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>
86 #include <vm/vm_page2.h>
88 #include <machine/cputypes.h>
89 #include <machine/md_var.h>
90 #include <machine/specialreg.h>
91 #include <machine/smp.h>
92 #include <machine_base/apic/apicreg.h>
93 #include <machine/globaldata.h>
94 #include <machine/pmap.h>
95 #include <machine/pmap_inval.h>
97 #define PMAP_KEEP_PDIRS
98 #ifndef PMAP_SHPGPERPROC
99 #define PMAP_SHPGPERPROC 200
100 #define PMAP_PVLIMIT 1400000 /* i386 kvm problems */
103 #if defined(DIAGNOSTIC)
104 #define PMAP_DIAGNOSTIC
109 #if !defined(PMAP_DIAGNOSTIC)
110 #define PMAP_INLINE __inline
116 * Get PDEs and PTEs for user/kernel address space
118 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
119 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
121 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
122 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
123 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
124 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
125 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
128 * Given a map and a machine independent protection code,
129 * convert to a vax protection code.
131 #define pte_prot(m, p) \
132 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
133 static int protection_codes[8];
135 struct pmap kernel_pmap;
136 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
138 vm_paddr_t avail_start; /* PA of first available physical page */
139 vm_paddr_t avail_end; /* PA of last available physical page */
140 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
141 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
142 vm_offset_t virtual2_start;
143 vm_offset_t virtual2_end;
144 vm_offset_t KvaStart; /* VA start of KVA space */
145 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
146 vm_offset_t KvaSize; /* max size of kernel virtual address space */
147 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
148 static int pgeflag; /* PG_G or-in */
149 static int pseflag; /* PG_PS or-in */
152 static vm_object_t kptobj;
155 vm_offset_t kernel_vm_end;
157 #define PAT_INDEX_SIZE 8
158 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE]; /* PAT -> PG_ bits */
159 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
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 = NULL, *ptmmap;
185 caddr_t CADDR1 = NULL, ptvmmap = NULL;
186 static pt_entry_t *msgbufmap;
187 struct msgbuf *msgbufp=NULL;
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 void 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 void 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, pv_entry_t pv,
212 vm_offset_t va, 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 void 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 void pmap_hold(pmap_t pmap);
224 static void pmap_drop(pmap_t pmap);
225 static void pmap_wait(pmap_t pmap, int count);
227 static unsigned pdir4mb;
230 * Move the kernel virtual free pointer to the next
231 * 4MB. This is used to help improve performance
232 * by using a large (4MB) page for much of the kernel
233 * (.text, .data, .bss)
237 pmap_kmem_choose(vm_offset_t addr)
239 vm_offset_t newaddr = addr;
241 if (cpu_feature & CPUID_PSE) {
242 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
249 * This function returns a pointer to the pte entry in the pmap and has
250 * the side effect of potentially retaining a cached mapping of the pmap.
252 * The caller must hold vm_token and the returned value is only valid
253 * until the caller blocks or releases the token.
257 pmap_pte(pmap_t pmap, vm_offset_t va)
261 ASSERT_LWKT_TOKEN_HELD(&vm_token);
263 pdeaddr = (unsigned *) pmap_pde(pmap, va);
264 if (*pdeaddr & PG_PS)
267 return get_ptbase(pmap) + i386_btop(va);
273 * pmap_pte using the kernel_pmap
275 * Used for debugging, no requirements.
278 pmap_kernel_pte(vm_offset_t va)
282 pdeaddr = (unsigned *) pmap_pde(&kernel_pmap, va);
283 if (*pdeaddr & PG_PS)
286 return (unsigned *)vtopte(va);
293 * Super fast pmap_pte routine best used when scanning the pv lists.
294 * This eliminates many course-grained invltlb calls. Note that many of
295 * the pv list scans are across different pmaps and it is very wasteful
296 * to do an entire invltlb when checking a single mapping.
298 * Should only be called while in a critical section.
300 * The caller must hold vm_token and the returned value is only valid
301 * until the caller blocks or releases the token.
305 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
307 struct mdglobaldata *gd = mdcpu;
310 ASSERT_LWKT_TOKEN_HELD(&vm_token);
311 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
312 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
313 unsigned index = i386_btop(va);
314 /* are we current address space or kernel? */
315 if ((pmap == &kernel_pmap) ||
316 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
317 return (unsigned *) PTmap + index;
319 newpf = pde & PG_FRAME;
320 if (((*(unsigned *)gd->gd_PMAP1) & PG_FRAME) != newpf) {
321 *(unsigned *)gd->gd_PMAP1 = newpf | PG_RW | PG_V;
322 cpu_invlpg(gd->gd_PADDR1);
324 return gd->gd_PADDR1 + (index & (NPTEPG - 1));
331 * Bootstrap the system enough to run with virtual memory.
333 * On the i386 this is called after mapping has already been enabled
334 * and just syncs the pmap module with what has already been done.
335 * [We can't call it easily with mapping off since the kernel is not
336 * mapped with PA == VA, hence we would have to relocate every address
337 * from the linked base (virtual) address "KERNBASE" to the actual
338 * (physical) address starting relative to 0]
341 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
345 struct mdglobaldata *gd;
349 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
350 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
351 KvaEnd = KvaStart + KvaSize;
353 avail_start = firstaddr;
356 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
357 * too large. It should instead be correctly calculated in locore.s and
358 * not based on 'first' (which is a physical address, not a virtual
359 * address, for the start of unused physical memory). The kernel
360 * page tables are NOT double mapped and thus should not be included
361 * in this calculation.
363 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
364 virtual_start = pmap_kmem_choose(virtual_start);
365 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
368 * Initialize protection array.
370 i386_protection_init();
373 * The kernel's pmap is statically allocated so we don't have to use
374 * pmap_create, which is unlikely to work correctly at this part of
375 * the boot sequence (XXX and which no longer exists).
377 * The kernel_pmap's pm_pteobj is used only for locking and not
380 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
381 kernel_pmap.pm_count = 1;
382 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
383 kernel_pmap.pm_pteobj = &kernel_object;
384 TAILQ_INIT(&kernel_pmap.pm_pvlist);
385 TAILQ_INIT(&kernel_pmap.pm_pvlist_free);
386 spin_init(&kernel_pmap.pm_spin);
387 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
391 * Reserve some special page table entries/VA space for temporary
394 #define SYSMAP(c, p, v, n) \
395 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
398 pte = (pt_entry_t *) pmap_kernel_pte(va);
401 * CMAP1/CMAP2 are used for zeroing and copying pages.
403 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
408 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
411 * ptvmmap is used for reading arbitrary physical pages via
414 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
417 * msgbufp is used to map the system message buffer.
418 * XXX msgbufmap is not used.
420 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
421 atop(round_page(MSGBUF_SIZE)))
426 for (i = 0; i < NKPT; i++)
430 * PG_G is terribly broken on SMP because we IPI invltlb's in some
431 * cases rather then invl1pg. Actually, I don't even know why it
432 * works under UP because self-referential page table mappings
437 * Initialize the 4MB page size flag
441 * The 4MB page version of the initial
442 * kernel page mapping.
446 #if !defined(DISABLE_PSE)
447 if (cpu_feature & CPUID_PSE) {
450 * Note that we have enabled PSE mode
453 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
454 ptditmp &= ~(NBPDR - 1);
455 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
461 * We need to finish setting up the globaldata page for the BSP.
462 * locore has already populated the page table for the mdglobaldata
465 pg = MDGLOBALDATA_BASEALLOC_PAGES;
466 gd = &CPU_prvspace[0].mdglobaldata;
467 gd->gd_CMAP1 = &SMPpt[pg + 0];
468 gd->gd_CMAP2 = &SMPpt[pg + 1];
469 gd->gd_CMAP3 = &SMPpt[pg + 2];
470 gd->gd_PMAP1 = &SMPpt[pg + 3];
471 gd->gd_GDMAP1 = &PTD[APTDPTDI];
472 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
473 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
474 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
475 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
476 gd->gd_GDADDR1= (unsigned *)VADDR(APTDPTDI, 0);
480 /* Initialize the PAT MSR */
494 * Default values mapping PATi,PCD,PWT bits at system reset.
495 * The default values effectively ignore the PATi bit by
496 * repeating the encodings for 0-3 in 4-7, and map the PCD
497 * and PWT bit combinations to the expected PAT types.
499 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
500 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
501 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
502 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
503 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
504 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
505 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
506 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
507 pat_pte_index[PAT_WRITE_BACK] = 0;
508 pat_pte_index[PAT_WRITE_THROUGH]= 0 | PG_NC_PWT;
509 pat_pte_index[PAT_UNCACHED] = PG_NC_PCD;
510 pat_pte_index[PAT_UNCACHEABLE] = PG_NC_PCD | PG_NC_PWT;
511 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
512 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
514 if (cpu_feature & CPUID_PAT) {
516 * If we support the PAT then set-up entries for
517 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
520 pat_msr = (pat_msr & ~PAT_MASK(4)) |
521 PAT_VALUE(4, PAT_WRITE_PROTECTED);
522 pat_msr = (pat_msr & ~PAT_MASK(5)) |
523 PAT_VALUE(5, PAT_WRITE_COMBINING);
524 pat_pte_index[PAT_WRITE_PROTECTED] = PG_PTE_PAT | 0;
525 pat_pte_index[PAT_WRITE_COMBINING] = PG_PTE_PAT | PG_NC_PWT;
528 * Then enable the PAT
533 load_cr4(cr4 & ~CR4_PGE);
535 /* Disable caches (CD = 1, NW = 0). */
537 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
539 /* Flushes caches and TLBs. */
543 /* Update PAT and index table. */
544 wrmsr(MSR_PAT, pat_msr);
546 /* Flush caches and TLBs again. */
550 /* Restore caches and PGE. */
558 * Set 4mb pdir for mp startup
563 if (pseflag && (cpu_feature & CPUID_PSE)) {
564 load_cr4(rcr4() | CR4_PSE);
565 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
566 kernel_pmap.pm_pdir[KPTDI] =
567 PTD[KPTDI] = (pd_entry_t)pdir4mb;
574 * Initialize the pmap module, called by vm_init()
576 * Called from the low level boot code only.
