4 * Copyright (c) 1991 Regents of the University of California.
5 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
7 * Copyright (c) 2003 Peter Wemm
8 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
9 * Copyright (c) 2008, 2009 The DragonFly Project.
10 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * the Systems Programming Group of the University of Utah Computer
15 * Science Department and William Jolitz of UUNET Technologies Inc.
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
25 * 3. All advertising materials mentioning features or use of this software
26 * must display the following acknowledgement:
27 * This product includes software developed by the University of
28 * California, Berkeley and its contributors.
29 * 4. Neither the name of the University nor the names of its contributors
30 * may be used to endorse or promote products derived from this software
31 * without specific prior written permission.
33 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
34 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
35 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
36 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
37 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
38 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
39 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
41 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
42 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
46 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
50 * Manages physical address maps.
52 * In addition to hardware address maps, this
53 * module is called upon to provide software-use-only
54 * maps which may or may not be stored in the same
55 * form as hardware maps. These pseudo-maps are
56 * used to store intermediate results from copy
57 * operations to and from address spaces.
59 * Since the information managed by this module is
60 * also stored by the logical address mapping module,
61 * this module may throw away valid virtual-to-physical
62 * mappings at almost any time. However, invalidations
63 * of virtual-to-physical mappings must be done as
66 * In order to cope with hardware architectures which
67 * make virtual-to-physical map invalidates expensive,
68 * this module may delay invalidate or reduced protection
69 * operations until such time as they are actually
70 * necessary. This module is given full information as
71 * to which processors are currently using which maps,
72 * and to when physical maps must be made correct.
76 #include "opt_disable_pse.h"
79 #include "opt_msgbuf.h"
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/kernel.h>
85 #include <sys/msgbuf.h>
86 #include <sys/vmmeter.h>
90 #include <vm/vm_param.h>
91 #include <sys/sysctl.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_zone.h>
102 #include <sys/user.h>
103 #include <sys/thread2.h>
104 #include <sys/sysref2.h>
105 #include <sys/spinlock2.h>
106 #include <vm/vm_page2.h>
108 #include <machine/cputypes.h>
109 #include <machine/md_var.h>
110 #include <machine/specialreg.h>
111 #include <machine/smp.h>
112 #include <machine_base/apic/apicreg.h>
113 #include <machine/globaldata.h>
114 #include <machine/pmap.h>
115 #include <machine/pmap_inval.h>
119 #define PMAP_KEEP_PDIRS
120 #ifndef PMAP_SHPGPERPROC
121 #define PMAP_SHPGPERPROC 200
124 #if defined(DIAGNOSTIC)
125 #define PMAP_DIAGNOSTIC
131 * Get PDEs and PTEs for user/kernel address space
133 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
134 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
136 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
137 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
138 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
139 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
140 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
144 * Given a map and a machine independent protection code,
145 * convert to a vax protection code.
147 #define pte_prot(m, p) \
148 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
149 static int protection_codes[8];
151 struct pmap kernel_pmap;
152 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
154 vm_paddr_t avail_start; /* PA of first available physical page */
155 vm_paddr_t avail_end; /* PA of last available physical page */
156 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
157 vm_offset_t virtual2_end;
158 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
159 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
160 vm_offset_t KvaStart; /* VA start of KVA space */
161 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
162 vm_offset_t KvaSize; /* max size of kernel virtual address space */
163 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
164 static int pgeflag; /* PG_G or-in */
165 static int pseflag; /* PG_PS or-in */
167 static vm_object_t kptobj;
170 static vm_paddr_t dmaplimit;
172 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
174 static uint64_t KPTbase;
175 static uint64_t KPTphys;
176 static uint64_t KPDphys; /* phys addr of kernel level 2 */
177 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
178 uint64_t KPDPphys; /* phys addr of kernel level 3 */
179 uint64_t KPML4phys; /* phys addr of kernel level 4 */
181 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
182 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
185 * Data for the pv entry allocation mechanism
187 static vm_zone_t pvzone;
188 static struct vm_zone pvzone_store;
189 static struct vm_object pvzone_obj;
190 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
191 static int pmap_pagedaemon_waken = 0;
192 static struct pv_entry *pvinit;
195 * All those kernel PT submaps that BSD is so fond of
197 pt_entry_t *CMAP1 = 0, *ptmmap;
198 caddr_t CADDR1 = 0, ptvmmap = 0;
199 static pt_entry_t *msgbufmap;
200 struct msgbuf *msgbufp=0;
205 static pt_entry_t *pt_crashdumpmap;
206 static caddr_t crashdumpmap;
208 static int pmap_yield_count = 64;
209 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
210 &pmap_yield_count, 0, "Yield during init_pt/release");
214 static pv_entry_t get_pv_entry (void);
215 static void i386_protection_init (void);
216 static void create_pagetables(vm_paddr_t *firstaddr);
217 static void pmap_remove_all (vm_page_t m);
218 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
219 vm_offset_t sva, pmap_inval_info_t info);
220 static void pmap_remove_page (struct pmap *pmap,
221 vm_offset_t va, pmap_inval_info_t info);
222 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
223 vm_offset_t va, pmap_inval_info_t info);
224 static boolean_t pmap_testbit (vm_page_t m, int bit);
225 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
226 vm_page_t mpte, vm_page_t m);
228 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
230 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
231 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
232 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
233 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
234 static int pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
235 pmap_inval_info_t info);
236 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
238 static unsigned pdir4mb;
241 * Move the kernel virtual free pointer to the next
242 * 2MB. This is used to help improve performance
243 * by using a large (2MB) page for much of the kernel
244 * (.text, .data, .bss)
248 pmap_kmem_choose(vm_offset_t addr)
250 vm_offset_t newaddr = addr;
252 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
259 * Super fast pmap_pte routine best used when scanning the pv lists.
260 * This eliminates many course-grained invltlb calls. Note that many of
261 * the pv list scans are across different pmaps and it is very wasteful
262 * to do an entire invltlb when checking a single mapping.
264 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
268 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
270 return pmap_pte(pmap, va);
273 /* Return a non-clipped PD index for a given VA */
276 pmap_pde_pindex(vm_offset_t va)
278 return va >> PDRSHIFT;
281 /* Return various clipped indexes for a given VA */
284 pmap_pte_index(vm_offset_t va)
287 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
292 pmap_pde_index(vm_offset_t va)
295 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
300 pmap_pdpe_index(vm_offset_t va)
303 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
308 pmap_pml4e_index(vm_offset_t va)
311 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
314 /* Return a pointer to the PML4 slot that corresponds to a VA */
317 pmap_pml4e(pmap_t pmap, vm_offset_t va)
320 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
323 /* Return a pointer to the PDP slot that corresponds to a VA */
326 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
330 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
331 return (&pdpe[pmap_pdpe_index(va)]);
334 /* Return a pointer to the PDP slot that corresponds to a VA */
337 pmap_pdpe(pmap_t pmap, vm_offset_t va)
341 pml4e = pmap_pml4e(pmap, va);
342 if ((*pml4e & PG_V) == 0)
344 return (pmap_pml4e_to_pdpe(pml4e, va));
347 /* Return a pointer to the PD slot that corresponds to a VA */
350 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
354 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
355 return (&pde[pmap_pde_index(va)]);
358 /* Return a pointer to the PD slot that corresponds to a VA */
361 pmap_pde(pmap_t pmap, vm_offset_t va)
365 pdpe = pmap_pdpe(pmap, va);
366 if (pdpe == NULL || (*pdpe & PG_V) == 0)
368 return (pmap_pdpe_to_pde(pdpe, va));
371 /* Return a pointer to the PT slot that corresponds to a VA */
374 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
378 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
379 return (&pte[pmap_pte_index(va)]);
382 /* Return a pointer to the PT slot that corresponds to a VA */
385 pmap_pte(pmap_t pmap, vm_offset_t va)
389 pde = pmap_pde(pmap, va);
390 if (pde == NULL || (*pde & PG_V) == 0)
392 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
393 return ((pt_entry_t *)pde);
394 return (pmap_pde_to_pte(pde, va));
399 vtopte(vm_offset_t va)
401 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
402 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
404 return (PTmap + ((va >> PAGE_SHIFT) & mask));
409 vtopde(vm_offset_t va)
411 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
412 NPML4EPGSHIFT)) - 1);
414 return (PDmap + ((va >> PDRSHIFT) & mask));
418 allocpages(vm_paddr_t *firstaddr, long n)
423 bzero((void *)ret, n * PAGE_SIZE);
424 *firstaddr += n * PAGE_SIZE;
430 create_pagetables(vm_paddr_t *firstaddr)
432 long i; /* must be 64 bits */
437 * We are running (mostly) V=P at this point
439 * Calculate NKPT - number of kernel page tables. We have to
440 * accomodoate prealloction of the vm_page_array, dump bitmap,
441 * MSGBUF_SIZE, and other stuff. Be generous.
443 * Maxmem is in pages.
445 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
446 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
450 * Starting at the beginning of kvm (not KERNBASE).
452 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
453 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
454 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + ndmpdp) +
459 * Starting at KERNBASE - map 2G worth of page table pages.
460 * KERNBASE is offset -2G from the end of kvm.
462 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
467 KPTbase = allocpages(firstaddr, nkpt_base);
468 KPTphys = allocpages(firstaddr, nkpt_phys);
469 KPML4phys = allocpages(firstaddr, 1);
470 KPDPphys = allocpages(firstaddr, NKPML4E);
471 KPDphys = allocpages(firstaddr, NKPDPE);
474 * Calculate the page directory base for KERNBASE,
475 * that is where we start populating the page table pages.
476 * Basically this is the end - 2.
478 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
480 DMPDPphys = allocpages(firstaddr, NDMPML4E);
481 if ((amd_feature & AMDID_PAGE1GB) == 0)
482 DMPDphys = allocpages(firstaddr, ndmpdp);
483 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
486 * Fill in the underlying page table pages for the area around
487 * KERNBASE. This remaps low physical memory to KERNBASE.
489 * Read-only from zero to physfree
490 * XXX not fully used, underneath 2M pages
492 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
493 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
494 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
498 * Now map the initial kernel page tables. One block of page
499 * tables is placed at the beginning of kernel virtual memory,
500 * and another block is placed at KERNBASE to map the kernel binary,
501 * data, bss, and initial pre-allocations.
503 for (i = 0; i < nkpt_base; i++) {
504 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
505 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
507 for (i = 0; i < nkpt_phys; i++) {
508 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
509 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
513 * Map from zero to end of allocations using 2M pages as an
514 * optimization. This will bypass some of the KPTBase pages
515 * above in the KERNBASE area.
