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>
106 #include <machine/cputypes.h>
107 #include <machine/md_var.h>
108 #include <machine/specialreg.h>
109 #include <machine/smp.h>
110 #include <machine_base/apic/apicreg.h>
111 #include <machine/globaldata.h>
112 #include <machine/pmap.h>
113 #include <machine/pmap_inval.h>
117 #define PMAP_KEEP_PDIRS
118 #ifndef PMAP_SHPGPERPROC
119 #define PMAP_SHPGPERPROC 200
122 #if defined(DIAGNOSTIC)
123 #define PMAP_DIAGNOSTIC
129 * Get PDEs and PTEs for user/kernel address space
131 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
132 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
134 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
135 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
136 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
137 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
138 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
142 * Given a map and a machine independent protection code,
143 * convert to a vax protection code.
145 #define pte_prot(m, p) \
146 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
147 static int protection_codes[8];
149 struct pmap kernel_pmap;
150 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
152 vm_paddr_t avail_start; /* PA of first available physical page */
153 vm_paddr_t avail_end; /* PA of last available physical page */
154 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
155 vm_offset_t virtual2_end;
156 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
157 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
158 vm_offset_t KvaStart; /* VA start of KVA space */
159 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
160 vm_offset_t KvaSize; /* max size of kernel virtual address space */
161 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
162 static int pgeflag; /* PG_G or-in */
163 static int pseflag; /* PG_PS or-in */
165 static vm_object_t kptobj;
168 static vm_paddr_t dmaplimit;
170 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
172 static uint64_t KPTbase;
173 static uint64_t KPTphys;
174 static uint64_t KPDphys; /* phys addr of kernel level 2 */
175 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
176 uint64_t KPDPphys; /* phys addr of kernel level 3 */
177 uint64_t KPML4phys; /* phys addr of kernel level 4 */
179 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
180 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
183 * Data for the pv entry allocation mechanism
185 static vm_zone_t pvzone;
186 static struct vm_zone pvzone_store;
187 static struct vm_object pvzone_obj;
188 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
189 static int pmap_pagedaemon_waken = 0;
190 static struct pv_entry *pvinit;
193 * All those kernel PT submaps that BSD is so fond of
195 pt_entry_t *CMAP1 = 0, *ptmmap;
196 caddr_t CADDR1 = 0, ptvmmap = 0;
197 static pt_entry_t *msgbufmap;
198 struct msgbuf *msgbufp=0;
203 static pt_entry_t *pt_crashdumpmap;
204 static caddr_t crashdumpmap;
208 static pv_entry_t get_pv_entry (void);
209 static void i386_protection_init (void);
210 static void create_pagetables(vm_paddr_t *firstaddr);
211 static void pmap_remove_all (vm_page_t m);
212 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
213 vm_offset_t sva, pmap_inval_info_t info);
214 static void pmap_remove_page (struct pmap *pmap,
215 vm_offset_t va, pmap_inval_info_t info);
216 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
217 vm_offset_t va, pmap_inval_info_t info);
218 static boolean_t pmap_testbit (vm_page_t m, int bit);
219 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
220 vm_page_t mpte, vm_page_t m);
222 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
224 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
225 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
226 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
227 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
228 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
229 pmap_inval_info_t info);
230 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
231 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
233 static unsigned pdir4mb;
236 * Move the kernel virtual free pointer to the next
237 * 2MB. This is used to help improve performance
238 * by using a large (2MB) page for much of the kernel
239 * (.text, .data, .bss)
243 pmap_kmem_choose(vm_offset_t addr)
245 vm_offset_t newaddr = addr;
247 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
254 * Super fast pmap_pte routine best used when scanning the pv lists.
255 * This eliminates many course-grained invltlb calls. Note that many of
256 * the pv list scans are across different pmaps and it is very wasteful
257 * to do an entire invltlb when checking a single mapping.
259 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
263 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
265 return pmap_pte(pmap, va);
268 /* Return a non-clipped PD index for a given VA */
271 pmap_pde_pindex(vm_offset_t va)
273 return va >> PDRSHIFT;
276 /* Return various clipped indexes for a given VA */
279 pmap_pte_index(vm_offset_t va)
282 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
287 pmap_pde_index(vm_offset_t va)
290 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
295 pmap_pdpe_index(vm_offset_t va)
298 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
303 pmap_pml4e_index(vm_offset_t va)
306 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
309 /* Return a pointer to the PML4 slot that corresponds to a VA */
312 pmap_pml4e(pmap_t pmap, vm_offset_t va)
315 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
318 /* Return a pointer to the PDP slot that corresponds to a VA */
321 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
325 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
326 return (&pdpe[pmap_pdpe_index(va)]);
329 /* Return a pointer to the PDP slot that corresponds to a VA */
332 pmap_pdpe(pmap_t pmap, vm_offset_t va)
336 pml4e = pmap_pml4e(pmap, va);
337 if ((*pml4e & PG_V) == 0)
339 return (pmap_pml4e_to_pdpe(pml4e, va));
342 /* Return a pointer to the PD slot that corresponds to a VA */
345 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
349 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
350 return (&pde[pmap_pde_index(va)]);
353 /* Return a pointer to the PD slot that corresponds to a VA */
356 pmap_pde(pmap_t pmap, vm_offset_t va)
360 pdpe = pmap_pdpe(pmap, va);
361 if (pdpe == NULL || (*pdpe & PG_V) == 0)
363 return (pmap_pdpe_to_pde(pdpe, va));
366 /* Return a pointer to the PT slot that corresponds to a VA */
369 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
373 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
374 return (&pte[pmap_pte_index(va)]);
377 /* Return a pointer to the PT slot that corresponds to a VA */
380 pmap_pte(pmap_t pmap, vm_offset_t va)
384 pde = pmap_pde(pmap, va);
385 if (pde == NULL || (*pde & PG_V) == 0)
387 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
388 return ((pt_entry_t *)pde);
389 return (pmap_pde_to_pte(pde, va));
394 vtopte(vm_offset_t va)
396 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
398 return (PTmap + ((va >> PAGE_SHIFT) & mask));
403 vtopde(vm_offset_t va)
405 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
407 return (PDmap + ((va >> PDRSHIFT) & mask));
411 allocpages(vm_paddr_t *firstaddr, long n)
416 bzero((void *)ret, n * PAGE_SIZE);
417 *firstaddr += n * PAGE_SIZE;
423 create_pagetables(vm_paddr_t *firstaddr)
425 long i; /* must be 64 bits */
430 * We are running (mostly) V=P at this point
432 * Calculate NKPT - number of kernel page tables. We have to
433 * accomodoate prealloction of the vm_page_array, dump bitmap,
434 * MSGBUF_SIZE, and other stuff. Be generous.
436 * Maxmem is in pages.
438 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
439 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
443 * Starting at the beginning of kvm (not KERNBASE).
445 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
446 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
447 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + ndmpdp) +
452 * Starting at KERNBASE - map 2G worth of page table pages.
453 * KERNBASE is offset -2G from the end of kvm.
455 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
460 KPTbase = allocpages(firstaddr, nkpt_base);
461 KPTphys = allocpages(firstaddr, nkpt_phys);
462 KPML4phys = allocpages(firstaddr, 1);
463 KPDPphys = allocpages(firstaddr, NKPML4E);
464 KPDphys = allocpages(firstaddr, NKPDPE);
467 * Calculate the page directory base for KERNBASE,
468 * that is where we start populating the page table pages.
469 * Basically this is the end - 2.
471 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
473 DMPDPphys = allocpages(firstaddr, NDMPML4E);
474 if ((amd_feature & AMDID_PAGE1GB) == 0)
475 DMPDphys = allocpages(firstaddr, ndmpdp);
476 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
479 * Fill in the underlying page table pages for the area around
480 * KERNBASE. This remaps low physical memory to KERNBASE.
482 * Read-only from zero to physfree
483 * XXX not fully used, underneath 2M pages
485 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
486 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
487 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
491 * Now map the initial kernel page tables. One block of page
492 * tables is placed at the beginning of kernel virtual memory,
493 * and another block is placed at KERNBASE to map the kernel binary,
494 * data, bss, and initial pre-allocations.
496 for (i = 0; i < nkpt_base; i++) {
497 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
498 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
500 for (i = 0; i < nkpt_phys; i++) {
501 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
502 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
506 * Map from zero to end of allocations using 2M pages as an
507 * optimization. This will bypass some of the KPTBase pages
508 * above in the KERNBASE area.
510 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
511 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
512 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
516 * And connect up the PD to the PDP. The kernel pmap is expected
517 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
519 for (i = 0; i < NKPDPE; i++) {
520 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
521 KPDphys + (i << PAGE_SHIFT);
522 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
526 /* Now set up the direct map space using either 2MB or 1GB pages */
527 /* Preset PG_M and PG_A because demotion expects it */
528 if ((amd_feature & AMDID_PAGE1GB) == 0) {
529 for (i = 0; i < NPDEPG * ndmpdp; i++) {
530 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
531 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
534 /* And the direct map space's PDP */
535 for (i = 0; i < ndmpdp; i++) {
536 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
538 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
541 for (i = 0; i < ndmpdp; i++) {
542 ((pdp_entry_t *)DMPDPphys)[i] =
543 (vm_paddr_t)i << PDPSHIFT;
544 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
549 /* And recursively map PML4 to itself in order to get PTmap */
550 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
551 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
553 /* Connect the Direct Map slot up to the PML4 */
554 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
555 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
557 /* Connect the KVA slot up to the PML4 */
558 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
559 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
563 * Bootstrap the system enough to run with virtual memory.
