2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
9 * Copyright (c) 2011 Matthew Dillon
10 * All rights reserved.
12 * This code is derived from software contributed to Berkeley by
13 * the Systems Programming Group of the University of Utah Computer
14 * Science Department and William Jolitz of UUNET Technologies Inc.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * Manage physical address maps for x86-64 systems.
49 #include "opt_disable_pse.h"
52 #include "opt_msgbuf.h"
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/kernel.h>
58 #include <sys/msgbuf.h>
59 #include <sys/vmmeter.h>
63 #include <vm/vm_param.h>
64 #include <sys/sysctl.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_pageout.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_zone.h>
76 #include <sys/thread2.h>
77 #include <sys/sysref2.h>
78 #include <sys/spinlock2.h>
79 #include <vm/vm_page2.h>
81 #include <machine/cputypes.h>
82 #include <machine/md_var.h>
83 #include <machine/specialreg.h>
84 #include <machine/smp.h>
85 #include <machine_base/apic/apicreg.h>
86 #include <machine/globaldata.h>
87 #include <machine/pmap.h>
88 #include <machine/pmap_inval.h>
89 #include <machine/inttypes.h>
93 #define PMAP_KEEP_PDIRS
94 #ifndef PMAP_SHPGPERPROC
95 #define PMAP_SHPGPERPROC 200
98 #if defined(DIAGNOSTIC)
99 #define PMAP_DIAGNOSTIC
105 * pmap debugging will report who owns a pv lock when blocking.
109 #define PMAP_DEBUG_DECL ,const char *func, int lineno
110 #define PMAP_DEBUG_ARGS , __func__, __LINE__
111 #define PMAP_DEBUG_COPY , func, lineno
113 #define pv_get(pmap, pindex) _pv_get(pmap, pindex \
115 #define pv_lock(pv) _pv_lock(pv \
117 #define pv_hold_try(pv) _pv_hold_try(pv \
119 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \
124 #define PMAP_DEBUG_DECL
125 #define PMAP_DEBUG_ARGS
126 #define PMAP_DEBUG_COPY
128 #define pv_get(pmap, pindex) _pv_get(pmap, pindex)
129 #define pv_lock(pv) _pv_lock(pv)
130 #define pv_hold_try(pv) _pv_hold_try(pv)
131 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp)
136 * Get PDEs and PTEs for user/kernel address space
138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
140 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
141 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
142 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
143 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
144 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
147 * Given a map and a machine independent protection code,
148 * convert to a vax protection code.
150 #define pte_prot(m, p) \
151 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
152 static int protection_codes[8];
154 struct pmap kernel_pmap;
155 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
157 vm_paddr_t avail_start; /* PA of first available physical page */
158 vm_paddr_t avail_end; /* PA of last available physical page */
159 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
160 vm_offset_t virtual2_end;
161 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
162 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
163 vm_offset_t KvaStart; /* VA start of KVA space */
164 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
165 vm_offset_t KvaSize; /* max size of kernel virtual address space */
166 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
167 static int pgeflag; /* PG_G or-in */
168 static int pseflag; /* PG_PS or-in */
171 static vm_paddr_t dmaplimit;
173 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
175 static uint64_t KPTbase;
176 static uint64_t KPTphys;
177 static uint64_t KPDphys; /* phys addr of kernel level 2 */
178 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
179 uint64_t KPDPphys; /* phys addr of kernel level 3 */
180 uint64_t KPML4phys; /* phys addr of kernel level 4 */
182 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
183 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
186 * Data for the pv entry allocation mechanism
188 static vm_zone_t pvzone;
189 static struct vm_zone pvzone_store;
190 static struct vm_object pvzone_obj;
191 static int pv_entry_max=0, pv_entry_high_water=0;
192 static int pmap_pagedaemon_waken = 0;
193 static struct pv_entry *pvinit;
196 * All those kernel PT submaps that BSD is so fond of
198 pt_entry_t *CMAP1 = 0, *ptmmap;
199 caddr_t CADDR1 = 0, ptvmmap = 0;
200 static pt_entry_t *msgbufmap;
201 struct msgbuf *msgbufp=0;
206 static pt_entry_t *pt_crashdumpmap;
207 static caddr_t crashdumpmap;
209 static int pmap_yield_count = 64;
210 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
211 &pmap_yield_count, 0, "Yield during init_pt/release");
215 static void pv_hold(pv_entry_t pv);
216 static int _pv_hold_try(pv_entry_t pv
218 static void pv_drop(pv_entry_t pv);
219 static void _pv_lock(pv_entry_t pv
221 static void pv_unlock(pv_entry_t pv);
222 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
224 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
226 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
227 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
228 static void pv_put(pv_entry_t pv);
229 static void pv_free(pv_entry_t pv);
230 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
231 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
233 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
234 struct pmap_inval_info *info);
235 static vm_page_t pmap_remove_pv_page(pv_entry_t pv, int holdpg);
237 static void pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
238 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
239 pt_entry_t *ptep, void *arg __unused);
240 static void pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
241 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
242 pt_entry_t *ptep, void *arg __unused);
244 static void i386_protection_init (void);
245 static void create_pagetables(vm_paddr_t *firstaddr);
246 static void pmap_remove_all (vm_page_t m);
247 static boolean_t pmap_testbit (vm_page_t m, int bit);
249 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
250 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
252 static unsigned pdir4mb;
255 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
257 if (pv1->pv_pindex < pv2->pv_pindex)
259 if (pv1->pv_pindex > pv2->pv_pindex)
264 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
265 pv_entry_compare, vm_pindex_t, pv_pindex);
268 * Move the kernel virtual free pointer to the next
269 * 2MB. This is used to help improve performance
270 * by using a large (2MB) page for much of the kernel
271 * (.text, .data, .bss)
275 pmap_kmem_choose(vm_offset_t addr)
277 vm_offset_t newaddr = addr;
279 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
286 * Super fast pmap_pte routine best used when scanning the pv lists.
287 * This eliminates many course-grained invltlb calls. Note that many of
288 * the pv list scans are across different pmaps and it is very wasteful
289 * to do an entire invltlb when checking a single mapping.
291 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
295 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
297 return pmap_pte(pmap, va);
301 * Returns the pindex of a page table entry (representing a terminal page).
302 * There are NUPTE_TOTAL page table entries possible (a huge number)
304 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
305 * We want to properly translate negative KVAs.
309 pmap_pte_pindex(vm_offset_t va)
311 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
315 * Returns the pindex of a page table.
319 pmap_pt_pindex(vm_offset_t va)
321 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
325 * Returns the pindex of a page directory.
329 pmap_pd_pindex(vm_offset_t va)
331 return (NUPTE_TOTAL + NUPT_TOTAL +
332 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
337 pmap_pdp_pindex(vm_offset_t va)
339 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
340 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
345 pmap_pml4_pindex(void)
347 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
351 * Return various clipped indexes for a given VA
353 * Returns the index of a pte in a page table, representing a terminal
358 pmap_pte_index(vm_offset_t va)
360 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
364 * Returns the index of a pt in a page directory, representing a page
369 pmap_pt_index(vm_offset_t va)
371 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
375 * Returns the index of a pd in a page directory page, representing a page
380 pmap_pd_index(vm_offset_t va)
382 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
386 * Returns the index of a pdp in the pml4 table, representing a page
391 pmap_pdp_index(vm_offset_t va)
393 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
397 * Generic procedure to index a pte from a pt, pd, or pdp.
401 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
405 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
406 return(&pte[pindex]);
410 * Return pointer to PDP slot in the PML4
414 pmap_pdp(pmap_t pmap, vm_offset_t va)
416 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
420 * Return pointer to PD slot in the PDP given a pointer to the PDP
424 pmap_pdp_to_pd(pml4_entry_t *pdp, vm_offset_t va)
428 pd = (pdp_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
429 return (&pd[pmap_pd_index(va)]);
433 * Return pointer to PD slot in the PDP
437 pmap_pd(pmap_t pmap, vm_offset_t va)
441 pdp = pmap_pdp(pmap, va);
442 if ((*pdp & PG_V) == 0)
444 return (pmap_pdp_to_pd(pdp, va));
448 * Return pointer to PT slot in the PD given a pointer to the PD
452 pmap_pd_to_pt(pdp_entry_t *pd, vm_offset_t va)
456 pt = (pd_entry_t *)PHYS_TO_DMAP(*pd & PG_FRAME);
457 return (&pt[pmap_pt_index(va)]);
461 * Return pointer to PT slot in the PD
465 pmap_pt(pmap_t pmap, vm_offset_t va)
469 pd = pmap_pd(pmap, va);
470 if (pd == NULL || (*pd & PG_V) == 0)
472 return (pmap_pd_to_pt(pd, va));
476 * Return pointer to PTE slot in the PT given a pointer to the PT
480 pmap_pt_to_pte(pd_entry_t *pt, vm_offset_t va)
484 pte = (pt_entry_t *)PHYS_TO_DMAP(*pt & PG_FRAME);
485 return (&pte[pmap_pte_index(va)]);
489 * Return pointer to PTE slot in the PT
493 pmap_pte(pmap_t pmap, vm_offset_t va)
497 pt = pmap_pt(pmap, va);
498 if (pt == NULL || (*pt & PG_V) == 0)
500 if ((*pt & PG_PS) != 0)
501 return ((pt_entry_t *)pt);
502 return (pmap_pt_to_pte(pt, va));
506 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
507 * the PT layer. This will speed up core pmap operations considerably.
