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 */
173 static vm_paddr_t dmaplimit;
175 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
177 static uint64_t KPTbase;
178 static uint64_t KPTphys;
179 static uint64_t KPDphys; /* phys addr of kernel level 2 */
180 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
181 uint64_t KPDPphys; /* phys addr of kernel level 3 */
182 uint64_t KPML4phys; /* phys addr of kernel level 4 */
184 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
185 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
188 * Data for the pv entry allocation mechanism
190 static vm_zone_t pvzone;
191 static struct vm_zone pvzone_store;
192 static struct vm_object pvzone_obj;
193 static int pv_entry_max=0, pv_entry_high_water=0;
194 static int pmap_pagedaemon_waken = 0;
195 static struct pv_entry *pvinit;
198 * All those kernel PT submaps that BSD is so fond of
200 pt_entry_t *CMAP1 = NULL, *ptmmap;
201 caddr_t CADDR1 = 0, ptvmmap = 0;
202 static pt_entry_t *msgbufmap;
203 struct msgbuf *msgbufp=NULL;
208 static pt_entry_t *pt_crashdumpmap;
209 static caddr_t crashdumpmap;
211 static int pmap_yield_count = 64;
212 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
213 &pmap_yield_count, 0, "Yield during init_pt/release");
214 static int pmap_mmu_optimize = 0;
215 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
216 &pmap_mmu_optimize, 0, "Share page table pages when possible");
220 static void pv_hold(pv_entry_t pv);
221 static int _pv_hold_try(pv_entry_t pv
223 static void pv_drop(pv_entry_t pv);
224 static void _pv_lock(pv_entry_t pv
226 static void pv_unlock(pv_entry_t pv);
227 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
229 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
231 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
232 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
233 static void pv_put(pv_entry_t pv);
234 static void pv_free(pv_entry_t pv);
235 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
236 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
238 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
239 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
240 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
241 struct pmap_inval_info *info);
242 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
243 static int pmap_release_pv(pv_entry_t pv);
245 static void pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
246 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
247 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
248 static void pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
249 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
250 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
252 static void i386_protection_init (void);
253 static void create_pagetables(vm_paddr_t *firstaddr);
254 static void pmap_remove_all (vm_page_t m);
255 static boolean_t pmap_testbit (vm_page_t m, int bit);
257 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
258 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
260 static unsigned pdir4mb;
263 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
265 if (pv1->pv_pindex < pv2->pv_pindex)
267 if (pv1->pv_pindex > pv2->pv_pindex)
272 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
273 pv_entry_compare, vm_pindex_t, pv_pindex);
276 * Move the kernel virtual free pointer to the next
277 * 2MB. This is used to help improve performance
278 * by using a large (2MB) page for much of the kernel
279 * (.text, .data, .bss)
283 pmap_kmem_choose(vm_offset_t addr)
285 vm_offset_t newaddr = addr;
287 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
294 * Super fast pmap_pte routine best used when scanning the pv lists.
295 * This eliminates many course-grained invltlb calls. Note that many of
296 * the pv list scans are across different pmaps and it is very wasteful
297 * to do an entire invltlb when checking a single mapping.
299 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
303 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
305 return pmap_pte(pmap, va);
309 * Returns the pindex of a page table entry (representing a terminal page).
310 * There are NUPTE_TOTAL page table entries possible (a huge number)
312 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
313 * We want to properly translate negative KVAs.
317 pmap_pte_pindex(vm_offset_t va)
319 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
323 * Returns the pindex of a page table.
327 pmap_pt_pindex(vm_offset_t va)
329 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
333 * Returns the pindex of a page directory.
