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 = NULL, *ptmmap;
199 caddr_t CADDR1 = 0, ptvmmap = 0;
200 static pt_entry_t *msgbufmap;
201 struct msgbuf *msgbufp=NULL;
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);
1364 KKASSERT(pmap->pm_pml4[255] == 0);
1365 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1366 KKASSERT(pv->pv_entry.rbe_left == NULL);
1367 KKASSERT(pv->pv_entry.rbe_right == NULL);
1371 * Clean up a pmap structure so it can be physically freed. This routine
1372 * is called by the vmspace dtor function. A great deal of pmap data is
1373 * left passively mapped to improve vmspace management so we have a bit
1374 * of cleanup work to do here.
1377 pmap_puninit(pmap_t pmap)
1382 KKASSERT(pmap->pm_active == 0);
1383 if ((pv = pmap->pm_pmlpv) != NULL) {
1384 if (pv_hold_try(pv) == 0)
1386 p = pmap_remove_pv_page(pv);
1388 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1389 vm_page_busy_wait(p, FALSE, "pgpun");
1390 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1391 vm_page_unwire(p, 0);
1392 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1395 * XXX eventually clean out PML4 static entries and
1396 * use vm_page_free_zero()
1399 pmap->pm_pmlpv = NULL;
1401 if (pmap->pm_pml4) {
1402 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1403 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1404 pmap->pm_pml4 = NULL;
1406 KKASSERT(pmap->pm_stats.resident_count == 0);
1407 KKASSERT(pmap->pm_stats.wired_count == 0);
1411 * Wire in kernel global address entries. To avoid a race condition
1412 * between pmap initialization and pmap_growkernel, this procedure
1413 * adds the pmap to the master list (which growkernel scans to update),
1414 * then copies the template.
1417 pmap_pinit2(struct pmap *pmap)
1420 * XXX copies current process, does not fill in MPPTDI
1422 spin_lock(&pmap_spin);
1423 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1424 spin_unlock(&pmap_spin);
1428 * This routine is called when various levels in the page table need to
1429 * be populated. This routine cannot fail.
1431 * This function returns two locked pv_entry's, one representing the
1432 * requested pv and one representing the requested pv's parent pv. If
1433 * the pv did not previously exist it will be mapped into its parent
1434 * and wired, otherwise no additional wire count will be added.
1438 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1443 vm_pindex_t pt_pindex;
1448 * If the pv already exists and we aren't being asked for the
1449 * parent page table page we can just return it. A locked+held pv
1452 pv = pv_alloc(pmap, ptepindex, &isnew);
1453 if (isnew == 0 && pvpp == NULL)
1457 * This is a new PV, we have to resolve its parent page table and
1458 * add an additional wiring to the page if necessary.
1462 * Special case terminal PVs. These are not page table pages so
1463 * no vm_page is allocated (the caller supplied the vm_page). If
1464 * pvpp is non-NULL we are being asked to also removed the pt_pv
1467 * Note that pt_pv's are only returned for user VAs. We assert that
1468 * a pt_pv is not being requested for kernel VAs.
1470 if (ptepindex < pmap_pt_pindex(0)) {
1471 if (ptepindex >= NUPTE_USER)
1472 KKASSERT(pvpp == NULL);
1474 KKASSERT(pvpp != NULL);
1476 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1477 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1479 vm_page_wire_quick(pvp->pv_m);
1488 * Non-terminal PVs allocate a VM page to represent the page table,
1489 * so we have to resolve pvp and calculate ptepindex for the pvp
1490 * and then for the page table entry index in the pvp for
1493 if (ptepindex < pmap_pd_pindex(0)) {
1495 * pv is PT, pvp is PD
1497 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1498 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1499 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1506 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1507 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1508 } else if (ptepindex < pmap_pdp_pindex(0)) {
1510 * pv is PD, pvp is PDP
1512 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1513 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1514 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1521 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1522 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1523 } else if (ptepindex < pmap_pml4_pindex()) {
1525 * pv is PDP, pvp is the root pml4 table
1527 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1534 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1535 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1538 * pv represents the top-level PML4, there is no parent.
1546 * This code is only reached if isnew is TRUE and this is not a
1547 * terminal PV. We need to allocate a vm_page for the page table
1548 * at this level and enter it into the parent page table.
1550 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1553 m = vm_page_alloc(NULL, pv->pv_pindex,
1554 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1555 VM_ALLOC_INTERRUPT);
1560 vm_page_spin_lock(m);
1561 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1563 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1564 vm_page_spin_unlock(m);
1565 vm_page_unmanage(m); /* m must be spinunlocked */
1567 if ((m->flags & PG_ZERO) == 0) {
1568 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1572 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1575 m->valid = VM_PAGE_BITS_ALL;
1576 vm_page_flag_clear(m, PG_ZERO);
1577 vm_page_wire(m); /* wire for mapping in parent */
1580 * Wire the page into pvp, bump the wire-count for pvp's page table
1581 * page. Bump the resident_count for the pmap. There is no pvp
1582 * for the top level, address the pm_pml4[] array directly.
1584 * If the caller wants the parent we return it, otherwise
1585 * we just put it away.
1587 * No interlock is needed for pte 0 -> non-zero.
1590 vm_page_wire_quick(pvp->pv_m);
1591 ptep = pv_pte_lookup(pvp, ptepindex);
1592 KKASSERT((*ptep & PG_V) == 0);
1593 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1606 * Release any resources held by the given physical map.
1608 * Called when a pmap initialized by pmap_pinit is being released. Should
1609 * only be called if the map contains no valid mappings.
1611 * Caller must hold pmap->pm_token
1613 struct pmap_release_info {
1618 static int pmap_release_callback(pv_entry_t pv, void *data);
1621 pmap_release(struct pmap *pmap)
1623 struct pmap_release_info info;
1625 KASSERT(pmap->pm_active == 0,
1626 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1628 spin_lock(&pmap_spin);
1629 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1630 spin_unlock(&pmap_spin);
1633 * Pull pv's off the RB tree in order from low to high and release
1639 spin_lock(&pmap->pm_spin);
1640 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
1641 pmap_release_callback, &info);
1642 spin_unlock(&pmap->pm_spin);
1643 } while (info.retry);
1647 * One resident page (the pml4 page) should remain.
1648 * No wired pages should remain.
1650 KKASSERT(pmap->pm_stats.resident_count == 1);
1651 KKASSERT(pmap->pm_stats.wired_count == 0);
1655 pmap_release_callback(pv_entry_t pv, void *data)
1657 struct pmap_release_info *info = data;
1658 pmap_t pmap = info->pmap;
1661 if (pv_hold_try(pv)) {
1662 spin_unlock(&pmap->pm_spin);
1664 spin_unlock(&pmap->pm_spin);
1666 if (pv->pv_pmap != pmap) {
1668 spin_lock(&pmap->pm_spin);
1675 * The pmap is currently not spinlocked, pv is held+locked.
