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 $
48 * Manages physical address maps.
50 * In most cases we hold page table pages busy in order to manipulate them.
53 * PMAP_DEBUG - see platform/pc32/include/pmap.h
56 #include "opt_disable_pse.h"
58 #include "opt_msgbuf.h"
60 #include <sys/param.h>
61 #include <sys/systm.h>
62 #include <sys/kernel.h>
64 #include <sys/msgbuf.h>
65 #include <sys/vmmeter.h>
67 #include <sys/thread.h>
70 #include <vm/vm_param.h>
71 #include <sys/sysctl.h>
73 #include <vm/vm_kern.h>
74 #include <vm/vm_page.h>
75 #include <vm/vm_map.h>
76 #include <vm/vm_object.h>
77 #include <vm/vm_extern.h>
78 #include <vm/vm_pageout.h>
79 #include <vm/vm_pager.h>
80 #include <vm/vm_zone.h>
83 #include <sys/thread2.h>
84 #include <sys/sysref2.h>
85 #include <sys/spinlock2.h>
87 #include <machine/cputypes.h>
88 #include <machine/md_var.h>
89 #include <machine/specialreg.h>
90 #include <machine/smp.h>
91 #include <machine_base/apic/apicreg.h>
92 #include <machine/globaldata.h>
93 #include <machine/pmap.h>
94 #include <machine/pmap_inval.h>
96 #define PMAP_KEEP_PDIRS
97 #ifndef PMAP_SHPGPERPROC
98 #define PMAP_SHPGPERPROC 200
99 #define PMAP_PVLIMIT 1400000 /* i386 kvm problems */
102 #if defined(DIAGNOSTIC)
103 #define PMAP_DIAGNOSTIC
108 #if !defined(PMAP_DIAGNOSTIC)
109 #define PMAP_INLINE __inline
115 * Get PDEs and PTEs for user/kernel address space
117 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
118 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
120 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
121 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
122 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
123 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
124 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
127 * Given a map and a machine independent protection code,
128 * convert to a vax protection code.
130 #define pte_prot(m, p) \
131 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
132 static int protection_codes[8];
134 struct pmap kernel_pmap;
135 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
137 vm_paddr_t avail_start; /* PA of first available physical page */
138 vm_paddr_t avail_end; /* PA of last available physical page */
139 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
140 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
141 vm_offset_t virtual2_start;
142 vm_offset_t virtual2_end;
143 vm_offset_t KvaStart; /* VA start of KVA space */
144 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
145 vm_offset_t KvaSize; /* max size of kernel virtual address space */
146 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
147 static int pgeflag; /* PG_G or-in */
148 static int pseflag; /* PG_PS or-in */
150 static vm_object_t kptobj;
153 vm_offset_t kernel_vm_end;
156 * Data for the pv entry allocation mechanism
158 static vm_zone_t pvzone;
159 static struct vm_zone pvzone_store;
160 static struct vm_object pvzone_obj;
161 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
162 static int pmap_pagedaemon_waken = 0;
163 static struct pv_entry *pvinit;
166 * Considering all the issues I'm having with pmap caching, if breakage
167 * continues to occur, and for debugging, I've added a sysctl that will
168 * just do an unconditional invltlb.
170 static int dreadful_invltlb;
172 SYSCTL_INT(_vm, OID_AUTO, dreadful_invltlb,
173 CTLFLAG_RW, &dreadful_invltlb, 0, "Debugging sysctl to force invltlb on pmap operations");
176 * All those kernel PT submaps that BSD is so fond of
178 pt_entry_t *CMAP1 = 0, *ptmmap;
179 caddr_t CADDR1 = 0, ptvmmap = 0;
180 static pt_entry_t *msgbufmap;
181 struct msgbuf *msgbufp=0;
186 static pt_entry_t *pt_crashdumpmap;
187 static caddr_t crashdumpmap;
189 extern pt_entry_t *SMPpt;
191 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
192 static unsigned * get_ptbase (pmap_t pmap);
193 static pv_entry_t get_pv_entry (void);
194 static void i386_protection_init (void);
195 static __inline void pmap_clearbit (vm_page_t m, int bit);
197 static void pmap_remove_all (vm_page_t m);
198 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
199 vm_offset_t sva, pmap_inval_info_t info);
200 static void pmap_remove_page (struct pmap *pmap,
201 vm_offset_t va, pmap_inval_info_t info);
202 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
203 vm_offset_t va, pmap_inval_info_t info);
204 static boolean_t pmap_testbit (vm_page_t m, int bit);
205 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
206 vm_page_t mpte, vm_page_t m);
208 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
210 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
211 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
212 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
213 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
214 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
215 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
217 static unsigned pdir4mb;
220 * Move the kernel virtual free pointer to the next
221 * 4MB. This is used to help improve performance
222 * by using a large (4MB) page for much of the kernel
223 * (.text, .data, .bss)
227 pmap_kmem_choose(vm_offset_t addr)
229 vm_offset_t newaddr = addr;
231 if (cpu_feature & CPUID_PSE) {
232 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
239 * This function returns a pointer to the pte entry in the pmap and has
240 * the side effect of potentially retaining a cached mapping of the pmap.
242 * The caller must hold vm_token and the returned value is only valid
243 * until the caller blocks or releases the token.
247 pmap_pte(pmap_t pmap, vm_offset_t va)
251 ASSERT_LWKT_TOKEN_HELD(&vm_token);
253 pdeaddr = (unsigned *) pmap_pde(pmap, va);
254 if (*pdeaddr & PG_PS)
257 return get_ptbase(pmap) + i386_btop(va);
263 * pmap_pte using the kernel_pmap
265 * Used for debugging, no requirements.
268 pmap_kernel_pte(vm_offset_t va)
272 pdeaddr = (unsigned *) pmap_pde(&kernel_pmap, va);
273 if (*pdeaddr & PG_PS)
276 return (unsigned *)vtopte(va);
283 * Super fast pmap_pte routine best used when scanning the pv lists.
284 * This eliminates many course-grained invltlb calls. Note that many of
285 * the pv list scans are across different pmaps and it is very wasteful
286 * to do an entire invltlb when checking a single mapping.
288 * Should only be called while in a critical section.
290 * The caller must hold vm_token and the returned value is only valid
291 * until the caller blocks or releases the token.
295 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
297 struct mdglobaldata *gd = mdcpu;
300 ASSERT_LWKT_TOKEN_HELD(&vm_token);
301 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
302 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
303 unsigned index = i386_btop(va);
304 /* are we current address space or kernel? */
305 if ((pmap == &kernel_pmap) ||
306 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
307 return (unsigned *) PTmap + index;
309 newpf = pde & PG_FRAME;
310 if (((*(unsigned *)gd->gd_PMAP1) & PG_FRAME) != newpf) {
311 *(unsigned *)gd->gd_PMAP1 = newpf | PG_RW | PG_V;
312 cpu_invlpg(gd->gd_PADDR1);
314 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
321 * Bootstrap the system enough to run with virtual memory.
323 * On the i386 this is called after mapping has already been enabled
324 * and just syncs the pmap module with what has already been done.
325 * [We can't call it easily with mapping off since the kernel is not
326 * mapped with PA == VA, hence we would have to relocate every address
327 * from the linked base (virtual) address "KERNBASE" to the actual
328 * (physical) address starting relative to 0]
331 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
335 struct mdglobaldata *gd;
339 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
340 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
341 KvaEnd = KvaStart + KvaSize;
343 avail_start = firstaddr;
346 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
347 * too large. It should instead be correctly calculated in locore.s and
348 * not based on 'first' (which is a physical address, not a virtual
349 * address, for the start of unused physical memory). The kernel
350 * page tables are NOT double mapped and thus should not be included
351 * in this calculation.
353 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
354 virtual_start = pmap_kmem_choose(virtual_start);
355 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
358 * Initialize protection array.
360 i386_protection_init();
363 * The kernel's pmap is statically allocated so we don't have to use
364 * pmap_create, which is unlikely to work correctly at this part of
365 * the boot sequence (XXX and which no longer exists).
367 * The kernel_pmap's pm_pteobj is used only for locking and not
370 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
371 kernel_pmap.pm_count = 1;
372 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
373 kernel_pmap.pm_pteobj = &kernel_object;
374 TAILQ_INIT(&kernel_pmap.pm_pvlist);
375 TAILQ_INIT(&kernel_pmap.pm_pvlist_free);
376 spin_init(&kernel_pmap.pm_spin);
377 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
381 * Reserve some special page table entries/VA space for temporary
384 #define SYSMAP(c, p, v, n) \
385 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
388 pte = (pt_entry_t *) pmap_kernel_pte(va);
391 * CMAP1/CMAP2 are used for zeroing and copying pages.
393 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
398 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
401 * ptvmmap is used for reading arbitrary physical pages via
404 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
407 * msgbufp is used to map the system message buffer.
408 * XXX msgbufmap is not used.
