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 2000
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);
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 */
565 * We are running (mostly) V=P at this point
567 * Calculate NKPT - number of kernel page tables. We have to
568 * accomodoate prealloction of the vm_page_array, dump bitmap,
569 * MSGBUF_SIZE, and other stuff. Be generous.
571 * Maxmem is in pages.
573 * ndmpdp is the number of 1GB pages we wish to map.
575 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
576 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
578 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
581 * Starting at the beginning of kvm (not KERNBASE).
583 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
584 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
585 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
586 ndmpdp) + 511) / 512;
590 * Starting at KERNBASE - map 2G worth of page table pages.
591 * KERNBASE is offset -2G from the end of kvm.
593 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
598 KPTbase = allocpages(firstaddr, nkpt_base);
599 KPTphys = allocpages(firstaddr, nkpt_phys);
600 KPML4phys = allocpages(firstaddr, 1);
601 KPDPphys = allocpages(firstaddr, NKPML4E);
602 KPDphys = allocpages(firstaddr, NKPDPE);
605 * Calculate the page directory base for KERNBASE,
606 * that is where we start populating the page table pages.
607 * Basically this is the end - 2.
609 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
611 DMPDPphys = allocpages(firstaddr, NDMPML4E);
612 if ((amd_feature & AMDID_PAGE1GB) == 0)
613 DMPDphys = allocpages(firstaddr, ndmpdp);
614 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
617 * Fill in the underlying page table pages for the area around
618 * KERNBASE. This remaps low physical memory to KERNBASE.
620 * Read-only from zero to physfree
621 * XXX not fully used, underneath 2M pages
623 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
624 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
625 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
629 * Now map the initial kernel page tables. One block of page
630 * tables is placed at the beginning of kernel virtual memory,
631 * and another block is placed at KERNBASE to map the kernel binary,
632 * data, bss, and initial pre-allocations.
634 for (i = 0; i < nkpt_base; i++) {
635 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
636 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
638 for (i = 0; i < nkpt_phys; i++) {
639 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
640 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
644 * Map from zero to end of allocations using 2M pages as an
645 * optimization. This will bypass some of the KPTBase pages
646 * above in the KERNBASE area.
648 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
649 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
650 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
654 * And connect up the PD to the PDP. The kernel pmap is expected
655 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
657 for (i = 0; i < NKPDPE; i++) {
658 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
659 KPDphys + (i << PAGE_SHIFT);
660 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
665 * Now set up the direct map space using either 2MB or 1GB pages
666 * Preset PG_M and PG_A because demotion expects it.
668 * When filling in entries in the PD pages make sure any excess
669 * entries are set to zero as we allocated enough PD pages
671 if ((amd_feature & AMDID_PAGE1GB) == 0) {
672 for (i = 0; i < NPDEPG * ndmpdp; i++) {
673 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
674 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
679 * And the direct map space's PDP
681 for (i = 0; i < ndmpdp; i++) {
682 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
684 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
687 for (i = 0; i < ndmpdp; i++) {
688 ((pdp_entry_t *)DMPDPphys)[i] =
689 (vm_paddr_t)i << PDPSHIFT;
690 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
695 /* And recursively map PML4 to itself in order to get PTmap */
696 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
697 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
700 * Connect the Direct Map slots up to the PML4
702 for (j = 0; j < NDMPML4E; ++j) {
703 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
704 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
709 * Connect the KVA slot up to the PML4
711 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
712 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
716 * Bootstrap the system enough to run with virtual memory.
718 * On the i386 this is called after mapping has already been enabled
719 * and just syncs the pmap module with what has already been done.
720 * [We can't call it easily with mapping off since the kernel is not
721 * mapped with PA == VA, hence we would have to relocate every address
722 * from the linked base (virtual) address "KERNBASE" to the actual
723 * (physical) address starting relative to 0]
726 pmap_bootstrap(vm_paddr_t *firstaddr)
730 struct mdglobaldata *gd;
733 KvaStart = VM_MIN_KERNEL_ADDRESS;
734 KvaEnd = VM_MAX_KERNEL_ADDRESS;
735 KvaSize = KvaEnd - KvaStart;
737 avail_start = *firstaddr;
740 * Create an initial set of page tables to run the kernel in.
742 create_pagetables(firstaddr);
744 virtual2_start = KvaStart;
745 virtual2_end = PTOV_OFFSET;
747 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
748 virtual_start = pmap_kmem_choose(virtual_start);
750 virtual_end = VM_MAX_KERNEL_ADDRESS;
752 /* XXX do %cr0 as well */
753 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
757 * Initialize protection array.
759 i386_protection_init();
762 * The kernel's pmap is statically allocated so we don't have to use
763 * pmap_create, which is unlikely to work correctly at this part of
764 * the boot sequence (XXX and which no longer exists).
766 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
767 kernel_pmap.pm_count = 1;
768 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
769 RB_INIT(&kernel_pmap.pm_pvroot);
770 spin_init(&kernel_pmap.pm_spin);
771 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
774 * Reserve some special page table entries/VA space for temporary
777 #define SYSMAP(c, p, v, n) \
778 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
784 * CMAP1/CMAP2 are used for zeroing and copying pages.
786 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
791 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
794 * ptvmmap is used for reading arbitrary physical pages via
797 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
800 * msgbufp is used to map the system message buffer.
801 * XXX msgbufmap is not used.
803 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
804 atop(round_page(MSGBUF_SIZE)))
811 * PG_G is terribly broken on SMP because we IPI invltlb's in some
812 * cases rather then invl1pg. Actually, I don't even know why it
813 * works under UP because self-referential page table mappings
818 if (cpu_feature & CPUID_PGE)
823 * Initialize the 4MB page size flag
827 * The 4MB page version of the initial
828 * kernel page mapping.
832 #if !defined(DISABLE_PSE)
833 if (cpu_feature & CPUID_PSE) {
836 * Note that we have enabled PSE mode
839 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
840 ptditmp &= ~(NBPDR - 1);
841 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
846 * Enable the PSE mode. If we are SMP we can't do this
847 * now because the APs will not be able to use it when
850 load_cr4(rcr4() | CR4_PSE);
853 * We can do the mapping here for the single processor
854 * case. We simply ignore the old page table page from
858 * For SMP, we still need 4K pages to bootstrap APs,
859 * PSE will be enabled as soon as all APs are up.
861 PTD[KPTDI] = (pd_entry_t)ptditmp;
868 * We need to finish setting up the globaldata page for the BSP.
869 * locore has already populated the page table for the mdglobaldata
872 pg = MDGLOBALDATA_BASEALLOC_PAGES;
873 gd = &CPU_prvspace[0].mdglobaldata;
880 * Set 4mb pdir for mp startup
885 if (pseflag && (cpu_feature & CPUID_PSE)) {
886 load_cr4(rcr4() | CR4_PSE);
887 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
895 * Initialize the pmap module.
896 * Called by vm_init, to initialize any structures that the pmap
897 * system needs to map virtual memory.
898 * pmap_init has been enhanced to support in a fairly consistant
899 * way, discontiguous physical memory.
908 * Allocate memory for random pmap data structures. Includes the
912 for (i = 0; i < vm_page_array_size; i++) {
915 m = &vm_page_array[i];
916 TAILQ_INIT(&m->md.pv_list);
920 * init the pv free list
922 initial_pvs = vm_page_array_size;
923 if (initial_pvs < MINPV)
925 pvzone = &pvzone_store;
926 pvinit = (void *)kmem_alloc(&kernel_map,
927 initial_pvs * sizeof (struct pv_entry));
928 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
929 pvinit, initial_pvs);
932 * Now it is safe to enable pv_table recording.
934 pmap_initialized = TRUE;
938 * Initialize the address space (zone) for the pv_entries. Set a
939 * high water mark so that the system can recover from excessive
940 * numbers of pv entries.
945 int shpgperproc = PMAP_SHPGPERPROC;
948 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
949 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
950 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
951 pv_entry_high_water = 9 * (pv_entry_max / 10);
954 * Subtract out pages already installed in the zone (hack)
956 entry_max = pv_entry_max - vm_page_array_size;
960 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
964 /***************************************************
965 * Low level helper routines.....