1676 * Remove the pv's page from its parent's page table. The
1677 * parent's page table page's wire_count will be decremented.
1679 pmap_remove_pv_pte(pv, NULL, NULL);
1682 * Terminal pvs are unhooked from their vm_pages. Because
1683 * terminal pages aren't page table pages they aren't wired
1684 * by us, so we have to be sure not to unwire them either.
1686 if (pv->pv_pindex < pmap_pt_pindex(0)) {
1687 pmap_remove_pv_page(pv);
1692 * We leave the top-level page table page cached, wired, and
1693 * mapped in the pmap until the dtor function (pmap_puninit())
1696 * Since we are leaving the top-level pv intact we need
1697 * to break out of what would otherwise be an infinite loop.
1699 if (pv->pv_pindex == pmap_pml4_pindex()) {
1701 spin_lock(&pmap->pm_spin);
1706 * For page table pages (other than the top-level page),
1707 * remove and free the vm_page. The representitive mapping
1708 * removed above by pmap_remove_pv_pte() did not undo the
1709 * last wire_count so we have to do that as well.
1711 p = pmap_remove_pv_page(pv);
1712 vm_page_busy_wait(p, FALSE, "pmaprl");
1713 if (p->wire_count != 1) {
1714 kprintf("p->wire_count was %016lx %d\n",
1715 pv->pv_pindex, p->wire_count);
1717 KKASSERT(p->wire_count == 1);
1718 KKASSERT(p->flags & PG_UNMANAGED);
1720 vm_page_unwire(p, 0);
1721 KKASSERT(p->wire_count == 0);
1722 /* JG eventually revert to using vm_page_free_zero() */
1726 spin_lock(&pmap->pm_spin);
1731 * This function will remove the pte associated with a pv from its parent.
1732 * Terminal pv's are supported. The removal will be interlocked if info
1733 * is non-NULL. The caller must dispose of pv instead of just unlocking
1736 * The wire count will be dropped on the parent page table. The wire
1737 * count on the page being removed (pv->pv_m) from the parent page table
1738 * is NOT touched. Note that terminal pages will not have any additional
1739 * wire counts while page table pages will have at least one representing
1740 * the mapping, plus others representing sub-mappings.
1742 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
1743 * pages and user page table and terminal pages.
1745 * The pv must be locked.
1747 * XXX must lock parent pv's if they exist to remove pte XXX
1751 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
1753 vm_pindex_t ptepindex = pv->pv_pindex;
1754 pmap_t pmap = pv->pv_pmap;
1760 if (ptepindex == pmap_pml4_pindex()) {
1762 * We are the top level pml4 table, there is no parent.
1764 p = pmap->pm_pmlpv->pv_m;
1765 } else if (ptepindex >= pmap_pdp_pindex(0)) {
1767 * Remove a PDP page from the pml4e. This can only occur
1768 * with user page tables. We do not have to lock the
1769 * pml4 PV so just ignore pvp.
1771 vm_pindex_t pml4_pindex;
1772 vm_pindex_t pdp_index;
1775 pdp_index = ptepindex - pmap_pdp_pindex(0);
1777 pml4_pindex = pmap_pml4_pindex();
1778 pvp = pv_get(pv->pv_pmap, pml4_pindex);
1781 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
1782 KKASSERT((*pdp & PG_V) != 0);
1783 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1785 KKASSERT(info == NULL);
1786 } else if (ptepindex >= pmap_pd_pindex(0)) {
1788 * Remove a PD page from the pdp
1790 vm_pindex_t pdp_pindex;
1791 vm_pindex_t pd_index;
1794 pd_index = ptepindex - pmap_pd_pindex(0);
1797 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
1798 (pd_index >> NPML4EPGSHIFT);
1799 pvp = pv_get(pv->pv_pmap, pdp_pindex);
1802 pd = pv_pte_lookup(pvp, pd_index & ((1ul << NPDPEPGSHIFT) - 1));
1803 KKASSERT((*pd & PG_V) != 0);
1804 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
1806 KKASSERT(info == NULL);
1807 } else if (ptepindex >= pmap_pt_pindex(0)) {
1809 * Remove a PT page from the pd
1811 vm_pindex_t pd_pindex;
1812 vm_pindex_t pt_index;
1815 pt_index = ptepindex - pmap_pt_pindex(0);
1818 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
1819 (pt_index >> NPDPEPGSHIFT);
1820 pvp = pv_get(pv->pv_pmap, pd_pindex);
1823 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
1824 KKASSERT((*pt & PG_V) != 0);
1825 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
1827 KKASSERT(info == NULL);
1830 * Remove a PTE from the PT page
1832 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
1833 * pv is a pte_pv so we can safely lock pt_pv.
1835 vm_pindex_t pt_pindex;
1840 pt_pindex = ptepindex >> NPTEPGSHIFT;
1841 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
1843 if (ptepindex >= NUPTE_USER) {
1844 ptep = vtopte(ptepindex << PAGE_SHIFT);
1845 KKASSERT(pvp == NULL);
1848 pt_pindex = NUPTE_TOTAL +
1849 (ptepindex >> NPDPEPGSHIFT);
1850 pvp = pv_get(pv->pv_pmap, pt_pindex);
1853 ptep = pv_pte_lookup(pvp, ptepindex &
1854 ((1ul << NPDPEPGSHIFT) - 1));
1858 pmap_inval_interlock(info, pmap, va);
1859 pte = pte_load_clear(ptep);
1861 pmap_inval_deinterlock(info, pmap);
1863 cpu_invlpg((void *)va);
1866 * Now update the vm_page_t
1868 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
1869 kprintf("remove_pte badpte %016lx %016lx %d\n",
1871 pv->pv_pindex < pmap_pt_pindex(0));
1873 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
1874 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
1877 if (pmap_track_modified(ptepindex))
1881 vm_page_flag_set(p, PG_REFERENCED);
1884 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1886 cpu_invlpg((void *)va);
1890 * Unwire the parent page table page. The wire_count cannot go below
1891 * 1 here because the parent page table page is itself still mapped.
1893 * XXX remove the assertions later.
1895 KKASSERT(pv->pv_m == p);
1896 if (pvp && vm_page_unwire_quick(pvp->pv_m))
1897 panic("pmap_remove_pv_pte: Insufficient wire_count");
1905 pmap_remove_pv_page(pv_entry_t pv)
1911 vm_page_spin_lock(m);
1913 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1916 atomic_add_int(&m->object->agg_pv_list_count, -1);
1918 if (TAILQ_EMPTY(&m->md.pv_list))
1919 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1920 vm_page_spin_unlock(m);
1925 * Grow the number of kernel page table entries, if needed.