410 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
411 atop(round_page(MSGBUF_SIZE)))
416 for (i = 0; i < NKPT; i++)
420 * PG_G is terribly broken on SMP because we IPI invltlb's in some
421 * cases rather then invl1pg. Actually, I don't even know why it
422 * works under UP because self-referential page table mappings
427 if (cpu_feature & CPUID_PGE)
432 * Initialize the 4MB page size flag
436 * The 4MB page version of the initial
437 * kernel page mapping.
441 #if !defined(DISABLE_PSE)
442 if (cpu_feature & CPUID_PSE) {
445 * Note that we have enabled PSE mode
448 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
449 ptditmp &= ~(NBPDR - 1);
450 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
455 * Enable the PSE mode. If we are SMP we can't do this
456 * now because the APs will not be able to use it when
459 load_cr4(rcr4() | CR4_PSE);
462 * We can do the mapping here for the single processor
463 * case. We simply ignore the old page table page from
467 * For SMP, we still need 4K pages to bootstrap APs,
468 * PSE will be enabled as soon as all APs are up.
470 PTD[KPTDI] = (pd_entry_t)ptditmp;
471 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
478 * We need to finish setting up the globaldata page for the BSP.
479 * locore has already populated the page table for the mdglobaldata
482 pg = MDGLOBALDATA_BASEALLOC_PAGES;
483 gd = &CPU_prvspace[0].mdglobaldata;
484 gd->gd_CMAP1 = &SMPpt[pg + 0];
485 gd->gd_CMAP2 = &SMPpt[pg + 1];
486 gd->gd_CMAP3 = &SMPpt[pg + 2];
487 gd->gd_PMAP1 = &SMPpt[pg + 3];
488 gd->gd_GDMAP1 = &PTD[APTDPTDI];
489 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
490 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
491 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
492 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
493 gd->gd_GDADDR1= (unsigned *)VADDR(APTDPTDI, 0);
500 * Set 4mb pdir for mp startup
505 if (pseflag && (cpu_feature & CPUID_PSE)) {
506 load_cr4(rcr4() | CR4_PSE);
507 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
508 kernel_pmap.pm_pdir[KPTDI] =
509 PTD[KPTDI] = (pd_entry_t)pdir4mb;
517 * Initialize the pmap module, called by vm_init()
519 * Called from the low level boot code only.
528 * object for kernel page table pages
530 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
533 * Allocate memory for random pmap data structures. Includes the
537 for(i = 0; i < vm_page_array_size; i++) {
540 m = &vm_page_array[i];
541 TAILQ_INIT(&m->md.pv_list);
542 m->md.pv_list_count = 0;
546 * init the pv free list
548 initial_pvs = vm_page_array_size;
549 if (initial_pvs < MINPV)
551 pvzone = &pvzone_store;
552 pvinit = (void *)kmem_alloc(&kernel_map,
553 initial_pvs * sizeof (struct pv_entry));
554 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
555 pvinit, initial_pvs);
558 * Now it is safe to enable pv_table recording.
560 pmap_initialized = TRUE;
564 * Initialize the address space (zone) for the pv_entries. Set a
565 * high water mark so that the system can recover from excessive
566 * numbers of pv entries.
568 * Called from the low level boot code only.
573 int shpgperproc = PMAP_SHPGPERPROC;
576 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
577 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
581 * Horrible hack for systems with a lot of memory running i386.
582 * the calculated pv_entry_max can wind up eating a ton of KVM
583 * so put a cap on the number of entries if the user did not
584 * change any of the values. This saves about 44MB of KVM on
585 * boxes with 3+GB of ram.
587 * On the flip side, this makes it more likely that some setups
588 * will run out of pv entries. Those sysads will have to bump
589 * the limit up with vm.pamp.pv_entries or vm.pmap.shpgperproc.
591 if (shpgperproc == PMAP_SHPGPERPROC) {
592 if (pv_entry_max > PMAP_PVLIMIT)
593 pv_entry_max = PMAP_PVLIMIT;
596 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
597 pv_entry_high_water = 9 * (pv_entry_max / 10);
600 * Subtract out pages already installed in the zone (hack)
602 entry_max = pv_entry_max - vm_page_array_size;
606 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
610 /***************************************************
611 * Low level helper routines.....
612 ***************************************************/
617 test_m_maps_pv(vm_page_t m, pv_entry_t pv)
623 KKASSERT(pv->pv_m == m);
625 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
632 panic("test_m_maps_pv: failed m %p pv %p\n", m, pv);
636 ptbase_assert(struct pmap *pmap)
638 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
640 /* are we current address space or kernel? */
641 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME))
643 KKASSERT(frame == (*mdcpu->gd_GDMAP1 & PG_FRAME));
648 #define test_m_maps_pv(m, pv)
649 #define ptbase_assert(pmap)
653 #if defined(PMAP_DIAGNOSTIC)
656 * This code checks for non-writeable/modified pages.
657 * This should be an invalid condition.
660 pmap_nw_modified(pt_entry_t ptea)
666 if ((pte & (PG_M|PG_RW)) == PG_M)
675 * This routine defines the region(s) of memory that should not be tested
676 * for the modified bit.
680 static PMAP_INLINE int
681 pmap_track_modified(vm_offset_t va)
683 if ((va < clean_sva) || (va >= clean_eva))
690 * Retrieve the mapped page table base for a particular pmap. Use our self
691 * mapping for the kernel_pmap or our current pmap.
693 * For foreign pmaps we use the per-cpu page table map. Since this involves
694 * installing a ptd it's actually (per-process x per-cpu). However, we
695 * still cannot depend on our mapping to survive thread switches because
696 * the process might be threaded and switching to another thread for the
697 * same process on the same cpu will allow that other thread to make its
700 * This could be a bit confusing but the jist is for something like the
701 * vkernel which uses foreign pmaps all the time this represents a pretty
702 * good cache that avoids unnecessary invltlb()s.
704 * The caller must hold vm_token and the returned value is only valid
705 * until the caller blocks or releases the token.
708 get_ptbase(pmap_t pmap)
710 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
711 struct mdglobaldata *gd = mdcpu;
713 ASSERT_LWKT_TOKEN_HELD(&vm_token);
716 * We can use PTmap if the pmap is our current address space or
717 * the kernel address space.
719 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
720 return (unsigned *) PTmap;
724 * Otherwise we use the per-cpu alternative page table map. Each
725 * cpu gets its own map. Because of this we cannot use this map
726 * from interrupts or threads which can preempt.
728 * Even if we already have the map cached we may still have to
729 * invalidate the TLB if another cpu modified a PDE in the map.
731 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
732 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
734 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
735 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
736 pmap->pm_cached |= gd->mi.gd_cpumask;
738 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
739 pmap->pm_cached |= gd->mi.gd_cpumask;
741 } else if (dreadful_invltlb) {
744 return ((unsigned *)gd->gd_GDADDR1);
750 * Extract the physical page address associated with the map/VA pair.
752 * The caller may hold vm_token if it desires non-blocking operation.
755 pmap_extract(pmap_t pmap, vm_offset_t va)
758 vm_offset_t pdirindex;
760 lwkt_gettoken(&vm_token);
761 pdirindex = va >> PDRSHIFT;
762 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
764 if ((rtval & PG_PS) != 0) {
765 rtval &= ~(NBPDR - 1);
766 rtval |= va & (NBPDR - 1);
768 pte = get_ptbase(pmap) + i386_btop(va);
769 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
774 lwkt_reltoken(&vm_token);
778 /***************************************************
779 * Low level mapping routines.....
780 ***************************************************/
783 * Map a wired VM page to a KVA, fully SMP synchronized.
785 * No requirements, non blocking.
788 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
792 pmap_inval_info info;
794 pmap_inval_init(&info);
795 npte = pa | PG_RW | PG_V | pgeflag;
796 pte = (unsigned *)vtopte(va);
797 pmap_inval_interlock(&info, &kernel_pmap, va);
799 pmap_inval_deinterlock(&info, &kernel_pmap);
800 pmap_inval_done(&info);
804 * Map a wired VM page to a KVA, synchronized on current cpu only.
806 * No requirements, non blocking.
809 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
814 npte = pa | PG_RW | PG_V | pgeflag;
815 pte = (unsigned *)vtopte(va);
817 cpu_invlpg((void *)va);
821 * Synchronize a previously entered VA on all cpus.
823 * No requirements, non blocking.
826 pmap_kenter_sync(vm_offset_t va)
828 pmap_inval_info info;
830 pmap_inval_init(&info);
831 pmap_inval_interlock(&info, &kernel_pmap, va);
832 pmap_inval_deinterlock(&info, &kernel_pmap);
833 pmap_inval_done(&info);
837 * Synchronize a previously entered VA on the current cpu only.
839 * No requirements, non blocking.
842 pmap_kenter_sync_quick(vm_offset_t va)
844 cpu_invlpg((void *)va);
848 * Remove a page from the kernel pagetables, fully SMP synchronized.
850 * No requirements, non blocking.