337 pmap_pd_pindex(vm_offset_t va)
339 return (NUPTE_TOTAL + NUPT_TOTAL +
340 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
345 pmap_pdp_pindex(vm_offset_t va)
347 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
348 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
353 pmap_pml4_pindex(void)
355 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
359 * Return various clipped indexes for a given VA
361 * Returns the index of a pte in a page table, representing a terminal
366 pmap_pte_index(vm_offset_t va)
368 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
372 * Returns the index of a pt in a page directory, representing a page
377 pmap_pt_index(vm_offset_t va)
379 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
383 * Returns the index of a pd in a page directory page, representing a page
388 pmap_pd_index(vm_offset_t va)
390 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
394 * Returns the index of a pdp in the pml4 table, representing a page
399 pmap_pdp_index(vm_offset_t va)
401 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
405 * Generic procedure to index a pte from a pt, pd, or pdp.
407 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
408 * a page table page index but is instead of PV lookup index.
412 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
416 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
417 return(&pte[pindex]);
421 * Return pointer to PDP slot in the PML4
425 pmap_pdp(pmap_t pmap, vm_offset_t va)
427 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
431 * Return pointer to PD slot in the PDP given a pointer to the PDP
435 pmap_pdp_to_pd(pml4_entry_t *pdp, vm_offset_t va)
439 pd = (pdp_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
440 return (&pd[pmap_pd_index(va)]);
444 * Return pointer to PD slot in the PDP
448 pmap_pd(pmap_t pmap, vm_offset_t va)
452 pdp = pmap_pdp(pmap, va);
453 if ((*pdp & PG_V) == 0)
455 return (pmap_pdp_to_pd(pdp, va));
459 * Return pointer to PT slot in the PD given a pointer to the PD
463 pmap_pd_to_pt(pdp_entry_t *pd, vm_offset_t va)
467 pt = (pd_entry_t *)PHYS_TO_DMAP(*pd & PG_FRAME);
468 return (&pt[pmap_pt_index(va)]);
472 * Return pointer to PT slot in the PD
476 pmap_pt(pmap_t pmap, vm_offset_t va)
480 pd = pmap_pd(pmap, va);
481 if (pd == NULL || (*pd & PG_V) == 0)
483 return (pmap_pd_to_pt(pd, va));
487 * Return pointer to PTE slot in the PT given a pointer to the PT
491 pmap_pt_to_pte(pd_entry_t *pt, vm_offset_t va)
495 pte = (pt_entry_t *)PHYS_TO_DMAP(*pt & PG_FRAME);
496 return (&pte[pmap_pte_index(va)]);
500 * Return pointer to PTE slot in the PT
504 pmap_pte(pmap_t pmap, vm_offset_t va)
508 pt = pmap_pt(pmap, va);
509 if (pt == NULL || (*pt & PG_V) == 0)
511 if ((*pt & PG_PS) != 0)
512 return ((pt_entry_t *)pt);
513 return (pmap_pt_to_pte(pt, va));
517 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
518 * the PT layer. This will speed up core pmap operations considerably.
522 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
524 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
525 pv->pv_pmap->pm_pvhint = pv;
530 * KVM - return address of PT slot in PD
534 vtopt(vm_offset_t va)
536 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
537 NPML4EPGSHIFT)) - 1);
539 return (PDmap + ((va >> PDRSHIFT) & mask));
543 * KVM - return address of PTE slot in PT
547 vtopte(vm_offset_t va)
549 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
550 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
552 return (PTmap + ((va >> PAGE_SHIFT) & mask));
556 allocpages(vm_paddr_t *firstaddr, long n)
561 bzero((void *)ret, n * PAGE_SIZE);
562 *firstaddr += n * PAGE_SIZE;
568 create_pagetables(vm_paddr_t *firstaddr)
570 long i; /* must be 64 bits */
576 * We are running (mostly) V=P at this point
578 * Calculate NKPT - number of kernel page tables. We have to
579 * accomodoate prealloction of the vm_page_array, dump bitmap,
580 * MSGBUF_SIZE, and other stuff. Be generous.
582 * Maxmem is in pages.
584 * ndmpdp is the number of 1GB pages we wish to map.
586 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
587 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
589 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
592 * Starting at the beginning of kvm (not KERNBASE).
