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, 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 */
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 ***************************************************/
968 #if defined(PMAP_DIAGNOSTIC)
971 * This code checks for non-writeable/modified pages.
972 * This should be an invalid condition.
976 pmap_nw_modified(pt_entry_t pte)
978 if ((pte & (PG_M|PG_RW)) == PG_M)
987 * this routine defines the region(s) of memory that should
988 * not be tested for the modified bit.
992 pmap_track_modified(vm_pindex_t pindex)
994 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
995 if ((va < clean_sva) || (va >= clean_eva))
1002 * Extract the physical page address associated with the map/VA pair.
1003 * The page must be wired for this to work reliably.
1005 * XXX for the moment we're using pv_find() instead of pv_get(), as
1006 * callers might be expecting non-blocking operation.
1009 pmap_extract(pmap_t pmap, vm_offset_t va)
1016 if (va >= VM_MAX_USER_ADDRESS) {
1018 * Kernel page directories might be direct-mapped and
1019 * there is typically no PV tracking of pte's
1023 pt = pmap_pt(pmap, va);
1024 if (pt && (*pt & PG_V)) {
1026 rtval = *pt & PG_PS_FRAME;
1027 rtval |= va & PDRMASK;
1029 ptep = pmap_pt_to_pte(pt, va);
1031 rtval = *ptep & PG_FRAME;
1032 rtval |= va & PAGE_MASK;
1038 * User pages currently do not direct-map the page directory
1039 * and some pages might not used managed PVs. But all PT's
1042 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1044 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1046 rtval = *ptep & PG_FRAME;
1047 rtval |= va & PAGE_MASK;
1056 * Extract the physical page address associated kernel virtual address.
1059 pmap_kextract(vm_offset_t va)
1061 pd_entry_t pt; /* pt entry in pd */
1064 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1065 pa = DMAP_TO_PHYS(va);
1069 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1072 * Beware of a concurrent promotion that changes the
1073 * PDE at this point! For example, vtopte() must not
1074 * be used to access the PTE because it would use the
1075 * new PDE. It is, however, safe to use the old PDE
1076 * because the page table page is preserved by the
1079 pa = *pmap_pt_to_pte(&pt, va);
1080 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1086 /***************************************************
1087 * Low level mapping routines.....
1088 ***************************************************/
1091 * Routine: pmap_kenter
1093 * Add a wired page to the KVA
1094 * NOTE! note that in order for the mapping to take effect -- you
1095 * should do an invltlb after doing the pmap_kenter().
1098 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1102 pmap_inval_info info;
1104 pmap_inval_init(&info); /* XXX remove */
1105 npte = pa | PG_RW | PG_V | pgeflag;
1107 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1109 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1110 pmap_inval_done(&info); /* XXX remove */
1114 * Routine: pmap_kenter_quick
1116 * Similar to pmap_kenter(), except we only invalidate the
1117 * mapping on the current CPU.
1120 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1125 npte = pa | PG_RW | PG_V | pgeflag;
1128 cpu_invlpg((void *)va);
1132 pmap_kenter_sync(vm_offset_t va)
1134 pmap_inval_info info;
1136 pmap_inval_init(&info);
1137 pmap_inval_interlock(&info, &kernel_pmap, va);
1138 pmap_inval_deinterlock(&info, &kernel_pmap);
1139 pmap_inval_done(&info);
1143 pmap_kenter_sync_quick(vm_offset_t va)
1145 cpu_invlpg((void *)va);
1149 * remove a page from the kernel pagetables
1152 pmap_kremove(vm_offset_t va)
1155 pmap_inval_info info;
1157 pmap_inval_init(&info);
1159 pmap_inval_interlock(&info, &kernel_pmap, va);
1161 pmap_inval_deinterlock(&info, &kernel_pmap);
1162 pmap_inval_done(&info);
1166 pmap_kremove_quick(vm_offset_t va)
1171 cpu_invlpg((void *)va);
1175 * XXX these need to be recoded. They are not used in any critical path.
1178 pmap_kmodify_rw(vm_offset_t va)
1180 atomic_set_long(vtopte(va), PG_RW);
1181 cpu_invlpg((void *)va);
1185 pmap_kmodify_nc(vm_offset_t va)
1187 atomic_set_long(vtopte(va), PG_N);
1188 cpu_invlpg((void *)va);
1192 * Used to map a range of physical addresses into kernel virtual
1193 * address space during the low level boot, typically to map the
1194 * dump bitmap, message buffer, and vm_page_array.
1196 * These mappings are typically made at some pointer after the end of the
1199 * We could return PHYS_TO_DMAP(start) here and not allocate any
1200 * via (*virtp), but then kmem from userland and kernel dumps won't
1201 * have access to the related pointers.
1204 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1207 vm_offset_t va_start;
1209 /*return PHYS_TO_DMAP(start);*/
1214 while (start < end) {
1215 pmap_kenter_quick(va, start);
1225 * Add a list of wired pages to the kva
1226 * this routine is only used for temporary
1227 * kernel mappings that do not need to have
1228 * page modification or references recorded.
1229 * Note that old mappings are simply written
1230 * over. The page *must* be wired.
1233 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1237 end_va = va + count * PAGE_SIZE;
1239 while (va < end_va) {
1243 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1244 cpu_invlpg((void *)va);
1252 * This routine jerks page mappings from the
1253 * kernel -- it is meant only for temporary mappings.
1255 * MPSAFE, INTERRUPT SAFE (cluster callback)
1258 pmap_qremove(vm_offset_t va, int count)
1262 end_va = va + count * PAGE_SIZE;
1264 while (va < end_va) {
1269 cpu_invlpg((void *)va);
1276 * Create a new thread and optionally associate it with a (new) process.
1277 * NOTE! the new thread's cpu may not equal the current cpu.
1280 pmap_init_thread(thread_t td)
1282 /* enforce pcb placement & alignment */
1283 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1284 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1285 td->td_savefpu = &td->td_pcb->pcb_save;
1286 td->td_sp = (char *)td->td_pcb; /* no -16 */
1290 * This routine directly affects the fork perf for a process.
1293 pmap_init_proc(struct proc *p)
1298 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1299 * it, and IdlePTD, represents the template used to update all other pmaps.
1301 * On architectures where the kernel pmap is not integrated into the user
1302 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1303 * kernel_pmap should be used to directly access the kernel_pmap.
1306 pmap_pinit0(struct pmap *pmap)
1308 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1310 pmap->pm_active = 0;
1311 pmap->pm_pvhint = NULL;
1312 RB_INIT(&pmap->pm_pvroot);
1313 spin_init(&pmap->pm_spin);
1314 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1315 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1319 * Initialize a preallocated and zeroed pmap structure,
1320 * such as one in a vmspace structure.
1323 pmap_pinit(struct pmap *pmap)
1329 * Misc initialization
1332 pmap->pm_active = 0;
1333 pmap->pm_pvhint = NULL;
1334 if (pmap->pm_pmlpv == NULL) {
1335 RB_INIT(&pmap->pm_pvroot);
1336 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1337 spin_init(&pmap->pm_spin);
1338 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1342 * No need to allocate page table space yet but we do need a valid
1343 * page directory table.
1345 if (pmap->pm_pml4 == NULL) {
1347 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1351 * Allocate the page directory page, which wires it even though
1352 * it isn't being entered into some higher level page table (it
1353 * being the highest level). If one is already cached we don't
1354 * have to do anything.
1356 if ((pv = pmap->pm_pmlpv) == NULL) {
1357 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1358 pmap->pm_pmlpv = pv;
1359 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1360 VM_PAGE_TO_PHYS(pv->pv_m));
1364 * Install DMAP and KMAP.