517 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
518 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
519 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
523 * And connect up the PD to the PDP. The kernel pmap is expected
524 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
526 for (i = 0; i < NKPDPE; i++) {
527 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
528 KPDphys + (i << PAGE_SHIFT);
529 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
533 /* Now set up the direct map space using either 2MB or 1GB pages */
534 /* Preset PG_M and PG_A because demotion expects it */
535 if ((amd_feature & AMDID_PAGE1GB) == 0) {
536 for (i = 0; i < NPDEPG * ndmpdp; i++) {
537 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
538 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
541 /* And the direct map space's PDP */
542 for (i = 0; i < ndmpdp; i++) {
543 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
545 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
548 for (i = 0; i < ndmpdp; i++) {
549 ((pdp_entry_t *)DMPDPphys)[i] =
550 (vm_paddr_t)i << PDPSHIFT;
551 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
556 /* And recursively map PML4 to itself in order to get PTmap */
557 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
558 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
560 /* Connect the Direct Map slot up to the PML4 */
561 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
562 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
564 /* Connect the KVA slot up to the PML4 */
565 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
566 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
570 * Bootstrap the system enough to run with virtual memory.
572 * On the i386 this is called after mapping has already been enabled
573 * and just syncs the pmap module with what has already been done.
574 * [We can't call it easily with mapping off since the kernel is not
575 * mapped with PA == VA, hence we would have to relocate every address
576 * from the linked base (virtual) address "KERNBASE" to the actual
577 * (physical) address starting relative to 0]
580 pmap_bootstrap(vm_paddr_t *firstaddr)
584 struct mdglobaldata *gd;
587 KvaStart = VM_MIN_KERNEL_ADDRESS;
588 KvaEnd = VM_MAX_KERNEL_ADDRESS;
589 KvaSize = KvaEnd - KvaStart;
591 avail_start = *firstaddr;
594 * Create an initial set of page tables to run the kernel in.
596 create_pagetables(firstaddr);
598 virtual2_start = KvaStart;
599 virtual2_end = PTOV_OFFSET;
601 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
602 virtual_start = pmap_kmem_choose(virtual_start);
604 virtual_end = VM_MAX_KERNEL_ADDRESS;
606 /* XXX do %cr0 as well */
607 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
611 * Initialize protection array.
613 i386_protection_init();
616 * The kernel's pmap is statically allocated so we don't have to use
617 * pmap_create, which is unlikely to work correctly at this part of
618 * the boot sequence (XXX and which no longer exists).
620 * The kernel_pmap's pm_pteobj is used only for locking and not
623 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
624 kernel_pmap.pm_count = 1;
625 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
626 kernel_pmap.pm_pteobj = &kernel_object;
627 TAILQ_INIT(&kernel_pmap.pm_pvlist);
628 TAILQ_INIT(&kernel_pmap.pm_pvlist_free);
629 kernel_pmap.pm_hold = 0;
630 spin_init(&kernel_pmap.pm_spin);
631 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
634 * Reserve some special page table entries/VA space for temporary
637 #define SYSMAP(c, p, v, n) \
638 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
644 * CMAP1/CMAP2 are used for zeroing and copying pages.
646 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
651 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
654 * ptvmmap is used for reading arbitrary physical pages via
657 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
660 * msgbufp is used to map the system message buffer.
661 * XXX msgbufmap is not used.
663 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
664 atop(round_page(MSGBUF_SIZE)))
671 * PG_G is terribly broken on SMP because we IPI invltlb's in some
672 * cases rather then invl1pg. Actually, I don't even know why it
673 * works under UP because self-referential page table mappings
678 if (cpu_feature & CPUID_PGE)
683 * Initialize the 4MB page size flag
687 * The 4MB page version of the initial
688 * kernel page mapping.
692 #if !defined(DISABLE_PSE)
693 if (cpu_feature & CPUID_PSE) {
696 * Note that we have enabled PSE mode
699 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
700 ptditmp &= ~(NBPDR - 1);
701 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
706 * Enable the PSE mode. If we are SMP we can't do this
707 * now because the APs will not be able to use it when
710 load_cr4(rcr4() | CR4_PSE);
713 * We can do the mapping here for the single processor
714 * case. We simply ignore the old page table page from
718 * For SMP, we still need 4K pages to bootstrap APs,
719 * PSE will be enabled as soon as all APs are up.
721 PTD[KPTDI] = (pd_entry_t)ptditmp;
728 * We need to finish setting up the globaldata page for the BSP.
729 * locore has already populated the page table for the mdglobaldata
732 pg = MDGLOBALDATA_BASEALLOC_PAGES;
733 gd = &CPU_prvspace[0].mdglobaldata;
740 * Set 4mb pdir for mp startup
745 if (pseflag && (cpu_feature & CPUID_PSE)) {
746 load_cr4(rcr4() | CR4_PSE);
747 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
755 * Initialize the pmap module.
756 * Called by vm_init, to initialize any structures that the pmap
757 * system needs to map virtual memory.
758 * pmap_init has been enhanced to support in a fairly consistant
759 * way, discontiguous physical memory.
768 * object for kernel page table pages
770 /* JG I think the number can be arbitrary */
771 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
774 * Allocate memory for random pmap data structures. Includes the
778 for(i = 0; i < vm_page_array_size; i++) {
781 m = &vm_page_array[i];
782 TAILQ_INIT(&m->md.pv_list);
783 m->md.pv_list_count = 0;
787 * init the pv free list
789 initial_pvs = vm_page_array_size;
790 if (initial_pvs < MINPV)
792 pvzone = &pvzone_store;
793 pvinit = (void *)kmem_alloc(&kernel_map,
794 initial_pvs * sizeof (struct pv_entry));
795 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
796 pvinit, initial_pvs);
799 * Now it is safe to enable pv_table recording.
801 pmap_initialized = TRUE;
805 * Initialize the address space (zone) for the pv_entries. Set a
806 * high water mark so that the system can recover from excessive
807 * numbers of pv entries.
812 int shpgperproc = PMAP_SHPGPERPROC;
815 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
816 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
817 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
818 pv_entry_high_water = 9 * (pv_entry_max / 10);
821 * Subtract out pages already installed in the zone (hack)
823 entry_max = pv_entry_max - vm_page_array_size;
827 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
831 /***************************************************
832 * Low level helper routines.....
833 ***************************************************/
835 #if defined(PMAP_DIAGNOSTIC)
838 * This code checks for non-writeable/modified pages.
839 * This should be an invalid condition.
843 pmap_nw_modified(pt_entry_t pte)
845 if ((pte & (PG_M|PG_RW)) == PG_M)
854 * this routine defines the region(s) of memory that should
855 * not be tested for the modified bit.
859 pmap_track_modified(vm_offset_t va)
861 if ((va < clean_sva) || (va >= clean_eva))
868 * Extract the physical page address associated with the map/VA pair.
870 * The caller must hold pmap->pm_token if non-blocking operation is desired.
873 pmap_extract(pmap_t pmap, vm_offset_t va)
877 pd_entry_t pde, *pdep;
879 lwkt_gettoken(&pmap->pm_token);
881 pdep = pmap_pde(pmap, va);
885 if ((pde & PG_PS) != 0) {
886 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
888 pte = pmap_pde_to_pte(pdep, va);
889 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
893 lwkt_reltoken(&pmap->pm_token);
898 * Extract the physical page address associated kernel virtual address.
901 pmap_kextract(vm_offset_t va)
906 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
907 pa = DMAP_TO_PHYS(va);
911 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
914 * Beware of a concurrent promotion that changes the
915 * PDE at this point! For example, vtopte() must not
916 * be used to access the PTE because it would use the
917 * new PDE. It is, however, safe to use the old PDE
918 * because the page table page is preserved by the
921 pa = *pmap_pde_to_pte(&pde, va);
922 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
928 /***************************************************
929 * Low level mapping routines.....
930 ***************************************************/
933 * Routine: pmap_kenter
935 * Add a wired page to the KVA
936 * NOTE! note that in order for the mapping to take effect -- you
937 * should do an invltlb after doing the pmap_kenter().
940 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
944 pmap_inval_info info;
946 pmap_inval_init(&info);
947 npte = pa | PG_RW | PG_V | pgeflag;
949 pmap_inval_interlock(&info, &kernel_pmap, va);
951 pmap_inval_deinterlock(&info, &kernel_pmap);
952 pmap_inval_done(&info);
956 * Routine: pmap_kenter_quick
958 * Similar to pmap_kenter(), except we only invalidate the
959 * mapping on the current CPU.
962 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
967 npte = pa | PG_RW | PG_V | pgeflag;
970 cpu_invlpg((void *)va);
974 pmap_kenter_sync(vm_offset_t va)
976 pmap_inval_info info;
978 pmap_inval_init(&info);
979 pmap_inval_interlock(&info, &kernel_pmap, va);
980 pmap_inval_deinterlock(&info, &kernel_pmap);
981 pmap_inval_done(&info);
985 pmap_kenter_sync_quick(vm_offset_t va)
987 cpu_invlpg((void *)va);
991 * remove a page from the kernel pagetables
994 pmap_kremove(vm_offset_t va)
997 pmap_inval_info info;
999 pmap_inval_init(&info);
1001 pmap_inval_interlock(&info, &kernel_pmap, va);
1003 pmap_inval_deinterlock(&info, &kernel_pmap);
1004 pmap_inval_done(&info);
1008 pmap_kremove_quick(vm_offset_t va)
1013 cpu_invlpg((void *)va);
1017 * XXX these need to be recoded. They are not used in any critical path.
1020 pmap_kmodify_rw(vm_offset_t va)
1022 *vtopte(va) |= PG_RW;
1023 cpu_invlpg((void *)va);
1027 pmap_kmodify_nc(vm_offset_t va)
1029 *vtopte(va) |= PG_N;
1030 cpu_invlpg((void *)va);
1034 * Used to map a range of physical addresses into kernel virtual
1035 * address space during the low level boot, typically to map the
1036 * dump bitmap, message buffer, and vm_page_array.
1038 * These mappings are typically made at some pointer after the end of the
1041 * We could return PHYS_TO_DMAP(start) here and not allocate any
1042 * via (*virtp), but then kmem from userland and kernel dumps won't
1043 * have access to the related pointers.
1046 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1049 vm_offset_t va_start;
1051 /*return PHYS_TO_DMAP(start);*/
1056 while (start < end) {
1057 pmap_kenter_quick(va, start);
1067 * Add a list of wired pages to the kva
1068 * this routine is only used for temporary
1069 * kernel mappings that do not need to have
1070 * page modification or references recorded.
1071 * Note that old mappings are simply written
1072 * over. The page *must* be wired.
1075 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1079 end_va = va + count * PAGE_SIZE;
1081 while (va < end_va) {
1085 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1086 cpu_invlpg((void *)va);
1094 * This routine jerks page mappings from the
1095 * kernel -- it is meant only for temporary mappings.
