 vm_offset_t KvaStart; /* VA start of KVA space */
150 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
151 vm_offset_t KvaSize; /* max size of kernel virtual address space */
152 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
153 static int pgeflag; /* PG_G or-in */
154 static int pseflag; /* PG_PS or-in */
156 static vm_object_t kptobj;
159 vm_offset_t kernel_vm_end;
162 * Data for the pv entry allocation mechanism
164 static vm_zone_t pvzone;
165 static struct vm_zone pvzone_store;
166 static struct vm_object pvzone_obj;
167 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
168 static int pmap_pagedaemon_waken = 0;
169 static struct pv_entry *pvinit;
172 * Considering all the issues I'm having with pmap caching, if breakage
173 * continues to occur, and for debugging, I've added a sysctl that will
174 * just do an unconditional invltlb.
176 static int dreadful_invltlb;
178 SYSCTL_INT(_vm, OID_AUTO, dreadful_invltlb,
179 CTLFLAG_RW, &dreadful_invltlb, 0, "Debugging sysctl to force invltlb on pmap operations");
182 * All those kernel PT submaps that BSD is so fond of
184 pt_entry_t *CMAP1 = 0, *ptmmap;
185 caddr_t CADDR1 = 0, ptvmmap = 0;
186 static pt_entry_t *msgbufmap;
187 struct msgbuf *msgbufp=0;
192 static pt_entry_t *pt_crashdumpmap;
193 static caddr_t crashdumpmap;
195 extern pt_entry_t *SMPpt;
197 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
198 static unsigned * get_ptbase (pmap_t pmap);
199 static pv_entry_t get_pv_entry (void);
200 static void i386_protection_init (void);
201 static __inline void pmap_clearbit (vm_page_t m, int bit);
203 static void pmap_remove_all (vm_page_t m);
204 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
205 vm_offset_t sva, pmap_inval_info_t info);
206 static void pmap_remove_page (struct pmap *pmap,
207 vm_offset_t va, pmap_inval_info_t info);
208 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
209 vm_offset_t va, pmap_inval_info_t info);
210 static boolean_t pmap_testbit (vm_page_t m, int bit);
211 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
212 vm_page_t mpte, vm_page_t m);
214 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
216 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
217 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
218 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
219 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
220 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
221 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
223 static unsigned pdir4mb;
226 * Move the kernel virtual free pointer to the next
227 * 4MB. This is used to help improve performance
228 * by using a large (4MB) page for much of the kernel
229 * (.text, .data, .bss)
233 pmap_kmem_choose(vm_offset_t addr)
235 vm_offset_t newaddr = addr;
237 if (cpu_feature & CPUID_PSE) {
238 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
245 * This function returns a pointer to the pte entry in the pmap and has
246 * the side effect of potentially retaining a cached mapping of the pmap.
248 * The caller must hold vm_token and the returned value is only valid
249 * until the caller blocks or releases the token.
253 pmap_pte(pmap_t pmap, vm_offset_t va)
257 ASSERT_LWKT_TOKEN_HELD(&vm_token);
259 pdeaddr = (unsigned *) pmap_pde(pmap, va);
260 if (*pdeaddr & PG_PS)
263 return get_ptbase(pmap) + i386_btop(va);
269 * pmap_pte using the kernel_pmap
271 * Used for debugging, no requirements.
274 pmap_kernel_pte(vm_offset_t va)
278 pdeaddr = (unsigned *) pmap_pde(&kernel_pmap, va);
279 if (*pdeaddr & PG_PS)
282 return (unsigned *)vtopte(va);
289 * Super fast pmap_pte routine best used when scanning the pv lists.
290 * This eliminates many course-grained invltlb calls. Note that many of
291 * the pv list scans are across different pmaps and it is very wasteful
292 * to do an entire invltlb when checking a single mapping.
294 * Should only be called while in a critical section.
296 * The caller must hold vm_token and the returned value is only valid
297 * until the caller blocks or releases the token.
301 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
303 struct mdglobaldata *gd = mdcpu;
306 ASSERT_LWKT_TOKEN_HELD(&vm_token);
307 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
308 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
309 unsigned index = i386_btop(va);
310 /* are we current address space or kernel? */
311 if ((pmap == &kernel_pmap) ||
312 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
313 return (unsigned *) PTmap + index;
315 newpf = pde & PG_FRAME;
316 if (((*(unsigned *)gd->gd_PMAP1) & PG_FRAME) != newpf) {
317 *(unsigned *)gd->gd_PMAP1 = newpf | PG_RW | PG_V;
318 cpu_invlpg(gd->gd_PADDR1);
320 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
327 * Bootstrap the system enough to run with virtual memory.
329 * On the i386 this is called after mapping has already been enabled
330 * and just syncs the pmap module with what has already been done.
331 * [We can't call it easily with mapping off since the kernel is not
332 * mapped with PA == VA, hence we would have to relocate every address
333 * from the linked base (virtual) address "KERNBASE" to the actual
334 * (physical) address starting relative to 0]
337 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
341 struct mdglobaldata *gd;
345 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
346 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
347 KvaEnd = KvaStart + KvaSize;
349 avail_start = firstaddr;
352 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
353 * too large. It should instead be correctly calculated in locore.s and
354 * not based on 'first' (which is a physical address, not a virtual
355 * address, for the start of unused physical memory). The kernel
356 * page tables are NOT double mapped and thus should not be included
357 * in this calculation.
359 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
360 virtual_start = pmap_kmem_choose(virtual_start);
361 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
364 * Initialize protection array.
366 i386_protection_init();
369 * The kernel's pmap is statically allocated so we don't have to use
370 * pmap_create, which is unlikely to work correctly at this part of
371 * the boot sequence (XXX and which no longer exists).
373 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
374 kernel_pmap.pm_count = 1;
375 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
376 TAILQ_INIT(&kernel_pmap.pm_pvlist);
380 * Reserve some special page table entries/VA space for temporary
383 #define SYSMAP(c, p, v, n) \
384 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
387 pte = (pt_entry_t *) pmap_kernel_pte(va);
390 * CMAP1/CMAP2 are used for zeroing and copying pages.
392 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
397 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
400 * ptvmmap is used for reading arbitrary physical pages via
403 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
406 * msgbufp is used to map the system message buffer.
407 * XXX msgbufmap is not used.
409 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
410 atop(round_page(MSGBUF_SIZE)))
415 for (i = 0; i < NKPT; i++)
419 * PG_G is terribly broken on SMP because we IPI invltlb's in some
420 * cases rather then invl1pg. Actually, I don't even know why it
421 * works under UP because self-referential page table mappings
426 if (cpu_feature & CPUID_PGE)
431 * Initialize the 4MB page size flag
435 * The 4MB page version of the initial
436 * kernel page mapping.
440 #if !defined(DISABLE_PSE)
441 if (cpu_feature & CPUID_PSE) {
444 * Note that we have enabled PSE mode
447 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
448 ptditmp &= ~(NBPDR - 1);
449 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
454 * Enable the PSE mode. If we are SMP we can't do this
455 * now because the APs will not be able to use it when
458 load_cr4(rcr4() | CR4_PSE);
461 * We can do the mapping here for the single processor
462 * case. We simply ignore the old page table page from
466 * For SMP, we still need 4K pages to bootstrap APs,
467 * PSE will be enabled as soon as all APs are up.
469 PTD[KPTDI] = (pd_entry_t)ptditmp;
470 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
477 * We need to finish setting up the globaldata page for the BSP.
478 * locore has already populated the page table for the mdglobaldata
481 pg = MDGLOBALDATA_BASEALLOC_PAGES;
482 gd = &CPU_prvspace[0].mdglobaldata;
483 gd->gd_CMAP1 = &SMPpt[pg + 0];
484 gd->gd_CMAP2 = &SMPpt[pg + 1];
485 gd->gd_CMAP3 = &SMPpt[pg + 2];
486 gd->gd_PMAP1 = &SMPpt[pg + 3];
487 gd->gd_GDMAP1 = &PTD[APTDPTDI];
488 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
489 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
490 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
491 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
492 gd->gd_GDADDR1= (unsigned *)VADDR(APTDPTDI, 0);
499 * Set 4mb pdir for mp startup
504 if (pseflag && (cpu_feature & CPUID_PSE)) {
505 load_cr4(rcr4() | CR4_PSE);
506 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
507 kernel_pmap.pm_pdir[KPTDI] =
508 PTD[KPTDI] = (pd_entry_t)pdir4mb;
516 * Initialize the pmap module, called by vm_init()
518 * Called from the low level boot code only.