1366 for (j = 0; j < NDMPML4E; ++j) {
1367 pmap->pm_pml4[DMPML4I + j] =
1368 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1369 PG_RW | PG_V | PG_U;
1371 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1374 * install self-referential address mapping entry
1376 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1377 PG_V | PG_RW | PG_A | PG_M;
1379 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1380 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1385 * Clean up a pmap structure so it can be physically freed. This routine
1386 * is called by the vmspace dtor function. A great deal of pmap data is
1387 * left passively mapped to improve vmspace management so we have a bit
1388 * of cleanup work to do here.
1391 pmap_puninit(pmap_t pmap)
1396 KKASSERT(pmap->pm_active == 0);
1397 if ((pv = pmap->pm_pmlpv) != NULL) {
1398 if (pv_hold_try(pv) == 0)
1400 p = pmap_remove_pv_page(pv, 1);
1402 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1403 vm_page_busy_wait(p, FALSE, "pgpun");
1405 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1406 vm_page_unwire(p, 0);
1407 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1410 * XXX eventually clean out PML4 static entries and
1411 * use vm_page_free_zero()
1414 pmap->pm_pmlpv = NULL;
1416 if (pmap->pm_pml4) {
1417 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1418 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1419 pmap->pm_pml4 = NULL;
1421 KKASSERT(pmap->pm_stats.resident_count == 0);
1422 KKASSERT(pmap->pm_stats.wired_count == 0);
1426 * Wire in kernel global address entries. To avoid a race condition
1427 * between pmap initialization and pmap_growkernel, this procedure
1428 * adds the pmap to the master list (which growkernel scans to update),
1429 * then copies the template.
1432 pmap_pinit2(struct pmap *pmap)
1435 * XXX copies current process, does not fill in MPPTDI
1437 spin_lock(&pmap_spin);
1438 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1439 spin_unlock(&pmap_spin);
1443 * This routine is called when various levels in the page table need to
1444 * be populated. This routine cannot fail.
1446 * This function returns two locked pv_entry's, one representing the
1447 * requested pv and one representing the requested pv's parent pv. If
1448 * the pv did not previously exist it will be mapped into its parent
1449 * and wired, otherwise no additional wire count will be added.
1453 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1458 vm_pindex_t pt_pindex;
1463 * If the pv already exists and we aren't being asked for the
1464 * parent page table page we can just return it. A locked+held pv
1467 pv = pv_alloc(pmap, ptepindex, &isnew);
1468 if (isnew == 0 && pvpp == NULL)
1472 * This is a new PV, we have to resolve its parent page table and
1473 * add an additional wiring to the page if necessary.
1477 * Special case terminal PVs. These are not page table pages so
1478 * no vm_page is allocated (the caller supplied the vm_page). If
1479 * pvpp is non-NULL we are being asked to also removed the pt_pv
1482 * Note that pt_pv's are only returned for user VAs. We assert that
1483 * a pt_pv is not being requested for kernel VAs.
1485 if (ptepindex < pmap_pt_pindex(0)) {
1486 if (ptepindex >= NUPTE_USER)
1487 KKASSERT(pvpp == NULL);
1489 KKASSERT(pvpp != NULL);
1491 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1492 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1494 vm_page_wire_quick(pvp->pv_m);
1503 * Non-terminal PVs allocate a VM page to represent the page table,
1504 * so we have to resolve pvp and calculate ptepindex for the pvp
1505 * and then for the page table entry index in the pvp for
1508 if (ptepindex < pmap_pd_pindex(0)) {
1510 * pv is PT, pvp is PD
1512 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1513 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1514 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1521 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1522 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1523 } else if (ptepindex < pmap_pdp_pindex(0)) {
1525 * pv is PD, pvp is PDP
1527 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1528 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1529 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1536 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1537 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1538 } else if (ptepindex < pmap_pml4_pindex()) {
1540 * pv is PDP, pvp is the root pml4 table
1542 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1549 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1550 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1553 * pv represents the top-level PML4, there is no parent.
1561 * This code is only reached if isnew is TRUE and this is not a
1562 * terminal PV. We need to allocate a vm_page for the page table
1563 * at this level and enter it into the parent page table.
1565 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1568 m = vm_page_alloc(NULL, pv->pv_pindex,
1569 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1570 VM_ALLOC_INTERRUPT);
1575 vm_page_spin_lock(m);
1576 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1578 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1579 vm_page_spin_unlock(m);
1580 vm_page_unmanage(m); /* m must be spinunlocked */
1582 if ((m->flags & PG_ZERO) == 0) {
1583 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1587 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1590 m->valid = VM_PAGE_BITS_ALL;
1591 vm_page_flag_clear(m, PG_ZERO);
1592 vm_page_wire(m); /* wire for mapping in parent */
1595 * Wire the page into pvp, bump the wire-count for pvp's page table
1596 * page. Bump the resident_count for the pmap. There is no pvp
1597 * for the top level, address the pm_pml4[] array directly.
1599 * If the caller wants the parent we return it, otherwise
1600 * we just put it away.
1602 * No interlock is needed for pte 0 -> non-zero.
1605 vm_page_wire_quick(pvp->pv_m);
1606 ptep = pv_pte_lookup(pvp, ptepindex);
1607 KKASSERT((*ptep & PG_V) == 0);
1608 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1621 * Release any resources held by the given physical map.
1623 * Called when a pmap initialized by pmap_pinit is being released. Should
1624 * only be called if the map contains no valid mappings.
1626 * Caller must hold pmap->pm_token
1628 struct pmap_release_info {
1633 static int pmap_release_callback(pv_entry_t pv, void *data);
1636 pmap_release(struct pmap *pmap)
1638 struct pmap_release_info info;
1640 KASSERT(pmap->pm_active == 0,
1641 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1643 spin_lock(&pmap_spin);
1644 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1645 spin_unlock(&pmap_spin);
1648 * Pull pv's off the RB tree in order from low to high and release
1654 spin_lock(&pmap->pm_spin);
1655 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
1656 pmap_release_callback, &info);
1657 spin_unlock(&pmap->pm_spin);
1658 } while (info.retry);
1662 * One resident page (the pml4 page) should remain.
1663 * No wired pages should remain.
1665 KKASSERT(pmap->pm_stats.resident_count == 1);
1666 KKASSERT(pmap->pm_stats.wired_count == 0);
1670 pmap_release_callback(pv_entry_t pv, void *data)
1672 struct pmap_release_info *info = data;
1673 pmap_t pmap = info->pmap;
1676 if (pv_hold_try(pv)) {
1677 spin_unlock(&pmap->pm_spin);
1679 spin_unlock(&pmap->pm_spin);
1681 if (pv->pv_pmap != pmap) {
1683 spin_lock(&pmap->pm_spin);
1690 * The pmap is currently not spinlocked, pv is held+locked.
1691 * Remove the pv's page from its parent's page table. The
1692 * parent's page table page's wire_count will be decremented.
1694 pmap_remove_pv_pte(pv, NULL, NULL);
1697 * Terminal pvs are unhooked from their vm_pages. Because
1698 * terminal pages aren't page table pages they aren't wired
1699 * by us, so we have to be sure not to unwire them either.
1701 if (pv->pv_pindex < pmap_pt_pindex(0)) {
1702 pmap_remove_pv_page(pv, 0);
1707 * We leave the top-level page table page cached, wired, and
1708 * mapped in the pmap until the dtor function (pmap_puninit())
1711 * Since we are leaving the top-level pv intact we need
1712 * to break out of what would otherwise be an infinite loop.
