2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
9 * Copyright (c) 2011-2012 Matthew Dillon
10 * All rights reserved.
12 * This code is derived from software contributed to Berkeley by
13 * the Systems Programming Group of the University of Utah Computer
14 * Science Department and William Jolitz of UUNET Technologies Inc.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * Manage physical address maps for x86-64 systems.
49 #include "opt_disable_pse.h"
52 #include "opt_msgbuf.h"
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/kernel.h>
58 #include <sys/msgbuf.h>
59 #include <sys/vmmeter.h>
63 #include <vm/vm_param.h>
64 #include <sys/sysctl.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_pageout.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_zone.h>
76 #include <sys/thread2.h>
77 #include <sys/sysref2.h>
78 #include <sys/spinlock2.h>
79 #include <vm/vm_page2.h>
81 #include <machine/cputypes.h>
82 #include <machine/md_var.h>
83 #include <machine/specialreg.h>
84 #include <machine/smp.h>
85 #include <machine_base/apic/apicreg.h>
86 #include <machine/globaldata.h>
87 #include <machine/pmap.h>
88 #include <machine/pmap_inval.h>
89 #include <machine/inttypes.h>
93 #define PMAP_KEEP_PDIRS
94 #ifndef PMAP_SHPGPERPROC
95 #define PMAP_SHPGPERPROC 2000
98 #if defined(DIAGNOSTIC)
99 #define PMAP_DIAGNOSTIC
105 * pmap debugging will report who owns a pv lock when blocking.
109 #define PMAP_DEBUG_DECL ,const char *func, int lineno
110 #define PMAP_DEBUG_ARGS , __func__, __LINE__
111 #define PMAP_DEBUG_COPY , func, lineno
113 #define pv_get(pmap, pindex) _pv_get(pmap, pindex \
115 #define pv_lock(pv) _pv_lock(pv \
117 #define pv_hold_try(pv) _pv_hold_try(pv \
119 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \
124 #define PMAP_DEBUG_DECL
125 #define PMAP_DEBUG_ARGS
126 #define PMAP_DEBUG_COPY
128 #define pv_get(pmap, pindex) _pv_get(pmap, pindex)
129 #define pv_lock(pv) _pv_lock(pv)
130 #define pv_hold_try(pv) _pv_hold_try(pv)
131 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp)
136 * Get PDEs and PTEs for user/kernel address space
138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
140 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
141 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
142 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
143 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
144 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
147 * Given a map and a machine independent protection code,
148 * convert to a vax protection code.
150 #define pte_prot(m, p) \
151 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
152 static int protection_codes[8];
154 struct pmap kernel_pmap;
155 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
159 vm_paddr_t avail_start; /* PA of first available physical page */
160 vm_paddr_t avail_end; /* PA of last available physical page */
161 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
162 vm_offset_t virtual2_end;
163 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
164 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
165 vm_offset_t KvaStart; /* VA start of KVA space */
166 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
167 vm_offset_t KvaSize; /* max size of kernel virtual address space */
168 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
169 static int pgeflag; /* PG_G or-in */
170 static int pseflag; /* PG_PS or-in */
174 static vm_paddr_t dmaplimit;
176 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
178 #define PAT_INDEX_SIZE 8
179 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE]; /* PAT -> PG_ bits */
180 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
182 static uint64_t KPTbase;
183 static uint64_t KPTphys;
184 static uint64_t KPDphys; /* phys addr of kernel level 2 */
185 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
186 uint64_t KPDPphys; /* phys addr of kernel level 3 */
187 uint64_t KPML4phys; /* phys addr of kernel level 4 */
189 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
190 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
193 * Data for the pv entry allocation mechanism
195 static vm_zone_t pvzone;
196 static struct vm_zone pvzone_store;
197 static struct vm_object pvzone_obj;
198 static int pv_entry_max=0, pv_entry_high_water=0;
199 static int pmap_pagedaemon_waken = 0;
200 static struct pv_entry *pvinit;
203 * All those kernel PT submaps that BSD is so fond of
205 pt_entry_t *CMAP1 = NULL, *ptmmap;
206 caddr_t CADDR1 = NULL, ptvmmap = NULL;
207 static pt_entry_t *msgbufmap;
208 struct msgbuf *msgbufp=NULL;
213 static pt_entry_t *pt_crashdumpmap;
214 static caddr_t crashdumpmap;
216 static int pmap_yield_count = 64;
217 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
218 &pmap_yield_count, 0, "Yield during init_pt/release");
219 static int pmap_mmu_optimize = 0;
220 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
221 &pmap_mmu_optimize, 0, "Share page table pages when possible");
225 static void pv_hold(pv_entry_t pv);
226 static int _pv_hold_try(pv_entry_t pv
228 static void pv_drop(pv_entry_t pv);
229 static void _pv_lock(pv_entry_t pv
231 static void pv_unlock(pv_entry_t pv);
232 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
234 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
236 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
237 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
238 static void pv_put(pv_entry_t pv);
239 static void pv_free(pv_entry_t pv);
240 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
241 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
243 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
244 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
245 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
246 struct pmap_inval_info *info);
247 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
248 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp);
250 struct pmap_scan_info;
251 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
252 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
253 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
254 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
255 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
256 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
258 static void i386_protection_init (void);
259 static void create_pagetables(vm_paddr_t *firstaddr);
260 static void pmap_remove_all (vm_page_t m);
261 static boolean_t pmap_testbit (vm_page_t m, int bit);
263 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
264 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
266 static unsigned pdir4mb;
269 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
271 if (pv1->pv_pindex < pv2->pv_pindex)
273 if (pv1->pv_pindex > pv2->pv_pindex)
278 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
279 pv_entry_compare, vm_pindex_t, pv_pindex);
282 * Move the kernel virtual free pointer to the next
283 * 2MB. This is used to help improve performance
284 * by using a large (2MB) page for much of the kernel
285 * (.text, .data, .bss)
289 pmap_kmem_choose(vm_offset_t addr)
291 vm_offset_t newaddr = addr;
293 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
300 * Super fast pmap_pte routine best used when scanning the pv lists.
301 * This eliminates many course-grained invltlb calls. Note that many of
302 * the pv list scans are across different pmaps and it is very wasteful
303 * to do an entire invltlb when checking a single mapping.
305 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
309 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
311 return pmap_pte(pmap, va);
315 * Returns the pindex of a page table entry (representing a terminal page).
316 * There are NUPTE_TOTAL page table entries possible (a huge number)
318 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
319 * We want to properly translate negative KVAs.
323 pmap_pte_pindex(vm_offset_t va)
325 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
329 * Returns the pindex of a page table.
333 pmap_pt_pindex(vm_offset_t va)
335 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
339 * Returns the pindex of a page directory.
343 pmap_pd_pindex(vm_offset_t va)
345 return (NUPTE_TOTAL + NUPT_TOTAL +
346 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
351 pmap_pdp_pindex(vm_offset_t va)
353 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
354 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
359 pmap_pml4_pindex(void)
361 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
365 * Return various clipped indexes for a given VA
367 * Returns the index of a pte in a page table, representing a terminal
372 pmap_pte_index(vm_offset_t va)
374 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
378 * Returns the index of a pt in a page directory, representing a page
383 pmap_pt_index(vm_offset_t va)
385 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
389 * Returns the index of a pd in a page directory page, representing a page
394 pmap_pd_index(vm_offset_t va)
396 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
400 * Returns the index of a pdp in the pml4 table, representing a page
405 pmap_pdp_index(vm_offset_t va)
407 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
411 * Generic procedure to index a pte from a pt, pd, or pdp.
413 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
414 * a page table page index but is instead of PV lookup index.
418 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
422 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
423 return(&pte[pindex]);
427 * Return pointer to PDP slot in the PML4
431 pmap_pdp(pmap_t pmap, vm_offset_t va)
433 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
437 * Return pointer to PD slot in the PDP given a pointer to the PDP
441 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
445 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
446 return (&pd[pmap_pd_index(va)]);
450 * Return pointer to PD slot in the PDP.
454 pmap_pd(pmap_t pmap, vm_offset_t va)
458 pdp = pmap_pdp(pmap, va);
459 if ((*pdp & PG_V) == 0)
461 return (pmap_pdp_to_pd(*pdp, va));
465 * Return pointer to PT slot in the PD given a pointer to the PD
469 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
473 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
474 return (&pt[pmap_pt_index(va)]);
478 * Return pointer to PT slot in the PD
480 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
481 * so we cannot lookup the PD via the PDP. Instead we
482 * must look it up via the pmap.
486 pmap_pt(pmap_t pmap, vm_offset_t va)
490 vm_pindex_t pd_pindex;
492 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
493 pd_pindex = pmap_pd_pindex(va);
494 spin_lock(&pmap->pm_spin);
495 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
496 spin_unlock(&pmap->pm_spin);
497 if (pv == NULL || pv->pv_m == NULL)
499 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
501 pd = pmap_pd(pmap, va);
502 if (pd == NULL || (*pd & PG_V) == 0)
504 return (pmap_pd_to_pt(*pd, va));
509 * Return pointer to PTE slot in the PT given a pointer to the PT
513 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
517 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
518 return (&pte[pmap_pte_index(va)]);
522 * Return pointer to PTE slot in the PT
526 pmap_pte(pmap_t pmap, vm_offset_t va)
530 pt = pmap_pt(pmap, va);
531 if (pt == NULL || (*pt & PG_V) == 0)
533 if ((*pt & PG_PS) != 0)
534 return ((pt_entry_t *)pt);
535 return (pmap_pt_to_pte(*pt, va));
539 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
540 * the PT layer. This will speed up core pmap operations considerably.
544 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
546 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
547 pv->pv_pmap->pm_pvhint = pv;
552 * KVM - return address of PT slot in PD
556 vtopt(vm_offset_t va)
558 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
559 NPML4EPGSHIFT)) - 1);
561 return (PDmap + ((va >> PDRSHIFT) & mask));
565 * KVM - return address of PTE slot in PT
569 vtopte(vm_offset_t va)
571 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
572 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
574 return (PTmap + ((va >> PAGE_SHIFT) & mask));
578 allocpages(vm_paddr_t *firstaddr, long n)
583 bzero((void *)ret, n * PAGE_SIZE);
584 *firstaddr += n * PAGE_SIZE;
590 create_pagetables(vm_paddr_t *firstaddr)
592 long i; /* must be 64 bits */
598 * We are running (mostly) V=P at this point
600 * Calculate NKPT - number of kernel page tables. We have to
601 * accomodoate prealloction of the vm_page_array, dump bitmap,
602 * MSGBUF_SIZE, and other stuff. Be generous.
604 * Maxmem is in pages.
606 * ndmpdp is the number of 1GB pages we wish to map.
608 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
609 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
611 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
614 * Starting at the beginning of kvm (not KERNBASE).
616 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
617 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
618 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
619 ndmpdp) + 511) / 512;
623 * Starting at KERNBASE - map 2G worth of page table pages.
624 * KERNBASE is offset -2G from the end of kvm.
626 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
631 KPTbase = allocpages(firstaddr, nkpt_base);
632 KPTphys = allocpages(firstaddr, nkpt_phys);
633 KPML4phys = allocpages(firstaddr, 1);
634 KPDPphys = allocpages(firstaddr, NKPML4E);
635 KPDphys = allocpages(firstaddr, NKPDPE);
638 * Calculate the page directory base for KERNBASE,
639 * that is where we start populating the page table pages.
640 * Basically this is the end - 2.
642 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
644 DMPDPphys = allocpages(firstaddr, NDMPML4E);
645 if ((amd_feature & AMDID_PAGE1GB) == 0)
646 DMPDphys = allocpages(firstaddr, ndmpdp);
647 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
650 * Fill in the underlying page table pages for the area around
651 * KERNBASE. This remaps low physical memory to KERNBASE.
653 * Read-only from zero to physfree
654 * XXX not fully used, underneath 2M pages
656 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
657 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
658 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
662 * Now map the initial kernel page tables. One block of page
663 * tables is placed at the beginning of kernel virtual memory,
664 * and another block is placed at KERNBASE to map the kernel binary,
665 * data, bss, and initial pre-allocations.