585 * object for kernel page table pages
587 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
590 * Allocate memory for random pmap data structures. Includes the
594 for(i = 0; i < vm_page_array_size; i++) {
597 m = &vm_page_array[i];
598 TAILQ_INIT(&m->md.pv_list);
599 m->md.pv_list_count = 0;
603 * init the pv free list
605 initial_pvs = vm_page_array_size;
606 if (initial_pvs < MINPV)
608 pvzone = &pvzone_store;
609 pvinit = (void *)kmem_alloc(&kernel_map,
610 initial_pvs * sizeof (struct pv_entry));
611 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
612 pvinit, initial_pvs);
615 * Now it is safe to enable pv_table recording.
617 pmap_initialized = TRUE;
621 * Initialize the address space (zone) for the pv_entries. Set a
622 * high water mark so that the system can recover from excessive
623 * numbers of pv entries.
625 * Called from the low level boot code only.
630 int shpgperproc = PMAP_SHPGPERPROC;
633 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
634 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
638 * Horrible hack for systems with a lot of memory running i386.
639 * the calculated pv_entry_max can wind up eating a ton of KVM
640 * so put a cap on the number of entries if the user did not
641 * change any of the values. This saves about 44MB of KVM on
642 * boxes with 3+GB of ram.
644 * On the flip side, this makes it more likely that some setups
645 * will run out of pv entries. Those sysads will have to bump
646 * the limit up with vm.pamp.pv_entries or vm.pmap.shpgperproc.
648 if (shpgperproc == PMAP_SHPGPERPROC) {
649 if (pv_entry_max > PMAP_PVLIMIT)
650 pv_entry_max = PMAP_PVLIMIT;
653 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
654 pv_entry_high_water = 9 * (pv_entry_max / 10);
657 * Subtract out pages already installed in the zone (hack)
659 entry_max = pv_entry_max - vm_page_array_size;
663 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
667 * Typically used to initialize a fictitious page by vm/device_pager.c
670 pmap_page_init(struct vm_page *m)
673 TAILQ_INIT(&m->md.pv_list);
676 /***************************************************
677 * Low level helper routines.....
678 ***************************************************/
683 test_m_maps_pv(vm_page_t m, pv_entry_t pv)
689 KKASSERT(pv->pv_m == m);
691 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
698 panic("test_m_maps_pv: failed m %p pv %p", m, pv);
702 ptbase_assert(struct pmap *pmap)
704 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
706 /* are we current address space or kernel? */
707 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME))
709 KKASSERT(frame == (*mdcpu->gd_GDMAP1 & PG_FRAME));
714 #define test_m_maps_pv(m, pv)
715 #define ptbase_assert(pmap)
719 #if defined(PMAP_DIAGNOSTIC)
722 * This code checks for non-writeable/modified pages.
723 * This should be an invalid condition.
726 pmap_nw_modified(pt_entry_t ptea)
732 if ((pte & (PG_M|PG_RW)) == PG_M)
741 * This routine defines the region(s) of memory that should not be tested
742 * for the modified bit.
746 static PMAP_INLINE int
747 pmap_track_modified(vm_offset_t va)
749 if ((va < clean_sva) || (va >= clean_eva))
756 * Retrieve the mapped page table base for a particular pmap. Use our self
757 * mapping for the kernel_pmap or our current pmap.
759 * For foreign pmaps we use the per-cpu page table map. Since this involves
760 * installing a ptd it's actually (per-process x per-cpu). However, we
761 * still cannot depend on our mapping to survive thread switches because
762 * the process might be threaded and switching to another thread for the
763 * same process on the same cpu will allow that other thread to make its
766 * This could be a bit confusing but the jist is for something like the
767 * vkernel which uses foreign pmaps all the time this represents a pretty
768 * good cache that avoids unnecessary invltlb()s.
770 * The caller must hold vm_token and the returned value is only valid
771 * until the caller blocks or releases the token.
774 get_ptbase(pmap_t pmap)
776 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
777 struct mdglobaldata *gd = mdcpu;
779 ASSERT_LWKT_TOKEN_HELD(&vm_token);
782 * We can use PTmap if the pmap is our current address space or
783 * the kernel address space.
785 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
786 return (unsigned *) PTmap;
790 * Otherwise we use the per-cpu alternative page table map. Each
791 * cpu gets its own map. Because of this we cannot use this map
792 * from interrupts or threads which can preempt.
794 * Even if we already have the map cached we may still have to
795 * invalidate the TLB if another cpu modified a PDE in the map.
797 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
798 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
800 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
801 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
802 pmap->pm_cached |= gd->mi.gd_cpumask;
804 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
805 pmap->pm_cached |= gd->mi.gd_cpumask;
807 } else if (dreadful_invltlb) {
810 return ((unsigned *)gd->gd_GDADDR1);
816 * Extract the physical page address associated with the map/VA pair.
818 * The caller may hold vm_token if it desires non-blocking operation.
821 pmap_extract(pmap_t pmap, vm_offset_t va)
824 vm_offset_t pdirindex;
826 lwkt_gettoken(&vm_token);
827 pdirindex = va >> PDRSHIFT;
828 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
830 if ((rtval & PG_PS) != 0) {
831 rtval &= ~(NBPDR - 1);
832 rtval |= va & (NBPDR - 1);
834 pte = get_ptbase(pmap) + i386_btop(va);
835 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
840 lwkt_reltoken(&vm_token);
845 * Similar to extract but checks protections, SMP-friendly short-cut for
846 * vm_fault_page[_quick]().
849 pmap_fault_page_quick(pmap_t pmap __unused, vm_offset_t vaddr __unused,
850 vm_prot_t prot __unused)
855 /***************************************************
856 * Low level mapping routines.....
857 ***************************************************/
860 * Map a wired VM page to a KVA, fully SMP synchronized.
862 * No requirements, non blocking.
865 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
869 pmap_inval_info info;
871 pmap_inval_init(&info);
872 npte = pa | PG_RW | PG_V | pgeflag;
873 pte = (unsigned *)vtopte(va);
874 pmap_inval_interlock(&info, &kernel_pmap, va);
876 pmap_inval_deinterlock(&info, &kernel_pmap);
877 pmap_inval_done(&info);
881 * Map a wired VM page to a KVA, synchronized on current cpu only.
883 * No requirements, non blocking.
886 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
891 npte = pa | PG_RW | PG_V | pgeflag;
892 pte = (unsigned *)vtopte(va);
894 cpu_invlpg((void *)va);
898 * Synchronize a previously entered VA on all cpus.
900 * No requirements, non blocking.
903 pmap_kenter_sync(vm_offset_t va)
905 pmap_inval_info info;
907 pmap_inval_init(&info);
908 pmap_inval_interlock(&info, &kernel_pmap, va);
909 pmap_inval_deinterlock(&info, &kernel_pmap);
910 pmap_inval_done(&info);
914 * Synchronize a previously entered VA on the current cpu only.
916 * No requirements, non blocking.
919 pmap_kenter_sync_quick(vm_offset_t va)
921 cpu_invlpg((void *)va);
925 * Remove a page from the kernel pagetables, fully SMP synchronized.
927 * No requirements, non blocking.
930 pmap_kremove(vm_offset_t va)
933 pmap_inval_info info;
935 pmap_inval_init(&info);
936 pte = (unsigned *)vtopte(va);
937 pmap_inval_interlock(&info, &kernel_pmap, va);
939 pmap_inval_deinterlock(&info, &kernel_pmap);
940 pmap_inval_done(&info);
944 * Remove a page from the kernel pagetables, synchronized on current cpu only.
946 * No requirements, non blocking.
949 pmap_kremove_quick(vm_offset_t va)
952 pte = (unsigned *)vtopte(va);
954 cpu_invlpg((void *)va);
958 * Adjust the permissions of a page in the kernel page table,
959 * synchronized on the current cpu only.
961 * No requirements, non blocking.
964 pmap_kmodify_rw(vm_offset_t va)
966 atomic_set_int(vtopte(va), PG_RW);
967 cpu_invlpg((void *)va);
971 * Adjust the permissions of a page in the kernel page table,
972 * synchronized on the current cpu only.
974 * No requirements, non blocking.
977 pmap_kmodify_nc(vm_offset_t va)
979 atomic_set_int(vtopte(va), PG_N);
980 cpu_invlpg((void *)va);
984 * Map a range of physical addresses into kernel virtual address space.
986 * No requirements, non blocking.
989 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
991 vm_offset_t sva, virt;
994 while (start < end) {
995 pmap_kenter(virt, start);
1003 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1006 * Remove the specified set of pages from the data and instruction caches.
1008 * In contrast to pmap_invalidate_cache_range(), this function does not
1009 * rely on the CPU's self-snoop feature, because it is intended for use
1010 * when moving pages into a different cache domain.
1013 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1015 wbinvd(); /* XXX: not optimal */
1019 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1021 KASSERT((sva & PAGE_MASK) == 0,
1022 ("pmap_invalidate_cache_range: sva not page-aligned"));
1023 KASSERT((eva & PAGE_MASK) == 0,
1024 ("pmap_invalidate_cache_range: eva not page-aligned"));
1026 if (cpu_feature & CPUID_SS) {
1027 ; /* If "Self Snoop" is supported, do nothing. */
1029 /* Globally invalidate caches */
1030 cpu_wbinvd_on_all_cpus();
1035 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1037 cpu_wbinvd_on_all_cpus(); /* XXX not optimal */
1041 * Add a list of wired pages to the kva
1042 * this routine is only used for temporary
1043 * kernel mappings that do not need to have
1044 * page modification or references recorded.