1714 if (pv->pv_pindex == pmap_pml4_pindex()) {
1716 spin_lock(&pmap->pm_spin);
1721 * For page table pages (other than the top-level page),
1722 * remove and free the vm_page. The representitive mapping
1723 * removed above by pmap_remove_pv_pte() did not undo the
1724 * last wire_count so we have to do that as well.
1726 p = pmap_remove_pv_page(pv, 1);
1727 vm_page_busy_wait(p, FALSE, "pmaprl");
1729 if (p->wire_count != 1) {
1730 kprintf("p->wire_count was %016lx %d\n",
1731 pv->pv_pindex, p->wire_count);
1733 KKASSERT(p->wire_count == 1);
1734 KKASSERT(p->flags & PG_UNMANAGED);
1736 vm_page_unwire(p, 0);
1737 KKASSERT(p->wire_count == 0);
1738 /* JG eventually revert to using vm_page_free_zero() */
1742 spin_lock(&pmap->pm_spin);
1747 * This function will remove the pte associated with a pv from its parent.
1748 * Terminal pv's are supported. The removal will be interlocked if info
1749 * is non-NULL. The caller must dispose of pv instead of just unlocking
1752 * The wire count will be dropped on the parent page table. The wire
1753 * count on the page being removed (pv->pv_m) from the parent page table
1754 * is NOT touched. Note that terminal pages will not have any additional
1755 * wire counts while page table pages will have at least one representing
1756 * the mapping, plus others representing sub-mappings.
1758 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
1759 * pages and user page table and terminal pages.
1761 * The pv must be locked.
1763 * XXX must lock parent pv's if they exist to remove pte XXX
1767 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
1769 vm_pindex_t ptepindex = pv->pv_pindex;
1770 pmap_t pmap = pv->pv_pmap;
1776 if (ptepindex == pmap_pml4_pindex()) {
1778 * We are the top level pml4 table, there is no parent.
1780 p = pmap->pm_pmlpv->pv_m;
1781 } else if (ptepindex >= pmap_pdp_pindex(0)) {
1783 * Remove a PDP page from the pml4e. This can only occur
1784 * with user page tables. We do not have to lock the
1785 * pml4 PV so just ignore pvp.
1787 vm_pindex_t pml4_pindex;
1788 vm_pindex_t pdp_index;
1791 pdp_index = ptepindex - pmap_pdp_pindex(0);
1793 pml4_pindex = pmap_pml4_pindex();
1794 pvp = pv_get(pv->pv_pmap, pml4_pindex);
1797 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
1798 KKASSERT((*pdp & PG_V) != 0);
1799 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1801 KKASSERT(info == NULL);
1802 } else if (ptepindex >= pmap_pd_pindex(0)) {
1804 * Remove a PD page from the pdp
1806 vm_pindex_t pdp_pindex;
1807 vm_pindex_t pd_index;
1810 pd_index = ptepindex - pmap_pd_pindex(0);
1813 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
1814 (pd_index >> NPML4EPGSHIFT);
1815 pvp = pv_get(pv->pv_pmap, pdp_pindex);
1818 pd = pv_pte_lookup(pvp, pd_index & ((1ul << NPDPEPGSHIFT) - 1));
1819 KKASSERT((*pd & PG_V) != 0);
1820 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
1822 KKASSERT(info == NULL);
1823 } else if (ptepindex >= pmap_pt_pindex(0)) {
1825 * Remove a PT page from the pd
1827 vm_pindex_t pd_pindex;
1828 vm_pindex_t pt_index;
1831 pt_index = ptepindex - pmap_pt_pindex(0);
1834 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
1835 (pt_index >> NPDPEPGSHIFT);
1836 pvp = pv_get(pv->pv_pmap, pd_pindex);
1839 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
1840 KKASSERT((*pt & PG_V) != 0);
1841 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
1843 KKASSERT(info == NULL);
1846 * Remove a PTE from the PT page
1848 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
1849 * pv is a pte_pv so we can safely lock pt_pv.
1851 vm_pindex_t pt_pindex;
1856 pt_pindex = ptepindex >> NPTEPGSHIFT;
1857 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
1859 if (ptepindex >= NUPTE_USER) {
1860 ptep = vtopte(ptepindex << PAGE_SHIFT);
1861 KKASSERT(pvp == NULL);
1864 pt_pindex = NUPTE_TOTAL +
1865 (ptepindex >> NPDPEPGSHIFT);
1866 pvp = pv_get(pv->pv_pmap, pt_pindex);
1869 ptep = pv_pte_lookup(pvp, ptepindex &
1870 ((1ul << NPDPEPGSHIFT) - 1));
1874 pmap_inval_interlock(info, pmap, va);
1875 pte = pte_load_clear(ptep);
1877 pmap_inval_deinterlock(info, pmap);
1880 * Now update the vm_page_t
1882 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
1883 kprintf("remove_pte badpte %016lx %016lx %d\n",
1885 pv->pv_pindex < pmap_pt_pindex(0));
1887 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
1888 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
1891 if (pmap_track_modified(ptepindex))
1895 vm_page_flag_set(p, PG_REFERENCED);
1898 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1900 cpu_invlpg((void *)va);
1904 * Unwire the parent page table page. The wire_count cannot go below
1905 * 1 here because the parent page table page is itself still mapped.
1907 * XXX remove the assertions later.
1909 KKASSERT(pv->pv_m == p);
1910 if (pvp && vm_page_unwire_quick(pvp->pv_m))
1911 panic("pmap_remove_pv_pte: Insufficient wire_count");
1919 pmap_remove_pv_page(pv_entry_t pv, int holdpg)
1927 vm_page_spin_lock(m);
1929 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1932 atomic_add_int(&m->object->agg_pv_list_count, -1);
1934 if (TAILQ_EMPTY(&m->md.pv_list))
1935 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1936 vm_page_spin_unlock(m);
1943 * Grow the number of kernel page table entries, if needed.
1945 * This routine is always called to validate any address space
1946 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1947 * space below KERNBASE.
1950 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1953 vm_offset_t ptppaddr;
1955 pd_entry_t *pt, newpt;
1957 int update_kernel_vm_end;
1960 * bootstrap kernel_vm_end on first real VM use
1962 if (kernel_vm_end == 0) {
1963 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1965 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1966 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1967 ~(PAGE_SIZE * NPTEPG - 1);
1969 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1970 kernel_vm_end = kernel_map.max_offset;
1977 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1978 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1979 * do not want to force-fill 128G worth of page tables.
1981 if (kstart < KERNBASE) {
1982 if (kstart > kernel_vm_end)
1983 kstart = kernel_vm_end;
1984 KKASSERT(kend <= KERNBASE);
1985 update_kernel_vm_end = 1;
1987 update_kernel_vm_end = 0;
1990 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1991 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1993 if (kend - 1 >= kernel_map.max_offset)
1994 kend = kernel_map.max_offset;
1996 while (kstart < kend) {
1997 pt = pmap_pt(&kernel_pmap, kstart);
1999 /* We need a new PDP entry */
2000 nkpg = vm_page_alloc(NULL, nkpt,
2003 VM_ALLOC_INTERRUPT);
2005 panic("pmap_growkernel: no memory to grow "
2008 paddr = VM_PAGE_TO_PHYS(nkpg);
2009 if ((nkpg->flags & PG_ZERO) == 0)
2010 pmap_zero_page(paddr);
2011 vm_page_flag_clear(nkpg, PG_ZERO);
2012 newpd = (pdp_entry_t)
2013 (paddr | PG_V | PG_RW | PG_A | PG_M);
2014 *pmap_pd(&kernel_pmap, kstart) = newpd;
2016 continue; /* try again */
2018 if ((*pt & PG_V) != 0) {
2019 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2020 ~(PAGE_SIZE * NPTEPG - 1);
2021 if (kstart - 1 >= kernel_map.max_offset) {
2022 kstart = kernel_map.max_offset;
2029 * This index is bogus, but out of the way
2031 nkpg = vm_page_alloc(NULL, nkpt,
2034 VM_ALLOC_INTERRUPT);
2036 panic("pmap_growkernel: no memory to grow kernel");
2039 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2040 pmap_zero_page(ptppaddr);
2041 vm_page_flag_clear(nkpg, PG_ZERO);
2042 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2043 *pmap_pt(&kernel_pmap, kstart) = newpt;
2046 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2047 ~(PAGE_SIZE * NPTEPG - 1);
2049 if (kstart - 1 >= kernel_map.max_offset) {
2050 kstart = kernel_map.max_offset;
2056 * Only update kernel_vm_end for areas below KERNBASE.