1097 * MPSAFE, INTERRUPT SAFE (cluster callback)
1100 pmap_qremove(vm_offset_t va, int count)
1104 end_va = va + count * PAGE_SIZE;
1106 while (va < end_va) {
1111 cpu_invlpg((void *)va);
1118 * This routine works like vm_page_lookup() but also blocks as long as the
1119 * page is busy. This routine does not busy the page it returns.
1121 * The call should be made with the governing object held so the page's
1122 * object association remains valid on return.
1124 * This function can block!
1128 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1132 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1133 m = vm_page_lookup_busy_wait(object, pindex, FALSE, "pplookp");
1139 * Create a new thread and optionally associate it with a (new) process.
1140 * NOTE! the new thread's cpu may not equal the current cpu.
1143 pmap_init_thread(thread_t td)
1145 /* enforce pcb placement & alignment */
1146 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1147 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1148 td->td_savefpu = &td->td_pcb->pcb_save;
1149 td->td_sp = (char *)td->td_pcb; /* no -16 */
1153 * This routine directly affects the fork perf for a process.
1156 pmap_init_proc(struct proc *p)
1161 * Dispose the UPAGES for a process that has exited.
1162 * This routine directly impacts the exit perf of a process.
1165 pmap_dispose_proc(struct proc *p)
1167 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1170 /***************************************************
1171 * Page table page management routines.....
1172 ***************************************************/
1175 * After removing a page table entry, this routine is used to
1176 * conditionally free the page, and manage the hold/wire counts.
1178 * This routine reduces the wire_count on a page. If the wire_count
1179 * would drop to zero we remove the PT, PD, or PDP from its parent page
1180 * table. Under normal operation this only occurs with PT pages.
1182 * mpte is never NULL for a user va, even for unmanaged pages. mpte should
1183 * always be NULL for a kernel va.
1187 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1188 pmap_inval_info_t info)
1192 if (!vm_page_unwire_quick(mpte))
1196 * Wait until we can busy the page ourselves. We cannot have
1197 * any active flushes if we block. We own one hold count on the
1198 * page so it cannot be freed out from under us.
1200 vm_page_busy_wait(mpte, FALSE, "pmuwpt");
1201 KASSERT(mpte->queue == PQ_NONE,
1202 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", mpte));
1205 * New references can bump the wire_count while we were blocked,
1206 * try to unwire quickly again (e.g. 2->1).
1208 if (vm_page_unwire_quick(mpte) == 0) {
1209 vm_page_wakeup(mpte);
1214 * Unmap the page table page
1216 KKASSERT(mpte->wire_count == 1);
1217 pmap_inval_interlock(info, pmap, -1);
1219 if (mpte->pindex >= (NUPDE + NUPDPE)) {
1222 pml4 = pmap_pml4e(pmap, va);
1225 } else if (mpte->pindex >= NUPDE) {
1228 pdp = pmap_pdpe(pmap, va);
1234 pd = pmap_pde(pmap, va);
1239 KKASSERT(pmap->pm_stats.resident_count > 0);
1240 --pmap->pm_stats.resident_count;
1242 if (pmap->pm_ptphint == mpte)
1243 pmap->pm_ptphint = NULL;
1244 pmap_inval_deinterlock(info, pmap);
1246 if (mpte->pindex < NUPDE) {
1247 /* We just released a PT, unhold the matching PD */
1250 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1251 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1253 if (mpte->pindex >= NUPDE && mpte->pindex < (NUPDE + NUPDPE)) {
1254 /* We just released a PD, unhold the matching PDP */
1257 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1258 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1262 * This was our wiring.
1264 KKASSERT(mpte->flags & PG_UNMANAGED);
1265 vm_page_unwire(mpte, 0);
1266 KKASSERT(mpte->wire_count == 0);
1267 vm_page_flag_clear(mpte, PG_MAPPED | PG_WRITEABLE);
1268 vm_page_flash(mpte);
1269 vm_page_free_zero(mpte);
1275 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1276 * it, and IdlePTD, represents the template used to update all other pmaps.
1278 * On architectures where the kernel pmap is not integrated into the user
1279 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1280 * kernel_pmap should be used to directly access the kernel_pmap.
1283 pmap_pinit0(struct pmap *pmap)
1285 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1287 pmap->pm_active = 0;
1288 pmap->pm_ptphint = NULL;
1289 TAILQ_INIT(&pmap->pm_pvlist);
1290 TAILQ_INIT(&pmap->pm_pvlist_free);
1292 spin_init(&pmap->pm_spin);
1293 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1294 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1298 * Initialize a preallocated and zeroed pmap structure,
1299 * such as one in a vmspace structure.
1302 pmap_pinit(struct pmap *pmap)
1307 * No need to allocate page table space yet but we do need a valid
1308 * page directory table.
1310 if (pmap->pm_pml4 == NULL) {
1312 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1316 * Allocate an object for the ptes
1318 if (pmap->pm_pteobj == NULL) {
1319 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT,
1320 NUPDE + NUPDPE + PML4PML4I + 1);
1324 * Allocate the page directory page, unless we already have
1325 * one cached. If we used the cached page the wire_count will
1326 * already be set appropriately.
1328 if ((pml4pg = pmap->pm_pdirm) == NULL) {
1329 pml4pg = vm_page_grab(pmap->pm_pteobj,
1330 NUPDE + NUPDPE + PML4PML4I,
1331 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1332 pmap->pm_pdirm = pml4pg;
1333 vm_page_unmanage(pml4pg);
1334 vm_page_flag_clear(pml4pg, PG_MAPPED);
1335 pml4pg->valid = VM_PAGE_BITS_ALL;
1336 vm_page_wire(pml4pg);
1337 vm_page_wakeup(pml4pg);
1338 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1339 VM_PAGE_TO_PHYS(pml4pg));
1341 if ((pml4pg->flags & PG_ZERO) == 0)
1342 bzero(pmap->pm_pml4, PAGE_SIZE);
1345 pmap_page_assertzero(VM_PAGE_TO_PHYS(pml4pg));
1347 vm_page_flag_clear(pml4pg, PG_ZERO);
1349 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1350 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1352 /* install self-referential address mapping entry */
1353 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pml4pg) |
1354 PG_V | PG_RW | PG_A | PG_M;
1357 pmap->pm_active = 0;
1358 pmap->pm_ptphint = NULL;
1359 TAILQ_INIT(&pmap->pm_pvlist);
1360 TAILQ_INIT(&pmap->pm_pvlist_free);
1362 spin_init(&pmap->pm_spin);
1363 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1364 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1365 pmap->pm_stats.resident_count = 1;
1369 * Clean up a pmap structure so it can be physically freed. This routine
1370 * is called by the vmspace dtor function. A great deal of pmap data is
1371 * left passively mapped to improve vmspace management so we have a bit
1372 * of cleanup work to do here.
1375 pmap_puninit(pmap_t pmap)
1379 KKASSERT(pmap->pm_active == 0);
1380 if ((p = pmap->pm_pdirm) != NULL) {
1381 KKASSERT(pmap->pm_pml4 != NULL);
1382 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1383 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1384 vm_page_busy_wait(p, FALSE, "pgpun");
1385 KKASSERT(p->flags & PG_UNMANAGED);
1386 vm_page_unwire(p, 0);
1387 vm_page_free_zero(p);
1388 pmap->pm_pdirm = NULL;
1390 if (pmap->pm_pml4) {
1391 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1392 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1393 pmap->pm_pml4 = NULL;
1395 if (pmap->pm_pteobj) {
1396 vm_object_deallocate(pmap->pm_pteobj);
1397 pmap->pm_pteobj = NULL;
1402 * Wire in kernel global address entries. To avoid a race condition
1403 * between pmap initialization and pmap_growkernel, this procedure
1404 * adds the pmap to the master list (which growkernel scans to update),
1405 * then copies the template.
1408 pmap_pinit2(struct pmap *pmap)
1411 * XXX copies current process, does not fill in MPPTDI
1413 spin_lock(&pmap_spin);
1414 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1415 spin_unlock(&pmap_spin);
1419 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1420 * 0 on failure (if the procedure had to sleep).
1422 * When asked to remove the page directory page itself, we actually just
1423 * leave it cached so we do not have to incur the SMP inval overhead of
1424 * removing the kernel mapping. pmap_puninit() will take care of it.
1428 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1431 * This code optimizes the case of freeing non-busy
1432 * page-table pages. Those pages are zero now, and
1433 * might as well be placed directly into the zero queue.
1435 if (vm_page_busy_try(p, FALSE)) {
1436 vm_page_sleep_busy(p, FALSE, "pmaprl");
1441 * Remove the page table page from the processes address space.
1443 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1445 * We are the pml4 table itself.
1447 /* XXX anything to do here? */
1448 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1450 * Remove a PDP page from the PML4. We do not maintain
1451 * wire counts on the PML4 page.
1457 m4 = vm_page_lookup(pmap->pm_pteobj,
1458 NUPDE + NUPDPE + PML4PML4I);
1459 KKASSERT(m4 != NULL);
1460 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1461 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1462 KKASSERT(pml4[idx] != 0);
1464 } else if (p->pindex >= NUPDE) {
1466 * Remove a PD page from the PDP and drop the wire count
1467 * on the PDP. The PDP has a wire_count just from being
1468 * mapped so the wire_count should never drop to 0 here.
1474 m3 = vm_page_lookup(pmap->pm_pteobj,
1475 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1476 KKASSERT(m3 != NULL);
1477 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1478 idx = (p->pindex - NUPDE) % NPDPEPG;
1479 KKASSERT(pdp[idx] != 0);
1481 if (vm_page_unwire_quick(m3))
1482 panic("pmap_release_free_page: m3 wire_count 1->0");
1485 * Remove a PT page from the PD and drop the wire count
1486 * on the PD. The PD has a wire_count just from being
1487 * mapped so the wire_count should never drop to 0 here.
1493 m2 = vm_page_lookup(pmap->pm_pteobj,
1494 NUPDE + p->pindex / NPDEPG);
1495 KKASSERT(m2 != NULL);
1496 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1497 idx = p->pindex % NPDEPG;
1499 if (vm_page_unwire_quick(m2))
1500 panic("pmap_release_free_page: m2 wire_count 1->0");
1504 * p's wire_count should be transitioning from 1 to 0 here.
1506 KKASSERT(p->wire_count == 1);
1507 KKASSERT(p->flags & PG_UNMANAGED);
1508 KKASSERT(pmap->pm_stats.resident_count > 0);
1509 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1510 --pmap->pm_stats.resident_count;
1511 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1512 pmap->pm_ptphint = NULL;
1515 * We leave the top-level page table page cached, wired, and mapped in
1516 * the pmap until the dtor function (pmap_puninit()) gets called.
1517 * However, still clean it up so we can set PG_ZERO.
1519 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1520 bzero(pmap->pm_pml4, PAGE_SIZE);
1521 vm_page_flag_set(p, PG_ZERO);
1524 vm_page_unwire(p, 0);
1525 KKASSERT(p->wire_count == 0);
1526 /* JG eventually revert to using vm_page_free_zero() */
1533 * This routine is called when various levels in the page table need to
1534 * be populated. This routine cannot fail.