511 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
513 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
514 pv->pv_pmap->pm_pvhint = pv;
519 * KVM - return address of PT slot in PD
523 vtopt(vm_offset_t va)
525 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
526 NPML4EPGSHIFT)) - 1);
528 return (PDmap + ((va >> PDRSHIFT) & mask));
532 * KVM - return address of PTE slot in PT
536 vtopte(vm_offset_t va)
538 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
539 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
541 return (PTmap + ((va >> PAGE_SHIFT) & mask));
545 allocpages(vm_paddr_t *firstaddr, long n)
550 bzero((void *)ret, n * PAGE_SIZE);
551 *firstaddr += n * PAGE_SIZE;
557 create_pagetables(vm_paddr_t *firstaddr)
559 long i; /* must be 64 bits */
564 * We are running (mostly) V=P at this point
566 * Calculate NKPT - number of kernel page tables. We have to
567 * accomodoate prealloction of the vm_page_array, dump bitmap,
568 * MSGBUF_SIZE, and other stuff. Be generous.
570 * Maxmem is in pages.
572 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
573 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
577 * Starting at the beginning of kvm (not KERNBASE).
579 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
580 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
581 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + ndmpdp) +
586 * Starting at KERNBASE - map 2G worth of page table pages.
587 * KERNBASE is offset -2G from the end of kvm.
589 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
594 KPTbase = allocpages(firstaddr, nkpt_base);
595 KPTphys = allocpages(firstaddr, nkpt_phys);
596 KPML4phys = allocpages(firstaddr, 1);
597 KPDPphys = allocpages(firstaddr, NKPML4E);
598 KPDphys = allocpages(firstaddr, NKPDPE);
601 * Calculate the page directory base for KERNBASE,
602 * that is where we start populating the page table pages.
603 * Basically this is the end - 2.
605 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
607 DMPDPphys = allocpages(firstaddr, NDMPML4E);
608 if ((amd_feature & AMDID_PAGE1GB) == 0)
609 DMPDphys = allocpages(firstaddr, ndmpdp);
610 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
613 * Fill in the underlying page table pages for the area around
614 * KERNBASE. This remaps low physical memory to KERNBASE.
616 * Read-only from zero to physfree
617 * XXX not fully used, underneath 2M pages
619 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
620 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
621 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
625 * Now map the initial kernel page tables. One block of page
626 * tables is placed at the beginning of kernel virtual memory,
627 * and another block is placed at KERNBASE to map the kernel binary,
628 * data, bss, and initial pre-allocations.
630 for (i = 0; i < nkpt_base; i++) {
631 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
632 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
634 for (i = 0; i < nkpt_phys; i++) {
635 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
636 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
640 * Map from zero to end of allocations using 2M pages as an
641 * optimization. This will bypass some of the KPTBase pages
642 * above in the KERNBASE area.
644 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
645 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
646 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
650 * And connect up the PD to the PDP. The kernel pmap is expected
651 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
653 for (i = 0; i < NKPDPE; i++) {
654 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
655 KPDphys + (i << PAGE_SHIFT);
656 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
660 /* Now set up the direct map space using either 2MB or 1GB pages */
661 /* Preset PG_M and PG_A because demotion expects it */
662 if ((amd_feature & AMDID_PAGE1GB) == 0) {
663 for (i = 0; i < NPDEPG * ndmpdp; i++) {
664 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
665 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
668 /* And the direct map space's PDP */
669 for (i = 0; i < ndmpdp; i++) {
670 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
672 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
675 for (i = 0; i < ndmpdp; i++) {
676 ((pdp_entry_t *)DMPDPphys)[i] =
677 (vm_paddr_t)i << PDPSHIFT;
678 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
683 /* And recursively map PML4 to itself in order to get PTmap */
684 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
685 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
687 /* Connect the Direct Map slot up to the PML4 */
688 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
689 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
691 /* Connect the KVA slot up to the PML4 */
692 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
693 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
697 * Bootstrap the system enough to run with virtual memory.
699 * On the i386 this is called after mapping has already been enabled
700 * and just syncs the pmap module with what has already been done.
701 * [We can't call it easily with mapping off since the kernel is not
702 * mapped with PA == VA, hence we would have to relocate every address
703 * from the linked base (virtual) address "KERNBASE" to the actual
704 * (physical) address starting relative to 0]
707 pmap_bootstrap(vm_paddr_t *firstaddr)
711 struct mdglobaldata *gd;
714 KvaStart = VM_MIN_KERNEL_ADDRESS;
715 KvaEnd = VM_MAX_KERNEL_ADDRESS;
716 KvaSize = KvaEnd - KvaStart;
718 avail_start = *firstaddr;
721 * Create an initial set of page tables to run the kernel in.
723 create_pagetables(firstaddr);
725 virtual2_start = KvaStart;
726 virtual2_end = PTOV_OFFSET;
728 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
729 virtual_start = pmap_kmem_choose(virtual_start);
731 virtual_end = VM_MAX_KERNEL_ADDRESS;
733 /* XXX do %cr0 as well */
734 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
738 * Initialize protection array.
740 i386_protection_init();
743 * The kernel's pmap is statically allocated so we don't have to use
744 * pmap_create, which is unlikely to work correctly at this part of
745 * the boot sequence (XXX and which no longer exists).
747 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
748 kernel_pmap.pm_count = 1;
749 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
750 RB_INIT(&kernel_pmap.pm_pvroot);
751 spin_init(&kernel_pmap.pm_spin);
752 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
755 * Reserve some special page table entries/VA space for temporary
758 #define SYSMAP(c, p, v, n) \
759 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
765 * CMAP1/CMAP2 are used for zeroing and copying pages.
767 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
772 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
775 * ptvmmap is used for reading arbitrary physical pages via
778 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
781 * msgbufp is used to map the system message buffer.
782 * XXX msgbufmap is not used.
784 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
785 atop(round_page(MSGBUF_SIZE)))
792 * PG_G is terribly broken on SMP because we IPI invltlb's in some
793 * cases rather then invl1pg. Actually, I don't even know why it
794 * works under UP because self-referential page table mappings
799 if (cpu_feature & CPUID_PGE)
804 * Initialize the 4MB page size flag
808 * The 4MB page version of the initial
809 * kernel page mapping.
813 #if !defined(DISABLE_PSE)
814 if (cpu_feature & CPUID_PSE) {
817 * Note that we have enabled PSE mode
820 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
821 ptditmp &= ~(NBPDR - 1);
822 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
827 * Enable the PSE mode. If we are SMP we can't do this
828 * now because the APs will not be able to use it when
831 load_cr4(rcr4() | CR4_PSE);
834 * We can do the mapping here for the single processor
835 * case. We simply ignore the old page table page from
839 * For SMP, we still need 4K pages to bootstrap APs,
840 * PSE will be enabled as soon as all APs are up.
842 PTD[KPTDI] = (pd_entry_t)ptditmp;
849 * We need to finish setting up the globaldata page for the BSP.
850 * locore has already populated the page table for the mdglobaldata
853 pg = MDGLOBALDATA_BASEALLOC_PAGES;
854 gd = &CPU_prvspace[0].mdglobaldata;
861 * Set 4mb pdir for mp startup
866 if (pseflag && (cpu_feature & CPUID_PSE)) {
867 load_cr4(rcr4() | CR4_PSE);
868 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
876 * Initialize the pmap module.
877 * Called by vm_init, to initialize any structures that the pmap
878 * system needs to map virtual memory.
879 * pmap_init has been enhanced to support in a fairly consistant
880 * way, discontiguous physical memory.
889 * Allocate memory for random pmap data structures. Includes the
893 for (i = 0; i < vm_page_array_size; i++) {
896 m = &vm_page_array[i];
897 TAILQ_INIT(&m->md.pv_list);
901 * init the pv free list
903 initial_pvs = vm_page_array_size;
904 if (initial_pvs < MINPV)
906 pvzone = &pvzone_store;
907 pvinit = (void *)kmem_alloc(&kernel_map,
908 initial_pvs * sizeof (struct pv_entry));
909 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
910 pvinit, initial_pvs);
913 * Now it is safe to enable pv_table recording.
915 pmap_initialized = TRUE;
919 * Initialize the address space (zone) for the pv_entries. Set a
920 * high water mark so that the system can recover from excessive
921 * numbers of pv entries.
926 int shpgperproc = PMAP_SHPGPERPROC;
929 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
930 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
931 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
932 pv_entry_high_water = 9 * (pv_entry_max / 10);
935 * Subtract out pages already installed in the zone (hack)
937 entry_max = pv_entry_max - vm_page_array_size;
941 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
945 /***************************************************
946 * Low level helper routines.....
947 ***************************************************/
949 #if defined(PMAP_DIAGNOSTIC)
952 * This code checks for non-writeable/modified pages.
953 * This should be an invalid condition.
957 pmap_nw_modified(pt_entry_t pte)
959 if ((pte & (PG_M|PG_RW)) == PG_M)
968 * this routine defines the region(s) of memory that should
969 * not be tested for the modified bit.
973 pmap_track_modified(vm_pindex_t pindex)
975 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
976 if ((va < clean_sva) || (va >= clean_eva))
983 * Extract the physical page address associated with the map/VA pair.
984 * The page must be wired for this to work reliably.
986 * XXX for the moment we're using pv_find() instead of pv_get(), as
987 * callers might be expecting non-blocking operation.
990 pmap_extract(pmap_t pmap, vm_offset_t va)
997 if (va >= VM_MAX_USER_ADDRESS) {
999 * Kernel page directories might be direct-mapped and
1000 * there is typically no PV tracking of pte's
1004 pt = pmap_pt(pmap, va);
1005 if (pt && (*pt & PG_V)) {
1007 rtval = *pt & PG_PS_FRAME;
1008 rtval |= va & PDRMASK;
1010 ptep = pmap_pt_to_pte(pt, va);
1012 rtval = *ptep & PG_FRAME;
1013 rtval |= va & PAGE_MASK;
1019 * User pages currently do not direct-map the page directory
1020 * and some pages might not used managed PVs. But all PT's
1023 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1025 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1027 rtval = *ptep & PG_FRAME;
1028 rtval |= va & PAGE_MASK;
1037 * Extract the physical page address associated kernel virtual address.