594 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
595 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
596 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
597 ndmpdp) + 511) / 512;
601 * Starting at KERNBASE - map 2G worth of page table pages.
602 * KERNBASE is offset -2G from the end of kvm.
604 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
609 KPTbase = allocpages(firstaddr, nkpt_base);
610 KPTphys = allocpages(firstaddr, nkpt_phys);
611 KPML4phys = allocpages(firstaddr, 1);
612 KPDPphys = allocpages(firstaddr, NKPML4E);
613 KPDphys = allocpages(firstaddr, NKPDPE);
616 * Calculate the page directory base for KERNBASE,
617 * that is where we start populating the page table pages.
618 * Basically this is the end - 2.
620 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
622 DMPDPphys = allocpages(firstaddr, NDMPML4E);
623 if ((amd_feature & AMDID_PAGE1GB) == 0)
624 DMPDphys = allocpages(firstaddr, ndmpdp);
625 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
628 * Fill in the underlying page table pages for the area around
629 * KERNBASE. This remaps low physical memory to KERNBASE.
631 * Read-only from zero to physfree
632 * XXX not fully used, underneath 2M pages
634 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
635 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
636 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
640 * Now map the initial kernel page tables. One block of page
641 * tables is placed at the beginning of kernel virtual memory,
642 * and another block is placed at KERNBASE to map the kernel binary,
643 * data, bss, and initial pre-allocations.
645 for (i = 0; i < nkpt_base; i++) {
646 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
647 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
649 for (i = 0; i < nkpt_phys; i++) {
650 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
651 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
655 * Map from zero to end of allocations using 2M pages as an
656 * optimization. This will bypass some of the KPTBase pages
657 * above in the KERNBASE area.
659 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
660 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
661 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
665 * And connect up the PD to the PDP. The kernel pmap is expected
666 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
668 for (i = 0; i < NKPDPE; i++) {
669 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
670 KPDphys + (i << PAGE_SHIFT);
671 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
676 * Now set up the direct map space using either 2MB or 1GB pages
677 * Preset PG_M and PG_A because demotion expects it.
679 * When filling in entries in the PD pages make sure any excess
680 * entries are set to zero as we allocated enough PD pages
682 if ((amd_feature & AMDID_PAGE1GB) == 0) {
683 for (i = 0; i < NPDEPG * ndmpdp; i++) {
684 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
685 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
690 * And the direct map space's PDP
692 for (i = 0; i < ndmpdp; i++) {
693 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
695 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
698 for (i = 0; i < ndmpdp; i++) {
699 ((pdp_entry_t *)DMPDPphys)[i] =
700 (vm_paddr_t)i << PDPSHIFT;
701 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
706 /* And recursively map PML4 to itself in order to get PTmap */
707 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
708 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
711 * Connect the Direct Map slots up to the PML4
713 for (j = 0; j < NDMPML4E; ++j) {
714 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
715 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
720 * Connect the KVA slot up to the PML4
722 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
723 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
727 * Bootstrap the system enough to run with virtual memory.
729 * On the i386 this is called after mapping has already been enabled
730 * and just syncs the pmap module with what has already been done.
731 * [We can't call it easily with mapping off since the kernel is not
732 * mapped with PA == VA, hence we would have to relocate every address
733 * from the linked base (virtual) address "KERNBASE" to the actual
734 * (physical) address starting relative to 0]
737 pmap_bootstrap(vm_paddr_t *firstaddr)
741 struct mdglobaldata *gd;
744 KvaStart = VM_MIN_KERNEL_ADDRESS;
745 KvaEnd = VM_MAX_KERNEL_ADDRESS;
746 KvaSize = KvaEnd - KvaStart;
748 avail_start = *firstaddr;
751 * Create an initial set of page tables to run the kernel in.
753 create_pagetables(firstaddr);
755 virtual2_start = KvaStart;
756 virtual2_end = PTOV_OFFSET;
758 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
759 virtual_start = pmap_kmem_choose(virtual_start);
761 virtual_end = VM_MAX_KERNEL_ADDRESS;
763 /* XXX do %cr0 as well */
764 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
768 * Initialize protection array.