527 * object for kernel page table pages
529 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
532 * Allocate memory for random pmap data structures. Includes the
536 for(i = 0; i < vm_page_array_size; i++) {
539 m = &vm_page_array[i];
540 TAILQ_INIT(&m->md.pv_list);
541 m->md.pv_list_count = 0;
545 * init the pv free list
547 initial_pvs = vm_page_array_size;
548 if (initial_pvs < MINPV)
550 pvzone = &pvzone_store;
551 pvinit = (void *)kmem_alloc(&kernel_map,
552 initial_pvs * sizeof (struct pv_entry));
553 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
554 pvinit, initial_pvs);
557 * Now it is safe to enable pv_table recording.
559 pmap_initialized = TRUE;
563 * Initialize the address space (zone) for the pv_entries. Set a
564 * high water mark so that the system can recover from excessive
565 * numbers of pv entries.
567 * Called from the low level boot code only.
572 int shpgperproc = PMAP_SHPGPERPROC;
575 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
576 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
580 * Horrible hack for systems with a lot of memory running i386.
581 * the calculated pv_entry_max can wind up eating a ton of KVM
582 * so put a cap on the number of entries if the user did not
583 * change any of the values. This saves about 44MB of KVM on
584 * boxes with 3+GB of ram.
586 * On the flip side, this makes it more likely that some setups
587 * will run out of pv entries. Those sysads will have to bump
588 * the limit up with vm.pamp.pv_entries or vm.pmap.shpgperproc.
590 if (shpgperproc == PMAP_SHPGPERPROC) {
591 if (pv_entry_max > PMAP_PVLIMIT)
592 pv_entry_max = PMAP_PVLIMIT;
595 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
596 pv_entry_high_water = 9 * (pv_entry_max / 10);
599 * Subtract out pages already installed in the zone (hack)
601 entry_max = pv_entry_max - vm_page_array_size;
605 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
609 /***************************************************
610 * Low level helper routines.....
611 ***************************************************/
616 test_m_maps_pv(vm_page_t m, pv_entry_t pv)
622 KKASSERT(pv->pv_m == m);
624 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
631 panic("test_m_maps_pv: failed m %p pv %p\n", m, pv);
635 ptbase_assert(struct pmap *pmap)
637 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
639 /* are we current address space or kernel? */
640 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME))
642 KKASSERT(frame == (*mdcpu->gd_GDMAP1 & PG_FRAME));
647 #define test_m_maps_pv(m, pv)
648 #define ptbase_assert(pmap)
652 #if defined(PMAP_DIAGNOSTIC)
655 * This code checks for non-writeable/modified pages.
656 * This should be an invalid condition.
659 pmap_nw_modified(pt_entry_t ptea)
665 if ((pte & (PG_M|PG_RW)) == PG_M)
674 * This routine defines the region(s) of memory that should not be tested
675 * for the modified bit.
679 static PMAP_INLINE int
680 pmap_track_modified(vm_offset_t va)
682 if ((va < clean_sva) || (va >= clean_eva))
689 * Retrieve the mapped page table base for a particular pmap. Use our self
690 * mapping for the kernel_pmap or our current pmap.
692 * For foreign pmaps we use the per-cpu page table map. Since this involves
693 * installing a ptd it's actually (per-process x per-cpu). However, we
694 * still cannot depend on our mapping to survive thread switches because
695 * the process might be threaded and switching to another thread for the
696 * same process on the same cpu will allow that other thread to make its
699 * This could be a bit confusing but the jist is for something like the
700 * vkernel which uses foreign pmaps all the time this represents a pretty
701 * good cache that avoids unnecessary invltlb()s.
703 * The caller must hold vm_token and the returned value is only valid
704 * until the caller blocks or releases the token.
707 get_ptbase(pmap_t pmap)
709 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
710 struct mdglobaldata *gd = mdcpu;
712 ASSERT_LWKT_TOKEN_HELD(&vm_token);
715 * We can use PTmap if the pmap is our current address space or
716 * the kernel address space.
718 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
719 return (unsigned *) PTmap;
723 * Otherwise we use the per-cpu alternative page table map. Each
724 * cpu gets its own map. Because of this we cannot use this map
725 * from interrupts or threads which can preempt.
727 * Even if we already have the map cached we may still have to
728 * invalidate the TLB if another cpu modified a PDE in the map.
730 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
731 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
733 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
734 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
735 pmap->pm_cached |= gd->mi.gd_cpumask;
737 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
738 pmap->pm_cached |= gd->mi.gd_cpumask;
740 } else if (dreadful_invltlb) {
743 return ((unsigned *)gd->gd_GDADDR1);
749 * Extract the physical page address associated with the map/VA pair.
751 * The caller may hold vm_token if it desires non-blocking operation.
754 pmap_extract(pmap_t pmap, vm_offset_t va)
757 vm_offset_t pdirindex;
759 lwkt_gettoken(&vm_token);
760 pdirindex = va >> PDRSHIFT;
761 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
763 if ((rtval & PG_PS) != 0) {
764 rtval &= ~(NBPDR - 1);
765 rtval |= va & (NBPDR - 1);
767 pte = get_ptbase(pmap) + i386_btop(va);
768 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
773 lwkt_reltoken(&vm_token);
777 /***************************************************
778 * Low level mapping routines.....
779 ***************************************************/
782 * Map a wired VM page to a KVA, fully SMP synchronized.
784 * No requirements, non blocking.
787 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
791 pmap_inval_info info;
793 pmap_inval_init(&info);
794 npte = pa | PG_RW | PG_V | pgeflag;
795 pte = (unsigned *)vtopte(va);
796 pmap_inval_interlock(&info, &kernel_pmap, va);
798 pmap_inval_deinterlock(&info, &kernel_pmap);
799 pmap_inval_done(&info);
803 * Map a wired VM page to a KVA, synchronized on current cpu only.
805 * No requirements, non blocking.
808 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
813 npte = pa | PG_RW | PG_V | pgeflag;
814 pte = (unsigned *)vtopte(va);
816 cpu_invlpg((void *)va);
820 * Synchronize a previously entered VA on all cpus.
822 * No requirements, non blocking.
825 pmap_kenter_sync(vm_offset_t va)
827 pmap_inval_info info;
829 pmap_inval_init(&info);
830 pmap_inval_interlock(&info, &kernel_pmap, va);
831 pmap_inval_deinterlock(&info, &kernel_pmap);
832 pmap_inval_done(&info);
836 * Synchronize a previously entered VA on the current cpu only.
838 * No requirements, non blocking.
841 pmap_kenter_sync_quick(vm_offset_t va)
843 cpu_invlpg((void *)va);
847 * Remove a page from the kernel pagetables, fully SMP synchronized.
849 * No requirements, non blocking.
852 pmap_kremove(vm_offset_t va)
855 pmap_inval_info info;
857 pmap_inval_init(&info);
858 pte = (unsigned *)vtopte(va);
859 pmap_inval_interlock(&info, &kernel_pmap, va);
861 pmap_inval_deinterlock(&info, &kernel_pmap);
862 pmap_inval_done(&info);
866 * Remove a page from the kernel pagetables, synchronized on current cpu only.
868 * No requirements, non blocking.
871 pmap_kremove_quick(vm_offset_t va)
874 pte = (unsigned *)vtopte(va);
876 cpu_invlpg((void *)va);
880 * Adjust the permissions of a page in the kernel page table,
881 * synchronized on the current cpu only.
883 * No requirements, non blocking.
886 pmap_kmodify_rw(vm_offset_t va)
888 atomic_set_int(vtopte(va), PG_RW);
889 cpu_invlpg((void *)va);
893 * Adjust the permissions of a page in the kernel page table,
894 * synchronized on the current cpu only.
896 * No requirements, non blocking.
899 pmap_kmodify_nc(vm_offset_t va)
901 atomic_set_int(vtopte(va), PG_N);
902 cpu_invlpg((void *)va);
906 * Map a range of physical addresses into kernel virtual address space.
908 * No requirements, non blocking.
911 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
913 vm_offset_t sva, virt;
916 while (start < end) {
917 pmap_kenter(virt, start);
926 * Add a list of wired pages to the kva, fully SMP synchronized.
928 * No requirements, non blocking.
931 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
935 end_va = va + count * PAGE_SIZE;
937 while (va < end_va) {
940 pte = (unsigned *)vtopte(va);
941 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
942 cpu_invlpg((void *)va);
947 smp_invltlb(); /* XXX */
952 * Remove pages from KVA, fully SMP synchronized.
954 * No requirements, non blocking.