966 ***************************************************/
969 * this routine defines the region(s) of memory that should
970 * not be tested for the modified bit.
974 pmap_track_modified(vm_pindex_t pindex)
976 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
977 if ((va < clean_sva) || (va >= clean_eva))
984 * Extract the physical page address associated with the map/VA pair.
985 * The page must be wired for this to work reliably.
987 * XXX for the moment we're using pv_find() instead of pv_get(), as
988 * callers might be expecting non-blocking operation.
991 pmap_extract(pmap_t pmap, vm_offset_t va)
998 if (va >= VM_MAX_USER_ADDRESS) {
1000 * Kernel page directories might be direct-mapped and
1001 * there is typically no PV tracking of pte's
1005 pt = pmap_pt(pmap, va);
1006 if (pt && (*pt & PG_V)) {
1008 rtval = *pt & PG_PS_FRAME;
1009 rtval |= va & PDRMASK;
1011 ptep = pmap_pt_to_pte(pt, va);
1013 rtval = *ptep & PG_FRAME;
1014 rtval |= va & PAGE_MASK;
1020 * User pages currently do not direct-map the page directory
1021 * and some pages might not used managed PVs. But all PT's
1024 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1026 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1028 rtval = *ptep & PG_FRAME;
1029 rtval |= va & PAGE_MASK;
1038 * Extract the physical page address associated kernel virtual address.
1041 pmap_kextract(vm_offset_t va)
1043 pd_entry_t pt; /* pt entry in pd */
1046 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1047 pa = DMAP_TO_PHYS(va);
1051 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1054 * Beware of a concurrent promotion that changes the
1055 * PDE at this point! For example, vtopte() must not
1056 * be used to access the PTE because it would use the
1057 * new PDE. It is, however, safe to use the old PDE
1058 * because the page table page is preserved by the
1061 pa = *pmap_pt_to_pte(&pt, va);
1062 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1068 /***************************************************
1069 * Low level mapping routines.....
1070 ***************************************************/
1073 * Routine: pmap_kenter
1075 * Add a wired page to the KVA
1076 * NOTE! note that in order for the mapping to take effect -- you
1077 * should do an invltlb after doing the pmap_kenter().
1080 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1084 pmap_inval_info info;
1086 pmap_inval_init(&info); /* XXX remove */
1087 npte = pa | PG_RW | PG_V | pgeflag;
1089 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1091 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1092 pmap_inval_done(&info); /* XXX remove */
1096 * Routine: pmap_kenter_quick
1098 * Similar to pmap_kenter(), except we only invalidate the
1099 * mapping on the current CPU.
1102 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1107 npte = pa | PG_RW | PG_V | pgeflag;
1110 cpu_invlpg((void *)va);
1114 pmap_kenter_sync(vm_offset_t va)
1116 pmap_inval_info info;
1118 pmap_inval_init(&info);
1119 pmap_inval_interlock(&info, &kernel_pmap, va);
1120 pmap_inval_deinterlock(&info, &kernel_pmap);
1121 pmap_inval_done(&info);
1125 pmap_kenter_sync_quick(vm_offset_t va)
1127 cpu_invlpg((void *)va);
1131 * remove a page from the kernel pagetables
1134 pmap_kremove(vm_offset_t va)
1137 pmap_inval_info info;
1139 pmap_inval_init(&info);
1141 pmap_inval_interlock(&info, &kernel_pmap, va);
1142 (void)pte_load_clear(pte);
1143 pmap_inval_deinterlock(&info, &kernel_pmap);
1144 pmap_inval_done(&info);
1148 pmap_kremove_quick(vm_offset_t va)
1152 (void)pte_load_clear(pte);
1153 cpu_invlpg((void *)va);
1157 * XXX these need to be recoded. They are not used in any critical path.
1160 pmap_kmodify_rw(vm_offset_t va)
1162 atomic_set_long(vtopte(va), PG_RW);
1163 cpu_invlpg((void *)va);
1167 pmap_kmodify_nc(vm_offset_t va)
1169 atomic_set_long(vtopte(va), PG_N);
1170 cpu_invlpg((void *)va);
1174 * Used to map a range of physical addresses into kernel virtual
1175 * address space during the low level boot, typically to map the
1176 * dump bitmap, message buffer, and vm_page_array.
1178 * These mappings are typically made at some pointer after the end of the
1181 * We could return PHYS_TO_DMAP(start) here and not allocate any
1182 * via (*virtp), but then kmem from userland and kernel dumps won't
1183 * have access to the related pointers.
1186 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1189 vm_offset_t va_start;
1191 /*return PHYS_TO_DMAP(start);*/
1196 while (start < end) {
1197 pmap_kenter_quick(va, start);
1207 * Add a list of wired pages to the kva
1208 * this routine is only used for temporary
1209 * kernel mappings that do not need to have
1210 * page modification or references recorded.
1211 * Note that old mappings are simply written
1212 * over. The page *must* be wired.
1215 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1219 end_va = va + count * PAGE_SIZE;
1221 while (va < end_va) {
1225 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1226 cpu_invlpg((void *)va);
1234 * This routine jerks page mappings from the
1235 * kernel -- it is meant only for temporary mappings.
1237 * MPSAFE, INTERRUPT SAFE (cluster callback)
1240 pmap_qremove(vm_offset_t va, int count)
1244 end_va = va + count * PAGE_SIZE;
1246 while (va < end_va) {
1250 (void)pte_load_clear(pte);
1251 cpu_invlpg((void *)va);
1258 * Create a new thread and optionally associate it with a (new) process.
1259 * NOTE! the new thread's cpu may not equal the current cpu.
1262 pmap_init_thread(thread_t td)
1264 /* enforce pcb placement & alignment */
1265 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1266 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1267 td->td_savefpu = &td->td_pcb->pcb_save;
1268 td->td_sp = (char *)td->td_pcb; /* no -16 */
1272 * This routine directly affects the fork perf for a process.
1275 pmap_init_proc(struct proc *p)
1280 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1281 * it, and IdlePTD, represents the template used to update all other pmaps.
1283 * On architectures where the kernel pmap is not integrated into the user
1284 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1285 * kernel_pmap should be used to directly access the kernel_pmap.
1288 pmap_pinit0(struct pmap *pmap)
1290 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1292 pmap->pm_active = 0;
1293 pmap->pm_pvhint = NULL;
1294 RB_INIT(&pmap->pm_pvroot);
1295 spin_init(&pmap->pm_spin);
1296 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1297 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1301 * Initialize a preallocated and zeroed pmap structure,
1302 * such as one in a vmspace structure.
1305 pmap_pinit(struct pmap *pmap)
1311 * Misc initialization
1314 pmap->pm_active = 0;
1315 pmap->pm_pvhint = NULL;
1316 if (pmap->pm_pmlpv == NULL) {
1317 RB_INIT(&pmap->pm_pvroot);
1318 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1319 spin_init(&pmap->pm_spin);
1320 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1324 * No need to allocate page table space yet but we do need a valid
1325 * page directory table.
1327 if (pmap->pm_pml4 == NULL) {
1329 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1333 * Allocate the page directory page, which wires it even though
1334 * it isn't being entered into some higher level page table (it
1335 * being the highest level). If one is already cached we don't
1336 * have to do anything.
1338 if ((pv = pmap->pm_pmlpv) == NULL) {
1339 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1340 pmap->pm_pmlpv = pv;
1341 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1342 VM_PAGE_TO_PHYS(pv->pv_m));
1346 * Install DMAP and KMAP.
1348 for (j = 0; j < NDMPML4E; ++j) {
1349 pmap->pm_pml4[DMPML4I + j] =
1350 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1351 PG_RW | PG_V | PG_U;
1353 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1356 * install self-referential address mapping entry
1358 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1359 PG_V | PG_RW | PG_A | PG_M;
1361 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1362 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1367 * Clean up a pmap structure so it can be physically freed. This routine
1368 * is called by the vmspace dtor function. A great deal of pmap data is
1369 * left passively mapped to improve vmspace management so we have a bit
1370 * of cleanup work to do here.