1927 * This routine is always called to validate any address space
1928 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1929 * space below KERNBASE.
1932 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1935 vm_offset_t ptppaddr;
1937 pd_entry_t *pt, newpt;
1939 int update_kernel_vm_end;
1942 * bootstrap kernel_vm_end on first real VM use
1944 if (kernel_vm_end == 0) {
1945 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1947 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1948 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1949 ~(PAGE_SIZE * NPTEPG - 1);
1951 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1952 kernel_vm_end = kernel_map.max_offset;
1959 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1960 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1961 * do not want to force-fill 128G worth of page tables.
1963 if (kstart < KERNBASE) {
1964 if (kstart > kernel_vm_end)
1965 kstart = kernel_vm_end;
1966 KKASSERT(kend <= KERNBASE);
1967 update_kernel_vm_end = 1;
1969 update_kernel_vm_end = 0;
1972 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1973 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1975 if (kend - 1 >= kernel_map.max_offset)
1976 kend = kernel_map.max_offset;
1978 while (kstart < kend) {
1979 pt = pmap_pt(&kernel_pmap, kstart);
1981 /* We need a new PDP entry */
1982 nkpg = vm_page_alloc(NULL, nkpt,
1985 VM_ALLOC_INTERRUPT);
1987 panic("pmap_growkernel: no memory to grow "
1990 paddr = VM_PAGE_TO_PHYS(nkpg);
1991 if ((nkpg->flags & PG_ZERO) == 0)
1992 pmap_zero_page(paddr);
1993 vm_page_flag_clear(nkpg, PG_ZERO);
1994 newpd = (pdp_entry_t)
1995 (paddr | PG_V | PG_RW | PG_A | PG_M);
1996 *pmap_pd(&kernel_pmap, kstart) = newpd;
1998 continue; /* try again */
2000 if ((*pt & PG_V) != 0) {
2001 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2002 ~(PAGE_SIZE * NPTEPG - 1);
2003 if (kstart - 1 >= kernel_map.max_offset) {
2004 kstart = kernel_map.max_offset;
2011 * This index is bogus, but out of the way
2013 nkpg = vm_page_alloc(NULL, nkpt,
2016 VM_ALLOC_INTERRUPT);
2018 panic("pmap_growkernel: no memory to grow kernel");
2021 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2022 pmap_zero_page(ptppaddr);
2023 vm_page_flag_clear(nkpg, PG_ZERO);
2024 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2025 *pmap_pt(&kernel_pmap, kstart) = newpt;
2028 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2029 ~(PAGE_SIZE * NPTEPG - 1);
2031 if (kstart - 1 >= kernel_map.max_offset) {
2032 kstart = kernel_map.max_offset;
2038 * Only update kernel_vm_end for areas below KERNBASE.
2040 if (update_kernel_vm_end && kernel_vm_end < kstart)
2041 kernel_vm_end = kstart;
2045 * Retire the given physical map from service.
2046 * Should only be called if the map contains
2047 * no valid mappings.
2050 pmap_destroy(pmap_t pmap)
2057 lwkt_gettoken(&pmap->pm_token);
2058 count = --pmap->pm_count;
2060 pmap_release(pmap); /* eats pm_token */
2061 panic("destroying a pmap is not yet implemented");
2063 lwkt_reltoken(&pmap->pm_token);
2067 * Add a reference to the specified pmap.
2070 pmap_reference(pmap_t pmap)
2073 lwkt_gettoken(&pmap->pm_token);
2075 lwkt_reltoken(&pmap->pm_token);
2079 /***************************************************
2080 * page management routines.
2081 ***************************************************/
2084 * Hold a pv without locking it
2087 pv_hold(pv_entry_t pv)
2091 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2095 count = pv->pv_hold;
2097 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2104 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2105 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2108 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2109 * pv list via its page) must be held by the caller.
2112 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2116 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2119 pv->pv_line = lineno;
2125 count = pv->pv_hold;
2127 if ((count & PV_HOLD_LOCKED) == 0) {
2128 if (atomic_cmpset_int(&pv->pv_hold, count,
2129 (count + 1) | PV_HOLD_LOCKED)) {
2132 pv->pv_line = lineno;
2137 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2145 * Drop a previously held pv_entry which could not be locked, allowing its
2148 * Must not be called with a spinlock held as we might zfree() the pv if it
2149 * is no longer associated with a pmap and this was the last hold count.
2152 pv_drop(pv_entry_t pv)
2156 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2157 if (pv->pv_pmap == NULL)
2163 count = pv->pv_hold;
2165 KKASSERT((count & PV_HOLD_MASK) > 0);
2166 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2167 (PV_HOLD_LOCKED | 1));
2168 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2169 if (count == 1 && pv->pv_pmap == NULL)
2178 * Find or allocate the requested PV entry, returning a locked pv
2182 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2185 pv_entry_t pnew = NULL;
2187 spin_lock(&pmap->pm_spin);
2189 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2190 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2195 spin_unlock(&pmap->pm_spin);
2196 pnew = zalloc(pvzone);
2197 spin_lock(&pmap->pm_spin);
2200 pnew->pv_pmap = pmap;
2201 pnew->pv_pindex = pindex;
2202 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2204 pnew->pv_func = func;
2205 pnew->pv_line = lineno;
2207 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2208 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2209 spin_unlock(&pmap->pm_spin);
2214 spin_unlock(&pmap->pm_spin);
2215 zfree(pvzone, pnew);
2217 spin_lock(&pmap->pm_spin);
2220 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2221 spin_unlock(&pmap->pm_spin);
2225 spin_unlock(&pmap->pm_spin);
2226 _pv_lock(pv PMAP_DEBUG_COPY);
2227 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2232 spin_lock(&pmap->pm_spin);
2239 * Find the requested PV entry, returning a locked+held pv or NULL
2243 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2247 spin_lock(&pmap->pm_spin);
2252 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2253 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2257 spin_unlock(&pmap->pm_spin);
2260 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2261 pv_cache(pv, pindex);
2262 spin_unlock(&pmap->pm_spin);
2265 spin_unlock(&pmap->pm_spin);
2266 _pv_lock(pv PMAP_DEBUG_COPY);
2267 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2270 spin_lock(&pmap->pm_spin);
2275 * Lookup, hold, and attempt to lock (pmap,pindex).
2277 * If the entry does not exist NULL is returned and *errorp is set to 0
2279 * If the entry exists and could be successfully locked it is returned and
2280 * errorp is set to 0.
2282 * If the entry exists but could NOT be successfully locked it is returned
2283 * held and *errorp is set to 1.