957 pmap_qremove(vm_offset_t va, int count)
961 end_va = va + count*PAGE_SIZE;
963 while (va < end_va) {
966 pte = (unsigned *)vtopte(va);
968 cpu_invlpg((void *)va);
977 * This routine works like vm_page_lookup() but also blocks as long as the
978 * page is busy. This routine does not busy the page it returns.
980 * The caller must hold vm_token.
983 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
987 ASSERT_LWKT_TOKEN_HELD(&vm_token);
989 m = vm_page_lookup(object, pindex);
990 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
996 * Create a new thread and optionally associate it with a (new) process.
997 * NOTE! the new thread's cpu may not equal the current cpu.
1000 pmap_init_thread(thread_t td)
1002 /* enforce pcb placement */
1003 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1004 td->td_savefpu = &td->td_pcb->pcb_save;
1005 td->td_sp = (char *)td->td_pcb - 16;
1009 * This routine directly affects the fork perf for a process.
1012 pmap_init_proc(struct proc *p)
1017 * Dispose the UPAGES for a process that has exited.
1018 * This routine directly impacts the exit perf of a process.
1021 pmap_dispose_proc(struct proc *p)
1023 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1026 /***************************************************
1027 * Page table page management routines.....
1028 ***************************************************/
1031 * This routine unholds page table pages, and if the hold count
1032 * drops to zero, then it decrements the wire count.
1034 * The caller must hold vm_token.
1035 * This function can block.
1038 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1041 * Wait until we can busy the page ourselves. We cannot have
1042 * any active flushes if we block.
1044 if (m->flags & PG_BUSY) {
1045 pmap_inval_flush(info);
1046 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1049 KASSERT(m->queue == PQ_NONE,
1050 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1052 if (m->hold_count == 1) {
1054 * Unmap the page table page.
1056 * NOTE: We must clear pm_cached for all cpus, including
1057 * the current one, when clearing a page directory
1061 pmap_inval_interlock(info, pmap, -1);
1062 KKASSERT(pmap->pm_pdir[m->pindex]);
1063 pmap->pm_pdir[m->pindex] = 0;
1064 pmap->pm_cached = 0;
1065 pmap_inval_deinterlock(info, pmap);
1067 KKASSERT(pmap->pm_stats.resident_count > 0);
1068 --pmap->pm_stats.resident_count;
1070 if (pmap->pm_ptphint == m)
1071 pmap->pm_ptphint = NULL;
1074 * This was our last hold, the page had better be unwired
1075 * after we decrement wire_count.
1077 * FUTURE NOTE: shared page directory page could result in
1078 * multiple wire counts.
1082 KKASSERT(m->wire_count == 0);
1083 --vmstats.v_wire_count;
1084 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1086 vm_page_free_zero(m);
1089 KKASSERT(m->hold_count > 1);
1096 * The caller must hold vm_token.
1097 * This function can block.
1099 static PMAP_INLINE int
1100 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1102 KKASSERT(m->hold_count > 0);
1103 if (m->hold_count > 1) {
1107 return _pmap_unwire_pte_hold(pmap, m, info);
1112 * After removing a (user) page table entry, this routine is used to
1113 * conditionally free the page, and manage the hold/wire counts.
1115 * The caller must hold vm_token.
1116 * This function can block regardless.
1119 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1120 pmap_inval_info_t info)
1124 if (va >= UPT_MIN_ADDRESS)
1128 ptepindex = (va >> PDRSHIFT);
1129 if (pmap->pm_ptphint &&
1130 (pmap->pm_ptphint->pindex == ptepindex)) {
1131 mpte = pmap->pm_ptphint;
1133 pmap_inval_flush(info);
1134 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1135 pmap->pm_ptphint = mpte;
1139 return pmap_unwire_pte_hold(pmap, mpte, info);
1143 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1144 * it, and IdlePTD, represents the template used to update all other pmaps.
1146 * On architectures where the kernel pmap is not integrated into the user
1147 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1148 * kernel_pmap should be used to directly access the kernel_pmap.
1153 pmap_pinit0(struct pmap *pmap)
1156 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1157 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1159 pmap->pm_active = 0;
1160 pmap->pm_cached = 0;
1161 pmap->pm_ptphint = NULL;
1162 TAILQ_INIT(&pmap->pm_pvlist);
1163 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1167 * Initialize a preallocated and zeroed pmap structure,
1168 * such as one in a vmspace structure.
1173 pmap_pinit(struct pmap *pmap)
1178 * No need to allocate page table space yet but we do need a valid
1179 * page directory table.
1181 if (pmap->pm_pdir == NULL) {
1183 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1187 * Allocate an object for the ptes
1189 if (pmap->pm_pteobj == NULL)
1190 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1193 * Allocate the page directory page, unless we already have
1194 * one cached. If we used the cached page the wire_count will
1195 * already be set appropriately.
1197 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1198 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1199 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1200 pmap->pm_pdirm = ptdpg;
1201 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1202 ptdpg->valid = VM_PAGE_BITS_ALL;
1203 ptdpg->wire_count = 1;
1204 ++vmstats.v_wire_count;
1205 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1207 if ((ptdpg->flags & PG_ZERO) == 0)
1208 bzero(pmap->pm_pdir, PAGE_SIZE);
1211 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1214 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1216 /* install self-referential address mapping entry */
1217 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1218 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1221 pmap->pm_active = 0;
1222 pmap->pm_cached = 0;
1223 pmap->pm_ptphint = NULL;
1224 TAILQ_INIT(&pmap->pm_pvlist);
1225 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1226 pmap->pm_stats.resident_count = 1;
1230 * Clean up a pmap structure so it can be physically freed. This routine
1231 * is called by the vmspace dtor function. A great deal of pmap data is
1232 * left passively mapped to improve vmspace management so we have a bit
1233 * of cleanup work to do here.
1238 pmap_puninit(pmap_t pmap)
1242 KKASSERT(pmap->pm_active == 0);
1243 lwkt_gettoken(&vm_token);
1244 if ((p = pmap->pm_pdirm) != NULL) {
1245 KKASSERT(pmap->pm_pdir != NULL);
1246 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1248 vmstats.v_wire_count--;
1249 KKASSERT((p->flags & PG_BUSY) == 0);
1251 vm_page_free_zero(p);
1252 pmap->pm_pdirm = NULL;
1254 lwkt_reltoken(&vm_token);
1255 if (pmap->pm_pdir) {
1256 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1257 pmap->pm_pdir = NULL;
1259 if (pmap->pm_pteobj) {
1260 vm_object_deallocate(pmap->pm_pteobj);
1261 pmap->pm_pteobj = NULL;
1266 * Wire in kernel global address entries. To avoid a race condition
1267 * between pmap initialization and pmap_growkernel, this procedure
1268 * adds the pmap to the master list (which growkernel scans to update),
1269 * then copies the template.
1274 pmap_pinit2(struct pmap *pmap)
1277 lwkt_gettoken(&vm_token);
1278 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1279 /* XXX copies current process, does not fill in MPPTDI */
1280 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1281 lwkt_reltoken(&vm_token);
1286 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1287 * 0 on failure (if the procedure had to sleep).
1289 * When asked to remove the page directory page itself, we actually just
1290 * leave it cached so we do not have to incur the SMP inval overhead of
1291 * removing the kernel mapping. pmap_puninit() will take care of it.
1293 * The caller must hold vm_token.
1294 * This function can block regardless.
1297 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1299 unsigned *pde = (unsigned *) pmap->pm_pdir;
1302 * This code optimizes the case of freeing non-busy
1303 * page-table pages. Those pages are zero now, and
1304 * might as well be placed directly into the zero queue.
1306 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1312 * Remove the page table page from the processes address space.
1314 KKASSERT(pmap->pm_stats.resident_count > 0);
1315 KKASSERT(pde[p->pindex]);
1317 --pmap->pm_stats.resident_count;
1318 pmap->pm_cached = 0;
1320 if (p->hold_count) {
1321 panic("pmap_release: freeing held page table page");
1323 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1324 pmap->pm_ptphint = NULL;
1327 * We leave the page directory page cached, wired, and mapped in
1328 * the pmap until the dtor function (pmap_puninit()) gets called.
1329 * However, still clean it up so we can set PG_ZERO.
1331 * The pmap has already been removed from the pmap_list in the
1334 if (p->pindex == PTDPTDI) {
1335 bzero(pde + KPTDI, nkpt * PTESIZE);
1336 bzero(pde + MPPTDI, (NPDEPG - MPPTDI) * PTESIZE);
1337 vm_page_flag_set(p, PG_ZERO);
1341 vmstats.v_wire_count--;
1342 vm_page_free_zero(p);
1348 * This routine is called if the page table page is not mapped correctly.