2058 if (update_kernel_vm_end && kernel_vm_end < kstart)
2059 kernel_vm_end = kstart;
2063 * Retire the given physical map from service.
2064 * Should only be called if the map contains
2065 * no valid mappings.
2068 pmap_destroy(pmap_t pmap)
2075 lwkt_gettoken(&pmap->pm_token);
2076 count = --pmap->pm_count;
2078 pmap_release(pmap); /* eats pm_token */
2079 panic("destroying a pmap is not yet implemented");
2081 lwkt_reltoken(&pmap->pm_token);
2085 * Add a reference to the specified pmap.
2088 pmap_reference(pmap_t pmap)
2091 lwkt_gettoken(&pmap->pm_token);
2093 lwkt_reltoken(&pmap->pm_token);
2097 /***************************************************
2098 * page management routines.
2099 ***************************************************/
2102 * Hold a pv without locking it
2105 pv_hold(pv_entry_t pv)
2109 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2113 count = pv->pv_hold;
2115 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2122 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2123 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2126 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2127 * pv list via its page) must be held by the caller.
2130 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2134 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2137 pv->pv_line = lineno;
2143 count = pv->pv_hold;
2145 if ((count & PV_HOLD_LOCKED) == 0) {
2146 if (atomic_cmpset_int(&pv->pv_hold, count,
2147 (count + 1) | PV_HOLD_LOCKED)) {
2150 pv->pv_line = lineno;
2155 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2163 * Drop a previously held pv_entry which could not be locked, allowing its
2166 * Must not be called with a spinlock held as we might zfree() the pv if it
2167 * is no longer associated with a pmap and this was the last hold count.
2170 pv_drop(pv_entry_t pv)
2174 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2175 if (pv->pv_pmap == NULL)
2181 count = pv->pv_hold;
2183 KKASSERT((count & PV_HOLD_MASK) > 0);
2184 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2185 (PV_HOLD_LOCKED | 1));
2186 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2187 if (count == 1 && pv->pv_pmap == NULL)
2196 * Find or allocate the requested PV entry, returning a locked pv
2200 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2203 pv_entry_t pnew = NULL;
2205 spin_lock(&pmap->pm_spin);
2207 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2208 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2213 spin_unlock(&pmap->pm_spin);
2214 pnew = zalloc(pvzone);
2215 spin_lock(&pmap->pm_spin);
2218 pnew->pv_pmap = pmap;
2219 pnew->pv_pindex = pindex;
2220 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2222 pnew->pv_func = func;
2223 pnew->pv_line = lineno;
2225 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2226 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2227 spin_unlock(&pmap->pm_spin);
2232 spin_unlock(&pmap->pm_spin);
2233 zfree(pvzone, pnew);
2235 spin_lock(&pmap->pm_spin);
2238 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2239 spin_unlock(&pmap->pm_spin);
2243 spin_unlock(&pmap->pm_spin);
2244 _pv_lock(pv PMAP_DEBUG_COPY);
2245 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2250 spin_lock(&pmap->pm_spin);
2257 * Find the requested PV entry, returning a locked+held pv or NULL
2261 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2265 spin_lock(&pmap->pm_spin);
2270 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2271 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2275 spin_unlock(&pmap->pm_spin);
2278 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2279 pv_cache(pv, pindex);
2280 spin_unlock(&pmap->pm_spin);
2283 spin_unlock(&pmap->pm_spin);
2284 _pv_lock(pv PMAP_DEBUG_COPY);
2285 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2288 spin_lock(&pmap->pm_spin);
2293 * Lookup, hold, and attempt to lock (pmap,pindex).
2295 * If the entry does not exist NULL is returned and *errorp is set to 0
2297 * If the entry exists and could be successfully locked it is returned and
2298 * errorp is set to 0.
2300 * If the entry exists but could NOT be successfully locked it is returned
2301 * held and *errorp is set to 1.
2305 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2309 spin_lock(&pmap->pm_spin);
2310 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2311 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2313 spin_unlock(&pmap->pm_spin);
2317 if (pv_hold_try(pv)) {
2318 pv_cache(pv, pindex);
2319 spin_unlock(&pmap->pm_spin);
2321 return(pv); /* lock succeeded */
2323 spin_unlock(&pmap->pm_spin);
2325 return (pv); /* lock failed */
2329 * Find the requested PV entry, returning a held pv or NULL
2333 pv_find(pmap_t pmap, vm_pindex_t pindex)
2337 spin_lock(&pmap->pm_spin);
2339 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2340 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2342 spin_unlock(&pmap->pm_spin);
2346 pv_cache(pv, pindex);
2347 spin_unlock(&pmap->pm_spin);
2352 * Lock a held pv, keeping the hold count
2356 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2361 count = pv->pv_hold;
2363 if ((count & PV_HOLD_LOCKED) == 0) {
2364 if (atomic_cmpset_int(&pv->pv_hold, count,
2365 count | PV_HOLD_LOCKED)) {
2368 pv->pv_line = lineno;
2374 tsleep_interlock(pv, 0);
2375 if (atomic_cmpset_int(&pv->pv_hold, count,
2376 count | PV_HOLD_WAITING)) {
2378 kprintf("pv waiting on %s:%d\n",
2379 pv->pv_func, pv->pv_line);
2381 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2388 * Unlock a held and locked pv, keeping the hold count.
2392 pv_unlock(pv_entry_t pv)
2396 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2400 count = pv->pv_hold;
2402 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2403 (PV_HOLD_LOCKED | 1));
2404 if (atomic_cmpset_int(&pv->pv_hold, count,
2406 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2407 if (count & PV_HOLD_WAITING)
2415 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2416 * and the hold count drops to zero we will free it.
2418 * Caller should not hold any spin locks. We are protected from hold races
2419 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2420 * lock held. A pv cannot be located otherwise.
2424 pv_put(pv_entry_t pv)
2426 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2427 if (pv->pv_pmap == NULL)
2436 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2437 * pmap. Any pte operations must have already been completed.
2441 pv_free(pv_entry_t pv)
2445 KKASSERT(pv->pv_m == NULL);
2446 if ((pmap = pv->pv_pmap) != NULL) {
2447 spin_lock(&pmap->pm_spin);
2448 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2449 if (pmap->pm_pvhint == pv)
2450 pmap->pm_pvhint = NULL;
2451 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2454 spin_unlock(&pmap->pm_spin);
2460 * This routine is very drastic, but can save the system
2468 static int warningdone=0;
2470 if (pmap_pagedaemon_waken == 0)
2472 pmap_pagedaemon_waken = 0;
2473 if (warningdone < 5) {
2474 kprintf("pmap_collect: collecting pv entries -- "
2475 "suggest increasing PMAP_SHPGPERPROC\n");
2479 for (i = 0; i < vm_page_array_size; i++) {
2480 m = &vm_page_array[i];
2481 if (m->wire_count || m->hold_count)
2483 if (vm_page_busy_try(m, TRUE) == 0) {
2484 if (m->wire_count == 0 && m->hold_count == 0) {
2493 * Scan the pmap for active page table entries and issue a callback.