1045 * Note that old mappings are simply written
1046 * over. The page *must* be wired.
1049 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1053 end_va = va + count * PAGE_SIZE;
1055 while (va < end_va) {
1059 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V |
1060 pat_pte_index[(*m)->pat_mode] | pgeflag;
1061 cpu_invlpg((void *)va);
1065 smp_invltlb(); /* XXX */
1069 * Remove pages from KVA, fully SMP synchronized.
1071 * No requirements, non blocking.
1074 pmap_qremove(vm_offset_t va, int count)
1078 end_va = va + count*PAGE_SIZE;
1080 while (va < end_va) {
1083 pte = (unsigned *)vtopte(va);
1085 cpu_invlpg((void *)va);
1092 * This routine works like vm_page_lookup() but also blocks as long as the
1093 * page is busy. This routine does not busy the page it returns.
1095 * The caller must hold the object.
1098 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1102 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1103 m = vm_page_lookup_busy_wait(object, pindex, FALSE, "pplookp");
1109 * Create a new thread and optionally associate it with a (new) process.
1110 * NOTE! the new thread's cpu may not equal the current cpu.
1113 pmap_init_thread(thread_t td)
1115 /* enforce pcb placement */
1116 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1117 td->td_savefpu = &td->td_pcb->pcb_save;
1118 td->td_sp = (char *)td->td_pcb - 16;
1122 * This routine directly affects the fork perf for a process.
1125 pmap_init_proc(struct proc *p)
1129 /***************************************************
1130 * Page table page management routines.....
1131 ***************************************************/
1134 * This routine unwires page table pages, removing and freeing the page
1135 * tale page when the wire count drops to 0.
1137 * The caller must hold vm_token.
1138 * This function can block.
1141 _pmap_unwire_pte(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1144 * Wait until we can busy the page ourselves. We cannot have
1145 * any active flushes if we block.
1147 vm_page_busy_wait(m, FALSE, "pmuwpt");
1148 KASSERT(m->queue == PQ_NONE,
1149 ("_pmap_unwire_pte: %p->queue != PQ_NONE", m));
1151 if (m->wire_count == 1) {
1153 * Unmap the page table page.
1155 * NOTE: We must clear pm_cached for all cpus, including
1156 * the current one, when clearing a page directory
1159 pmap_inval_interlock(info, pmap, -1);
1160 KKASSERT(pmap->pm_pdir[m->pindex]);
1161 pmap->pm_pdir[m->pindex] = 0;
1162 pmap->pm_cached = 0;
1163 pmap_inval_deinterlock(info, pmap);
1165 KKASSERT(pmap->pm_stats.resident_count > 0);
1166 --pmap->pm_stats.resident_count;
1168 if (pmap->pm_ptphint == m)
1169 pmap->pm_ptphint = NULL;
1172 * This was our last hold, the page had better be unwired
1173 * after we decrement wire_count.
1175 * FUTURE NOTE: shared page directory page could result in
1176 * multiple wire counts.
1178 vm_page_unwire(m, 0);
1179 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1181 vm_page_free_zero(m);
1184 KKASSERT(m->wire_count > 1);
1185 if (vm_page_unwire_quick(m))
1186 panic("pmap_unwire_pte: Insufficient wire_count");
1193 * The caller must hold vm_token.
1195 * This function can block.
1197 * This function can race the wire_count 2->1 case because the page
1198 * is not busied during the unwire_quick operation. An eventual
1199 * pmap_release() will catch the case.
1201 static PMAP_INLINE int
1202 pmap_unwire_pte(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1204 KKASSERT(m->wire_count > 0);
1205 if (m->wire_count > 1) {
1206 if (vm_page_unwire_quick(m))
1207 panic("pmap_unwire_pte: Insufficient wire_count");
1210 return _pmap_unwire_pte(pmap, m, info);
1215 * After removing a (user) page table entry, this routine is used to
1216 * conditionally free the page, and manage the hold/wire counts.
1218 * The caller must hold vm_token.
1219 * This function can block regardless.
1222 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1223 pmap_inval_info_t info)
1227 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1229 if (va >= UPT_MIN_ADDRESS)
1233 ptepindex = (va >> PDRSHIFT);
1234 if ((mpte = pmap->pm_ptphint) != NULL &&
1235 mpte->pindex == ptepindex &&
1236 (mpte->flags & PG_BUSY) == 0) {
1239 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1240 pmap->pm_ptphint = mpte;
1241 vm_page_wakeup(mpte);
1244 pmap_unwire_pte(pmap, mpte, info);
1248 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1249 * it, and IdlePTD, represents the template used to update all other pmaps.
1251 * On architectures where the kernel pmap is not integrated into the user
1252 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1253 * kernel_pmap should be used to directly access the kernel_pmap.
1258 pmap_pinit0(struct pmap *pmap)
1261 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1262 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1264 pmap->pm_active = 0;
1265 pmap->pm_cached = 0;
1266 pmap->pm_ptphint = NULL;
1267 TAILQ_INIT(&pmap->pm_pvlist);
1268 TAILQ_INIT(&pmap->pm_pvlist_free);
1269 spin_init(&pmap->pm_spin);
1270 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1271 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1275 * Initialize a preallocated and zeroed pmap structure,
1276 * such as one in a vmspace structure.
1281 pmap_pinit(struct pmap *pmap)
1286 * No need to allocate page table space yet but we do need a valid
1287 * page directory table.
1289 if (pmap->pm_pdir == NULL) {
1291 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1295 * Allocate an object for the ptes
1297 if (pmap->pm_pteobj == NULL)
1298 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1301 * Allocate the page directory page, unless we already have
1302 * one cached. If we used the cached page the wire_count will
1303 * already be set appropriately.
1305 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1306 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1307 VM_ALLOC_NORMAL | VM_ALLOC_RETRY |
1309 pmap->pm_pdirm = ptdpg;
1310 vm_page_flag_clear(ptdpg, PG_MAPPED);
1311 vm_page_wire(ptdpg);
1312 KKASSERT(ptdpg->valid == VM_PAGE_BITS_ALL);
1313 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1314 vm_page_wakeup(ptdpg);
1316 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1318 /* install self-referential address mapping entry */
1319 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1320 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1323 pmap->pm_active = 0;
1324 pmap->pm_cached = 0;
1325 pmap->pm_ptphint = NULL;
1326 TAILQ_INIT(&pmap->pm_pvlist);
1327 TAILQ_INIT(&pmap->pm_pvlist_free);
1328 spin_init(&pmap->pm_spin);
1329 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1330 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1331 pmap->pm_stats.resident_count = 1;
1335 * Clean up a pmap structure so it can be physically freed. This routine
1336 * is called by the vmspace dtor function. A great deal of pmap data is
1337 * left passively mapped to improve vmspace management so we have a bit
1338 * of cleanup work to do here.
1343 pmap_puninit(pmap_t pmap)
1347 pmap_wait(pmap, -1);
1348 KKASSERT(pmap->pm_active == 0);
1349 if ((p = pmap->pm_pdirm) != NULL) {
1350 KKASSERT(pmap->pm_pdir != NULL);
1351 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1352 vm_page_busy_wait(p, FALSE, "pgpun");
1353 vm_page_unwire(p, 0);
1354 vm_page_free_zero(p);
1355 pmap->pm_pdirm = NULL;
1357 if (pmap->pm_pdir) {
1358 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1359 pmap->pm_pdir = NULL;
1361 if (pmap->pm_pteobj) {
1362 vm_object_deallocate(pmap->pm_pteobj);
1363 pmap->pm_pteobj = NULL;
1368 * Wire in kernel global address entries. To avoid a race condition
1369 * between pmap initialization and pmap_growkernel, this procedure
1370 * adds the pmap to the master list (which growkernel scans to update),
1371 * then copies the template.
1376 pmap_pinit2(struct pmap *pmap)
1379 * XXX copies current process, does not fill in MPPTDI
1381 spin_lock(&pmap_spin);
1382 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1383 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1384 spin_unlock(&pmap_spin);
1388 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1389 * 0 on failure (if the procedure had to sleep).
1391 * When asked to remove the page directory page itself, we actually just
1392 * leave it cached so we do not have to incur the SMP inval overhead of
1393 * removing the kernel mapping. pmap_puninit() will take care of it.
1395 * The caller must hold vm_token.
1396 * This function can block regardless.
1399 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1401 unsigned *pde = (unsigned *) pmap->pm_pdir;
1404 * This code optimizes the case of freeing non-busy
1405 * page-table pages. Those pages are zero now, and
1406 * might as well be placed directly into the zero queue.
1408 if (vm_page_busy_try(p, FALSE)) {
1409 vm_page_sleep_busy(p, FALSE, "pmaprl");
1413 KKASSERT(pmap->pm_stats.resident_count > 0);
1414 KKASSERT(pde[p->pindex]);
1417 * page table page's wire_count must be 1. Caller is the pmap
1418 * termination code which holds the pm_pteobj, there is a race
1419 * if someone else is trying to hold the VM object in order to
1420 * clean up a wire_count.
1422 if (p->wire_count != 1) {
1423 if (pmap->pm_pteobj->hold_count <= 1)
1424 panic("pmap_release: freeing wired page table page");
1425 kprintf("pmap_release_free_page: unwire race detected\n");
1427 tsleep(p, 0, "pmapx", 1);
1432 * Remove the page table page from the processes address space.
1434 pmap->pm_cached = 0;
1436 --pmap->pm_stats.resident_count;
1437 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1438 pmap->pm_ptphint = NULL;
1441 * We leave the page directory page cached, wired, and mapped in
1442 * the pmap until the dtor function (pmap_puninit()) gets called.