565 * On the i386 this is called after mapping has already been enabled
566 * and just syncs the pmap module with what has already been done.
567 * [We can't call it easily with mapping off since the kernel is not
568 * mapped with PA == VA, hence we would have to relocate every address
569 * from the linked base (virtual) address "KERNBASE" to the actual
570 * (physical) address starting relative to 0]
573 pmap_bootstrap(vm_paddr_t *firstaddr)
577 struct mdglobaldata *gd;
580 KvaStart = VM_MIN_KERNEL_ADDRESS;
581 KvaEnd = VM_MAX_KERNEL_ADDRESS;
582 KvaSize = KvaEnd - KvaStart;
584 avail_start = *firstaddr;
587 * Create an initial set of page tables to run the kernel in.
589 create_pagetables(firstaddr);
591 virtual2_start = KvaStart;
592 virtual2_end = PTOV_OFFSET;
594 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
595 virtual_start = pmap_kmem_choose(virtual_start);
597 virtual_end = VM_MAX_KERNEL_ADDRESS;
599 /* XXX do %cr0 as well */
600 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
604 * Initialize protection array.
606 i386_protection_init();
609 * The kernel's pmap is statically allocated so we don't have to use
610 * pmap_create, which is unlikely to work correctly at this part of
611 * the boot sequence (XXX and which no longer exists).
613 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
614 kernel_pmap.pm_count = 1;
615 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
616 TAILQ_INIT(&kernel_pmap.pm_pvlist);
619 * Reserve some special page table entries/VA space for temporary
622 #define SYSMAP(c, p, v, n) \
623 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
629 * CMAP1/CMAP2 are used for zeroing and copying pages.
631 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
636 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
639 * ptvmmap is used for reading arbitrary physical pages via
642 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
645 * msgbufp is used to map the system message buffer.
646 * XXX msgbufmap is not used.
648 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
649 atop(round_page(MSGBUF_SIZE)))
656 * PG_G is terribly broken on SMP because we IPI invltlb's in some
657 * cases rather then invl1pg. Actually, I don't even know why it
658 * works under UP because self-referential page table mappings
663 if (cpu_feature & CPUID_PGE)
668 * Initialize the 4MB page size flag
672 * The 4MB page version of the initial
673 * kernel page mapping.
677 #if !defined(DISABLE_PSE)
678 if (cpu_feature & CPUID_PSE) {
681 * Note that we have enabled PSE mode
684 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
685 ptditmp &= ~(NBPDR - 1);
686 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
691 * Enable the PSE mode. If we are SMP we can't do this
692 * now because the APs will not be able to use it when
695 load_cr4(rcr4() | CR4_PSE);
698 * We can do the mapping here for the single processor
699 * case. We simply ignore the old page table page from
703 * For SMP, we still need 4K pages to bootstrap APs,
704 * PSE will be enabled as soon as all APs are up.
706 PTD[KPTDI] = (pd_entry_t)ptditmp;
713 * We need to finish setting up the globaldata page for the BSP.
714 * locore has already populated the page table for the mdglobaldata
717 pg = MDGLOBALDATA_BASEALLOC_PAGES;
718 gd = &CPU_prvspace[0].mdglobaldata;
725 * Set 4mb pdir for mp startup
730 if (pseflag && (cpu_feature & CPUID_PSE)) {
731 load_cr4(rcr4() | CR4_PSE);
732 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
740 * Initialize the pmap module.
741 * Called by vm_init, to initialize any structures that the pmap
742 * system needs to map virtual memory.
743 * pmap_init has been enhanced to support in a fairly consistant
744 * way, discontiguous physical memory.
753 * object for kernel page table pages
755 /* JG I think the number can be arbitrary */
756 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
759 * Allocate memory for random pmap data structures. Includes the
763 for(i = 0; i < vm_page_array_size; i++) {
766 m = &vm_page_array[i];
767 TAILQ_INIT(&m->md.pv_list);
768 m->md.pv_list_count = 0;
772 * init the pv free list
774 initial_pvs = vm_page_array_size;
775 if (initial_pvs < MINPV)
777 pvzone = &pvzone_store;
778 pvinit = (void *)kmem_alloc(&kernel_map,
779 initial_pvs * sizeof (struct pv_entry));
780 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
781 pvinit, initial_pvs);
784 * Now it is safe to enable pv_table recording.
786 pmap_initialized = TRUE;
790 * Initialize the address space (zone) for the pv_entries. Set a
791 * high water mark so that the system can recover from excessive
792 * numbers of pv entries.
797 int shpgperproc = PMAP_SHPGPERPROC;
800 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
801 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
802 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
803 pv_entry_high_water = 9 * (pv_entry_max / 10);
806 * Subtract out pages already installed in the zone (hack)
808 entry_max = pv_entry_max - vm_page_array_size;
812 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
816 /***************************************************
817 * Low level helper routines.....
818 ***************************************************/
820 #if defined(PMAP_DIAGNOSTIC)
823 * This code checks for non-writeable/modified pages.
824 * This should be an invalid condition.
828 pmap_nw_modified(pt_entry_t pte)
830 if ((pte & (PG_M|PG_RW)) == PG_M)
839 * this routine defines the region(s) of memory that should
840 * not be tested for the modified bit.
844 pmap_track_modified(vm_offset_t va)
846 if ((va < clean_sva) || (va >= clean_eva))
853 * Extract the physical page address associated with the map/VA pair.
855 * The caller must hold vm_token if non-blocking operation is desired.
858 pmap_extract(pmap_t pmap, vm_offset_t va)
862 pd_entry_t pde, *pdep;
864 lwkt_gettoken(&vm_token);
866 pdep = pmap_pde(pmap, va);
870 if ((pde & PG_PS) != 0) {
871 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
873 pte = pmap_pde_to_pte(pdep, va);
874 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
878 lwkt_reltoken(&vm_token);
883 * Extract the physical page address associated kernel virtual address.
886 pmap_kextract(vm_offset_t va)
891 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
892 pa = DMAP_TO_PHYS(va);
896 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
899 * Beware of a concurrent promotion that changes the
900 * PDE at this point! For example, vtopte() must not
901 * be used to access the PTE because it would use the
902 * new PDE. It is, however, safe to use the old PDE
903 * because the page table page is preserved by the
906 pa = *pmap_pde_to_pte(&pde, va);
907 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
913 /***************************************************
914 * Low level mapping routines.....
915 ***************************************************/
918 * Routine: pmap_kenter
920 * Add a wired page to the KVA
921 * NOTE! note that in order for the mapping to take effect -- you
922 * should do an invltlb after doing the pmap_kenter().
925 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
929 pmap_inval_info info;
931 pmap_inval_init(&info);
932 npte = pa | PG_RW | PG_V | pgeflag;
934 pmap_inval_interlock(&info, &kernel_pmap, va);
936 pmap_inval_deinterlock(&info, &kernel_pmap);
937 pmap_inval_done(&info);
941 * Routine: pmap_kenter_quick
943 * Similar to pmap_kenter(), except we only invalidate the
944 * mapping on the current CPU.
947 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
952 npte = pa | PG_RW | PG_V | pgeflag;
955 cpu_invlpg((void *)va);
959 pmap_kenter_sync(vm_offset_t va)
961 pmap_inval_info info;
963 pmap_inval_init(&info);
964 pmap_inval_interlock(&info, &kernel_pmap, va);
965 pmap_inval_deinterlock(&info, &kernel_pmap);
966 pmap_inval_done(&info);
970 pmap_kenter_sync_quick(vm_offset_t va)
972 cpu_invlpg((void *)va);
976 * remove a page from the kernel pagetables
979 pmap_kremove(vm_offset_t va)
982 pmap_inval_info info;
984 pmap_inval_init(&info);
986 pmap_inval_interlock(&info, &kernel_pmap, va);
988 pmap_inval_deinterlock(&info, &kernel_pmap);
989 pmap_inval_done(&info);
993 pmap_kremove_quick(vm_offset_t va)
998 cpu_invlpg((void *)va);
1002 * XXX these need to be recoded. They are not used in any critical path.
1005 pmap_kmodify_rw(vm_offset_t va)
1007 *vtopte(va) |= PG_RW;
1008 cpu_invlpg((void *)va);
1012 pmap_kmodify_nc(vm_offset_t va)
1014 *vtopte(va) |= PG_N;
1015 cpu_invlpg((void *)va);
1019 * Used to map a range of physical addresses into kernel virtual
1020 * address space during the low level boot, typically to map the
1021 * dump bitmap, message buffer, and vm_page_array.
1023 * These mappings are typically made at some pointer after the end of the
1026 * We could return PHYS_TO_DMAP(start) here and not allocate any
1027 * via (*virtp), but then kmem from userland and kernel dumps won't
1028 * have access to the related pointers.
1031 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1034 vm_offset_t va_start;
1036 /*return PHYS_TO_DMAP(start);*/
1041 while (start < end) {
1042 pmap_kenter_quick(va, start);
1052 * Add a list of wired pages to the kva
1053 * this routine is only used for temporary
1054 * kernel mappings that do not need to have
1055 * page modification or references recorded.
1056 * Note that old mappings are simply written
1057 * over. The page *must* be wired.