1040 pmap_kextract(vm_offset_t va)
1042 pd_entry_t pt; /* pt entry in pd */
1045 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1046 pa = DMAP_TO_PHYS(va);
1050 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1053 * Beware of a concurrent promotion that changes the
1054 * PDE at this point! For example, vtopte() must not
1055 * be used to access the PTE because it would use the
1056 * new PDE. It is, however, safe to use the old PDE
1057 * because the page table page is preserved by the
1060 pa = *pmap_pt_to_pte(&pt, va);
1061 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1067 /***************************************************
1068 * Low level mapping routines.....
1069 ***************************************************/
1072 * Routine: pmap_kenter
1074 * Add a wired page to the KVA
1075 * NOTE! note that in order for the mapping to take effect -- you
1076 * should do an invltlb after doing the pmap_kenter().
1079 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1083 pmap_inval_info info;
1085 pmap_inval_init(&info); /* XXX remove */
1086 npte = pa | PG_RW | PG_V | pgeflag;
1088 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1090 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1091 pmap_inval_done(&info); /* XXX remove */
1095 * Routine: pmap_kenter_quick
1097 * Similar to pmap_kenter(), except we only invalidate the
1098 * mapping on the current CPU.
1101 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1106 npte = pa | PG_RW | PG_V | pgeflag;
1109 cpu_invlpg((void *)va);
1113 pmap_kenter_sync(vm_offset_t va)
1115 pmap_inval_info info;
1117 pmap_inval_init(&info);
1118 pmap_inval_interlock(&info, &kernel_pmap, va);
1119 pmap_inval_deinterlock(&info, &kernel_pmap);
1120 pmap_inval_done(&info);
1124 pmap_kenter_sync_quick(vm_offset_t va)
1126 cpu_invlpg((void *)va);
1130 * remove a page from the kernel pagetables
1133 pmap_kremove(vm_offset_t va)
1136 pmap_inval_info info;
1138 pmap_inval_init(&info);
1140 pmap_inval_interlock(&info, &kernel_pmap, va);
1142 pmap_inval_deinterlock(&info, &kernel_pmap);
1143 pmap_inval_done(&info);
1147 pmap_kremove_quick(vm_offset_t va)
1152 cpu_invlpg((void *)va);
1156 * XXX these need to be recoded. They are not used in any critical path.
1159 pmap_kmodify_rw(vm_offset_t va)
1161 atomic_set_long(vtopte(va), PG_RW);
1162 cpu_invlpg((void *)va);
1166 pmap_kmodify_nc(vm_offset_t va)
1168 atomic_set_long(vtopte(va), PG_N);
1169 cpu_invlpg((void *)va);
1173 * Used to map a range of physical addresses into kernel virtual
1174 * address space during the low level boot, typically to map the
1175 * dump bitmap, message buffer, and vm_page_array.
1177 * These mappings are typically made at some pointer after the end of the
1180 * We could return PHYS_TO_DMAP(start) here and not allocate any
1181 * via (*virtp), but then kmem from userland and kernel dumps won't
1182 * have access to the related pointers.
1185 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1188 vm_offset_t va_start;
1190 /*return PHYS_TO_DMAP(start);*/
1195 while (start < end) {
1196 pmap_kenter_quick(va, start);
1206 * Add a list of wired pages to the kva
1207 * this routine is only used for temporary
1208 * kernel mappings that do not need to have
1209 * page modification or references recorded.
1210 * Note that old mappings are simply written
1211 * over. The page *must* be wired.
1214 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1218 end_va = va + count * PAGE_SIZE;
1220 while (va < end_va) {
1224 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1225 cpu_invlpg((void *)va);
1233 * This routine jerks page mappings from the
1234 * kernel -- it is meant only for temporary mappings.
1236 * MPSAFE, INTERRUPT SAFE (cluster callback)
1239 pmap_qremove(vm_offset_t va, int count)
1243 end_va = va + count * PAGE_SIZE;
1245 while (va < end_va) {
1250 cpu_invlpg((void *)va);
1257 * Create a new thread and optionally associate it with a (new) process.
1258 * NOTE! the new thread's cpu may not equal the current cpu.
1261 pmap_init_thread(thread_t td)
1263 /* enforce pcb placement & alignment */
1264 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1265 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1266 td->td_savefpu = &td->td_pcb->pcb_save;
1267 td->td_sp = (char *)td->td_pcb; /* no -16 */
1271 * This routine directly affects the fork perf for a process.
1274 pmap_init_proc(struct proc *p)
1279 * Dispose the UPAGES for a process that has exited.
1280 * This routine directly impacts the exit perf of a process.
1283 pmap_dispose_proc(struct proc *p)
1285 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1289 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1290 * it, and IdlePTD, represents the template used to update all other pmaps.
1292 * On architectures where the kernel pmap is not integrated into the user
1293 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1294 * kernel_pmap should be used to directly access the kernel_pmap.
1297 pmap_pinit0(struct pmap *pmap)
1299 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1301 pmap->pm_active = 0;
1302 pmap->pm_pvhint = NULL;
1303 RB_INIT(&pmap->pm_pvroot);
1304 spin_init(&pmap->pm_spin);
1305 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1306 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1310 * Initialize a preallocated and zeroed pmap structure,
1311 * such as one in a vmspace structure.
1314 pmap_pinit(struct pmap *pmap)
1319 * Misc initialization
1322 pmap->pm_active = 0;
1323 pmap->pm_pvhint = NULL;
1324 if (pmap->pm_pmlpv == NULL) {
1325 RB_INIT(&pmap->pm_pvroot);
1326 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1327 spin_init(&pmap->pm_spin);
1328 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1332 * No need to allocate page table space yet but we do need a valid
1333 * page directory table.
1335 if (pmap->pm_pml4 == NULL) {
1337 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1341 * Allocate the page directory page, which wires it even though
1342 * it isn't being entered into some higher level page table (it
1343 * being the highest level). If one is already cached we don't
1344 * have to do anything.
1346 if ((pv = pmap->pm_pmlpv) == NULL) {
1347 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1348 pmap->pm_pmlpv = pv;
1349 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1350 VM_PAGE_TO_PHYS(pv->pv_m));
1352 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1353 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1355 /* install self-referential address mapping entry */
1356 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1357 PG_V | PG_RW | PG_A | PG_M;
1359 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1360 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1365 * Clean up a pmap structure so it can be physically freed. This routine
1366 * is called by the vmspace dtor function. A great deal of pmap data is
1367 * left passively mapped to improve vmspace management so we have a bit
1368 * of cleanup work to do here.
1371 pmap_puninit(pmap_t pmap)
1376 KKASSERT(pmap->pm_active == 0);
1377 if ((pv = pmap->pm_pmlpv) != NULL) {
1378 if (pv_hold_try(pv) == 0)
1380 p = pmap_remove_pv_page(pv, 1);
1382 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1383 vm_page_busy_wait(p, FALSE, "pgpun");
1385 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1386 vm_page_unwire(p, 0);
1387 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1390 * XXX eventually clean out PML4 static entries and
1391 * use vm_page_free_zero()
1394 pmap->pm_pmlpv = NULL;
1396 if (pmap->pm_pml4) {
1397 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1398 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1399 pmap->pm_pml4 = NULL;
1401 KKASSERT(pmap->pm_stats.resident_count == 0);
1402 KKASSERT(pmap->pm_stats.wired_count == 0);
1406 * Wire in kernel global address entries. To avoid a race condition
1407 * between pmap initialization and pmap_growkernel, this procedure
1408 * adds the pmap to the master list (which growkernel scans to update),
1409 * then copies the template.
1412 pmap_pinit2(struct pmap *pmap)
1415 * XXX copies current process, does not fill in MPPTDI
1417 spin_lock(&pmap_spin);
1418 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1419 spin_unlock(&pmap_spin);
1423 * This routine is called when various levels in the page table need to
1424 * be populated. This routine cannot fail.
1426 * This function returns two locked pv_entry's, one representing the
1427 * requested pv and one representing the requested pv's parent pv. If
1428 * the pv did not previously exist it will be mapped into its parent
1429 * and wired, otherwise no additional wire count will be added.
1433 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1438 vm_pindex_t pt_pindex;
1443 * If the pv already exists and we aren't being asked for the
1444 * parent page table page we can just return it. A locked+held pv
1447 pv = pv_alloc(pmap, ptepindex, &isnew);
1448 if (isnew == 0 && pvpp == NULL)
1452 * This is a new PV, we have to resolve its parent page table and
1453 * add an additional wiring to the page if necessary.
1457 * Special case terminal PVs. These are not page table pages so
1458 * no vm_page is allocated (the caller supplied the vm_page). If
1459 * pvpp is non-NULL we are being asked to also removed the pt_pv
1462 * Note that pt_pv's are only returned for user VAs. We assert that
1463 * a pt_pv is not being requested for kernel VAs.
1465 if (ptepindex < pmap_pt_pindex(0)) {
1466 if (ptepindex >= NUPTE_USER)
1467 KKASSERT(pvpp == NULL);
1469 KKASSERT(pvpp != NULL);
1471 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1472 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1474 vm_page_wire_quick(pvp->pv_m);
1483 * Non-terminal PVs allocate a VM page to represent the page table,
1484 * so we have to resolve pvp and calculate ptepindex for the pvp
1485 * and then for the page table entry index in the pvp for
1488 if (ptepindex < pmap_pd_pindex(0)) {
1490 * pv is PT, pvp is PD
1492 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1493 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1494 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1501 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1502 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1503 } else if (ptepindex < pmap_pdp_pindex(0)) {
1505 * pv is PD, pvp is PDP
1507 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1508 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1509 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1516 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1517 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1518 } else if (ptepindex < pmap_pml4_pindex()) {
1520 * pv is PDP, pvp is the root pml4 table
1522 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1529 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1530 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1533 * pv represents the top-level PML4, there is no parent.