1443 * However, still clean it up so we can set PG_ZERO.
1445 * The pmap has already been removed from the pmap_list in the
1448 if (p->pindex == PTDPTDI) {
1449 bzero(pde + KPTDI, nkpt * PTESIZE);
1450 bzero(pde + MPPTDI, (NPDEPG - MPPTDI) * PTESIZE);
1451 vm_page_flag_set(p, PG_ZERO);
1455 * This case can occur if a pmap_unwire_pte() loses a race
1456 * while the page is unbusied.
1458 /*panic("pmap_release: page should already be gone %p", p);*/
1459 vm_page_flag_clear(p, PG_MAPPED);
1460 vm_page_unwire(p, 0);
1461 vm_page_free_zero(p);
1467 * This routine is called if the page table page is not mapped correctly.
1469 * The caller must hold vm_token.
1472 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1478 * Find or fabricate a new pagetable page. Setting VM_ALLOC_ZERO
1479 * will zero any new page and mark it valid.
1481 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1482 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1484 KASSERT(m->queue == PQ_NONE,
1485 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1488 * Increment the wire count for the page we will be returning to
1494 * It is possible that someone else got in and mapped by the page
1495 * directory page while we were blocked, if so just unbusy and
1496 * return the wired page.
1498 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1499 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1505 * Map the pagetable page into the process address space, if
1506 * it isn't already there.
1508 * NOTE: For safety clear pm_cached for all cpus including the
1509 * current one when adding a PDE to the map.
1511 ++pmap->pm_stats.resident_count;
1513 ptepa = VM_PAGE_TO_PHYS(m);
1514 pmap->pm_pdir[ptepindex] =
1515 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1516 pmap->pm_cached = 0;
1519 * Set the page table hint
1521 pmap->pm_ptphint = m;
1522 vm_page_flag_set(m, PG_MAPPED);
1529 * Allocate a page table entry for a va.
1531 * The caller must hold vm_token.
1534 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1540 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1543 * Calculate pagetable page index
1545 ptepindex = va >> PDRSHIFT;
1548 * Get the page directory entry
1550 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1553 * This supports switching from a 4MB page to a
1556 if (ptepa & PG_PS) {
1557 pmap->pm_pdir[ptepindex] = 0;
1564 * If the page table page is mapped, we just increment the
1565 * wire count, and activate it.
1569 * In order to get the page table page, try the
1572 if ((mpte = pmap->pm_ptphint) != NULL &&
1573 (mpte->pindex == ptepindex) &&
1574 (mpte->flags & PG_BUSY) == 0) {
1575 vm_page_wire_quick(mpte);
1577 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1578 pmap->pm_ptphint = mpte;
1579 vm_page_wire_quick(mpte);
1580 vm_page_wakeup(mpte);
1585 * Here if the pte page isn't mapped, or if it has been deallocated.
1587 return _pmap_allocpte(pmap, ptepindex);
1591 /***************************************************
1592 * Pmap allocation/deallocation routines.
1593 ***************************************************/
1596 * Release any resources held by the given physical map.
1597 * Called when a pmap initialized by pmap_pinit is being released.
1598 * Should only be called if the map contains no valid mappings.
1600 * Caller must hold pmap->pm_token
1602 static int pmap_release_callback(struct vm_page *p, void *data);
1605 pmap_release(struct pmap *pmap)
1607 vm_object_t object = pmap->pm_pteobj;
1608 struct rb_vm_page_scan_info info;
1610 KASSERT(pmap->pm_active == 0,
1611 ("pmap still active! %08x", pmap->pm_active));
1612 #if defined(DIAGNOSTIC)
1613 if (object->ref_count != 1)
1614 panic("pmap_release: pteobj reference count != 1");
1618 info.object = object;
1620 spin_lock(&pmap_spin);
1621 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1622 spin_unlock(&pmap_spin);
1624 vm_object_hold(object);
1625 /*lwkt_gettoken(&vm_token);*/
1629 info.limit = object->generation;
1631 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1632 pmap_release_callback, &info);
1633 if (info.error == 0 && info.mpte) {
1634 if (!pmap_release_free_page(pmap, info.mpte))
1637 } while (info.error);
1638 /*lwkt_reltoken(&vm_token);*/
1639 vm_object_drop(object);
1641 pmap->pm_cached = 0;
1645 * The caller must hold vm_token.
1648 pmap_release_callback(struct vm_page *p, void *data)
1650 struct rb_vm_page_scan_info *info = data;
1652 if (p->pindex == PTDPTDI) {
1656 if (!pmap_release_free_page(info->pmap, p)) {
1660 if (info->object->generation != info->limit) {
1668 * Grow the number of kernel page table entries, if needed.
1673 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1675 vm_offset_t addr = kend;
1677 vm_offset_t ptppaddr;
1681 vm_object_hold(kptobj);
1682 if (kernel_vm_end == 0) {
1683 kernel_vm_end = KERNBASE;
1685 while (pdir_pde(PTD, kernel_vm_end)) {
1686 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1687 ~(PAGE_SIZE * NPTEPG - 1);
1691 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1692 while (kernel_vm_end < addr) {
1693 if (pdir_pde(PTD, kernel_vm_end)) {
1694 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1695 ~(PAGE_SIZE * NPTEPG - 1);
1700 * This index is bogus, but out of the way
1702 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_NORMAL |
1704 VM_ALLOC_INTERRUPT);
1706 panic("pmap_growkernel: no memory to grow kernel");
1709 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1710 pmap_zero_page(ptppaddr);
1711 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1712 pdir_pde(PTD, kernel_vm_end) = newpdir;
1713 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1717 * This update must be interlocked with pmap_pinit2.
1719 spin_lock(&pmap_spin);
1720 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1721 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1723 spin_unlock(&pmap_spin);
1724 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1725 ~(PAGE_SIZE * NPTEPG - 1);
1727 vm_object_drop(kptobj);
1731 * Add a reference to the specified pmap.
1736 pmap_reference(pmap_t pmap)
1739 lwkt_gettoken(&vm_token);
1741 lwkt_reltoken(&vm_token);
1746 * vm_token must be held
1750 pmap_hold(pmap_t pmap)
1756 * vm_token must be held
1760 pmap_drop(pmap_t pmap)
1763 if (pmap->pm_count == (int)0x80000000)
1769 pmap_wait(pmap_t pmap, int count)
1771 lwkt_gettoken(&vm_token);
1772 pmap->pm_count += count;
1773 if (pmap->pm_count & 0x7FFFFFFF) {
1774 while (pmap->pm_count & 0x7FFFFFFF) {
1775 pmap->pm_count |= 0x80000000;
1776 tsleep(pmap, 0, "pmapd", 0);
1777 pmap->pm_count &= ~0x80000000;
1778 kprintf("pmap_wait: race averted\n");
1781 lwkt_reltoken(&vm_token);
1784 /***************************************************
1785 * page management routines.
1786 ***************************************************/
1789 * free the pv_entry back to the free list. This function may be
1790 * called from an interrupt.
1792 * The caller must hold vm_token.
1794 static PMAP_INLINE void
1795 free_pv_entry(pv_entry_t pv)
1797 struct mdglobaldata *gd;
1800 KKASSERT(pv->pv_m != NULL);
1805 if (gd->gd_freepv == NULL)
1812 * get a new pv_entry, allocating a block from the system
1813 * when needed. This function may be called from an interrupt thread.
1815 * THIS FUNCTION CAN BLOCK ON THE ZALLOC TOKEN, serialization of other
1816 * tokens (aka vm_token) to be temporarily lost.
1818 * The caller must hold vm_token.
1823 struct mdglobaldata *gd;
1827 if (pv_entry_high_water &&
1828 (pv_entry_count > pv_entry_high_water) &&
1829 (pmap_pagedaemon_waken == 0)) {
1830 pmap_pagedaemon_waken = 1;
1831 wakeup (&vm_pages_needed);
1834 if ((pv = gd->gd_freepv) != NULL)
1835 gd->gd_freepv = NULL;
1837 pv = zalloc(pvzone);
1842 * This routine is very drastic, but can save the system
1852 static int warningdone=0;
1854 if (pmap_pagedaemon_waken == 0)
1856 lwkt_gettoken(&vm_token);
1857 pmap_pagedaemon_waken = 0;
1859 if (warningdone < 5) {
1860 kprintf("pmap_collect: collecting pv entries -- "
1861 "suggest increasing PMAP_SHPGPERPROC\n");
1865 for (i = 0; i < vm_page_array_size; i++) {
1866 m = &vm_page_array[i];
1867 if (m->wire_count || m->hold_count)
1869 if (vm_page_busy_try(m, TRUE) == 0) {
1870 if (m->wire_count == 0 && m->hold_count == 0) {
1876 lwkt_reltoken(&vm_token);
1881 * Remove the pv entry and unwire the page table page related to the
1882 * pte the caller has cleared from the page table.
1884 * The caller must hold vm_token.
1887 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1888 vm_offset_t va, pmap_inval_info_t info)
1895 ASSERT_LWKT_TOKEN_HELD(&vm_token);
1896 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1897 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1898 if (pmap == pv->pv_pmap && va == pv->pv_va)
1902 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1904 KKASSERT(pv->pv_pmap == pmap);
1906 if (va == pv->pv_va)
1915 test_m_maps_pv(m, pv);
1916 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1917 m->md.pv_list_count--;
1919 atomic_add_int(&m->object->agg_pv_list_count, -1);
1920 if (TAILQ_EMPTY(&m->md.pv_list))
1921 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1922 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1923 ++pmap->pm_generation;
1928 vm_object_hold(pmap->pm_pteobj);
1929 pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1930 vm_object_drop(pmap->pm_pteobj);
1935 * Create a pv entry for page at pa for (pmap, va).