1350 * The caller must hold vm_token.
1353 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1355 vm_offset_t pteva, ptepa;
1359 * Find or fabricate a new pagetable page
1361 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1362 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1364 KASSERT(m->queue == PQ_NONE,
1365 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1368 * Increment the hold count for the page we will be returning to
1374 * It is possible that someone else got in and mapped by the page
1375 * directory page while we were blocked, if so just unbusy and
1376 * return the held page.
1378 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1379 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1384 if (m->wire_count == 0)
1385 vmstats.v_wire_count++;
1390 * Map the pagetable page into the process address space, if
1391 * it isn't already there.
1393 * NOTE: For safety clear pm_cached for all cpus including the
1394 * current one when adding a PDE to the map.
1396 ++pmap->pm_stats.resident_count;
1398 ptepa = VM_PAGE_TO_PHYS(m);
1399 pmap->pm_pdir[ptepindex] =
1400 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1401 pmap->pm_cached = 0;
1404 * Set the page table hint
1406 pmap->pm_ptphint = m;
1409 * Try to use the new mapping, but if we cannot, then
1410 * do it with the routine that maps the page explicitly.
1412 if ((m->flags & PG_ZERO) == 0) {
1413 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1414 (((unsigned) PTDpde) & PG_FRAME)) {
1415 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1416 bzero((caddr_t) pteva, PAGE_SIZE);
1418 pmap_zero_page(ptepa);
1423 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1427 m->valid = VM_PAGE_BITS_ALL;
1428 vm_page_flag_clear(m, PG_ZERO);
1429 vm_page_flag_set(m, PG_MAPPED);
1436 * Allocate a page table entry for a va.
1438 * The caller must hold vm_token.
1441 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1448 * Calculate pagetable page index
1450 ptepindex = va >> PDRSHIFT;
1453 * Get the page directory entry
1455 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1458 * This supports switching from a 4MB page to a
1461 if (ptepa & PG_PS) {
1462 pmap->pm_pdir[ptepindex] = 0;
1469 * If the page table page is mapped, we just increment the
1470 * hold count, and activate it.
1474 * In order to get the page table page, try the
1477 if (pmap->pm_ptphint &&
1478 (pmap->pm_ptphint->pindex == ptepindex)) {
1479 m = pmap->pm_ptphint;
1481 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1482 pmap->pm_ptphint = m;
1488 * Here if the pte page isn't mapped, or if it has been deallocated.
1490 return _pmap_allocpte(pmap, ptepindex);
1494 /***************************************************
1495 * Pmap allocation/deallocation routines.
1496 ***************************************************/
1499 * Release any resources held by the given physical map.
1500 * Called when a pmap initialized by pmap_pinit is being released.
1501 * Should only be called if the map contains no valid mappings.
1505 static int pmap_release_callback(struct vm_page *p, void *data);
1508 pmap_release(struct pmap *pmap)
1510 vm_object_t object = pmap->pm_pteobj;
1511 struct rb_vm_page_scan_info info;
1513 KASSERT(pmap->pm_active == 0,
1514 ("pmap still active! %08x", pmap->pm_active));
1515 #if defined(DIAGNOSTIC)
1516 if (object->ref_count != 1)
1517 panic("pmap_release: pteobj reference count != 1");
1521 info.object = object;
1523 lwkt_gettoken(&vm_token);
1524 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1531 info.limit = object->generation;
1533 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1534 pmap_release_callback, &info);
1535 if (info.error == 0 && info.mpte) {
1536 if (!pmap_release_free_page(pmap, info.mpte))
1540 } while (info.error);
1541 pmap->pm_cached = 0;
1542 lwkt_reltoken(&vm_token);
1546 * The caller must hold vm_token.
1549 pmap_release_callback(struct vm_page *p, void *data)
1551 struct rb_vm_page_scan_info *info = data;
1553 if (p->pindex == PTDPTDI) {
1557 if (!pmap_release_free_page(info->pmap, p)) {
1561 if (info->object->generation != info->limit) {
1569 * Grow the number of kernel page table entries, if needed.
1574 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1576 vm_offset_t addr = kend;
1578 vm_offset_t ptppaddr;
1583 lwkt_gettoken(&vm_token);
1584 if (kernel_vm_end == 0) {
1585 kernel_vm_end = KERNBASE;
1587 while (pdir_pde(PTD, kernel_vm_end)) {
1588 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1589 ~(PAGE_SIZE * NPTEPG - 1);
1593 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1594 while (kernel_vm_end < addr) {
1595 if (pdir_pde(PTD, kernel_vm_end)) {
1596 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1597 ~(PAGE_SIZE * NPTEPG - 1);
1602 * This index is bogus, but out of the way
1604 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_NORMAL |
1606 VM_ALLOC_INTERRUPT);
1608 panic("pmap_growkernel: no memory to grow kernel");
1611 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1612 pmap_zero_page(ptppaddr);
1613 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1614 pdir_pde(PTD, kernel_vm_end) = newpdir;
1615 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1619 * This update must be interlocked with pmap_pinit2.
1621 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1622 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1624 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1625 ~(PAGE_SIZE * NPTEPG - 1);
1627 lwkt_reltoken(&vm_token);
1632 * Retire the given physical map from service.
1634 * Should only be called if the map contains no valid mappings.
1639 pmap_destroy(pmap_t pmap)
1644 lwkt_gettoken(&vm_token);
1645 if (--pmap->pm_count == 0) {
1647 panic("destroying a pmap is not yet implemented");
1649 lwkt_reltoken(&vm_token);
1653 * Add a reference to the specified pmap.
1658 pmap_reference(pmap_t pmap)
1661 lwkt_gettoken(&vm_token);
1663 lwkt_reltoken(&vm_token);
1667 /***************************************************
1668 * page management routines.
1669 ***************************************************/
1672 * free the pv_entry back to the free list. This function may be
1673 * called from an interrupt.
1675 * The caller must hold vm_token.
1677 static PMAP_INLINE void
1678 free_pv_entry(pv_entry_t pv)
1681 KKASSERT(pv->pv_m != NULL);
1689 * get a new pv_entry, allocating a block from the system
1690 * when needed. This function may be called from an interrupt.
1692 * The caller must hold vm_token.
1698 if (pv_entry_high_water &&
1699 (pv_entry_count > pv_entry_high_water) &&
1700 (pmap_pagedaemon_waken == 0)) {
1701 pmap_pagedaemon_waken = 1;
1702 wakeup (&vm_pages_needed);
1704 return zalloc(pvzone);
1708 * This routine is very drastic, but can save the system
1718 static int warningdone=0;
1720 if (pmap_pagedaemon_waken == 0)
1722 lwkt_gettoken(&vm_token);
1723 pmap_pagedaemon_waken = 0;
1725 if (warningdone < 5) {
1726 kprintf("pmap_collect: collecting pv entries -- "
1727 "suggest increasing PMAP_SHPGPERPROC\n");
1731 for(i = 0; i < vm_page_array_size; i++) {
1732 m = &vm_page_array[i];
1733 if (m->wire_count || m->hold_count || m->busy ||
1734 (m->flags & PG_BUSY)) {
1739 lwkt_reltoken(&vm_token);
1744 * If it is the first entry on the list, it is actually
1745 * in the header and we must copy the following entry up
1746 * to the header. Otherwise we must search the list for
1747 * the entry. In either case we free the now unused entry.
1749 * The caller must hold vm_token.
1752 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1753 vm_offset_t va, pmap_inval_info_t info)
1759 ASSERT_LWKT_TOKEN_HELD(&vm_token);
1760 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1761 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1762 if (pmap == pv->pv_pmap && va == pv->pv_va)
1766 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1768 KKASSERT(pv->pv_pmap == pmap);
1770 if (va == pv->pv_va)
1777 test_m_maps_pv(m, pv);
1778 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1779 m->md.pv_list_count--;
1780 m->object->agg_pv_list_count--;
1781 if (TAILQ_EMPTY(&m->md.pv_list))
1782 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1783 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1784 ++pmap->pm_generation;
1785 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1792 * Create a pv entry for page at pa for (pmap, va).
1794 * The caller must hold vm_token.
1797 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1802 pv = get_pv_entry();
1804 KKASSERT(pv->pv_m == NULL);
1811 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1812 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1813 ++pmap->pm_generation;
1814 m->md.pv_list_count++;
1815 m->object->agg_pv_list_count++;
1821 * pmap_remove_pte: do the things to unmap a page in a process.