667 for (i = 0; i < nkpt_base; i++) {
668 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
669 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
671 for (i = 0; i < nkpt_phys; i++) {
672 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
673 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
677 * Map from zero to end of allocations using 2M pages as an
678 * optimization. This will bypass some of the KPTBase pages
679 * above in the KERNBASE area.
681 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
682 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
683 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
687 * And connect up the PD to the PDP. The kernel pmap is expected
688 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
690 for (i = 0; i < NKPDPE; i++) {
691 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
692 KPDphys + (i << PAGE_SHIFT);
693 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
698 * Now set up the direct map space using either 2MB or 1GB pages
699 * Preset PG_M and PG_A because demotion expects it.
701 * When filling in entries in the PD pages make sure any excess
702 * entries are set to zero as we allocated enough PD pages
704 if ((amd_feature & AMDID_PAGE1GB) == 0) {
705 for (i = 0; i < NPDEPG * ndmpdp; i++) {
706 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
707 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
712 * And the direct map space's PDP
714 for (i = 0; i < ndmpdp; i++) {
715 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
717 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
720 for (i = 0; i < ndmpdp; i++) {
721 ((pdp_entry_t *)DMPDPphys)[i] =
722 (vm_paddr_t)i << PDPSHIFT;
723 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
728 /* And recursively map PML4 to itself in order to get PTmap */
729 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
730 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
733 * Connect the Direct Map slots up to the PML4
735 for (j = 0; j < NDMPML4E; ++j) {
736 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
737 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
742 * Connect the KVA slot up to the PML4
744 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
745 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
749 * Bootstrap the system enough to run with virtual memory.
751 * On the i386 this is called after mapping has already been enabled
752 * and just syncs the pmap module with what has already been done.
753 * [We can't call it easily with mapping off since the kernel is not
754 * mapped with PA == VA, hence we would have to relocate every address
755 * from the linked base (virtual) address "KERNBASE" to the actual
756 * (physical) address starting relative to 0]
759 pmap_bootstrap(vm_paddr_t *firstaddr)
764 KvaStart = VM_MIN_KERNEL_ADDRESS;
765 KvaEnd = VM_MAX_KERNEL_ADDRESS;
766 KvaSize = KvaEnd - KvaStart;
768 avail_start = *firstaddr;
771 * Create an initial set of page tables to run the kernel in.
773 create_pagetables(firstaddr);
775 virtual2_start = KvaStart;
776 virtual2_end = PTOV_OFFSET;
778 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
779 virtual_start = pmap_kmem_choose(virtual_start);
781 virtual_end = VM_MAX_KERNEL_ADDRESS;
783 /* XXX do %cr0 as well */
784 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
788 * Initialize protection array.
790 i386_protection_init();
793 * The kernel's pmap is statically allocated so we don't have to use
794 * pmap_create, which is unlikely to work correctly at this part of
795 * the boot sequence (XXX and which no longer exists).
797 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
798 kernel_pmap.pm_count = 1;
799 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
800 RB_INIT(&kernel_pmap.pm_pvroot);
801 spin_init(&kernel_pmap.pm_spin);
802 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
805 * Reserve some special page table entries/VA space for temporary
808 #define SYSMAP(c, p, v, n) \
809 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
815 * CMAP1/CMAP2 are used for zeroing and copying pages.
817 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
822 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
825 * ptvmmap is used for reading arbitrary physical pages via
828 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
831 * msgbufp is used to map the system message buffer.
832 * XXX msgbufmap is not used.
834 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
835 atop(round_page(MSGBUF_SIZE)))
842 * PG_G is terribly broken on SMP because we IPI invltlb's in some
843 * cases rather then invl1pg. Actually, I don't even know why it
844 * works under UP because self-referential page table mappings
849 * Initialize the 4MB page size flag
853 * The 4MB page version of the initial
854 * kernel page mapping.
858 #if !defined(DISABLE_PSE)
859 if (cpu_feature & CPUID_PSE) {
862 * Note that we have enabled PSE mode
865 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
866 ptditmp &= ~(NBPDR - 1);
867 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
873 /* Initialize the PAT MSR */
887 * Default values mapping PATi,PCD,PWT bits at system reset.
888 * The default values effectively ignore the PATi bit by
889 * repeating the encodings for 0-3 in 4-7, and map the PCD
890 * and PWT bit combinations to the expected PAT types.
892 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
893 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
894 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
895 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
896 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
897 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
898 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
899 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
900 pat_pte_index[PAT_WRITE_BACK] = 0;
901 pat_pte_index[PAT_WRITE_THROUGH]= 0 | PG_NC_PWT;
902 pat_pte_index[PAT_UNCACHED] = PG_NC_PCD;
903 pat_pte_index[PAT_UNCACHEABLE] = PG_NC_PCD | PG_NC_PWT;
904 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
905 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
907 if (cpu_feature & CPUID_PAT) {
909 * If we support the PAT then set-up entries for
910 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
913 pat_msr = (pat_msr & ~PAT_MASK(4)) |
914 PAT_VALUE(4, PAT_WRITE_PROTECTED);
915 pat_msr = (pat_msr & ~PAT_MASK(5)) |
916 PAT_VALUE(5, PAT_WRITE_COMBINING);
917 pat_pte_index[PAT_WRITE_PROTECTED] = PG_PTE_PAT | 0;
918 pat_pte_index[PAT_WRITE_COMBINING] = PG_PTE_PAT | PG_NC_PWT;
921 * Then enable the PAT
926 load_cr4(cr4 & ~CR4_PGE);
928 /* Disable caches (CD = 1, NW = 0). */
930 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
932 /* Flushes caches and TLBs. */
936 /* Update PAT and index table. */
937 wrmsr(MSR_PAT, pat_msr);
939 /* Flush caches and TLBs again. */
943 /* Restore caches and PGE. */
951 * Set 4mb pdir for mp startup
956 if (pseflag && (cpu_feature & CPUID_PSE)) {
957 load_cr4(rcr4() | CR4_PSE);
958 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
965 * Initialize the pmap module.
966 * Called by vm_init, to initialize any structures that the pmap
967 * system needs to map virtual memory.
968 * pmap_init has been enhanced to support in a fairly consistant
969 * way, discontiguous physical memory.
978 * Allocate memory for random pmap data structures. Includes the
982 for (i = 0; i < vm_page_array_size; i++) {
985 m = &vm_page_array[i];
986 TAILQ_INIT(&m->md.pv_list);
990 * init the pv free list
992 initial_pvs = vm_page_array_size;
993 if (initial_pvs < MINPV)
995 pvzone = &pvzone_store;
996 pvinit = (void *)kmem_alloc(&kernel_map,
997 initial_pvs * sizeof (struct pv_entry));
998 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
999 pvinit, initial_pvs);
1002 * Now it is safe to enable pv_table recording.
1004 pmap_initialized = TRUE;
1008 * Initialize the address space (zone) for the pv_entries. Set a
1009 * high water mark so that the system can recover from excessive
1010 * numbers of pv entries.
1015 int shpgperproc = PMAP_SHPGPERPROC;
1018 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1019 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1020 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1021 pv_entry_high_water = 9 * (pv_entry_max / 10);
1024 * Subtract out pages already installed in the zone (hack)
1026 entry_max = pv_entry_max - vm_page_array_size;
1030 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
1034 /***************************************************
1035 * Low level helper routines.....
1036 ***************************************************/
1039 * this routine defines the region(s) of memory that should
1040 * not be tested for the modified bit.
1044 pmap_track_modified(vm_pindex_t pindex)
1046 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1047 if ((va < clean_sva) || (va >= clean_eva))
1054 * Extract the physical page address associated with the map/VA pair.
1055 * The page must be wired for this to work reliably.
1057 * XXX for the moment we're using pv_find() instead of pv_get(), as
1058 * callers might be expecting non-blocking operation.
1061 pmap_extract(pmap_t pmap, vm_offset_t va)
1068 if (va >= VM_MAX_USER_ADDRESS) {
1070 * Kernel page directories might be direct-mapped and
1071 * there is typically no PV tracking of pte's
1075 pt = pmap_pt(pmap, va);
1076 if (pt && (*pt & PG_V)) {
1078 rtval = *pt & PG_PS_FRAME;
1079 rtval |= va & PDRMASK;
1081 ptep = pmap_pt_to_pte(*pt, va);
1083 rtval = *ptep & PG_FRAME;
1084 rtval |= va & PAGE_MASK;
1090 * User pages currently do not direct-map the page directory
1091 * and some pages might not used managed PVs. But all PT's
1094 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1096 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1098 rtval = *ptep & PG_FRAME;
1099 rtval |= va & PAGE_MASK;
1108 * Extract the physical page address associated kernel virtual address.
1111 pmap_kextract(vm_offset_t va)
1113 pd_entry_t pt; /* pt entry in pd */
1116 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1117 pa = DMAP_TO_PHYS(va);
1121 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1124 * Beware of a concurrent promotion that changes the
1125 * PDE at this point! For example, vtopte() must not
1126 * be used to access the PTE because it would use the
1127 * new PDE. It is, however, safe to use the old PDE
1128 * because the page table page is preserved by the
1131 pa = *pmap_pt_to_pte(pt, va);
1132 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1138 /***************************************************
1139 * Low level mapping routines.....
1140 ***************************************************/
1143 * Routine: pmap_kenter
1145 * Add a wired page to the KVA
1146 * NOTE! note that in order for the mapping to take effect -- you
1147 * should do an invltlb after doing the pmap_kenter().
1150 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1154 pmap_inval_info info;
1156 pmap_inval_init(&info); /* XXX remove */
1157 npte = pa | PG_RW | PG_V | pgeflag;
1159 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1161 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1162 pmap_inval_done(&info); /* XXX remove */
1166 * Routine: pmap_kenter_quick
1168 * Similar to pmap_kenter(), except we only invalidate the
1169 * mapping on the current CPU.
1172 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1177 npte = pa | PG_RW | PG_V | pgeflag;
1180 cpu_invlpg((void *)va);
1184 pmap_kenter_sync(vm_offset_t va)
1186 pmap_inval_info info;
1188 pmap_inval_init(&info);
1189 pmap_inval_interlock(&info, &kernel_pmap, va);
1190 pmap_inval_deinterlock(&info, &kernel_pmap);
1191 pmap_inval_done(&info);
1195 pmap_kenter_sync_quick(vm_offset_t va)
1197 cpu_invlpg((void *)va);
1201 * remove a page from the kernel pagetables
1204 pmap_kremove(vm_offset_t va)
1207 pmap_inval_info info;
1209 pmap_inval_init(&info);
1211 pmap_inval_interlock(&info, &kernel_pmap, va);
1212 (void)pte_load_clear(pte);
1213 pmap_inval_deinterlock(&info, &kernel_pmap);
1214 pmap_inval_done(&info);
1218 pmap_kremove_quick(vm_offset_t va)
1222 (void)pte_load_clear(pte);
1223 cpu_invlpg((void *)va);
1227 * XXX these need to be recoded. They are not used in any critical path.
1230 pmap_kmodify_rw(vm_offset_t va)
1232 atomic_set_long(vtopte(va), PG_RW);
1233 cpu_invlpg((void *)va);
1237 pmap_kmodify_nc(vm_offset_t va)
1239 atomic_set_long(vtopte(va), PG_N);
1240 cpu_invlpg((void *)va);
1244 * Used to map a range of physical addresses into kernel virtual
1245 * address space during the low level boot, typically to map the
1246 * dump bitmap, message buffer, and vm_page_array.
1248 * These mappings are typically made at some pointer after the end of the
1251 * We could return PHYS_TO_DMAP(start) here and not allocate any
1252 * via (*virtp), but then kmem from userland and kernel dumps won't
1253 * have access to the related pointers.
1256 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1259 vm_offset_t va_start;
1261 /*return PHYS_TO_DMAP(start);*/
1266 while (start < end) {
1267 pmap_kenter_quick(va, start);
1275 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1278 * Remove the specified set of pages from the data and instruction caches.