1937 * The caller must hold vm_token.
1940 pmap_insert_entry(pmap_t pmap, pv_entry_t pv, vm_offset_t va,
1941 vm_page_t mpte, vm_page_t m)
1944 KKASSERT(pv->pv_m == NULL);
1951 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1952 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1953 ++pmap->pm_generation;
1954 m->md.pv_list_count++;
1956 atomic_add_int(&m->object->agg_pv_list_count, 1);
1960 * pmap_remove_pte: do the things to unmap a page in a process.
1962 * The caller must hold vm_token.
1964 * WARNING! As with most other pmap functions this one can block, so
1965 * callers using temporary page table mappings must reload
1969 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1970 pmap_inval_info_t info)
1975 ptbase_assert(pmap);
1976 pmap_inval_interlock(info, pmap, va);
1977 ptbase_assert(pmap);
1978 oldpte = loadandclear(ptq);
1980 pmap->pm_stats.wired_count -= 1;
1981 pmap_inval_deinterlock(info, pmap);
1982 KKASSERT(oldpte & PG_V);
1984 * Machines that don't support invlpg, also don't support
1985 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1989 cpu_invlpg((void *)va);
1990 KKASSERT(pmap->pm_stats.resident_count > 0);
1991 --pmap->pm_stats.resident_count;
1992 if (oldpte & PG_MANAGED) {
1993 m = PHYS_TO_VM_PAGE(oldpte);
1994 if (oldpte & PG_M) {
1995 #if defined(PMAP_DIAGNOSTIC)
1996 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1997 kprintf("pmap_remove: modified page not "
1998 "writable: va: %p, pte: 0x%lx\n",
1999 (void *)va, (long)oldpte);
2002 if (pmap_track_modified(va))
2006 vm_page_flag_set(m, PG_REFERENCED);
2007 pmap_remove_entry(pmap, m, va, info);
2009 pmap_unuse_pt(pmap, va, NULL, info);
2014 * Remove a single page from a process address space.
2016 * The caller must hold vm_token.
2019 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2024 * If there is no pte for this address, just skip it!!! Otherwise
2025 * get a local va for mappings for this pmap and remove the entry.
2027 if (*pmap_pde(pmap, va) != 0) {
2028 ptq = get_ptbase(pmap) + i386_btop(va);
2030 pmap_remove_pte(pmap, ptq, va, info);
2037 * Remove the given range of addresses from the specified map.
2039 * It is assumed that the start and end are properly rounded to the page
2045 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2049 vm_offset_t ptpaddr;
2050 vm_offset_t sindex, eindex;
2051 struct pmap_inval_info info;
2056 vm_object_hold(pmap->pm_pteobj);
2057 lwkt_gettoken(&vm_token);
2058 if (pmap->pm_stats.resident_count == 0) {
2059 lwkt_reltoken(&vm_token);
2060 vm_object_drop(pmap->pm_pteobj);
2064 pmap_inval_init(&info);
2067 * special handling of removing one page. a very
2068 * common operation and easy to short circuit some
2071 if (((sva + PAGE_SIZE) == eva) &&
2072 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
2073 pmap_remove_page(pmap, sva, &info);
2074 pmap_inval_done(&info);
2075 lwkt_reltoken(&vm_token);
2076 vm_object_drop(pmap->pm_pteobj);
2081 * Get a local virtual address for the mappings that are being
2084 sindex = i386_btop(sva);
2085 eindex = i386_btop(eva);
2087 while (sindex < eindex) {
2091 * Stop scanning if no pages are left
2093 if (pmap->pm_stats.resident_count == 0)
2097 * Calculate index for next page table, limited by eindex.
2099 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2103 pdirindex = sindex / NPDEPG;
2104 ptpaddr = (unsigned)pmap->pm_pdir[pdirindex];
2105 if (ptpaddr & PG_PS) {
2106 pmap_inval_interlock(&info, pmap, -1);
2107 pmap->pm_pdir[pdirindex] = 0;
2108 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2109 pmap->pm_cached = 0;
2110 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.
2125 * Sub-scan the page table page. pmap_remove_pte() can
2126 * block on us, invalidating ptbase, so we must reload
2127 * ptbase and we must also check whether the page directory
2128 * page is still present.
2130 while (sindex < pdnxt) {
2133 ptbase = get_ptbase(pmap);
2134 if (ptbase[sindex]) {
2135 va = i386_ptob(sindex);
2136 pmap_remove_pte(pmap, ptbase + sindex,
2139 if (pmap->pm_pdir[pdirindex] == 0 ||
2140 (pmap->pm_pdir[pdirindex] & PG_PS)) {
2146 pmap_inval_done(&info);
2147 lwkt_reltoken(&vm_token);
2148 vm_object_drop(pmap->pm_pteobj);
2152 * Removes this physical page from all physical maps in which it resides.
2153 * Reflects back modify bits to the pager.
2155 * vm_token must be held by caller.
2158 pmap_remove_all(vm_page_t m)
2160 struct pmap_inval_info info;
2161 unsigned *pte, tpte;
2165 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2167 if (TAILQ_EMPTY(&m->md.pv_list))
2170 pmap_inval_init(&info);
2171 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2173 KKASSERT(pmap->pm_stats.resident_count > 0);
2175 vm_object_hold(pmap->pm_pteobj);
2177 if (pv != TAILQ_FIRST(&m->md.pv_list)) {
2178 vm_object_drop(pmap->pm_pteobj);
2183 --pmap->pm_stats.resident_count;
2184 pte = pmap_pte_quick(pmap, pv->pv_va);
2185 pmap_inval_interlock(&info, pmap, pv->pv_va);
2186 tpte = loadandclear(pte);
2188 pmap->pm_stats.wired_count--;
2189 pmap_inval_deinterlock(&info, pmap);
2191 vm_page_flag_set(m, PG_REFERENCED);
2192 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
2195 * Update the vm_page_t clean and reference bits.
2198 #if defined(PMAP_DIAGNOSTIC)
2199 if (pmap_nw_modified((pt_entry_t) tpte)) {
2200 kprintf("pmap_remove_all: modified page "
2201 "not writable: va: %p, pte: 0x%lx\n",
2202 (void *)pv->pv_va, (long)tpte);
2205 if (pmap_track_modified(pv->pv_va))
2209 KKASSERT(pv->pv_m == m);
2211 KKASSERT(pv == TAILQ_FIRST(&m->md.pv_list));
2212 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2213 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2214 ++pmap->pm_generation;
2215 m->md.pv_list_count--;
2217 atomic_add_int(&m->object->agg_pv_list_count, -1);
2218 if (TAILQ_EMPTY(&m->md.pv_list))
2219 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2220 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2221 vm_object_drop(pmap->pm_pteobj);
2225 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2226 pmap_inval_done(&info);
2230 * Set the physical protection on the specified range of this map
2236 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2239 vm_offset_t pdnxt, ptpaddr;
2240 vm_pindex_t sindex, eindex;
2241 pmap_inval_info info;
2246 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2247 pmap_remove(pmap, sva, eva);
2251 if (prot & VM_PROT_WRITE)
2254 lwkt_gettoken(&vm_token);
2255 pmap_inval_init(&info);
2257 ptbase = get_ptbase(pmap);
2259 sindex = i386_btop(sva);
2260 eindex = i386_btop(eva);
2262 for (; sindex < eindex; sindex = pdnxt) {
2265 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2267 pdirindex = sindex / NPDEPG;
2268 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
2269 pmap_inval_interlock(&info, pmap, -1);
2270 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2271 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2272 pmap_inval_deinterlock(&info, pmap);
2277 * Weed out invalid mappings. Note: we assume that the page
2278 * directory table is always allocated, and in kernel virtual.
2283 if (pdnxt > eindex) {
2287 for (; sindex != pdnxt; sindex++) {
2295 pmap_inval_interlock(&info, pmap, i386_ptob(sindex));
2297 pbits = ptbase[sindex];
2300 if (pbits & PG_MANAGED) {
2303 m = PHYS_TO_VM_PAGE(pbits);
2304 vm_page_flag_set(m, PG_REFERENCED);
2308 if (pmap_track_modified(i386_ptob(sindex))) {
2310 m = PHYS_TO_VM_PAGE(pbits);
2317 if (pbits != cbits &&
2318 !atomic_cmpset_int(ptbase + sindex, pbits, cbits)) {
2321 pmap_inval_deinterlock(&info, pmap);
2324 pmap_inval_done(&info);
2325 lwkt_reltoken(&vm_token);
2329 * Insert the given physical page (p) at the specified virtual address (v)
2330 * in the target physical map with the protection requested.
2332 * If specified, the page will be wired down, meaning that the related pte
2333 * cannot be reclaimed.
2338 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2339 boolean_t wired, vm_map_entry_t entry __unused)
2344 vm_offset_t origpte, newpte;
2346 pmap_inval_info info;
2353 #ifdef PMAP_DIAGNOSTIC
2355 panic("pmap_enter: toobig");
2356 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2357 panic("pmap_enter: invalid to pmap_enter page "
2358 "table pages (va: %p)", (void *)va);
2361 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2362 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2363 print_backtrace(-1);
2365 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2366 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2367 print_backtrace(-1);
2370 vm_object_hold(pmap->pm_pteobj);
2371 lwkt_gettoken(&vm_token);
2374 * This can block, get it before we do anything important.
2376 if (pmap_initialized &&
2377 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2378 pv = get_pv_entry();
2384 * In the case that a page table page is not
2385 * resident, we are creating it here.