770 i386_protection_init();
773 * The kernel's pmap is statically allocated so we don't have to use
774 * pmap_create, which is unlikely to work correctly at this part of
775 * the boot sequence (XXX and which no longer exists).
777 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
778 kernel_pmap.pm_count = 1;
779 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
780 RB_INIT(&kernel_pmap.pm_pvroot);
781 spin_init(&kernel_pmap.pm_spin);
782 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
785 * Reserve some special page table entries/VA space for temporary
788 #define SYSMAP(c, p, v, n) \
789 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
795 * CMAP1/CMAP2 are used for zeroing and copying pages.
797 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
802 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
805 * ptvmmap is used for reading arbitrary physical pages via
808 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
811 * msgbufp is used to map the system message buffer.
812 * XXX msgbufmap is not used.
814 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
815 atop(round_page(MSGBUF_SIZE)))
822 * PG_G is terribly broken on SMP because we IPI invltlb's in some
823 * cases rather then invl1pg. Actually, I don't even know why it
824 * works under UP because self-referential page table mappings
829 if (cpu_feature & CPUID_PGE)
834 * Initialize the 4MB page size flag
838 * The 4MB page version of the initial
839 * kernel page mapping.
843 #if !defined(DISABLE_PSE)
844 if (cpu_feature & CPUID_PSE) {
847 * Note that we have enabled PSE mode
850 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
851 ptditmp &= ~(NBPDR - 1);
852 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
857 * Enable the PSE mode. If we are SMP we can't do this
858 * now because the APs will not be able to use it when
861 load_cr4(rcr4() | CR4_PSE);
864 * We can do the mapping here for the single processor
865 * case. We simply ignore the old page table page from
869 * For SMP, we still need 4K pages to bootstrap APs,
870 * PSE will be enabled as soon as all APs are up.
872 PTD[KPTDI] = (pd_entry_t)ptditmp;
879 * We need to finish setting up the globaldata page for the BSP.
880 * locore has already populated the page table for the mdglobaldata
883 pg = MDGLOBALDATA_BASEALLOC_PAGES;
884 gd = &CPU_prvspace[0].mdglobaldata;
891 * Set 4mb pdir for mp startup
896 if (pseflag && (cpu_feature & CPUID_PSE)) {
897 load_cr4(rcr4() | CR4_PSE);
898 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
906 * Initialize the pmap module.
907 * Called by vm_init, to initialize any structures that the pmap
908 * system needs to map virtual memory.
909 * pmap_init has been enhanced to support in a fairly consistant
910 * way, discontiguous physical memory.
919 * Allocate memory for random pmap data structures. Includes the
923 for (i = 0; i < vm_page_array_size; i++) {
926 m = &vm_page_array[i];
927 TAILQ_INIT(&m->md.pv_list);
931 * init the pv free list
933 initial_pvs = vm_page_array_size;
934 if (initial_pvs < MINPV)
936 pvzone = &pvzone_store;
937 pvinit = (void *)kmem_alloc(&kernel_map,
938 initial_pvs * sizeof (struct pv_entry));
939 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
940 pvinit, initial_pvs);
943 * Now it is safe to enable pv_table recording.
945 pmap_initialized = TRUE;
949 * Initialize the address space (zone) for the pv_entries. Set a
950 * high water mark so that the system can recover from excessive
951 * numbers of pv entries.
956 int shpgperproc = PMAP_SHPGPERPROC;
959 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
960 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
961 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
962 pv_entry_high_water = 9 * (pv_entry_max / 10);
965 * Subtract out pages already installed in the zone (hack)
967 entry_max = pv_entry_max - vm_page_array_size;
971 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
975 /***************************************************
976 * Low level helper routines.....
977 ***************************************************/
980 * this routine defines the region(s) of memory that should
981 * not be tested for the modified bit.