2494 * The callback must dispose of pte_pv.
2496 * NOTE: Unmanaged page table entries will not have a pte_pv
2498 * NOTE: Kernel page table entries will not have a pt_pv. That is, wiring
2499 * counts are not tracked in kernel page table pages.
2501 * It is assumed that the start and end are properly rounded to the page size.
2504 pmap_scan(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva,
2505 void (*func)(pmap_t, struct pmap_inval_info *,
2506 pv_entry_t, pv_entry_t, vm_offset_t,
2507 pt_entry_t *, void *),
2510 pv_entry_t pdp_pv; /* A page directory page PV */
2511 pv_entry_t pd_pv; /* A page directory PV */
2512 pv_entry_t pt_pv; /* A page table PV */
2513 pv_entry_t pte_pv; /* A page table entry PV */
2515 vm_offset_t va_next;
2516 struct pmap_inval_info info;
2523 * Hold the token for stability; if the pmap is empty we have nothing
2526 lwkt_gettoken(&pmap->pm_token);
2528 if (pmap->pm_stats.resident_count == 0) {
2529 lwkt_reltoken(&pmap->pm_token);
2534 pmap_inval_init(&info);
2537 * Special handling for removing one page, which is a very common
2538 * operation (it is?).
2539 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
2541 if (sva + PAGE_SIZE == eva) {
2542 if (sva >= VM_MAX_USER_ADDRESS) {
2544 * Kernel mappings do not track wire counts on
2548 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2552 * User mappings may or may not have a pte_pv but
2553 * will always have a pt_pv if the page is present.
2555 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2556 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2557 if (pt_pv == NULL) {
2558 KKASSERT(pte_pv == NULL);
2561 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
2565 * Unlike the pv_find() case below we actually
2566 * acquired a locked pv in this case so any
2567 * race should have been resolved. It is expected
2570 KKASSERT(pte_pv == NULL);
2571 } else if (pte_pv) {
2572 KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED|
2574 ("bad *ptep %016lx sva %016lx pte_pv %p",
2575 *ptep, sva, pte_pv));
2576 func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg);
2578 KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V,
2579 ("bad *ptep %016lx sva %016lx pte_pv NULL",
2581 func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg);
2586 pmap_inval_done(&info);
2587 lwkt_reltoken(&pmap->pm_token);
2592 * NOTE: kernel mappings do not track page table pages, only
2595 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
2596 * However, for the scan to be efficient we try to
2597 * cache items top-down.
2603 for (; sva < eva; sva = va_next) {
2605 if (sva >= VM_MAX_USER_ADDRESS) {
2616 if (pdp_pv == NULL) {
2617 pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva));
2618 } else if (pdp_pv->pv_pindex != pmap_pdp_pindex(sva)) {
2620 pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva));
2622 if (pdp_pv == NULL) {
2623 va_next = (sva + NBPML4) & ~PML4MASK;
2632 if (pd_pv == NULL) {
2637 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
2638 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
2644 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
2646 if (pd_pv == NULL) {
2647 va_next = (sva + NBPDP) & ~PDPMASK;
2656 if (pt_pv == NULL) {
2665 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2666 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
2676 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2680 * We will scan or skip a page table page so adjust va_next
2683 if (pt_pv == NULL) {
2684 va_next = (sva + NBPDR) & ~PDRMASK;
2691 * From this point in the loop testing pt_pv for non-NULL
2692 * means we are in UVM, else if it is NULL we are in KVM.
2695 va_next = (sva + NBPDR) & ~PDRMASK;
2700 * Limit our scan to either the end of the va represented
2701 * by the current page table page, or to the end of the
2702 * range being removed.
2704 * Scan the page table for pages. Some pages may not be
2705 * managed (might not have a pv_entry).
2707 * There is no page table management for kernel pages so
2708 * pt_pv will be NULL in that case, but otherwise pt_pv
2709 * is non-NULL, locked, and referenced.
2715 * At this point a non-NULL pt_pv means a UVA, and a NULL
2716 * pt_pv means a KVA.
2719 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
2723 while (sva < va_next) {
2725 * Acquire the related pte_pv, if any. If *ptep == 0
2726 * the related pte_pv should not exist, but if *ptep
2727 * is not zero the pte_pv may or may not exist (e.g.
2728 * will not exist for an unmanaged page).
2730 * However a multitude of races are possible here.
2732 * In addition, the (pt_pv, pte_pv) lock order is
2733 * backwards, so we have to be careful in aquiring
2734 * a properly locked pte_pv.
2738 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
2749 pv_put(pt_pv); /* must be non-NULL */
2751 pv_lock(pte_pv); /* safe to block now */
2754 pt_pv = pv_get(pmap,
2755 pmap_pt_pindex(sva));
2759 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2763 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
2767 kprintf("Unexpected non-NULL pte_pv "
2768 "%p pt_pv %p *ptep = %016lx\n",
2769 pte_pv, pt_pv, *ptep);
2770 panic("Unexpected non-NULL pte_pv");
2778 * Ready for the callback. The locked pte_pv (if any)
2779 * is consumed by the callback. pte_pv will exist if
2780 * the page is managed, and will not exist if it
2784 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
2786 ("bad *ptep %016lx sva %016lx "
2788 *ptep, sva, pte_pv));
2789 func(pmap, &info, pte_pv, pt_pv, sva,
2792 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
2794 ("bad *ptep %016lx sva %016lx "
2797 func(pmap, &info, pte_pv, pt_pv, sva,
2817 pmap_inval_done(&info);
2818 lwkt_reltoken(&pmap->pm_token);
2822 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2824 pmap_scan(pmap, sva, eva, pmap_remove_callback, NULL);
2828 pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
2829 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
2830 pt_entry_t *ptep, void *arg __unused)
2836 * This will also drop pt_pv's wire_count. Note that
2837 * terminal pages are not wired based on mmu presence.
2839 pmap_remove_pv_pte(pte_pv, pt_pv, info);
2840 pmap_remove_pv_page(pte_pv, 0);
2844 * pt_pv's wire_count is still bumped by unmanaged pages
2845 * so we must decrement it manually.
2847 pmap_inval_interlock(info, pmap, va);
2848 pte = pte_load_clear(ptep);
2849 pmap_inval_deinterlock(info, pmap);
2851 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2852 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2853 if (pt_pv && vm_page_unwire_quick(pt_pv->pv_m))
2854 panic("pmap_remove: insufficient wirecount");
2859 * Removes this physical page from all physical maps in which it resides.
2860 * Reflects back modify bits to the pager.
2862 * This routine may not be called from an interrupt.
2866 pmap_remove_all(vm_page_t m)
2868 struct pmap_inval_info info;
2871 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2874 pmap_inval_init(&info);
2875 vm_page_spin_lock(m);
2876 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2877 KKASSERT(pv->pv_m == m);
2878 if (pv_hold_try(pv)) {
2879 vm_page_spin_unlock(m);
2881 vm_page_spin_unlock(m);
2883 if (pv->pv_m != m) {
2885 vm_page_spin_lock(m);
2890 * Holding no spinlocks, pv is locked.