1373 pmap_puninit(pmap_t pmap)
1378 KKASSERT(pmap->pm_active == 0);
1379 if ((pv = pmap->pm_pmlpv) != NULL) {
1380 if (pv_hold_try(pv) == 0)
1382 p = pmap_remove_pv_page(pv);
1384 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1385 vm_page_busy_wait(p, FALSE, "pgpun");
1386 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1387 vm_page_unwire(p, 0);
1388 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1391 * XXX eventually clean out PML4 static entries and
1392 * use vm_page_free_zero()
1395 pmap->pm_pmlpv = NULL;
1397 if (pmap->pm_pml4) {
1398 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1399 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1400 pmap->pm_pml4 = NULL;
1402 KKASSERT(pmap->pm_stats.resident_count == 0);
1403 KKASSERT(pmap->pm_stats.wired_count == 0);
1407 * Wire in kernel global address entries. To avoid a race condition
1408 * between pmap initialization and pmap_growkernel, this procedure
1409 * adds the pmap to the master list (which growkernel scans to update),
1410 * then copies the template.
1413 pmap_pinit2(struct pmap *pmap)
1416 * XXX copies current process, does not fill in MPPTDI
1418 spin_lock(&pmap_spin);
1419 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1420 spin_unlock(&pmap_spin);
1424 * This routine is called when various levels in the page table need to
1425 * be populated. This routine cannot fail.
1427 * This function returns two locked pv_entry's, one representing the
1428 * requested pv and one representing the requested pv's parent pv. If
1429 * the pv did not previously exist it will be mapped into its parent
1430 * and wired, otherwise no additional wire count will be added.
1434 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1439 vm_pindex_t pt_pindex;
1444 * If the pv already exists and we aren't being asked for the
1445 * parent page table page we can just return it. A locked+held pv
1448 pv = pv_alloc(pmap, ptepindex, &isnew);
1449 if (isnew == 0 && pvpp == NULL)
1453 * This is a new PV, we have to resolve its parent page table and
1454 * add an additional wiring to the page if necessary.
1458 * Special case terminal PVs. These are not page table pages so
1459 * no vm_page is allocated (the caller supplied the vm_page). If
1460 * pvpp is non-NULL we are being asked to also removed the pt_pv
1463 * Note that pt_pv's are only returned for user VAs. We assert that
1464 * a pt_pv is not being requested for kernel VAs.
1466 if (ptepindex < pmap_pt_pindex(0)) {
1467 if (ptepindex >= NUPTE_USER)
1468 KKASSERT(pvpp == NULL);
1470 KKASSERT(pvpp != NULL);
1472 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1473 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1475 vm_page_wire_quick(pvp->pv_m);
1484 * Non-terminal PVs allocate a VM page to represent the page table,
1485 * so we have to resolve pvp and calculate ptepindex for the pvp
1486 * and then for the page table entry index in the pvp for
1489 if (ptepindex < pmap_pd_pindex(0)) {
1491 * pv is PT, pvp is PD
1493 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1494 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1495 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1502 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1503 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1504 } else if (ptepindex < pmap_pdp_pindex(0)) {
1506 * pv is PD, pvp is PDP
1508 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1509 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1510 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1517 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1518 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1519 } else if (ptepindex < pmap_pml4_pindex()) {
1521 * pv is PDP, pvp is the root pml4 table
1523 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1530 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1531 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1534 * pv represents the top-level PML4, there is no parent.
1542 * This code is only reached if isnew is TRUE and this is not a
1543 * terminal PV. We need to allocate a vm_page for the page table
1544 * at this level and enter it into the parent page table.
1546 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1549 m = vm_page_alloc(NULL, pv->pv_pindex,
1550 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1551 VM_ALLOC_INTERRUPT);
1556 vm_page_spin_lock(m);
1557 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1559 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1560 vm_page_spin_unlock(m);
1561 vm_page_unmanage(m); /* m must be spinunlocked */
1563 if ((m->flags & PG_ZERO) == 0) {
1564 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1568 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1571 m->valid = VM_PAGE_BITS_ALL;
1572 vm_page_flag_clear(m, PG_ZERO);
1573 vm_page_wire(m); /* wire for mapping in parent */
1576 * Wire the page into pvp, bump the wire-count for pvp's page table
1577 * page. Bump the resident_count for the pmap. There is no pvp
1578 * for the top level, address the pm_pml4[] array directly.
1580 * If the caller wants the parent we return it, otherwise
1581 * we just put it away.
1583 * No interlock is needed for pte 0 -> non-zero.
1586 vm_page_wire_quick(pvp->pv_m);
1587 ptep = pv_pte_lookup(pvp, ptepindex);
1588 KKASSERT((*ptep & PG_V) == 0);
1589 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1602 * Release any resources held by the given physical map.
1604 * Called when a pmap initialized by pmap_pinit is being released. Should
1605 * only be called if the map contains no valid mappings.
1607 * Caller must hold pmap->pm_token
1609 struct pmap_release_info {
1614 static int pmap_release_callback(pv_entry_t pv, void *data);
1617 pmap_release(struct pmap *pmap)
1619 struct pmap_release_info info;
1621 KASSERT(pmap->pm_active == 0,
1622 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1624 spin_lock(&pmap_spin);
1625 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1626 spin_unlock(&pmap_spin);
1629 * Pull pv's off the RB tree in order from low to high and release
1635 spin_lock(&pmap->pm_spin);
1636 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
1637 pmap_release_callback, &info);
1638 spin_unlock(&pmap->pm_spin);
1639 } while (info.retry);
1643 * One resident page (the pml4 page) should remain.
1644 * No wired pages should remain.
1646 KKASSERT(pmap->pm_stats.resident_count == 1);
1647 KKASSERT(pmap->pm_stats.wired_count == 0);
1651 pmap_release_callback(pv_entry_t pv, void *data)
1653 struct pmap_release_info *info = data;
1654 pmap_t pmap = info->pmap;
1657 if (pv_hold_try(pv)) {
1658 spin_unlock(&pmap->pm_spin);
1660 spin_unlock(&pmap->pm_spin);
1662 if (pv->pv_pmap != pmap) {
1664 spin_lock(&pmap->pm_spin);
1671 * The pmap is currently not spinlocked, pv is held+locked.
1672 * Remove the pv's page from its parent's page table. The
1673 * parent's page table page's wire_count will be decremented.
1675 pmap_remove_pv_pte(pv, NULL, NULL);
1678 * Terminal pvs are unhooked from their vm_pages. Because
1679 * terminal pages aren't page table pages they aren't wired
1680 * by us, so we have to be sure not to unwire them either.
1682 if (pv->pv_pindex < pmap_pt_pindex(0)) {
1683 pmap_remove_pv_page(pv);
1688 * We leave the top-level page table page cached, wired, and
1689 * mapped in the pmap until the dtor function (pmap_puninit())
1692 * Since we are leaving the top-level pv intact we need
1693 * to break out of what would otherwise be an infinite loop.
1695 if (pv->pv_pindex == pmap_pml4_pindex()) {
1697 spin_lock(&pmap->pm_spin);
1702 * For page table pages (other than the top-level page),
1703 * remove and free the vm_page. The representitive mapping
1704 * removed above by pmap_remove_pv_pte() did not undo the
1705 * last wire_count so we have to do that as well.
1707 p = pmap_remove_pv_page(pv);
1708 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);
1859 cpu_invlpg((void *)va);
1862 * Now update the vm_page_t
1864 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
1865 kprintf("remove_pte badpte %016lx %016lx %d\n",
1867 pv->pv_pindex < pmap_pt_pindex(0));
1869 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
1870 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
1873 if (pmap_track_modified(ptepindex))
1877 vm_page_flag_set(p, PG_REFERENCED);
1880 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1882 cpu_invlpg((void *)va);
1886 * Unwire the parent page table page. The wire_count cannot go below
1887 * 1 here because the parent page table page is itself still mapped.