1060 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1064 end_va = va + count * PAGE_SIZE;
1066 while (va < end_va) {
1070 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1071 cpu_invlpg((void *)va);
1079 * This routine jerks page mappings from the
1080 * kernel -- it is meant only for temporary mappings.
1082 * MPSAFE, INTERRUPT SAFE (cluster callback)
1085 pmap_qremove(vm_offset_t va, int count)
1089 end_va = va + count * PAGE_SIZE;
1091 while (va < end_va) {
1096 cpu_invlpg((void *)va);
1103 * This routine works like vm_page_lookup() but also blocks as long as the
1104 * page is busy. This routine does not busy the page it returns.
1106 * Unless the caller is managing objects whos pages are in a known state,
1107 * the call should be made with both vm_token held and the governing object
1108 * and its token held so the page's object association remains valid on
1111 * This function can block!
1115 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1120 m = vm_page_lookup(object, pindex);
1121 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1127 * Create a new thread and optionally associate it with a (new) process.
1128 * NOTE! the new thread's cpu may not equal the current cpu.
1131 pmap_init_thread(thread_t td)
1133 /* enforce pcb placement & alignment */
1134 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1135 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1136 td->td_savefpu = &td->td_pcb->pcb_save;
1137 td->td_sp = (char *)td->td_pcb; /* no -16 */
1141 * This routine directly affects the fork perf for a process.
1144 pmap_init_proc(struct proc *p)
1149 * Dispose the UPAGES for a process that has exited.
1150 * This routine directly impacts the exit perf of a process.
1153 pmap_dispose_proc(struct proc *p)
1155 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1158 /***************************************************
1159 * Page table page management routines.....
1160 ***************************************************/
1163 * This routine unholds page table pages, and if the hold count
1164 * drops to zero, then it decrements the wire count.
1168 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1169 pmap_inval_info_t info)
1171 KKASSERT(m->hold_count > 0);
1172 if (m->hold_count > 1) {
1176 return _pmap_unwire_pte_hold(pmap, va, m, info);
1182 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1183 pmap_inval_info_t info)
1186 * Wait until we can busy the page ourselves. We cannot have
1187 * any active flushes if we block. We own one hold count on the
1188 * page so it cannot be freed out from under us.
1190 if (m->flags & PG_BUSY) {
1191 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1194 KASSERT(m->queue == PQ_NONE,
1195 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1198 * This case can occur if new references were acquired while
1201 if (m->hold_count > 1) {
1202 KKASSERT(m->hold_count > 1);
1208 * Unmap the page table page
1210 KKASSERT(m->hold_count == 1);
1212 pmap_inval_interlock(info, pmap, -1);
1214 if (m->pindex >= (NUPDE + NUPDPE)) {
1217 pml4 = pmap_pml4e(pmap, va);
1219 } else if (m->pindex >= NUPDE) {
1222 pdp = pmap_pdpe(pmap, va);
1227 pd = pmap_pde(pmap, va);
1231 KKASSERT(pmap->pm_stats.resident_count > 0);
1232 --pmap->pm_stats.resident_count;
1234 if (pmap->pm_ptphint == m)
1235 pmap->pm_ptphint = NULL;
1236 pmap_inval_deinterlock(info, pmap);
1238 if (m->pindex < NUPDE) {
1239 /* We just released a PT, unhold the matching PD */
1242 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1243 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1245 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1246 /* We just released a PD, unhold the matching PDP */
1249 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1250 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1254 * This was our last hold, the page had better be unwired
1255 * after we decrement wire_count.
1257 * FUTURE NOTE: shared page directory page could result in
1258 * multiple wire counts.
1262 KKASSERT(m->wire_count == 0);
1263 --vmstats.v_wire_count;
1264 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1266 vm_page_free_zero(m);
1272 * After removing a page table entry, this routine is used to
1273 * conditionally free the page, and manage the hold/wire counts.
1277 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1278 pmap_inval_info_t info)
1280 vm_pindex_t ptepindex;
1282 if (va >= VM_MAX_USER_ADDRESS)
1286 ptepindex = pmap_pde_pindex(va);
1288 if (pmap->pm_ptphint &&
1289 (pmap->pm_ptphint->pindex == ptepindex)) {
1290 mpte = pmap->pm_ptphint;
1293 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1294 pmap->pm_ptphint = mpte;
1299 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1303 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1304 * it, and IdlePTD, represents the template used to update all other pmaps.
1306 * On architectures where the kernel pmap is not integrated into the user
1307 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1308 * kernel_pmap should be used to directly access the kernel_pmap.
1311 pmap_pinit0(struct pmap *pmap)
1313 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1315 pmap->pm_active = 0;
1316 pmap->pm_ptphint = NULL;
1317 TAILQ_INIT(&pmap->pm_pvlist);
1318 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1322 * Initialize a preallocated and zeroed pmap structure,
1323 * such as one in a vmspace structure.
1326 pmap_pinit(struct pmap *pmap)
1331 * No need to allocate page table space yet but we do need a valid
1332 * page directory table.
1334 if (pmap->pm_pml4 == NULL) {
1336 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1340 * Allocate an object for the ptes
1342 if (pmap->pm_pteobj == NULL)
1343 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1346 * Allocate the page directory page, unless we already have
1347 * one cached. If we used the cached page the wire_count will
1348 * already be set appropriately.
1350 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1351 ptdpg = vm_page_grab(pmap->pm_pteobj,
1352 NUPDE + NUPDPE + PML4PML4I,
1353 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1354 pmap->pm_pdirm = ptdpg;
1355 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1356 ptdpg->valid = VM_PAGE_BITS_ALL;
1357 if (ptdpg->wire_count == 0)
1358 ++vmstats.v_wire_count;
1359 ptdpg->wire_count = 1;
1360 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1362 if ((ptdpg->flags & PG_ZERO) == 0)
1363 bzero(pmap->pm_pml4, PAGE_SIZE);
1366 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1369 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1370 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1372 /* install self-referential address mapping entry */
1373 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1376 pmap->pm_active = 0;
1377 pmap->pm_ptphint = NULL;
1378 TAILQ_INIT(&pmap->pm_pvlist);
1379 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1380 pmap->pm_stats.resident_count = 1;
1384 * Clean up a pmap structure so it can be physically freed. This routine
1385 * is called by the vmspace dtor function. A great deal of pmap data is
1386 * left passively mapped to improve vmspace management so we have a bit
1387 * of cleanup work to do here.
1390 pmap_puninit(pmap_t pmap)
1394 KKASSERT(pmap->pm_active == 0);
1395 lwkt_gettoken(&vm_token);
1396 if ((p = pmap->pm_pdirm) != NULL) {
1397 KKASSERT(pmap->pm_pml4 != NULL);
1398 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1399 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1401 vmstats.v_wire_count--;
1402 KKASSERT((p->flags & PG_BUSY) == 0);
1404 vm_page_free_zero(p);
1405 pmap->pm_pdirm = NULL;
1407 if (pmap->pm_pml4) {
1408 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1409 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1410 pmap->pm_pml4 = NULL;
1412 if (pmap->pm_pteobj) {
1413 vm_object_deallocate(pmap->pm_pteobj);
1414 pmap->pm_pteobj = NULL;
1416 lwkt_reltoken(&vm_token);
1420 * Wire in kernel global address entries. To avoid a race condition
1421 * between pmap initialization and pmap_growkernel, this procedure
1422 * adds the pmap to the master list (which growkernel scans to update),
1423 * then copies the template.
1426 pmap_pinit2(struct pmap *pmap)
1428 lwkt_gettoken(&vm_token);
1429 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1430 /* XXX copies current process, does not fill in MPPTDI */
1431 lwkt_reltoken(&vm_token);
1435 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1436 * 0 on failure (if the procedure had to sleep).
1438 * When asked to remove the page directory page itself, we actually just
1439 * leave it cached so we do not have to incur the SMP inval overhead of
1440 * removing the kernel mapping. pmap_puninit() will take care of it.
1444 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1447 * This code optimizes the case of freeing non-busy
1448 * page-table pages. Those pages are zero now, and
1449 * might as well be placed directly into the zero queue.
1451 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1457 * Remove the page table page from the processes address space.
1459 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1461 * We are the pml4 table itself.
1463 /* XXX anything to do here? */
1464 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1466 * Remove a PDP page from the PML4. We do not maintain
1467 * hold counts on the PML4 page.
1473 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1474 KKASSERT(m4 != NULL);
1475 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1476 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1477 KKASSERT(pml4[idx] != 0);
1479 } else if (p->pindex >= NUPDE) {
1481 * Remove a PD page from the PDP and drop the hold count
1482 * on the PDP. The PDP is left cached in the pmap if
1483 * the hold count drops to 0 so the wire count remains
1490 m3 = vm_page_lookup(pmap->pm_pteobj,
1491 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1492 KKASSERT(m3 != NULL);
1493 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1494 idx = (p->pindex - NUPDE) % NPDPEPG;
1495 KKASSERT(pdp[idx] != 0);
1500 * Remove a PT page from the PD and drop the hold count
1501 * on the PD. The PD is left cached in the pmap if
1502 * the hold count drops to 0 so the wire count remains
1509 m2 = vm_page_lookup(pmap->pm_pteobj,
1510 NUPDE + p->pindex / NPDEPG);
1511 KKASSERT(m2 != NULL);
1512 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1513 idx = p->pindex % NPDEPG;
1519 * One fewer mappings in the pmap. p's hold count had better
1522 KKASSERT(pmap->pm_stats.resident_count > 0);
1523 --pmap->pm_stats.resident_count;
1525 panic("pmap_release: freeing held page table page");
1526 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1527 pmap->pm_ptphint = NULL;
1530 * We leave the top-level page table page cached, wired, and mapped in
1531 * the pmap until the dtor function (pmap_puninit()) gets called.