853 pmap_kremove(vm_offset_t va)
856 pmap_inval_info info;
858 pmap_inval_init(&info);
859 pte = (unsigned *)vtopte(va);
860 pmap_inval_interlock(&info, &kernel_pmap, va);
862 pmap_inval_deinterlock(&info, &kernel_pmap);
863 pmap_inval_done(&info);
867 * Remove a page from the kernel pagetables, synchronized on current cpu only.
869 * No requirements, non blocking.
872 pmap_kremove_quick(vm_offset_t va)
875 pte = (unsigned *)vtopte(va);
877 cpu_invlpg((void *)va);
881 * Adjust the permissions of a page in the kernel page table,
882 * synchronized on the current cpu only.
884 * No requirements, non blocking.
887 pmap_kmodify_rw(vm_offset_t va)
889 atomic_set_int(vtopte(va), PG_RW);
890 cpu_invlpg((void *)va);
894 * Adjust the permissions of a page in the kernel page table,
895 * synchronized on the current cpu only.
897 * No requirements, non blocking.
900 pmap_kmodify_nc(vm_offset_t va)
902 atomic_set_int(vtopte(va), PG_N);
903 cpu_invlpg((void *)va);
907 * Map a range of physical addresses into kernel virtual address space.
909 * No requirements, non blocking.
912 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
914 vm_offset_t sva, virt;
917 while (start < end) {
918 pmap_kenter(virt, start);
927 * Add a list of wired pages to the kva, fully SMP synchronized.
929 * No requirements, non blocking.
932 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
936 end_va = va + count * PAGE_SIZE;
938 while (va < end_va) {
941 pte = (unsigned *)vtopte(va);
942 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
943 cpu_invlpg((void *)va);
948 smp_invltlb(); /* XXX */
953 * Remove pages from KVA, fully SMP synchronized.
955 * No requirements, non blocking.
958 pmap_qremove(vm_offset_t va, int count)
962 end_va = va + count*PAGE_SIZE;
964 while (va < end_va) {
967 pte = (unsigned *)vtopte(va);
969 cpu_invlpg((void *)va);
978 * This routine works like vm_page_lookup() but also blocks as long as the
979 * page is busy. This routine does not busy the page it returns.
981 * The caller must hold the object.
984 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
988 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
989 m = vm_page_lookup_busy_wait(object, pindex, FALSE, "pplookp");
995 * Create a new thread and optionally associate it with a (new) process.
996 * NOTE! the new thread's cpu may not equal the current cpu.
999 pmap_init_thread(thread_t td)
1001 /* enforce pcb placement */
1002 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1003 td->td_savefpu = &td->td_pcb->pcb_save;
1004 td->td_sp = (char *)td->td_pcb - 16;
1008 * This routine directly affects the fork perf for a process.
1011 pmap_init_proc(struct proc *p)
1016 * Dispose the UPAGES for a process that has exited.
1017 * This routine directly impacts the exit perf of a process.
1020 pmap_dispose_proc(struct proc *p)
1022 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1025 /***************************************************
1026 * Page table page management routines.....
1027 ***************************************************/
1030 * This routine unholds page table pages, and if the hold count
1031 * drops to zero, then it decrements the wire count.
1033 * The caller must hold vm_token.
1034 * This function can block.
1037 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1040 * Wait until we can busy the page ourselves. We cannot have
1041 * any active flushes if we block.
1043 vm_page_busy_wait(m, FALSE, "pmuwpt");
1044 KASSERT(m->queue == PQ_NONE,
1045 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1047 if (m->hold_count == 1) {
1049 * Unmap the page table page.
1051 * NOTE: We must clear pm_cached for all cpus, including
1052 * the current one, when clearing a page directory
1055 pmap_inval_interlock(info, pmap, -1);
1056 KKASSERT(pmap->pm_pdir[m->pindex]);
1057 pmap->pm_pdir[m->pindex] = 0;
1058 pmap->pm_cached = 0;
1059 pmap_inval_deinterlock(info, pmap);
1061 KKASSERT(pmap->pm_stats.resident_count > 0);
1062 --pmap->pm_stats.resident_count;
1064 if (pmap->pm_ptphint == m)
1065 pmap->pm_ptphint = NULL;
1068 * This was our last hold, the page had better be unwired
1069 * after we decrement wire_count.
1071 * FUTURE NOTE: shared page directory page could result in
1072 * multiple wire counts.
1076 KKASSERT(m->wire_count == 0);
1077 atomic_add_int(&vmstats.v_wire_count, -1);
1078 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1080 vm_page_free_zero(m);
1083 KKASSERT(m->hold_count > 1);
1091 * The caller must hold vm_token.
1092 * This function can block.
1094 static PMAP_INLINE int
1095 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1097 KKASSERT(m->hold_count > 0);
1098 if (m->hold_count > 1) {
1102 return _pmap_unwire_pte_hold(pmap, m, info);
1107 * After removing a (user) page table entry, this routine is used to
1108 * conditionally free the page, and manage the hold/wire counts.
1110 * The caller must hold vm_token.
1111 * This function can block regardless.
1114 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1115 pmap_inval_info_t info)
1119 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1121 if (va >= UPT_MIN_ADDRESS)
1125 ptepindex = (va >> PDRSHIFT);
1126 if (pmap->pm_ptphint &&
1127 (pmap->pm_ptphint->pindex == ptepindex)) {
1128 mpte = pmap->pm_ptphint;
1130 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1131 pmap->pm_ptphint = mpte;
1132 vm_page_wakeup(mpte);
1136 return pmap_unwire_pte_hold(pmap, mpte, info);
1140 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1141 * it, and IdlePTD, represents the template used to update all other pmaps.
1143 * On architectures where the kernel pmap is not integrated into the user
1144 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1145 * kernel_pmap should be used to directly access the kernel_pmap.
1150 pmap_pinit0(struct pmap *pmap)
1153 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1154 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1156 pmap->pm_active = 0;
1157 pmap->pm_cached = 0;
1158 pmap->pm_ptphint = NULL;
1159 TAILQ_INIT(&pmap->pm_pvlist);
1160 TAILQ_INIT(&pmap->pm_pvlist_free);
1161 spin_init(&pmap->pm_spin);
1162 lwkt_token_init(&pmap->pm_token, "pmap_tok");
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 |
1201 pmap->pm_pdirm = ptdpg;
1202 vm_page_flag_clear(ptdpg, PG_MAPPED);
1203 vm_page_wire(ptdpg);
1204 KKASSERT(ptdpg->valid == VM_PAGE_BITS_ALL);
1205 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1206 vm_page_wakeup(ptdpg);
1208 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1210 /* install self-referential address mapping entry */
1211 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1212 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1215 pmap->pm_active = 0;
1216 pmap->pm_cached = 0;
1217 pmap->pm_ptphint = NULL;
1218 TAILQ_INIT(&pmap->pm_pvlist);
1219 TAILQ_INIT(&pmap->pm_pvlist_free);
1220 spin_init(&pmap->pm_spin);
1221 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1222 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1223 pmap->pm_stats.resident_count = 1;
1227 * Clean up a pmap structure so it can be physically freed. This routine
1228 * is called by the vmspace dtor function. A great deal of pmap data is
1229 * left passively mapped to improve vmspace management so we have a bit
1230 * of cleanup work to do here.
1235 pmap_puninit(pmap_t pmap)
1239 KKASSERT(pmap->pm_active == 0);
1240 if ((p = pmap->pm_pdirm) != NULL) {
1241 KKASSERT(pmap->pm_pdir != NULL);
1242 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1243 vm_page_busy_wait(p, FALSE, "pgpun");
1245 atomic_add_int(&vmstats.v_wire_count, -1);
1246 vm_page_free_zero(p);
1247 pmap->pm_pdirm = NULL;
1249 if (pmap->pm_pdir) {
1250 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1251 pmap->pm_pdir = NULL;
1253 if (pmap->pm_pteobj) {
1254 vm_object_deallocate(pmap->pm_pteobj);
1255 pmap->pm_pteobj = NULL;
1260 * Wire in kernel global address entries. To avoid a race condition
1261 * between pmap initialization and pmap_growkernel, this procedure
1262 * adds the pmap to the master list (which growkernel scans to update),
1263 * then copies the template.
1268 pmap_pinit2(struct pmap *pmap)
1271 * XXX copies current process, does not fill in MPPTDI
1273 spin_lock(&pmap_spin);
1274 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1275 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1276 spin_unlock(&pmap_spin);
1280 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1281 * 0 on failure (if the procedure had to sleep).
1283 * When asked to remove the page directory page itself, we actually just
1284 * leave it cached so we do not have to incur the SMP inval overhead of
1285 * removing the kernel mapping. pmap_puninit() will take care of it.
1287 * The caller must hold vm_token.
1288 * This function can block regardless.
1291 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1293 unsigned *pde = (unsigned *) pmap->pm_pdir;
1296 * This code optimizes the case of freeing non-busy
1297 * page-table pages. Those pages are zero now, and
1298 * might as well be placed directly into the zero queue.
1300 if (vm_page_busy_try(p, FALSE)) {
1301 vm_page_sleep_busy(p, FALSE, "pmaprl");
1306 * Remove the page table page from the processes address space.