2387 if (va < UPT_MIN_ADDRESS)
2388 mpte = pmap_allocpte(pmap, va);
2392 if ((prot & VM_PROT_NOSYNC) == 0)
2393 pmap_inval_init(&info);
2394 pte = pmap_pte(pmap, va);
2397 * Page Directory table entry not valid, we need a new PT page
2400 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p",
2401 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
2404 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2405 origpte = *(vm_offset_t *)pte;
2406 opa = origpte & PG_FRAME;
2408 if (origpte & PG_PS)
2409 panic("pmap_enter: attempted pmap_enter on 4MB page");
2412 * Mapping has not changed, must be protection or wiring change.
2414 if (origpte && (opa == pa)) {
2416 * Wiring change, just update stats. We don't worry about
2417 * wiring PT pages as they remain resident as long as there
2418 * are valid mappings in them. Hence, if a user page is wired,
2419 * the PT page will be also.
2421 if (wired && ((origpte & PG_W) == 0))
2422 pmap->pm_stats.wired_count++;
2423 else if (!wired && (origpte & PG_W))
2424 pmap->pm_stats.wired_count--;
2426 #if defined(PMAP_DIAGNOSTIC)
2427 if (pmap_nw_modified((pt_entry_t) origpte)) {
2428 kprintf("pmap_enter: modified page not "
2429 "writable: va: %p, pte: 0x%lx\n",
2430 (void *)va, (long )origpte);
2435 * We might be turning off write access to the page,
2436 * so we go ahead and sense modify status.
2438 if (origpte & PG_MANAGED) {
2439 if ((origpte & PG_M) && pmap_track_modified(va)) {
2441 om = PHYS_TO_VM_PAGE(opa);
2445 KKASSERT(m->flags & PG_MAPPED);
2450 * Mapping has changed, invalidate old range and fall through to
2451 * handle validating new mapping.
2453 * Since we have a ref on the page directory page pmap_pte()
2454 * will always return non-NULL.
2456 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2457 * to get wiped. reload the ptbase. I'm not sure if it is
2458 * also possible to race another pmap_enter() but check for
2462 KKASSERT((origpte & PG_FRAME) ==
2463 (*(vm_offset_t *)pte & PG_FRAME));
2464 if (prot & VM_PROT_NOSYNC) {
2465 prot &= ~VM_PROT_NOSYNC;
2466 pmap_inval_init(&info);
2468 pmap_remove_pte(pmap, pte, va, &info);
2469 pte = pmap_pte(pmap, va);
2470 origpte = *(vm_offset_t *)pte;
2471 opa = origpte & PG_FRAME;
2473 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2479 * Enter on the PV list if part of our managed memory. Note that we
2480 * raise IPL while manipulating pv_table since pmap_enter can be
2481 * called at interrupt time.
2483 if (pmap_initialized &&
2484 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2485 pmap_insert_entry(pmap, pv, va, mpte, m);
2487 ptbase_assert(pmap);
2489 vm_page_flag_set(m, PG_MAPPED);
2493 * Increment counters
2495 ++pmap->pm_stats.resident_count;
2497 pmap->pm_stats.wired_count++;
2498 KKASSERT(*pte == 0);
2502 * Now validate mapping with desired protection/wiring.
2504 ptbase_assert(pmap);
2505 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2509 if (va < UPT_MIN_ADDRESS)
2511 if (pmap == &kernel_pmap)
2513 newpte |= pat_pte_index[m->pat_mode];
2516 * If the mapping or permission bits are different, we need
2517 * to update the pte. If the pte is already present we have
2518 * to get rid of the extra wire-count on mpte we had obtained
2521 * mpte has a new wire_count, which also serves to prevent the
2522 * page table page from getting ripped out while we work. If we
2523 * are modifying an existing pte instead of installing a new one
2524 * we have to drop it.
2526 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2527 if (prot & VM_PROT_NOSYNC)
2528 cpu_invlpg((void *)va);
2530 pmap_inval_interlock(&info, pmap, va);
2531 ptbase_assert(pmap);
2534 KKASSERT((*pte & PG_FRAME) == (newpte & PG_FRAME));
2535 if (mpte && vm_page_unwire_quick(mpte))
2536 panic("pmap_enter: Insufficient wire_count");
2539 *pte = newpte | PG_A;
2540 if ((prot & VM_PROT_NOSYNC) == 0)
2541 pmap_inval_deinterlock(&info, pmap);
2543 vm_page_flag_set(m, PG_WRITEABLE);
2546 KKASSERT((*pte & PG_FRAME) == (newpte & PG_FRAME));
2547 if (mpte && vm_page_unwire_quick(mpte))
2548 panic("pmap_enter: Insufficient wire_count");
2553 * NOTE: mpte invalid after this point if we block.
2555 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2556 if ((prot & VM_PROT_NOSYNC) == 0)
2557 pmap_inval_done(&info);
2560 lwkt_reltoken(&vm_token);
2561 vm_object_drop(pmap->pm_pteobj);
2565 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2566 * This code also assumes that the pmap has no pre-existing entry for this
2569 * This code currently may only be used on user pmaps, not kernel_pmap.
2574 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2581 pmap_inval_info info;
2584 vm_object_hold(pmap->pm_pteobj);
2585 lwkt_gettoken(&vm_token);
2588 * This can block, get it before we do anything important.
2590 if (pmap_initialized &&
2591 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2592 pv = get_pv_entry();
2597 pmap_inval_init(&info);
2599 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2600 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2601 print_backtrace(-1);
2603 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2604 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2605 print_backtrace(-1);
2608 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2611 * Calculate the page table page (mpte), allocating it if necessary.
2613 * A held page table page (mpte), or NULL, is passed onto the
2614 * section following.
2616 if (va < UPT_MIN_ADDRESS) {
2618 * Calculate pagetable page index
2620 ptepindex = va >> PDRSHIFT;
2624 * Get the page directory entry
2626 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2629 * If the page table page is mapped, we just increment
2630 * the wire count, and activate it.
2634 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2635 if ((mpte = pmap->pm_ptphint) != NULL &&
2636 (mpte->pindex == ptepindex) &&
2637 (mpte->flags & PG_BUSY) == 0) {
2638 vm_page_wire_quick(mpte);
2640 mpte = pmap_page_lookup(pmap->pm_pteobj,
2642 pmap->pm_ptphint = mpte;
2643 vm_page_wire_quick(mpte);
2644 vm_page_wakeup(mpte);
2647 mpte = _pmap_allocpte(pmap, ptepindex);
2649 } while (mpte == NULL);
2652 /* this code path is not yet used */
2656 * With a valid (and held) page directory page, we can just use
2657 * vtopte() to get to the pte. If the pte is already present
2658 * we do not disturb it.
2660 pte = (unsigned *)vtopte(va);
2662 KKASSERT(*pte & PG_V);
2663 pa = VM_PAGE_TO_PHYS(m);
2664 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2665 pmap_inval_done(&info);
2667 pmap_unwire_pte(pmap, mpte, &info);
2672 lwkt_reltoken(&vm_token);
2673 vm_object_drop(pmap->pm_pteobj);
2678 * Enter on the PV list if part of our managed memory
2680 if (pmap_initialized &&
2681 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2682 pmap_insert_entry(pmap, pv, va, mpte, m);
2684 vm_page_flag_set(m, PG_MAPPED);
2688 * Increment counters
2690 ++pmap->pm_stats.resident_count;
2692 pa = VM_PAGE_TO_PHYS(m);
2695 * Now validate mapping with RO protection
2697 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2698 *pte = pa | PG_V | PG_U;
2700 *pte = pa | PG_V | PG_U | PG_MANAGED;
2701 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2702 pmap_inval_done(&info);
2707 lwkt_reltoken(&vm_token);
2708 vm_object_drop(pmap->pm_pteobj);
2712 * Make a temporary mapping for a physical address. This is only intended
2713 * to be used for panic dumps.
2715 * The caller is responsible for calling smp_invltlb().
2720 pmap_kenter_temporary(vm_paddr_t pa, long i)
2722 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2723 return ((void *)crashdumpmap);
2726 #define MAX_INIT_PT (96)
2729 * This routine preloads the ptes for a given object into the specified pmap.
2730 * This eliminates the blast of soft faults on process startup and
2731 * immediately after an mmap.
2735 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2738 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2739 vm_object_t object, vm_pindex_t pindex,
2740 vm_size_t size, int limit)
2742 struct rb_vm_page_scan_info info;
2747 * We can't preinit if read access isn't set or there is no pmap
2750 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2754 * We can't preinit if the pmap is not the current pmap
2756 lp = curthread->td_lwp;
2757 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2760 psize = i386_btop(size);
2762 if ((object->type != OBJT_VNODE) ||
2763 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2764 (object->resident_page_count > MAX_INIT_PT))) {
2768 if (psize + pindex > object->size) {
2769 if (object->size < pindex)
2771 psize = object->size - pindex;
2778 * Use a red-black scan to traverse the requested range and load
2779 * any valid pages found into the pmap.
2781 * We cannot safely scan the object's memq unless we are in a
2782 * critical section since interrupts can remove pages from objects.
2784 info.start_pindex = pindex;
2785 info.end_pindex = pindex + psize - 1;
2791 vm_object_hold_shared(object);
2792 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2793 pmap_object_init_pt_callback, &info);
2794 vm_object_drop(object);
2798 * The caller must hold vm_token.
2802 pmap_object_init_pt_callback(vm_page_t p, void *data)
2804 struct rb_vm_page_scan_info *info = data;
2805 vm_pindex_t rel_index;
2807 * don't allow an madvise to blow away our really
2808 * free pages allocating pv entries.
2810 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2811 vmstats.v_free_count < vmstats.v_free_reserved) {
2816 * Ignore list markers and ignore pages we cannot instantly
2817 * busy (while holding the object token).