1536 * We returned a page wired for the caller. If we had to map the page into
1537 * a parent page table it will receive an additional wire_count. For example,
1538 * an empty page table directory which is still mapped into its pdp will
1539 * retain a wire_count of 1.
1543 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1548 * Find or fabricate a new pagetable page. This will busy the page.
1550 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1551 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1554 * The grab may have blocked and raced another thread populating
1555 * the same page table. m->valid will be 0 on a newly allocated page
1556 * so use this to determine if we have to zero it out or not. We
1557 * don't want to zero-out a raced page as this would desynchronize
1558 * the pv_entry's for the related pte's and cause pmap_remove_all()
1561 * Page table pages are unmanaged (do not use the normal PQ_s)
1563 if (m->valid == 0) {
1564 vm_page_unmanage(m);
1565 if ((m->flags & PG_ZERO) == 0) {
1566 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1570 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1573 m->valid = VM_PAGE_BITS_ALL;
1574 vm_page_flag_clear(m, PG_ZERO);
1578 KKASSERT((m->flags & PG_ZERO) == 0);
1582 KASSERT(m->queue == PQ_NONE,
1583 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1586 * Increment the wire_count for the page we will be returning to
1592 * Map the pagetable page into the process address space, if
1593 * it isn't already there.
1595 * It is possible that someone else got in and mapped the page
1596 * directory page while we were blocked, if so just unbusy and
1597 * return the held page.
1599 if (ptepindex >= (NUPDE + NUPDPE)) {
1601 * Wire up a new PDP page in the PML4.
1603 * (m) is busied so we cannot race another thread trying
1604 * to map the PDP entry in the PML4.
1606 vm_pindex_t pml4index;
1609 pml4index = ptepindex - (NUPDE + NUPDPE);
1610 pml4 = &pmap->pm_pml4[pml4index];
1611 if ((*pml4 & PG_V) == 0) {
1612 *pml4 = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1614 ++pmap->pm_stats.resident_count;
1615 vm_page_wire_quick(m); /* wire for mapping */
1617 /* return (m) wired for the caller */
1618 } else if (ptepindex >= NUPDE) {
1620 * Wire up a new PD page in the PDP
1622 vm_pindex_t pml4index;
1623 vm_pindex_t pdpindex;
1628 pdpindex = ptepindex - NUPDE;
1629 pml4index = pdpindex >> NPML4EPGSHIFT;
1632 * Once mapped the PDP is not unmapped during normal operation
1633 * so we only need to handle races in the unmapped case.
1635 * Mapping a PD into the PDP requires an additional wiring
1638 pml4 = &pmap->pm_pml4[pml4index];
1639 if ((*pml4 & PG_V) == 0) {
1640 pdppg = _pmap_allocpte(pmap,
1641 NUPDE + NUPDPE + pml4index);
1642 /* pdppg wired for the map and also wired for return */
1644 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1645 vm_page_wire_quick(pdppg);
1647 /* we have an extra ref on pdppg now for our use */
1650 * Now find the PD entry in the PDP and map it.
1652 * (m) is busied so we cannot race another thread trying
1653 * to map the PD entry in the PDP.
1655 * If the PD entry is already mapped we have to drop one
1656 * wire count on the pdppg that we had bumped above.
1658 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1659 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1661 if ((*pdp & PG_V) == 0) {
1662 *pdp = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1664 vm_page_wire_quick(m); /* wire for mapping */
1665 ++pmap->pm_stats.resident_count;
1666 /* eat extra pdppg wiring for mapping */
1668 if (vm_page_unwire_quick(pdppg))
1669 panic("pmap_allocpte: unwire case 1");
1671 /* return (m) wired for the caller */
1674 * Wire up the new PT page in the PD
1676 vm_pindex_t pml4index;
1677 vm_pindex_t pdpindex;
1684 pdpindex = ptepindex >> NPDPEPGSHIFT;
1685 pml4index = pdpindex >> NPML4EPGSHIFT;
1688 * Locate the PDP page in the PML4
1690 * Once mapped the PDP is not unmapped during normal operation
1691 * so we only need to handle races in the unmapped case.
1693 pml4 = &pmap->pm_pml4[pml4index];
1694 if ((*pml4 & PG_V) == 0) {
1695 pdppg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1697 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1698 vm_page_wire_quick(pdppg);
1700 /* we have an extra ref on pdppg now for our use */
1703 * Locate the PD page in the PDP
1705 * Once mapped the PDP is not unmapped during normal operation
1706 * so we only need to handle races in the unmapped case.
1708 * We can scrap the extra reference on pdppg not needed if
1709 * *pdp is already mapped and also not needed if it wasn't
1710 * because the _pmap_allocpte() picked up the case for us.
1712 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1713 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1715 if ((*pdp & PG_V) == 0) {
1716 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1718 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1719 vm_page_wire_quick(pdpg);
1721 vm_page_unwire_quick(pdppg);
1722 /* we have an extra ref on pdpg now for our use */
1725 * Locate the PT page in the PD.
1727 * (m) is busied so we cannot race another thread trying
1728 * to map the PT page in the PD.
1730 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1731 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1732 if ((*pd & PG_V) == 0) {
1733 *pd = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1735 ++pmap->pm_stats.resident_count;
1736 vm_page_wire_quick(m); /* wire for mapping */
1737 /* eat extra pdpg wiring for mapping */
1739 if (vm_page_unwire_quick(pdpg))
1740 panic("pmap_allocpte: unwire case 2");
1742 /* return (m) wired for the caller */
1746 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1747 * valid bits, mapped flag, unbusy, and we're done.
1749 pmap->pm_ptphint = m;
1752 m->valid = VM_PAGE_BITS_ALL;
1753 vm_page_flag_clear(m, PG_ZERO);
1755 vm_page_flag_set(m, PG_MAPPED);
1763 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1765 vm_pindex_t ptepindex;
1769 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1772 * Calculate pagetable page index
1774 ptepindex = pmap_pde_pindex(va);
1777 * Get the page directory entry
1779 pd = pmap_pde(pmap, va);
1782 * This supports switching from a 2MB page to a
1785 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1786 panic("no promotion/demotion yet");
1794 * If the page table page is mapped, we just increment the
1795 * wire count, and activate it.
1797 if (pd != NULL && (*pd & PG_V) != 0) {
1798 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1799 pmap->pm_ptphint = m;
1800 vm_page_wire_quick(m);
1805 * Here if the pte page isn't mapped, or if it has been deallocated.
1807 return _pmap_allocpte(pmap, ptepindex);
1811 /***************************************************
1812 * Pmap allocation/deallocation routines.
1813 ***************************************************/
1816 * Release any resources held by the given physical map.
1817 * Called when a pmap initialized by pmap_pinit is being released.
1818 * Should only be called if the map contains no valid mappings.
1820 * Caller must hold pmap->pm_token
1822 static int pmap_release_callback(struct vm_page *p, void *data);
1826 pmap_auto_yield(struct rb_vm_page_scan_info *info)
1828 if (++info->desired >= pmap_yield_count) {
1835 pmap_release(struct pmap *pmap)
1837 vm_object_t object = pmap->pm_pteobj;
1838 struct rb_vm_page_scan_info info;
1840 KASSERT(pmap->pm_active == 0,
1841 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1842 #if defined(DIAGNOSTIC)
1843 if (object->ref_count != 1)
1844 panic("pmap_release: pteobj reference count != 1");
1848 info.object = object;
1850 spin_lock(&pmap_spin);
1851 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1852 spin_unlock(&pmap_spin);
1855 vm_object_hold(object);
1859 info.limit = object->generation;
1861 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1862 pmap_release_callback, &info);
1863 if (info.error == 0 && info.mpte) {
1864 if (!pmap_release_free_page(pmap, info.mpte))
1867 } while (info.error);
1868 vm_object_drop(object);
1870 while (pmap->pm_hold)
1871 tsleep(pmap, 0, "pmapx", 1);
1876 pmap_release_callback(struct vm_page *p, void *data)
1878 struct rb_vm_page_scan_info *info = data;
1880 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1884 if (!pmap_release_free_page(info->pmap, p)) {
1886 pmap_auto_yield(info);
1889 if (info->object->generation != info->limit) {
1891 pmap_auto_yield(info);
1898 * Grow the number of kernel page table entries, if needed.
1900 * This routine is always called to validate any address space
1901 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1902 * space below KERNBASE.
1905 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1908 vm_offset_t ptppaddr;
1910 pd_entry_t *pde, newpdir;
1912 int update_kernel_vm_end;
1914 vm_object_hold(kptobj);
1917 * bootstrap kernel_vm_end on first real VM use
1919 if (kernel_vm_end == 0) {
1920 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1922 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1923 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1924 ~(PAGE_SIZE * NPTEPG - 1);
1926 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1927 kernel_vm_end = kernel_map.max_offset;
1934 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1935 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1936 * do not want to force-fill 128G worth of page tables.
1938 if (kstart < KERNBASE) {
1939 if (kstart > kernel_vm_end)
1940 kstart = kernel_vm_end;
1941 KKASSERT(kend <= KERNBASE);
1942 update_kernel_vm_end = 1;
1944 update_kernel_vm_end = 0;
1947 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1948 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1950 if (kend - 1 >= kernel_map.max_offset)
1951 kend = kernel_map.max_offset;
1953 while (kstart < kend) {
1954 pde = pmap_pde(&kernel_pmap, kstart);
1956 /* We need a new PDP entry */
1957 nkpg = vm_page_alloc(kptobj, nkpt,
1960 VM_ALLOC_INTERRUPT);
1962 panic("pmap_growkernel: no memory to grow "
1965 paddr = VM_PAGE_TO_PHYS(nkpg);
1966 if ((nkpg->flags & PG_ZERO) == 0)
1967 pmap_zero_page(paddr);
1968 vm_page_flag_clear(nkpg, PG_ZERO);
1969 newpdp = (pdp_entry_t)
1970 (paddr | PG_V | PG_RW | PG_A | PG_M);
1971 *pmap_pdpe(&kernel_pmap, kstart) = newpdp;
1973 continue; /* try again */
1975 if ((*pde & PG_V) != 0) {
1976 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1977 ~(PAGE_SIZE * NPTEPG - 1);
1978 if (kstart - 1 >= kernel_map.max_offset) {
1979 kstart = kernel_map.max_offset;
1986 * This index is bogus, but out of the way
1988 nkpg = vm_page_alloc(kptobj, nkpt,
1991 VM_ALLOC_INTERRUPT);
1993 panic("pmap_growkernel: no memory to grow kernel");
1996 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1997 pmap_zero_page(ptppaddr);
1998 vm_page_flag_clear(nkpg, PG_ZERO);
1999 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2000 *pmap_pde(&kernel_pmap, kstart) = newpdir;
2003 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2004 ~(PAGE_SIZE * NPTEPG - 1);
2006 if (kstart - 1 >= kernel_map.max_offset) {
2007 kstart = kernel_map.max_offset;
2013 * Only update kernel_vm_end for areas below KERNBASE.