1889 * XXX remove the assertions later.
1891 KKASSERT(pv->pv_m == p);
1892 if (pvp && vm_page_unwire_quick(pvp->pv_m))
1893 panic("pmap_remove_pv_pte: Insufficient wire_count");
1901 pmap_remove_pv_page(pv_entry_t pv)
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);
1921 * Grow the number of kernel page table entries, if needed.
1923 * This routine is always called to validate any address space
1924 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1925 * space below KERNBASE.
1928 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1931 vm_offset_t ptppaddr;
1933 pd_entry_t *pt, newpt;
1935 int update_kernel_vm_end;
1938 * bootstrap kernel_vm_end on first real VM use
1940 if (kernel_vm_end == 0) {
1941 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1943 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1944 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1945 ~(PAGE_SIZE * NPTEPG - 1);
1947 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1948 kernel_vm_end = kernel_map.max_offset;
1955 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1956 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1957 * do not want to force-fill 128G worth of page tables.
1959 if (kstart < KERNBASE) {
1960 if (kstart > kernel_vm_end)
1961 kstart = kernel_vm_end;
1962 KKASSERT(kend <= KERNBASE);
1963 update_kernel_vm_end = 1;
1965 update_kernel_vm_end = 0;
1968 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1969 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1971 if (kend - 1 >= kernel_map.max_offset)
1972 kend = kernel_map.max_offset;
1974 while (kstart < kend) {
1975 pt = pmap_pt(&kernel_pmap, kstart);
1977 /* We need a new PDP entry */
1978 nkpg = vm_page_alloc(NULL, nkpt,
1981 VM_ALLOC_INTERRUPT);
1983 panic("pmap_growkernel: no memory to grow "
1986 paddr = VM_PAGE_TO_PHYS(nkpg);
1987 if ((nkpg->flags & PG_ZERO) == 0)
1988 pmap_zero_page(paddr);
1989 vm_page_flag_clear(nkpg, PG_ZERO);
1990 newpd = (pdp_entry_t)
1991 (paddr | PG_V | PG_RW | PG_A | PG_M);
1992 *pmap_pd(&kernel_pmap, kstart) = newpd;
1994 continue; /* try again */
1996 if ((*pt & PG_V) != 0) {
1997 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1998 ~(PAGE_SIZE * NPTEPG - 1);
1999 if (kstart - 1 >= kernel_map.max_offset) {
2000 kstart = kernel_map.max_offset;
2007 * This index is bogus, but out of the way
2009 nkpg = vm_page_alloc(NULL, nkpt,
2012 VM_ALLOC_INTERRUPT);
2014 panic("pmap_growkernel: no memory to grow kernel");
2017 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2018 pmap_zero_page(ptppaddr);
2019 vm_page_flag_clear(nkpg, PG_ZERO);
2020 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2021 *pmap_pt(&kernel_pmap, kstart) = newpt;
2024 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2025 ~(PAGE_SIZE * NPTEPG - 1);
2027 if (kstart - 1 >= kernel_map.max_offset) {
2028 kstart = kernel_map.max_offset;
2034 * Only update kernel_vm_end for areas below KERNBASE.
2036 if (update_kernel_vm_end && kernel_vm_end < kstart)
2037 kernel_vm_end = kstart;
2041 * Retire the given physical map from service.
2042 * Should only be called if the map contains
2043 * no valid mappings.
2046 pmap_destroy(pmap_t pmap)
2053 lwkt_gettoken(&pmap->pm_token);
2054 count = --pmap->pm_count;
2056 pmap_release(pmap); /* eats pm_token */
2057 panic("destroying a pmap is not yet implemented");
2059 lwkt_reltoken(&pmap->pm_token);
2063 * Add a reference to the specified pmap.
2066 pmap_reference(pmap_t pmap)
2069 lwkt_gettoken(&pmap->pm_token);
2071 lwkt_reltoken(&pmap->pm_token);
2075 /***************************************************
2076 * page management routines.
2077 ***************************************************/
2080 * Hold a pv without locking it
2083 pv_hold(pv_entry_t pv)
2087 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2091 count = pv->pv_hold;
2093 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2100 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2101 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2104 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2105 * pv list via its page) must be held by the caller.
2108 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2112 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2115 pv->pv_line = lineno;
2121 count = pv->pv_hold;
2123 if ((count & PV_HOLD_LOCKED) == 0) {
2124 if (atomic_cmpset_int(&pv->pv_hold, count,
2125 (count + 1) | PV_HOLD_LOCKED)) {
2128 pv->pv_line = lineno;
2133 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2141 * Drop a previously held pv_entry which could not be locked, allowing its
2144 * Must not be called with a spinlock held as we might zfree() the pv if it
2145 * is no longer associated with a pmap and this was the last hold count.
2148 pv_drop(pv_entry_t pv)
2152 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2153 if (pv->pv_pmap == NULL)
2159 count = pv->pv_hold;
2161 KKASSERT((count & PV_HOLD_MASK) > 0);
2162 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2163 (PV_HOLD_LOCKED | 1));
2164 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2165 if (count == 1 && pv->pv_pmap == NULL)
2174 * Find or allocate the requested PV entry, returning a locked pv
2178 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2181 pv_entry_t pnew = NULL;
2183 spin_lock(&pmap->pm_spin);
2185 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2186 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2191 spin_unlock(&pmap->pm_spin);
2192 pnew = zalloc(pvzone);
2193 spin_lock(&pmap->pm_spin);
2196 pnew->pv_pmap = pmap;
2197 pnew->pv_pindex = pindex;
2198 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2200 pnew->pv_func = func;
2201 pnew->pv_line = lineno;
2203 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2204 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2205 spin_unlock(&pmap->pm_spin);
2210 spin_unlock(&pmap->pm_spin);
2211 zfree(pvzone, pnew);
2213 spin_lock(&pmap->pm_spin);
2216 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2217 spin_unlock(&pmap->pm_spin);
2221 spin_unlock(&pmap->pm_spin);
2222 _pv_lock(pv PMAP_DEBUG_COPY);
2223 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2228 spin_lock(&pmap->pm_spin);
2235 * Find the requested PV entry, returning a locked+held pv or NULL
2239 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2243 spin_lock(&pmap->pm_spin);
2248 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2249 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2253 spin_unlock(&pmap->pm_spin);
2256 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2257 pv_cache(pv, pindex);
2258 spin_unlock(&pmap->pm_spin);
2261 spin_unlock(&pmap->pm_spin);
2262 _pv_lock(pv PMAP_DEBUG_COPY);
2263 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2266 spin_lock(&pmap->pm_spin);
2271 * Lookup, hold, and attempt to lock (pmap,pindex).
2273 * If the entry does not exist NULL is returned and *errorp is set to 0
2275 * If the entry exists and could be successfully locked it is returned and
2276 * errorp is set to 0.
2278 * If the entry exists but could NOT be successfully locked it is returned
2279 * held and *errorp is set to 1.
2283 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2287 spin_lock(&pmap->pm_spin);
2288 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2289 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2291 spin_unlock(&pmap->pm_spin);
2295 if (pv_hold_try(pv)) {
2296 pv_cache(pv, pindex);
2297 spin_unlock(&pmap->pm_spin);
2299 return(pv); /* lock succeeded */
2301 spin_unlock(&pmap->pm_spin);
2303 return (pv); /* lock failed */
2307 * Find the requested PV entry, returning a held pv or NULL
2311 pv_find(pmap_t pmap, vm_pindex_t pindex)
2315 spin_lock(&pmap->pm_spin);
2317 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2318 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2320 spin_unlock(&pmap->pm_spin);
2324 pv_cache(pv, pindex);
2325 spin_unlock(&pmap->pm_spin);
2330 * Lock a held pv, keeping the hold count
2334 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2339 count = pv->pv_hold;
2341 if ((count & PV_HOLD_LOCKED) == 0) {
2342 if (atomic_cmpset_int(&pv->pv_hold, count,
2343 count | PV_HOLD_LOCKED)) {
2346 pv->pv_line = lineno;
2352 tsleep_interlock(pv, 0);
2353 if (atomic_cmpset_int(&pv->pv_hold, count,
2354 count | PV_HOLD_WAITING)) {
2356 kprintf("pv waiting on %s:%d\n",
2357 pv->pv_func, pv->pv_line);
2359 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2366 * Unlock a held and locked pv, keeping the hold count.