2287 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2291 spin_lock(&pmap->pm_spin);
2292 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2293 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2295 spin_unlock(&pmap->pm_spin);
2299 if (pv_hold_try(pv)) {
2300 pv_cache(pv, pindex);
2301 spin_unlock(&pmap->pm_spin);
2303 return(pv); /* lock succeeded */
2305 spin_unlock(&pmap->pm_spin);
2307 return (pv); /* lock failed */
2311 * Find the requested PV entry, returning a held pv or NULL
2315 pv_find(pmap_t pmap, vm_pindex_t pindex)
2319 spin_lock(&pmap->pm_spin);
2321 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2322 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2324 spin_unlock(&pmap->pm_spin);
2328 pv_cache(pv, pindex);
2329 spin_unlock(&pmap->pm_spin);
2334 * Lock a held pv, keeping the hold count
2338 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2343 count = pv->pv_hold;
2345 if ((count & PV_HOLD_LOCKED) == 0) {
2346 if (atomic_cmpset_int(&pv->pv_hold, count,
2347 count | PV_HOLD_LOCKED)) {
2350 pv->pv_line = lineno;
2356 tsleep_interlock(pv, 0);
2357 if (atomic_cmpset_int(&pv->pv_hold, count,
2358 count | PV_HOLD_WAITING)) {
2360 kprintf("pv waiting on %s:%d\n",
2361 pv->pv_func, pv->pv_line);
2363 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2370 * Unlock a held and locked pv, keeping the hold count.
2374 pv_unlock(pv_entry_t pv)
2378 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2382 count = pv->pv_hold;
2384 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2385 (PV_HOLD_LOCKED | 1));
2386 if (atomic_cmpset_int(&pv->pv_hold, count,
2388 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2389 if (count & PV_HOLD_WAITING)
2397 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2398 * and the hold count drops to zero we will free it.
2400 * Caller should not hold any spin locks. We are protected from hold races
2401 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2402 * lock held. A pv cannot be located otherwise.
2406 pv_put(pv_entry_t pv)
2408 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2409 if (pv->pv_pmap == NULL)
2418 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2419 * pmap. Any pte operations must have already been completed.
2423 pv_free(pv_entry_t pv)
2427 KKASSERT(pv->pv_m == NULL);
2428 if ((pmap = pv->pv_pmap) != NULL) {
2429 spin_lock(&pmap->pm_spin);
2430 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2431 if (pmap->pm_pvhint == pv)
2432 pmap->pm_pvhint = NULL;
2433 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2436 spin_unlock(&pmap->pm_spin);
2442 * This routine is very drastic, but can save the system
2450 static int warningdone=0;
2452 if (pmap_pagedaemon_waken == 0)
2454 pmap_pagedaemon_waken = 0;
2455 if (warningdone < 5) {
2456 kprintf("pmap_collect: collecting pv entries -- "
2457 "suggest increasing PMAP_SHPGPERPROC\n");
2461 for (i = 0; i < vm_page_array_size; i++) {
2462 m = &vm_page_array[i];
2463 if (m->wire_count || m->hold_count)
2465 if (vm_page_busy_try(m, TRUE) == 0) {
2466 if (m->wire_count == 0 && m->hold_count == 0) {
2475 * Scan the pmap for active page table entries and issue a callback.
2476 * The callback must dispose of pte_pv.
2478 * NOTE: Unmanaged page table entries will not have a pte_pv
2480 * NOTE: Kernel page table entries will not have a pt_pv. That is, wiring
2481 * counts are not tracked in kernel page table pages.
2483 * It is assumed that the start and end are properly rounded to the page size.
2486 pmap_scan(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva,
2487 void (*func)(pmap_t, struct pmap_inval_info *,
2488 pv_entry_t, pv_entry_t, vm_offset_t,
2489 pt_entry_t *, void *),
2492 pv_entry_t pdp_pv; /* A page directory page PV */
2493 pv_entry_t pd_pv; /* A page directory PV */
2494 pv_entry_t pt_pv; /* A page table PV */
2495 pv_entry_t pte_pv; /* A page table entry PV */
2497 vm_offset_t va_next;
2498 struct pmap_inval_info info;
2505 * Hold the token for stability; if the pmap is empty we have nothing
2508 lwkt_gettoken(&pmap->pm_token);
2510 if (pmap->pm_stats.resident_count == 0) {
2511 lwkt_reltoken(&pmap->pm_token);
2516 pmap_inval_init(&info);
2519 * Special handling for removing one page, which is a very common
2520 * operation (it is?).
2521 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
2523 if (sva + PAGE_SIZE == eva) {
2524 if (sva >= VM_MAX_USER_ADDRESS) {
2526 * Kernel mappings do not track wire counts on
2530 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2534 * User mappings may or may not have a pte_pv but
2535 * will always have a pt_pv if the page is present.
2537 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2538 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2539 if (pt_pv == NULL) {
2540 KKASSERT(pte_pv == NULL);
2543 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
2547 * Unlike the pv_find() case below we actually
2548 * acquired a locked pv in this case so any
2549 * race should have been resolved. It is expected
2552 KKASSERT(pte_pv == NULL);
2553 } else if (pte_pv) {
2554 KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED|
2556 ("bad *ptep %016lx sva %016lx pte_pv %p",
2557 *ptep, sva, pte_pv));
2558 func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg);
2560 KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V,
2561 ("bad *ptep %016lx sva %016lx pte_pv NULL",
2563 func(pmap, &info, pte_pv, pt_pv, sva, ptep, arg);
2568 pmap_inval_done(&info);
2569 lwkt_reltoken(&pmap->pm_token);
2574 * NOTE: kernel mappings do not track page table pages, only
2577 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
2578 * However, for the scan to be efficient we try to
2579 * cache items top-down.
2585 for (; sva < eva; sva = va_next) {
2587 if (sva >= VM_MAX_USER_ADDRESS) {
2598 if (pdp_pv == NULL) {
2599 pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva));
2600 } else if (pdp_pv->pv_pindex != pmap_pdp_pindex(sva)) {
2602 pdp_pv = pv_get(pmap, pmap_pdp_pindex(sva));
2604 if (pdp_pv == NULL) {
2605 va_next = (sva + NBPML4) & ~PML4MASK;
2614 if (pd_pv == NULL) {
2619 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
2620 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
2626 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
2628 if (pd_pv == NULL) {
2629 va_next = (sva + NBPDP) & ~PDPMASK;
2638 if (pt_pv == NULL) {
2647 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2648 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
2658 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
2662 * We will scan or skip a page table page so adjust va_next
2665 if (pt_pv == NULL) {
2666 va_next = (sva + NBPDR) & ~PDRMASK;
2673 * From this point in the loop testing pt_pv for non-NULL
2674 * means we are in UVM, else if it is NULL we are in KVM.