1541 * This code is only reached if isnew is TRUE and this is not a
1542 * terminal PV. We need to allocate a vm_page for the page table
1543 * at this level and enter it into the parent page table.
1545 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1548 m = vm_page_alloc(NULL, pv->pv_pindex,
1549 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1550 VM_ALLOC_INTERRUPT);
1555 vm_page_spin_lock(m);
1556 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1558 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1559 vm_page_spin_unlock(m);
1560 vm_page_unmanage(m); /* m must be spinunlocked */
1562 if ((m->flags & PG_ZERO) == 0) {
1563 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1567 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1570 m->valid = VM_PAGE_BITS_ALL;
1571 vm_page_flag_clear(m, PG_ZERO);
1572 vm_page_wire(m); /* wire for mapping in parent */
1575 * Wire the page into pvp, bump the wire-count for pvp's page table
1576 * page. Bump the resident_count for the pmap. There is no pvp
1577 * for the top level, address the pm_pml4[] array directly.
1579 * If the caller wants the parent we return it, otherwise
1580 * we just put it away.
1582 * No interlock is needed for pte 0 -> non-zero.
1585 vm_page_wire_quick(pvp->pv_m);
1586 ptep = pv_pte_lookup(pvp, ptepindex);
1587 KKASSERT((*ptep & PG_V) == 0);
1588 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1601 * Release any resources held by the given physical map.
1603 * Called when a pmap initialized by pmap_pinit is being released. Should
1604 * only be called if the map contains no valid mappings.
1606 * Caller must hold pmap->pm_token
1608 struct pmap_release_info {
1613 static int pmap_release_callback(pv_entry_t pv, void *data);
1616 pmap_release(struct pmap *pmap)
1618 struct pmap_release_info info;
1620 KASSERT(pmap->pm_active == 0,
1621 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1623 spin_lock(&pmap_spin);
1624 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1625 spin_unlock(&pmap_spin);
1628 * Pull pv's off the RB tree in order from low to high and release
1634 spin_lock(&pmap->pm_spin);
1635 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
1636 pmap_release_callback, &info);
1637 spin_unlock(&pmap->pm_spin);
1638 } while (info.retry);
1642 * One resident page (the pml4 page) should remain.
1643 * No wired pages should remain.
1645 KKASSERT(pmap->pm_stats.resident_count == 1);
1646 KKASSERT(pmap->pm_stats.wired_count == 0);
1650 pmap_release_callback(pv_entry_t pv, void *data)
1652 struct pmap_release_info *info = data;
1653 pmap_t pmap = info->pmap;
1656 if (pv_hold_try(pv)) {
1657 spin_unlock(&pmap->pm_spin);
1659 spin_unlock(&pmap->pm_spin);
1661 if (pv->pv_pmap != pmap) {
1663 spin_lock(&pmap->pm_spin);
1670 * The pmap is currently not spinlocked, pv is held+locked.
1671 * Remove the pv's page from its parent's page table. The
1672 * parent's page table page's wire_count will be decremented.
1674 pmap_remove_pv_pte(pv, NULL, NULL);
1677 * Terminal pvs are unhooked from their vm_pages. Because
1678 * terminal pages aren't page table pages they aren't wired
1679 * by us, so we have to be sure not to unwire them either.
1681 if (pv->pv_pindex < pmap_pt_pindex(0)) {
1682 pmap_remove_pv_page(pv, 0);
1687 * We leave the top-level page table page cached, wired, and
1688 * mapped in the pmap until the dtor function (pmap_puninit())
1691 * Since we are leaving the top-level pv intact we need
1692 * to break out of what would otherwise be an infinite loop.
1694 if (pv->pv_pindex == pmap_pml4_pindex()) {
1696 spin_lock(&pmap->pm_spin);
1701 * For page table pages (other than the top-level page),
1702 * remove and free the vm_page. The representitive mapping
1703 * removed above by pmap_remove_pv_pte() did not undo the
1704 * last wire_count so we have to do that as well.
1706 p = pmap_remove_pv_page(pv, 1);
1707 vm_page_busy_wait(p, FALSE, "pmaprl");
1709 if (p->wire_count != 1) {
1710 kprintf("p->wire_count was %016lx %d\n",
1711 pv->pv_pindex, p->wire_count);
1713 KKASSERT(p->wire_count == 1);
1714 KKASSERT(p->flags & PG_UNMANAGED);
1716 vm_page_unwire(p, 0);
1717 KKASSERT(p->wire_count == 0);
1718 /* JG eventually revert to using vm_page_free_zero() */
1722 spin_lock(&pmap->pm_spin);
1727 * This function will remove the pte associated with a pv from its parent.
1728 * Terminal pv's are supported. The removal will be interlocked if info
1729 * is non-NULL. The caller must dispose of pv instead of just unlocking
1732 * The wire count will be dropped on the parent page table. The wire
1733 * count on the page being removed (pv->pv_m) from the parent page table
1734 * is NOT touched. Note that terminal pages will not have any additional
1735 * wire counts while page table pages will have at least one representing
1736 * the mapping, plus others representing sub-mappings.
1738 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
1739 * pages and user page table and terminal pages.
1741 * The pv must be locked.
1743 * XXX must lock parent pv's if they exist to remove pte XXX
1747 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
1749 vm_pindex_t ptepindex = pv->pv_pindex;
1750 pmap_t pmap = pv->pv_pmap;
1756 if (ptepindex == pmap_pml4_pindex()) {
1758 * We are the top level pml4 table, there is no parent.
1760 p = pmap->pm_pmlpv->pv_m;
1761 } else if (ptepindex >= pmap_pdp_pindex(0)) {
1763 * Remove a PDP page from the pml4e. This can only occur
1764 * with user page tables. We do not have to lock the
1765 * pml4 PV so just ignore pvp.
1767 vm_pindex_t pml4_pindex;
1768 vm_pindex_t pdp_index;
1771 pdp_index = ptepindex - pmap_pdp_pindex(0);
1773 pml4_pindex = pmap_pml4_pindex();
1774 pvp = pv_get(pv->pv_pmap, pml4_pindex);
1777 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
1778 KKASSERT((*pdp & PG_V) != 0);
1779 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1781 KKASSERT(info == NULL);
1782 } else if (ptepindex >= pmap_pd_pindex(0)) {
1784 * Remove a PD page from the pdp
1786 vm_pindex_t pdp_pindex;
1787 vm_pindex_t pd_index;
1790 pd_index = ptepindex - pmap_pd_pindex(0);
1793 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
1794 (pd_index >> NPML4EPGSHIFT);
1795 pvp = pv_get(pv->pv_pmap, pdp_pindex);
1798 pd = pv_pte_lookup(pvp, pd_index & ((1ul << NPDPEPGSHIFT) - 1));
1799 KKASSERT((*pd & PG_V) != 0);
1800 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
1802 KKASSERT(info == NULL);
1803 } else if (ptepindex >= pmap_pt_pindex(0)) {
1805 * Remove a PT page from the pd
1807 vm_pindex_t pd_pindex;
1808 vm_pindex_t pt_index;
1811 pt_index = ptepindex - pmap_pt_pindex(0);
1814 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
1815 (pt_index >> NPDPEPGSHIFT);
1816 pvp = pv_get(pv->pv_pmap, pd_pindex);
1819 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
1820 KKASSERT((*pt & PG_V) != 0);
1821 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
1823 KKASSERT(info == NULL);
1826 * Remove a PTE from the PT page
1828 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
1829 * pv is a pte_pv so we can safely lock pt_pv.
1831 vm_pindex_t pt_pindex;
1836 pt_pindex = ptepindex >> NPTEPGSHIFT;
1837 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
1839 if (ptepindex >= NUPTE_USER) {
1840 ptep = vtopte(ptepindex << PAGE_SHIFT);
1841 KKASSERT(pvp == NULL);
1844 pt_pindex = NUPTE_TOTAL +
1845 (ptepindex >> NPDPEPGSHIFT);
1846 pvp = pv_get(pv->pv_pmap, pt_pindex);
1849 ptep = pv_pte_lookup(pvp, ptepindex &
1850 ((1ul << NPDPEPGSHIFT) - 1));
1854 pmap_inval_interlock(info, pmap, va);
1855 pte = pte_load_clear(ptep);
1857 pmap_inval_deinterlock(info, pmap);
1860 * Now update the vm_page_t
1862 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
1863 kprintf("remove_pte badpte %016lx %016lx %d\n",
1865 pv->pv_pindex < pmap_pt_pindex(0));
1867 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
1868 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
1871 if (pmap_track_modified(ptepindex))
1875 vm_page_flag_set(p, PG_REFERENCED);
1878 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1880 cpu_invlpg((void *)va);
1884 * Unwire the parent page table page. The wire_count cannot go below
1885 * 1 here because the parent page table page is itself still mapped.
1887 * XXX remove the assertions later.
1889 KKASSERT(pv->pv_m == p);
1890 if (pvp && vm_page_unwire_quick(pvp->pv_m))
1891 panic("pmap_remove_pv_pte: Insufficient wire_count");
1899 pmap_remove_pv_page(pv_entry_t pv, int holdpg)
1907 vm_page_spin_lock(m);
1909 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1912 atomic_add_int(&m->object->agg_pv_list_count, -1);
1914 if (TAILQ_EMPTY(&m->md.pv_list))
1915 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1916 vm_page_spin_unlock(m);
1923 * Grow the number of kernel page table entries, if needed.
1925 * This routine is always called to validate any address space
1926 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1927 * space below KERNBASE.