2015 if (update_kernel_vm_end && kernel_vm_end < kstart)
2016 kernel_vm_end = kstart;
2018 vm_object_drop(kptobj);
2022 * Retire the given physical map from service.
2023 * Should only be called if the map contains
2024 * no valid mappings.
2027 pmap_destroy(pmap_t pmap)
2034 lwkt_gettoken(&pmap->pm_token);
2035 count = --pmap->pm_count;
2037 pmap_release(pmap); /* eats pm_token */
2038 panic("destroying a pmap is not yet implemented");
2040 lwkt_reltoken(&pmap->pm_token);
2044 * Add a reference to the specified pmap.
2047 pmap_reference(pmap_t pmap)
2050 lwkt_gettoken(&pmap->pm_token);
2052 lwkt_reltoken(&pmap->pm_token);
2056 /***************************************************
2057 * page management routines.
2058 ***************************************************/
2061 * free the pv_entry back to the free list. This function may be
2062 * called from an interrupt.
2066 free_pv_entry(pv_entry_t pv)
2068 atomic_add_int(&pv_entry_count, -1);
2069 KKASSERT(pv_entry_count >= 0);
2074 * get a new pv_entry, allocating a block from the system
2075 * when needed. This function may be called from an interrupt.
2081 atomic_add_int(&pv_entry_count, 1);
2082 if (pv_entry_high_water &&
2083 (pv_entry_count > pv_entry_high_water) &&
2084 (pmap_pagedaemon_waken == 0)) {
2085 pmap_pagedaemon_waken = 1;
2086 wakeup(&vm_pages_needed);
2088 return zalloc(pvzone);
2092 * This routine is very drastic, but can save the system
2100 static int warningdone=0;
2102 if (pmap_pagedaemon_waken == 0)
2104 pmap_pagedaemon_waken = 0;
2105 if (warningdone < 5) {
2106 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2110 for (i = 0; i < vm_page_array_size; i++) {
2111 m = &vm_page_array[i];
2112 if (m->wire_count || m->hold_count)
2114 if (vm_page_busy_try(m, TRUE) == 0) {
2115 if (m->wire_count == 0 && m->hold_count == 0) {
2125 * If it is the first entry on the list, it is actually in the header and
2126 * we must copy the following entry up to the header.
2128 * Otherwise we must search the list for the entry. In either case we
2129 * free the now unused entry.
2131 * Caller must hold pmap->pm_token
2135 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2136 vm_offset_t va, pmap_inval_info_t info)
2141 spin_lock(&pmap_spin);
2142 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2143 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2144 if (pmap == pv->pv_pmap && va == pv->pv_va)
2148 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2149 if (va == pv->pv_va)
2157 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2158 m->md.pv_generation++;
2159 m->md.pv_list_count--;
2160 vm_page_spin_lock(m);
2162 atomic_add_int(&m->object->agg_pv_list_count, -1);
2163 vm_page_spin_unlock(m);
2164 KKASSERT(m->md.pv_list_count >= 0);
2165 if (TAILQ_EMPTY(&m->md.pv_list))
2166 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2167 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2168 ++pmap->pm_generation;
2169 spin_unlock(&pmap_spin);
2171 rtval = pmap_unwire_pte_hold(pmap, va, pv->pv_ptem, info);
2178 * Create a pv entry for page at pa for (pmap, va).
2180 * Caller must hold pmap token
2184 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2188 pv = get_pv_entry();
2193 spin_lock(&pmap_spin);
2194 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2195 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2196 m->md.pv_generation++;
2197 m->md.pv_list_count++;
2198 vm_page_spin_lock(m);
2200 atomic_add_int(&m->object->agg_pv_list_count, 1);
2201 vm_page_spin_unlock(m);
2202 pmap->pm_generation++;
2203 spin_unlock(&pmap_spin);
2207 * pmap_remove_pte: do the things to unmap a page in a process
2209 * Caller must hold pmap token
2210 * Caller must hold pmap object
2214 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2215 pmap_inval_info_t info)
2220 vm_pindex_t ptepindex;
2222 ASSERT_LWKT_TOKEN_HELD(&pmap->pm_token);
2224 pmap_inval_interlock(info, pmap, va);
2225 oldpte = pte_load_clear(ptq);
2226 pmap_inval_deinterlock(info, pmap);
2228 pmap->pm_stats.wired_count -= 1;
2230 * Machines that don't support invlpg, also don't support
2231 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2235 cpu_invlpg((void *)va);
2236 KKASSERT(pmap->pm_stats.resident_count > 0);
2237 --pmap->pm_stats.resident_count;
2238 if (oldpte & PG_MANAGED) {
2239 m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME);
2240 if (oldpte & PG_M) {
2241 #if defined(PMAP_DIAGNOSTIC)
2242 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2244 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2248 if (pmap_track_modified(va))
2252 vm_page_flag_set(m, PG_REFERENCED);
2253 return pmap_remove_entry(pmap, m, va, info);
2257 * Unmanaged pages in userspace still wire the PT page, we have
2258 * to look up the mpte for the PDE page and pass it in.
2260 if (va < VM_MAX_USER_ADDRESS) {
2261 ptepindex = pmap_pde_pindex(va);
2262 mpte = vm_page_lookup(pmap->pm_pteobj, ptepindex);
2267 return pmap_unwire_pte_hold(pmap, va, mpte, info);
2271 * Remove a single page from a process address space.
2273 * This function may not be called from an interrupt if the pmap is
2276 * Caller must hold pmap->pm_token
2277 * Caller must hold pmap object
2281 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2285 ASSERT_LWKT_TOKEN_HELD(&pmap->pm_token);
2287 pte = pmap_pte(pmap, va);
2290 if ((*pte & PG_V) == 0)
2292 pmap_remove_pte(pmap, pte, va, info);
2296 * Remove the given range of addresses from the specified map.
2298 * It is assumed that the start and end are properly rounded to the page size.
2300 * This function may not be called from an interrupt if the pmap is not
2304 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2306 vm_offset_t va_next;
2307 pml4_entry_t *pml4e;
2309 pd_entry_t ptpaddr, *pde;
2311 struct pmap_inval_info info;
2316 vm_object_hold(pmap->pm_pteobj);
2317 lwkt_gettoken(&pmap->pm_token);
2318 if (pmap->pm_stats.resident_count == 0) {
2319 lwkt_reltoken(&pmap->pm_token);
2320 vm_object_drop(pmap->pm_pteobj);
2324 pmap_inval_init(&info);
2327 * special handling of removing one page. a very
2328 * common operation and easy to short circuit some
2331 if (sva + PAGE_SIZE == eva) {
2332 pde = pmap_pde(pmap, sva);
2333 if (pde && (*pde & PG_PS) == 0) {
2334 pmap_remove_page(pmap, sva, &info);
2335 pmap_inval_done(&info);
2336 lwkt_reltoken(&pmap->pm_token);
2337 vm_object_drop(pmap->pm_pteobj);
2342 for (; sva < eva; sva = va_next) {
2343 pml4e = pmap_pml4e(pmap, sva);
2344 if ((*pml4e & PG_V) == 0) {
2345 va_next = (sva + NBPML4) & ~PML4MASK;
2351 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2352 if ((*pdpe & PG_V) == 0) {
2353 va_next = (sva + NBPDP) & ~PDPMASK;
2360 * Calculate index for next page table.
2362 va_next = (sva + NBPDR) & ~PDRMASK;
2366 pde = pmap_pdpe_to_pde(pdpe, sva);
2370 * Weed out invalid mappings.
2376 * Check for large page.
2378 if ((ptpaddr & PG_PS) != 0) {
2379 /* JG FreeBSD has more complex treatment here */
2380 pmap_inval_interlock(&info, pmap, -1);
2382 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2383 pmap_inval_deinterlock(&info, pmap);
2388 * Limit our scan to either the end of the va represented
2389 * by the current page table page, or to the end of the
2390 * range being removed.
2396 * NOTE: pmap_remove_pte() can block.
2398 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2402 if (pmap_remove_pte(pmap, pte, sva, &info))
2406 pmap_inval_done(&info);
2407 lwkt_reltoken(&pmap->pm_token);
2408 vm_object_drop(pmap->pm_pteobj);
2412 * Removes this physical page from all physical maps in which it resides.
2413 * Reflects back modify bits to the pager.
2415 * This routine may not be called from an interrupt.
2419 pmap_remove_all(vm_page_t m)
2421 struct pmap_inval_info info;
2422 pt_entry_t *pte, tpte;
2426 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2429 pmap_inval_init(&info);
2430 spin_lock(&pmap_spin);
2431 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2433 * We have to be holding the pmap token to interlock
2434 * the pte destruction and pv removal. XXX need hold on
2438 spin_unlock(&pmap_spin);
2439 lwkt_gettoken(&pmap->pm_token); /* XXX hold race */
2440 spin_lock(&pmap_spin);
2441 if (pv != TAILQ_FIRST(&m->md.pv_list)) {
2442 spin_unlock(&pmap_spin);
2443 lwkt_reltoken(&pmap->pm_token);
2444 spin_lock(&pmap_spin);
2451 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2452 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2453 m->md.pv_generation++;
2454 m->md.pv_list_count--;
2455 vm_page_spin_lock(m);
2457 atomic_add_int(&m->object->agg_pv_list_count, -1);
2458 vm_page_spin_unlock(m);
2459 KKASSERT(m->md.pv_list_count >= 0);
2460 ++pv->pv_pmap->pm_generation;
2461 spin_unlock(&pmap_spin);
2464 * pv is now isolated
2466 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2467 --pv->pv_pmap->pm_stats.resident_count;
2469 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2470 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2471 tpte = pte_load_clear(pte);
2472 KKASSERT(tpte & PG_MANAGED);
2474 pv->pv_pmap->pm_stats.wired_count--;
2475 pmap_inval_deinterlock(&info, pv->pv_pmap);
2477 vm_page_flag_set(m, PG_REFERENCED);
2480 * Update the vm_page_t clean and reference bits.
2483 #if defined(PMAP_DIAGNOSTIC)
2484 if (pmap_nw_modified(tpte)) {
2485 kprintf("pmap_remove_all: modified page not "
2486 "writable: va: 0x%lx, pte: 0x%lx\n",
2490 if (pmap_track_modified(pv->pv_va))
2491 vm_page_dirty(m); /* XXX races(m) */
2494 spin_lock(&pmap_spin);
2495 if (TAILQ_EMPTY(&m->md.pv_list))
2496 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2497 spin_unlock(&pmap_spin);
2499 pmap_unwire_pte_hold(pv->pv_pmap, pv->pv_va,
2500 pv->pv_ptem, &info);
2501 lwkt_reltoken(&pv->pv_pmap->pm_token);
2504 spin_lock(&pmap_spin);
2506 spin_unlock(&pmap_spin);
2507 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2508 pmap_inval_done(&info);
2514 * Set the physical protection on the specified range of this map
2517 * This function may not be called from an interrupt if the map is
2518 * not the kernel_pmap.