1532 * However, still clean it up so we can set PG_ZERO.
1534 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1535 bzero(pmap->pm_pml4, PAGE_SIZE);
1536 vm_page_flag_set(p, PG_ZERO);
1540 KKASSERT(p->wire_count == 0);
1541 vmstats.v_wire_count--;
1542 /* JG eventually revert to using vm_page_free_zero() */
1549 * This routine is called when various levels in the page table need to
1550 * be populated. This routine cannot fail.
1554 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1559 * Find or fabricate a new pagetable page. This will busy the page.
1561 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1562 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1565 * The grab may have blocked and raced another thread populating
1566 * the same page table. m->valid will be 0 on a newly allocated page
1567 * so use this to determine if we have to zero it out or not. We
1568 * don't want to zero-out a raced page as this would desynchronize
1569 * the pv_entry's for the related pte's and cause pmap_remove_all()
1572 if (m->valid == 0) {
1573 if ((m->flags & PG_ZERO) == 0) {
1574 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1578 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1581 m->valid = VM_PAGE_BITS_ALL;
1582 vm_page_flag_clear(m, PG_ZERO);
1586 KKASSERT((m->flags & PG_ZERO) == 0);
1590 KASSERT(m->queue == PQ_NONE,
1591 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1594 * Increment the hold count for the page we will be returning to
1598 if (m->wire_count++ == 0)
1599 vmstats.v_wire_count++;
1602 * Map the pagetable page into the process address space, if
1603 * it isn't already there.
1605 * It is possible that someone else got in and mapped the page
1606 * directory page while we were blocked, if so just unbusy and
1607 * return the held page.
1609 if (ptepindex >= (NUPDE + NUPDPE)) {
1611 * Wire up a new PDP page in the PML4
1613 vm_pindex_t pml4index;
1616 pml4index = ptepindex - (NUPDE + NUPDPE);
1617 pml4 = &pmap->pm_pml4[pml4index];
1619 if (--m->wire_count == 0)
1620 --vmstats.v_wire_count;
1624 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1625 } else if (ptepindex >= NUPDE) {
1627 * Wire up a new PD page in the PDP
1629 vm_pindex_t pml4index;
1630 vm_pindex_t pdpindex;
1635 pdpindex = ptepindex - NUPDE;
1636 pml4index = pdpindex >> NPML4EPGSHIFT;
1638 pml4 = &pmap->pm_pml4[pml4index];
1639 if ((*pml4 & PG_V) == 0) {
1641 * Have to allocate a new PDP page, recurse.
1642 * This always succeeds. Returned page will
1645 pdppg = _pmap_allocpte(pmap,
1646 NUPDE + NUPDPE + pml4index);
1649 * Add a held reference to the PDP page.
1651 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1652 pdppg->hold_count++;
1656 * Now find the pdp_entry and map the PDP. If the PDP
1657 * has already been mapped unwind and return the
1658 * already-mapped PDP held.
1660 * pdppg is left held (hold_count is incremented for
1661 * each PD in the PDP).
1663 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1664 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1666 vm_page_unhold(pdppg);
1667 if (--m->wire_count == 0)
1668 --vmstats.v_wire_count;
1672 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1675 * Wire up the new PT page in the PD
1677 vm_pindex_t pml4index;
1678 vm_pindex_t pdpindex;
1684 pdpindex = ptepindex >> NPDPEPGSHIFT;
1685 pml4index = pdpindex >> NPML4EPGSHIFT;
1688 * Locate the PDP page in the PML4, then the PD page in
1689 * the PDP. If either does not exist we simply recurse
1692 * We can just recurse on the PD page as it will recurse
1693 * on the PDP if necessary.
1695 pml4 = &pmap->pm_pml4[pml4index];
1696 if ((*pml4 & PG_V) == 0) {
1697 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1698 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1699 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1701 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1702 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1703 if ((*pdp & PG_V) == 0) {
1704 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1706 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1712 * Now fill in the pte in the PD. If the pte already exists
1713 * (again, if we raced the grab), unhold pdpg and unwire
1714 * m, returning a held m.
1716 * pdpg is left held (hold_count is incremented for
1717 * each PT in the PD).
1719 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1720 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1722 vm_page_unhold(pdpg);
1723 if (--m->wire_count == 0)
1724 --vmstats.v_wire_count;
1728 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1732 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1733 * valid bits, mapped flag, unbusy, and we're done.
1735 pmap->pm_ptphint = m;
1736 ++pmap->pm_stats.resident_count;
1739 m->valid = VM_PAGE_BITS_ALL;
1740 vm_page_flag_clear(m, PG_ZERO);
1742 vm_page_flag_set(m, PG_MAPPED);
1750 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1752 vm_pindex_t ptepindex;
1757 * Calculate pagetable page index
1759 ptepindex = pmap_pde_pindex(va);
1762 * Get the page directory entry
1764 pd = pmap_pde(pmap, va);
1767 * This supports switching from a 2MB page to a
1770 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1771 panic("no promotion/demotion yet");
1779 * If the page table page is mapped, we just increment the
1780 * hold count, and activate it.
1782 if (pd != NULL && (*pd & PG_V) != 0) {
1783 /* YYY hint is used here on i386 */
1784 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1785 pmap->pm_ptphint = m;
1790 * Here if the pte page isn't mapped, or if it has been deallocated.
1792 return _pmap_allocpte(pmap, ptepindex);
1796 /***************************************************
1797 * Pmap allocation/deallocation routines.
1798 ***************************************************/
1801 * Release any resources held by the given physical map.
1802 * Called when a pmap initialized by pmap_pinit is being released.
1803 * Should only be called if the map contains no valid mappings.
1805 static int pmap_release_callback(struct vm_page *p, void *data);
1808 pmap_release(struct pmap *pmap)
1810 vm_object_t object = pmap->pm_pteobj;
1811 struct rb_vm_page_scan_info info;
1813 KASSERT(pmap->pm_active == 0,
1814 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1815 #if defined(DIAGNOSTIC)
1816 if (object->ref_count != 1)
1817 panic("pmap_release: pteobj reference count != 1");
1821 info.object = object;
1822 vm_object_hold(object);
1823 lwkt_gettoken(&vm_token);
1824 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1829 info.limit = object->generation;
1831 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1832 pmap_release_callback, &info);
1833 if (info.error == 0 && info.mpte) {
1834 if (!pmap_release_free_page(pmap, info.mpte))
1837 } while (info.error);
1838 lwkt_reltoken(&vm_token);
1839 vm_object_drop(object);
1844 pmap_release_callback(struct vm_page *p, void *data)
1846 struct rb_vm_page_scan_info *info = data;
1848 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1852 if (!pmap_release_free_page(info->pmap, p)) {
1856 if (info->object->generation != info->limit) {
1864 * Grow the number of kernel page table entries, if needed.
1866 * This routine is always called to validate any address space
1867 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1868 * space below KERNBASE.
1871 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1874 vm_offset_t ptppaddr;
1876 pd_entry_t *pde, newpdir;
1878 int update_kernel_vm_end;
1880 lwkt_gettoken(&vm_token);
1883 * bootstrap kernel_vm_end on first real VM use
1885 if (kernel_vm_end == 0) {
1886 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1888 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1889 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1890 ~(PAGE_SIZE * NPTEPG - 1);
1892 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1893 kernel_vm_end = kernel_map.max_offset;
1900 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1901 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1902 * do not want to force-fill 128G worth of page tables.
1904 if (kstart < KERNBASE) {
1905 if (kstart > kernel_vm_end)
1906 kstart = kernel_vm_end;
1907 KKASSERT(kend <= KERNBASE);
1908 update_kernel_vm_end = 1;
1910 update_kernel_vm_end = 0;
1913 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1914 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1916 if (kend - 1 >= kernel_map.max_offset)
1917 kend = kernel_map.max_offset;
1919 while (kstart < kend) {
1920 pde = pmap_pde(&kernel_pmap, kstart);
1922 /* We need a new PDP entry */
1923 nkpg = vm_page_alloc(kptobj, nkpt,
1926 VM_ALLOC_INTERRUPT);
1928 panic("pmap_growkernel: no memory to grow "
1931 paddr = VM_PAGE_TO_PHYS(nkpg);
1932 if ((nkpg->flags & PG_ZERO) == 0)
1933 pmap_zero_page(paddr);
1934 vm_page_flag_clear(nkpg, PG_ZERO);
1935 newpdp = (pdp_entry_t)
1936 (paddr | PG_V | PG_RW | PG_A | PG_M);
1937 *pmap_pdpe(&kernel_pmap, kstart) = newpdp;
1939 continue; /* try again */
1941 if ((*pde & PG_V) != 0) {
1942 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1943 ~(PAGE_SIZE * NPTEPG - 1);
1944 if (kstart - 1 >= kernel_map.max_offset) {
1945 kstart = kernel_map.max_offset;
1952 * This index is bogus, but out of the way
1954 nkpg = vm_page_alloc(kptobj, nkpt,
1957 VM_ALLOC_INTERRUPT);
1959 panic("pmap_growkernel: no memory to grow kernel");
1962 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1963 pmap_zero_page(ptppaddr);
1964 vm_page_flag_clear(nkpg, PG_ZERO);
1965 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1966 *pmap_pde(&kernel_pmap, kstart) = newpdir;
1969 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1970 ~(PAGE_SIZE * NPTEPG - 1);
1972 if (kstart - 1 >= kernel_map.max_offset) {
1973 kstart = kernel_map.max_offset;
1979 * Only update kernel_vm_end for areas below KERNBASE.