1930 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1933 vm_offset_t ptppaddr;
1935 pd_entry_t *pt, newpt;
1937 int update_kernel_vm_end;
1940 * bootstrap kernel_vm_end on first real VM use
1942 if (kernel_vm_end == 0) {
1943 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1945 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1946 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1947 ~(PAGE_SIZE * NPTEPG - 1);
1949 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1950 kernel_vm_end = kernel_map.max_offset;
1957 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1958 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1959 * do not want to force-fill 128G worth of page tables.
1961 if (kstart < KERNBASE) {
1962 if (kstart > kernel_vm_end)
1963 kstart = kernel_vm_end;
1964 KKASSERT(kend <= KERNBASE);
1965 update_kernel_vm_end = 1;
1967 update_kernel_vm_end = 0;
1970 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1971 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1973 if (kend - 1 >= kernel_map.max_offset)
1974 kend = kernel_map.max_offset;
1976 while (kstart < kend) {
1977 pt = pmap_pt(&kernel_pmap, kstart);
1979 /* We need a new PDP entry */
1980 nkpg = vm_page_alloc(NULL, nkpt,
1983 VM_ALLOC_INTERRUPT);
1985 panic("pmap_growkernel: no memory to grow "
1988 paddr = VM_PAGE_TO_PHYS(nkpg);
1989 if ((nkpg->flags & PG_ZERO) == 0)
1990 pmap_zero_page(paddr);
1991 vm_page_flag_clear(nkpg, PG_ZERO);
1992 newpd = (pdp_entry_t)
1993 (paddr | PG_V | PG_RW | PG_A | PG_M);
1994 *pmap_pd(&kernel_pmap, kstart) = newpd;
1996 continue; /* try again */
1998 if ((*pt & PG_V) != 0) {
1999 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2000 ~(PAGE_SIZE * NPTEPG - 1);
2001 if (kstart - 1 >= kernel_map.max_offset) {
2002 kstart = kernel_map.max_offset;
2009 * This index is bogus, but out of the way
2011 nkpg = vm_page_alloc(NULL, nkpt,
2014 VM_ALLOC_INTERRUPT);
2016 panic("pmap_growkernel: no memory to grow kernel");
2019 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2020 pmap_zero_page(ptppaddr);
2021 vm_page_flag_clear(nkpg, PG_ZERO);
2022 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2023 *pmap_pt(&kernel_pmap, kstart) = newpt;
2026 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2027 ~(PAGE_SIZE * NPTEPG - 1);
2029 if (kstart - 1 >= kernel_map.max_offset) {
2030 kstart = kernel_map.max_offset;
2036 * Only update kernel_vm_end for areas below KERNBASE.
2038 if (update_kernel_vm_end && kernel_vm_end < kstart)
2039 kernel_vm_end = kstart;
2043 * Retire the given physical map from service.
2044 * Should only be called if the map contains
2045 * no valid mappings.
2048 pmap_destroy(pmap_t pmap)
2055 lwkt_gettoken(&pmap->pm_token);
2056 count = --pmap->pm_count;
2058 pmap_release(pmap); /* eats pm_token */
2059 panic("destroying a pmap is not yet implemented");
2061 lwkt_reltoken(&pmap->pm_token);
2065 * Add a reference to the specified pmap.
2068 pmap_reference(pmap_t pmap)
2071 lwkt_gettoken(&pmap->pm_token);
2073 lwkt_reltoken(&pmap->pm_token);
2077 /***************************************************
2078 * page management routines.
2079 ***************************************************/
2082 * Hold a pv without locking it
2085 pv_hold(pv_entry_t pv)
2089 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2093 count = pv->pv_hold;
2095 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2102 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2103 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2106 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2107 * pv list via its page) must be held by the caller.
2110 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2114 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2117 pv->pv_line = lineno;
2123 count = pv->pv_hold;
2125 if ((count & PV_HOLD_LOCKED) == 0) {
2126 if (atomic_cmpset_int(&pv->pv_hold, count,
2127 (count + 1) | PV_HOLD_LOCKED)) {
2130 pv->pv_line = lineno;
2135 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2143 * Drop a previously held pv_entry which could not be locked, allowing its
2146 * Must not be called with a spinlock held as we might zfree() the pv if it
2147 * is no longer associated with a pmap and this was the last hold count.
2150 pv_drop(pv_entry_t pv)
2154 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2155 if (pv->pv_pmap == NULL)
2161 count = pv->pv_hold;
2163 KKASSERT((count & PV_HOLD_MASK) > 0);
2164 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2165 (PV_HOLD_LOCKED | 1));
2166 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2167 if (count == 1 && pv->pv_pmap == NULL)
2176 * Find or allocate the requested PV entry, returning a locked pv
2180 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2183 pv_entry_t pnew = NULL;
2185 spin_lock(&pmap->pm_spin);
2187 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2188 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2193 spin_unlock(&pmap->pm_spin);
2194 pnew = zalloc(pvzone);
2195 spin_lock(&pmap->pm_spin);
2198 pnew->pv_pmap = pmap;
2199 pnew->pv_pindex = pindex;
2200 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2202 pnew->pv_func = func;
2203 pnew->pv_line = lineno;
2205 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2206 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2207 spin_unlock(&pmap->pm_spin);
2212 spin_unlock(&pmap->pm_spin);
2213 zfree(pvzone, pnew);
2215 spin_lock(&pmap->pm_spin);
2218 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2219 spin_unlock(&pmap->pm_spin);
2223 spin_unlock(&pmap->pm_spin);
2224 _pv_lock(pv PMAP_DEBUG_COPY);
2225 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2230 spin_lock(&pmap->pm_spin);
2237 * Find the requested PV entry, returning a locked+held pv or NULL
2241 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2245 spin_lock(&pmap->pm_spin);
2250 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2251 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2255 spin_unlock(&pmap->pm_spin);
2258 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2259 pv_cache(pv, pindex);
2260 spin_unlock(&pmap->pm_spin);
2263 spin_unlock(&pmap->pm_spin);
2264 _pv_lock(pv PMAP_DEBUG_COPY);
2265 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2268 spin_lock(&pmap->pm_spin);
2273 * Lookup, hold, and attempt to lock (pmap,pindex).
2275 * If the entry does not exist NULL is returned and *errorp is set to 0
2277 * If the entry exists and could be successfully locked it is returned and
2278 * errorp is set to 0.
2280 * If the entry exists but could NOT be successfully locked it is returned
2281 * held and *errorp is set to 1.
2285 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2289 spin_lock(&pmap->pm_spin);
2290 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2291 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2293 spin_unlock(&pmap->pm_spin);
2297 if (pv_hold_try(pv)) {
2298 pv_cache(pv, pindex);
2299 spin_unlock(&pmap->pm_spin);
2301 return(pv); /* lock succeeded */
2303 spin_unlock(&pmap->pm_spin);
2305 return (pv); /* lock failed */
2309 * Find the requested PV entry, returning a held pv or NULL
2313 pv_find(pmap_t pmap, vm_pindex_t pindex)
2317 spin_lock(&pmap->pm_spin);
2319 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2320 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2322 spin_unlock(&pmap->pm_spin);
2326 pv_cache(pv, pindex);
2327 spin_unlock(&pmap->pm_spin);
2332 * Lock a held pv, keeping the hold count
2336 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2341 count = pv->pv_hold;
2343 if ((count & PV_HOLD_LOCKED) == 0) {
2344 if (atomic_cmpset_int(&pv->pv_hold, count,
2345 count | PV_HOLD_LOCKED)) {
2348 pv->pv_line = lineno;
2354 tsleep_interlock(pv, 0);
2355 if (atomic_cmpset_int(&pv->pv_hold, count,
2356 count | PV_HOLD_WAITING)) {
2358 kprintf("pv waiting on %s:%d\n",
2359 pv->pv_func, pv->pv_line);
2361 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2368 * Unlock a held and locked pv, keeping the hold count.
2372 pv_unlock(pv_entry_t pv)
2376 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2380 count = pv->pv_hold;
2382 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2383 (PV_HOLD_LOCKED | 1));
2384 if (atomic_cmpset_int(&pv->pv_hold, count,
2386 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2387 if (count & PV_HOLD_WAITING)
2395 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2396 * and the hold count drops to zero we will free it.
2398 * Caller should not hold any spin locks. We are protected from hold races
2399 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2400 * lock held. A pv cannot be located otherwise.
2404 pv_put(pv_entry_t pv)
2406 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2407 if (pv->pv_pmap == NULL)
2416 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2417 * pmap. Any pte operations must have already been completed.
2421 pv_free(pv_entry_t pv)
2425 KKASSERT(pv->pv_m == NULL);
2426 if ((pmap = pv->pv_pmap) != NULL) {
2427 spin_lock(&pmap->pm_spin);
2428 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2429 if (pmap->pm_pvhint == pv)
2430 pmap->pm_pvhint = NULL;
2431 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2434 spin_unlock(&pmap->pm_spin);
2440 * This routine is very drastic, but can save the system
2448 static int warningdone=0;
2450 if (pmap_pagedaemon_waken == 0)
2452 pmap_pagedaemon_waken = 0;
2453 if (warningdone < 5) {
2454 kprintf("pmap_collect: collecting pv entries -- "
2455 "suggest increasing PMAP_SHPGPERPROC\n");
2459 for (i = 0; i < vm_page_array_size; i++) {
2460 m = &vm_page_array[i];
2461 if (m->wire_count || m->hold_count)
2463 if (vm_page_busy_try(m, TRUE) == 0) {
2464 if (m->wire_count == 0 && m->hold_count == 0) {
2473 * Scan the pmap for active page table entries and issue a callback.
2474 * The callback must dispose of pte_pv.
2476 * NOTE: Unmanaged page table entries will not have a pte_pv
2478 * NOTE: Kernel page table entries will not have a pt_pv. That is, wiring
2479 * counts are not tracked in kernel page table pages.
2481 * It is assumed that the start and end are properly rounded to the page size.