2819 if (p->flags & PG_MARKER)
2821 if (vm_page_busy_try(p, TRUE))
2823 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2824 (p->flags & PG_FICTITIOUS) == 0) {
2825 if ((p->queue - p->pc) == PQ_CACHE)
2826 vm_page_deactivate(p);
2827 rel_index = p->pindex - info->start_pindex;
2828 pmap_enter_quick(info->pmap,
2829 info->addr + i386_ptob(rel_index), p);
2836 * Return TRUE if the pmap is in shape to trivially
2837 * pre-fault the specified address.
2839 * Returns FALSE if it would be non-trivial or if a
2840 * pte is already loaded into the slot.
2845 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2850 lwkt_gettoken(&vm_token);
2851 if ((*pmap_pde(pmap, addr)) == 0) {
2854 pte = (unsigned *) vtopte(addr);
2855 ret = (*pte) ? 0 : 1;
2857 lwkt_reltoken(&vm_token);
2862 * Change the wiring attribute for a map/virtual-adderss pair. The mapping
2863 * must already exist.
2868 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
2869 vm_map_entry_t entry __unused)
2876 lwkt_gettoken(&vm_token);
2877 pte = pmap_pte(pmap, va);
2879 if (wired && !pmap_pte_w(pte))
2880 pmap->pm_stats.wired_count++;
2881 else if (!wired && pmap_pte_w(pte))
2882 pmap->pm_stats.wired_count--;
2885 * Wiring is not a hardware characteristic so there is no need to
2886 * invalidate TLB. However, in an SMP environment we must use
2887 * a locked bus cycle to update the pte (if we are not using
2888 * the pmap_inval_*() API that is)... it's ok to do this for simple
2892 atomic_set_int(pte, PG_W);
2894 atomic_clear_int(pte, PG_W);
2895 lwkt_reltoken(&vm_token);
2899 * Copy the range specified by src_addr/len from the source map to the
2900 * range dst_addr/len in the destination map.
2902 * This routine is only advisory and need not do anything.
2907 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2908 vm_size_t len, vm_offset_t src_addr)
2914 * Zero the specified PA by mapping the page into KVM and clearing its
2920 pmap_zero_page(vm_paddr_t phys)
2922 struct mdglobaldata *gd = mdcpu;
2925 if (*(int *)gd->gd_CMAP3)
2926 panic("pmap_zero_page: CMAP3 busy");
2927 *(int *)gd->gd_CMAP3 =
2928 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2929 cpu_invlpg(gd->gd_CADDR3);
2930 bzero(gd->gd_CADDR3, PAGE_SIZE);
2931 *(int *) gd->gd_CMAP3 = 0;
2936 * Assert that a page is empty, panic if it isn't.
2941 pmap_page_assertzero(vm_paddr_t phys)
2943 struct mdglobaldata *gd = mdcpu;
2947 if (*(int *)gd->gd_CMAP3)
2948 panic("pmap_zero_page: CMAP3 busy");
2949 *(int *)gd->gd_CMAP3 =
2950 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2951 cpu_invlpg(gd->gd_CADDR3);
2952 for (i = 0; i < PAGE_SIZE; i += 4) {
2953 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2954 panic("pmap_page_assertzero() @ %p not zero!",
2955 (void *)gd->gd_CADDR3);
2958 *(int *) gd->gd_CMAP3 = 0;
2963 * Zero part of a physical page by mapping it into memory and clearing
2964 * its contents with bzero.
2966 * off and size may not cover an area beyond a single hardware page.
2971 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2973 struct mdglobaldata *gd = mdcpu;
2976 if (*(int *) gd->gd_CMAP3)
2977 panic("pmap_zero_page: CMAP3 busy");
2978 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2979 cpu_invlpg(gd->gd_CADDR3);
2980 bzero((char *)gd->gd_CADDR3 + off, size);
2981 *(int *) gd->gd_CMAP3 = 0;
2986 * Copy the physical page from the source PA to the target PA.
2987 * This function may be called from an interrupt. No locking
2993 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2995 struct mdglobaldata *gd = mdcpu;
2998 if (*(int *) gd->gd_CMAP1)
2999 panic("pmap_copy_page: CMAP1 busy");
3000 if (*(int *) gd->gd_CMAP2)
3001 panic("pmap_copy_page: CMAP2 busy");
3003 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
3004 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
3006 cpu_invlpg(gd->gd_CADDR1);
3007 cpu_invlpg(gd->gd_CADDR2);
3009 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
3011 *(int *) gd->gd_CMAP1 = 0;
3012 *(int *) gd->gd_CMAP2 = 0;
3017 * Copy the physical page from the source PA to the target PA.
3018 * This function may be called from an interrupt. No locking
3024 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3026 struct mdglobaldata *gd = mdcpu;
3029 if (*(int *) gd->gd_CMAP1)
3030 panic("pmap_copy_page: CMAP1 busy");
3031 if (*(int *) gd->gd_CMAP2)
3032 panic("pmap_copy_page: CMAP2 busy");
3034 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
3035 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
3037 cpu_invlpg(gd->gd_CADDR1);
3038 cpu_invlpg(gd->gd_CADDR2);
3040 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
3041 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
3044 *(int *) gd->gd_CMAP1 = 0;
3045 *(int *) gd->gd_CMAP2 = 0;
3050 * Returns true if the pmap's pv is one of the first
3051 * 16 pvs linked to from this page. This count may
3052 * be changed upwards or downwards in the future; it
3053 * is only necessary that true be returned for a small
3054 * subset of pmaps for proper page aging.
3059 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3064 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3067 lwkt_gettoken(&vm_token);
3068 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3069 if (pv->pv_pmap == pmap) {
3070 lwkt_reltoken(&vm_token);
3077 lwkt_reltoken(&vm_token);
3082 * Remove all pages from specified address space
3083 * this aids process exit speeds. Also, this code
3084 * is special cased for current process only, but
3085 * can have the more generic (and slightly slower)
3086 * mode enabled. This is much faster than pmap_remove
3087 * in the case of running down an entire address space.
3092 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3095 unsigned *pte, tpte;
3098 pmap_inval_info info;
3100 int32_t save_generation;
3102 lp = curthread->td_lwp;
3103 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3108 if (pmap->pm_pteobj)
3109 vm_object_hold(pmap->pm_pteobj);
3110 lwkt_gettoken(&vm_token);
3111 pmap_inval_init(&info);
3113 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3114 if (pv->pv_va >= eva || pv->pv_va < sva) {
3115 npv = TAILQ_NEXT(pv, pv_plist);
3119 KKASSERT(pmap == pv->pv_pmap);
3122 pte = (unsigned *)vtopte(pv->pv_va);
3124 pte = pmap_pte_quick(pmap, pv->pv_va);
3126 pmap_inval_interlock(&info, pmap, pv->pv_va);
3129 * We cannot remove wired pages from a process' mapping
3133 pmap_inval_deinterlock(&info, pmap);
3134 npv = TAILQ_NEXT(pv, pv_plist);
3138 tpte = loadandclear(pte);
3139 pmap_inval_deinterlock(&info, pmap);
3141 m = PHYS_TO_VM_PAGE(tpte);
3142 test_m_maps_pv(m, pv);
3144 KASSERT(m < &vm_page_array[vm_page_array_size],
3145 ("pmap_remove_pages: bad tpte %x", tpte));
3147 KKASSERT(pmap->pm_stats.resident_count > 0);
3148 --pmap->pm_stats.resident_count;
3151 * Update the vm_page_t clean and reference bits.
3157 npv = TAILQ_NEXT(pv, pv_plist);
3159 KKASSERT(pv->pv_m == m);
3160 KKASSERT(pv->pv_pmap == pmap);
3162 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3163 save_generation = ++pmap->pm_generation;
3165 m->md.pv_list_count--;
3167 atomic_add_int(&m->object->agg_pv_list_count, -1);
3168 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3169 if (TAILQ_EMPTY(&m->md.pv_list))
3170 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3172 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3176 * Restart the scan if we blocked during the unuse or free
3177 * calls and other removals were made.
3179 if (save_generation != pmap->pm_generation) {
3180 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3181 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3184 pmap_inval_done(&info);
3185 lwkt_reltoken(&vm_token);
3186 if (pmap->pm_pteobj)
3187 vm_object_drop(pmap->pm_pteobj);
3191 * pmap_testbit tests bits in pte's
3192 * note that the testbit/clearbit routines are inline,
3193 * and a lot of things compile-time evaluate.
3195 * The caller must hold vm_token.
3198 pmap_testbit(vm_page_t m, int bit)
3203 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3206 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3209 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3211 * if the bit being tested is the modified bit, then
3212 * mark clean_map and ptes as never
3215 if (bit & (PG_A|PG_M)) {
3216 if (!pmap_track_modified(pv->pv_va))
3220 #if defined(PMAP_DIAGNOSTIC)
3222 kprintf("Null pmap (tb) at va: %p\n",
3227 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3236 * This routine is used to modify bits in ptes
3238 * The caller must hold vm_token.
3240 static __inline void
3241 pmap_clearbit(vm_page_t m, int bit)
3243 struct pmap_inval_info info;
3248 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3251 pmap_inval_init(&info);
3254 * Loop over all current mappings setting/clearing as appropos If
3255 * setting RO do we need to clear the VAC?
3257 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3259 * don't write protect pager mappings
3262 if (!pmap_track_modified(pv->pv_va))
3266 #if defined(PMAP_DIAGNOSTIC)
3268 kprintf("Null pmap (cb) at va: %p\n",
3275 * Careful here. We can use a locked bus instruction to
3276 * clear PG_A or PG_M safely but we need to synchronize
3277 * with the target cpus when we mess with PG_RW.
3279 * We do not have to force synchronization when clearing
3280 * PG_M even for PTEs generated via virtual memory maps,
3281 * because the virtual kernel will invalidate the pmap
3282 * entry when/if it needs to resynchronize the Modify bit.