1280 * In contrast to pmap_invalidate_cache_range(), this function does not
1281 * rely on the CPU's self-snoop feature, because it is intended for use
1282 * when moving pages into a different cache domain.
1285 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1287 vm_offset_t daddr, eva;
1290 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1291 (cpu_feature & CPUID_CLFSH) == 0)
1295 for (i = 0; i < count; i++) {
1296 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1297 eva = daddr + PAGE_SIZE;
1298 for (; daddr < eva; daddr += cpu_clflush_line_size)
1306 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1308 KASSERT((sva & PAGE_MASK) == 0,
1309 ("pmap_invalidate_cache_range: sva not page-aligned"));
1310 KASSERT((eva & PAGE_MASK) == 0,
1311 ("pmap_invalidate_cache_range: eva not page-aligned"));
1313 if (cpu_feature & CPUID_SS) {
1314 ; /* If "Self Snoop" is supported, do nothing. */
1316 /* Globally invalidate caches */
1317 cpu_wbinvd_on_all_cpus();
1321 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1323 smp_invlpg_range(pmap->pm_active, sva, eva);
1327 * Add a list of wired pages to the kva
1328 * this routine is only used for temporary
1329 * kernel mappings that do not need to have
1330 * page modification or references recorded.
1331 * Note that old mappings are simply written
1332 * over. The page *must* be wired.
1335 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1339 end_va = va + count * PAGE_SIZE;
1341 while (va < end_va) {
1345 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V |
1346 pat_pte_index[(*m)->pat_mode] | pgeflag;
1347 cpu_invlpg((void *)va);
1355 * This routine jerks page mappings from the
1356 * kernel -- it is meant only for temporary mappings.
1358 * MPSAFE, INTERRUPT SAFE (cluster callback)
1361 pmap_qremove(vm_offset_t va, int count)
1365 end_va = va + count * PAGE_SIZE;
1367 while (va < end_va) {
1371 (void)pte_load_clear(pte);
1372 cpu_invlpg((void *)va);
1379 * Create a new thread and optionally associate it with a (new) process.
1380 * NOTE! the new thread's cpu may not equal the current cpu.
1383 pmap_init_thread(thread_t td)
1385 /* enforce pcb placement & alignment */
1386 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1387 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1388 td->td_savefpu = &td->td_pcb->pcb_save;
1389 td->td_sp = (char *)td->td_pcb; /* no -16 */
1393 * This routine directly affects the fork perf for a process.
1396 pmap_init_proc(struct proc *p)
1401 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1402 * it, and IdlePTD, represents the template used to update all other pmaps.
1404 * On architectures where the kernel pmap is not integrated into the user
1405 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1406 * kernel_pmap should be used to directly access the kernel_pmap.
1409 pmap_pinit0(struct pmap *pmap)
1411 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1413 pmap->pm_active = 0;
1414 pmap->pm_pvhint = NULL;
1415 RB_INIT(&pmap->pm_pvroot);
1416 spin_init(&pmap->pm_spin);
1417 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1418 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1422 * Initialize a preallocated and zeroed pmap structure,
1423 * such as one in a vmspace structure.
1426 pmap_pinit_simple(struct pmap *pmap)
1429 * Misc initialization
1432 pmap->pm_active = 0;
1433 pmap->pm_pvhint = NULL;
1434 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1437 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1440 if (pmap->pm_pmlpv == NULL) {
1441 RB_INIT(&pmap->pm_pvroot);
1442 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1443 spin_init(&pmap->pm_spin);
1444 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1449 pmap_pinit(struct pmap *pmap)
1454 pmap_pinit_simple(pmap);
1455 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1458 * No need to allocate page table space yet but we do need a valid
1459 * page directory table.
1461 if (pmap->pm_pml4 == NULL) {
1463 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1467 * Allocate the page directory page, which wires it even though
1468 * it isn't being entered into some higher level page table (it
1469 * being the highest level). If one is already cached we don't
1470 * have to do anything.
1472 if ((pv = pmap->pm_pmlpv) == NULL) {
1473 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1474 pmap->pm_pmlpv = pv;
1475 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1476 VM_PAGE_TO_PHYS(pv->pv_m));
1480 * Install DMAP and KMAP.
1482 for (j = 0; j < NDMPML4E; ++j) {
1483 pmap->pm_pml4[DMPML4I + j] =
1484 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1485 PG_RW | PG_V | PG_U;
1487 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1490 * install self-referential address mapping entry
1492 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1493 PG_V | PG_RW | PG_A | PG_M;
1495 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1496 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1498 KKASSERT(pmap->pm_pml4[255] == 0);
1499 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1500 KKASSERT(pv->pv_entry.rbe_left == NULL);
1501 KKASSERT(pv->pv_entry.rbe_right == NULL);
1505 * Clean up a pmap structure so it can be physically freed. This routine
1506 * is called by the vmspace dtor function. A great deal of pmap data is
1507 * left passively mapped to improve vmspace management so we have a bit
1508 * of cleanup work to do here.
1511 pmap_puninit(pmap_t pmap)
1516 KKASSERT(pmap->pm_active == 0);
1517 if ((pv = pmap->pm_pmlpv) != NULL) {
1518 if (pv_hold_try(pv) == 0)
1520 p = pmap_remove_pv_page(pv);
1522 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1523 vm_page_busy_wait(p, FALSE, "pgpun");
1524 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1525 vm_page_unwire(p, 0);
1526 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1529 * XXX eventually clean out PML4 static entries and
1530 * use vm_page_free_zero()
1533 pmap->pm_pmlpv = NULL;
1535 if (pmap->pm_pml4) {
1536 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1537 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1538 pmap->pm_pml4 = NULL;
1540 KKASSERT(pmap->pm_stats.resident_count == 0);
1541 KKASSERT(pmap->pm_stats.wired_count == 0);
1545 * Wire in kernel global address entries. To avoid a race condition
1546 * between pmap initialization and pmap_growkernel, this procedure
1547 * adds the pmap to the master list (which growkernel scans to update),
1548 * then copies the template.
1551 pmap_pinit2(struct pmap *pmap)
1553 spin_lock(&pmap_spin);
1554 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1555 spin_unlock(&pmap_spin);
1559 * This routine is called when various levels in the page table need to
1560 * be populated. This routine cannot fail.
1562 * This function returns two locked pv_entry's, one representing the
1563 * requested pv and one representing the requested pv's parent pv. If
1564 * the pv did not previously exist it will be mapped into its parent
1565 * and wired, otherwise no additional wire count will be added.
1569 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1574 vm_pindex_t pt_pindex;
1580 * If the pv already exists and we aren't being asked for the
1581 * parent page table page we can just return it. A locked+held pv
1585 pv = pv_alloc(pmap, ptepindex, &isnew);
1586 if (isnew == 0 && pvpp == NULL)
1590 * This is a new PV, we have to resolve its parent page table and
1591 * add an additional wiring to the page if necessary.
1595 * Special case terminal PVs. These are not page table pages so
1596 * no vm_page is allocated (the caller supplied the vm_page). If
1597 * pvpp is non-NULL we are being asked to also removed the pt_pv
1600 * Note that pt_pv's are only returned for user VAs. We assert that
1601 * a pt_pv is not being requested for kernel VAs.
1603 if (ptepindex < pmap_pt_pindex(0)) {
1604 if (ptepindex >= NUPTE_USER)
1605 KKASSERT(pvpp == NULL);
1607 KKASSERT(pvpp != NULL);
1609 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1610 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1612 vm_page_wire_quick(pvp->pv_m);
1621 * Non-terminal PVs allocate a VM page to represent the page table,
1622 * so we have to resolve pvp and calculate ptepindex for the pvp
1623 * and then for the page table entry index in the pvp for
1626 if (ptepindex < pmap_pd_pindex(0)) {
1628 * pv is PT, pvp is PD
1630 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1631 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1632 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1639 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1640 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1642 } else if (ptepindex < pmap_pdp_pindex(0)) {
1644 * pv is PD, pvp is PDP
1646 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1649 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1650 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1652 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1653 KKASSERT(pvpp == NULL);
1656 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1664 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1665 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1666 } else if (ptepindex < pmap_pml4_pindex()) {
1668 * pv is PDP, pvp is the root pml4 table
1670 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1677 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1678 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1681 * pv represents the top-level PML4, there is no parent.
1689 * This code is only reached if isnew is TRUE and this is not a
1690 * terminal PV. We need to allocate a vm_page for the page table
1691 * at this level and enter it into the parent page table.
1693 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1696 m = vm_page_alloc(NULL, pv->pv_pindex,
1697 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1698 VM_ALLOC_INTERRUPT);
1703 vm_page_spin_lock(m);
1704 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1706 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1707 vm_page_spin_unlock(m);
1708 vm_page_unmanage(m); /* m must be spinunlocked */
1710 if ((m->flags & PG_ZERO) == 0) {
1711 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1715 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1718 m->valid = VM_PAGE_BITS_ALL;
1719 vm_page_flag_clear(m, PG_ZERO);
1720 vm_page_wire(m); /* wire for mapping in parent */
1723 * Wire the page into pvp, bump the wire-count for pvp's page table
1724 * page. Bump the resident_count for the pmap. There is no pvp
1725 * for the top level, address the pm_pml4[] array directly.
1727 * If the caller wants the parent we return it, otherwise
1728 * we just put it away.
1730 * No interlock is needed for pte 0 -> non-zero.
1732 * In the situation where *ptep is valid we might have an unmanaged
1733 * page table page shared from another page table which we need to
1734 * unshare before installing our private page table page.
1737 ptep = pv_pte_lookup(pvp, ptepindex);
1740 pmap_inval_info info;
1743 panic("pmap_allocpte: unexpected pte %p/%d",
1744 pvp, (int)ptepindex);
1746 pmap_inval_init(&info);
1747 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1748 pte = pte_load_clear(ptep);
1749 pmap_inval_deinterlock(&info, pmap);
1750 pmap_inval_done(&info);
1751 if (vm_page_unwire_quick(
1752 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
1753 panic("pmap_allocpte: shared pgtable "
1754 "pg bad wirecount");
1756 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1758 vm_page_wire_quick(pvp->pv_m);
1760 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1773 * This version of pmap_allocpte() checks for possible segment optimizations
1774 * that would allow page-table sharing. It can be called for terminal
1775 * page or page table page ptepindex's.
1777 * The function is called with page table page ptepindex's for fictitious
1778 * and unmanaged terminal pages. That is, we don't want to allocate a
1779 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
1782 * This function can return a pv and *pvpp associated with the passed in pmap
1783 * OR a pv and *pvpp associated with the shared pmap. In the latter case
1784 * an unmanaged page table page will be entered into the pass in pmap.
1788 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
1789 vm_map_entry_t entry, vm_offset_t va)
1791 struct pmap_inval_info info;
1796 pv_entry_t pte_pv; /* in original or shared pmap */
1797 pv_entry_t pt_pv; /* in original or shared pmap */
1798 pv_entry_t proc_pd_pv; /* in original pmap */
1799 pv_entry_t proc_pt_pv; /* in original pmap */
1800 pv_entry_t xpv; /* PT in shared pmap */
1801 pd_entry_t *pt; /* PT entry in PD of original pmap */
1802 pd_entry_t opte; /* contents of *pt */
1803 pd_entry_t npte; /* contents of *pt */
1808 * Basic tests, require a non-NULL vm_map_entry, require proper
1809 * alignment and type for the vm_map_entry, require that the
1810 * underlying object already be allocated.
1812 * We currently allow any type of object to use this optimization.
1813 * The object itself does NOT have to be sized to a multiple of the
1814 * segment size, but the memory mapping does.
1816 if (entry == NULL ||
1817 pmap_mmu_optimize == 0 || /* not enabled */
1818 ptepindex >= pmap_pd_pindex(0) || /* not terminal */
1819 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
1820 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
1821 entry->object.vm_object == NULL || /* needs VM object */
1822 (entry->offset & SEG_MASK) || /* must be aligned */
1823 (entry->start & SEG_MASK)) {
1824 return(pmap_allocpte(pmap, ptepindex, pvpp));
1828 * Make sure the full segment can be represented.