2677 va_next = (sva + NBPDR) & ~PDRMASK;
2682 * Limit our scan to either the end of the va represented
2683 * by the current page table page, or to the end of the
2684 * range being removed.
2686 * Scan the page table for pages. Some pages may not be
2687 * managed (might not have a pv_entry).
2689 * There is no page table management for kernel pages so
2690 * pt_pv will be NULL in that case, but otherwise pt_pv
2691 * is non-NULL, locked, and referenced.
2697 * At this point a non-NULL pt_pv means a UVA, and a NULL
2698 * pt_pv means a KVA.
2701 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
2705 while (sva < va_next) {
2707 * Acquire the related pte_pv, if any. If *ptep == 0
2708 * the related pte_pv should not exist, but if *ptep
2709 * is not zero the pte_pv may or may not exist (e.g.
2710 * will not exist for an unmanaged page).
2712 * However a multitude of races are possible here.
2714 * In addition, the (pt_pv, pte_pv) lock order is
2715 * backwards, so we have to be careful in aquiring
2716 * a properly locked pte_pv.
2720 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
2731 pv_put(pt_pv); /* must be non-NULL */
2733 pv_lock(pte_pv); /* safe to block now */
2736 pt_pv = pv_get(pmap,
2737 pmap_pt_pindex(sva));
2741 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
2745 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
2749 kprintf("Unexpected non-NULL pte_pv "
2750 "%p pt_pv %p *ptep = %016lx\n",
2751 pte_pv, pt_pv, *ptep);
2752 panic("Unexpected non-NULL pte_pv");
2760 * Ready for the callback. The locked pte_pv (if any)
2761 * is consumed by the callback. pte_pv will exist if
2762 * the page is managed, and will not exist if it
2766 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
2768 ("bad *ptep %016lx sva %016lx "
2770 *ptep, sva, pte_pv));
2771 func(pmap, &info, pte_pv, pt_pv, sva,
2774 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
2776 ("bad *ptep %016lx sva %016lx "
2779 func(pmap, &info, pte_pv, pt_pv, sva,
2799 pmap_inval_done(&info);
2800 lwkt_reltoken(&pmap->pm_token);
2804 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2806 pmap_scan(pmap, sva, eva, pmap_remove_callback, NULL);
2810 pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
2811 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
2812 pt_entry_t *ptep, void *arg __unused)
2818 * This will also drop pt_pv's wire_count. Note that
2819 * terminal pages are not wired based on mmu presence.
2821 pmap_remove_pv_pte(pte_pv, pt_pv, info);
2822 pmap_remove_pv_page(pte_pv);
2826 * pt_pv's wire_count is still bumped by unmanaged pages
2827 * so we must decrement it manually.
2829 pmap_inval_interlock(info, pmap, va);
2830 pte = pte_load_clear(ptep);
2831 pmap_inval_deinterlock(info, pmap);
2833 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2834 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2835 if (pt_pv && vm_page_unwire_quick(pt_pv->pv_m))
2836 panic("pmap_remove: insufficient wirecount");
2841 * Removes this physical page from all physical maps in which it resides.
2842 * Reflects back modify bits to the pager.
2844 * This routine may not be called from an interrupt.
2848 pmap_remove_all(vm_page_t m)
2850 struct pmap_inval_info info;
2853 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2856 pmap_inval_init(&info);
2857 vm_page_spin_lock(m);
2858 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2859 KKASSERT(pv->pv_m == m);
2860 if (pv_hold_try(pv)) {
2861 vm_page_spin_unlock(m);
2863 vm_page_spin_unlock(m);
2865 if (pv->pv_m != m) {
2867 vm_page_spin_lock(m);
2872 * Holding no spinlocks, pv is locked.
2874 pmap_remove_pv_pte(pv, NULL, &info);
2875 pmap_remove_pv_page(pv);
2877 vm_page_spin_lock(m);
2879 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2880 vm_page_spin_unlock(m);
2881 pmap_inval_done(&info);
2887 * Set the physical protection on the specified range of this map
2890 * This function may not be called from an interrupt if the map is
2891 * not the kernel_pmap.
2894 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2896 /* JG review for NX */
2900 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2901 pmap_remove(pmap, sva, eva);
2904 if (prot & VM_PROT_WRITE)
2906 pmap_scan(pmap, sva, eva, pmap_protect_callback, &prot);
2911 pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
2912 pv_entry_t pte_pv, pv_entry_t pt_pv, vm_offset_t va,
2913 pt_entry_t *ptep, void *arg __unused)
2922 pmap_inval_interlock(info, pmap, va);
2929 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2930 KKASSERT(m == pte_pv->pv_m);
2931 vm_page_flag_set(m, PG_REFERENCED);
2935 if (pmap_track_modified(pte_pv->pv_pindex)) {
2937 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2944 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
2947 pmap_inval_deinterlock(info, pmap);
2953 * Insert the vm_page (m) at the virtual address (va), replacing any prior
2954 * mapping at that address. Set protection and wiring as requested.
2956 * NOTE: This routine MUST insert the page into the pmap now, it cannot
2960 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2963 pmap_inval_info info;
2964 pv_entry_t pt_pv; /* page table */
2965 pv_entry_t pte_pv; /* page table entry */
2968 pt_entry_t origpte, newpte;
2973 va = trunc_page(va);
2974 #ifdef PMAP_DIAGNOSTIC
2976 panic("pmap_enter: toobig");
2977 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2978 panic("pmap_enter: invalid to pmap_enter page table "
2979 "pages (va: 0x%lx)", va);
2981 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2982 kprintf("Warning: pmap_enter called on UVA with "
2985 db_print_backtrace();
2988 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2989 kprintf("Warning: pmap_enter called on KVA without"
2992 db_print_backtrace();
2997 * Get locked PV entries for our new page table entry (pte_pv)
2998 * and for its parent page table (pt_pv). We need the parent
2999 * so we can resolve the location of the ptep.
3001 * Only hardware MMU actions can modify the ptep out from
3004 * if (m) is fictitious or unmanaged we do not create a managing
3005 * pte_pv for it. Any pre-existing page's management state must
3006 * match (avoiding code complexity).
3008 * If the pmap is still being initialized we assume existing
3011 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3012 * pmap_allocpte() checks the
3014 if (pmap_initialized == FALSE) {
3018 } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) {
3020 if (va >= VM_MAX_USER_ADDRESS) {
3024 pt_pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
3025 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3027 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0);
3029 if (va >= VM_MAX_USER_ADDRESS) {
3031 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3034 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va),
3036 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3038 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED));
3041 pa = VM_PAGE_TO_PHYS(m);
3043 opa = origpte & PG_FRAME;
3045 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V | PG_A);
3048 if (va < VM_MAX_USER_ADDRESS)
3051 newpte |= PG_MANAGED;
3052 if (pmap == &kernel_pmap)
3056 * It is possible for multiple faults to occur in threaded
3057 * environments, the existing pte might be correct.