1823 * The caller must hold vm_token.
1825 * WARNING! As with most other pmap functions this one can block, so
1826 * callers using temporary page table mappings must reload
1830 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1831 pmap_inval_info_t info)
1836 ptbase_assert(pmap);
1837 pmap_inval_interlock(info, pmap, va);
1838 ptbase_assert(pmap);
1839 oldpte = loadandclear(ptq);
1841 pmap->pm_stats.wired_count -= 1;
1842 pmap_inval_deinterlock(info, pmap);
1845 * Machines that don't support invlpg, also don't support
1846 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1850 cpu_invlpg((void *)va);
1851 KKASSERT(pmap->pm_stats.resident_count > 0);
1852 --pmap->pm_stats.resident_count;
1853 if (oldpte & PG_MANAGED) {
1854 m = PHYS_TO_VM_PAGE(oldpte);
1855 if (oldpte & PG_M) {
1856 #if defined(PMAP_DIAGNOSTIC)
1857 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1858 kprintf("pmap_remove: modified page not "
1859 "writable: va: %p, pte: 0x%lx\n",
1860 (void *)va, (long)oldpte);
1863 if (pmap_track_modified(va))
1867 vm_page_flag_set(m, PG_REFERENCED);
1868 return pmap_remove_entry(pmap, m, va, info);
1870 return pmap_unuse_pt(pmap, va, NULL, info);
1877 * Remove a single page from a process address space.
1879 * The caller must hold vm_token.
1882 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1887 * if there is no pte for this address, just skip it!!! Otherwise
1888 * get a local va for mappings for this pmap and remove the entry.
1890 if (*pmap_pde(pmap, va) != 0) {
1891 ptq = get_ptbase(pmap) + i386_btop(va);
1893 pmap_remove_pte(pmap, ptq, va, info);
1900 * Remove the given range of addresses from the specified map.
1902 * It is assumed that the start and end are properly rounded to the page
1908 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1912 vm_offset_t ptpaddr;
1913 vm_offset_t sindex, eindex;
1914 struct pmap_inval_info info;
1919 lwkt_gettoken(&vm_token);
1920 if (pmap->pm_stats.resident_count == 0) {
1921 lwkt_reltoken(&vm_token);
1925 pmap_inval_init(&info);
1928 * special handling of removing one page. a very
1929 * common operation and easy to short circuit some
1932 if (((sva + PAGE_SIZE) == eva) &&
1933 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1934 pmap_remove_page(pmap, sva, &info);
1935 pmap_inval_done(&info);
1936 lwkt_reltoken(&vm_token);
1941 * Get a local virtual address for the mappings that are being
1944 sindex = i386_btop(sva);
1945 eindex = i386_btop(eva);
1947 for (; sindex < eindex; sindex = pdnxt) {
1951 * Calculate index for next page table.
1953 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1954 if (pmap->pm_stats.resident_count == 0)
1957 pdirindex = sindex / NPDEPG;
1958 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1959 pmap_inval_interlock(&info, pmap, -1);
1960 pmap->pm_pdir[pdirindex] = 0;
1961 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1962 pmap->pm_cached = 0;
1963 pmap_inval_deinterlock(&info, pmap);
1968 * Weed out invalid mappings. Note: we assume that the page
1969 * directory table is always allocated, and in kernel virtual.
1975 * Limit our scan to either the end of the va represented
1976 * by the current page table page, or to the end of the
1977 * range being removed.
1979 if (pdnxt > eindex) {
1984 * NOTE: pmap_remove_pte() can block and wipe the temporary
1987 for (; sindex != pdnxt; sindex++) {
1990 ptbase = get_ptbase(pmap);
1991 if (ptbase[sindex] == 0)
1993 va = i386_ptob(sindex);
1994 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1998 pmap_inval_done(&info);
1999 lwkt_reltoken(&vm_token);
2003 * Removes this physical page from all physical maps in which it resides.
2004 * Reflects back modify bits to the pager.
2009 pmap_remove_all(vm_page_t m)
2011 struct pmap_inval_info info;
2012 unsigned *pte, tpte;
2015 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2018 lwkt_gettoken(&vm_token);
2019 pmap_inval_init(&info);
2020 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2021 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2022 --pv->pv_pmap->pm_stats.resident_count;
2024 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2025 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2026 tpte = loadandclear(pte);
2028 pv->pv_pmap->pm_stats.wired_count--;
2029 pmap_inval_deinterlock(&info, pv->pv_pmap);
2031 vm_page_flag_set(m, PG_REFERENCED);
2033 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
2037 * Update the vm_page_t clean and reference bits.
2040 #if defined(PMAP_DIAGNOSTIC)
2041 if (pmap_nw_modified((pt_entry_t) tpte)) {
2042 kprintf("pmap_remove_all: modified page "
2043 "not writable: va: %p, pte: 0x%lx\n",
2044 (void *)pv->pv_va, (long)tpte);
2047 if (pmap_track_modified(pv->pv_va))
2051 KKASSERT(pv->pv_m == m);
2053 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2054 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2055 ++pv->pv_pmap->pm_generation;
2056 m->md.pv_list_count--;
2057 m->object->agg_pv_list_count--;
2058 if (TAILQ_EMPTY(&m->md.pv_list))
2059 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2060 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2063 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2064 pmap_inval_done(&info);
2065 lwkt_reltoken(&vm_token);
2069 * Set the physical protection on the specified range of this map
2075 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2078 vm_offset_t pdnxt, ptpaddr;
2079 vm_pindex_t sindex, eindex;
2080 pmap_inval_info info;
2085 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2086 pmap_remove(pmap, sva, eva);
2090 if (prot & VM_PROT_WRITE)
2093 lwkt_gettoken(&vm_token);
2094 pmap_inval_init(&info);
2096 ptbase = get_ptbase(pmap);
2098 sindex = i386_btop(sva);
2099 eindex = i386_btop(eva);
2101 for (; sindex < eindex; sindex = pdnxt) {
2104 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2106 pdirindex = sindex / NPDEPG;
2107 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
2108 pmap_inval_interlock(&info, pmap, -1);
2109 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2110 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2111 pmap_inval_deinterlock(&info, pmap);
2116 * Weed out invalid mappings. Note: we assume that the page
2117 * directory table is always allocated, and in kernel virtual.
2122 if (pdnxt > eindex) {
2126 for (; sindex != pdnxt; sindex++) {
2132 * XXX non-optimal. Note also that there can be
2133 * no pmap_inval_flush() calls until after we modify
2134 * ptbase[sindex] (or otherwise we have to do another
2135 * pmap_inval_interlock() call).
2137 pmap_inval_interlock(&info, pmap, i386_ptob(sindex));
2139 pbits = ptbase[sindex];
2142 if (pbits & PG_MANAGED) {
2145 m = PHYS_TO_VM_PAGE(pbits);
2146 vm_page_flag_set(m, PG_REFERENCED);
2150 if (pmap_track_modified(i386_ptob(sindex))) {
2152 m = PHYS_TO_VM_PAGE(pbits);
2159 if (pbits != cbits &&
2160 !atomic_cmpset_int(ptbase + sindex, pbits, cbits)) {
2163 pmap_inval_deinterlock(&info, pmap);
2166 pmap_inval_done(&info);
2167 lwkt_reltoken(&vm_token);
2171 * Insert the given physical page (p) at the specified virtual address (v)
2172 * in the target physical map with the protection requested.
2174 * If specified, the page will be wired down, meaning that the related pte
2175 * cannot be reclaimed.
2180 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2186 vm_offset_t origpte, newpte;
2188 pmap_inval_info info;
2194 #ifdef PMAP_DIAGNOSTIC
2196 panic("pmap_enter: toobig");
2197 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2198 panic("pmap_enter: invalid to pmap_enter page "
2199 "table pages (va: %p)", (void *)va);
2202 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2203 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2204 print_backtrace(-1);
2206 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2207 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2208 print_backtrace(-1);
2211 lwkt_gettoken(&vm_token);
2214 * In the case that a page table page is not
2215 * resident, we are creating it here.
2217 if (va < UPT_MIN_ADDRESS)
2218 mpte = pmap_allocpte(pmap, va);
2222 pmap_inval_init(&info);
2223 pte = pmap_pte(pmap, va);
2226 * Page Directory table entry not valid, we need a new PT page
2229 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p\n",
2230 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
2233 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2234 origpte = *(vm_offset_t *)pte;
2235 opa = origpte & PG_FRAME;
2237 if (origpte & PG_PS)
2238 panic("pmap_enter: attempted pmap_enter on 4MB page");
2241 * Mapping has not changed, must be protection or wiring change.