1830 b = va & ~(vm_offset_t)SEG_MASK;
1831 if (b < entry->start && b + SEG_SIZE > entry->end)
1832 return(pmap_allocpte(pmap, ptepindex, pvpp));
1835 * If the full segment can be represented dive the VM object's
1836 * shared pmap, allocating as required.
1838 object = entry->object.vm_object;
1840 if (entry->protection & VM_PROT_WRITE)
1841 obpmapp = &object->md.pmap_rw;
1843 obpmapp = &object->md.pmap_ro;
1846 * We allocate what appears to be a normal pmap but because portions
1847 * of this pmap are shared with other unrelated pmaps we have to
1848 * set pm_active to point to all cpus.
1850 * XXX Currently using pmap_spin to interlock the update, can't use
1851 * vm_object_hold/drop because the token might already be held
1852 * shared OR exclusive and we don't know.
1854 while ((obpmap = *obpmapp) == NULL) {
1855 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
1856 pmap_pinit_simple(obpmap);
1857 pmap_pinit2(obpmap);
1858 spin_lock(&pmap_spin);
1859 if (*obpmapp != NULL) {
1863 spin_unlock(&pmap_spin);
1864 pmap_release(obpmap);
1865 pmap_puninit(obpmap);
1866 kfree(obpmap, M_OBJPMAP);
1868 obpmap->pm_active = smp_active_mask;
1870 spin_unlock(&pmap_spin);
1875 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
1876 * pte/pt using the shared pmap from the object but also adjust
1877 * the process pmap's page table page as a side effect.
1881 * Resolve the terminal PTE and PT in the shared pmap. This is what
1882 * we will return. This is true if ptepindex represents a terminal
1883 * page, otherwise pte_pv is actually the PT and pt_pv is actually
1887 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
1888 if (ptepindex >= pmap_pt_pindex(0))
1894 * Resolve the PD in the process pmap so we can properly share the
1895 * page table page. Lock order is bottom-up (leaf first)!
1897 * NOTE: proc_pt_pv can be NULL.
1899 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
1900 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
1903 * xpv is the page table page pv from the shared object
1904 * (for convenience).
1906 * Calculate the pte value for the PT to load into the process PD.
1907 * If we have to change it we must properly dispose of the previous
1910 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1911 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
1912 (PG_U | PG_RW | PG_V | PG_A | PG_M);
1915 * Dispose of previous page table page if it was local to the
1916 * process pmap. If the old pt is not empty we cannot dispose of it
1917 * until we clean it out. This case should not arise very often so
1918 * it is not optimized.
1921 if (proc_pt_pv->pv_m->wire_count != 1) {
1927 va & ~(vm_offset_t)SEG_MASK,
1928 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
1931 pmap_release_pv(proc_pt_pv, proc_pd_pv);
1934 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1938 * Handle remaining cases.
1942 vm_page_wire_quick(xpv->pv_m);
1943 vm_page_wire_quick(proc_pd_pv->pv_m);
1944 atomic_add_long(&pmap->pm_stats.resident_count, 1);
1945 } else if (*pt != npte) {
1946 pmap_inval_init(&info);
1947 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1949 opte = pte_load_clear(pt);
1950 KKASSERT(opte && opte != npte);
1953 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
1956 * Clean up opte, bump the wire_count for the process
1957 * PD page representing the new entry if it was
1960 * If the entry was not previously empty and we have
1961 * a PT in the proc pmap then opte must match that
1962 * pt. The proc pt must be retired (this is done
1963 * later on in this procedure).
1965 * NOTE: replacing valid pte, wire_count on proc_pd_pv
1968 KKASSERT(opte & PG_V);
1969 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
1970 if (vm_page_unwire_quick(m)) {
1971 panic("pmap_allocpte_seg: "
1972 "bad wire count %p",
1976 pmap_inval_deinterlock(&info, pmap);
1977 pmap_inval_done(&info);
1981 * The existing process page table was replaced and must be destroyed
1995 * Release any resources held by the given physical map.
1997 * Called when a pmap initialized by pmap_pinit is being released. Should
1998 * only be called if the map contains no valid mappings.
2000 * Caller must hold pmap->pm_token
2002 struct pmap_release_info {
2007 static int pmap_release_callback(pv_entry_t pv, void *data);
2010 pmap_release(struct pmap *pmap)
2012 struct pmap_release_info info;
2014 KASSERT(pmap->pm_active == 0,
2015 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
2017 spin_lock(&pmap_spin);
2018 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2019 spin_unlock(&pmap_spin);
2022 * Pull pv's off the RB tree in order from low to high and release
2028 spin_lock(&pmap->pm_spin);
2029 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2030 pmap_release_callback, &info);
2031 spin_unlock(&pmap->pm_spin);
2032 } while (info.retry);
2036 * One resident page (the pml4 page) should remain.
2037 * No wired pages should remain.
2039 KKASSERT(pmap->pm_stats.resident_count ==
2040 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2042 KKASSERT(pmap->pm_stats.wired_count == 0);
2046 pmap_release_callback(pv_entry_t pv, void *data)
2048 struct pmap_release_info *info = data;
2049 pmap_t pmap = info->pmap;
2052 if (pv_hold_try(pv)) {
2053 spin_unlock(&pmap->pm_spin);
2055 spin_unlock(&pmap->pm_spin);
2057 if (pv->pv_pmap != pmap) {
2059 spin_lock(&pmap->pm_spin);
2064 r = pmap_release_pv(pv, NULL);
2065 spin_lock(&pmap->pm_spin);
2070 * Called with held (i.e. also locked) pv. This function will dispose of
2071 * the lock along with the pv.
2073 * If the caller already holds the locked parent page table for pv it
2074 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2075 * pass NULL for pvp.
2078 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp)
2083 * The pmap is currently not spinlocked, pv is held+locked.
2084 * Remove the pv's page from its parent's page table. The
2085 * parent's page table page's wire_count will be decremented.
2087 pmap_remove_pv_pte(pv, pvp, NULL);
2090 * Terminal pvs are unhooked from their vm_pages. Because
2091 * terminal pages aren't page table pages they aren't wired
2092 * by us, so we have to be sure not to unwire them either.
2094 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2095 pmap_remove_pv_page(pv);
2100 * We leave the top-level page table page cached, wired, and
2101 * mapped in the pmap until the dtor function (pmap_puninit())
2104 * Since we are leaving the top-level pv intact we need
2105 * to break out of what would otherwise be an infinite loop.
2107 if (pv->pv_pindex == pmap_pml4_pindex()) {
2113 * For page table pages (other than the top-level page),
2114 * remove and free the vm_page. The representitive mapping
2115 * removed above by pmap_remove_pv_pte() did not undo the
2116 * last wire_count so we have to do that as well.
2118 p = pmap_remove_pv_page(pv);
2119 vm_page_busy_wait(p, FALSE, "pmaprl");
2120 if (p->wire_count != 1) {
2121 kprintf("p->wire_count was %016lx %d\n",
2122 pv->pv_pindex, p->wire_count);
2124 KKASSERT(p->wire_count == 1);
2125 KKASSERT(p->flags & PG_UNMANAGED);
2127 vm_page_unwire(p, 0);
2128 KKASSERT(p->wire_count == 0);
2131 * Theoretically this page, if not the pml4 page, should contain
2132 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2142 * This function will remove the pte associated with a pv from its parent.
2143 * Terminal pv's are supported. The removal will be interlocked if info
2144 * is non-NULL. The caller must dispose of pv instead of just unlocking
2147 * The wire count will be dropped on the parent page table. The wire
2148 * count on the page being removed (pv->pv_m) from the parent page table
2149 * is NOT touched. Note that terminal pages will not have any additional
2150 * wire counts while page table pages will have at least one representing
2151 * the mapping, plus others representing sub-mappings.
2153 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2154 * pages and user page table and terminal pages.
2156 * The pv must be locked.
2158 * XXX must lock parent pv's if they exist to remove pte XXX
2162 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2164 vm_pindex_t ptepindex = pv->pv_pindex;
2165 pmap_t pmap = pv->pv_pmap;
2171 if (ptepindex == pmap_pml4_pindex()) {
2173 * We are the top level pml4 table, there is no parent.
2175 p = pmap->pm_pmlpv->pv_m;
2176 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2178 * Remove a PDP page from the pml4e. This can only occur
2179 * with user page tables. We do not have to lock the
2180 * pml4 PV so just ignore pvp.
2182 vm_pindex_t pml4_pindex;
2183 vm_pindex_t pdp_index;
2186 pdp_index = ptepindex - pmap_pdp_pindex(0);
2188 pml4_pindex = pmap_pml4_pindex();
2189 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2193 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2194 KKASSERT((*pdp & PG_V) != 0);
2195 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2197 KKASSERT(info == NULL);
2198 } else if (ptepindex >= pmap_pd_pindex(0)) {
2200 * Remove a PD page from the pdp
2202 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2203 * of a simple pmap because it stops at
2206 vm_pindex_t pdp_pindex;
2207 vm_pindex_t pd_index;
2210 pd_index = ptepindex - pmap_pd_pindex(0);
2213 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2214 (pd_index >> NPML4EPGSHIFT);
2215 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2220 pd = pv_pte_lookup(pvp, pd_index &
2221 ((1ul << NPDPEPGSHIFT) - 1));
2222 KKASSERT((*pd & PG_V) != 0);
2223 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2226 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2227 p = pv->pv_m; /* degenerate test later */
2229 KKASSERT(info == NULL);
2230 } else if (ptepindex >= pmap_pt_pindex(0)) {
2232 * Remove a PT page from the pd
2234 vm_pindex_t pd_pindex;
2235 vm_pindex_t pt_index;
2238 pt_index = ptepindex - pmap_pt_pindex(0);
2241 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2242 (pt_index >> NPDPEPGSHIFT);
2243 pvp = pv_get(pv->pv_pmap, pd_pindex);
2247 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2248 KKASSERT((*pt & PG_V) != 0);
2249 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2251 KKASSERT(info == NULL);
2254 * Remove a PTE from the PT page
2256 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2257 * pv is a pte_pv so we can safely lock pt_pv.
2259 vm_pindex_t pt_pindex;
2264 pt_pindex = ptepindex >> NPTEPGSHIFT;
2265 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2267 if (ptepindex >= NUPTE_USER) {
2268 ptep = vtopte(ptepindex << PAGE_SHIFT);
2269 KKASSERT(pvp == NULL);
2272 pt_pindex = NUPTE_TOTAL +
2273 (ptepindex >> NPDPEPGSHIFT);
2274 pvp = pv_get(pv->pv_pmap, pt_pindex);
2278 ptep = pv_pte_lookup(pvp, ptepindex &
2279 ((1ul << NPDPEPGSHIFT) - 1));
2283 pmap_inval_interlock(info, pmap, va);
2284 pte = pte_load_clear(ptep);
2286 pmap_inval_deinterlock(info, pmap);
2288 cpu_invlpg((void *)va);
2291 * Now update the vm_page_t
2293 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
2294 kprintf("remove_pte badpte %016lx %016lx %d\n",
2296 pv->pv_pindex < pmap_pt_pindex(0));
2298 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2299 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2302 if (pmap_track_modified(ptepindex))
2306 vm_page_flag_set(p, PG_REFERENCED);
2309 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2311 cpu_invlpg((void *)va);
2315 * Unwire the parent page table page. The wire_count cannot go below
2316 * 1 here because the parent page table page is itself still mapped.
2318 * XXX remove the assertions later.
2320 KKASSERT(pv->pv_m == p);
2321 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2322 panic("pmap_remove_pv_pte: Insufficient wire_count");
2330 pmap_remove_pv_page(pv_entry_t pv)
2336 vm_page_spin_lock(m);
2338 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2341 atomic_add_int(&m->object->agg_pv_list_count, -1);
2343 if (TAILQ_EMPTY(&m->md.pv_list))
2344 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2345 vm_page_spin_unlock(m);
2350 * Grow the number of kernel page table entries, if needed.