1981 if (update_kernel_vm_end && kernel_vm_end < kstart)
1982 kernel_vm_end = kstart;
1984 lwkt_reltoken(&vm_token);
1988 * Retire the given physical map from service.
1989 * Should only be called if the map contains
1990 * no valid mappings.
1993 pmap_destroy(pmap_t pmap)
2000 lwkt_gettoken(&vm_token);
2001 count = --pmap->pm_count;
2004 panic("destroying a pmap is not yet implemented");
2006 lwkt_reltoken(&vm_token);
2010 * Add a reference to the specified pmap.
2013 pmap_reference(pmap_t pmap)
2016 lwkt_gettoken(&vm_token);
2018 lwkt_reltoken(&vm_token);
2022 /***************************************************
2023 * page management routines.
2024 ***************************************************/
2027 * free the pv_entry back to the free list. This function may be
2028 * called from an interrupt.
2032 free_pv_entry(pv_entry_t pv)
2035 KKASSERT(pv_entry_count >= 0);
2040 * get a new pv_entry, allocating a block from the system
2041 * when needed. This function may be called from an interrupt.
2048 if (pv_entry_high_water &&
2049 (pv_entry_count > pv_entry_high_water) &&
2050 (pmap_pagedaemon_waken == 0)) {
2051 pmap_pagedaemon_waken = 1;
2052 wakeup(&vm_pages_needed);
2054 return zalloc(pvzone);
2058 * This routine is very drastic, but can save the system
2066 static int warningdone=0;
2068 if (pmap_pagedaemon_waken == 0)
2070 lwkt_gettoken(&vm_token);
2071 if (warningdone < 5) {
2072 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2076 for(i = 0; i < vm_page_array_size; i++) {
2077 m = &vm_page_array[i];
2078 if (m->wire_count || m->hold_count || m->busy ||
2079 (m->flags & PG_BUSY))
2083 pmap_pagedaemon_waken = 0;
2084 lwkt_reltoken(&vm_token);
2089 * If it is the first entry on the list, it is actually in the header and
2090 * we must copy the following entry up to the header.
2092 * Otherwise we must search the list for the entry. In either case we
2093 * free the now unused entry.
2095 * Caller must hold vm_token
2099 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2100 vm_offset_t va, pmap_inval_info_t info)
2105 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2106 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2107 if (pmap == pv->pv_pmap && va == pv->pv_va)
2111 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2112 if (va == pv->pv_va)
2120 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2121 m->md.pv_list_count--;
2122 m->object->agg_pv_list_count--;
2123 KKASSERT(m->md.pv_list_count >= 0);
2124 if (TAILQ_EMPTY(&m->md.pv_list))
2125 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2126 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2127 ++pmap->pm_generation;
2128 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2135 * Create a pv entry for page at pa for (pmap, va).
2137 * Caller must hold vm_token
2141 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2145 pv = get_pv_entry();
2150 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2151 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2152 ++pmap->pm_generation;
2153 m->md.pv_list_count++;
2154 m->object->agg_pv_list_count++;
2158 * pmap_remove_pte: do the things to unmap a page in a process
2160 * Caller must hold vm_token
2164 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2165 pmap_inval_info_t info)
2170 pmap_inval_interlock(info, pmap, va);
2171 oldpte = pte_load_clear(ptq);
2172 pmap_inval_deinterlock(info, pmap);
2174 pmap->pm_stats.wired_count -= 1;
2176 * Machines that don't support invlpg, also don't support
2177 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2181 cpu_invlpg((void *)va);
2182 KKASSERT(pmap->pm_stats.resident_count > 0);
2183 --pmap->pm_stats.resident_count;
2184 if (oldpte & PG_MANAGED) {
2185 m = PHYS_TO_VM_PAGE(oldpte);
2186 if (oldpte & PG_M) {
2187 #if defined(PMAP_DIAGNOSTIC)
2188 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2190 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2194 if (pmap_track_modified(va))
2198 vm_page_flag_set(m, PG_REFERENCED);
2199 return pmap_remove_entry(pmap, m, va, info);
2201 return pmap_unuse_pt(pmap, va, NULL, info);
2208 * Remove a single page from a process address space.
2210 * This function may not be called from an interrupt if the pmap is
2213 * Caller must hold vm_token
2217 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2221 pte = pmap_pte(pmap, va);
2224 if ((*pte & PG_V) == 0)
2226 pmap_remove_pte(pmap, pte, va, info);
2230 * Remove the given range of addresses from the specified map.
2232 * It is assumed that the start and end are properly rounded to the page size.
2234 * This function may not be called from an interrupt if the pmap is not
2238 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2240 vm_offset_t va_next;
2241 pml4_entry_t *pml4e;
2243 pd_entry_t ptpaddr, *pde;
2245 struct pmap_inval_info info;
2250 lwkt_gettoken(&vm_token);
2251 if (pmap->pm_stats.resident_count == 0) {
2252 lwkt_reltoken(&vm_token);
2256 pmap_inval_init(&info);
2259 * special handling of removing one page. a very
2260 * common operation and easy to short circuit some
2263 if (sva + PAGE_SIZE == eva) {
2264 pde = pmap_pde(pmap, sva);
2265 if (pde && (*pde & PG_PS) == 0) {
2266 pmap_remove_page(pmap, sva, &info);
2267 pmap_inval_done(&info);
2268 lwkt_reltoken(&vm_token);
2273 for (; sva < eva; sva = va_next) {
2274 pml4e = pmap_pml4e(pmap, sva);
2275 if ((*pml4e & PG_V) == 0) {
2276 va_next = (sva + NBPML4) & ~PML4MASK;
2282 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2283 if ((*pdpe & PG_V) == 0) {
2284 va_next = (sva + NBPDP) & ~PDPMASK;
2291 * Calculate index for next page table.
2293 va_next = (sva + NBPDR) & ~PDRMASK;
2297 pde = pmap_pdpe_to_pde(pdpe, sva);
2301 * Weed out invalid mappings.
2307 * Check for large page.
2309 if ((ptpaddr & PG_PS) != 0) {
2310 /* JG FreeBSD has more complex treatment here */
2311 pmap_inval_interlock(&info, pmap, -1);
2313 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2314 pmap_inval_deinterlock(&info, pmap);
2319 * Limit our scan to either the end of the va represented
2320 * by the current page table page, or to the end of the
2321 * range being removed.
2327 * NOTE: pmap_remove_pte() can block.
2329 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2333 if (pmap_remove_pte(pmap, pte, sva, &info))
2337 pmap_inval_done(&info);
2338 lwkt_reltoken(&vm_token);
2344 * Removes this physical page from all physical maps in which it resides.
2345 * Reflects back modify bits to the pager.
2347 * This routine may not be called from an interrupt.
2352 pmap_remove_all(vm_page_t m)
2354 struct pmap_inval_info info;
2355 pt_entry_t *pte, tpte;
2358 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2361 lwkt_gettoken(&vm_token);
2362 pmap_inval_init(&info);
2363 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2364 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2365 --pv->pv_pmap->pm_stats.resident_count;
2367 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2368 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2369 tpte = pte_load_clear(pte);
2370 KKASSERT(tpte & PG_MANAGED);
2372 pv->pv_pmap->pm_stats.wired_count--;
2373 pmap_inval_deinterlock(&info, pv->pv_pmap);
2375 vm_page_flag_set(m, PG_REFERENCED);
2378 * Update the vm_page_t clean and reference bits.
2381 #if defined(PMAP_DIAGNOSTIC)
2382 if (pmap_nw_modified(tpte)) {
2384 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2388 if (pmap_track_modified(pv->pv_va))
2391 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2392 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2393 ++pv->pv_pmap->pm_generation;
2394 m->md.pv_list_count--;
2395 m->object->agg_pv_list_count--;
2396 KKASSERT(m->md.pv_list_count >= 0);
2397 if (TAILQ_EMPTY(&m->md.pv_list))
2398 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2399 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2402 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2403 pmap_inval_done(&info);
2404 lwkt_reltoken(&vm_token);
2410 * Set the physical protection on the specified range of this map
2413 * This function may not be called from an interrupt if the map is
2414 * not the kernel_pmap.
2417 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2419 vm_offset_t va_next;
2420 pml4_entry_t *pml4e;
2422 pd_entry_t ptpaddr, *pde;
2424 pmap_inval_info info;
2426 /* JG review for NX */
2431 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2432 pmap_remove(pmap, sva, eva);
2436 if (prot & VM_PROT_WRITE)
2439 lwkt_gettoken(&vm_token);
2440 pmap_inval_init(&info);
2442 for (; sva < eva; sva = va_next) {
2444 pml4e = pmap_pml4e(pmap, sva);
2445 if ((*pml4e & PG_V) == 0) {
2446 va_next = (sva + NBPML4) & ~PML4MASK;
2452 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2453 if ((*pdpe & PG_V) == 0) {
2454 va_next = (sva + NBPDP) & ~PDPMASK;
2460 va_next = (sva + NBPDR) & ~PDRMASK;
2464 pde = pmap_pdpe_to_pde(pdpe, sva);
2468 * Check for large page.