2521 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2523 vm_offset_t va_next;
2524 pml4_entry_t *pml4e;
2526 pd_entry_t ptpaddr, *pde;
2528 pmap_inval_info info;
2530 /* JG review for NX */
2535 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2536 pmap_remove(pmap, sva, eva);
2540 if (prot & VM_PROT_WRITE)
2543 lwkt_gettoken(&pmap->pm_token);
2544 pmap_inval_init(&info);
2546 for (; sva < eva; sva = va_next) {
2547 pml4e = pmap_pml4e(pmap, sva);
2548 if ((*pml4e & PG_V) == 0) {
2549 va_next = (sva + NBPML4) & ~PML4MASK;
2555 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2556 if ((*pdpe & PG_V) == 0) {
2557 va_next = (sva + NBPDP) & ~PDPMASK;
2563 va_next = (sva + NBPDR) & ~PDRMASK;
2567 pde = pmap_pdpe_to_pde(pdpe, sva);
2571 * Check for large page.
2573 if ((ptpaddr & PG_PS) != 0) {
2574 pmap_inval_interlock(&info, pmap, -1);
2575 *pde &= ~(PG_M|PG_RW);
2576 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2577 pmap_inval_deinterlock(&info, pmap);
2582 * Weed out invalid mappings. Note: we assume that the page
2583 * directory table is always allocated, and in kernel virtual.
2591 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2600 pmap_inval_interlock(&info, pmap, sva);
2604 if ((pbits & PG_V) == 0) {
2605 pmap_inval_deinterlock(&info, pmap);
2608 if (pbits & PG_MANAGED) {
2611 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2612 vm_page_flag_set(m, PG_REFERENCED);
2616 if (pmap_track_modified(sva)) {
2618 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2625 if (pbits != cbits &&
2626 !atomic_cmpset_long(pte, pbits, cbits)) {
2629 pmap_inval_deinterlock(&info, pmap);
2632 pmap_inval_done(&info);
2633 lwkt_reltoken(&pmap->pm_token);
2637 * Insert the given physical page (p) at
2638 * the specified virtual address (v) in the
2639 * target physical map with the protection requested.
2641 * If specified, the page will be wired down, meaning
2642 * that the related pte can not be reclaimed.
2644 * NB: This is the only routine which MAY NOT lazy-evaluate
2645 * or lose information. That is, this routine must actually
2646 * insert this page into the given map NOW.
2649 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2656 pt_entry_t origpte, newpte;
2658 pmap_inval_info info;
2663 va = trunc_page(va);
2664 #ifdef PMAP_DIAGNOSTIC
2666 panic("pmap_enter: toobig");
2667 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2668 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2670 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2671 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2673 db_print_backtrace();
2676 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2677 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2679 db_print_backtrace();
2683 vm_object_hold(pmap->pm_pteobj);
2684 lwkt_gettoken(&pmap->pm_token);
2687 * In the case that a page table page is not
2688 * resident, we are creating it here.
2690 if (va < VM_MAX_USER_ADDRESS)
2691 mpte = pmap_allocpte(pmap, va);
2695 if ((prot & VM_PROT_NOSYNC) == 0)
2696 pmap_inval_init(&info);
2697 pde = pmap_pde(pmap, va);
2698 if (pde != NULL && (*pde & PG_V) != 0) {
2699 if ((*pde & PG_PS) != 0)
2700 panic("pmap_enter: attempted pmap_enter on 2MB page");
2701 pte = pmap_pde_to_pte(pde, va);
2703 panic("pmap_enter: invalid page directory va=%#lx", va);
2706 KKASSERT(pte != NULL);
2707 pa = VM_PAGE_TO_PHYS(m);
2709 opa = origpte & PG_FRAME;
2712 * Mapping has not changed, must be protection or wiring change.
2714 if (origpte && (opa == pa)) {
2716 * Wiring change, just update stats. We don't worry about
2717 * wiring PT pages as they remain resident as long as there
2718 * are valid mappings in them. Hence, if a user page is wired,
2719 * the PT page will be also.
2721 if (wired && ((origpte & PG_W) == 0))
2722 pmap->pm_stats.wired_count++;
2723 else if (!wired && (origpte & PG_W))
2724 pmap->pm_stats.wired_count--;
2726 #if defined(PMAP_DIAGNOSTIC)
2727 if (pmap_nw_modified(origpte)) {
2729 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2735 * Remove the extra pte reference. Note that we cannot
2736 * optimize the RO->RW case because we have adjusted the
2737 * wiring count above and may need to adjust the wiring
2741 vm_page_unwire_quick(mpte);
2744 * We might be turning off write access to the page,
2745 * so we go ahead and sense modify status.
2747 if (origpte & PG_MANAGED) {
2748 if ((origpte & PG_M) && pmap_track_modified(va)) {
2750 om = PHYS_TO_VM_PAGE(opa);
2754 KKASSERT(m->flags & PG_MAPPED);
2759 * Mapping has changed, invalidate old range and fall through to
2760 * handle validating new mapping.
2764 err = pmap_remove_pte(pmap, pte, va, &info);
2766 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2768 opa = origpte & PG_FRAME;
2770 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2776 * Enter on the PV list if part of our managed memory. Note that we
2777 * raise IPL while manipulating pv_table since pmap_enter can be
2778 * called at interrupt time.
2780 * The new mapping covers mpte's new wiring count so we don't
2783 if (pmap_initialized &&
2784 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2785 pmap_insert_entry(pmap, va, mpte, m);
2787 vm_page_flag_set(m, PG_MAPPED);
2791 * Increment counters
2793 ++pmap->pm_stats.resident_count;
2795 pmap->pm_stats.wired_count++;
2799 * Now validate mapping with desired protection/wiring.
2801 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2805 if (va < VM_MAX_USER_ADDRESS)
2807 if (pmap == &kernel_pmap)
2811 * if the mapping or permission bits are different, we need
2812 * to update the pte.
2814 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2815 if ((prot & VM_PROT_NOSYNC) == 0)
2816 pmap_inval_interlock(&info, pmap, va);
2817 *pte = newpte | PG_A;
2818 if (prot & VM_PROT_NOSYNC)
2819 cpu_invlpg((void *)va);
2821 pmap_inval_deinterlock(&info, pmap);
2823 vm_page_flag_set(m, PG_WRITEABLE);
2825 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2826 if ((prot & VM_PROT_NOSYNC) == 0)
2827 pmap_inval_done(&info);
2828 lwkt_reltoken(&pmap->pm_token);
2829 vm_object_drop(pmap->pm_pteobj);
2833 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2834 * This code also assumes that the pmap has no pre-existing entry for this
2837 * This code currently may only be used on user pmaps, not kernel_pmap.
2840 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2845 pmap_inval_info info;
2847 lwkt_gettoken(&pmap->pm_token);
2848 vm_object_hold(pmap->pm_pteobj);
2849 pmap_inval_init(&info);
2851 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2852 kprintf("Warning: pmap_enter_quick called on UVA with"
2855 db_print_backtrace();
2858 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2859 kprintf("Warning: pmap_enter_quick called on KVA without"
2862 db_print_backtrace();
2866 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2869 * Calculate the page table page (mpte), allocating it if necessary.
2871 * A wired page table page (mpte), or NULL, is passed onto the
2872 * section following.
2874 if (va < VM_MAX_USER_ADDRESS) {
2875 mpte = pmap_allocpte(pmap, va);
2878 /* this code path is not yet used */
2882 * With a valid (and held) page directory page, we can just use
2883 * vtopte() to get to the pte. If the pte is already present
2884 * we do not disturb it.
2888 pa = VM_PAGE_TO_PHYS(m);
2889 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2890 pmap_inval_done(&info);
2892 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2893 vm_object_drop(pmap->pm_pteobj);
2894 lwkt_reltoken(&pmap->pm_token);
2899 * Enter on the PV list if part of our managed memory.
2901 * The new mapping covers mpte's new wiring count so we don't
2904 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2905 pmap_insert_entry(pmap, va, mpte, m);
2906 vm_page_flag_set(m, PG_MAPPED);
2910 * Increment counters
2912 ++pmap->pm_stats.resident_count;
2914 pa = VM_PAGE_TO_PHYS(m);
2917 * Now validate mapping with RO protection
2919 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2920 *pte = pa | PG_V | PG_U;
2922 *pte = pa | PG_V | PG_U | PG_MANAGED;
2923 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2924 pmap_inval_done(&info);
2925 vm_object_drop(pmap->pm_pteobj);
2926 lwkt_reltoken(&pmap->pm_token);
2930 * Make a temporary mapping for a physical address. This is only intended
2931 * to be used for panic dumps.
2933 * The caller is responsible for calling smp_invltlb().
2936 pmap_kenter_temporary(vm_paddr_t pa, long i)
2938 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2939 return ((void *)crashdumpmap);
2942 #define MAX_INIT_PT (96)
2945 * This routine preloads the ptes for a given object into the specified pmap.
2946 * This eliminates the blast of soft faults on process startup and
2947 * immediately after an mmap.
2949 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2952 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2953 vm_object_t object, vm_pindex_t pindex,
2954 vm_size_t size, int limit)
2956 struct rb_vm_page_scan_info info;
2961 * We can't preinit if read access isn't set or there is no pmap
2964 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2968 * We can't preinit if the pmap is not the current pmap
2970 lp = curthread->td_lwp;
2971 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2974 psize = x86_64_btop(size);
2976 if ((object->type != OBJT_VNODE) ||
2977 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2978 (object->resident_page_count > MAX_INIT_PT))) {
2982 if (psize + pindex > object->size) {
2983 if (object->size < pindex)
2985 psize = object->size - pindex;
2992 * Use a red-black scan to traverse the requested range and load
2993 * any valid pages found into the pmap.
2995 * We cannot safely scan the object's memq without holding the
2998 info.start_pindex = pindex;
2999 info.end_pindex = pindex + psize - 1;
3006 vm_object_hold(object);
3007 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3008 pmap_object_init_pt_callback, &info);
3009 vm_object_drop(object);
3014 pmap_object_init_pt_callback(vm_page_t p, void *data)
3016 struct rb_vm_page_scan_info *info = data;
3017 vm_pindex_t rel_index;
3020 * don't allow an madvise to blow away our really
3021 * free pages allocating pv entries.