2352 * This routine is always called to validate any address space
2353 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2354 * space below KERNBASE.
2357 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2360 vm_offset_t ptppaddr;
2362 pd_entry_t *pt, newpt;
2364 int update_kernel_vm_end;
2367 * bootstrap kernel_vm_end on first real VM use
2369 if (kernel_vm_end == 0) {
2370 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2372 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
2373 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2374 ~(PAGE_SIZE * NPTEPG - 1);
2376 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2377 kernel_vm_end = kernel_map.max_offset;
2384 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2385 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2386 * do not want to force-fill 128G worth of page tables.
2388 if (kstart < KERNBASE) {
2389 if (kstart > kernel_vm_end)
2390 kstart = kernel_vm_end;
2391 KKASSERT(kend <= KERNBASE);
2392 update_kernel_vm_end = 1;
2394 update_kernel_vm_end = 0;
2397 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2398 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2400 if (kend - 1 >= kernel_map.max_offset)
2401 kend = kernel_map.max_offset;
2403 while (kstart < kend) {
2404 pt = pmap_pt(&kernel_pmap, kstart);
2406 /* We need a new PDP entry */
2407 nkpg = vm_page_alloc(NULL, nkpt,
2410 VM_ALLOC_INTERRUPT);
2412 panic("pmap_growkernel: no memory to grow "
2415 paddr = VM_PAGE_TO_PHYS(nkpg);
2416 if ((nkpg->flags & PG_ZERO) == 0)
2417 pmap_zero_page(paddr);
2418 vm_page_flag_clear(nkpg, PG_ZERO);
2419 newpd = (pdp_entry_t)
2420 (paddr | PG_V | PG_RW | PG_A | PG_M);
2421 *pmap_pd(&kernel_pmap, kstart) = newpd;
2423 continue; /* try again */
2425 if ((*pt & PG_V) != 0) {
2426 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2427 ~(PAGE_SIZE * NPTEPG - 1);
2428 if (kstart - 1 >= kernel_map.max_offset) {
2429 kstart = kernel_map.max_offset;
2436 * This index is bogus, but out of the way
2438 nkpg = vm_page_alloc(NULL, nkpt,
2441 VM_ALLOC_INTERRUPT);
2443 panic("pmap_growkernel: no memory to grow kernel");
2446 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2447 pmap_zero_page(ptppaddr);
2448 vm_page_flag_clear(nkpg, PG_ZERO);
2449 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2450 *pmap_pt(&kernel_pmap, kstart) = newpt;
2453 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2454 ~(PAGE_SIZE * NPTEPG - 1);
2456 if (kstart - 1 >= kernel_map.max_offset) {
2457 kstart = kernel_map.max_offset;
2463 * Only update kernel_vm_end for areas below KERNBASE.
2465 if (update_kernel_vm_end && kernel_vm_end < kstart)
2466 kernel_vm_end = kstart;
2470 * Add a reference to the specified pmap.
2473 pmap_reference(pmap_t pmap)
2476 lwkt_gettoken(&pmap->pm_token);
2478 lwkt_reltoken(&pmap->pm_token);
2482 /***************************************************
2483 * page management routines.
2484 ***************************************************/
2487 * Hold a pv without locking it
2490 pv_hold(pv_entry_t pv)
2494 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2498 count = pv->pv_hold;
2500 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2507 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2508 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2511 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2512 * pv list via its page) must be held by the caller.
2515 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2519 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2522 pv->pv_line = lineno;
2528 count = pv->pv_hold;
2530 if ((count & PV_HOLD_LOCKED) == 0) {
2531 if (atomic_cmpset_int(&pv->pv_hold, count,
2532 (count + 1) | PV_HOLD_LOCKED)) {
2535 pv->pv_line = lineno;
2540 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2548 * Drop a previously held pv_entry which could not be locked, allowing its
2551 * Must not be called with a spinlock held as we might zfree() the pv if it
2552 * is no longer associated with a pmap and this was the last hold count.
2555 pv_drop(pv_entry_t pv)
2559 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2560 if (pv->pv_pmap == NULL)
2566 count = pv->pv_hold;
2568 KKASSERT((count & PV_HOLD_MASK) > 0);
2569 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2570 (PV_HOLD_LOCKED | 1));
2571 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2572 if (count == 1 && pv->pv_pmap == NULL)
2581 * Find or allocate the requested PV entry, returning a locked pv
2585 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2588 pv_entry_t pnew = NULL;
2590 spin_lock(&pmap->pm_spin);
2592 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2593 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2598 spin_unlock(&pmap->pm_spin);
2599 pnew = zalloc(pvzone);
2600 spin_lock(&pmap->pm_spin);
2603 pnew->pv_pmap = pmap;
2604 pnew->pv_pindex = pindex;
2605 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2607 pnew->pv_func = func;
2608 pnew->pv_line = lineno;
2610 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2611 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2612 spin_unlock(&pmap->pm_spin);
2617 spin_unlock(&pmap->pm_spin);
2618 zfree(pvzone, pnew);
2620 spin_lock(&pmap->pm_spin);
2623 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2624 spin_unlock(&pmap->pm_spin);
2628 spin_unlock(&pmap->pm_spin);
2629 _pv_lock(pv PMAP_DEBUG_COPY);
2630 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2635 spin_lock(&pmap->pm_spin);
2642 * Find the requested PV entry, returning a locked+held pv or NULL
2646 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2650 spin_lock(&pmap->pm_spin);
2655 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2656 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2660 spin_unlock(&pmap->pm_spin);
2663 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2664 pv_cache(pv, pindex);
2665 spin_unlock(&pmap->pm_spin);
2668 spin_unlock(&pmap->pm_spin);
2669 _pv_lock(pv PMAP_DEBUG_COPY);
2670 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2673 spin_lock(&pmap->pm_spin);
2678 * Lookup, hold, and attempt to lock (pmap,pindex).
2680 * If the entry does not exist NULL is returned and *errorp is set to 0
2682 * If the entry exists and could be successfully locked it is returned and
2683 * errorp is set to 0.
2685 * If the entry exists but could NOT be successfully locked it is returned
2686 * held and *errorp is set to 1.
2690 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2694 spin_lock(&pmap->pm_spin);
2695 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2696 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2698 spin_unlock(&pmap->pm_spin);
2702 if (pv_hold_try(pv)) {
2703 pv_cache(pv, pindex);
2704 spin_unlock(&pmap->pm_spin);
2706 return(pv); /* lock succeeded */
2708 spin_unlock(&pmap->pm_spin);
2710 return (pv); /* lock failed */
2714 * Find the requested PV entry, returning a held pv or NULL
2718 pv_find(pmap_t pmap, vm_pindex_t pindex)
2722 spin_lock(&pmap->pm_spin);
2724 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2725 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2727 spin_unlock(&pmap->pm_spin);
2731 pv_cache(pv, pindex);
2732 spin_unlock(&pmap->pm_spin);
2737 * Lock a held pv, keeping the hold count
2741 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2746 count = pv->pv_hold;
2748 if ((count & PV_HOLD_LOCKED) == 0) {
2749 if (atomic_cmpset_int(&pv->pv_hold, count,
2750 count | PV_HOLD_LOCKED)) {
2753 pv->pv_line = lineno;
2759 tsleep_interlock(pv, 0);
2760 if (atomic_cmpset_int(&pv->pv_hold, count,
2761 count | PV_HOLD_WAITING)) {
2763 kprintf("pv waiting on %s:%d\n",
2764 pv->pv_func, pv->pv_line);
2766 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2773 * Unlock a held and locked pv, keeping the hold count.
2777 pv_unlock(pv_entry_t pv)
2781 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2785 count = pv->pv_hold;
2787 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2788 (PV_HOLD_LOCKED | 1));
2789 if (atomic_cmpset_int(&pv->pv_hold, count,
2791 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2792 if (count & PV_HOLD_WAITING)
2800 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2801 * and the hold count drops to zero we will free it.
2803 * Caller should not hold any spin locks. We are protected from hold races
2804 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2805 * lock held. A pv cannot be located otherwise.
2809 pv_put(pv_entry_t pv)
2811 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2812 if (pv->pv_pmap == NULL)
2821 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2822 * pmap. Any pte operations must have already been completed.
2826 pv_free(pv_entry_t pv)
2830 KKASSERT(pv->pv_m == NULL);
2831 if ((pmap = pv->pv_pmap) != NULL) {
2832 spin_lock(&pmap->pm_spin);
2833 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2834 if (pmap->pm_pvhint == pv)
2835 pmap->pm_pvhint = NULL;
2836 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2839 spin_unlock(&pmap->pm_spin);
2845 * This routine is very drastic, but can save the system
2853 static int warningdone=0;
2855 if (pmap_pagedaemon_waken == 0)
2857 pmap_pagedaemon_waken = 0;
2858 if (warningdone < 5) {
2859 kprintf("pmap_collect: collecting pv entries -- "
2860 "suggest increasing PMAP_SHPGPERPROC\n");
2864 for (i = 0; i < vm_page_array_size; i++) {
2865 m = &vm_page_array[i];
2866 if (m->wire_count || m->hold_count)
2868 if (vm_page_busy_try(m, TRUE) == 0) {
2869 if (m->wire_count == 0 && m->hold_count == 0) {
2878 * Scan the pmap for active page table entries and issue a callback.
2879 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
2880 * its parent page table.
2882 * pte_pv will be NULL if the page or page table is unmanaged.
2883 * pt_pv will point to the page table page containing the pte for the page.
2885 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
2886 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
2887 * process pmap's PD and page to the callback function. This can be
2888 * confusing because the pt_pv is really a pd_pv, and the target page
2889 * table page is simply aliased by the pmap and not owned by it.
2891 * It is assumed that the start and end are properly rounded to the page size.
2893 * It is assumed that PD pages and above are managed and thus in the RB tree,
2894 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
2896 struct pmap_scan_info {
2900 vm_pindex_t sva_pd_pindex;
2901 vm_pindex_t eva_pd_pindex;
2902 void (*func)(pmap_t, struct pmap_scan_info *,
2903 pv_entry_t, pv_entry_t, int, vm_offset_t,
2904 pt_entry_t *, void *);
2907 struct pmap_inval_info inval;
2910 static int pmap_scan_cmp(pv_entry_t pv, void *data);
2911 static int pmap_scan_callback(pv_entry_t pv, void *data);
2914 pmap_scan(struct pmap_scan_info *info)
2916 struct pmap *pmap = info->pmap;
2917 pv_entry_t pd_pv; /* A page directory PV */
2918 pv_entry_t pt_pv; /* A page table PV */
2919 pv_entry_t pte_pv; /* A page table entry PV */
2921 struct pv_entry dummy_pv;
2927 * Hold the token for stability; if the pmap is empty we have nothing
2930 lwkt_gettoken(&pmap->pm_token);
2932 if (pmap->pm_stats.resident_count == 0) {
2933 lwkt_reltoken(&pmap->pm_token);
2938 pmap_inval_init(&info->inval);
2941 * Special handling for scanning one page, which is a very common
2942 * operation (it is?).
2944 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
2946 if (info->sva + PAGE_SIZE == info->eva) {
2947 if (info->sva >= VM_MAX_USER_ADDRESS) {
2949 * Kernel mappings do not track wire counts on
2950 * page table pages and only maintain pd_pv and
2951 * pte_pv levels so pmap_scan() works.
2954 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
2955 ptep = vtopte(info->sva);
2958 * User pages which are unmanaged will not have a
2959 * pte_pv. User page table pages which are unmanaged
2960 * (shared from elsewhere) will also not have a pt_pv.
2961 * The func() callback will pass both pte_pv and pt_pv
2962 * as NULL in that case.