2470 if ((ptpaddr & PG_PS) != 0) {
2471 pmap_inval_interlock(&info, pmap, -1);
2472 *pde &= ~(PG_M|PG_RW);
2473 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2474 pmap_inval_deinterlock(&info, pmap);
2479 * Weed out invalid mappings. Note: we assume that the page
2480 * directory table is always allocated, and in kernel virtual.
2488 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2497 pmap_inval_interlock(&info, pmap, sva);
2501 if ((pbits & PG_V) == 0) {
2502 pmap_inval_deinterlock(&info, pmap);
2505 if (pbits & PG_MANAGED) {
2508 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2509 vm_page_flag_set(m, PG_REFERENCED);
2513 if (pmap_track_modified(sva)) {
2515 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2522 if (pbits != cbits &&
2523 !atomic_cmpset_long(pte, pbits, cbits)) {
2526 pmap_inval_deinterlock(&info, pmap);
2529 pmap_inval_done(&info);
2530 lwkt_reltoken(&vm_token);
2534 * Insert the given physical page (p) at
2535 * the specified virtual address (v) in the
2536 * target physical map with the protection requested.
2538 * If specified, the page will be wired down, meaning
2539 * that the related pte can not be reclaimed.
2541 * NB: This is the only routine which MAY NOT lazy-evaluate
2542 * or lose information. That is, this routine must actually
2543 * insert this page into the given map NOW.
2546 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2553 pt_entry_t origpte, newpte;
2555 pmap_inval_info info;
2560 va = trunc_page(va);
2561 #ifdef PMAP_DIAGNOSTIC
2563 panic("pmap_enter: toobig");
2564 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2565 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2567 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2568 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2570 db_print_backtrace();
2573 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2574 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2576 db_print_backtrace();
2580 lwkt_gettoken(&vm_token);
2583 * In the case that a page table page is not
2584 * resident, we are creating it here.
2586 if (va < VM_MAX_USER_ADDRESS)
2587 mpte = pmap_allocpte(pmap, va);
2591 pmap_inval_init(&info);
2592 pde = pmap_pde(pmap, va);
2593 if (pde != NULL && (*pde & PG_V) != 0) {
2594 if ((*pde & PG_PS) != 0)
2595 panic("pmap_enter: attempted pmap_enter on 2MB page");
2596 pte = pmap_pde_to_pte(pde, va);
2598 panic("pmap_enter: invalid page directory va=%#lx", va);
2600 KKASSERT(pte != NULL);
2601 pa = VM_PAGE_TO_PHYS(m);
2603 opa = origpte & PG_FRAME;
2606 * Mapping has not changed, must be protection or wiring change.
2608 if (origpte && (opa == pa)) {
2610 * Wiring change, just update stats. We don't worry about
2611 * wiring PT pages as they remain resident as long as there
2612 * are valid mappings in them. Hence, if a user page is wired,
2613 * the PT page will be also.
2615 if (wired && ((origpte & PG_W) == 0))
2616 pmap->pm_stats.wired_count++;
2617 else if (!wired && (origpte & PG_W))
2618 pmap->pm_stats.wired_count--;
2620 #if defined(PMAP_DIAGNOSTIC)
2621 if (pmap_nw_modified(origpte)) {
2623 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2629 * Remove the extra pte reference. Note that we cannot
2630 * optimize the RO->RW case because we have adjusted the
2631 * wiring count above and may need to adjust the wiring
2638 * We might be turning off write access to the page,
2639 * so we go ahead and sense modify status.
2641 if (origpte & PG_MANAGED) {
2642 if ((origpte & PG_M) && pmap_track_modified(va)) {
2644 om = PHYS_TO_VM_PAGE(opa);
2648 KKASSERT(m->flags & PG_MAPPED);
2653 * Mapping has changed, invalidate old range and fall through to
2654 * handle validating new mapping.
2658 err = pmap_remove_pte(pmap, pte, va, &info);
2660 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2662 opa = origpte & PG_FRAME;
2664 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2670 * Enter on the PV list if part of our managed memory. Note that we
2671 * raise IPL while manipulating pv_table since pmap_enter can be
2672 * called at interrupt time.
2674 if (pmap_initialized &&
2675 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2676 pmap_insert_entry(pmap, va, mpte, m);
2678 vm_page_flag_set(m, PG_MAPPED);
2682 * Increment counters
2684 ++pmap->pm_stats.resident_count;
2686 pmap->pm_stats.wired_count++;
2690 * Now validate mapping with desired protection/wiring.
2692 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2696 if (va < VM_MAX_USER_ADDRESS)
2698 if (pmap == &kernel_pmap)
2702 * if the mapping or permission bits are different, we need
2703 * to update the pte.
2705 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2706 pmap_inval_interlock(&info, pmap, va);
2707 *pte = newpte | PG_A;
2708 pmap_inval_deinterlock(&info, pmap);
2710 vm_page_flag_set(m, PG_WRITEABLE);
2712 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2713 pmap_inval_done(&info);
2714 lwkt_reltoken(&vm_token);
2718 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2719 * This code also assumes that the pmap has no pre-existing entry for this
2722 * This code currently may only be used on user pmaps, not kernel_pmap.
2725 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2730 vm_pindex_t ptepindex;
2732 pmap_inval_info info;
2734 lwkt_gettoken(&vm_token);
2735 pmap_inval_init(&info);
2737 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2738 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2740 db_print_backtrace();
2743 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2744 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2746 db_print_backtrace();
2750 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2753 * Calculate the page table page (mpte), allocating it if necessary.
2755 * A held page table page (mpte), or NULL, is passed onto the
2756 * section following.
2758 if (va < VM_MAX_USER_ADDRESS) {
2760 * Calculate pagetable page index
2762 ptepindex = pmap_pde_pindex(va);
2766 * Get the page directory entry
2768 ptepa = pmap_pde(pmap, va);
2771 * If the page table page is mapped, we just increment
2772 * the hold count, and activate it.
2774 if (ptepa && (*ptepa & PG_V) != 0) {
2776 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2777 // if (pmap->pm_ptphint &&
2778 // (pmap->pm_ptphint->pindex == ptepindex)) {
2779 // mpte = pmap->pm_ptphint;
2781 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2782 pmap->pm_ptphint = mpte;
2787 mpte = _pmap_allocpte(pmap, ptepindex);
2789 } while (mpte == NULL);
2792 /* this code path is not yet used */
2796 * With a valid (and held) page directory page, we can just use
2797 * vtopte() to get to the pte. If the pte is already present
2798 * we do not disturb it.
2803 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2804 pa = VM_PAGE_TO_PHYS(m);
2805 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2806 pmap_inval_done(&info);
2807 lwkt_reltoken(&vm_token);
2812 * Enter on the PV list if part of our managed memory
2814 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2815 pmap_insert_entry(pmap, va, mpte, m);
2816 vm_page_flag_set(m, PG_MAPPED);
2820 * Increment counters
2822 ++pmap->pm_stats.resident_count;
2824 pa = VM_PAGE_TO_PHYS(m);
2827 * Now validate mapping with RO protection
2829 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2830 *pte = pa | PG_V | PG_U;
2832 *pte = pa | PG_V | PG_U | PG_MANAGED;
2833 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2834 pmap_inval_done(&info);
2835 lwkt_reltoken(&vm_token);
2839 * Make a temporary mapping for a physical address. This is only intended
2840 * to be used for panic dumps.
2842 * The caller is responsible for calling smp_invltlb().
2845 pmap_kenter_temporary(vm_paddr_t pa, long i)
2847 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2848 return ((void *)crashdumpmap);
2851 #define MAX_INIT_PT (96)
2854 * This routine preloads the ptes for a given object into the specified pmap.
2855 * This eliminates the blast of soft faults on process startup and
2856 * immediately after an mmap.
2858 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2861 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2862 vm_object_t object, vm_pindex_t pindex,
2863 vm_size_t size, int limit)
2865 struct rb_vm_page_scan_info info;
2870 * We can't preinit if read access isn't set or there is no pmap
2873 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2877 * We can't preinit if the pmap is not the current pmap
2879 lp = curthread->td_lwp;
2880 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2883 psize = x86_64_btop(size);
2885 if ((object->type != OBJT_VNODE) ||
2886 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2887 (object->resident_page_count > MAX_INIT_PT))) {
2891 if (psize + pindex > object->size) {
2892 if (object->size < pindex)
2894 psize = object->size - pindex;
2901 * Use a red-black scan to traverse the requested range and load
2902 * any valid pages found into the pmap.
2904 * We cannot safely scan the object's memq without holding the
2907 info.start_pindex = pindex;
2908 info.end_pindex = pindex + psize - 1;
2914 vm_object_hold(object);
2915 lwkt_gettoken(&vm_token);
2916 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2917 pmap_object_init_pt_callback, &info);
2918 lwkt_reltoken(&vm_token);
2919 vm_object_drop(object);
2924 pmap_object_init_pt_callback(vm_page_t p, void *data)
2926 struct rb_vm_page_scan_info *info = data;
2927 vm_pindex_t rel_index;
2929 * don't allow an madvise to blow away our really
2930 * free pages allocating pv entries.