1308 KKASSERT(pmap->pm_stats.resident_count > 0);
1309 KKASSERT(pde[p->pindex]);
1311 --pmap->pm_stats.resident_count;
1312 pmap->pm_cached = 0;
1314 if (p->hold_count) {
1315 panic("pmap_release: freeing held page table page");
1317 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1318 pmap->pm_ptphint = NULL;
1321 * We leave the page directory page cached, wired, and mapped in
1322 * the pmap until the dtor function (pmap_puninit()) gets called.
1323 * However, still clean it up so we can set PG_ZERO.
1325 * The pmap has already been removed from the pmap_list in the
1328 if (p->pindex == PTDPTDI) {
1329 bzero(pde + KPTDI, nkpt * PTESIZE);
1330 bzero(pde + MPPTDI, (NPDEPG - MPPTDI) * PTESIZE);
1331 vm_page_flag_set(p, PG_ZERO);
1335 atomic_add_int(&vmstats.v_wire_count, -1);
1336 vm_page_free_zero(p);
1342 * This routine is called if the page table page is not mapped correctly.
1344 * The caller must hold vm_token.
1347 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1353 * Find or fabricate a new pagetable page. Setting VM_ALLOC_ZERO
1354 * will zero any new page and mark it valid.
1356 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1357 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1359 KASSERT(m->queue == PQ_NONE,
1360 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1363 * Increment the hold count for the page we will be returning to
1369 * It is possible that someone else got in and mapped by the page
1370 * directory page while we were blocked, if so just unbusy and
1371 * return the held page.
1373 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1374 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1379 if (m->wire_count == 0)
1380 atomic_add_int(&vmstats.v_wire_count, 1);
1385 * Map the pagetable page into the process address space, if
1386 * it isn't already there.
1388 * NOTE: For safety clear pm_cached for all cpus including the
1389 * current one when adding a PDE to the map.
1391 ++pmap->pm_stats.resident_count;
1393 ptepa = VM_PAGE_TO_PHYS(m);
1394 pmap->pm_pdir[ptepindex] =
1395 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1396 pmap->pm_cached = 0;
1399 * Set the page table hint
1401 pmap->pm_ptphint = m;
1402 vm_page_flag_set(m, PG_MAPPED);
1409 * Allocate a page table entry for a va.
1411 * The caller must hold vm_token.
1414 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1420 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1423 * Calculate pagetable page index
1425 ptepindex = va >> PDRSHIFT;
1428 * Get the page directory entry
1430 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1433 * This supports switching from a 4MB page to a
1436 if (ptepa & PG_PS) {
1437 pmap->pm_pdir[ptepindex] = 0;
1444 * If the page table page is mapped, we just increment the
1445 * hold count, and activate it.
1449 * In order to get the page table page, try the
1452 if (pmap->pm_ptphint &&
1453 (pmap->pm_ptphint->pindex == ptepindex)) {
1454 m = pmap->pm_ptphint;
1456 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1457 pmap->pm_ptphint = m;
1464 * Here if the pte page isn't mapped, or if it has been deallocated.
1466 return _pmap_allocpte(pmap, ptepindex);
1470 /***************************************************
1471 * Pmap allocation/deallocation routines.
1472 ***************************************************/
1475 * Release any resources held by the given physical map.
1476 * Called when a pmap initialized by pmap_pinit is being released.
1477 * Should only be called if the map contains no valid mappings.
1479 * Caller must hold pmap->pm_token
1481 static int pmap_release_callback(struct vm_page *p, void *data);
1484 pmap_release(struct pmap *pmap)
1486 vm_object_t object = pmap->pm_pteobj;
1487 struct rb_vm_page_scan_info info;
1489 KASSERT(pmap->pm_active == 0,
1490 ("pmap still active! %08x", pmap->pm_active));
1491 #if defined(DIAGNOSTIC)
1492 if (object->ref_count != 1)
1493 panic("pmap_release: pteobj reference count != 1");
1497 info.object = object;
1499 spin_lock(&pmap_spin);
1500 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1501 spin_unlock(&pmap_spin);
1503 vm_object_hold(object);
1507 info.limit = object->generation;
1509 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1510 pmap_release_callback, &info);
1511 if (info.error == 0 && info.mpte) {
1512 if (!pmap_release_free_page(pmap, info.mpte))
1515 } while (info.error);
1516 vm_object_drop(object);
1518 pmap->pm_cached = 0;
1522 * The caller must hold vm_token.
1525 pmap_release_callback(struct vm_page *p, void *data)
1527 struct rb_vm_page_scan_info *info = data;
1529 if (p->pindex == PTDPTDI) {
1533 if (!pmap_release_free_page(info->pmap, p)) {
1537 if (info->object->generation != info->limit) {
1545 * Grow the number of kernel page table entries, if needed.
1550 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1552 vm_offset_t addr = kend;
1554 vm_offset_t ptppaddr;
1558 vm_object_hold(kptobj);
1559 if (kernel_vm_end == 0) {
1560 kernel_vm_end = KERNBASE;
1562 while (pdir_pde(PTD, kernel_vm_end)) {
1563 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1564 ~(PAGE_SIZE * NPTEPG - 1);
1568 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1569 while (kernel_vm_end < addr) {
1570 if (pdir_pde(PTD, kernel_vm_end)) {
1571 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1572 ~(PAGE_SIZE * NPTEPG - 1);
1577 * This index is bogus, but out of the way
1579 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_NORMAL |
1581 VM_ALLOC_INTERRUPT);
1583 panic("pmap_growkernel: no memory to grow kernel");
1586 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1587 pmap_zero_page(ptppaddr);
1588 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1589 pdir_pde(PTD, kernel_vm_end) = newpdir;
1590 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1594 * This update must be interlocked with pmap_pinit2.
1596 spin_lock(&pmap_spin);
1597 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1598 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1600 spin_unlock(&pmap_spin);
1601 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1602 ~(PAGE_SIZE * NPTEPG - 1);
1604 vm_object_drop(kptobj);
1608 * Retire the given physical map from service.
1610 * Should only be called if the map contains no valid mappings.
1615 pmap_destroy(pmap_t pmap)
1620 lwkt_gettoken(&vm_token);
1621 if (--pmap->pm_count == 0) {
1623 panic("destroying a pmap is not yet implemented");
1625 lwkt_reltoken(&vm_token);
1629 * Add a reference to the specified pmap.
1634 pmap_reference(pmap_t pmap)
1637 lwkt_gettoken(&vm_token);
1639 lwkt_reltoken(&vm_token);
1643 /***************************************************
1644 * page management routines.
1645 ***************************************************/
1648 * free the pv_entry back to the free list. This function may be
1649 * called from an interrupt.
1651 * The caller must hold vm_token.
1653 static PMAP_INLINE void
1654 free_pv_entry(pv_entry_t pv)
1657 KKASSERT(pv->pv_m != NULL);
1665 * get a new pv_entry, allocating a block from the system
1666 * when needed. This function may be called from an interrupt.
1668 * The caller must hold vm_token.
1674 if (pv_entry_high_water &&
1675 (pv_entry_count > pv_entry_high_water) &&
1676 (pmap_pagedaemon_waken == 0)) {
1677 pmap_pagedaemon_waken = 1;
1678 wakeup (&vm_pages_needed);
1680 return zalloc(pvzone);
1684 * This routine is very drastic, but can save the system
1694 static int warningdone=0;
1696 if (pmap_pagedaemon_waken == 0)
1698 lwkt_gettoken(&vm_token);
1699 pmap_pagedaemon_waken = 0;
1701 if (warningdone < 5) {
1702 kprintf("pmap_collect: collecting pv entries -- "
1703 "suggest increasing PMAP_SHPGPERPROC\n");
1707 for (i = 0; i < vm_page_array_size; i++) {
1708 m = &vm_page_array[i];
1709 if (m->wire_count || m->hold_count)
1711 if (vm_page_busy_try(m, TRUE) == 0) {
1712 if (m->wire_count == 0 && m->hold_count == 0) {
1718 lwkt_reltoken(&vm_token);
1723 * If it is the first entry on the list, it is actually
1724 * in the header and we must copy the following entry up
1725 * to the header. Otherwise we must search the list for
1726 * the entry. In either case we free the now unused entry.
1728 * The caller must hold vm_token.
1731 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1732 vm_offset_t va, pmap_inval_info_t info)
1737 ASSERT_LWKT_TOKEN_HELD(&vm_token);
1738 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1739 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1740 if (pmap == pv->pv_pmap && va == pv->pv_va)
1744 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1746 KKASSERT(pv->pv_pmap == pmap);
1748 if (va == pv->pv_va)
1755 test_m_maps_pv(m, pv);
1756 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1757 m->md.pv_list_count--;
1759 atomic_add_int(&m->object->agg_pv_list_count, -1);
1760 if (TAILQ_EMPTY(&m->md.pv_list))
1761 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1762 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1763 ++pmap->pm_generation;
1764 vm_object_hold(pmap->pm_pteobj);
1765 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1766 vm_object_drop(pmap->pm_pteobj);
1773 * Create a pv entry for page at pa for (pmap, va).