2892 pmap_remove_pv_pte(pv, NULL, &info);
2893 pmap_remove_pv_page(pv, 0);
2895 vm_page_spin_lock(m);
2897 vm_page_spin_unlock(m);
2898 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2899 pmap_inval_done(&info);
2905 * Set the physical protection on the specified range of this map
2908 * This function may not be called from an interrupt if the map is
2909 * not the kernel_pmap.
2912 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2914 /* JG review for NX */
2918 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2919 pmap_remove(pmap, sva, eva);
2922 if (prot & VM_PROT_WRITE)
2924 pmap_scan(pmap, sva, eva, pmap_protect_callback, &prot);
2929 pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
2930 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
2931 pt_entry_t *ptep, void *arg __unused)
2940 pmap_inval_interlock(info, pmap, va);
2947 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2948 KKASSERT(m == pte_pv->pv_m);
2949 vm_page_flag_set(m, PG_REFERENCED);
2953 if (pmap_track_modified(pte_pv->pv_pindex)) {
2955 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2962 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
2965 pmap_inval_deinterlock(info, pmap);
2971 * Insert the vm_page (m) at the virtual address (va), replacing any prior
2972 * mapping at that address. Set protection and wiring as requested.
2974 * NOTE: This routine MUST insert the page into the pmap now, it cannot
2978 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2981 pmap_inval_info info;
2982 pv_entry_t pt_pv; /* page table */
2983 pv_entry_t pte_pv; /* page table entry */
2986 pt_entry_t origpte, newpte;
2991 va = trunc_page(va);
2992 #ifdef PMAP_DIAGNOSTIC
2994 panic("pmap_enter: toobig");
2995 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2996 panic("pmap_enter: invalid to pmap_enter page table "
2997 "pages (va: 0x%lx)", va);
2999 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
3000 kprintf("Warning: pmap_enter called on UVA with "
3003 db_print_backtrace();
3006 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
3007 kprintf("Warning: pmap_enter called on KVA without"
3010 db_print_backtrace();
3015 * Get locked PV entries for our new page table entry (pte_pv)
3016 * and for its parent page table (pt_pv). We need the parent
3017 * so we can resolve the location of the ptep.
3019 * Only hardware MMU actions can modify the ptep out from
3022 * if (m) is fictitious or unmanaged we do not create a managing
3023 * pte_pv for it. Any pre-existing page's management state must
3024 * match (avoiding code complexity).
3026 * If the pmap is still being initialized we assume existing
3029 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3030 * pmap_allocpte() checks the
3032 if (pmap_initialized == FALSE) {
3036 } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) {
3038 if (va >= VM_MAX_USER_ADDRESS) {
3042 pt_pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
3043 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3045 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0);
3047 if (va >= VM_MAX_USER_ADDRESS) {
3049 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3052 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va),
3054 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3056 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED));
3059 if ((prot & VM_PROT_NOSYNC) == 0)
3060 pmap_inval_init(&info);
3062 pa = VM_PAGE_TO_PHYS(m);
3064 opa = origpte & PG_FRAME;
3067 * Mapping has not changed, must be protection or wiring change.
3069 if (origpte && (opa == pa)) {
3071 * Wiring change, just update stats. We don't worry about
3072 * wiring PT pages as they remain resident as long as there
3073 * are valid mappings in them. Hence, if a user page is wired,
3074 * the PT page will be also.
3076 KKASSERT(pte_pv == NULL || m == pte_pv->pv_m);
3077 if (wired && ((origpte & PG_W) == 0))
3078 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3079 else if (!wired && (origpte & PG_W))
3080 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3082 #if defined(PMAP_DIAGNOSTIC)
3083 if (pmap_nw_modified(origpte)) {
3084 kprintf("pmap_enter: modified page not writable: "
3085 "va: 0x%lx, pte: 0x%lx\n", va, origpte);
3090 * We might be turning off write access to the page,
3091 * so we go ahead and sense modify status.
3094 if ((origpte & PG_M) &&
3095 pmap_track_modified(pte_pv->pv_pindex)) {
3098 KKASSERT(PHYS_TO_VM_PAGE(opa) == om);
3107 * Mapping has changed, invalidate old range and fall through to
3108 * handle validating new mapping.
3110 * We always interlock pte removals.
3114 /* XXX pmap_remove_pv_pte() unwires pt_pv */
3115 vm_page_wire_quick(pt_pv->pv_m);
3116 if (prot & VM_PROT_NOSYNC)
3117 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3119 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3121 pmap_remove_pv_page(pte_pv, 0);
3122 } else if (prot & VM_PROT_NOSYNC) {
3124 cpu_invlpg((void *)va);
3125 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3127 pmap_inval_interlock(&info, pmap, va);
3129 pmap_inval_deinterlock(&info, pmap);
3130 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3132 KKASSERT(*ptep == 0);
3136 * Enter on the PV list if part of our managed memory. Wiring is
3137 * handled automatically.
3140 KKASSERT(pte_pv->pv_m == NULL);
3141 vm_page_spin_lock(m);
3143 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3146 atomic_add_int(&m->object->agg_pv_list_count, 1);
3148 vm_page_flag_set(m, PG_MAPPED);
3149 vm_page_spin_unlock(m);
3151 } else if (pt_pv && opa == 0) {
3152 vm_page_wire_quick(pt_pv->pv_m);
3156 * Increment counters
3159 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3163 * Now validate mapping with desired protection/wiring.
3165 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V);
3169 if (va < VM_MAX_USER_ADDRESS)
3171 if (pmap == &kernel_pmap)
3175 * If the mapping or permission bits are different, we need
3176 * to update the pte.
3178 * We do not have to interlock pte insertions as no other
3179 * cpu will have a TLB entry.
3181 if ((origpte & ~(PG_M|PG_A)) != newpte) {
3183 if ((prot & VM_PROT_NOSYNC) == 0)
3184 pmap_inval_interlock(&info, pmap, va);
3186 *ptep = newpte | PG_A;
3187 cpu_invlpg((void *)va);
3189 if (prot & VM_PROT_NOSYNC)
3190 cpu_invlpg((void *)va);
3192 pmap_inval_deinterlock(&info, pmap);
3195 vm_page_flag_set(m, PG_WRITEABLE);
3197 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3200 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
3201 if ((prot & VM_PROT_NOSYNC) == 0)
3202 pmap_inval_done(&info);
3205 * Cleanup the pv entry, allowing other accessors.
3214 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
3215 * This code also assumes that the pmap has no pre-existing entry for this
3218 * This code currently may only be used on user pmaps, not kernel_pmap.
3221 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
3223 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE);
3227 * Make a temporary mapping for a physical address. This is only intended
3228 * to be used for panic dumps.
3230 * The caller is responsible for calling smp_invltlb().
3233 pmap_kenter_temporary(vm_paddr_t pa, long i)
3235 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
3236 return ((void *)crashdumpmap);
3239 #define MAX_INIT_PT (96)
3242 * This routine preloads the ptes for a given object into the specified pmap.
3243 * This eliminates the blast of soft faults on process startup and
3244 * immediately after an mmap.
3246 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
3249 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
3250 vm_object_t object, vm_pindex_t pindex,
3251 vm_size_t size, int limit)
3253 struct rb_vm_page_scan_info info;
3258 * We can't preinit if read access isn't set or there is no pmap
3261 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
3265 * We can't preinit if the pmap is not the current pmap
3267 lp = curthread->td_lwp;
3268 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
3271 psize = x86_64_btop(size);
3273 if ((object->type != OBJT_VNODE) ||
3274 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
3275 (object->resident_page_count > MAX_INIT_PT))) {
3279 if (pindex + psize > object->size) {
3280 if (object->size < pindex)
3282 psize = object->size - pindex;
3289 * Use a red-black scan to traverse the requested range and load
3290 * any valid pages found into the pmap.