2484 pmap_scan(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva,
2485 void (*func)(pmap_t, struct pmap_inval_info *,
2486 pv_entry_t, pv_entry_t, vm_offset_t,
2487 pt_entry_t *, void *),
2490 pv_entry_t pdp_pv; /* A page directory page PV */
2491 pv_entry_t pd_pv; /* A page directory PV */
2492 pv_entry_t pt_pv; /* A page table PV */
2493 pv_entry_t pte_pv; /* A page table entry PV */
2495 vm_offset_t va_next;
2496 struct pmap_inval_info info;
2503 * Hold the token for stability; if the pmap is empty we have nothing
2506 lwkt_gettoken(&pmap->pm_token);
2508 if (pmap->pm_stats.resident_count == 0) {
2509 lwkt_reltoken(&pmap->pm_token);
2514 pmap_inval_init(&info);
2517 * Special handling for removing one page, which is a very common
2518 * operation (it is?).
2519 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
2521 if (sva + PAGE_SIZE == eva) {
2522 if (sva >= VM_MAX_USER_ADDRESS) {
2524 * Kernel mappings do not track wire counts on
2528 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2532 * User mappings may or may not have a pte_pv but
2533 * will always have a pt_pv if the page is present.
2535 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2536 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2537 if (pt_pv == NULL) {
2538 KKASSERT(pte_pv == NULL);
2541 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
2545 * Unlike the pv_find() case below we actually
2546 * acquired a locked pv in this case so any
2547 * race should have been resolved. It is expected
2550 KKASSERT(pte_pv == NULL);
2551 } else if (pte_pv) {
2552 KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED|
2554 ("bad *ptep %016lx sva %016lx pte_pv %p",
2555 *ptep, sva, pte_pv));
2556 func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg);
2558 KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V,
2559 ("bad *ptep %016lx sva %016lx pte_pv NULL",
2561 func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg);
2566 pmap_inval_done(&info);
2567 lwkt_reltoken(&pmap->pm_token);
2572 * NOTE: kernel mappings do not track page table pages, only
2575 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
2576 * However, for the scan to be efficient we try to
2577 * cache items top-down.
2583 for (; sva < eva; sva = va_next) {
2585 if (sva >= VM_MAX_USER_ADDRESS) {
2596 if (pdp_pv == NULL) {
2597 pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva));
2598 } else if (pdp_pv->pv_pindex != pmap_pdp_pindex(sva)) {
2600 pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva));
2602 if (pdp_pv == NULL) {
2603 va_next = (sva + NBPML4) & ~PML4MASK;
2612 if (pd_pv == NULL) {
2617 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
2618 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
2624 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
2626 if (pd_pv == NULL) {
2627 va_next = (sva + NBPDP) & ~PDPMASK;
2636 if (pt_pv == NULL) {
2645 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2646 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
2656 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2660 * We will scan or skip a page table page so adjust va_next
2663 if (pt_pv == NULL) {
2664 va_next = (sva + NBPDR) & ~PDRMASK;
2671 * From this point in the loop testing pt_pv for non-NULL
2672 * means we are in UVM, else if it is NULL we are in KVM.
2675 va_next = (sva + NBPDR) & ~PDRMASK;
2680 * Limit our scan to either the end of the va represented
2681 * by the current page table page, or to the end of the
2682 * range being removed.
2684 * Scan the page table for pages. Some pages may not be
2685 * managed (might not have a pv_entry).
2687 * There is no page table management for kernel pages so
2688 * pt_pv will be NULL in that case, but otherwise pt_pv
2689 * is non-NULL, locked, and referenced.
2695 * At this point a non-NULL pt_pv means a UVA, and a NULL
2696 * pt_pv means a KVA.
2699 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
2703 while (sva < va_next) {
2705 * Acquire the related pte_pv, if any. If *ptep == 0
2706 * the related pte_pv should not exist, but if *ptep
2707 * is not zero the pte_pv may or may not exist (e.g.
2708 * will not exist for an unmanaged page).
2710 * However a multitude of races are possible here.
2712 * In addition, the (pt_pv, pte_pv) lock order is
2713 * backwards, so we have to be careful in aquiring
2714 * a properly locked pte_pv.
2718 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
2729 pv_put(pt_pv); /* must be non-NULL */
2731 pv_lock(pte_pv); /* safe to block now */
2734 pt_pv = pv_get(pmap,
2735 pmap_pt_pindex(sva));
2739 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2743 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
2747 kprintf("Unexpected non-NULL pte_pv "
2748 "%p pt_pv %p *ptep = %016lx\n",
2749 pte_pv, pt_pv, *ptep);
2750 panic("Unexpected non-NULL pte_pv");
2758 * Ready for the callback. The locked pte_pv (if any)
2759 * is consumed by the callback. pte_pv will exist if
2760 * the page is managed, and will not exist if it
2764 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
2766 ("bad *ptep %016lx sva %016lx "
2768 *ptep, sva, pte_pv));
2769 func(pmap, &info, pte_pv, pt_pv, sva,
2772 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
2774 ("bad *ptep %016lx sva %016lx "
2777 func(pmap, &info, pte_pv, pt_pv, sva,
2797 pmap_inval_done(&info);
2798 lwkt_reltoken(&pmap->pm_token);
2802 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2804 pmap_scan(pmap, sva, eva, pmap_remove_callback, NULL);
2808 pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
2809 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
2810 pt_entry_t *ptep, void *arg __unused)
2816 * This will also drop pt_pv's wire_count. Note that
2817 * terminal pages are not wired based on mmu presence.
2819 pmap_remove_pv_pte(pte_pv, pt_pv, info);
2820 pmap_remove_pv_page(pte_pv, 0);
2824 * pt_pv's wire_count is still bumped by unmanaged pages
2825 * so we must decrement it manually.
2827 pmap_inval_interlock(info, pmap, va);
2828 pte = pte_load_clear(ptep);
2829 pmap_inval_deinterlock(info, pmap);
2831 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2832 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2833 if (pt_pv && vm_page_unwire_quick(pt_pv->pv_m))
2834 panic("pmap_remove: insufficient wirecount");
2839 * Removes this physical page from all physical maps in which it resides.
2840 * Reflects back modify bits to the pager.
2842 * This routine may not be called from an interrupt.
2846 pmap_remove_all(vm_page_t m)
2848 struct pmap_inval_info info;
2851 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2854 pmap_inval_init(&info);
2855 vm_page_spin_lock(m);
2856 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2857 KKASSERT(pv->pv_m == m);
2858 if (pv_hold_try(pv)) {
2859 vm_page_spin_unlock(m);
2861 vm_page_spin_unlock(m);
2863 if (pv->pv_m != m) {
2865 vm_page_spin_lock(m);
2870 * Holding no spinlocks, pv is locked.
2872 pmap_remove_pv_pte(pv, NULL, &info);
2873 pmap_remove_pv_page(pv, 0);
2875 vm_page_spin_lock(m);
2877 vm_page_spin_unlock(m);
2878 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2879 pmap_inval_done(&info);
2885 * Set the physical protection on the specified range of this map
2888 * This function may not be called from an interrupt if the map is
2889 * not the kernel_pmap.
2892 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2894 /* JG review for NX */
2898 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2899 pmap_remove(pmap, sva, eva);
2902 if (prot & VM_PROT_WRITE)
2904 pmap_scan(pmap, sva, eva, pmap_protect_callback, &prot);
2909 pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
2910 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
2911 pt_entry_t *ptep, void *arg __unused)
2920 pmap_inval_interlock(info, pmap, va);
2927 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2928 KKASSERT(m == pte_pv->pv_m);
2929 vm_page_flag_set(m, PG_REFERENCED);
2933 if (pmap_track_modified(pte_pv->pv_pindex)) {
2935 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2942 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
2945 pmap_inval_deinterlock(info, pmap);
2951 * Insert the vm_page (m) at the virtual address (va), replacing any prior
2952 * mapping at that address. Set protection and wiring as requested.
2954 * NOTE: This routine MUST insert the page into the pmap now, it cannot
2958 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2961 pmap_inval_info info;
2962 pv_entry_t pt_pv; /* page table */
2963 pv_entry_t pte_pv; /* page table entry */
2966 pt_entry_t origpte, newpte;
2971 va = trunc_page(va);
2972 #ifdef PMAP_DIAGNOSTIC
2974 panic("pmap_enter: toobig");
2975 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2976 panic("pmap_enter: invalid to pmap_enter page table "
2977 "pages (va: 0x%lx)", va);
2979 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2980 kprintf("Warning: pmap_enter called on UVA with "
2983 db_print_backtrace();
2986 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2987 kprintf("Warning: pmap_enter called on KVA without"
2990 db_print_backtrace();
2995 * Get locked PV entries for our new page table entry (pte_pv)
2996 * and for its parent page table (pt_pv). We need the parent
2997 * so we can resolve the location of the ptep.
2999 * Only hardware MMU actions can modify the ptep out from
3002 * if (m) is fictitious or unmanaged we do not create a managing
3003 * pte_pv for it. Any pre-existing page's management state must
3004 * match (avoiding code complexity).
3006 * If the pmap is still being initialized we assume existing
3009 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3010 * pmap_allocpte() checks the
3012 if (pmap_initialized == FALSE) {
3016 } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) {
3018 if (va >= VM_MAX_USER_ADDRESS) {
3022 pt_pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
3023 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3025 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0);
3027 if (va >= VM_MAX_USER_ADDRESS) {
3029 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3032 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va),
3034 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3036 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED));
3039 if ((prot & VM_PROT_NOSYNC) == 0)
3040 pmap_inval_init(&info);
3042 pa = VM_PAGE_TO_PHYS(m);
3044 opa = origpte & PG_FRAME;
3047 * Mapping has not changed, must be protection or wiring change.
3049 if (origpte && (opa == pa)) {
3051 * Wiring change, just update stats. We don't worry about
3052 * wiring PT pages as they remain resident as long as there
3053 * are valid mappings in them. Hence, if a user page is wired,
3054 * the PT page will be also.