2370 pv_unlock(pv_entry_t pv)
2374 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2378 count = pv->pv_hold;
2380 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2381 (PV_HOLD_LOCKED | 1));
2382 if (atomic_cmpset_int(&pv->pv_hold, count,
2384 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2385 if (count & PV_HOLD_WAITING)
2393 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2394 * and the hold count drops to zero we will free it.
2396 * Caller should not hold any spin locks. We are protected from hold races
2397 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2398 * lock held. A pv cannot be located otherwise.
2402 pv_put(pv_entry_t pv)
2404 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2405 if (pv->pv_pmap == NULL)
2414 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2415 * pmap. Any pte operations must have already been completed.
2419 pv_free(pv_entry_t pv)
2423 KKASSERT(pv->pv_m == NULL);
2424 if ((pmap = pv->pv_pmap) != NULL) {
2425 spin_lock(&pmap->pm_spin);
2426 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2427 if (pmap->pm_pvhint == pv)
2428 pmap->pm_pvhint = NULL;
2429 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2432 spin_unlock(&pmap->pm_spin);
2438 * This routine is very drastic, but can save the system
2446 static int warningdone=0;
2448 if (pmap_pagedaemon_waken == 0)
2450 pmap_pagedaemon_waken = 0;
2451 if (warningdone < 5) {
2452 kprintf("pmap_collect: collecting pv entries -- "
2453 "suggest increasing PMAP_SHPGPERPROC\n");
2457 for (i = 0; i < vm_page_array_size; i++) {
2458 m = &vm_page_array[i];
2459 if (m->wire_count || m->hold_count)
2461 if (vm_page_busy_try(m, TRUE) == 0) {
2462 if (m->wire_count == 0 && m->hold_count == 0) {
2471 * Scan the pmap for active page table entries and issue a callback.
2472 * The callback must dispose of pte_pv.
2474 * NOTE: Unmanaged page table entries will not have a pte_pv
2476 * NOTE: Kernel page table entries will not have a pt_pv. That is, wiring
2477 * counts are not tracked in kernel page table pages.
2479 * It is assumed that the start and end are properly rounded to the page size.
2482 pmap_scan(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva,
2483 void (*func)(pmap_t, struct pmap_inval_info *,
2484 pv_entry_t, pv_entry_t, vm_offset_t,
2485 pt_entry_t *, void *),
2488 pv_entry_t pdp_pv; /* A page directory page PV */
2489 pv_entry_t pd_pv; /* A page directory PV */
2490 pv_entry_t pt_pv; /* A page table PV */
2491 pv_entry_t pte_pv; /* A page table entry PV */
2493 vm_offset_t va_next;
2494 struct pmap_inval_info info;
2501 * Hold the token for stability; if the pmap is empty we have nothing
2504 lwkt_gettoken(&pmap->pm_token);
2506 if (pmap->pm_stats.resident_count == 0) {
2507 lwkt_reltoken(&pmap->pm_token);
2512 pmap_inval_init(&info);
2515 * Special handling for removing one page, which is a very common
2516 * operation (it is?).
2517 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
2519 if (sva + PAGE_SIZE == eva) {
2520 if (sva >= VM_MAX_USER_ADDRESS) {
2522 * Kernel mappings do not track wire counts on
2526 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2530 * User mappings may or may not have a pte_pv but
2531 * will always have a pt_pv if the page is present.
2533 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2534 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2535 if (pt_pv == NULL) {
2536 KKASSERT(pte_pv == NULL);
2539 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
2543 * Unlike the pv_find() case below we actually
2544 * acquired a locked pv in this case so any
2545 * race should have been resolved. It is expected
2548 KKASSERT(pte_pv == NULL);
2549 } else if (pte_pv) {
2550 KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED|
2552 ("bad *ptep %016lx sva %016lx pte_pv %p",
2553 *ptep, sva, pte_pv));
2554 func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg);
2556 KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V,
2557 ("bad *ptep %016lx sva %016lx pte_pv NULL",
2559 func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg);
2564 pmap_inval_done(&info);
2565 lwkt_reltoken(&pmap->pm_token);
2570 * NOTE: kernel mappings do not track page table pages, only
2573 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
2574 * However, for the scan to be efficient we try to
2575 * cache items top-down.
2581 for (; sva < eva; sva = va_next) {
2583 if (sva >= VM_MAX_USER_ADDRESS) {
2594 if (pdp_pv == NULL) {
2595 pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva));
2596 } else if (pdp_pv->pv_pindex != pmap_pdp_pindex(sva)) {
2598 pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva));
2600 if (pdp_pv == NULL) {
2601 va_next = (sva + NBPML4) & ~PML4MASK;
2610 if (pd_pv == NULL) {
2615 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
2616 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
2622 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
2624 if (pd_pv == NULL) {
2625 va_next = (sva + NBPDP) & ~PDPMASK;
2634 if (pt_pv == NULL) {
2643 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2644 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
2654 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2658 * We will scan or skip a page table page so adjust va_next
2661 if (pt_pv == NULL) {
2662 va_next = (sva + NBPDR) & ~PDRMASK;
2669 * From this point in the loop testing pt_pv for non-NULL
2670 * means we are in UVM, else if it is NULL we are in KVM.
2673 va_next = (sva + NBPDR) & ~PDRMASK;
2678 * Limit our scan to either the end of the va represented
2679 * by the current page table page, or to the end of the
2680 * range being removed.
2682 * Scan the page table for pages. Some pages may not be
2683 * managed (might not have a pv_entry).
2685 * There is no page table management for kernel pages so
2686 * pt_pv will be NULL in that case, but otherwise pt_pv
2687 * is non-NULL, locked, and referenced.
2693 * At this point a non-NULL pt_pv means a UVA, and a NULL
2694 * pt_pv means a KVA.
2697 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
2701 while (sva < va_next) {
2703 * Acquire the related pte_pv, if any. If *ptep == 0
2704 * the related pte_pv should not exist, but if *ptep
2705 * is not zero the pte_pv may or may not exist (e.g.
2706 * will not exist for an unmanaged page).
2708 * However a multitude of races are possible here.
2710 * In addition, the (pt_pv, pte_pv) lock order is
2711 * backwards, so we have to be careful in aquiring
2712 * a properly locked pte_pv.
2716 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
2727 pv_put(pt_pv); /* must be non-NULL */
2729 pv_lock(pte_pv); /* safe to block now */
2732 pt_pv = pv_get(pmap,
2733 pmap_pt_pindex(sva));
2737 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2741 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
2745 kprintf("Unexpected non-NULL pte_pv "
2746 "%p pt_pv %p *ptep = %016lx\n",
2747 pte_pv, pt_pv, *ptep);
2748 panic("Unexpected non-NULL pte_pv");
2756 * Ready for the callback. The locked pte_pv (if any)
2757 * is consumed by the callback. pte_pv will exist if
2758 * the page is managed, and will not exist if it
2762 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
2764 ("bad *ptep %016lx sva %016lx "
2766 *ptep, sva, pte_pv));
2767 func(pmap, &info, pte_pv, pt_pv, sva,
2770 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
2772 ("bad *ptep %016lx sva %016lx "
2775 func(pmap, &info, pte_pv, pt_pv, sva,
2795 pmap_inval_done(&info);
2796 lwkt_reltoken(&pmap->pm_token);
2800 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2802 pmap_scan(pmap, sva, eva, pmap_remove_callback, NULL);
2806 pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
2807 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
2808 pt_entry_t *ptep, void *arg __unused)
2814 * This will also drop pt_pv's wire_count. Note that
2815 * terminal pages are not wired based on mmu presence.
2817 pmap_remove_pv_pte(pte_pv, pt_pv, info);
2818 pmap_remove_pv_page(pte_pv);
2822 * pt_pv's wire_count is still bumped by unmanaged pages
2823 * so we must decrement it manually.