3292 * We cannot safely scan the object's memq without holding the
3295 info.start_pindex = pindex;
3296 info.end_pindex = pindex + psize - 1;
3302 vm_object_hold(object);
3303 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3304 pmap_object_init_pt_callback, &info);
3305 vm_object_drop(object);
3310 pmap_object_init_pt_callback(vm_page_t p, void *data)
3312 struct rb_vm_page_scan_info *info = data;
3313 vm_pindex_t rel_index;
3316 * don't allow an madvise to blow away our really
3317 * free pages allocating pv entries.
3319 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3320 vmstats.v_free_count < vmstats.v_free_reserved) {
3323 if (vm_page_busy_try(p, TRUE))
3325 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3326 (p->flags & PG_FICTITIOUS) == 0) {
3327 if ((p->queue - p->pc) == PQ_CACHE)
3328 vm_page_deactivate(p);
3329 rel_index = p->pindex - info->start_pindex;
3330 pmap_enter_quick(info->pmap,
3331 info->addr + x86_64_ptob(rel_index), p);
3339 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
3342 * Returns FALSE if it would be non-trivial or if a pte is already loaded
3346 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3350 spin_lock(&pmap->pm_spin);
3351 if ((pte = pmap_pte(pmap, addr)) != NULL) {
3353 spin_unlock(&pmap->pm_spin);
3357 spin_unlock(&pmap->pm_spin);
3362 * Change the wiring attribute for a pmap/va pair. The mapping must already
3363 * exist in the pmap. The mapping may or may not be managed.
3366 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
3373 lwkt_gettoken(&pmap->pm_token);
3374 pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
3375 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
3377 if (wired && !pmap_pte_w(ptep))
3378 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3379 else if (!wired && pmap_pte_w(ptep))
3380 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3383 * Wiring is not a hardware characteristic so there is no need to
3384 * invalidate TLB. However, in an SMP environment we must use
3385 * a locked bus cycle to update the pte (if we are not using
3386 * the pmap_inval_*() API that is)... it's ok to do this for simple
3391 atomic_set_long(ptep, PG_W);
3393 atomic_clear_long(ptep, PG_W);
3396 atomic_set_long_nonlocked(ptep, PG_W);
3398 atomic_clear_long_nonlocked(ptep, PG_W);
3401 lwkt_reltoken(&pmap->pm_token);
3407 * Copy the range specified by src_addr/len from the source map to
3408 * the range dst_addr/len in the destination map.
3410 * This routine is only advisory and need not do anything.
3413 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3414 vm_size_t len, vm_offset_t src_addr)
3421 * Zero the specified physical page.
3423 * This function may be called from an interrupt and no locking is
3427 pmap_zero_page(vm_paddr_t phys)
3429 vm_offset_t va = PHYS_TO_DMAP(phys);
3431 pagezero((void *)va);
3435 * pmap_page_assertzero:
3437 * Assert that a page is empty, panic if it isn't.
3440 pmap_page_assertzero(vm_paddr_t phys)
3442 vm_offset_t va = PHYS_TO_DMAP(phys);
3445 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3446 if (*(long *)((char *)va + i) != 0) {
3447 panic("pmap_page_assertzero() @ %p not zero!\n",
3448 (void *)(intptr_t)va);
3456 * Zero part of a physical page by mapping it into memory and clearing
3457 * its contents with bzero.
3459 * off and size may not cover an area beyond a single hardware page.
3462 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3464 vm_offset_t virt = PHYS_TO_DMAP(phys);
3466 bzero((char *)virt + off, size);
3472 * Copy the physical page from the source PA to the target PA.
3473 * This function may be called from an interrupt. No locking
3477 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3479 vm_offset_t src_virt, dst_virt;
3481 src_virt = PHYS_TO_DMAP(src);
3482 dst_virt = PHYS_TO_DMAP(dst);
3483 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3487 * pmap_copy_page_frag:
3489 * Copy the physical page from the source PA to the target PA.
3490 * This function may be called from an interrupt. No locking
3494 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3496 vm_offset_t src_virt, dst_virt;
3498 src_virt = PHYS_TO_DMAP(src);
3499 dst_virt = PHYS_TO_DMAP(dst);
3501 bcopy((char *)src_virt + (src & PAGE_MASK),
3502 (char *)dst_virt + (dst & PAGE_MASK),
3507 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
3508 * this page. This count may be changed upwards or downwards in the future;
3509 * it is only necessary that true be returned for a small subset of pmaps
3510 * for proper page aging.
3513 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3518 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3521 vm_page_spin_lock(m);
3522 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3523 if (pv->pv_pmap == pmap) {
3524 vm_page_spin_unlock(m);
3531 vm_page_spin_unlock(m);
3536 * Remove all pages from specified address space this aids process exit
3537 * speeds. Also, this code may be special cased for the current process
3541 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3543 pmap_remove(pmap, sva, eva);
3547 * pmap_testbit tests bits in pte's note that the testbit/clearbit
3548 * routines are inline, and a lot of things compile-time evaluate.
3552 pmap_testbit(vm_page_t m, int bit)
3557 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3560 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3562 vm_page_spin_lock(m);
3563 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
3564 vm_page_spin_unlock(m);
3568 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3570 * if the bit being tested is the modified bit, then
3571 * mark clean_map and ptes as never
3574 if (bit & (PG_A|PG_M)) {
3575 if (!pmap_track_modified(pv->pv_pindex))
3579 #if defined(PMAP_DIAGNOSTIC)
3580 if (pv->pv_pmap == NULL) {
3581 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
3586 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3588 vm_page_spin_unlock(m);
3592 vm_page_spin_unlock(m);
3597 * This routine is used to modify bits in ptes
3599 * Caller must NOT hold any spin locks
3603 pmap_clearbit(vm_page_t m, int bit)
3605 struct pmap_inval_info info;
3609 vm_pindex_t save_pindex;
3613 vm_page_flag_clear(m, PG_WRITEABLE);
3614 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
3618 pmap_inval_init(&info);
3621 * Loop over all current mappings setting/clearing as appropos If
3622 * setting RO do we need to clear the VAC?
3624 vm_page_spin_lock(m);
3626 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3628 * don't write protect pager mappings
3631 if (!pmap_track_modified(pv->pv_pindex))
3635 #if defined(PMAP_DIAGNOSTIC)
3636 if (pv->pv_pmap == NULL) {
3637 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
3644 * Careful here. We can use a locked bus instruction to
3645 * clear PG_A or PG_M safely but we need to synchronize
3646 * with the target cpus when we mess with PG_RW.
3648 * We do not have to force synchronization when clearing
3649 * PG_M even for PTEs generated via virtual memory maps,
3650 * because the virtual kernel will invalidate the pmap
3651 * entry when/if it needs to resynchronize the Modify bit.
3654 save_pmap = pv->pv_pmap;
3655 save_pindex = pv->pv_pindex;
3657 vm_page_spin_unlock(m);
3658 pmap_inval_interlock(&info, save_pmap,
3659 (vm_offset_t)save_pindex << PAGE_SHIFT);
3660 vm_page_spin_lock(m);
3661 if (pv->pv_pmap == NULL) {
3667 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3674 atomic_clear_long(pte, PG_M|PG_RW);
3677 * The cpu may be trying to set PG_M
3678 * simultaniously with our clearing
3681 if (!atomic_cmpset_long(pte, pbits,
3685 } else if (bit == PG_M) {
3687 * We could also clear PG_RW here to force
3688 * a fault on write to redetect PG_M for
3689 * virtual kernels, but it isn't necessary
3690 * since virtual kernels invalidate the pte
3691 * when they clear the VPTE_M bit in their
3692 * virtual page tables.