3056 KKASSERT(pte_pv == NULL || m == pte_pv->pv_m);
3057 if (wired && ((origpte & PG_W) == 0))
3058 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3059 else if (!wired && (origpte & PG_W))
3060 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3062 #if defined(PMAP_DIAGNOSTIC)
3063 if (pmap_nw_modified(origpte)) {
3064 kprintf("pmap_enter: modified page not writable: "
3065 "va: 0x%lx, pte: 0x%lx\n", va, origpte);
3070 * We might be turning off write access to the page,
3071 * so we go ahead and sense modify status.
3074 if ((origpte & PG_M) &&
3075 pmap_track_modified(pte_pv->pv_pindex)) {
3078 KKASSERT(PHYS_TO_VM_PAGE(opa) == om);
3087 * Mapping has changed, invalidate old range and fall through to
3088 * handle validating new mapping.
3090 * We always interlock pte removals.
3094 /* XXX pmap_remove_pv_pte() unwires pt_pv */
3095 vm_page_wire_quick(pt_pv->pv_m);
3096 if (prot & VM_PROT_NOSYNC)
3097 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3099 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3101 pmap_remove_pv_page(pte_pv, 0);
3102 } else if (prot & VM_PROT_NOSYNC) {
3104 cpu_invlpg((void *)va);
3105 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3107 pmap_inval_interlock(&info, pmap, va);
3109 pmap_inval_deinterlock(&info, pmap);
3110 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3112 KKASSERT(*ptep == 0);
3116 * Enter on the PV list if part of our managed memory. Wiring is
3117 * handled automatically.
3120 KKASSERT(pte_pv->pv_m == NULL);
3121 vm_page_spin_lock(m);
3123 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3126 atomic_add_int(&m->object->agg_pv_list_count, 1);
3128 vm_page_flag_set(m, PG_MAPPED);
3129 vm_page_spin_unlock(m);
3131 } else if (pt_pv && opa == 0) {
3132 vm_page_wire_quick(pt_pv->pv_m);
3136 * Increment counters
3139 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3143 * Now validate mapping with desired protection/wiring.
3145 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V);
3149 if (va < VM_MAX_USER_ADDRESS)
3151 if (pmap == &kernel_pmap)
3155 * If the mapping or permission bits are different, we need
3156 * to update the pte.
3158 * We do not have to interlock pte insertions as no other
3159 * cpu will have a TLB entry.
3161 if ((origpte & ~(PG_M|PG_A)) != newpte) {
3163 if ((prot & VM_PROT_NOSYNC) == 0)
3164 pmap_inval_interlock(&info, pmap, va);
3166 *ptep = newpte | PG_A;
3167 cpu_invlpg((void *)va);
3169 if (prot & VM_PROT_NOSYNC)
3170 cpu_invlpg((void *)va);
3172 pmap_inval_deinterlock(&info, pmap);
3175 vm_page_flag_set(m, PG_WRITEABLE);
3177 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3180 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
3181 if ((prot & VM_PROT_NOSYNC) == 0)
3182 pmap_inval_done(&info);
3185 * Cleanup the pv entry, allowing other accessors.
3194 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
3195 * This code also assumes that the pmap has no pre-existing entry for this
3198 * This code currently may only be used on user pmaps, not kernel_pmap.
3201 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
3203 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE);
3207 * Make a temporary mapping for a physical address. This is only intended
3208 * to be used for panic dumps.
3210 * The caller is responsible for calling smp_invltlb().
3213 pmap_kenter_temporary(vm_paddr_t pa, long i)
3215 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
3216 return ((void *)crashdumpmap);
3219 #define MAX_INIT_PT (96)
3222 * This routine preloads the ptes for a given object into the specified pmap.
3223 * This eliminates the blast of soft faults on process startup and
3224 * immediately after an mmap.
3226 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
3229 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
3230 vm_object_t object, vm_pindex_t pindex,
3231 vm_size_t size, int limit)
3233 struct rb_vm_page_scan_info info;
3238 * We can't preinit if read access isn't set or there is no pmap
3241 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
3245 * We can't preinit if the pmap is not the current pmap
3247 lp = curthread->td_lwp;
3248 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
3251 psize = x86_64_btop(size);
3253 if ((object->type != OBJT_VNODE) ||
3254 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
3255 (object->resident_page_count > MAX_INIT_PT))) {
3259 if (pindex + psize > object->size) {
3260 if (object->size < pindex)
3262 psize = object->size - pindex;
3269 * Use a red-black scan to traverse the requested range and load
3270 * any valid pages found into the pmap.
3272 * We cannot safely scan the object's memq without holding the
3275 info.start_pindex = pindex;
3276 info.end_pindex = pindex + psize - 1;
3282 vm_object_hold(object);
3283 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3284 pmap_object_init_pt_callback, &info);
3285 vm_object_drop(object);
3290 pmap_object_init_pt_callback(vm_page_t p, void *data)
3292 struct rb_vm_page_scan_info *info = data;
3293 vm_pindex_t rel_index;
3296 * don't allow an madvise to blow away our really
3297 * free pages allocating pv entries.
3299 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3300 vmstats.v_free_count < vmstats.v_free_reserved) {
3303 if (vm_page_busy_try(p, TRUE))
3305 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3306 (p->flags & PG_FICTITIOUS) == 0) {
3307 if ((p->queue - p->pc) == PQ_CACHE)
3308 vm_page_deactivate(p);
3309 rel_index = p->pindex - info->start_pindex;
3310 pmap_enter_quick(info->pmap,
3311 info->addr + x86_64_ptob(rel_index), p);
3319 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
3322 * Returns FALSE if it would be non-trivial or if a pte is already loaded
3326 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3330 spin_lock(&pmap->pm_spin);
3331 if ((pte = pmap_pte(pmap, addr)) != NULL) {
3333 spin_unlock(&pmap->pm_spin);
3337 spin_unlock(&pmap->pm_spin);
3342 * Change the wiring attribute for a pmap/va pair. The mapping must already
3343 * exist in the pmap. The mapping may or may not be managed.
3346 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
3353 lwkt_gettoken(&pmap->pm_token);
3354 pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
3355 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
3357 if (wired && !pmap_pte_w(ptep))
3358 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3359 else if (!wired && pmap_pte_w(ptep))
3360 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3363 * Wiring is not a hardware characteristic so there is no need to
3364 * invalidate TLB. However, in an SMP environment we must use
3365 * a locked bus cycle to update the pte (if we are not using
3366 * the pmap_inval_*() API that is)... it's ok to do this for simple
3371 atomic_set_long(ptep, PG_W);
3373 atomic_clear_long(ptep, PG_W);
3376 atomic_set_long_nonlocked(ptep, PG_W);
3378 atomic_clear_long_nonlocked(ptep, PG_W);
3381 lwkt_reltoken(&pmap->pm_token);
3387 * Copy the range specified by src_addr/len from the source map to
3388 * the range dst_addr/len in the destination map.
3390 * This routine is only advisory and need not do anything.
3393 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3394 vm_size_t len, vm_offset_t src_addr)
3401 * Zero the specified physical page.
3403 * This function may be called from an interrupt and no locking is
3407 pmap_zero_page(vm_paddr_t phys)
3409 vm_offset_t va = PHYS_TO_DMAP(phys);
3411 pagezero((void *)va);
3415 * pmap_page_assertzero:
3417 * Assert that a page is empty, panic if it isn't.
3420 pmap_page_assertzero(vm_paddr_t phys)
3422 vm_offset_t va = PHYS_TO_DMAP(phys);
3425 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3426 if (*(long *)((char *)va + i) != 0) {
3427 panic("pmap_page_assertzero() @ %p not zero!\n",
3428 (void *)(intptr_t)va);
3436 * Zero part of a physical page by mapping it into memory and clearing
3437 * its contents with bzero.
3439 * off and size may not cover an area beyond a single hardware page.
3442 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3444 vm_offset_t virt = PHYS_TO_DMAP(phys);
3446 bzero((char *)virt + off, size);
3452 * Copy the physical page from the source PA to the target PA.
3453 * This function may be called from an interrupt. No locking
3457 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3459 vm_offset_t src_virt, dst_virt;
3461 src_virt = PHYS_TO_DMAP(src);
3462 dst_virt = PHYS_TO_DMAP(dst);
3463 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3467 * pmap_copy_page_frag:
3469 * Copy the physical page from the source PA to the target PA.
3470 * This function may be called from an interrupt. No locking
3474 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3476 vm_offset_t src_virt, dst_virt;
3478 src_virt = PHYS_TO_DMAP(src);
3479 dst_virt = PHYS_TO_DMAP(dst);
3481 bcopy((char *)src_virt + (src & PAGE_MASK),
3482 (char *)dst_virt + (dst & PAGE_MASK),
3487 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
3488 * this page. This count may be changed upwards or downwards in the future;
3489 * it is only necessary that true be returned for a small subset of pmaps
3490 * for proper page aging.
3493 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3498 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3501 vm_page_spin_lock(m);
3502 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3503 if (pv->pv_pmap == pmap) {
3504 vm_page_spin_unlock(m);
3511 vm_page_spin_unlock(m);
3516 * Remove all pages from specified address space this aids process exit
3517 * speeds. Also, this code may be special cased for the current process
3521 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3523 pmap_remove(pmap, sva, eva);
3527 * pmap_testbit tests bits in pte's note that the testbit/clearbit
3528 * routines are inline, and a lot of things compile-time evaluate.