2825 pmap_inval_interlock(info, pmap, va);
2826 pte = pte_load_clear(ptep);
2827 pmap_inval_deinterlock(info, pmap);
2829 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2830 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2831 if (pt_pv && vm_page_unwire_quick(pt_pv->pv_m))
2832 panic("pmap_remove: insufficient wirecount");
2837 * Removes this physical page from all physical maps in which it resides.
2838 * Reflects back modify bits to the pager.
2840 * This routine may not be called from an interrupt.
2844 pmap_remove_all(vm_page_t m)
2846 struct pmap_inval_info info;
2849 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2852 pmap_inval_init(&info);
2853 vm_page_spin_lock(m);
2854 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2855 KKASSERT(pv->pv_m == m);
2856 if (pv_hold_try(pv)) {
2857 vm_page_spin_unlock(m);
2859 vm_page_spin_unlock(m);
2861 if (pv->pv_m != m) {
2863 vm_page_spin_lock(m);
2868 * Holding no spinlocks, pv is locked.
2870 pmap_remove_pv_pte(pv, NULL, &info);
2871 pmap_remove_pv_page(pv);
2873 vm_page_spin_lock(m);
2875 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2876 vm_page_spin_unlock(m);
2877 pmap_inval_done(&info);
2883 * Set the physical protection on the specified range of this map
2886 * This function may not be called from an interrupt if the map is
2887 * not the kernel_pmap.
2890 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2892 /* JG review for NX */
2896 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2897 pmap_remove(pmap, sva, eva);
2900 if (prot & VM_PROT_WRITE)
2902 pmap_scan(pmap, sva, eva, pmap_protect_callback, &prot);
2907 pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
2908 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
2909 pt_entry_t *ptep, void *arg __unused)
2918 pmap_inval_interlock(info, pmap, va);
2925 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2926 KKASSERT(m == pte_pv->pv_m);
2927 vm_page_flag_set(m, PG_REFERENCED);
2931 if (pmap_track_modified(pte_pv->pv_pindex)) {
2933 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2940 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
2943 pmap_inval_deinterlock(info, pmap);
2949 * Insert the vm_page (m) at the virtual address (va), replacing any prior
2950 * mapping at that address. Set protection and wiring as requested.
2952 * NOTE: This routine MUST insert the page into the pmap now, it cannot
2956 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2959 pmap_inval_info info;
2960 pv_entry_t pt_pv; /* page table */
2961 pv_entry_t pte_pv; /* page table entry */
2964 pt_entry_t origpte, newpte;
2969 va = trunc_page(va);
2970 #ifdef PMAP_DIAGNOSTIC
2972 panic("pmap_enter: toobig");
2973 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2974 panic("pmap_enter: invalid to pmap_enter page table "
2975 "pages (va: 0x%lx)", va);
2977 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2978 kprintf("Warning: pmap_enter called on UVA with "
2981 db_print_backtrace();
2984 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2985 kprintf("Warning: pmap_enter called on KVA without"
2988 db_print_backtrace();
2993 * Get locked PV entries for our new page table entry (pte_pv)
2994 * and for its parent page table (pt_pv). We need the parent
2995 * so we can resolve the location of the ptep.
2997 * Only hardware MMU actions can modify the ptep out from
3000 * if (m) is fictitious or unmanaged we do not create a managing
3001 * pte_pv for it. Any pre-existing page's management state must
3002 * match (avoiding code complexity).
3004 * If the pmap is still being initialized we assume existing
3007 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3008 * pmap_allocpte() checks the
3010 if (pmap_initialized == FALSE) {
3014 } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) {
3016 if (va >= VM_MAX_USER_ADDRESS) {
3020 pt_pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
3021 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3023 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0);
3025 if (va >= VM_MAX_USER_ADDRESS) {
3027 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3030 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va),
3032 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3034 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED));
3037 pa = VM_PAGE_TO_PHYS(m);
3039 opa = origpte & PG_FRAME;
3041 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V | PG_A);
3044 if (va < VM_MAX_USER_ADDRESS)
3047 newpte |= PG_MANAGED;
3048 if (pmap == &kernel_pmap)
3052 * It is possible for multiple faults to occur in threaded
3053 * environments, the existing pte might be correct.
3055 if (((origpte ^ newpte) & ~(pt_entry_t)(PG_M|PG_A)) == 0)
3058 if ((prot & VM_PROT_NOSYNC) == 0)
3059 pmap_inval_init(&info);
3062 * Ok, either the address changed or the protection or wiring
3065 * Clear the current entry, interlocking the removal. For managed
3066 * pte's this will also flush the modified state to the vm_page.
3067 * Atomic ops are mandatory in order to ensure that PG_M events are
3068 * not lost during any transition.
3073 * pmap_remove_pv_pte() unwires pt_pv and assumes
3074 * we will free pte_pv, but since we are reusing
3075 * pte_pv we want to retain the wire count.
3077 * pt_pv won't exist for a kernel page (managed or
3081 vm_page_wire_quick(pt_pv->pv_m);
3082 if (prot & VM_PROT_NOSYNC)
3083 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3085 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3087 pmap_remove_pv_page(pte_pv);
3088 } else if (prot & VM_PROT_NOSYNC) {
3089 /* leave wire count on PT page intact */
3090 (void)pte_load_clear(ptep);
3091 cpu_invlpg((void *)va);
3092 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3094 /* leave wire count on PT page intact */
3095 pmap_inval_interlock(&info, pmap, va);
3096 (void)pte_load_clear(ptep);
3097 pmap_inval_deinterlock(&info, pmap);
3098 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3100 KKASSERT(*ptep == 0);
3105 * Enter on the PV list if part of our managed memory.
3106 * Wiring of the PT page is already handled.
3108 KKASSERT(pte_pv->pv_m == NULL);
3109 vm_page_spin_lock(m);
3111 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3114 atomic_add_int(&m->object->agg_pv_list_count, 1);
3116 vm_page_flag_set(m, PG_MAPPED);
3117 vm_page_spin_unlock(m);
3118 } else if (pt_pv && opa == 0) {
3120 * We have to adjust the wire count on the PT page ourselves
3121 * for unmanaged entries. If opa was non-zero we retained
3122 * the existing wire count from the removal.
3124 vm_page_wire_quick(pt_pv->pv_m);
3128 * Ok, for UVM (pt_pv != NULL) we don't need to interlock or
3129 * invalidate anything, the TLB won't have any stale entries to
3132 * For KVM there appear to still be issues. Theoretically we
3133 * should be able to scrap the interlocks entirely but we
3136 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3137 pmap_inval_interlock(&info, pmap, va);
3138 *(volatile pt_entry_t *)ptep = newpte;
3140 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3141 pmap_inval_deinterlock(&info, pmap);
3142 else if (pt_pv == NULL)
3143 cpu_invlpg((void *)va);
3146 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3148 vm_page_flag_set(m, PG_WRITEABLE);
3150 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3155 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
3156 pmap_inval_done(&info);
3158 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
3161 * Cleanup the pv entry, allowing other accessors.
3170 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
3171 * This code also assumes that the pmap has no pre-existing entry for this
3174 * This code currently may only be used on user pmaps, not kernel_pmap.
3177 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
3179 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE);
3183 * Make a temporary mapping for a physical address. This is only intended
3184 * to be used for panic dumps.
3186 * The caller is responsible for calling smp_invltlb().
3189 pmap_kenter_temporary(vm_paddr_t pa, long i)
3191 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
3192 return ((void *)crashdumpmap);
3195 #define MAX_INIT_PT (96)
3198 * This routine preloads the ptes for a given object into the specified pmap.
3199 * This eliminates the blast of soft faults on process startup and
3200 * immediately after an mmap.