1775 * The caller must hold vm_token.
1778 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1782 pv = get_pv_entry();
1784 KKASSERT(pv->pv_m == NULL);
1791 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1792 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1793 ++pmap->pm_generation;
1794 m->md.pv_list_count++;
1796 atomic_add_int(&m->object->agg_pv_list_count, 1);
1800 * pmap_remove_pte: do the things to unmap a page in a process.
1802 * The caller must hold vm_token.
1804 * WARNING! As with most other pmap functions this one can block, so
1805 * callers using temporary page table mappings must reload
1809 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1810 pmap_inval_info_t info)
1815 ptbase_assert(pmap);
1816 pmap_inval_interlock(info, pmap, va);
1817 ptbase_assert(pmap);
1818 oldpte = loadandclear(ptq);
1820 pmap->pm_stats.wired_count -= 1;
1821 pmap_inval_deinterlock(info, pmap);
1824 * Machines that don't support invlpg, also don't support
1825 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1829 cpu_invlpg((void *)va);
1830 KKASSERT(pmap->pm_stats.resident_count > 0);
1831 --pmap->pm_stats.resident_count;
1832 if (oldpte & PG_MANAGED) {
1833 m = PHYS_TO_VM_PAGE(oldpte);
1834 if (oldpte & PG_M) {
1835 #if defined(PMAP_DIAGNOSTIC)
1836 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1837 kprintf("pmap_remove: modified page not "
1838 "writable: va: %p, pte: 0x%lx\n",
1839 (void *)va, (long)oldpte);
1842 if (pmap_track_modified(va))
1846 vm_page_flag_set(m, PG_REFERENCED);
1847 return pmap_remove_entry(pmap, m, va, info);
1849 return pmap_unuse_pt(pmap, va, NULL, info);
1856 * Remove a single page from a process address space.
1858 * The caller must hold vm_token.
1861 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1866 * if there is no pte for this address, just skip it!!! Otherwise
1867 * get a local va for mappings for this pmap and remove the entry.
1869 if (*pmap_pde(pmap, va) != 0) {
1870 ptq = get_ptbase(pmap) + i386_btop(va);
1872 pmap_remove_pte(pmap, ptq, va, info);
1879 * Remove the given range of addresses from the specified map.
1881 * It is assumed that the start and end are properly rounded to the page
1887 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1891 vm_offset_t ptpaddr;
1892 vm_offset_t sindex, eindex;
1893 struct pmap_inval_info info;
1898 vm_object_hold(pmap->pm_pteobj);
1899 lwkt_gettoken(&vm_token);
1900 if (pmap->pm_stats.resident_count == 0) {
1901 lwkt_reltoken(&vm_token);
1902 vm_object_drop(pmap->pm_pteobj);
1906 pmap_inval_init(&info);
1909 * special handling of removing one page. a very
1910 * common operation and easy to short circuit some
1913 if (((sva + PAGE_SIZE) == eva) &&
1914 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1915 pmap_remove_page(pmap, sva, &info);
1916 pmap_inval_done(&info);
1917 lwkt_reltoken(&vm_token);
1918 vm_object_drop(pmap->pm_pteobj);
1923 * Get a local virtual address for the mappings that are being
1926 sindex = i386_btop(sva);
1927 eindex = i386_btop(eva);
1929 for (; sindex < eindex; sindex = pdnxt) {
1933 * Calculate index for next page table.
1935 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1936 if (pmap->pm_stats.resident_count == 0)
1939 pdirindex = sindex / NPDEPG;
1940 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1941 pmap_inval_interlock(&info, pmap, -1);
1942 pmap->pm_pdir[pdirindex] = 0;
1943 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1944 pmap->pm_cached = 0;
1945 pmap_inval_deinterlock(&info, pmap);
1950 * Weed out invalid mappings. Note: we assume that the page
1951 * directory table is always allocated, and in kernel virtual.
1957 * Limit our scan to either the end of the va represented
1958 * by the current page table page, or to the end of the
1959 * range being removed.
1961 if (pdnxt > eindex) {
1966 * NOTE: pmap_remove_pte() can block and wipe the temporary
1969 for (; sindex != pdnxt; sindex++) {
1972 ptbase = get_ptbase(pmap);
1973 if (ptbase[sindex] == 0)
1975 va = i386_ptob(sindex);
1976 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1980 pmap_inval_done(&info);
1981 lwkt_reltoken(&vm_token);
1982 vm_object_drop(pmap->pm_pteobj);
1986 * Removes this physical page from all physical maps in which it resides.
1987 * Reflects back modify bits to the pager.
1992 pmap_remove_all(vm_page_t m)
1994 struct pmap_inval_info info;
1995 unsigned *pte, tpte;
1998 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2001 pmap_inval_init(&info);
2002 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2003 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2004 --pv->pv_pmap->pm_stats.resident_count;
2006 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2007 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2008 tpte = loadandclear(pte);
2010 pv->pv_pmap->pm_stats.wired_count--;
2011 pmap_inval_deinterlock(&info, pv->pv_pmap);
2013 vm_page_flag_set(m, PG_REFERENCED);
2015 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
2019 * Update the vm_page_t clean and reference bits.
2022 #if defined(PMAP_DIAGNOSTIC)
2023 if (pmap_nw_modified((pt_entry_t) tpte)) {
2024 kprintf("pmap_remove_all: modified page "
2025 "not writable: va: %p, pte: 0x%lx\n",
2026 (void *)pv->pv_va, (long)tpte);
2029 if (pmap_track_modified(pv->pv_va))
2033 KKASSERT(pv->pv_m == m);
2035 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2036 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2037 ++pv->pv_pmap->pm_generation;
2038 m->md.pv_list_count--;
2040 atomic_add_int(&m->object->agg_pv_list_count, -1);
2041 if (TAILQ_EMPTY(&m->md.pv_list))
2042 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2043 vm_object_hold(pv->pv_pmap->pm_pteobj);
2044 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2045 vm_object_drop(pv->pv_pmap->pm_pteobj);
2048 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2049 pmap_inval_done(&info);
2053 * Set the physical protection on the specified range of this map
2059 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2062 vm_offset_t pdnxt, ptpaddr;
2063 vm_pindex_t sindex, eindex;
2064 pmap_inval_info info;
2069 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2070 pmap_remove(pmap, sva, eva);
2074 if (prot & VM_PROT_WRITE)
2077 lwkt_gettoken(&vm_token);
2078 pmap_inval_init(&info);
2080 ptbase = get_ptbase(pmap);
2082 sindex = i386_btop(sva);
2083 eindex = i386_btop(eva);
2085 for (; sindex < eindex; sindex = pdnxt) {
2088 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2090 pdirindex = sindex / NPDEPG;
2091 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
2092 pmap_inval_interlock(&info, pmap, -1);
2093 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2094 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2095 pmap_inval_deinterlock(&info, pmap);
2100 * Weed out invalid mappings. Note: we assume that the page
2101 * directory table is always allocated, and in kernel virtual.
2106 if (pdnxt > eindex) {
2110 for (; sindex != pdnxt; sindex++) {
2118 pmap_inval_interlock(&info, pmap, i386_ptob(sindex));
2120 pbits = ptbase[sindex];
2123 if (pbits & PG_MANAGED) {
2126 m = PHYS_TO_VM_PAGE(pbits);
2127 vm_page_flag_set(m, PG_REFERENCED);
2131 if (pmap_track_modified(i386_ptob(sindex))) {
2133 m = PHYS_TO_VM_PAGE(pbits);
2140 if (pbits != cbits &&
2141 !atomic_cmpset_int(ptbase + sindex, pbits, cbits)) {
2144 pmap_inval_deinterlock(&info, pmap);
2147 pmap_inval_done(&info);
2148 lwkt_reltoken(&vm_token);
2152 * Insert the given physical page (p) at the specified virtual address (v)
2153 * in the target physical map with the protection requested.
2155 * If specified, the page will be wired down, meaning that the related pte
2156 * cannot be reclaimed.
2161 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2167 vm_offset_t origpte, newpte;
2169 pmap_inval_info info;
2175 #ifdef PMAP_DIAGNOSTIC
2177 panic("pmap_enter: toobig");
2178 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2179 panic("pmap_enter: invalid to pmap_enter page "
2180 "table pages (va: %p)", (void *)va);
2183 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2184 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2185 print_backtrace(-1);
2187 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2188 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2189 print_backtrace(-1);
2192 vm_object_hold(pmap->pm_pteobj);
2193 lwkt_gettoken(&vm_token);
2196 * In the case that a page table page is not
2197 * resident, we are creating it here.
2199 if (va < UPT_MIN_ADDRESS)
2200 mpte = pmap_allocpte(pmap, va);
2204 if ((prot & VM_PROT_NOSYNC) == 0)
2205 pmap_inval_init(&info);
2206 pte = pmap_pte(pmap, va);
2209 * Page Directory table entry not valid, we need a new PT page
2212 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p\n",
2213 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
2216 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2217 origpte = *(vm_offset_t *)pte;
2218 opa = origpte & PG_FRAME;
2220 if (origpte & PG_PS)
2221 panic("pmap_enter: attempted pmap_enter on 4MB page");
2224 * Mapping has not changed, must be protection or wiring change.