2932 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2933 vmstats.v_free_count < vmstats.v_free_reserved) {
2936 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2937 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2939 if ((p->queue - p->pc) == PQ_CACHE)
2940 vm_page_deactivate(p);
2941 rel_index = p->pindex - info->start_pindex;
2942 pmap_enter_quick(info->pmap,
2943 info->addr + x86_64_ptob(rel_index), p);
2950 * Return TRUE if the pmap is in shape to trivially
2951 * pre-fault the specified address.
2953 * Returns FALSE if it would be non-trivial or if a
2954 * pte is already loaded into the slot.
2957 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2963 lwkt_gettoken(&vm_token);
2964 pde = pmap_pde(pmap, addr);
2965 if (pde == NULL || *pde == 0) {
2969 ret = (*pte) ? 0 : 1;
2971 lwkt_reltoken(&vm_token);
2976 * Routine: pmap_change_wiring
2977 * Function: Change the wiring attribute for a map/virtual-address
2979 * In/out conditions:
2980 * The mapping must already exist in the pmap.
2983 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2990 lwkt_gettoken(&vm_token);
2991 pte = pmap_pte(pmap, va);
2993 if (wired && !pmap_pte_w(pte))
2994 pmap->pm_stats.wired_count++;
2995 else if (!wired && pmap_pte_w(pte))
2996 pmap->pm_stats.wired_count--;
2999 * Wiring is not a hardware characteristic so there is no need to
3000 * invalidate TLB. However, in an SMP environment we must use
3001 * a locked bus cycle to update the pte (if we are not using
3002 * the pmap_inval_*() API that is)... it's ok to do this for simple
3007 atomic_set_long(pte, PG_W);
3009 atomic_clear_long(pte, PG_W);
3012 atomic_set_long_nonlocked(pte, PG_W);
3014 atomic_clear_long_nonlocked(pte, PG_W);
3016 lwkt_reltoken(&vm_token);
3022 * Copy the range specified by src_addr/len from the source map to
3023 * the range dst_addr/len in the destination map.
3025 * This routine is only advisory and need not do anything.
3028 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3029 vm_size_t len, vm_offset_t src_addr)
3033 pmap_inval_info info;
3035 vm_offset_t end_addr = src_addr + len;
3037 pd_entry_t src_frame, dst_frame;
3040 if (dst_addr != src_addr)
3043 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3044 if (src_frame != (PTDpde & PG_FRAME)) {
3048 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3049 if (dst_frame != (APTDpde & PG_FRAME)) {
3050 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3051 /* The page directory is not shared between CPUs */
3055 pmap_inval_init(&info);
3056 pmap_inval_add(&info, dst_pmap, -1);
3057 pmap_inval_add(&info, src_pmap, -1);
3060 * vm_token section protection is required to maintain the page/object
3063 lwkt_gettoken(&vm_token);
3064 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3065 pt_entry_t *src_pte, *dst_pte;
3066 vm_page_t dstmpte, srcmpte;
3067 vm_offset_t srcptepaddr;
3068 vm_pindex_t ptepindex;
3070 if (addr >= UPT_MIN_ADDRESS)
3071 panic("pmap_copy: invalid to pmap_copy page tables\n");
3074 * Don't let optional prefaulting of pages make us go
3075 * way below the low water mark of free pages or way
3076 * above high water mark of used pv entries.
3078 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3079 pv_entry_count > pv_entry_high_water)
3082 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3083 ptepindex = addr >> PDRSHIFT;
3086 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3088 if (srcptepaddr == 0)
3091 if (srcptepaddr & PG_PS) {
3093 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3094 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3095 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3101 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3102 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3103 (srcmpte->flags & PG_BUSY)) {
3107 if (pdnxt > end_addr)
3110 src_pte = vtopte(addr);
3112 dst_pte = avtopte(addr);
3114 while (addr < pdnxt) {
3119 * we only virtual copy managed pages
3121 if ((ptetemp & PG_MANAGED) != 0) {
3123 * We have to check after allocpte for the
3124 * pte still being around... allocpte can
3127 * pmap_allocpte() can block. If we lose
3128 * our page directory mappings we stop.
3130 dstmpte = pmap_allocpte(dst_pmap, addr);
3133 if (src_frame != (PTDpde & PG_FRAME) ||
3134 dst_frame != (APTDpde & PG_FRAME)
3136 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3137 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3139 } else if ((*dst_pte == 0) &&
3140 (ptetemp = *src_pte) != 0 &&
3141 (ptetemp & PG_MANAGED)) {
3143 * Clear the modified and
3144 * accessed (referenced) bits
3147 m = PHYS_TO_VM_PAGE(ptetemp);
3148 *dst_pte = ptetemp & ~(PG_M | PG_A);
3149 ++dst_pmap->pm_stats.resident_count;
3150 pmap_insert_entry(dst_pmap, addr,
3152 KKASSERT(m->flags & PG_MAPPED);
3154 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3155 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3159 if (dstmpte->hold_count >= srcmpte->hold_count)
3168 lwkt_reltoken(&vm_token);
3169 pmap_inval_done(&info);
3176 * Zero the specified physical page.
3178 * This function may be called from an interrupt and no locking is
3182 pmap_zero_page(vm_paddr_t phys)
3184 vm_offset_t va = PHYS_TO_DMAP(phys);
3186 pagezero((void *)va);
3190 * pmap_page_assertzero:
3192 * Assert that a page is empty, panic if it isn't.
3195 pmap_page_assertzero(vm_paddr_t phys)
3197 vm_offset_t va = PHYS_TO_DMAP(phys);
3200 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3201 if (*(long *)((char *)va + i) != 0) {
3202 panic("pmap_page_assertzero() @ %p not zero!\n",
3203 (void *)(intptr_t)va);
3211 * Zero part of a physical page by mapping it into memory and clearing
3212 * its contents with bzero.
3214 * off and size may not cover an area beyond a single hardware page.
3217 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3219 vm_offset_t virt = PHYS_TO_DMAP(phys);
3221 bzero((char *)virt + off, size);
3227 * Copy the physical page from the source PA to the target PA.
3228 * This function may be called from an interrupt. No locking
3232 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3234 vm_offset_t src_virt, dst_virt;
3236 src_virt = PHYS_TO_DMAP(src);
3237 dst_virt = PHYS_TO_DMAP(dst);
3238 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3242 * pmap_copy_page_frag:
3244 * Copy the physical page from the source PA to the target PA.
3245 * This function may be called from an interrupt. No locking
3249 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3251 vm_offset_t src_virt, dst_virt;
3253 src_virt = PHYS_TO_DMAP(src);
3254 dst_virt = PHYS_TO_DMAP(dst);
3256 bcopy((char *)src_virt + (src & PAGE_MASK),
3257 (char *)dst_virt + (dst & PAGE_MASK),
3262 * Returns true if the pmap's pv is one of the first
3263 * 16 pvs linked to from this page. This count may
3264 * be changed upwards or downwards in the future; it
3265 * is only necessary that true be returned for a small
3266 * subset of pmaps for proper page aging.
3269 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3274 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3277 lwkt_gettoken(&vm_token);
3279 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3280 if (pv->pv_pmap == pmap) {
3281 lwkt_reltoken(&vm_token);
3288 lwkt_reltoken(&vm_token);
3293 * Remove all pages from specified address space
3294 * this aids process exit speeds. Also, this code
3295 * is special cased for current process only, but
3296 * can have the more generic (and slightly slower)
3297 * mode enabled. This is much faster than pmap_remove
3298 * in the case of running down an entire address space.
3301 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3304 pt_entry_t *pte, tpte;
3307 pmap_inval_info info;
3309 int save_generation;
3311 lp = curthread->td_lwp;
3312 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3317 lwkt_gettoken(&vm_token);
3318 pmap_inval_init(&info);
3319 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3320 if (pv->pv_va >= eva || pv->pv_va < sva) {
3321 npv = TAILQ_NEXT(pv, pv_plist);
3325 KKASSERT(pmap == pv->pv_pmap);
3328 pte = vtopte(pv->pv_va);
3330 pte = pmap_pte_quick(pmap, pv->pv_va);
3331 pmap_inval_interlock(&info, pmap, pv->pv_va);
3334 * We cannot remove wired pages from a process' mapping
3338 pmap_inval_deinterlock(&info, pmap);
3339 npv = TAILQ_NEXT(pv, pv_plist);
3342 tpte = pte_load_clear(pte);
3343 KKASSERT(tpte & PG_MANAGED);
3345 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3347 KASSERT(m < &vm_page_array[vm_page_array_size],
3348 ("pmap_remove_pages: bad tpte %lx", tpte));
3350 KKASSERT(pmap->pm_stats.resident_count > 0);
3351 --pmap->pm_stats.resident_count;
3352 pmap_inval_deinterlock(&info, pmap);
3355 * Update the vm_page_t clean and reference bits.
3361 npv = TAILQ_NEXT(pv, pv_plist);
3362 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3363 save_generation = ++pmap->pm_generation;
3365 m->md.pv_list_count--;
3366 m->object->agg_pv_list_count--;
3367 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3368 if (TAILQ_EMPTY(&m->md.pv_list))
3369 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3371 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3375 * Restart the scan if we blocked during the unuse or free
3376 * calls and other removals were made.
3378 if (save_generation != pmap->pm_generation) {
3379 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3380 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3383 pmap_inval_done(&info);
3384 lwkt_reltoken(&vm_token);
3388 * pmap_testbit tests bits in pte's note that the testbit/clearbit
3389 * routines are inline, and a lot of things compile-time evaluate.