2243 if (origpte && (opa == pa)) {
2245 * Wiring change, just update stats. We don't worry about
2246 * wiring PT pages as they remain resident as long as there
2247 * are valid mappings in them. Hence, if a user page is wired,
2248 * the PT page will be also.
2250 if (wired && ((origpte & PG_W) == 0))
2251 pmap->pm_stats.wired_count++;
2252 else if (!wired && (origpte & PG_W))
2253 pmap->pm_stats.wired_count--;
2255 #if defined(PMAP_DIAGNOSTIC)
2256 if (pmap_nw_modified((pt_entry_t) origpte)) {
2257 kprintf("pmap_enter: modified page not "
2258 "writable: va: %p, pte: 0x%lx\n",
2259 (void *)va, (long )origpte);
2264 * Remove the extra pte reference. Note that we cannot
2265 * optimize the RO->RW case because we have adjusted the
2266 * wiring count above and may need to adjust the wiring
2273 * We might be turning off write access to the page,
2274 * so we go ahead and sense modify status.
2276 if (origpte & PG_MANAGED) {
2277 if ((origpte & PG_M) && pmap_track_modified(va)) {
2279 om = PHYS_TO_VM_PAGE(opa);
2283 KKASSERT(m->flags & PG_MAPPED);
2288 * Mapping has changed, invalidate old range and fall through to
2289 * handle validating new mapping.
2291 * Since we have a ref on the page directory page pmap_pte()
2292 * will always return non-NULL.
2294 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2295 * to get wiped. reload the ptbase. I'm not sure if it is
2296 * also possible to race another pmap_enter() but check for
2302 KKASSERT((origpte & PG_FRAME) ==
2303 (*(vm_offset_t *)pte & PG_FRAME));
2304 err = pmap_remove_pte(pmap, pte, va, &info);
2306 panic("pmap_enter: pte vanished, va: %p", (void *)va);
2307 pte = pmap_pte(pmap, va);
2308 origpte = *(vm_offset_t *)pte;
2309 opa = origpte & PG_FRAME;
2311 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2317 * Enter on the PV list if part of our managed memory. Note that we
2318 * raise IPL while manipulating pv_table since pmap_enter can be
2319 * called at interrupt time.
2321 if (pmap_initialized &&
2322 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2323 pmap_insert_entry(pmap, va, mpte, m);
2324 ptbase_assert(pmap);
2326 vm_page_flag_set(m, PG_MAPPED);
2330 * Increment counters
2332 ++pmap->pm_stats.resident_count;
2334 pmap->pm_stats.wired_count++;
2335 KKASSERT(*pte == 0);
2339 * Now validate mapping with desired protection/wiring.
2341 ptbase_assert(pmap);
2342 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2346 if (va < UPT_MIN_ADDRESS)
2348 if (pmap == &kernel_pmap)
2352 * if the mapping or permission bits are different, we need
2353 * to update the pte.
2355 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2356 pmap_inval_interlock(&info, pmap, va);
2357 ptbase_assert(pmap);
2358 KKASSERT(*pte == 0 ||
2359 (*pte & PG_FRAME) == (newpte & PG_FRAME));
2360 *pte = newpte | PG_A;
2361 pmap_inval_deinterlock(&info, pmap);
2363 vm_page_flag_set(m, PG_WRITEABLE);
2365 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2366 pmap_inval_done(&info);
2367 lwkt_reltoken(&vm_token);
2371 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2372 * This code also assumes that the pmap has no pre-existing entry for this
2375 * This code currently may only be used on user pmaps, not kernel_pmap.
2380 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2387 pmap_inval_info info;
2389 lwkt_gettoken(&vm_token);
2390 pmap_inval_init(&info);
2392 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2393 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2394 print_backtrace(-1);
2396 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2397 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2398 print_backtrace(-1);
2401 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2404 * Calculate the page table page (mpte), allocating it if necessary.
2406 * A held page table page (mpte), or NULL, is passed onto the
2407 * section following.
2409 if (va < UPT_MIN_ADDRESS) {
2411 * Calculate pagetable page index
2413 ptepindex = va >> PDRSHIFT;
2417 * Get the page directory entry
2419 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2422 * If the page table page is mapped, we just increment
2423 * the hold count, and activate it.
2427 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2428 if (pmap->pm_ptphint &&
2429 (pmap->pm_ptphint->pindex == ptepindex)) {
2430 mpte = pmap->pm_ptphint;
2432 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2433 pmap->pm_ptphint = mpte;
2438 mpte = _pmap_allocpte(pmap, ptepindex);
2440 } while (mpte == NULL);
2443 /* this code path is not yet used */
2447 * With a valid (and held) page directory page, we can just use
2448 * vtopte() to get to the pte. If the pte is already present
2449 * we do not disturb it.
2451 pte = (unsigned *)vtopte(va);
2454 pmap_unwire_pte_hold(pmap, mpte, &info);
2455 pa = VM_PAGE_TO_PHYS(m);
2456 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2457 pmap_inval_done(&info);
2458 lwkt_reltoken(&vm_token);
2463 * Enter on the PV list if part of our managed memory
2465 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2466 pmap_insert_entry(pmap, va, mpte, m);
2467 vm_page_flag_set(m, PG_MAPPED);
2471 * Increment counters
2473 ++pmap->pm_stats.resident_count;
2475 pa = VM_PAGE_TO_PHYS(m);
2478 * Now validate mapping with RO protection
2480 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2481 *pte = pa | PG_V | PG_U;
2483 *pte = pa | PG_V | PG_U | PG_MANAGED;
2484 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2485 pmap_inval_done(&info);
2486 lwkt_reltoken(&vm_token);
2490 * Make a temporary mapping for a physical address. This is only intended
2491 * to be used for panic dumps.
2496 pmap_kenter_temporary(vm_paddr_t pa, long i)
2498 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2499 return ((void *)crashdumpmap);
2502 #define MAX_INIT_PT (96)
2505 * This routine preloads the ptes for a given object into the specified pmap.
2506 * This eliminates the blast of soft faults on process startup and
2507 * immediately after an mmap.
2511 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2514 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2515 vm_object_t object, vm_pindex_t pindex,
2516 vm_size_t size, int limit)
2518 struct rb_vm_page_scan_info info;
2523 * We can't preinit if read access isn't set or there is no pmap
2526 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2530 * We can't preinit if the pmap is not the current pmap
2532 lp = curthread->td_lwp;
2533 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2536 psize = i386_btop(size);
2538 if ((object->type != OBJT_VNODE) ||
2539 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2540 (object->resident_page_count > MAX_INIT_PT))) {
2544 if (psize + pindex > object->size) {
2545 if (object->size < pindex)
2547 psize = object->size - pindex;
2554 * Use a red-black scan to traverse the requested range and load
2555 * any valid pages found into the pmap.
2557 * We cannot safely scan the object's memq unless we are in a
2558 * critical section since interrupts can remove pages from objects.
2560 info.start_pindex = pindex;
2561 info.end_pindex = pindex + psize - 1;
2568 lwkt_gettoken(&vm_token);
2569 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2570 pmap_object_init_pt_callback, &info);
2571 lwkt_reltoken(&vm_token);
2576 * The caller must hold vm_token.
2580 pmap_object_init_pt_callback(vm_page_t p, void *data)
2582 struct rb_vm_page_scan_info *info = data;
2583 vm_pindex_t rel_index;
2585 * don't allow an madvise to blow away our really
2586 * free pages allocating pv entries.
2588 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2589 vmstats.v_free_count < vmstats.v_free_reserved) {
2592 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2593 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2595 if ((p->queue - p->pc) == PQ_CACHE)
2596 vm_page_deactivate(p);
2597 rel_index = p->pindex - info->start_pindex;
2598 pmap_enter_quick(info->pmap,
2599 info->addr + i386_ptob(rel_index), p);
2606 * Return TRUE if the pmap is in shape to trivially
2607 * pre-fault the specified address.
2609 * Returns FALSE if it would be non-trivial or if a
2610 * pte is already loaded into the slot.
2615 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2620 lwkt_gettoken(&vm_token);
2621 if ((*pmap_pde(pmap, addr)) == 0) {
2624 pte = (unsigned *) vtopte(addr);
2625 ret = (*pte) ? 0 : 1;
2627 lwkt_reltoken(&vm_token);
2632 * Change the wiring attribute for a map/virtual-adderss pair. The mapping
2633 * must already exist.