3532 pmap_testbit(vm_page_t m, int bit)
3537 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3540 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3542 vm_page_spin_lock(m);
3543 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
3544 vm_page_spin_unlock(m);
3548 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3550 * if the bit being tested is the modified bit, then
3551 * mark clean_map and ptes as never
3554 if (bit & (PG_A|PG_M)) {
3555 if (!pmap_track_modified(pv->pv_pindex))
3559 #if defined(PMAP_DIAGNOSTIC)
3560 if (pv->pv_pmap == NULL) {
3561 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
3566 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3568 vm_page_spin_unlock(m);
3572 vm_page_spin_unlock(m);
3577 * This routine is used to modify bits in ptes
3579 * Caller must NOT hold any spin locks
3583 pmap_clearbit(vm_page_t m, int bit)
3585 struct pmap_inval_info info;
3589 vm_pindex_t save_pindex;
3593 vm_page_flag_clear(m, PG_WRITEABLE);
3594 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
3598 pmap_inval_init(&info);
3601 * Loop over all current mappings setting/clearing as appropos If
3602 * setting RO do we need to clear the VAC?
3604 vm_page_spin_lock(m);
3606 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3608 * don't write protect pager mappings
3611 if (!pmap_track_modified(pv->pv_pindex))
3615 #if defined(PMAP_DIAGNOSTIC)
3616 if (pv->pv_pmap == NULL) {
3617 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
3624 * Careful here. We can use a locked bus instruction to
3625 * clear PG_A or PG_M safely but we need to synchronize
3626 * with the target cpus when we mess with PG_RW.
3628 * We do not have to force synchronization when clearing
3629 * PG_M even for PTEs generated via virtual memory maps,
3630 * because the virtual kernel will invalidate the pmap
3631 * entry when/if it needs to resynchronize the Modify bit.
3634 save_pmap = pv->pv_pmap;
3635 save_pindex = pv->pv_pindex;
3637 vm_page_spin_unlock(m);
3638 pmap_inval_interlock(&info, save_pmap,
3639 (vm_offset_t)save_pindex << PAGE_SHIFT);
3640 vm_page_spin_lock(m);
3641 if (pv->pv_pmap == NULL) {
3647 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3654 atomic_clear_long(pte, PG_M|PG_RW);
3657 * The cpu may be trying to set PG_M
3658 * simultaniously with our clearing
3661 if (!atomic_cmpset_long(pte, pbits,
3665 } else if (bit == PG_M) {
3667 * We could also clear PG_RW here to force
3668 * a fault on write to redetect PG_M for
3669 * virtual kernels, but it isn't necessary
3670 * since virtual kernels invalidate the pte
3671 * when they clear the VPTE_M bit in their
3672 * virtual page tables.
3674 atomic_clear_long(pte, PG_M);
3676 atomic_clear_long(pte, bit);
3680 save_pmap = pv->pv_pmap;
3682 vm_page_spin_unlock(m);
3683 pmap_inval_deinterlock(&info, save_pmap);
3684 vm_page_spin_lock(m);
3685 if (pv->pv_pmap == NULL) {
3692 vm_page_spin_unlock(m);
3693 pmap_inval_done(&info);
3697 * Lower the permission for all mappings to a given page.
3699 * Page must be busied by caller.
3702 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3704 /* JG NX support? */
3705 if ((prot & VM_PROT_WRITE) == 0) {
3706 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3708 * NOTE: pmap_clearbit(.. PG_RW) also clears
3709 * the PG_WRITEABLE flag in (m).
3711 pmap_clearbit(m, PG_RW);
3719 pmap_phys_address(vm_pindex_t ppn)
3721 return (x86_64_ptob(ppn));
3725 * Return a count of reference bits for a page, clearing those bits.
3726 * It is not necessary for every reference bit to be cleared, but it
3727 * is necessary that 0 only be returned when there are truly no
3728 * reference bits set.
3730 * XXX: The exact number of bits to check and clear is a matter that
3731 * should be tested and standardized at some point in the future for
3732 * optimal aging of shared pages.
3734 * This routine may not block.
3737 pmap_ts_referenced(vm_page_t m)
3743 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3746 vm_page_spin_lock(m);
3747 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3748 if (!pmap_track_modified(pv->pv_pindex))
3750 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3751 if (pte && (*pte & PG_A)) {
3753 atomic_clear_long(pte, PG_A);
3755 atomic_clear_long_nonlocked(pte, PG_A);
3762 vm_page_spin_unlock(m);
3769 * Return whether or not the specified physical page was modified
3770 * in any physical maps.
3773 pmap_is_modified(vm_page_t m)
3777 res = pmap_testbit(m, PG_M);
3782 * Clear the modify bits on the specified physical page.
3785 pmap_clear_modify(vm_page_t m)
3787 pmap_clearbit(m, PG_M);
3791 * pmap_clear_reference:
3793 * Clear the reference bit on the specified physical page.
3796 pmap_clear_reference(vm_page_t m)
3798 pmap_clearbit(m, PG_A);
3802 * Miscellaneous support routines follow
3807 i386_protection_init(void)
3811 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3812 kp = protection_codes;
3813 for (prot = 0; prot < 8; prot++) {
3815 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3817 * Read access is also 0. There isn't any execute bit,
3818 * so just make it readable.
3820 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3821 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3822 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3825 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3826 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3827 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3828 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3836 * Map a set of physical memory pages into the kernel virtual
3837 * address space. Return a pointer to where it is mapped. This
3838 * routine is intended to be used for mapping device memory,
3841 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3845 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3847 vm_offset_t va, tmpva, offset;
3850 offset = pa & PAGE_MASK;
3851 size = roundup(offset + size, PAGE_SIZE);
3853 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3855 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3857 pa = pa & ~PAGE_MASK;
3858 for (tmpva = va; size > 0;) {
3859 pte = vtopte(tmpva);
3860 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3868 return ((void *)(va + offset));
3872 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3874 vm_offset_t va, tmpva, offset;
3877 offset = pa & PAGE_MASK;
3878 size = roundup(offset + size, PAGE_SIZE);
3880 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3882 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3884 pa = pa & ~PAGE_MASK;
3885 for (tmpva = va; size > 0;) {
3886 pte = vtopte(tmpva);
3887 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3895 return ((void *)(va + offset));
3899 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3901 vm_offset_t base, offset;
3903 base = va & ~PAGE_MASK;
3904 offset = va & PAGE_MASK;
3905 size = roundup(offset + size, PAGE_SIZE);
3906 pmap_qremove(va, size >> PAGE_SHIFT);
3907 kmem_free(&kernel_map, base, size);
3911 * perform the pmap work for mincore
3914 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3916 pt_entry_t *ptep, pte;
3920 lwkt_gettoken(&pmap->pm_token);
3921 ptep = pmap_pte(pmap, addr);
3923 if (ptep && (pte = *ptep) != 0) {
3926 val = MINCORE_INCORE;
3927 if ((pte & PG_MANAGED) == 0)
3930 pa = pte & PG_FRAME;
3932 m = PHYS_TO_VM_PAGE(pa);
3938 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3940 * Modified by someone
3942 else if (m->dirty || pmap_is_modified(m))
3943 val |= MINCORE_MODIFIED_OTHER;
3948 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3951 * Referenced by someone
3953 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3954 val |= MINCORE_REFERENCED_OTHER;
3955 vm_page_flag_set(m, PG_REFERENCED);
3959 lwkt_reltoken(&pmap->pm_token);
3965 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3966 * vmspace will be ref'd and the old one will be deref'd.
3968 * The vmspace for all lwps associated with the process will be adjusted
3969 * and cr3 will be reloaded if any lwp is the current lwp.
3971 * The process must hold the vmspace->vm_map.token for oldvm and newvm
3974 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3976 struct vmspace *oldvm;
3979 oldvm = p->p_vmspace;
3980 if (oldvm != newvm) {
3982 sysref_get(&newvm->vm_sysref);
3983 p->p_vmspace = newvm;
3984 KKASSERT(p->p_nthreads == 1);
3985 lp = RB_ROOT(&p->p_lwp_tree);
3986 pmap_setlwpvm(lp, newvm);
3988 sysref_put(&oldvm->vm_sysref);
3993 * Set the vmspace for a LWP. The vmspace is almost universally set the
3994 * same as the process vmspace, but virtual kernels need to swap out contexts
3995 * on a per-lwp basis.
3997 * Caller does not necessarily hold any vmspace tokens. Caller must control
3998 * the lwp (typically be in the context of the lwp). We use a critical
3999 * section to protect against statclock and hardclock (statistics collection).
4002 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
4004 struct vmspace *oldvm;
4007 oldvm = lp->lwp_vmspace;
4009 if (oldvm != newvm) {
4011 lp->lwp_vmspace = newvm;
4012 if (curthread->td_lwp == lp) {
4013 pmap = vmspace_pmap(newvm);
4015 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4016 if (pmap->pm_active & CPUMASK_LOCK)
4017 pmap_interlock_wait(newvm);
4019 pmap->pm_active |= 1;
4021 #if defined(SWTCH_OPTIM_STATS)
4024 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4025 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
4026 load_cr3(curthread->td_pcb->pcb_cr3);
4027 pmap = vmspace_pmap(oldvm);
4029 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4031 pmap->pm_active &= ~(cpumask_t)1;
4041 * Called when switching to a locked pmap, used to interlock against pmaps
4042 * undergoing modifications to prevent us from activating the MMU for the
4043 * target pmap until all such modifications have completed. We have to do
4044 * this because the thread making the modifications has already set up its
4045 * SMP synchronization mask.
4050 pmap_interlock_wait(struct vmspace *vm)
4052 struct pmap *pmap = &vm->vm_pmap;
4054 if (pmap->pm_active & CPUMASK_LOCK) {
4056 DEBUG_PUSH_INFO("pmap_interlock_wait");
4057 while (pmap->pm_active & CPUMASK_LOCK) {
4059 lwkt_process_ipiq();
4069 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4072 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4076 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4081 * Used by kmalloc/kfree, page already exists at va
4084 pmap_kvtom(vm_offset_t va)
4086 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));