3202 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
3205 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
3206 vm_object_t object, vm_pindex_t pindex,
3207 vm_size_t size, int limit)
3209 struct rb_vm_page_scan_info info;
3214 * We can't preinit if read access isn't set or there is no pmap
3217 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
3221 * We can't preinit if the pmap is not the current pmap
3223 lp = curthread->td_lwp;
3224 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
3227 psize = x86_64_btop(size);
3229 if ((object->type != OBJT_VNODE) ||
3230 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
3231 (object->resident_page_count > MAX_INIT_PT))) {
3235 if (pindex + psize > object->size) {
3236 if (object->size < pindex)
3238 psize = object->size - pindex;
3245 * Use a red-black scan to traverse the requested range and load
3246 * any valid pages found into the pmap.
3248 * We cannot safely scan the object's memq without holding the
3251 info.start_pindex = pindex;
3252 info.end_pindex = pindex + psize - 1;
3258 vm_object_hold_shared(object);
3259 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3260 pmap_object_init_pt_callback, &info);
3261 vm_object_drop(object);
3266 pmap_object_init_pt_callback(vm_page_t p, void *data)
3268 struct rb_vm_page_scan_info *info = data;
3269 vm_pindex_t rel_index;
3272 * don't allow an madvise to blow away our really
3273 * free pages allocating pv entries.
3275 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3276 vmstats.v_free_count < vmstats.v_free_reserved) {
3281 * Ignore list markers and ignore pages we cannot instantly
3282 * busy (while holding the object token).
3284 if (p->flags & PG_MARKER)
3286 if (vm_page_busy_try(p, TRUE))
3288 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3289 (p->flags & PG_FICTITIOUS) == 0) {
3290 if ((p->queue - p->pc) == PQ_CACHE)
3291 vm_page_deactivate(p);
3292 rel_index = p->pindex - info->start_pindex;
3293 pmap_enter_quick(info->pmap,
3294 info->addr + x86_64_ptob(rel_index), p);
3302 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
3305 * Returns FALSE if it would be non-trivial or if a pte is already loaded
3308 * XXX This is safe only because page table pages are not freed.
3311 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3315 /*spin_lock(&pmap->pm_spin);*/
3316 if ((pte = pmap_pte(pmap, addr)) != NULL) {
3318 /*spin_unlock(&pmap->pm_spin);*/
3322 /*spin_unlock(&pmap->pm_spin);*/
3327 * Change the wiring attribute for a pmap/va pair. The mapping must already
3328 * exist in the pmap. The mapping may or may not be managed.
3331 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
3338 lwkt_gettoken(&pmap->pm_token);
3339 pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
3340 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
3342 if (wired && !pmap_pte_w(ptep))
3343 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3344 else if (!wired && pmap_pte_w(ptep))
3345 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3348 * Wiring is not a hardware characteristic so there is no need to
3349 * invalidate TLB. However, in an SMP environment we must use
3350 * a locked bus cycle to update the pte (if we are not using
3351 * the pmap_inval_*() API that is)... it's ok to do this for simple
3356 atomic_set_long(ptep, PG_W);
3358 atomic_clear_long(ptep, PG_W);
3361 atomic_set_long_nonlocked(ptep, PG_W);
3363 atomic_clear_long_nonlocked(ptep, PG_W);
3366 lwkt_reltoken(&pmap->pm_token);
3372 * Copy the range specified by src_addr/len from the source map to
3373 * the range dst_addr/len in the destination map.
3375 * This routine is only advisory and need not do anything.
3378 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3379 vm_size_t len, vm_offset_t src_addr)
3386 * Zero the specified physical page.
3388 * This function may be called from an interrupt and no locking is
3392 pmap_zero_page(vm_paddr_t phys)
3394 vm_offset_t va = PHYS_TO_DMAP(phys);
3396 pagezero((void *)va);
3400 * pmap_page_assertzero:
3402 * Assert that a page is empty, panic if it isn't.
3405 pmap_page_assertzero(vm_paddr_t phys)
3407 vm_offset_t va = PHYS_TO_DMAP(phys);
3410 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3411 if (*(long *)((char *)va + i) != 0) {
3412 panic("pmap_page_assertzero() @ %p not zero!\n",
3413 (void *)(intptr_t)va);
3421 * Zero part of a physical page by mapping it into memory and clearing
3422 * its contents with bzero.
3424 * off and size may not cover an area beyond a single hardware page.
3427 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3429 vm_offset_t virt = PHYS_TO_DMAP(phys);
3431 bzero((char *)virt + off, size);
3437 * Copy the physical page from the source PA to the target PA.
3438 * This function may be called from an interrupt. No locking
3442 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3444 vm_offset_t src_virt, dst_virt;
3446 src_virt = PHYS_TO_DMAP(src);
3447 dst_virt = PHYS_TO_DMAP(dst);
3448 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3452 * pmap_copy_page_frag:
3454 * Copy the physical page from the source PA to the target PA.
3455 * This function may be called from an interrupt. No locking
3459 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3461 vm_offset_t src_virt, dst_virt;
3463 src_virt = PHYS_TO_DMAP(src);
3464 dst_virt = PHYS_TO_DMAP(dst);
3466 bcopy((char *)src_virt + (src & PAGE_MASK),
3467 (char *)dst_virt + (dst & PAGE_MASK),
3472 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
3473 * this page. This count may be changed upwards or downwards in the future;
3474 * it is only necessary that true be returned for a small subset of pmaps
3475 * for proper page aging.
3478 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3483 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3486 vm_page_spin_lock(m);
3487 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3488 if (pv->pv_pmap == pmap) {
3489 vm_page_spin_unlock(m);
3496 vm_page_spin_unlock(m);
3501 * Remove all pages from specified address space this aids process exit
3502 * speeds. Also, this code may be special cased for the current process
3506 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3508 pmap_remove(pmap, sva, eva);
3512 * pmap_testbit tests bits in pte's note that the testbit/clearbit
3513 * routines are inline, and a lot of things compile-time evaluate.
3517 pmap_testbit(vm_page_t m, int bit)
3522 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3525 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3527 vm_page_spin_lock(m);
3528 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
3529 vm_page_spin_unlock(m);
3533 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3535 * if the bit being tested is the modified bit, then
3536 * mark clean_map and ptes as never
3539 if (bit & (PG_A|PG_M)) {
3540 if (!pmap_track_modified(pv->pv_pindex))
3544 #if defined(PMAP_DIAGNOSTIC)
3545 if (pv->pv_pmap == NULL) {
3546 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
3551 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3553 vm_page_spin_unlock(m);
3557 vm_page_spin_unlock(m);
3562 * This routine is used to modify bits in ptes. Only one bit should be
3563 * specified. PG_RW requires special handling.
3565 * Caller must NOT hold any spin locks
3569 pmap_clearbit(vm_page_t m, int bit)
3571 struct pmap_inval_info info;
3578 vm_page_flag_clear(m, PG_WRITEABLE);
3579 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
3586 * Loop over all current mappings setting/clearing as appropos If
3587 * setting RO do we need to clear the VAC?
3589 * NOTE: When clearing PG_M we could also (not implemented) drop
3590 * through to the PG_RW code and clear PG_RW too, forcing
3591 * a fault on write to redetect PG_M for virtual kernels, but
3592 * it isn't necessary since virtual kernels invalidate the
3593 * pte when they clear the VPTE_M bit in their virtual page
3596 * NOTE: Does not re-dirty the page when clearing only PG_M.
3598 if ((bit & PG_RW) == 0) {
3599 vm_page_spin_lock(m);
3600 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3601 #if defined(PMAP_DIAGNOSTIC)
3602 if (pv->pv_pmap == NULL) {
3603 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
3608 pte = pmap_pte_quick(pv->pv_pmap,
3609 pv->pv_pindex << PAGE_SHIFT);
3612 atomic_clear_long(pte, bit);
3614 vm_page_spin_unlock(m);
3619 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
3622 pmap_inval_init(&info);
3625 vm_page_spin_lock(m);
3626 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3628 * don't write protect pager mappings
3630 if (!pmap_track_modified(pv->pv_pindex))
3633 #if defined(PMAP_DIAGNOSTIC)
3634 if (pv->pv_pmap == NULL) {
3635 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
3641 * Skip pages which do not have PG_RW set.