2638 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2645 lwkt_gettoken(&vm_token);
2646 pte = pmap_pte(pmap, va);
2648 if (wired && !pmap_pte_w(pte))
2649 pmap->pm_stats.wired_count++;
2650 else if (!wired && pmap_pte_w(pte))
2651 pmap->pm_stats.wired_count--;
2654 * Wiring is not a hardware characteristic so there is no need to
2655 * invalidate TLB. However, in an SMP environment we must use
2656 * a locked bus cycle to update the pte (if we are not using
2657 * the pmap_inval_*() API that is)... it's ok to do this for simple
2662 atomic_set_int(pte, PG_W);
2664 atomic_clear_int(pte, PG_W);
2667 atomic_set_int_nonlocked(pte, PG_W);
2669 atomic_clear_int_nonlocked(pte, PG_W);
2671 lwkt_reltoken(&vm_token);
2675 * Copy the range specified by src_addr/len from the source map to the
2676 * range dst_addr/len in the destination map.
2678 * This routine is only advisory and need not do anything.
2683 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2684 vm_size_t len, vm_offset_t src_addr)
2690 * Zero the specified PA by mapping the page into KVM and clearing its
2696 pmap_zero_page(vm_paddr_t phys)
2698 struct mdglobaldata *gd = mdcpu;
2701 if (*(int *)gd->gd_CMAP3)
2702 panic("pmap_zero_page: CMAP3 busy");
2703 *(int *)gd->gd_CMAP3 =
2704 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2705 cpu_invlpg(gd->gd_CADDR3);
2707 #if defined(I686_CPU)
2708 if (cpu_class == CPUCLASS_686)
2709 i686_pagezero(gd->gd_CADDR3);
2712 bzero(gd->gd_CADDR3, PAGE_SIZE);
2713 *(int *) gd->gd_CMAP3 = 0;
2718 * Assert that a page is empty, panic if it isn't.
2723 pmap_page_assertzero(vm_paddr_t phys)
2725 struct mdglobaldata *gd = mdcpu;
2729 if (*(int *)gd->gd_CMAP3)
2730 panic("pmap_zero_page: CMAP3 busy");
2731 *(int *)gd->gd_CMAP3 =
2732 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2733 cpu_invlpg(gd->gd_CADDR3);
2734 for (i = 0; i < PAGE_SIZE; i += 4) {
2735 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2736 panic("pmap_page_assertzero() @ %p not zero!\n",
2737 (void *)gd->gd_CADDR3);
2740 *(int *) gd->gd_CMAP3 = 0;
2745 * Zero part of a physical page by mapping it into memory and clearing
2746 * its contents with bzero.
2748 * off and size may not cover an area beyond a single hardware page.
2753 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2755 struct mdglobaldata *gd = mdcpu;
2758 if (*(int *) gd->gd_CMAP3)
2759 panic("pmap_zero_page: CMAP3 busy");
2760 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2761 cpu_invlpg(gd->gd_CADDR3);
2763 #if defined(I686_CPU)
2764 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2765 i686_pagezero(gd->gd_CADDR3);
2768 bzero((char *)gd->gd_CADDR3 + off, size);
2769 *(int *) gd->gd_CMAP3 = 0;
2774 * Copy the physical page from the source PA to the target PA.
2775 * This function may be called from an interrupt. No locking
2781 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2783 struct mdglobaldata *gd = mdcpu;
2786 if (*(int *) gd->gd_CMAP1)
2787 panic("pmap_copy_page: CMAP1 busy");
2788 if (*(int *) gd->gd_CMAP2)
2789 panic("pmap_copy_page: CMAP2 busy");
2791 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2792 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2794 cpu_invlpg(gd->gd_CADDR1);
2795 cpu_invlpg(gd->gd_CADDR2);
2797 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2799 *(int *) gd->gd_CMAP1 = 0;
2800 *(int *) gd->gd_CMAP2 = 0;
2805 * Copy the physical page from the source PA to the target PA.
2806 * This function may be called from an interrupt. No locking
2812 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2814 struct mdglobaldata *gd = mdcpu;
2817 if (*(int *) gd->gd_CMAP1)
2818 panic("pmap_copy_page: CMAP1 busy");
2819 if (*(int *) gd->gd_CMAP2)
2820 panic("pmap_copy_page: CMAP2 busy");
2822 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2823 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2825 cpu_invlpg(gd->gd_CADDR1);
2826 cpu_invlpg(gd->gd_CADDR2);
2828 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2829 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2832 *(int *) gd->gd_CMAP1 = 0;
2833 *(int *) gd->gd_CMAP2 = 0;
2838 * Returns true if the pmap's pv is one of the first
2839 * 16 pvs linked to from this page. This count may
2840 * be changed upwards or downwards in the future; it
2841 * is only necessary that true be returned for a small
2842 * subset of pmaps for proper page aging.
2847 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2852 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2856 lwkt_gettoken(&vm_token);
2857 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2858 if (pv->pv_pmap == pmap) {
2859 lwkt_reltoken(&vm_token);
2867 lwkt_reltoken(&vm_token);
2873 * Remove all pages from specified address space
2874 * this aids process exit speeds. Also, this code
2875 * is special cased for current process only, but
2876 * can have the more generic (and slightly slower)
2877 * mode enabled. This is much faster than pmap_remove
2878 * in the case of running down an entire address space.
2883 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2886 unsigned *pte, tpte;
2889 pmap_inval_info info;
2891 int32_t save_generation;
2893 lp = curthread->td_lwp;
2894 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2899 lwkt_gettoken(&vm_token);
2900 pmap_inval_init(&info);
2901 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2902 if (pv->pv_va >= eva || pv->pv_va < sva) {
2903 npv = TAILQ_NEXT(pv, pv_plist);
2907 KKASSERT(pmap == pv->pv_pmap);
2910 pte = (unsigned *)vtopte(pv->pv_va);
2912 pte = pmap_pte_quick(pmap, pv->pv_va);
2914 pmap_inval_interlock(&info, pmap, pv->pv_va);
2917 * We cannot remove wired pages from a process' mapping
2921 pmap_inval_deinterlock(&info, pmap);
2922 npv = TAILQ_NEXT(pv, pv_plist);
2926 tpte = loadandclear(pte);
2927 pmap_inval_deinterlock(&info, pmap);
2929 m = PHYS_TO_VM_PAGE(tpte);
2930 test_m_maps_pv(m, pv);
2932 KASSERT(m < &vm_page_array[vm_page_array_size],
2933 ("pmap_remove_pages: bad tpte %x", tpte));
2935 KKASSERT(pmap->pm_stats.resident_count > 0);
2936 --pmap->pm_stats.resident_count;
2939 * Update the vm_page_t clean and reference bits.
2945 npv = TAILQ_NEXT(pv, pv_plist);
2947 KKASSERT(pv->pv_m == m);
2948 KKASSERT(pv->pv_pmap == pmap);
2950 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2951 save_generation = ++pmap->pm_generation;
2953 m->md.pv_list_count--;
2954 m->object->agg_pv_list_count--;
2955 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2956 if (TAILQ_EMPTY(&m->md.pv_list))
2957 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2959 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2963 * Restart the scan if we blocked during the unuse or free
2964 * calls and other removals were made.
2966 if (save_generation != pmap->pm_generation) {
2967 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2968 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2971 pmap_inval_done(&info);
2972 lwkt_reltoken(&vm_token);
2976 * pmap_testbit tests bits in pte's
2977 * note that the testbit/clearbit routines are inline,
2978 * and a lot of things compile-time evaluate.
2980 * The caller must hold vm_token.
2983 pmap_testbit(vm_page_t m, int bit)
2988 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2991 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2995 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2997 * if the bit being tested is the modified bit, then
2998 * mark clean_map and ptes as never
3001 if (bit & (PG_A|PG_M)) {
3002 if (!pmap_track_modified(pv->pv_va))
3006 #if defined(PMAP_DIAGNOSTIC)
3008 kprintf("Null pmap (tb) at va: %p\n",
3013 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3024 * This routine is used to modify bits in ptes
3026 * The caller must hold vm_token.
3028 static __inline void
3029 pmap_clearbit(vm_page_t m, int bit)
3031 struct pmap_inval_info info;
3036 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3039 pmap_inval_init(&info);
3042 * Loop over all current mappings setting/clearing as appropos If
3043 * setting RO do we need to clear the VAC?