2226 if (origpte && (opa == pa)) {
2228 * Wiring change, just update stats. We don't worry about
2229 * wiring PT pages as they remain resident as long as there
2230 * are valid mappings in them. Hence, if a user page is wired,
2231 * the PT page will be also.
2233 if (wired && ((origpte & PG_W) == 0))
2234 pmap->pm_stats.wired_count++;
2235 else if (!wired && (origpte & PG_W))
2236 pmap->pm_stats.wired_count--;
2238 #if defined(PMAP_DIAGNOSTIC)
2239 if (pmap_nw_modified((pt_entry_t) origpte)) {
2240 kprintf("pmap_enter: modified page not "
2241 "writable: va: %p, pte: 0x%lx\n",
2242 (void *)va, (long )origpte);
2247 * Remove the extra pte reference. Note that we cannot
2248 * optimize the RO->RW case because we have adjusted the
2249 * wiring count above and may need to adjust the wiring
2256 * We might be turning off write access to the page,
2257 * so we go ahead and sense modify status.
2259 if (origpte & PG_MANAGED) {
2260 if ((origpte & PG_M) && pmap_track_modified(va)) {
2262 om = PHYS_TO_VM_PAGE(opa);
2266 KKASSERT(m->flags & PG_MAPPED);
2271 * Mapping has changed, invalidate old range and fall through to
2272 * handle validating new mapping.
2274 * Since we have a ref on the page directory page pmap_pte()
2275 * will always return non-NULL.
2277 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2278 * to get wiped. reload the ptbase. I'm not sure if it is
2279 * also possible to race another pmap_enter() but check for
2285 KKASSERT((origpte & PG_FRAME) ==
2286 (*(vm_offset_t *)pte & PG_FRAME));
2287 err = pmap_remove_pte(pmap, pte, va, &info);
2289 panic("pmap_enter: pte vanished, va: %p", (void *)va);
2290 pte = pmap_pte(pmap, va);
2291 origpte = *(vm_offset_t *)pte;
2292 opa = origpte & PG_FRAME;
2294 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2300 * Enter on the PV list if part of our managed memory. Note that we
2301 * raise IPL while manipulating pv_table since pmap_enter can be
2302 * called at interrupt time.
2304 if (pmap_initialized &&
2305 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2306 pmap_insert_entry(pmap, va, mpte, m);
2307 ptbase_assert(pmap);
2309 vm_page_flag_set(m, PG_MAPPED);
2313 * Increment counters
2315 ++pmap->pm_stats.resident_count;
2317 pmap->pm_stats.wired_count++;
2318 KKASSERT(*pte == 0);
2322 * Now validate mapping with desired protection/wiring.
2324 ptbase_assert(pmap);
2325 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2329 if (va < UPT_MIN_ADDRESS)
2331 if (pmap == &kernel_pmap)
2335 * if the mapping or permission bits are different, we need
2336 * to update the pte.
2338 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2339 if (prot & VM_PROT_NOSYNC)
2340 cpu_invlpg((void *)va);
2342 pmap_inval_interlock(&info, pmap, va);
2343 ptbase_assert(pmap);
2344 KKASSERT(*pte == 0 ||
2345 (*pte & PG_FRAME) == (newpte & PG_FRAME));
2346 *pte = newpte | PG_A;
2347 if ((prot & VM_PROT_NOSYNC) == 0)
2348 pmap_inval_deinterlock(&info, pmap);
2350 vm_page_flag_set(m, PG_WRITEABLE);
2352 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2353 if ((prot & VM_PROT_NOSYNC) == 0)
2354 pmap_inval_done(&info);
2355 lwkt_reltoken(&vm_token);
2356 vm_object_drop(pmap->pm_pteobj);
2360 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2361 * This code also assumes that the pmap has no pre-existing entry for this
2364 * This code currently may only be used on user pmaps, not kernel_pmap.
2369 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2376 pmap_inval_info info;
2378 vm_object_hold(pmap->pm_pteobj);
2379 lwkt_gettoken(&vm_token);
2380 pmap_inval_init(&info);
2382 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2383 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2384 print_backtrace(-1);
2386 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2387 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2388 print_backtrace(-1);
2391 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2394 * Calculate the page table page (mpte), allocating it if necessary.
2396 * A held page table page (mpte), or NULL, is passed onto the
2397 * section following.
2399 if (va < UPT_MIN_ADDRESS) {
2401 * Calculate pagetable page index
2403 ptepindex = va >> PDRSHIFT;
2407 * Get the page directory entry
2409 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2412 * If the page table page is mapped, we just increment
2413 * the hold count, and activate it.
2417 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2418 if (pmap->pm_ptphint &&
2419 (pmap->pm_ptphint->pindex == ptepindex)) {
2420 mpte = pmap->pm_ptphint;
2422 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
2423 pmap->pm_ptphint = mpte;
2424 vm_page_wakeup(mpte);
2429 mpte = _pmap_allocpte(pmap, ptepindex);
2431 } while (mpte == NULL);
2434 /* this code path is not yet used */
2438 * With a valid (and held) page directory page, we can just use
2439 * vtopte() to get to the pte. If the pte is already present
2440 * we do not disturb it.
2442 pte = (unsigned *)vtopte(va);
2445 pmap_unwire_pte_hold(pmap, mpte, &info);
2446 pa = VM_PAGE_TO_PHYS(m);
2447 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2448 pmap_inval_done(&info);
2449 lwkt_reltoken(&vm_token);
2450 vm_object_drop(pmap->pm_pteobj);
2455 * Enter on the PV list if part of our managed memory
2457 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2458 pmap_insert_entry(pmap, va, mpte, m);
2459 vm_page_flag_set(m, PG_MAPPED);
2463 * Increment counters
2465 ++pmap->pm_stats.resident_count;
2467 pa = VM_PAGE_TO_PHYS(m);
2470 * Now validate mapping with RO protection
2472 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2473 *pte = pa | PG_V | PG_U;
2475 *pte = pa | PG_V | PG_U | PG_MANAGED;
2476 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2477 pmap_inval_done(&info);
2478 lwkt_reltoken(&vm_token);
2479 vm_object_drop(pmap->pm_pteobj);
2483 * Make a temporary mapping for a physical address. This is only intended
2484 * to be used for panic dumps.
2486 * The caller is responsible for calling smp_invltlb().
2491 pmap_kenter_temporary(vm_paddr_t pa, long i)
2493 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2494 return ((void *)crashdumpmap);
2497 #define MAX_INIT_PT (96)
2500 * This routine preloads the ptes for a given object into the specified pmap.
2501 * This eliminates the blast of soft faults on process startup and
2502 * immediately after an mmap.
2506 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2509 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2510 vm_object_t object, vm_pindex_t pindex,
2511 vm_size_t size, int limit)
2513 struct rb_vm_page_scan_info info;
2518 * We can't preinit if read access isn't set or there is no pmap
2521 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2525 * We can't preinit if the pmap is not the current pmap
2527 lp = curthread->td_lwp;
2528 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2531 psize = i386_btop(size);
2533 if ((object->type != OBJT_VNODE) ||
2534 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2535 (object->resident_page_count > MAX_INIT_PT))) {
2539 if (psize + pindex > object->size) {
2540 if (object->size < pindex)
2542 psize = object->size - pindex;
2549 * Use a red-black scan to traverse the requested range and load
2550 * any valid pages found into the pmap.
2552 * We cannot safely scan the object's memq unless we are in a
2553 * critical section since interrupts can remove pages from objects.
2555 info.start_pindex = pindex;
2556 info.end_pindex = pindex + psize - 1;
2562 vm_object_hold(object);
2563 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2564 pmap_object_init_pt_callback, &info);
2565 vm_object_drop(object);
2569 * The caller must hold vm_token.
2573 pmap_object_init_pt_callback(vm_page_t p, void *data)
2575 struct rb_vm_page_scan_info *info = data;
2576 vm_pindex_t rel_index;
2578 * don't allow an madvise to blow away our really
2579 * free pages allocating pv entries.
2581 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2582 vmstats.v_free_count < vmstats.v_free_reserved) {
2585 if (vm_page_busy_try(p, TRUE))
2587 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2588 (p->flags & PG_FICTITIOUS) == 0) {
2589 if ((p->queue - p->pc) == PQ_CACHE)
2590 vm_page_deactivate(p);
2591 rel_index = p->pindex - info->start_pindex;
2592 pmap_enter_quick(info->pmap,
2593 info->addr + i386_ptob(rel_index), p);
2600 * Return TRUE if the pmap is in shape to trivially
2601 * pre-fault the specified address.
2603 * Returns FALSE if it would be non-trivial or if a
2604 * pte is already loaded into the slot.