3059 if (((origpte ^ newpte) & ~(pt_entry_t)(PG_M|PG_A)) == 0)
3062 if ((prot & VM_PROT_NOSYNC) == 0)
3063 pmap_inval_init(&info);
3066 * Ok, either the address changed or the protection or wiring
3069 * Clear the current entry, interlocking the removal. For managed
3070 * pte's this will also flush the modified state to the vm_page.
3071 * Atomic ops are mandatory in order to ensure that PG_M events are
3072 * not lost during any transition.
3077 * pmap_remove_pv_pte() unwires pt_pv and assumes
3078 * we will free pte_pv, but since we are reusing
3079 * pte_pv we want to retain the wire count.
3081 * pt_pv won't exist for a kernel page (managed or
3085 vm_page_wire_quick(pt_pv->pv_m);
3086 if (prot & VM_PROT_NOSYNC)
3087 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3089 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3091 pmap_remove_pv_page(pte_pv);
3092 } else if (prot & VM_PROT_NOSYNC) {
3093 /* leave wire count on PT page intact */
3094 (void)pte_load_clear(ptep);
3095 cpu_invlpg((void *)va);
3096 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3098 /* leave wire count on PT page intact */
3099 pmap_inval_interlock(&info, pmap, va);
3100 (void)pte_load_clear(ptep);
3101 pmap_inval_deinterlock(&info, pmap);
3102 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3104 KKASSERT(*ptep == 0);
3109 * Enter on the PV list if part of our managed memory.
3110 * Wiring of the PT page is already handled.
3112 KKASSERT(pte_pv->pv_m == NULL);
3113 vm_page_spin_lock(m);
3115 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3118 atomic_add_int(&m->object->agg_pv_list_count, 1);
3120 vm_page_flag_set(m, PG_MAPPED);
3121 vm_page_spin_unlock(m);
3122 } else if (pt_pv && opa == 0) {
3124 * We have to adjust the wire count on the PT page ourselves
3125 * for unmanaged entries. If opa was non-zero we retained
3126 * the existing wire count from the removal.
3128 vm_page_wire_quick(pt_pv->pv_m);
3132 * Ok, for UVM (pt_pv != NULL) we don't need to interlock or
3133 * invalidate anything, the TLB won't have any stale entries to
3136 * For KVM there appear to still be issues. Theoretically we
3137 * should be able to scrap the interlocks entirely but we
3140 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3141 pmap_inval_interlock(&info, pmap, va);
3142 *(volatile pt_entry_t *)ptep = newpte;
3144 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3145 pmap_inval_deinterlock(&info, pmap);
3146 else if (pt_pv == NULL)
3147 cpu_invlpg((void *)va);
3150 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3152 vm_page_flag_set(m, PG_WRITEABLE);
3154 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3159 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
3160 pmap_inval_done(&info);
3162 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
3165 * Cleanup the pv entry, allowing other accessors.
3174 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
3175 * This code also assumes that the pmap has no pre-existing entry for this
3178 * This code currently may only be used on user pmaps, not kernel_pmap.
3181 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
3183 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE);
3187 * Make a temporary mapping for a physical address. This is only intended
3188 * to be used for panic dumps.
3190 * The caller is responsible for calling smp_invltlb().
3193 pmap_kenter_temporary(vm_paddr_t pa, long i)
3195 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
3196 return ((void *)crashdumpmap);
3199 #define MAX_INIT_PT (96)
3202 * This routine preloads the ptes for a given object into the specified pmap.
3203 * This eliminates the blast of soft faults on process startup and
3204 * immediately after an mmap.
3206 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
3209 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
3210 vm_object_t object, vm_pindex_t pindex,
3211 vm_size_t size, int limit)
3213 struct rb_vm_page_scan_info info;
3218 * We can't preinit if read access isn't set or there is no pmap
3221 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
3225 * We can't preinit if the pmap is not the current pmap
3227 lp = curthread->td_lwp;
3228 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
3231 psize = x86_64_btop(size);
3233 if ((object->type != OBJT_VNODE) ||
3234 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
3235 (object->resident_page_count > MAX_INIT_PT))) {
3239 if (pindex + psize > object->size) {
3240 if (object->size < pindex)
3242 psize = object->size - pindex;
3249 * Use a red-black scan to traverse the requested range and load
3250 * any valid pages found into the pmap.
3252 * We cannot safely scan the object's memq without holding the
3255 info.start_pindex = pindex;
3256 info.end_pindex = pindex + psize - 1;
3262 vm_object_hold_shared(object);
3263 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3264 pmap_object_init_pt_callback, &info);
3265 vm_object_drop(object);
3270 pmap_object_init_pt_callback(vm_page_t p, void *data)
3272 struct rb_vm_page_scan_info *info = data;
3273 vm_pindex_t rel_index;
3276 * don't allow an madvise to blow away our really
3277 * free pages allocating pv entries.
3279 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3280 vmstats.v_free_count < vmstats.v_free_reserved) {
3285 * Ignore list markers and ignore pages we cannot instantly
3286 * busy (while holding the object token).
3288 if (p->flags & PG_MARKER)
3290 if (vm_page_busy_try(p, TRUE))
3292 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3293 (p->flags & PG_FICTITIOUS) == 0) {
3294 if ((p->queue - p->pc) == PQ_CACHE)
3295 vm_page_deactivate(p);
3296 rel_index = p->pindex - info->start_pindex;
3297 pmap_enter_quick(info->pmap,
3298 info->addr + x86_64_ptob(rel_index), p);
3306 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
3309 * Returns FALSE if it would be non-trivial or if a pte is already loaded
3312 * XXX This is safe only because page table pages are not freed.
3315 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3319 /*spin_lock(&pmap->pm_spin);*/
3320 if ((pte = pmap_pte(pmap, addr)) != NULL) {
3322 /*spin_unlock(&pmap->pm_spin);*/
3326 /*spin_unlock(&pmap->pm_spin);*/
3331 * Change the wiring attribute for a pmap/va pair. The mapping must already
3332 * exist in the pmap. The mapping may or may not be managed.
3335 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
3342 lwkt_gettoken(&pmap->pm_token);
3343 pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
3344 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
3346 if (wired && !pmap_pte_w(ptep))
3347 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3348 else if (!wired && pmap_pte_w(ptep))
3349 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3352 * Wiring is not a hardware characteristic so there is no need to
3353 * invalidate TLB. However, in an SMP environment we must use
3354 * a locked bus cycle to update the pte (if we are not using
3355 * the pmap_inval_*() API that is)... it's ok to do this for simple
3360 atomic_set_long(ptep, PG_W);
3362 atomic_clear_long(ptep, PG_W);
3365 atomic_set_long_nonlocked(ptep, PG_W);
3367 atomic_clear_long_nonlocked(ptep, PG_W);
3370 lwkt_reltoken(&pmap->pm_token);
3376 * Copy the range specified by src_addr/len from the source map to
3377 * the range dst_addr/len in the destination map.