3023 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3024 vmstats.v_free_count < vmstats.v_free_reserved) {
3027 if (vm_page_busy_try(p, TRUE))
3029 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3030 (p->flags & PG_FICTITIOUS) == 0) {
3031 if ((p->queue - p->pc) == PQ_CACHE)
3032 vm_page_deactivate(p);
3033 rel_index = p->pindex - info->start_pindex;
3034 pmap_enter_quick(info->pmap,
3035 info->addr + x86_64_ptob(rel_index), p);
3038 pmap_auto_yield(info);
3043 * Return TRUE if the pmap is in shape to trivially
3044 * pre-fault the specified address.
3046 * Returns FALSE if it would be non-trivial or if a
3047 * pte is already loaded into the slot.
3050 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3056 lwkt_gettoken(&pmap->pm_token);
3057 pde = pmap_pde(pmap, addr);
3058 if (pde == NULL || *pde == 0) {
3062 ret = (*pte) ? 0 : 1;
3064 lwkt_reltoken(&pmap->pm_token);
3069 * Routine: pmap_change_wiring
3070 * Function: Change the wiring attribute for a map/virtual-address
3072 * In/out conditions:
3073 * The mapping must already exist in the pmap.
3076 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
3083 lwkt_gettoken(&pmap->pm_token);
3084 pte = pmap_pte(pmap, va);
3086 if (wired && !pmap_pte_w(pte))
3087 pmap->pm_stats.wired_count++;
3088 else if (!wired && pmap_pte_w(pte))
3089 pmap->pm_stats.wired_count--;
3092 * Wiring is not a hardware characteristic so there is no need to
3093 * invalidate TLB. However, in an SMP environment we must use
3094 * a locked bus cycle to update the pte (if we are not using
3095 * the pmap_inval_*() API that is)... it's ok to do this for simple
3100 atomic_set_long(pte, PG_W);
3102 atomic_clear_long(pte, PG_W);
3105 atomic_set_long_nonlocked(pte, PG_W);
3107 atomic_clear_long_nonlocked(pte, PG_W);
3109 lwkt_reltoken(&pmap->pm_token);
3115 * Copy the range specified by src_addr/len from the source map to
3116 * the range dst_addr/len in the destination map.
3118 * This routine is only advisory and need not do anything.
3121 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3122 vm_size_t len, vm_offset_t src_addr)
3126 pmap_inval_info info;
3128 vm_offset_t end_addr = src_addr + len;
3130 pd_entry_t src_frame, dst_frame;
3133 if (dst_addr != src_addr)
3136 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3137 if (src_frame != (PTDpde & PG_FRAME)) {
3141 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3142 if (dst_frame != (APTDpde & PG_FRAME)) {
3143 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3144 /* The page directory is not shared between CPUs */
3148 pmap_inval_init(&info);
3149 pmap_inval_add(&info, dst_pmap, -1);
3150 pmap_inval_add(&info, src_pmap, -1);
3152 lwkt_gettoken(&src_pmap->pm_token);
3153 lwkt_gettoken(&dst_pmap->pm_token);
3154 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3155 pt_entry_t *src_pte, *dst_pte;
3156 vm_page_t dstmpte, srcmpte;
3157 vm_offset_t srcptepaddr;
3158 vm_pindex_t ptepindex;
3160 if (addr >= UPT_MIN_ADDRESS)
3161 panic("pmap_copy: invalid to pmap_copy page tables\n");
3164 * Don't let optional prefaulting of pages make us go
3165 * way below the low water mark of free pages or way
3166 * above high water mark of used pv entries.
3168 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3169 pv_entry_count > pv_entry_high_water)
3172 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3173 ptepindex = addr >> PDRSHIFT;
3176 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3178 if (srcptepaddr == 0)
3181 if (srcptepaddr & PG_PS) {
3183 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3184 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3185 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3194 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3195 if (srcmpte == NULL || srcmpte->wire_count == 1 ||
3196 (srcmpte->flags & PG_BUSY)) {
3200 if (pdnxt > end_addr)
3203 src_pte = vtopte(addr);
3205 dst_pte = avtopte(addr);
3207 while (addr < pdnxt) {
3212 * we only virtual copy managed pages
3214 if ((ptetemp & PG_MANAGED) != 0) {
3216 * We have to check after allocpte for the
3217 * pte still being around... allocpte can
3220 * pmap_allocpte() can block. If we lose
3221 * our page directory mappings we stop.
3223 dstmpte = pmap_allocpte(dst_pmap, addr);
3226 if (src_frame != (PTDpde & PG_FRAME) ||
3227 dst_frame != (APTDpde & PG_FRAME)
3229 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3230 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3232 } else if ((*dst_pte == 0) &&
3233 (ptetemp = *src_pte) != 0 &&
3234 (ptetemp & PG_MANAGED)) {
3236 * Clear the modified and
3237 * accessed (referenced) bits
3240 m = PHYS_TO_VM_PAGE(ptetemp);
3241 *dst_pte = ptetemp & ~(PG_M | PG_A);
3242 ++dst_pmap->pm_stats.resident_count;
3243 pmap_insert_entry(dst_pmap, addr,
3245 KKASSERT(m->flags & PG_MAPPED);
3247 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3248 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3252 if (dstmpte->hold_count >= srcmpte->hold_count)
3261 lwkt_reltoken(&dst_pmap->pm_token);
3262 lwkt_reltoken(&src_pmap->pm_token);
3263 pmap_inval_done(&info);
3270 * Zero the specified physical page.
3272 * This function may be called from an interrupt and no locking is
3276 pmap_zero_page(vm_paddr_t phys)
3278 vm_offset_t va = PHYS_TO_DMAP(phys);
3280 pagezero((void *)va);
3284 * pmap_page_assertzero:
3286 * Assert that a page is empty, panic if it isn't.
3289 pmap_page_assertzero(vm_paddr_t phys)
3291 vm_offset_t va = PHYS_TO_DMAP(phys);
3294 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3295 if (*(long *)((char *)va + i) != 0) {
3296 panic("pmap_page_assertzero() @ %p not zero!\n",
3297 (void *)(intptr_t)va);
3305 * Zero part of a physical page by mapping it into memory and clearing
3306 * its contents with bzero.
3308 * off and size may not cover an area beyond a single hardware page.
3311 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3313 vm_offset_t virt = PHYS_TO_DMAP(phys);
3315 bzero((char *)virt + off, size);
3321 * Copy the physical page from the source PA to the target PA.
3322 * This function may be called from an interrupt. No locking
3326 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3328 vm_offset_t src_virt, dst_virt;
3330 src_virt = PHYS_TO_DMAP(src);
3331 dst_virt = PHYS_TO_DMAP(dst);
3332 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3336 * pmap_copy_page_frag:
3338 * Copy the physical page from the source PA to the target PA.
3339 * This function may be called from an interrupt. No locking
3343 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3345 vm_offset_t src_virt, dst_virt;
3347 src_virt = PHYS_TO_DMAP(src);
3348 dst_virt = PHYS_TO_DMAP(dst);
3350 bcopy((char *)src_virt + (src & PAGE_MASK),
3351 (char *)dst_virt + (dst & PAGE_MASK),
3356 * Returns true if the pmap's pv is one of the first
3357 * 16 pvs linked to from this page. This count may
3358 * be changed upwards or downwards in the future; it
3359 * is only necessary that true be returned for a small
3360 * subset of pmaps for proper page aging.
3363 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3368 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3371 spin_lock(&pmap_spin);
3372 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3373 if (pv->pv_pmap == pmap) {
3374 spin_unlock(&pmap_spin);
3381 spin_unlock(&pmap_spin);
3386 * Remove all pages from specified address space this aids process exit
3387 * speeds. Also, this code is special cased for current process only, but
3388 * can have the more generic (and slightly slower) mode enabled. This
3389 * is much faster than pmap_remove in the case of running down an entire
3393 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3396 pt_entry_t *pte, tpte;
3400 pmap_inval_info info;
3402 int save_generation;
3404 lp = curthread->td_lwp;
3405 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3410 if (pmap->pm_pteobj)
3411 vm_object_hold(pmap->pm_pteobj);
3412 lwkt_gettoken(&pmap->pm_token);
3413 pmap_inval_init(&info);
3415 spin_lock(&pmap_spin);
3416 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3418 * Validate the pv. We have to interlock the address with
3419 * pmap_spin unlocked.
3421 if (pv->pv_va >= eva || pv->pv_va < sva) {
3422 npv = TAILQ_NEXT(pv, pv_plist);
3426 KKASSERT(pmap == pv->pv_pmap);
3428 pte = vtopte(pv->pv_va);
3430 pte = pmap_pte_quick(pmap, pv->pv_va);
3433 * We cannot remove wired pages from a process' mapping
3434 * at this time. This does not require an invaldiation
3435 * interlock as PG_W cannot be set by the MMU.
3438 npv = TAILQ_NEXT(pv, pv_plist);
3443 * Interlock the pte so we can safely remove it
3445 save_generation = pmap->pm_generation;
3447 spin_unlock(&pmap_spin);
3449 pmap_inval_interlock(&info, pmap, va);
3452 * Restart the scan if the pv list changed out from under us.
3454 spin_lock(&pmap_spin);
3455 if (save_generation != pmap->pm_generation) {
3456 spin_unlock(&pmap_spin);
3457 pmap_inval_deinterlock(&info, pmap);
3458 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3459 spin_lock(&pmap_spin);
3460 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3463 KKASSERT(pmap == pv->pv_pmap && va == pv->pv_va);
3466 * Extract the pte and clear its memory
3468 tpte = pte_load_clear(pte);
3469 KKASSERT(tpte & PG_MANAGED);
3471 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3472 KASSERT(m < &vm_page_array[vm_page_array_size],
3473 ("pmap_remove_pages: bad tpte %lx", tpte));
3476 * Remove the entry, set npv
3478 npv = TAILQ_NEXT(pv, pv_plist);
3479 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3480 m->md.pv_generation++;
3481 m->md.pv_list_count--;
3482 vm_page_spin_lock(m);
3484 atomic_add_int(&m->object->agg_pv_list_count, -1);
3485 vm_page_spin_unlock(m);
3486 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3487 if (TAILQ_EMPTY(&m->md.pv_list))
3488 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3489 save_generation = ++pmap->pm_generation;
3491 spin_unlock(&pmap_spin);
3494 * Adjust the pmap and cleanup the tpte and related vm_page
3496 KKASSERT(pmap->pm_stats.resident_count > 0);
3497 --pmap->pm_stats.resident_count;
3498 pmap_inval_deinterlock(&info, pmap);
3501 * Update the vm_page_t clean and reference bits.
3507 pmap_unwire_pte_hold(pmap, pv->pv_va, pv->pv_ptem, &info);
3511 * Restart the scan if we blocked during the unuse or free
3512 * calls and other removals were made.