985 pmap_track_modified(vm_pindex_t pindex)
987 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
988 if ((va < clean_sva) || (va >= clean_eva))
995 * Extract the physical page address associated with the map/VA pair.
996 * The page must be wired for this to work reliably.
998 * XXX for the moment we're using pv_find() instead of pv_get(), as
999 * callers might be expecting non-blocking operation.
1002 pmap_extract(pmap_t pmap, vm_offset_t va)
1009 if (va >= VM_MAX_USER_ADDRESS) {
1011 * Kernel page directories might be direct-mapped and
1012 * there is typically no PV tracking of pte's
1016 pt = pmap_pt(pmap, va);
1017 if (pt && (*pt & PG_V)) {
1019 rtval = *pt & PG_PS_FRAME;
1020 rtval |= va & PDRMASK;
1022 ptep = pmap_pt_to_pte(pt, va);
1024 rtval = *ptep & PG_FRAME;
1025 rtval |= va & PAGE_MASK;
1031 * User pages currently do not direct-map the page directory
1032 * and some pages might not used managed PVs. But all PT's
1035 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1037 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1039 rtval = *ptep & PG_FRAME;
1040 rtval |= va & PAGE_MASK;
1049 * Extract the physical page address associated kernel virtual address.
1052 pmap_kextract(vm_offset_t va)
1054 pd_entry_t pt; /* pt entry in pd */
1057 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1058 pa = DMAP_TO_PHYS(va);
1062 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1065 * Beware of a concurrent promotion that changes the
1066 * PDE at this point! For example, vtopte() must not
1067 * be used to access the PTE because it would use the
1068 * new PDE. It is, however, safe to use the old PDE
1069 * because the page table page is preserved by the
1072 pa = *pmap_pt_to_pte(&pt, va);
1073 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1079 /***************************************************
1080 * Low level mapping routines.....
1081 ***************************************************/
1084 * Routine: pmap_kenter
1086 * Add a wired page to the KVA
1087 * NOTE! note that in order for the mapping to take effect -- you
1088 * should do an invltlb after doing the pmap_kenter().
1091 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1095 pmap_inval_info info;
1097 pmap_inval_init(&info); /* XXX remove */
1098 npte = pa | PG_RW | PG_V | pgeflag;
1100 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1102 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1103 pmap_inval_done(&info); /* XXX remove */
1107 * Routine: pmap_kenter_quick
1109 * Similar to pmap_kenter(), except we only invalidate the
1110 * mapping on the current CPU.
1113 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1118 npte = pa | PG_RW | PG_V | pgeflag;
1121 cpu_invlpg((void *)va);
1125 pmap_kenter_sync(vm_offset_t va)
1127 pmap_inval_info info;
1129 pmap_inval_init(&info);
1130 pmap_inval_interlock(&info, &kernel_pmap, va);
1131 pmap_inval_deinterlock(&info, &kernel_pmap);
1132 pmap_inval_done(&info);
1136 pmap_kenter_sync_quick(vm_offset_t va)
1138 cpu_invlpg((void *)va);
1142 * remove a page from the kernel pagetables
1145 pmap_kremove(vm_offset_t va)
1148 pmap_inval_info info;
1150 pmap_inval_init(&info);
1152 pmap_inval_interlock(&info, &kernel_pmap, va);
1153 (void)pte_load_clear(pte);
1154 pmap_inval_deinterlock(&info, &kernel_pmap);
1155 pmap_inval_done(&info);
1159 pmap_kremove_quick(vm_offset_t va)
1163 (void)pte_load_clear(pte);
1164 cpu_invlpg((void *)va);
1168 * XXX these need to be recoded. They are not used in any critical path.
1171 pmap_kmodify_rw(vm_offset_t va)
1173 atomic_set_long(vtopte(va), PG_RW);
1174 cpu_invlpg((void *)va);
1178 pmap_kmodify_nc(vm_offset_t va)
1180 atomic_set_long(vtopte(va), PG_N);
1181 cpu_invlpg((void *)va);
1185 * Used to map a range of physical addresses into kernel virtual
1186 * address space during the low level boot, typically to map the
1187 * dump bitmap, message buffer, and vm_page_array.