3379 * This routine is only advisory and need not do anything.
3382 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3383 vm_size_t len, vm_offset_t src_addr)
3390 * Zero the specified physical page.
3392 * This function may be called from an interrupt and no locking is
3396 pmap_zero_page(vm_paddr_t phys)
3398 vm_offset_t va = PHYS_TO_DMAP(phys);
3400 pagezero((void *)va);
3404 * pmap_page_assertzero:
3406 * Assert that a page is empty, panic if it isn't.
3409 pmap_page_assertzero(vm_paddr_t phys)
3411 vm_offset_t va = PHYS_TO_DMAP(phys);
3414 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3415 if (*(long *)((char *)va + i) != 0) {
3416 panic("pmap_page_assertzero() @ %p not zero!\n",
3417 (void *)(intptr_t)va);
3425 * Zero part of a physical page by mapping it into memory and clearing
3426 * its contents with bzero.
3428 * off and size may not cover an area beyond a single hardware page.
3431 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3433 vm_offset_t virt = PHYS_TO_DMAP(phys);
3435 bzero((char *)virt + off, size);
3441 * Copy the physical page from the source PA to the target PA.
3442 * This function may be called from an interrupt. No locking
3446 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3448 vm_offset_t src_virt, dst_virt;
3450 src_virt = PHYS_TO_DMAP(src);
3451 dst_virt = PHYS_TO_DMAP(dst);
3452 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3456 * pmap_copy_page_frag:
3458 * Copy the physical page from the source PA to the target PA.
3459 * This function may be called from an interrupt. No locking
3463 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3465 vm_offset_t src_virt, dst_virt;
3467 src_virt = PHYS_TO_DMAP(src);
3468 dst_virt = PHYS_TO_DMAP(dst);
3470 bcopy((char *)src_virt + (src & PAGE_MASK),
3471 (char *)dst_virt + (dst & PAGE_MASK),
3476 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
3477 * this page. This count may be changed upwards or downwards in the future;
3478 * it is only necessary that true be returned for a small subset of pmaps
3479 * for proper page aging.
3482 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3487 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3490 vm_page_spin_lock(m);
3491 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3492 if (pv->pv_pmap == pmap) {
3493 vm_page_spin_unlock(m);
3500 vm_page_spin_unlock(m);
3505 * Remove all pages from specified address space this aids process exit
3506 * speeds. Also, this code may be special cased for the current process
3510 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3512 pmap_remove(pmap, sva, eva);
3516 * pmap_testbit tests bits in pte's note that the testbit/clearbit
3517 * routines are inline, and a lot of things compile-time evaluate.
3521 pmap_testbit(vm_page_t m, int bit)
3526 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3529 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3531 vm_page_spin_lock(m);
3532 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
3533 vm_page_spin_unlock(m);
3537 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3539 * if the bit being tested is the modified bit, then
3540 * mark clean_map and ptes as never
3543 if (bit & (PG_A|PG_M)) {
3544 if (!pmap_track_modified(pv->pv_pindex))
3548 #if defined(PMAP_DIAGNOSTIC)
3549 if (pv->pv_pmap == NULL) {
3550 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
3555 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3557 vm_page_spin_unlock(m);
3561 vm_page_spin_unlock(m);
3566 * This routine is used to modify bits in ptes. Only one bit should be
3567 * specified. PG_RW requires special handling.
3569 * Caller must NOT hold any spin locks
3573 pmap_clearbit(vm_page_t m, int bit)
3575 struct pmap_inval_info info;
3582 vm_page_flag_clear(m, PG_WRITEABLE);
3583 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
3590 * Loop over all current mappings setting/clearing as appropos If
3591 * setting RO do we need to clear the VAC?
3593 * NOTE: When clearing PG_M we could also (not implemented) drop
3594 * through to the PG_RW code and clear PG_RW too, forcing
3595 * a fault on write to redetect PG_M for virtual kernels, but
3596 * it isn't necessary since virtual kernels invalidate the
3597 * pte when they clear the VPTE_M bit in their virtual page
3600 * NOTE: Does not re-dirty the page when clearing only PG_M.
3602 if ((bit & PG_RW) == 0) {
3603 vm_page_spin_lock(m);
3604 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3605 #if defined(PMAP_DIAGNOSTIC)
3606 if (pv->pv_pmap == NULL) {
3607 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
3612 pte = pmap_pte_quick(pv->pv_pmap,
3613 pv->pv_pindex << PAGE_SHIFT);
3616 atomic_clear_long(pte, bit);
3618 vm_page_spin_unlock(m);
3623 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
3626 pmap_inval_init(&info);
3629 vm_page_spin_lock(m);
3630 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3632 * don't write protect pager mappings
3634 if (!pmap_track_modified(pv->pv_pindex))
3637 #if defined(PMAP_DIAGNOSTIC)
3638 if (pv->pv_pmap == NULL) {
3639 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
3645 * Skip pages which do not have PG_RW set.
3647 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3648 if ((*pte & PG_RW) == 0)
3654 if (pv_hold_try(pv) == 0) {
3655 vm_page_spin_unlock(m);
3656 pv_lock(pv); /* held, now do a blocking lock */
3657 pv_put(pv); /* and release */
3658 goto restart; /* anything could have happened */
3661 save_pmap = pv->pv_pmap;
3662 vm_page_spin_unlock(m);
3663 pmap_inval_interlock(&info, save_pmap,
3664 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
3665 KKASSERT(pv->pv_pmap == save_pmap);
3669 if (atomic_cmpset_long(pte, pbits,
3670 pbits & ~(PG_RW|PG_M))) {
3674 pmap_inval_deinterlock(&info, save_pmap);
3675 vm_page_spin_lock(m);
3678 * If PG_M was found to be set while we were clearing PG_RW
3679 * we also clear PG_M (done above) and mark the page dirty.
3680 * Callers expect this behavior.
3686 vm_page_spin_unlock(m);
3687 pmap_inval_done(&info);
3691 * Lower the permission for all mappings to a given page.
3693 * Page must be busied by caller.
3696 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3698 /* JG NX support? */
3699 if ((prot & VM_PROT_WRITE) == 0) {
3700 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3702 * NOTE: pmap_clearbit(.. PG_RW) also clears
3703 * the PG_WRITEABLE flag in (m).