3045 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3047 * don't write protect pager mappings
3050 if (!pmap_track_modified(pv->pv_va))
3054 #if defined(PMAP_DIAGNOSTIC)
3056 kprintf("Null pmap (cb) at va: %p\n",
3063 * Careful here. We can use a locked bus instruction to
3064 * clear PG_A or PG_M safely but we need to synchronize
3065 * with the target cpus when we mess with PG_RW.
3067 * We do not have to force synchronization when clearing
3068 * PG_M even for PTEs generated via virtual memory maps,
3069 * because the virtual kernel will invalidate the pmap
3070 * entry when/if it needs to resynchronize the Modify bit.
3073 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3074 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3081 atomic_clear_int(pte, PG_M|PG_RW);
3084 * The cpu may be trying to set PG_M
3085 * simultaniously with our clearing
3088 if (!atomic_cmpset_int(pte, pbits,
3092 } else if (bit == PG_M) {
3094 * We could also clear PG_RW here to force
3095 * a fault on write to redetect PG_M for
3096 * virtual kernels, but it isn't necessary
3097 * since virtual kernels invalidate the pte
3098 * when they clear the VPTE_M bit in their
3099 * virtual page tables.
3101 atomic_clear_int(pte, PG_M);
3103 atomic_clear_int(pte, bit);
3107 pmap_inval_deinterlock(&info, pv->pv_pmap);
3109 pmap_inval_done(&info);
3113 * Lower the permission for all mappings to a given page.
3118 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3120 if ((prot & VM_PROT_WRITE) == 0) {
3121 lwkt_gettoken(&vm_token);
3122 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3123 pmap_clearbit(m, PG_RW);
3124 vm_page_flag_clear(m, PG_WRITEABLE);
3128 lwkt_reltoken(&vm_token);
3133 * Return the physical address given a physical page index.
3138 pmap_phys_address(vm_pindex_t ppn)
3140 return (i386_ptob(ppn));
3144 * Return a count of reference bits for a page, clearing those bits.
3145 * It is not necessary for every reference bit to be cleared, but it
3146 * is necessary that 0 only be returned when there are truly no
3147 * reference bits set.
3149 * XXX: The exact number of bits to check and clear is a matter that
3150 * should be tested and standardized at some point in the future for
3151 * optimal aging of shared pages.
3156 pmap_ts_referenced(vm_page_t m)
3158 pv_entry_t pv, pvf, pvn;
3162 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3166 lwkt_gettoken(&vm_token);
3168 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3173 pvn = TAILQ_NEXT(pv, pv_list);
3176 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3177 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3180 if (!pmap_track_modified(pv->pv_va))
3183 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3185 if (pte && (*pte & PG_A)) {
3187 atomic_clear_int(pte, PG_A);
3189 atomic_clear_int_nonlocked(pte, PG_A);
3196 } while ((pv = pvn) != NULL && pv != pvf);
3199 lwkt_reltoken(&vm_token);
3206 * Return whether or not the specified physical page was modified
3207 * in any physical maps.
3212 pmap_is_modified(vm_page_t m)
3216 lwkt_gettoken(&vm_token);
3217 res = pmap_testbit(m, PG_M);
3218 lwkt_reltoken(&vm_token);
3223 * Clear the modify bits on the specified physical page.
3228 pmap_clear_modify(vm_page_t m)
3230 lwkt_gettoken(&vm_token);
3231 pmap_clearbit(m, PG_M);
3232 lwkt_reltoken(&vm_token);
3236 * Clear the reference bit on the specified physical page.
3241 pmap_clear_reference(vm_page_t m)
3243 lwkt_gettoken(&vm_token);
3244 pmap_clearbit(m, PG_A);
3245 lwkt_reltoken(&vm_token);
3249 * Miscellaneous support routines follow
3251 * Called from the low level boot code only.
3254 i386_protection_init(void)
3258 kp = protection_codes;
3259 for (prot = 0; prot < 8; prot++) {
3261 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3263 * Read access is also 0. There isn't any execute bit,
3264 * so just make it readable.
3266 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3267 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3268 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3271 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3272 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3273 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3274 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3282 * Map a set of physical memory pages into the kernel virtual
3283 * address space. Return a pointer to where it is mapped. This
3284 * routine is intended to be used for mapping device memory,
3287 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3293 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3295 vm_offset_t va, tmpva, offset;
3298 offset = pa & PAGE_MASK;
3299 size = roundup(offset + size, PAGE_SIZE);
3301 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3303 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3306 for (tmpva = va; size > 0;) {
3307 pte = (unsigned *)vtopte(tmpva);
3308 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3316 return ((void *)(va + offset));
3323 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3325 vm_offset_t base, offset;
3327 base = va & PG_FRAME;
3328 offset = va & PAGE_MASK;
3329 size = roundup(offset + size, PAGE_SIZE);
3330 pmap_qremove(va, size >> PAGE_SHIFT);
3331 kmem_free(&kernel_map, base, size);
3335 * Perform the pmap work for mincore
3337 * The caller must hold vm_token if the caller wishes a stable result,
3338 * and even in that case some bits can change due to third party accesses
3344 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3346 unsigned *ptep, pte;
3350 lwkt_gettoken(&vm_token);
3351 ptep = pmap_pte(pmap, addr);
3353 if (ptep && (pte = *ptep) != 0) {
3356 val = MINCORE_INCORE;
3357 if ((pte & PG_MANAGED) == 0)
3360 pa = pte & PG_FRAME;
3362 m = PHYS_TO_VM_PAGE(pa);
3368 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3369 } else if (m->dirty || pmap_is_modified(m)) {
3371 * Modified by someone else
3373 val |= MINCORE_MODIFIED_OTHER;
3380 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3381 } else if ((m->flags & PG_REFERENCED) ||
3382 pmap_ts_referenced(m)) {
3384 * Referenced by someone else
3386 val |= MINCORE_REFERENCED_OTHER;
3387 vm_page_flag_set(m, PG_REFERENCED);
3391 lwkt_reltoken(&vm_token);
3396 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3397 * vmspace will be ref'd and the old one will be deref'd.
3399 * cr3 will be reloaded if any lwp is the current lwp.
3401 * Only called with new VM spaces.
3402 * The process must have only a single thread.
3403 * No other requirements.
3406 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3408 struct vmspace *oldvm;
3412 oldvm = p->p_vmspace;
3413 if (oldvm != newvm) {
3414 p->p_vmspace = newvm;
3415 KKASSERT(p->p_nthreads == 1);
3416 lp = RB_ROOT(&p->p_lwp_tree);
3417 pmap_setlwpvm(lp, newvm);
3419 sysref_get(&newvm->vm_sysref);
3420 sysref_put(&oldvm->vm_sysref);
3427 * Set the vmspace for a LWP. The vmspace is almost universally set the
3428 * same as the process vmspace, but virtual kernels need to swap out contexts
3429 * on a per-lwp basis.
3431 * Always called with a lp under the caller's direct control, either
3432 * unscheduled or the current lwp.
3437 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3439 struct vmspace *oldvm;
3443 oldvm = lp->lwp_vmspace;
3445 if (oldvm != newvm) {
3446 lp->lwp_vmspace = newvm;
3447 if (curthread->td_lwp == lp) {
3448 pmap = vmspace_pmap(newvm);
3450 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3451 if (pmap->pm_active & CPUMASK_LOCK)
3452 pmap_interlock_wait(newvm);
3454 pmap->pm_active |= 1;
3456 #if defined(SWTCH_OPTIM_STATS)
3459 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3460 load_cr3(curthread->td_pcb->pcb_cr3);
3461 pmap = vmspace_pmap(oldvm);
3463 atomic_clear_cpumask(&pmap->pm_active,
3466 pmap->pm_active &= ~(cpumask_t)1;
3475 * Called when switching to a locked pmap, used to interlock against pmaps
3476 * undergoing modifications to prevent us from activating the MMU for the
3477 * target pmap until all such modifications have completed. We have to do
3478 * this because the thread making the modifications has already set up its
3479 * SMP synchronization mask.
3484 pmap_interlock_wait(struct vmspace *vm)
3486 struct pmap *pmap = &vm->vm_pmap;
3488 if (pmap->pm_active & CPUMASK_LOCK) {
3489 while (pmap->pm_active & CPUMASK_LOCK) {
3492 lwkt_process_ipiq();
3500 * Return a page-directory alignment hint for device mappings which will
3501 * allow the use of super-pages for the mapping.
3506 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3509 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3513 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3518 * Return whether the PGE flag is supported globally.
3523 pmap_get_pgeflag(void)
3529 * Used by kmalloc/kfree, page already exists at va
3532 pmap_kvtom(vm_offset_t va)
3534 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));