3694 atomic_clear_long(pte, PG_M);
3696 atomic_clear_long(pte, bit);
3700 save_pmap = pv->pv_pmap;
3702 vm_page_spin_unlock(m);
3703 pmap_inval_deinterlock(&info, save_pmap);
3704 vm_page_spin_lock(m);
3705 if (pv->pv_pmap == NULL) {
3712 vm_page_spin_unlock(m);
3713 pmap_inval_done(&info);
3717 * Lower the permission for all mappings to a given page.
3719 * Page must be busied by caller.
3722 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3724 /* JG NX support? */
3725 if ((prot & VM_PROT_WRITE) == 0) {
3726 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3728 * NOTE: pmap_clearbit(.. PG_RW) also clears
3729 * the PG_WRITEABLE flag in (m).
3731 pmap_clearbit(m, PG_RW);
3739 pmap_phys_address(vm_pindex_t ppn)
3741 return (x86_64_ptob(ppn));
3745 * Return a count of reference bits for a page, clearing those bits.
3746 * It is not necessary for every reference bit to be cleared, but it
3747 * is necessary that 0 only be returned when there are truly no
3748 * reference bits set.
3750 * XXX: The exact number of bits to check and clear is a matter that
3751 * should be tested and standardized at some point in the future for
3752 * optimal aging of shared pages.
3754 * This routine may not block.
3757 pmap_ts_referenced(vm_page_t m)
3763 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3766 vm_page_spin_lock(m);
3767 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3768 if (!pmap_track_modified(pv->pv_pindex))
3770 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3771 if (pte && (*pte & PG_A)) {
3773 atomic_clear_long(pte, PG_A);
3775 atomic_clear_long_nonlocked(pte, PG_A);
3782 vm_page_spin_unlock(m);
3789 * Return whether or not the specified physical page was modified
3790 * in any physical maps.
3793 pmap_is_modified(vm_page_t m)
3797 res = pmap_testbit(m, PG_M);
3802 * Clear the modify bits on the specified physical page.
3805 pmap_clear_modify(vm_page_t m)
3807 pmap_clearbit(m, PG_M);
3811 * pmap_clear_reference:
3813 * Clear the reference bit on the specified physical page.
3816 pmap_clear_reference(vm_page_t m)
3818 pmap_clearbit(m, PG_A);
3822 * Miscellaneous support routines follow
3827 i386_protection_init(void)
3831 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3832 kp = protection_codes;
3833 for (prot = 0; prot < 8; prot++) {
3835 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3837 * Read access is also 0. There isn't any execute bit,
3838 * so just make it readable.
3840 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3841 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3842 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3845 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3846 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3847 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3848 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3856 * Map a set of physical memory pages into the kernel virtual
3857 * address space. Return a pointer to where it is mapped. This
3858 * routine is intended to be used for mapping device memory,
3861 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3865 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3867 vm_offset_t va, tmpva, offset;
3870 offset = pa & PAGE_MASK;
3871 size = roundup(offset + size, PAGE_SIZE);
3873 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3875 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3877 pa = pa & ~PAGE_MASK;
3878 for (tmpva = va; size > 0;) {
3879 pte = vtopte(tmpva);
3880 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3888 return ((void *)(va + offset));
3892 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3894 vm_offset_t va, tmpva, offset;
3897 offset = pa & PAGE_MASK;
3898 size = roundup(offset + size, PAGE_SIZE);
3900 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3902 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3904 pa = pa & ~PAGE_MASK;
3905 for (tmpva = va; size > 0;) {
3906 pte = vtopte(tmpva);
3907 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3915 return ((void *)(va + offset));
3919 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3921 vm_offset_t base, offset;
3923 base = va & ~PAGE_MASK;
3924 offset = va & PAGE_MASK;
3925 size = roundup(offset + size, PAGE_SIZE);
3926 pmap_qremove(va, size >> PAGE_SHIFT);
3927 kmem_free(&kernel_map, base, size);
3931 * perform the pmap work for mincore
3934 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3936 pt_entry_t *ptep, pte;
3940 lwkt_gettoken(&pmap->pm_token);
3941 ptep = pmap_pte(pmap, addr);
3943 if (ptep && (pte = *ptep) != 0) {
3946 val = MINCORE_INCORE;
3947 if ((pte & PG_MANAGED) == 0)
3950 pa = pte & PG_FRAME;
3952 m = PHYS_TO_VM_PAGE(pa);
3958 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3960 * Modified by someone
3962 else if (m->dirty || pmap_is_modified(m))
3963 val |= MINCORE_MODIFIED_OTHER;
3968 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3971 * Referenced by someone
3973 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3974 val |= MINCORE_REFERENCED_OTHER;
3975 vm_page_flag_set(m, PG_REFERENCED);
3979 lwkt_reltoken(&pmap->pm_token);
3985 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3986 * vmspace will be ref'd and the old one will be deref'd.
3988 * The vmspace for all lwps associated with the process will be adjusted
3989 * and cr3 will be reloaded if any lwp is the current lwp.
3991 * The process must hold the vmspace->vm_map.token for oldvm and newvm
3994 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3996 struct vmspace *oldvm;
3999 oldvm = p->p_vmspace;
4000 if (oldvm != newvm) {
4002 sysref_get(&newvm->vm_sysref);
4003 p->p_vmspace = newvm;
4004 KKASSERT(p->p_nthreads == 1);
4005 lp = RB_ROOT(&p->p_lwp_tree);
4006 pmap_setlwpvm(lp, newvm);
4008 sysref_put(&oldvm->vm_sysref);
4013 * Set the vmspace for a LWP. The vmspace is almost universally set the
4014 * same as the process vmspace, but virtual kernels need to swap out contexts
4015 * on a per-lwp basis.
4017 * Caller does not necessarily hold any vmspace tokens. Caller must control
4018 * the lwp (typically be in the context of the lwp). We use a critical
4019 * section to protect against statclock and hardclock (statistics collection).
4022 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
4024 struct vmspace *oldvm;
4027 oldvm = lp->lwp_vmspace;
4029 if (oldvm != newvm) {
4031 lp->lwp_vmspace = newvm;
4032 if (curthread->td_lwp == lp) {
4033 pmap = vmspace_pmap(newvm);
4035 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4036 if (pmap->pm_active & CPUMASK_LOCK)
4037 pmap_interlock_wait(newvm);
4039 pmap->pm_active |= 1;
4041 #if defined(SWTCH_OPTIM_STATS)
4044 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4045 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
4046 load_cr3(curthread->td_pcb->pcb_cr3);
4047 pmap = vmspace_pmap(oldvm);
4049 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4051 pmap->pm_active &= ~(cpumask_t)1;
4061 * Called when switching to a locked pmap, used to interlock against pmaps
4062 * undergoing modifications to prevent us from activating the MMU for the
4063 * target pmap until all such modifications have completed. We have to do
4064 * this because the thread making the modifications has already set up its
4065 * SMP synchronization mask.
4070 pmap_interlock_wait(struct vmspace *vm)
4072 struct pmap *pmap = &vm->vm_pmap;
4074 if (pmap->pm_active & CPUMASK_LOCK) {
4076 DEBUG_PUSH_INFO("pmap_interlock_wait");
4077 while (pmap->pm_active & CPUMASK_LOCK) {
4079 lwkt_process_ipiq();
4089 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4092 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4096 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4101 * Used by kmalloc/kfree, page already exists at va
4104 pmap_kvtom(vm_offset_t va)
4106 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));