2964 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
2965 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
2966 if (pt_pv == NULL) {
2967 KKASSERT(pte_pv == NULL);
2968 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
2970 ptep = pv_pte_lookup(pd_pv,
2971 pmap_pt_index(info->sva));
2973 info->func(pmap, info,
2982 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
2986 * Unlike the pv_find() case below we actually
2987 * acquired a locked pv in this case so any
2988 * race should have been resolved. It is expected
2991 KKASSERT(pte_pv == NULL);
2992 } else if (pte_pv) {
2993 KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED|
2995 ("bad *ptep %016lx sva %016lx pte_pv %p",
2996 *ptep, info->sva, pte_pv));
2997 info->func(pmap, info, pte_pv, pt_pv, 0,
2998 info->sva, ptep, info->arg);
3000 KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V,
3001 ("bad *ptep %016lx sva %016lx pte_pv NULL",
3003 info->func(pmap, info, NULL, pt_pv, 0,
3004 info->sva, ptep, info->arg);
3009 pmap_inval_done(&info->inval);
3010 lwkt_reltoken(&pmap->pm_token);
3015 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3018 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3019 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3021 if (info->sva >= VM_MAX_USER_ADDRESS) {
3023 * The kernel does not currently maintain any pv_entry's for
3024 * higher-level page tables.
3026 bzero(&dummy_pv, sizeof(dummy_pv));
3027 dummy_pv.pv_pindex = info->sva_pd_pindex;
3028 spin_lock(&pmap->pm_spin);
3029 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3030 pmap_scan_callback(&dummy_pv, info);
3031 ++dummy_pv.pv_pindex;
3033 spin_unlock(&pmap->pm_spin);
3036 * User page tables maintain local PML4, PDP, and PD
3037 * pv_entry's at the very least. PT pv's might be
3038 * unmanaged and thus not exist. PTE pv's might be
3039 * unmanaged and thus not exist.
3041 spin_lock(&pmap->pm_spin);
3042 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3043 pmap_scan_cmp, pmap_scan_callback, info);
3044 spin_unlock(&pmap->pm_spin);
3046 pmap_inval_done(&info->inval);
3047 lwkt_reltoken(&pmap->pm_token);
3051 * WARNING! pmap->pm_spin held
3054 pmap_scan_cmp(pv_entry_t pv, void *data)
3056 struct pmap_scan_info *info = data;
3057 if (pv->pv_pindex < info->sva_pd_pindex)
3059 if (pv->pv_pindex >= info->eva_pd_pindex)
3065 * WARNING! pmap->pm_spin held
3068 pmap_scan_callback(pv_entry_t pv, void *data)
3070 struct pmap_scan_info *info = data;
3071 struct pmap *pmap = info->pmap;
3072 pv_entry_t pd_pv; /* A page directory PV */
3073 pv_entry_t pt_pv; /* A page table PV */
3074 pv_entry_t pte_pv; /* A page table entry PV */
3078 vm_offset_t va_next;
3079 vm_pindex_t pd_pindex;
3083 * Pull the PD pindex from the pv before releasing the spinlock.
3085 * WARNING: pv is faked for kernel pmap scans.
3087 pd_pindex = pv->pv_pindex;
3088 spin_unlock(&pmap->pm_spin);
3089 pv = NULL; /* invalid after spinlock unlocked */
3092 * Calculate the page range within the PD. SIMPLE pmaps are
3093 * direct-mapped for the entire 2^64 address space. Normal pmaps
3094 * reflect the user and kernel address space which requires
3095 * cannonicalization w/regards to converting pd_pindex's back
3098 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3099 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3100 (sva & PML4_SIGNMASK)) {
3101 sva |= PML4_SIGNMASK;
3103 eva = sva + NBPDP; /* can overflow */
3104 if (sva < info->sva)
3106 if (eva < info->sva || eva > info->eva)
3110 * NOTE: kernel mappings do not track page table pages, only
3113 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3114 * However, for the scan to be efficient we try to
3115 * cache items top-down.
3120 for (; sva < eva; sva = va_next) {
3121 if (sva >= VM_MAX_USER_ADDRESS) {
3130 * PD cache (degenerate case if we skip). It is possible
3131 * for the PD to not exist due to races. This is ok.
3133 if (pd_pv == NULL) {
3134 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3135 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3137 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3139 if (pd_pv == NULL) {
3140 va_next = (sva + NBPDP) & ~PDPMASK;
3149 if (pt_pv == NULL) {
3154 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3155 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3161 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3165 * If pt_pv is NULL we either have an shared page table
3166 * page and must issue a callback specific to that case,
3167 * or there is no page table page.
3169 * Either way we can skip the page table page.
3171 if (pt_pv == NULL) {
3173 * Possible unmanaged (shared from another pmap)
3177 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3178 KKASSERT(pd_pv != NULL);
3179 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3181 info->func(pmap, info, NULL, pd_pv, 1,
3182 sva, ptep, info->arg);
3186 * Done, move to next page table page.
3188 va_next = (sva + NBPDR) & ~PDRMASK;
3195 * From this point in the loop testing pt_pv for non-NULL
3196 * means we are in UVM, else if it is NULL we are in KVM.
3198 * Limit our scan to either the end of the va represented
3199 * by the current page table page, or to the end of the
3200 * range being removed.
3203 va_next = (sva + NBPDR) & ~PDRMASK;
3210 * Scan the page table for pages. Some pages may not be
3211 * managed (might not have a pv_entry).
3213 * There is no page table management for kernel pages so
3214 * pt_pv will be NULL in that case, but otherwise pt_pv
3215 * is non-NULL, locked, and referenced.
3219 * At this point a non-NULL pt_pv means a UVA, and a NULL
3220 * pt_pv means a KVA.
3223 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3227 while (sva < va_next) {
3229 * Acquire the related pte_pv, if any. If *ptep == 0
3230 * the related pte_pv should not exist, but if *ptep
3231 * is not zero the pte_pv may or may not exist (e.g.
3232 * will not exist for an unmanaged page).
3234 * However a multitude of races are possible here.
3236 * In addition, the (pt_pv, pte_pv) lock order is
3237 * backwards, so we have to be careful in aquiring
3238 * a properly locked pte_pv.
3241 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3248 pv_put(pt_pv); /* must be non-NULL */
3250 pv_lock(pte_pv); /* safe to block now */
3253 pt_pv = pv_get(pmap,
3254 pmap_pt_pindex(sva));
3256 * pt_pv reloaded, need new ptep
3258 KKASSERT(pt_pv != NULL);
3259 ptep = pv_pte_lookup(pt_pv,
3260 pmap_pte_index(sva));
3264 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3268 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3272 kprintf("Unexpected non-NULL pte_pv "
3273 "%p pt_pv %p *ptep = %016lx\n",
3274 pte_pv, pt_pv, *ptep);
3275 panic("Unexpected non-NULL pte_pv");
3283 * Ready for the callback. The locked pte_pv (if any)
3284 * is consumed by the callback. pte_pv will exist if
3285 * the page is managed, and will not exist if it
3289 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3291 ("bad *ptep %016lx sva %016lx "
3293 *ptep, sva, pte_pv));
3294 info->func(pmap, info, pte_pv, pt_pv, 0,
3295 sva, ptep, info->arg);
3297 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3299 ("bad *ptep %016lx sva %016lx "
3302 info->func(pmap, info, NULL, pt_pv, 0,
3303 sva, ptep, info->arg);
3322 * Relock before returning.
3324 spin_lock(&pmap->pm_spin);
3329 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3331 struct pmap_scan_info info;
3336 info.func = pmap_remove_callback;
3338 info.doinval = 1; /* normal remove requires pmap inval */
3343 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3345 struct pmap_scan_info info;
3350 info.func = pmap_remove_callback;
3352 info.doinval = 0; /* normal remove requires pmap inval */
3357 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
3358 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3359 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3365 * This will also drop pt_pv's wire_count. Note that
3366 * terminal pages are not wired based on mmu presence.
3369 pmap_remove_pv_pte(pte_pv, pt_pv, &info->inval);
3371 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3372 pmap_remove_pv_page(pte_pv);
3374 } else if (sharept == 0) {
3378 * pt_pv's wire_count is still bumped by unmanaged pages
3379 * so we must decrement it manually.
3382 pmap_inval_interlock(&info->inval, pmap, va);
3383 pte = pte_load_clear(ptep);
3385 pmap_inval_deinterlock(&info->inval, pmap);
3387 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3388 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3389 if (vm_page_unwire_quick(pt_pv->pv_m))
3390 panic("pmap_remove: insufficient wirecount");
3393 * Unmanaged page table, pt_pv is actually the pd_pv
3394 * for our pmap (not the share object pmap).
3396 * We have to unwire the target page table page and we
3397 * have to unwire our page directory page.
3400 pmap_inval_interlock(&info->inval, pmap, va);
3401 pte = pte_load_clear(ptep);
3403 pmap_inval_deinterlock(&info->inval, pmap);
3404 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3405 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3406 panic("pmap_remove: shared pgtable1 bad wirecount");
3407 if (vm_page_unwire_quick(pt_pv->pv_m))
3408 panic("pmap_remove: shared pgtable2 bad wirecount");
3413 * Removes this physical page from all physical maps in which it resides.
3414 * Reflects back modify bits to the pager.
3416 * This routine may not be called from an interrupt.
3420 pmap_remove_all(vm_page_t m)
3422 struct pmap_inval_info info;
3425 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3428 pmap_inval_init(&info);
3429 vm_page_spin_lock(m);
3430 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3431 KKASSERT(pv->pv_m == m);
3432 if (pv_hold_try(pv)) {
3433 vm_page_spin_unlock(m);
3435 vm_page_spin_unlock(m);
3437 if (pv->pv_m != m) {
3439 vm_page_spin_lock(m);
3444 * Holding no spinlocks, pv is locked.
3446 pmap_remove_pv_pte(pv, NULL, &info);
3447 pmap_remove_pv_page(pv);
3449 vm_page_spin_lock(m);
3451 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3452 vm_page_spin_unlock(m);
3453 pmap_inval_done(&info);
3457 * Set the physical protection on the specified range of this map
3458 * as requested. This function is typically only used for debug watchpoints
3461 * This function may not be called from an interrupt if the map is
3462 * not the kernel_pmap.
3464 * NOTE! For shared page table pages we just unmap the page.
3467 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3469 struct pmap_scan_info info;
3470 /* JG review for NX */
3474 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3475 pmap_remove(pmap, sva, eva);
3478 if (prot & VM_PROT_WRITE)
3483 info.func = pmap_protect_callback;
3491 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
3492 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3493 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3503 pmap_inval_interlock(&info->inval, pmap, va);
3510 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3511 KKASSERT(m == pte_pv->pv_m);
3512 vm_page_flag_set(m, PG_REFERENCED);
3516 if (pmap_track_modified(pte_pv->pv_pindex)) {
3518 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3523 } else if (sharept) {
3525 * Unmanaged page table, pt_pv is actually the pd_pv
3526 * for our pmap (not the share object pmap).
3528 * When asked to protect something in a shared page table
3529 * page we just unmap the page table page. We have to
3530 * invalidate the tlb in this situation.
3532 pte = pte_load_clear(ptep);
3533 pmap_inval_invltlb(&info->inval);
3534 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3535 panic("pmap_protect: pgtable1 pg bad wirecount");
3536 if (vm_page_unwire_quick(pt_pv->pv_m))
3537 panic("pmap_protect: pgtable2 pg bad wirecount");
3540 /* else unmanaged page, adjust bits, no wire changes */
3544 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
3548 pmap_inval_deinterlock(&info->inval, pmap);
3554 * Insert the vm_page (m) at the virtual address (va), replacing any prior
3555 * mapping at that address. Set protection and wiring as requested.
3557 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
3558 * possible. If it is we enter the page into the appropriate shared pmap
3559 * hanging off the related VM object instead of the passed pmap, then we
3560 * share the page table page from the VM object's pmap into the current pmap.