3705 pmap_clearbit(m, PG_RW);
3713 pmap_phys_address(vm_pindex_t ppn)
3715 return (x86_64_ptob(ppn));
3719 * Return a count of reference bits for a page, clearing those bits.
3720 * It is not necessary for every reference bit to be cleared, but it
3721 * is necessary that 0 only be returned when there are truly no
3722 * reference bits set.
3724 * XXX: The exact number of bits to check and clear is a matter that
3725 * should be tested and standardized at some point in the future for
3726 * optimal aging of shared pages.
3728 * This routine may not block.
3731 pmap_ts_referenced(vm_page_t m)
3737 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3740 vm_page_spin_lock(m);
3741 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3742 if (!pmap_track_modified(pv->pv_pindex))
3744 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
3745 if (pte && (*pte & PG_A)) {
3747 atomic_clear_long(pte, PG_A);
3749 atomic_clear_long_nonlocked(pte, PG_A);
3756 vm_page_spin_unlock(m);
3763 * Return whether or not the specified physical page was modified
3764 * in any physical maps.
3767 pmap_is_modified(vm_page_t m)
3771 res = pmap_testbit(m, PG_M);
3776 * Clear the modify bits on the specified physical page.
3779 pmap_clear_modify(vm_page_t m)
3781 pmap_clearbit(m, PG_M);
3785 * pmap_clear_reference:
3787 * Clear the reference bit on the specified physical page.
3790 pmap_clear_reference(vm_page_t m)
3792 pmap_clearbit(m, PG_A);
3796 * Miscellaneous support routines follow
3801 i386_protection_init(void)
3805 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3806 kp = protection_codes;
3807 for (prot = 0; prot < 8; prot++) {
3809 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3811 * Read access is also 0. There isn't any execute bit,
3812 * so just make it readable.
3814 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3815 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3816 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3819 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3820 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3821 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3822 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3830 * Map a set of physical memory pages into the kernel virtual
3831 * address space. Return a pointer to where it is mapped. This
3832 * routine is intended to be used for mapping device memory,
3835 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3839 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3841 vm_offset_t va, tmpva, offset;
3844 offset = pa & PAGE_MASK;
3845 size = roundup(offset + size, PAGE_SIZE);
3847 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3849 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3851 pa = pa & ~PAGE_MASK;
3852 for (tmpva = va; size > 0;) {
3853 pte = vtopte(tmpva);
3854 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3862 return ((void *)(va + offset));
3866 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3868 vm_offset_t va, tmpva, offset;
3871 offset = pa & PAGE_MASK;
3872 size = roundup(offset + size, PAGE_SIZE);
3874 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3876 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3878 pa = pa & ~PAGE_MASK;
3879 for (tmpva = va; size > 0;) {
3880 pte = vtopte(tmpva);
3881 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3889 return ((void *)(va + offset));
3893 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3895 vm_offset_t base, offset;
3897 base = va & ~PAGE_MASK;
3898 offset = va & PAGE_MASK;
3899 size = roundup(offset + size, PAGE_SIZE);
3900 pmap_qremove(va, size >> PAGE_SHIFT);
3901 kmem_free(&kernel_map, base, size);
3905 * perform the pmap work for mincore
3908 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3910 pt_entry_t *ptep, pte;
3914 lwkt_gettoken(&pmap->pm_token);
3915 ptep = pmap_pte(pmap, addr);
3917 if (ptep && (pte = *ptep) != 0) {
3920 val = MINCORE_INCORE;
3921 if ((pte & PG_MANAGED) == 0)
3924 pa = pte & PG_FRAME;
3926 m = PHYS_TO_VM_PAGE(pa);
3932 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3934 * Modified by someone
3936 else if (m->dirty || pmap_is_modified(m))
3937 val |= MINCORE_MODIFIED_OTHER;
3942 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3945 * Referenced by someone
3947 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3948 val |= MINCORE_REFERENCED_OTHER;
3949 vm_page_flag_set(m, PG_REFERENCED);
3953 lwkt_reltoken(&pmap->pm_token);
3959 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3960 * vmspace will be ref'd and the old one will be deref'd.
3962 * The vmspace for all lwps associated with the process will be adjusted
3963 * and cr3 will be reloaded if any lwp is the current lwp.
3965 * The process must hold the vmspace->vm_map.token for oldvm and newvm
3968 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3970 struct vmspace *oldvm;
3973 oldvm = p->p_vmspace;
3974 if (oldvm != newvm) {
3976 sysref_get(&newvm->vm_sysref);
3977 p->p_vmspace = newvm;
3978 KKASSERT(p->p_nthreads == 1);
3979 lp = RB_ROOT(&p->p_lwp_tree);
3980 pmap_setlwpvm(lp, newvm);
3982 sysref_put(&oldvm->vm_sysref);
3987 * Set the vmspace for a LWP. The vmspace is almost universally set the
3988 * same as the process vmspace, but virtual kernels need to swap out contexts
3989 * on a per-lwp basis.
3991 * Caller does not necessarily hold any vmspace tokens. Caller must control
3992 * the lwp (typically be in the context of the lwp). We use a critical
3993 * section to protect against statclock and hardclock (statistics collection).
3996 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3998 struct vmspace *oldvm;
4001 oldvm = lp->lwp_vmspace;
4003 if (oldvm != newvm) {
4005 lp->lwp_vmspace = newvm;
4006 if (curthread->td_lwp == lp) {
4007 pmap = vmspace_pmap(newvm);
4009 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4010 if (pmap->pm_active & CPUMASK_LOCK)
4011 pmap_interlock_wait(newvm);
4013 pmap->pm_active |= 1;
4015 #if defined(SWTCH_OPTIM_STATS)
4018 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4019 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
4020 load_cr3(curthread->td_pcb->pcb_cr3);
4021 pmap = vmspace_pmap(oldvm);
4023 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4025 pmap->pm_active &= ~(cpumask_t)1;
4035 * Called when switching to a locked pmap, used to interlock against pmaps
4036 * undergoing modifications to prevent us from activating the MMU for the
4037 * target pmap until all such modifications have completed. We have to do
4038 * this because the thread making the modifications has already set up its
4039 * SMP synchronization mask.
4041 * This function cannot sleep!
4046 pmap_interlock_wait(struct vmspace *vm)
4048 struct pmap *pmap = &vm->vm_pmap;
4050 if (pmap->pm_active & CPUMASK_LOCK) {
4052 KKASSERT(curthread->td_critcount >= 2);
4053 DEBUG_PUSH_INFO("pmap_interlock_wait");
4054 while (pmap->pm_active & CPUMASK_LOCK) {
4056 lwkt_process_ipiq();
4066 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4069 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4073 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4078 * Used by kmalloc/kfree, page already exists at va
4081 pmap_kvtom(vm_offset_t va)
4083 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));