2609 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2614 lwkt_gettoken(&vm_token);
2615 if ((*pmap_pde(pmap, addr)) == 0) {
2618 pte = (unsigned *) vtopte(addr);
2619 ret = (*pte) ? 0 : 1;
2621 lwkt_reltoken(&vm_token);
2626 * Change the wiring attribute for a map/virtual-adderss pair. The mapping
2627 * must already exist.
2632 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2639 lwkt_gettoken(&vm_token);
2640 pte = pmap_pte(pmap, va);
2642 if (wired && !pmap_pte_w(pte))
2643 pmap->pm_stats.wired_count++;
2644 else if (!wired && pmap_pte_w(pte))
2645 pmap->pm_stats.wired_count--;
2648 * Wiring is not a hardware characteristic so there is no need to
2649 * invalidate TLB. However, in an SMP environment we must use
2650 * a locked bus cycle to update the pte (if we are not using
2651 * the pmap_inval_*() API that is)... it's ok to do this for simple
2656 atomic_set_int(pte, PG_W);
2658 atomic_clear_int(pte, PG_W);
2661 atomic_set_int_nonlocked(pte, PG_W);
2663 atomic_clear_int_nonlocked(pte, PG_W);
2665 lwkt_reltoken(&vm_token);
2669 * Copy the range specified by src_addr/len from the source map to the
2670 * range dst_addr/len in the destination map.
2672 * This routine is only advisory and need not do anything.
2677 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2678 vm_size_t len, vm_offset_t src_addr)
2684 * Zero the specified PA by mapping the page into KVM and clearing its
2690 pmap_zero_page(vm_paddr_t phys)
2692 struct mdglobaldata *gd = mdcpu;
2695 if (*(int *)gd->gd_CMAP3)
2696 panic("pmap_zero_page: CMAP3 busy");
2697 *(int *)gd->gd_CMAP3 =
2698 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2699 cpu_invlpg(gd->gd_CADDR3);
2701 #if defined(I686_CPU)
2702 if (cpu_class == CPUCLASS_686)
2703 i686_pagezero(gd->gd_CADDR3);
2706 bzero(gd->gd_CADDR3, PAGE_SIZE);
2707 *(int *) gd->gd_CMAP3 = 0;
2712 * Assert that a page is empty, panic if it isn't.
2717 pmap_page_assertzero(vm_paddr_t phys)
2719 struct mdglobaldata *gd = mdcpu;
2723 if (*(int *)gd->gd_CMAP3)
2724 panic("pmap_zero_page: CMAP3 busy");
2725 *(int *)gd->gd_CMAP3 =
2726 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2727 cpu_invlpg(gd->gd_CADDR3);
2728 for (i = 0; i < PAGE_SIZE; i += 4) {
2729 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2730 panic("pmap_page_assertzero() @ %p not zero!\n",
2731 (void *)gd->gd_CADDR3);
2734 *(int *) gd->gd_CMAP3 = 0;
2739 * Zero part of a physical page by mapping it into memory and clearing
2740 * its contents with bzero.
2742 * off and size may not cover an area beyond a single hardware page.
2747 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2749 struct mdglobaldata *gd = mdcpu;
2752 if (*(int *) gd->gd_CMAP3)
2753 panic("pmap_zero_page: CMAP3 busy");
2754 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2755 cpu_invlpg(gd->gd_CADDR3);
2757 #if defined(I686_CPU)
2758 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2759 i686_pagezero(gd->gd_CADDR3);
2762 bzero((char *)gd->gd_CADDR3 + off, size);
2763 *(int *) gd->gd_CMAP3 = 0;
2768 * Copy the physical page from the source PA to the target PA.
2769 * This function may be called from an interrupt. No locking
2775 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2777 struct mdglobaldata *gd = mdcpu;
2780 if (*(int *) gd->gd_CMAP1)
2781 panic("pmap_copy_page: CMAP1 busy");
2782 if (*(int *) gd->gd_CMAP2)
2783 panic("pmap_copy_page: CMAP2 busy");
2785 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2786 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2788 cpu_invlpg(gd->gd_CADDR1);
2789 cpu_invlpg(gd->gd_CADDR2);
2791 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2793 *(int *) gd->gd_CMAP1 = 0;
2794 *(int *) gd->gd_CMAP2 = 0;
2799 * Copy the physical page from the source PA to the target PA.
2800 * This function may be called from an interrupt. No locking
2806 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2808 struct mdglobaldata *gd = mdcpu;
2811 if (*(int *) gd->gd_CMAP1)
2812 panic("pmap_copy_page: CMAP1 busy");
2813 if (*(int *) gd->gd_CMAP2)
2814 panic("pmap_copy_page: CMAP2 busy");
2816 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2817 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2819 cpu_invlpg(gd->gd_CADDR1);
2820 cpu_invlpg(gd->gd_CADDR2);
2822 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2823 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2826 *(int *) gd->gd_CMAP1 = 0;
2827 *(int *) gd->gd_CMAP2 = 0;
2832 * Returns true if the pmap's pv is one of the first
2833 * 16 pvs linked to from this page. This count may
2834 * be changed upwards or downwards in the future; it
2835 * is only necessary that true be returned for a small
2836 * subset of pmaps for proper page aging.
2841 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2846 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2849 lwkt_gettoken(&vm_token);
2850 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2851 if (pv->pv_pmap == pmap) {
2852 lwkt_reltoken(&vm_token);
2859 lwkt_reltoken(&vm_token);
2864 * Remove all pages from specified address space
2865 * this aids process exit speeds. Also, this code
2866 * is special cased for current process only, but
2867 * can have the more generic (and slightly slower)
2868 * mode enabled. This is much faster than pmap_remove
2869 * in the case of running down an entire address space.
2874 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2877 unsigned *pte, tpte;
2880 pmap_inval_info info;
2882 int32_t save_generation;
2884 lp = curthread->td_lwp;
2885 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2890 if (pmap->pm_pteobj)
2891 vm_object_hold(pmap->pm_pteobj);
2892 lwkt_gettoken(&vm_token);
2893 pmap_inval_init(&info);
2895 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2896 if (pv->pv_va >= eva || pv->pv_va < sva) {
2897 npv = TAILQ_NEXT(pv, pv_plist);
2901 KKASSERT(pmap == pv->pv_pmap);
2904 pte = (unsigned *)vtopte(pv->pv_va);
2906 pte = pmap_pte_quick(pmap, pv->pv_va);
2908 pmap_inval_interlock(&info, pmap, pv->pv_va);
2911 * We cannot remove wired pages from a process' mapping
2915 pmap_inval_deinterlock(&info, pmap);
2916 npv = TAILQ_NEXT(pv, pv_plist);
2920 tpte = loadandclear(pte);
2921 pmap_inval_deinterlock(&info, pmap);
2923 m = PHYS_TO_VM_PAGE(tpte);
2924 test_m_maps_pv(m, pv);
2926 KASSERT(m < &vm_page_array[vm_page_array_size],
2927 ("pmap_remove_pages: bad tpte %x", tpte));
2929 KKASSERT(pmap->pm_stats.resident_count > 0);
2930 --pmap->pm_stats.resident_count;
2933 * Update the vm_page_t clean and reference bits.
2939 npv = TAILQ_NEXT(pv, pv_plist);
2941 KKASSERT(pv->pv_m == m);
2942 KKASSERT(pv->pv_pmap == pmap);
2944 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2945 save_generation = ++pmap->pm_generation;
2947 m->md.pv_list_count--;
2949 atomic_add_int(&m->object->agg_pv_list_count, -1);
2950 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2951 if (TAILQ_EMPTY(&m->md.pv_list))
2952 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2954 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2958 * Restart the scan if we blocked during the unuse or free
2959 * calls and other removals were made.
2961 if (save_generation != pmap->pm_generation) {
2962 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2963 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2966 pmap_inval_done(&info);
2967 lwkt_reltoken(&vm_token);
2968 if (pmap->pm_pteobj)
2969 vm_object_drop(pmap->pm_pteobj);
2973 * pmap_testbit tests bits in pte's
2974 * note that the testbit/clearbit routines are inline,
2975 * and a lot of things compile-time evaluate.
2977 * The caller must hold vm_token.
2980 pmap_testbit(vm_page_t m, int bit)
2985 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2988 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2991 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2993 * if the bit being tested is the modified bit, then
2994 * mark clean_map and ptes as never
2997 if (bit & (PG_A|PG_M)) {
2998 if (!pmap_track_modified(pv->pv_va))
3002 #if defined(PMAP_DIAGNOSTIC)
3004 kprintf("Null pmap (tb) at va: %p\n",
3009 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3018 * This routine is used to modify bits in ptes
3020 * The caller must hold vm_token.
3022 static __inline void
3023 pmap_clearbit(vm_page_t m, int bit)
3025 struct pmap_inval_info info;
3030 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3033 pmap_inval_init(&info);
3036 * Loop over all current mappings setting/clearing as appropos If
3037 * setting RO do we need to clear the VAC?