3562 * NOTE: This routine MUST insert the page into the pmap now, it cannot
3566 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3567 boolean_t wired, vm_map_entry_t entry)
3569 pmap_inval_info info;
3570 pv_entry_t pt_pv; /* page table */
3571 pv_entry_t pte_pv; /* page table entry */
3574 pt_entry_t origpte, newpte;
3579 va = trunc_page(va);
3580 #ifdef PMAP_DIAGNOSTIC
3582 panic("pmap_enter: toobig");
3583 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
3584 panic("pmap_enter: invalid to pmap_enter page table "
3585 "pages (va: 0x%lx)", va);
3587 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
3588 kprintf("Warning: pmap_enter called on UVA with "
3591 db_print_backtrace();
3594 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
3595 kprintf("Warning: pmap_enter called on KVA without"
3598 db_print_backtrace();
3603 * Get locked PV entries for our new page table entry (pte_pv)
3604 * and for its parent page table (pt_pv). We need the parent
3605 * so we can resolve the location of the ptep.
3607 * Only hardware MMU actions can modify the ptep out from
3610 * if (m) is fictitious or unmanaged we do not create a managing
3611 * pte_pv for it. Any pre-existing page's management state must
3612 * match (avoiding code complexity).
3614 * If the pmap is still being initialized we assume existing
3617 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3618 * pmap_allocpte() checks the
3620 if (pmap_initialized == FALSE) {
3624 } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) { /* XXX */
3626 if (va >= VM_MAX_USER_ADDRESS) {
3630 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
3632 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3634 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0);
3636 if (va >= VM_MAX_USER_ADDRESS) {
3638 * Kernel map, pv_entry-tracked.
3641 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3647 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
3649 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3651 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED));
3654 pa = VM_PAGE_TO_PHYS(m);
3656 opa = origpte & PG_FRAME;
3658 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V | PG_A);
3661 if (va < VM_MAX_USER_ADDRESS)
3664 newpte |= PG_MANAGED;
3665 if (pmap == &kernel_pmap)
3667 newpte |= pat_pte_index[m->pat_mode];
3670 * It is possible for multiple faults to occur in threaded
3671 * environments, the existing pte might be correct.
3673 if (((origpte ^ newpte) & ~(pt_entry_t)(PG_M|PG_A)) == 0)
3676 if ((prot & VM_PROT_NOSYNC) == 0)
3677 pmap_inval_init(&info);
3680 * Ok, either the address changed or the protection or wiring
3683 * Clear the current entry, interlocking the removal. For managed
3684 * pte's this will also flush the modified state to the vm_page.
3685 * Atomic ops are mandatory in order to ensure that PG_M events are
3686 * not lost during any transition.
3691 * pmap_remove_pv_pte() unwires pt_pv and assumes
3692 * we will free pte_pv, but since we are reusing
3693 * pte_pv we want to retain the wire count.
3695 * pt_pv won't exist for a kernel page (managed or
3699 vm_page_wire_quick(pt_pv->pv_m);
3700 if (prot & VM_PROT_NOSYNC)
3701 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3703 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3705 pmap_remove_pv_page(pte_pv);
3706 } else if (prot & VM_PROT_NOSYNC) {
3708 * Unmanaged page, NOSYNC (no mmu sync) requested.
3710 * Leave wire count on PT page intact.
3712 (void)pte_load_clear(ptep);
3713 cpu_invlpg((void *)va);
3714 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3717 * Unmanaged page, normal enter.
3719 * Leave wire count on PT page intact.
3721 pmap_inval_interlock(&info, pmap, va);
3722 (void)pte_load_clear(ptep);
3723 pmap_inval_deinterlock(&info, pmap);
3724 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3726 KKASSERT(*ptep == 0);
3731 * Enter on the PV list if part of our managed memory.
3732 * Wiring of the PT page is already handled.
3734 KKASSERT(pte_pv->pv_m == NULL);
3735 vm_page_spin_lock(m);
3737 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3740 atomic_add_int(&m->object->agg_pv_list_count, 1);
3742 vm_page_flag_set(m, PG_MAPPED);
3743 vm_page_spin_unlock(m);
3744 } else if (pt_pv && opa == 0) {
3746 * We have to adjust the wire count on the PT page ourselves
3747 * for unmanaged entries. If opa was non-zero we retained
3748 * the existing wire count from the removal.
3750 vm_page_wire_quick(pt_pv->pv_m);
3754 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
3756 * User VMAs do not because those will be zero->non-zero, so no
3757 * stale entries to worry about at this point.
3759 * For KVM there appear to still be issues. Theoretically we
3760 * should be able to scrap the interlocks entirely but we
3763 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3764 pmap_inval_interlock(&info, pmap, va);
3769 *(volatile pt_entry_t *)ptep = newpte;
3771 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3772 pmap_inval_deinterlock(&info, pmap);
3773 else if (pt_pv == NULL)
3774 cpu_invlpg((void *)va);
3778 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
3781 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3785 vm_page_flag_set(m, PG_WRITEABLE);
3788 * Unmanaged pages need manual resident_count tracking.
3790 if (pte_pv == NULL && pt_pv)
3791 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
3796 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
3797 pmap_inval_done(&info);
3799 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
3802 * Cleanup the pv entry, allowing other accessors.
3811 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
3812 * This code also assumes that the pmap has no pre-existing entry for this
3815 * This code currently may only be used on user pmaps, not kernel_pmap.
3818 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
3820 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
3824 * Make a temporary mapping for a physical address. This is only intended
3825 * to be used for panic dumps.
3827 * The caller is responsible for calling smp_invltlb().
3830 pmap_kenter_temporary(vm_paddr_t pa, long i)
3832 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
3833 return ((void *)crashdumpmap);
3836 #define MAX_INIT_PT (96)
3839 * This routine preloads the ptes for a given object into the specified pmap.
3840 * This eliminates the blast of soft faults on process startup and
3841 * immediately after an mmap.
3843 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
3846 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
3847 vm_object_t object, vm_pindex_t pindex,
3848 vm_size_t size, int limit)
3850 struct rb_vm_page_scan_info info;
3855 * We can't preinit if read access isn't set or there is no pmap
3858 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
3862 * We can't preinit if the pmap is not the current pmap
3864 lp = curthread->td_lwp;
3865 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
3869 * Misc additional checks
3871 psize = x86_64_btop(size);
3873 if ((object->type != OBJT_VNODE) ||
3874 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
3875 (object->resident_page_count > MAX_INIT_PT))) {
3879 if (pindex + psize > object->size) {
3880 if (object->size < pindex)
3882 psize = object->size - pindex;
3889 * If everything is segment-aligned do not pre-init here. Instead
3890 * allow the normal vm_fault path to pass a segment hint to
3891 * pmap_enter() which will then use an object-referenced shared
3894 if ((addr & SEG_MASK) == 0 &&
3895 (ctob(psize) & SEG_MASK) == 0 &&
3896 (ctob(pindex) & SEG_MASK) == 0) {
3901 * Use a red-black scan to traverse the requested range and load
3902 * any valid pages found into the pmap.
3904 * We cannot safely scan the object's memq without holding the
3907 info.start_pindex = pindex;
3908 info.end_pindex = pindex + psize - 1;
3914 vm_object_hold_shared(object);
3915 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3916 pmap_object_init_pt_callback, &info);
3917 vm_object_drop(object);
3922 pmap_object_init_pt_callback(vm_page_t p, void *data)
3924 struct rb_vm_page_scan_info *info = data;
3925 vm_pindex_t rel_index;
3928 * don't allow an madvise to blow away our really
3929 * free pages allocating pv entries.
3931 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3932 vmstats.v_free_count < vmstats.v_free_reserved) {
3937 * Ignore list markers and ignore pages we cannot instantly
3938 * busy (while holding the object token).
3940 if (p->flags & PG_MARKER)
3942 if (vm_page_busy_try(p, TRUE))
3944 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3945 (p->flags & PG_FICTITIOUS) == 0) {
3946 if ((p->queue - p->pc) == PQ_CACHE)
3947 vm_page_deactivate(p);
3948 rel_index = p->pindex - info->start_pindex;
3949 pmap_enter_quick(info->pmap,
3950 info->addr + x86_64_ptob(rel_index), p);
3958 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
3961 * Returns FALSE if it would be non-trivial or if a pte is already loaded
3964 * XXX This is safe only because page table pages are not freed.
3967 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3971 /*spin_lock(&pmap->pm_spin);*/
3972 if ((pte = pmap_pte(pmap, addr)) != NULL) {
3974 /*spin_unlock(&pmap->pm_spin);*/
3978 /*spin_unlock(&pmap->pm_spin);*/
3983 * Change the wiring attribute for a pmap/va pair. The mapping must already
3984 * exist in the pmap. The mapping may or may not be managed.
3987 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
3988 vm_map_entry_t entry)
3995 lwkt_gettoken(&pmap->pm_token);
3996 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
3997 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
3999 if (wired && !pmap_pte_w(ptep))
4000 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4001 else if (!wired && pmap_pte_w(ptep))
4002 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4005 * Wiring is not a hardware characteristic so there is no need to
4006 * invalidate TLB. However, in an SMP environment we must use
4007 * a locked bus cycle to update the pte (if we are not using
4008 * the pmap_inval_*() API that is)... it's ok to do this for simple
4012 atomic_set_long(ptep, PG_W);
4014 atomic_clear_long(ptep, PG_W);
4016 lwkt_reltoken(&pmap->pm_token);
4022 * Copy the range specified by src_addr/len from the source map to
4023 * the range dst_addr/len in the destination map.
4025 * This routine is only advisory and need not do anything.
4028 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4029 vm_size_t len, vm_offset_t src_addr)
4036 * Zero the specified physical page.
4038 * This function may be called from an interrupt and no locking is
4042 pmap_zero_page(vm_paddr_t phys)
4044 vm_offset_t va = PHYS_TO_DMAP(phys);
4046 pagezero((void *)va);
4050 * pmap_page_assertzero:
4052 * Assert that a page is empty, panic if it isn't.
4055 pmap_page_assertzero(vm_paddr_t phys)
4057 vm_offset_t va = PHYS_TO_DMAP(phys);
4060 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
4061 if (*(long *)((char *)va + i) != 0) {
4062 panic("pmap_page_assertzero() @ %p not zero!",
4063 (void *)(intptr_t)va);
4071 * Zero part of a physical page by mapping it into memory and clearing
4072 * its contents with bzero.
4074 * off and size may not cover an area beyond a single hardware page.
4077 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4079 vm_offset_t virt = PHYS_TO_DMAP(phys);
4081 bzero((char *)virt + off, size);
4087 * Copy the physical page from the source PA to the target PA.
4088 * This function may be called from an interrupt. No locking
4092 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4094 vm_offset_t src_virt, dst_virt;
4096 src_virt = PHYS_TO_DMAP(src);
4097 dst_virt = PHYS_TO_DMAP(dst);
4098 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4102 * pmap_copy_page_frag:
4104 * Copy the physical page from the source PA to the target PA.
4105 * This function may be called from an interrupt. No locking
4109 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4111 vm_offset_t src_virt, dst_virt;
4113 src_virt = PHYS_TO_DMAP(src);
4114 dst_virt = PHYS_TO_DMAP(dst);
4116 bcopy((char *)src_virt + (src & PAGE_MASK),
4117 (char *)dst_virt + (dst & PAGE_MASK),
4122 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4123 * this page. This count may be changed upwards or downwards in the future;
4124 * it is only necessary that true be returned for a small subset of pmaps
4125 * for proper page aging.
4128 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4133 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4136 vm_page_spin_lock(m);
4137 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4138 if (pv->pv_pmap == pmap) {
4139 vm_page_spin_unlock(m);
4146 vm_page_spin_unlock(m);
4151 * Remove all pages from specified address space this aids process exit
4152 * speeds. Also, this code may be special cased for the current process
4156 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4158 pmap_remove_noinval(pmap, sva, eva);
4163 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4164 * routines are inline, and a lot of things compile-time evaluate.