3643 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3644 if ((*pte & PG_RW) == 0)
3650 if (pv_hold_try(pv) == 0) {
3651 vm_page_spin_unlock(m);
3652 pv_lock(pv); /* held, now do a blocking lock */
3653 pv_put(pv); /* and release */
3654 goto restart; /* anything could have happened */
3657 save_pmap = pv->pv_pmap;
3658 vm_page_spin_unlock(m);
3659 pmap_inval_interlock(&info, save_pmap,
3660 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
3661 KKASSERT(pv->pv_pmap == save_pmap);
3665 if (atomic_cmpset_long(pte, pbits,
3666 pbits & ~(PG_RW|PG_M))) {
3670 pmap_inval_deinterlock(&info, save_pmap);
3671 vm_page_spin_lock(m);
3674 * If PG_M was found to be set while we were clearing PG_RW
3675 * we also clear PG_M (done above) and mark the page dirty.
3676 * Callers expect this behavior.
3682 vm_page_spin_unlock(m);
3683 pmap_inval_done(&info);
3687 * Lower the permission for all mappings to a given page.
3689 * Page must be busied by caller.
3692 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3694 /* JG NX support? */
3695 if ((prot & VM_PROT_WRITE) == 0) {
3696 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3698 * NOTE: pmap_clearbit(.. PG_RW) also clears
3699 * the PG_WRITEABLE flag in (m).
3701 pmap_clearbit(m, PG_RW);
3709 pmap_phys_address(vm_pindex_t ppn)
3711 return (x86_64_ptob(ppn));
3715 * Return a count of reference bits for a page, clearing those bits.
3716 * It is not necessary for every reference bit to be cleared, but it
3717 * is necessary that 0 only be returned when there are truly no
3718 * reference bits set.
3720 * XXX: The exact number of bits to check and clear is a matter that
3721 * should be tested and standardized at some point in the future for
3722 * optimal aging of shared pages.
3724 * This routine may not block.
3727 pmap_ts_referenced(vm_page_t m)
3733 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3736 vm_page_spin_lock(m);
3737 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3738 if (!pmap_track_modified(pv->pv_pindex))
3740 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3741 if (pte && (*pte & PG_A)) {
3743 atomic_clear_long(pte, PG_A);
3745 atomic_clear_long_nonlocked(pte, PG_A);
3752 vm_page_spin_unlock(m);
3759 * Return whether or not the specified physical page was modified
3760 * in any physical maps.
3763 pmap_is_modified(vm_page_t m)
3767 res = pmap_testbit(m, PG_M);
3772 * Clear the modify bits on the specified physical page.
3775 pmap_clear_modify(vm_page_t m)
3777 pmap_clearbit(m, PG_M);
3781 * pmap_clear_reference:
3783 * Clear the reference bit on the specified physical page.
3786 pmap_clear_reference(vm_page_t m)
3788 pmap_clearbit(m, PG_A);
3792 * Miscellaneous support routines follow
3797 i386_protection_init(void)
3801 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3802 kp = protection_codes;
3803 for (prot = 0; prot < 8; prot++) {
3805 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3807 * Read access is also 0. There isn't any execute bit,
3808 * so just make it readable.
3810 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3811 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3812 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3815 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3816 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3817 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3818 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3826 * Map a set of physical memory pages into the kernel virtual
3827 * address space. Return a pointer to where it is mapped. This
3828 * routine is intended to be used for mapping device memory,
3831 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3835 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3837 vm_offset_t va, tmpva, offset;
3840 offset = pa & PAGE_MASK;
3841 size = roundup(offset + size, PAGE_SIZE);
3843 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3845 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3847 pa = pa & ~PAGE_MASK;
3848 for (tmpva = va; size > 0;) {
3849 pte = vtopte(tmpva);
3850 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3858 return ((void *)(va + offset));
3862 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3864 vm_offset_t va, tmpva, offset;
3867 offset = pa & PAGE_MASK;
3868 size = roundup(offset + size, PAGE_SIZE);
3870 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3872 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3874 pa = pa & ~PAGE_MASK;
3875 for (tmpva = va; size > 0;) {
3876 pte = vtopte(tmpva);
3877 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3885 return ((void *)(va + offset));
3889 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3891 vm_offset_t base, offset;
3893 base = va & ~PAGE_MASK;
3894 offset = va & PAGE_MASK;
3895 size = roundup(offset + size, PAGE_SIZE);
3896 pmap_qremove(va, size >> PAGE_SHIFT);
3897 kmem_free(&kernel_map, base, size);
3901 * perform the pmap work for mincore
3904 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3906 pt_entry_t *ptep, pte;
3910 lwkt_gettoken(&pmap->pm_token);
3911 ptep = pmap_pte(pmap, addr);
3913 if (ptep && (pte = *ptep) != 0) {
3916 val = MINCORE_INCORE;
3917 if ((pte & PG_MANAGED) == 0)
3920 pa = pte & PG_FRAME;
3922 m = PHYS_TO_VM_PAGE(pa);
3928 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3930 * Modified by someone
3932 else if (m->dirty || pmap_is_modified(m))
3933 val |= MINCORE_MODIFIED_OTHER;
3938 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3941 * Referenced by someone
3943 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3944 val |= MINCORE_REFERENCED_OTHER;
3945 vm_page_flag_set(m, PG_REFERENCED);
3949 lwkt_reltoken(&pmap->pm_token);
3955 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3956 * vmspace will be ref'd and the old one will be deref'd.
3958 * The vmspace for all lwps associated with the process will be adjusted
3959 * and cr3 will be reloaded if any lwp is the current lwp.
3961 * The process must hold the vmspace->vm_map.token for oldvm and newvm
3964 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3966 struct vmspace *oldvm;
3969 oldvm = p->p_vmspace;
3970 if (oldvm != newvm) {
3972 sysref_get(&newvm->vm_sysref);
3973 p->p_vmspace = newvm;
3974 KKASSERT(p->p_nthreads == 1);
3975 lp = RB_ROOT(&p->p_lwp_tree);
3976 pmap_setlwpvm(lp, newvm);
3978 sysref_put(&oldvm->vm_sysref);
3983 * Set the vmspace for a LWP. The vmspace is almost universally set the
3984 * same as the process vmspace, but virtual kernels need to swap out contexts
3985 * on a per-lwp basis.
3987 * Caller does not necessarily hold any vmspace tokens. Caller must control
3988 * the lwp (typically be in the context of the lwp). We use a critical
3989 * section to protect against statclock and hardclock (statistics collection).
3992 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3994 struct vmspace *oldvm;
3997 oldvm = lp->lwp_vmspace;
3999 if (oldvm != newvm) {
4001 lp->lwp_vmspace = newvm;
4002 if (curthread->td_lwp == lp) {
4003 pmap = vmspace_pmap(newvm);
4005 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4006 if (pmap->pm_active & CPUMASK_LOCK)
4007 pmap_interlock_wait(newvm);
4009 pmap->pm_active |= 1;
4011 #if defined(SWTCH_OPTIM_STATS)
4014 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4015 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
4016 load_cr3(curthread->td_pcb->pcb_cr3);
4017 pmap = vmspace_pmap(oldvm);
4019 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4021 pmap->pm_active &= ~(cpumask_t)1;
4031 * Called when switching to a locked pmap, used to interlock against pmaps
4032 * undergoing modifications to prevent us from activating the MMU for the
4033 * target pmap until all such modifications have completed. We have to do
4034 * this because the thread making the modifications has already set up its
4035 * SMP synchronization mask.
4037 * This function cannot sleep!
4042 pmap_interlock_wait(struct vmspace *vm)
4044 struct pmap *pmap = &vm->vm_pmap;
4046 if (pmap->pm_active & CPUMASK_LOCK) {
4048 DEBUG_PUSH_INFO("pmap_interlock_wait");
4049 while (pmap->pm_active & CPUMASK_LOCK) {
4051 lwkt_process_ipiq();
4061 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4064 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4068 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4073 * Used by kmalloc/kfree, page already exists at va
4076 pmap_kvtom(vm_offset_t va)
4078 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));