3039 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3041 * don't write protect pager mappings
3044 if (!pmap_track_modified(pv->pv_va))
3048 #if defined(PMAP_DIAGNOSTIC)
3050 kprintf("Null pmap (cb) at va: %p\n",
3057 * Careful here. We can use a locked bus instruction to
3058 * clear PG_A or PG_M safely but we need to synchronize
3059 * with the target cpus when we mess with PG_RW.
3061 * We do not have to force synchronization when clearing
3062 * PG_M even for PTEs generated via virtual memory maps,
3063 * because the virtual kernel will invalidate the pmap
3064 * entry when/if it needs to resynchronize the Modify bit.
3067 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3068 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3075 atomic_clear_int(pte, PG_M|PG_RW);
3078 * The cpu may be trying to set PG_M
3079 * simultaniously with our clearing
3082 if (!atomic_cmpset_int(pte, pbits,
3086 } else if (bit == PG_M) {
3088 * We could also clear PG_RW here to force
3089 * a fault on write to redetect PG_M for
3090 * virtual kernels, but it isn't necessary
3091 * since virtual kernels invalidate the pte
3092 * when they clear the VPTE_M bit in their
3093 * virtual page tables.
3095 atomic_clear_int(pte, PG_M);
3097 atomic_clear_int(pte, bit);
3101 pmap_inval_deinterlock(&info, pv->pv_pmap);
3103 pmap_inval_done(&info);
3107 * Lower the permission for all mappings to a given page.
3112 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3114 if ((prot & VM_PROT_WRITE) == 0) {
3115 lwkt_gettoken(&vm_token);
3116 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3117 pmap_clearbit(m, PG_RW);
3118 vm_page_flag_clear(m, PG_WRITEABLE);
3122 lwkt_reltoken(&vm_token);
3127 * Return the physical address given a physical page index.
3132 pmap_phys_address(vm_pindex_t ppn)
3134 return (i386_ptob(ppn));
3138 * Return a count of reference bits for a page, clearing those bits.
3139 * It is not necessary for every reference bit to be cleared, but it
3140 * is necessary that 0 only be returned when there are truly no
3141 * reference bits set.
3146 pmap_ts_referenced(vm_page_t m)
3148 pv_entry_t pv, pvf, pvn;
3152 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3155 lwkt_gettoken(&vm_token);
3157 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3162 pvn = TAILQ_NEXT(pv, pv_list);
3164 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3165 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3167 if (!pmap_track_modified(pv->pv_va))
3170 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3172 if (pte && (*pte & PG_A)) {
3174 atomic_clear_int(pte, PG_A);
3176 atomic_clear_int_nonlocked(pte, PG_A);
3183 } while ((pv = pvn) != NULL && pv != pvf);
3186 lwkt_reltoken(&vm_token);
3192 * Return whether or not the specified physical page was modified
3193 * in any physical maps.
3198 pmap_is_modified(vm_page_t m)
3202 lwkt_gettoken(&vm_token);
3203 res = pmap_testbit(m, PG_M);
3204 lwkt_reltoken(&vm_token);
3209 * Clear the modify bits on the specified physical page.
3214 pmap_clear_modify(vm_page_t m)
3216 lwkt_gettoken(&vm_token);
3217 pmap_clearbit(m, PG_M);
3218 lwkt_reltoken(&vm_token);
3222 * Clear the reference bit on the specified physical page.
3227 pmap_clear_reference(vm_page_t m)
3229 lwkt_gettoken(&vm_token);
3230 pmap_clearbit(m, PG_A);
3231 lwkt_reltoken(&vm_token);
3235 * Miscellaneous support routines follow
3237 * Called from the low level boot code only.
3240 i386_protection_init(void)
3244 kp = protection_codes;
3245 for (prot = 0; prot < 8; prot++) {
3247 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3249 * Read access is also 0. There isn't any execute bit,
3250 * so just make it readable.
3252 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3253 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3254 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3257 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3258 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3259 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3260 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3268 * Map a set of physical memory pages into the kernel virtual
3269 * address space. Return a pointer to where it is mapped. This
3270 * routine is intended to be used for mapping device memory,
3273 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3279 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3281 vm_offset_t va, tmpva, offset;
3284 offset = pa & PAGE_MASK;
3285 size = roundup(offset + size, PAGE_SIZE);
3287 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3289 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3292 for (tmpva = va; size > 0;) {
3293 pte = (unsigned *)vtopte(tmpva);
3294 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3302 return ((void *)(va + offset));
3306 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3308 vm_offset_t va, tmpva, offset;
3311 offset = pa & PAGE_MASK;
3312 size = roundup(offset + size, PAGE_SIZE);
3314 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3316 panic("pmap_mapdev_uncacheable: "
3317 "Couldn't alloc kernel virtual memory");
3321 for (tmpva = va; size > 0;) {
3322 pte = (unsigned *)vtopte(tmpva);
3323 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3331 return ((void *)(va + offset));
3338 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3340 vm_offset_t base, offset;
3342 base = va & PG_FRAME;
3343 offset = va & PAGE_MASK;
3344 size = roundup(offset + size, PAGE_SIZE);
3345 pmap_qremove(va, size >> PAGE_SHIFT);
3346 kmem_free(&kernel_map, base, size);
3350 * Perform the pmap work for mincore
3352 * The caller must hold vm_token if the caller wishes a stable result,
3353 * and even in that case some bits can change due to third party accesses
3359 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3361 unsigned *ptep, pte;
3365 lwkt_gettoken(&vm_token);
3366 ptep = pmap_pte(pmap, addr);
3368 if (ptep && (pte = *ptep) != 0) {
3371 val = MINCORE_INCORE;
3372 if ((pte & PG_MANAGED) == 0)
3375 pa = pte & PG_FRAME;
3377 m = PHYS_TO_VM_PAGE(pa);
3383 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3384 } else if (m->dirty || pmap_is_modified(m)) {
3386 * Modified by someone else
3388 val |= MINCORE_MODIFIED_OTHER;
3395 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3396 } else if ((m->flags & PG_REFERENCED) ||
3397 pmap_ts_referenced(m)) {
3399 * Referenced by someone else
3401 val |= MINCORE_REFERENCED_OTHER;
3402 vm_page_flag_set(m, PG_REFERENCED);
3406 lwkt_reltoken(&vm_token);
3411 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3412 * vmspace will be ref'd and the old one will be deref'd.
3414 * cr3 will be reloaded if any lwp is the current lwp.
3416 * Only called with new VM spaces.
3417 * The process must have only a single thread.
3418 * The process must hold the vmspace->vm_map.token for oldvm and newvm
3419 * No other requirements.
3422 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3424 struct vmspace *oldvm;
3427 oldvm = p->p_vmspace;
3428 if (oldvm != newvm) {
3430 sysref_get(&newvm->vm_sysref);
3431 p->p_vmspace = newvm;
3432 KKASSERT(p->p_nthreads == 1);
3433 lp = RB_ROOT(&p->p_lwp_tree);
3434 pmap_setlwpvm(lp, newvm);
3436 sysref_put(&oldvm->vm_sysref);
3441 * Set the vmspace for a LWP. The vmspace is almost universally set the
3442 * same as the process vmspace, but virtual kernels need to swap out contexts
3443 * on a per-lwp basis.
3445 * Always called with a lp under the caller's direct control, either
3446 * unscheduled or the current lwp.
3451 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3453 struct vmspace *oldvm;
3456 oldvm = lp->lwp_vmspace;
3458 if (oldvm != newvm) {
3459 lp->lwp_vmspace = newvm;
3460 if (curthread->td_lwp == lp) {
3461 pmap = vmspace_pmap(newvm);
3463 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3464 if (pmap->pm_active & CPUMASK_LOCK)
3465 pmap_interlock_wait(newvm);
3467 pmap->pm_active |= 1;
3469 #if defined(SWTCH_OPTIM_STATS)
3472 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3473 load_cr3(curthread->td_pcb->pcb_cr3);
3474 pmap = vmspace_pmap(oldvm);
3476 atomic_clear_cpumask(&pmap->pm_active,
3479 pmap->pm_active &= ~(cpumask_t)1;
3487 * Called when switching to a locked pmap, used to interlock against pmaps
3488 * undergoing modifications to prevent us from activating the MMU for the
3489 * target pmap until all such modifications have completed. We have to do
3490 * this because the thread making the modifications has already set up its
3491 * SMP synchronization mask.
3496 pmap_interlock_wait(struct vmspace *vm)
3498 struct pmap *pmap = &vm->vm_pmap;
3500 if (pmap->pm_active & CPUMASK_LOCK) {
3502 DEBUG_PUSH_INFO("pmap_interlock_wait");
3503 while (pmap->pm_active & CPUMASK_LOCK) {
3505 lwkt_process_ipiq();
3515 * Return a page-directory alignment hint for device mappings which will
3516 * allow the use of super-pages for the mapping.
3521 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3524 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3528 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3533 * Return whether the PGE flag is supported globally.
3538 pmap_get_pgeflag(void)
3544 * Used by kmalloc/kfree, page already exists at va
3547 pmap_kvtom(vm_offset_t va)
3549 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));