4168 pmap_testbit(vm_page_t m, int bit)
4173 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4176 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4178 vm_page_spin_lock(m);
4179 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4180 vm_page_spin_unlock(m);
4184 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4186 * if the bit being tested is the modified bit, then
4187 * mark clean_map and ptes as never
4190 if (bit & (PG_A|PG_M)) {
4191 if (!pmap_track_modified(pv->pv_pindex))
4195 #if defined(PMAP_DIAGNOSTIC)
4196 if (pv->pv_pmap == NULL) {
4197 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4202 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4204 vm_page_spin_unlock(m);
4208 vm_page_spin_unlock(m);
4213 * This routine is used to modify bits in ptes. Only one bit should be
4214 * specified. PG_RW requires special handling.
4216 * Caller must NOT hold any spin locks
4220 pmap_clearbit(vm_page_t m, int bit)
4222 struct pmap_inval_info info;
4229 vm_page_flag_clear(m, PG_WRITEABLE);
4230 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4237 * Loop over all current mappings setting/clearing as appropos If
4238 * setting RO do we need to clear the VAC?
4240 * NOTE: When clearing PG_M we could also (not implemented) drop
4241 * through to the PG_RW code and clear PG_RW too, forcing
4242 * a fault on write to redetect PG_M for virtual kernels, but
4243 * it isn't necessary since virtual kernels invalidate the
4244 * pte when they clear the VPTE_M bit in their virtual page
4247 * NOTE: Does not re-dirty the page when clearing only PG_M.
4249 if ((bit & PG_RW) == 0) {
4250 vm_page_spin_lock(m);
4251 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4252 #if defined(PMAP_DIAGNOSTIC)
4253 if (pv->pv_pmap == NULL) {
4254 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4259 pte = pmap_pte_quick(pv->pv_pmap,
4260 pv->pv_pindex << PAGE_SHIFT);
4263 atomic_clear_long(pte, bit);
4265 vm_page_spin_unlock(m);
4270 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4273 pmap_inval_init(&info);
4276 vm_page_spin_lock(m);
4277 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4279 * don't write protect pager mappings
4281 if (!pmap_track_modified(pv->pv_pindex))
4284 #if defined(PMAP_DIAGNOSTIC)
4285 if (pv->pv_pmap == NULL) {
4286 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4292 * Skip pages which do not have PG_RW set.
4294 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4295 if ((*pte & PG_RW) == 0)
4301 if (pv_hold_try(pv) == 0) {
4302 vm_page_spin_unlock(m);
4303 pv_lock(pv); /* held, now do a blocking lock */
4304 pv_put(pv); /* and release */
4305 goto restart; /* anything could have happened */
4308 save_pmap = pv->pv_pmap;
4309 vm_page_spin_unlock(m);
4310 pmap_inval_interlock(&info, save_pmap,
4311 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
4312 KKASSERT(pv->pv_pmap == save_pmap);
4316 if (atomic_cmpset_long(pte, pbits,
4317 pbits & ~(PG_RW|PG_M))) {
4321 pmap_inval_deinterlock(&info, save_pmap);
4322 vm_page_spin_lock(m);
4325 * If PG_M was found to be set while we were clearing PG_RW
4326 * we also clear PG_M (done above) and mark the page dirty.
4327 * Callers expect this behavior.
4333 vm_page_spin_unlock(m);
4334 pmap_inval_done(&info);
4338 * Lower the permission for all mappings to a given page.
4340 * Page must be busied by caller.
4343 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4345 /* JG NX support? */
4346 if ((prot & VM_PROT_WRITE) == 0) {
4347 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4349 * NOTE: pmap_clearbit(.. PG_RW) also clears
4350 * the PG_WRITEABLE flag in (m).
4352 pmap_clearbit(m, PG_RW);
4360 pmap_phys_address(vm_pindex_t ppn)
4362 return (x86_64_ptob(ppn));
4366 * Return a count of reference bits for a page, clearing those bits.
4367 * It is not necessary for every reference bit to be cleared, but it
4368 * is necessary that 0 only be returned when there are truly no
4369 * reference bits set.
4371 * XXX: The exact number of bits to check and clear is a matter that
4372 * should be tested and standardized at some point in the future for
4373 * optimal aging of shared pages.
4375 * This routine may not block.
4378 pmap_ts_referenced(vm_page_t m)
4384 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4387 vm_page_spin_lock(m);
4388 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4389 if (!pmap_track_modified(pv->pv_pindex))
4391 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4392 if (pte && (*pte & PG_A)) {
4393 atomic_clear_long(pte, PG_A);
4399 vm_page_spin_unlock(m);
4406 * Return whether or not the specified physical page was modified
4407 * in any physical maps.
4410 pmap_is_modified(vm_page_t m)
4414 res = pmap_testbit(m, PG_M);
4419 * Clear the modify bits on the specified physical page.
4422 pmap_clear_modify(vm_page_t m)
4424 pmap_clearbit(m, PG_M);
4428 * pmap_clear_reference:
4430 * Clear the reference bit on the specified physical page.
4433 pmap_clear_reference(vm_page_t m)
4435 pmap_clearbit(m, PG_A);
4439 * Miscellaneous support routines follow
4444 i386_protection_init(void)
4448 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4449 kp = protection_codes;
4450 for (prot = 0; prot < 8; prot++) {
4452 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4454 * Read access is also 0. There isn't any execute bit,
4455 * so just make it readable.
4457 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4458 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4459 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4462 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4463 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4464 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4465 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4473 * Map a set of physical memory pages into the kernel virtual
4474 * address space. Return a pointer to where it is mapped. This
4475 * routine is intended to be used for mapping device memory,
4478 * NOTE: We can't use pgeflag unless we invalidate the pages one at
4481 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
4482 * work whether the cpu supports PAT or not. The remaining PAT
4483 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
4487 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
4489 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4493 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
4495 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
4499 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
4501 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4505 * Map a set of physical memory pages into the kernel virtual
4506 * address space. Return a pointer to where it is mapped. This
4507 * routine is intended to be used for mapping device memory,
4511 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
4513 vm_offset_t va, tmpva, offset;
4517 offset = pa & PAGE_MASK;
4518 size = roundup(offset + size, PAGE_SIZE);
4520 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4522 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4524 pa = pa & ~PAGE_MASK;
4525 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
4526 pte = vtopte(tmpva);
4527 *pte = pa | PG_RW | PG_V | /* pgeflag | */
4528 pat_pte_index[mode];
4529 tmpsize -= PAGE_SIZE;
4533 pmap_invalidate_range(&kernel_pmap, va, va + size);
4534 pmap_invalidate_cache_range(va, va + size);
4536 return ((void *)(va + offset));
4540 pmap_unmapdev(vm_offset_t va, vm_size_t size)
4542 vm_offset_t base, offset;
4544 base = va & ~PAGE_MASK;
4545 offset = va & PAGE_MASK;
4546 size = roundup(offset + size, PAGE_SIZE);
4547 pmap_qremove(va, size >> PAGE_SHIFT);
4548 kmem_free(&kernel_map, base, size);
4552 * Change the PAT attribute on an existing kernel memory map. Caller
4553 * must ensure that the virtual memory in question is not accessed
4554 * during the adjustment.
4557 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
4564 panic("pmap_change_attr: va is NULL");
4565 base = trunc_page(va);
4569 *pte = (*pte & ~(pt_entry_t)(PG_PTE_PAT | PG_NC_PCD |
4571 pat_pte_index[mode];
4576 changed = 1; /* XXX: not optimal */
4579 * Flush CPU caches if required to make sure any data isn't cached that
4580 * shouldn't be, etc.
4583 pmap_invalidate_range(&kernel_pmap, base, va);
4584 pmap_invalidate_cache_range(base, va);
4589 * perform the pmap work for mincore
4592 pmap_mincore(pmap_t pmap, vm_offset_t addr)
4594 pt_entry_t *ptep, pte;
4598 lwkt_gettoken(&pmap->pm_token);
4599 ptep = pmap_pte(pmap, addr);
4601 if (ptep && (pte = *ptep) != 0) {
4604 val = MINCORE_INCORE;
4605 if ((pte & PG_MANAGED) == 0)
4608 pa = pte & PG_FRAME;
4610 m = PHYS_TO_VM_PAGE(pa);
4616 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
4618 * Modified by someone
4620 else if (m->dirty || pmap_is_modified(m))
4621 val |= MINCORE_MODIFIED_OTHER;
4626 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
4629 * Referenced by someone
4631 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
4632 val |= MINCORE_REFERENCED_OTHER;
4633 vm_page_flag_set(m, PG_REFERENCED);
4637 lwkt_reltoken(&pmap->pm_token);
4643 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
4644 * vmspace will be ref'd and the old one will be deref'd.
4646 * The vmspace for all lwps associated with the process will be adjusted
4647 * and cr3 will be reloaded if any lwp is the current lwp.
4649 * The process must hold the vmspace->vm_map.token for oldvm and newvm
4652 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
4654 struct vmspace *oldvm;
4657 oldvm = p->p_vmspace;
4658 if (oldvm != newvm) {
4660 sysref_get(&newvm->vm_sysref);
4661 p->p_vmspace = newvm;
4662 KKASSERT(p->p_nthreads == 1);
4663 lp = RB_ROOT(&p->p_lwp_tree);
4664 pmap_setlwpvm(lp, newvm);
4666 sysref_put(&oldvm->vm_sysref);
4671 * Set the vmspace for a LWP. The vmspace is almost universally set the
4672 * same as the process vmspace, but virtual kernels need to swap out contexts
4673 * on a per-lwp basis.
4675 * Caller does not necessarily hold any vmspace tokens. Caller must control
4676 * the lwp (typically be in the context of the lwp). We use a critical
4677 * section to protect against statclock and hardclock (statistics collection).
4680 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
4682 struct vmspace *oldvm;
4685 oldvm = lp->lwp_vmspace;
4687 if (oldvm != newvm) {
4689 lp->lwp_vmspace = newvm;
4690 if (curthread->td_lwp == lp) {
4691 pmap = vmspace_pmap(newvm);
4692 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4693 if (pmap->pm_active & CPUMASK_LOCK)
4694 pmap_interlock_wait(newvm);
4695 #if defined(SWTCH_OPTIM_STATS)
4698 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4699 load_cr3(curthread->td_pcb->pcb_cr3);
4700 pmap = vmspace_pmap(oldvm);
4701 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4708 * Called when switching to a locked pmap, used to interlock against pmaps
4709 * undergoing modifications to prevent us from activating the MMU for the
4710 * target pmap until all such modifications have completed. We have to do
4711 * this because the thread making the modifications has already set up its
4712 * SMP synchronization mask.
4714 * This function cannot sleep!
4719 pmap_interlock_wait(struct vmspace *vm)
4721 struct pmap *pmap = &vm->vm_pmap;
4723 if (pmap->pm_active & CPUMASK_LOCK) {
4725 KKASSERT(curthread->td_critcount >= 2);
4726 DEBUG_PUSH_INFO("pmap_interlock_wait");
4727 while (pmap->pm_active & CPUMASK_LOCK) {
4729 lwkt_process_ipiq();
4737 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4740 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4744 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4749 * Used by kmalloc/kfree, page already exists at va
4752 pmap_kvtom(vm_offset_t va)
4754 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));
4758 * Initialize machine-specific shared page directory support. This
4759 * is executed when a VM object is created.
4762 pmap_object_init(vm_object_t object)
4764 object->md.pmap_rw = NULL;
4765 object->md.pmap_ro = NULL;
4769 * Clean up machine-specific shared page directory support. This
4770 * is executed when a VM object is destroyed.
4773 pmap_object_free(vm_object_t object)
4777 if ((pmap = object->md.pmap_rw) != NULL) {
4778 object->md.pmap_rw = NULL;
4779 pmap_remove_noinval(pmap,
4780 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
4781 pmap->pm_active = 0;
4784 kfree(pmap, M_OBJPMAP);
4786 if ((pmap = object->md.pmap_ro) != NULL) {
4787 object->md.pmap_ro = NULL;
4788 pmap_remove_noinval(pmap,
4789 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
4790 pmap->pm_active = 0;
4793 kfree(pmap, M_OBJPMAP);