3285 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3286 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3293 atomic_clear_int(pte, PG_M|PG_RW);
3296 * The cpu may be trying to set PG_M
3297 * simultaniously with our clearing
3300 if (!atomic_cmpset_int(pte, pbits,
3304 } else if (bit == PG_M) {
3306 * We could also clear PG_RW here to force
3307 * a fault on write to redetect PG_M for
3308 * virtual kernels, but it isn't necessary
3309 * since virtual kernels invalidate the pte
3310 * when they clear the VPTE_M bit in their
3311 * virtual page tables.
3313 atomic_clear_int(pte, PG_M);
3315 atomic_clear_int(pte, bit);
3319 pmap_inval_deinterlock(&info, pv->pv_pmap);
3321 pmap_inval_done(&info);
3325 * Lower the permission for all mappings to a given page.
3330 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3332 if ((prot & VM_PROT_WRITE) == 0) {
3333 lwkt_gettoken(&vm_token);
3334 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3335 pmap_clearbit(m, PG_RW);
3336 vm_page_flag_clear(m, PG_WRITEABLE);
3340 lwkt_reltoken(&vm_token);
3345 * Return the physical address given a physical page index.
3350 pmap_phys_address(vm_pindex_t ppn)
3352 return (i386_ptob(ppn));
3356 * Return a count of reference bits for a page, clearing those bits.
3357 * It is not necessary for every reference bit to be cleared, but it
3358 * is necessary that 0 only be returned when there are truly no
3359 * reference bits set.
3364 pmap_ts_referenced(vm_page_t m)
3366 pv_entry_t pv, pvf, pvn;
3370 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3373 lwkt_gettoken(&vm_token);
3375 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3380 pvn = TAILQ_NEXT(pv, pv_list);
3382 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3383 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3385 if (!pmap_track_modified(pv->pv_va))
3388 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3390 if (pte && (*pte & PG_A)) {
3391 atomic_clear_int(pte, PG_A);
3397 } while ((pv = pvn) != NULL && pv != pvf);
3400 lwkt_reltoken(&vm_token);
3406 * Return whether or not the specified physical page was modified
3407 * in any physical maps.
3412 pmap_is_modified(vm_page_t m)
3416 lwkt_gettoken(&vm_token);
3417 res = pmap_testbit(m, PG_M);
3418 lwkt_reltoken(&vm_token);
3423 * Clear the modify bits on the specified physical page.
3428 pmap_clear_modify(vm_page_t m)
3430 lwkt_gettoken(&vm_token);
3431 pmap_clearbit(m, PG_M);
3432 lwkt_reltoken(&vm_token);
3436 * Clear the reference bit on the specified physical page.
3441 pmap_clear_reference(vm_page_t m)
3443 lwkt_gettoken(&vm_token);
3444 pmap_clearbit(m, PG_A);
3445 lwkt_reltoken(&vm_token);
3449 * Miscellaneous support routines follow
3451 * Called from the low level boot code only.
3454 i386_protection_init(void)
3458 kp = protection_codes;
3459 for (prot = 0; prot < 8; prot++) {
3461 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3463 * Read access is also 0. There isn't any execute bit,
3464 * so just make it readable.
3466 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3467 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3468 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3471 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3472 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3473 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3474 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3482 * Map a set of physical memory pages into the kernel virtual
3483 * address space. Return a pointer to where it is mapped. This
3484 * routine is intended to be used for mapping device memory,
3487 * NOTE: We can't use pgeflag unless we invalidate the pages one at
3490 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
3491 * work whether the cpu supports PAT or not. The remaining PAT
3492 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
3496 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3498 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
3502 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3504 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
3508 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
3510 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
3514 * Map a set of physical memory pages into the kernel virtual
3515 * address space. Return a pointer to where it is mapped. This
3516 * routine is intended to be used for mapping device memory,
3520 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
3522 vm_offset_t va, tmpva, offset;
3526 offset = pa & PAGE_MASK;
3527 size = roundup(offset + size, PAGE_SIZE);
3529 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3531 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3533 pa = pa & ~PAGE_MASK;
3534 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
3535 pte = vtopte(tmpva);
3536 *pte = pa | PG_RW | PG_V | /* pgeflag | */
3537 pat_pte_index[mode];
3538 tmpsize -= PAGE_SIZE;
3542 pmap_invalidate_range(&kernel_pmap, va, va + size);
3543 pmap_invalidate_cache_range(va, va + size);
3545 return ((void *)(va + offset));
3552 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3554 vm_offset_t base, offset;
3556 base = va & PG_FRAME;
3557 offset = va & PAGE_MASK;
3558 size = roundup(offset + size, PAGE_SIZE);
3559 pmap_qremove(va, size >> PAGE_SHIFT);
3560 kmem_free(&kernel_map, base, size);
3564 * Sets the memory attribute for the specified page.
3567 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
3574 * The following code is NOP, until we get pmap_change_attr()
3579 * If "m" is a normal page, update its direct mapping. This update
3580 * can be relied upon to perform any cache operations that are
3581 * required for data coherence.
3583 if ((m->flags & PG_FICTITIOUS) == 0)
3584 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE,
3590 * Change the PAT attribute on an existing kernel memory map. Caller
3591 * must ensure that the virtual memory in question is not accessed
3592 * during the adjustment.
3595 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
3602 panic("pmap_change_attr: va is NULL");
3603 base = trunc_page(va);
3607 *pte = (*pte & ~(pt_entry_t)(PG_PTE_PAT | PG_NC_PCD |
3609 pat_pte_index[mode];
3614 changed = 1; /* XXX: not optimal */
3617 * Flush CPU caches if required to make sure any data isn't cached that
3618 * shouldn't be, etc.
3621 pmap_invalidate_range(&kernel_pmap, base, va);
3622 pmap_invalidate_cache_range(base, va);
3627 * Perform the pmap work for mincore
3629 * The caller must hold vm_token if the caller wishes a stable result,
3630 * and even in that case some bits can change due to third party accesses
3636 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3638 unsigned *ptep, pte;
3642 lwkt_gettoken(&vm_token);
3643 ptep = pmap_pte(pmap, addr);
3645 if (ptep && (pte = *ptep) != 0) {
3648 val = MINCORE_INCORE;
3649 if ((pte & PG_MANAGED) == 0)
3652 pa = pte & PG_FRAME;
3654 m = PHYS_TO_VM_PAGE(pa);
3660 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3661 } else if (m->dirty || pmap_is_modified(m)) {
3663 * Modified by someone else
3665 val |= MINCORE_MODIFIED_OTHER;
3672 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3673 } else if ((m->flags & PG_REFERENCED) ||
3674 pmap_ts_referenced(m)) {
3676 * Referenced by someone else
3678 val |= MINCORE_REFERENCED_OTHER;
3679 vm_page_flag_set(m, PG_REFERENCED);
3683 lwkt_reltoken(&vm_token);
3688 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3689 * vmspace will be ref'd and the old one will be deref'd.
3691 * cr3 will be reloaded if any lwp is the current lwp.
3693 * Only called with new VM spaces.
3694 * The process must have only a single thread.
3695 * The process must hold the vmspace->vm_map.token for oldvm and newvm
3696 * No other requirements.
3699 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3701 struct vmspace *oldvm;
3704 oldvm = p->p_vmspace;
3705 if (oldvm != newvm) {
3707 sysref_get(&newvm->vm_sysref);
3708 p->p_vmspace = newvm;
3709 KKASSERT(p->p_nthreads == 1);
3710 lp = RB_ROOT(&p->p_lwp_tree);
3711 pmap_setlwpvm(lp, newvm);
3713 sysref_put(&oldvm->vm_sysref);
3718 * Set the vmspace for a LWP. The vmspace is almost universally set the
3719 * same as the process vmspace, but virtual kernels need to swap out contexts
3720 * on a per-lwp basis.
3722 * Always called with a lp under the caller's direct control, either
3723 * unscheduled or the current lwp.
3728 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3730 struct vmspace *oldvm;
3733 oldvm = lp->lwp_vmspace;
3735 if (oldvm != newvm) {
3736 lp->lwp_vmspace = newvm;
3737 if (curthread->td_lwp == lp) {
3738 pmap = vmspace_pmap(newvm);
3739 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3740 if (pmap->pm_active & CPUMASK_LOCK)
3741 pmap_interlock_wait(newvm);
3742 #if defined(SWTCH_OPTIM_STATS)
3745 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3746 load_cr3(curthread->td_pcb->pcb_cr3);
3747 pmap = vmspace_pmap(oldvm);
3748 atomic_clear_cpumask(&pmap->pm_active,
3755 * Called when switching to a locked pmap, used to interlock against pmaps
3756 * undergoing modifications to prevent us from activating the MMU for the
3757 * target pmap until all such modifications have completed. We have to do
3758 * this because the thread making the modifications has already set up its
3759 * SMP synchronization mask.
3764 pmap_interlock_wait(struct vmspace *vm)
3766 struct pmap *pmap = &vm->vm_pmap;
3768 if (pmap->pm_active & CPUMASK_LOCK) {
3770 DEBUG_PUSH_INFO("pmap_interlock_wait");
3771 while (pmap->pm_active & CPUMASK_LOCK) {
3773 lwkt_process_ipiq();
3781 * Return a page-directory alignment hint for device mappings which will
3782 * allow the use of super-pages for the mapping.
3787 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3790 if ((obj == NULL) || (size < NBPDR) ||
3791 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
3795 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3800 * Return whether the PGE flag is supported globally.
3805 pmap_get_pgeflag(void)
3811 * Used by kmalloc/kfree, page already exists at va
3814 pmap_kvtom(vm_offset_t va)
3816 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));
3820 pmap_object_init(vm_object_t object)
3826 pmap_object_free(vm_object_t object)