3514 spin_lock(&pmap_spin);
3515 if (save_generation != pmap->pm_generation) {
3516 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3517 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3520 spin_unlock(&pmap_spin);
3521 pmap_inval_done(&info);
3522 lwkt_reltoken(&pmap->pm_token);
3523 if (pmap->pm_pteobj)
3524 vm_object_drop(pmap->pm_pteobj);
3528 * pmap_testbit tests bits in pte's note that the testbit/clearbit
3529 * routines are inline, and a lot of things compile-time evaluate.
3531 * Caller must hold pmap_spin
3535 pmap_testbit(vm_page_t m, int bit)
3540 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3543 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3546 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3548 * if the bit being tested is the modified bit, then
3549 * mark clean_map and ptes as never
3552 if (bit & (PG_A|PG_M)) {
3553 if (!pmap_track_modified(pv->pv_va))
3557 #if defined(PMAP_DIAGNOSTIC)
3558 if (pv->pv_pmap == NULL) {
3559 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3563 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3571 * This routine is used to modify bits in ptes
3573 * Caller must NOT hold pmap_spin
3577 pmap_clearbit(vm_page_t m, int bit)
3579 struct pmap_inval_info info;
3580 int save_generation;
3581 vm_offset_t save_va;
3582 struct pmap *save_pmap;
3588 vm_page_flag_clear(m, PG_WRITEABLE);
3589 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
3593 pmap_inval_init(&info);
3596 * Loop over all current mappings setting/clearing as appropos If
3597 * setting RO do we need to clear the VAC?
3599 spin_lock(&pmap_spin);
3601 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3603 * don't write protect pager mappings
3606 if (!pmap_track_modified(pv->pv_va))
3610 #if defined(PMAP_DIAGNOSTIC)
3611 if (pv->pv_pmap == NULL) {
3612 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3618 * Careful here. We can use a locked bus instruction to
3619 * clear PG_A or PG_M safely but we need to synchronize
3620 * with the target cpus when we mess with PG_RW.
3622 * We do not have to force synchronization when clearing
3623 * PG_M even for PTEs generated via virtual memory maps,
3624 * because the virtual kernel will invalidate the pmap
3625 * entry when/if it needs to resynchronize the Modify bit.
3627 * We have to restart our scan if m->md.pv_generation changes
3631 save_generation = m->md.pv_generation;
3632 save_pmap = pv->pv_pmap;
3633 save_va = pv->pv_va;
3634 spin_unlock(&pmap_spin);
3635 pmap_inval_interlock(&info, save_pmap, save_va);
3636 spin_lock(&pmap_spin);
3637 if (save_generation != m->md.pv_generation)
3640 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3647 atomic_clear_long(pte, PG_M|PG_RW);
3650 * The cpu may be trying to set PG_M
3651 * simultaniously with our clearing
3654 if (!atomic_cmpset_long(pte, pbits,
3658 } else if (bit == PG_M) {
3660 * We could also clear PG_RW here to force
3661 * a fault on write to redetect PG_M for
3662 * virtual kernels, but it isn't necessary
3663 * since virtual kernels invalidate the pte
3664 * when they clear the VPTE_M bit in their
3665 * virtual page tables.
3667 atomic_clear_long(pte, PG_M);
3669 atomic_clear_long(pte, bit);
3673 save_generation = m->md.pv_generation;
3674 save_pmap = pv->pv_pmap;
3675 spin_unlock(&pmap_spin);
3676 pmap_inval_deinterlock(&info, save_pmap);
3677 spin_lock(&pmap_spin);
3678 if (save_generation != m->md.pv_generation)
3682 spin_unlock(&pmap_spin);
3683 pmap_inval_done(&info);
3687 * Lower the permission for all mappings to a given page.
3689 * Page must be busied by caller.
3692 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3694 /* JG NX support? */
3695 if ((prot & VM_PROT_WRITE) == 0) {
3696 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3698 * NOTE: pmap_clearbit(.. PG_RW) also clears
3699 * the PG_WRITEABLE flag in (m).
3701 pmap_clearbit(m, PG_RW);
3709 pmap_phys_address(vm_pindex_t ppn)
3711 return (x86_64_ptob(ppn));
3715 * Return a count of reference bits for a page, clearing those bits.
3716 * It is not necessary for every reference bit to be cleared, but it
3717 * is necessary that 0 only be returned when there are truly no
3718 * reference bits set.
3720 * XXX: The exact number of bits to check and clear is a matter that
3721 * should be tested and standardized at some point in the future for
3722 * optimal aging of shared pages.
3724 * This routine may not block.
3727 pmap_ts_referenced(vm_page_t m)
3729 pv_entry_t pv, pvf, pvn;
3733 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3736 spin_lock(&pmap_spin);
3737 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3740 pvn = TAILQ_NEXT(pv, pv_list);
3742 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3743 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3744 /*++pv->pv_pmap->pm_generation; not needed */
3746 if (!pmap_track_modified(pv->pv_va))
3749 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3751 if (pte && (*pte & PG_A)) {
3753 atomic_clear_long(pte, PG_A);
3755 atomic_clear_long_nonlocked(pte, PG_A);
3762 } while ((pv = pvn) != NULL && pv != pvf);
3764 spin_unlock(&pmap_spin);
3772 * Return whether or not the specified physical page was modified
3773 * in any physical maps.
3776 pmap_is_modified(vm_page_t m)
3780 spin_lock(&pmap_spin);
3781 res = pmap_testbit(m, PG_M);
3782 spin_unlock(&pmap_spin);
3787 * Clear the modify bits on the specified physical page.
3790 pmap_clear_modify(vm_page_t m)
3792 pmap_clearbit(m, PG_M);
3796 * pmap_clear_reference:
3798 * Clear the reference bit on the specified physical page.
3801 pmap_clear_reference(vm_page_t m)
3803 pmap_clearbit(m, PG_A);
3807 * Miscellaneous support routines follow
3812 i386_protection_init(void)
3816 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3817 kp = protection_codes;
3818 for (prot = 0; prot < 8; prot++) {
3820 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3822 * Read access is also 0. There isn't any execute bit,
3823 * so just make it readable.
3825 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3826 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3827 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3830 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3831 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3832 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3833 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3841 * Map a set of physical memory pages into the kernel virtual
3842 * address space. Return a pointer to where it is mapped. This
3843 * routine is intended to be used for mapping device memory,
3846 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3850 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3852 vm_offset_t va, tmpva, offset;
3855 offset = pa & PAGE_MASK;
3856 size = roundup(offset + size, PAGE_SIZE);
3858 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3860 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3862 pa = pa & ~PAGE_MASK;
3863 for (tmpva = va; size > 0;) {
3864 pte = vtopte(tmpva);
3865 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3873 return ((void *)(va + offset));
3877 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3879 vm_offset_t va, tmpva, offset;
3882 offset = pa & PAGE_MASK;
3883 size = roundup(offset + size, PAGE_SIZE);
3885 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3887 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3889 pa = pa & ~PAGE_MASK;
3890 for (tmpva = va; size > 0;) {
3891 pte = vtopte(tmpva);
3892 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3900 return ((void *)(va + offset));
3904 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3906 vm_offset_t base, offset;
3908 base = va & ~PAGE_MASK;
3909 offset = va & PAGE_MASK;
3910 size = roundup(offset + size, PAGE_SIZE);
3911 pmap_qremove(va, size >> PAGE_SHIFT);
3912 kmem_free(&kernel_map, base, size);
3916 * perform the pmap work for mincore
3919 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3921 pt_entry_t *ptep, pte;
3925 lwkt_gettoken(&pmap->pm_token);
3926 ptep = pmap_pte(pmap, addr);
3928 if (ptep && (pte = *ptep) != 0) {
3931 val = MINCORE_INCORE;
3932 if ((pte & PG_MANAGED) == 0)
3935 pa = pte & PG_FRAME;
3937 m = PHYS_TO_VM_PAGE(pa);
3943 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3945 * Modified by someone
3947 else if (m->dirty || pmap_is_modified(m))
3948 val |= MINCORE_MODIFIED_OTHER;
3953 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3956 * Referenced by someone
3958 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3959 val |= MINCORE_REFERENCED_OTHER;
3960 vm_page_flag_set(m, PG_REFERENCED);
3964 lwkt_reltoken(&pmap->pm_token);
3970 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3971 * vmspace will be ref'd and the old one will be deref'd.
3973 * The vmspace for all lwps associated with the process will be adjusted
3974 * and cr3 will be reloaded if any lwp is the current lwp.
3976 * The process must hold the vmspace->vm_map.token for oldvm and newvm
3979 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3981 struct vmspace *oldvm;
3984 oldvm = p->p_vmspace;
3985 if (oldvm != newvm) {
3987 sysref_get(&newvm->vm_sysref);
3988 p->p_vmspace = newvm;
3989 KKASSERT(p->p_nthreads == 1);
3990 lp = RB_ROOT(&p->p_lwp_tree);
3991 pmap_setlwpvm(lp, newvm);
3993 sysref_put(&oldvm->vm_sysref);
3998 * Set the vmspace for a LWP. The vmspace is almost universally set the
3999 * same as the process vmspace, but virtual kernels need to swap out contexts
4000 * on a per-lwp basis.
4002 * Caller does not necessarily hold any vmspace tokens. Caller must control
4003 * the lwp (typically be in the context of the lwp). We use a critical
4004 * section to protect against statclock and hardclock (statistics collection).
4007 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
4009 struct vmspace *oldvm;
4012 oldvm = lp->lwp_vmspace;
4014 if (oldvm != newvm) {
4016 lp->lwp_vmspace = newvm;
4017 if (curthread->td_lwp == lp) {
4018 pmap = vmspace_pmap(newvm);
4020 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4021 if (pmap->pm_active & CPUMASK_LOCK)
4022 pmap_interlock_wait(newvm);
4024 pmap->pm_active |= 1;
4026 #if defined(SWTCH_OPTIM_STATS)
4029 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4030 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
4031 load_cr3(curthread->td_pcb->pcb_cr3);
4032 pmap = vmspace_pmap(oldvm);
4034 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4036 pmap->pm_active &= ~(cpumask_t)1;
4046 * Called when switching to a locked pmap, used to interlock against pmaps
4047 * undergoing modifications to prevent us from activating the MMU for the
4048 * target pmap until all such modifications have completed. We have to do
4049 * this because the thread making the modifications has already set up its
4050 * SMP synchronization mask.
4055 pmap_interlock_wait(struct vmspace *vm)
4057 struct pmap *pmap = &vm->vm_pmap;
4059 if (pmap->pm_active & CPUMASK_LOCK) {
4061 DEBUG_PUSH_INFO("pmap_interlock_wait");
4062 while (pmap->pm_active & CPUMASK_LOCK) {
4064 lwkt_process_ipiq();
4074 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4077 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4081 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4086 * Used by kmalloc/kfree, page already exists at va
4089 pmap_kvtom(vm_offset_t va)
4091 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));