3391 * Caller must hold vm_token
3395 pmap_testbit(vm_page_t m, int bit)
3400 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3403 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3406 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3408 * if the bit being tested is the modified bit, then
3409 * mark clean_map and ptes as never
3412 if (bit & (PG_A|PG_M)) {
3413 if (!pmap_track_modified(pv->pv_va))
3417 #if defined(PMAP_DIAGNOSTIC)
3418 if (pv->pv_pmap == NULL) {
3419 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3423 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3431 * this routine is used to modify bits in ptes
3435 pmap_clearbit(vm_page_t m, int bit)
3437 struct pmap_inval_info info;
3442 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3445 pmap_inval_init(&info);
3448 * Loop over all current mappings setting/clearing as appropos If
3449 * setting RO do we need to clear the VAC?
3451 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3453 * don't write protect pager mappings
3456 if (!pmap_track_modified(pv->pv_va))
3460 #if defined(PMAP_DIAGNOSTIC)
3461 if (pv->pv_pmap == NULL) {
3462 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3468 * Careful here. We can use a locked bus instruction to
3469 * clear PG_A or PG_M safely but we need to synchronize
3470 * with the target cpus when we mess with PG_RW.
3472 * We do not have to force synchronization when clearing
3473 * PG_M even for PTEs generated via virtual memory maps,
3474 * because the virtual kernel will invalidate the pmap
3475 * entry when/if it needs to resynchronize the Modify bit.
3478 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3479 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3486 atomic_clear_long(pte, PG_M|PG_RW);
3489 * The cpu may be trying to set PG_M
3490 * simultaniously with our clearing
3493 if (!atomic_cmpset_long(pte, pbits,
3497 } else if (bit == PG_M) {
3499 * We could also clear PG_RW here to force
3500 * a fault on write to redetect PG_M for
3501 * virtual kernels, but it isn't necessary
3502 * since virtual kernels invalidate the pte
3503 * when they clear the VPTE_M bit in their
3504 * virtual page tables.
3506 atomic_clear_long(pte, PG_M);
3508 atomic_clear_long(pte, bit);
3512 pmap_inval_deinterlock(&info, pv->pv_pmap);
3514 pmap_inval_done(&info);
3518 * pmap_page_protect:
3520 * Lower the permission for all mappings to a given page.
3523 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3525 /* JG NX support? */
3526 if ((prot & VM_PROT_WRITE) == 0) {
3527 lwkt_gettoken(&vm_token);
3528 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3529 pmap_clearbit(m, PG_RW);
3530 vm_page_flag_clear(m, PG_WRITEABLE);
3534 lwkt_reltoken(&vm_token);
3539 pmap_phys_address(vm_pindex_t ppn)
3541 return (x86_64_ptob(ppn));
3545 * pmap_ts_referenced:
3547 * Return a count of reference bits for a page, clearing those bits.
3548 * It is not necessary for every reference bit to be cleared, but it
3549 * is necessary that 0 only be returned when there are truly no
3550 * reference bits set.
3552 * XXX: The exact number of bits to check and clear is a matter that
3553 * should be tested and standardized at some point in the future for
3554 * optimal aging of shared pages.
3557 pmap_ts_referenced(vm_page_t m)
3559 pv_entry_t pv, pvf, pvn;
3563 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3566 lwkt_gettoken(&vm_token);
3568 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3571 pvn = TAILQ_NEXT(pv, pv_list);
3573 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3574 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3576 if (!pmap_track_modified(pv->pv_va))
3579 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3581 if (pte && (*pte & PG_A)) {
3583 atomic_clear_long(pte, PG_A);
3585 atomic_clear_long_nonlocked(pte, PG_A);
3592 } while ((pv = pvn) != NULL && pv != pvf);
3594 lwkt_reltoken(&vm_token);
3602 * Return whether or not the specified physical page was modified
3603 * in any physical maps.
3606 pmap_is_modified(vm_page_t m)
3610 lwkt_gettoken(&vm_token);
3611 res = pmap_testbit(m, PG_M);
3612 lwkt_reltoken(&vm_token);
3617 * Clear the modify bits on the specified physical page.
3620 pmap_clear_modify(vm_page_t m)
3622 lwkt_gettoken(&vm_token);
3623 pmap_clearbit(m, PG_M);
3624 lwkt_reltoken(&vm_token);
3628 * pmap_clear_reference:
3630 * Clear the reference bit on the specified physical page.
3633 pmap_clear_reference(vm_page_t m)
3635 lwkt_gettoken(&vm_token);
3636 pmap_clearbit(m, PG_A);
3637 lwkt_reltoken(&vm_token);
3641 * Miscellaneous support routines follow
3646 i386_protection_init(void)
3650 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3651 kp = protection_codes;
3652 for (prot = 0; prot < 8; prot++) {
3654 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3656 * Read access is also 0. There isn't any execute bit,
3657 * so just make it readable.
3659 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3660 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3661 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3664 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3665 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3666 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3667 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3675 * Map a set of physical memory pages into the kernel virtual
3676 * address space. Return a pointer to where it is mapped. This
3677 * routine is intended to be used for mapping device memory,
3680 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3684 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3686 vm_offset_t va, tmpva, offset;
3689 offset = pa & PAGE_MASK;
3690 size = roundup(offset + size, PAGE_SIZE);
3692 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3694 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3696 pa = pa & ~PAGE_MASK;
3697 for (tmpva = va; size > 0;) {
3698 pte = vtopte(tmpva);
3699 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3707 return ((void *)(va + offset));
3711 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3713 vm_offset_t va, tmpva, offset;
3716 offset = pa & PAGE_MASK;
3717 size = roundup(offset + size, PAGE_SIZE);
3719 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3721 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3723 pa = pa & ~PAGE_MASK;
3724 for (tmpva = va; size > 0;) {
3725 pte = vtopte(tmpva);
3726 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3734 return ((void *)(va + offset));
3738 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3740 vm_offset_t base, offset;
3742 base = va & ~PAGE_MASK;
3743 offset = va & PAGE_MASK;
3744 size = roundup(offset + size, PAGE_SIZE);
3745 pmap_qremove(va, size >> PAGE_SHIFT);
3746 kmem_free(&kernel_map, base, size);
3750 * perform the pmap work for mincore
3753 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3755 pt_entry_t *ptep, pte;
3759 lwkt_gettoken(&vm_token);
3760 ptep = pmap_pte(pmap, addr);
3762 if (ptep && (pte = *ptep) != 0) {
3765 val = MINCORE_INCORE;
3766 if ((pte & PG_MANAGED) == 0)
3769 pa = pte & PG_FRAME;
3771 m = PHYS_TO_VM_PAGE(pa);
3777 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3779 * Modified by someone
3781 else if (m->dirty || pmap_is_modified(m))
3782 val |= MINCORE_MODIFIED_OTHER;
3787 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3790 * Referenced by someone
3792 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3793 val |= MINCORE_REFERENCED_OTHER;
3794 vm_page_flag_set(m, PG_REFERENCED);
3798 lwkt_reltoken(&vm_token);
3803 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3804 * vmspace will be ref'd and the old one will be deref'd.
3806 * The vmspace for all lwps associated with the process will be adjusted
3807 * and cr3 will be reloaded if any lwp is the current lwp.
3809 * Caller must hold vmspace_token
3812 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3814 struct vmspace *oldvm;
3817 oldvm = p->p_vmspace;
3818 if (oldvm != newvm) {
3820 sysref_get(&newvm->vm_sysref);
3821 p->p_vmspace = newvm;
3822 KKASSERT(p->p_nthreads == 1);
3823 lp = RB_ROOT(&p->p_lwp_tree);
3824 pmap_setlwpvm(lp, newvm);
3826 sysref_put(&oldvm->vm_sysref);
3831 * Set the vmspace for a LWP. The vmspace is almost universally set the
3832 * same as the process vmspace, but virtual kernels need to swap out contexts
3833 * on a per-lwp basis.
3835 * Caller does not necessarily hold vmspace_token. Caller must control
3836 * the lwp (typically be in the context of the lwp). We use a critical
3837 * section to protect against statclock and hardclock (statistics collection).
3840 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3842 struct vmspace *oldvm;
3845 oldvm = lp->lwp_vmspace;
3847 if (oldvm != newvm) {
3849 lp->lwp_vmspace = newvm;
3850 if (curthread->td_lwp == lp) {
3851 pmap = vmspace_pmap(newvm);
3853 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3854 if (pmap->pm_active & CPUMASK_LOCK)
3855 pmap_interlock_wait(newvm);
3857 pmap->pm_active |= 1;
3859 #if defined(SWTCH_OPTIM_STATS)
3862 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3863 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
3864 load_cr3(curthread->td_pcb->pcb_cr3);
3865 pmap = vmspace_pmap(oldvm);
3867 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3869 pmap->pm_active &= ~(cpumask_t)1;
3879 * Called when switching to a locked pmap
3882 pmap_interlock_wait(struct vmspace *vm)
3884 struct pmap *pmap = &vm->vm_pmap;
3886 if (pmap->pm_active & CPUMASK_LOCK) {
3887 DEBUG_PUSH_INFO("pmap_interlock_wait");
3888 while (pmap->pm_active & CPUMASK_LOCK) {
3891 lwkt_process_ipiq();
3900 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3903 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3907 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3912 * Used by kmalloc/kfree, page already exists at va
3915 pmap_kvtom(vm_offset_t va)
3917 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));