1189 * These mappings are typically made at some pointer after the end of the
1192 * We could return PHYS_TO_DMAP(start) here and not allocate any
1193 * via (*virtp), but then kmem from userland and kernel dumps won't
1194 * have access to the related pointers.
1197 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1200 vm_offset_t va_start;
1202 /*return PHYS_TO_DMAP(start);*/
1207 while (start < end) {
1208 pmap_kenter_quick(va, start);
1218 * Add a list of wired pages to the kva
1219 * this routine is only used for temporary
1220 * kernel mappings that do not need to have
1221 * page modification or references recorded.
1222 * Note that old mappings are simply written
1223 * over. The page *must* be wired.
1226 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1230 end_va = va + count * PAGE_SIZE;
1232 while (va < end_va) {
1236 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1237 cpu_invlpg((void *)va);
1245 * This routine jerks page mappings from the
1246 * kernel -- it is meant only for temporary mappings.
1248 * MPSAFE, INTERRUPT SAFE (cluster callback)
1251 pmap_qremove(vm_offset_t va, int count)
1255 end_va = va + count * PAGE_SIZE;
1257 while (va < end_va) {
1261 (void)pte_load_clear(pte);
1262 cpu_invlpg((void *)va);
1269 * Create a new thread and optionally associate it with a (new) process.
1270 * NOTE! the new thread's cpu may not equal the current cpu.
1273 pmap_init_thread(thread_t td)
1275 /* enforce pcb placement & alignment */
1276 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1277 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1278 td->td_savefpu = &td->td_pcb->pcb_save;
1279 td->td_sp = (char *)td->td_pcb; /* no -16 */
1283 * This routine directly affects the fork perf for a process.
1286 pmap_init_proc(struct proc *p)
1291 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1292 * it, and IdlePTD, represents the template used to update all other pmaps.
1294 * On architectures where the kernel pmap is not integrated into the user
1295 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1296 * kernel_pmap should be used to directly access the kernel_pmap.
1299 pmap_pinit0(struct pmap *pmap)
1301 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1303 pmap->pm_active = 0;
1304 pmap->pm_pvhint = NULL;
1305 RB_INIT(&pmap->pm_pvroot);
1306 spin_init(&pmap->pm_spin);
1307 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1308 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1312 * Initialize a preallocated and zeroed pmap structure,
1313 * such as one in a vmspace structure.
1316 pmap_pinit_simple(struct pmap *pmap)
1319 * Misc initialization
1322 pmap->pm_active = 0;
1323 pmap->pm_pvhint = NULL;
1324 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1327 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1330 if (pmap->pm_pmlpv == NULL) {
1331 RB_INIT(&pmap->pm_pvroot);
1332 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1333 spin_init(&pmap->pm_spin);
1334 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1339 pmap_pinit(struct pmap *pmap)
1344 pmap_pinit_simple(pmap);
1345 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1348 * No need to allocate page table space yet but we do need a valid
1349 * page directory table.
1351 if (pmap->pm_pml4 == NULL) {
1353 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1357 * Allocate the page directory page, which wires it even though
1358 * it isn't being entered into some higher level page table (it
1359 * being the highest level). If one is already cached we don't
1360 * have to do anything.
1362 if ((pv = pmap->pm_pmlpv) == NULL) {
1363 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1364 pmap->pm_pmlpv = pv;
1365 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1366 VM_PAGE_TO_PHYS(pv->pv_m));
1370 * Install DMAP and KMAP.
1372 for (j = 0; j < NDMPML4E; ++j) {
1373 pmap->pm_pml4[DMPML4I + j] =
1374 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1375 PG_RW | PG_V | PG_U;
1377 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1380 * install self-referential address mapping entry
1382 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1383 PG_V | PG_RW | PG_A | PG_M;
1385 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1386 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1388 KKASSERT(pmap->pm_pml4[255] == 0);
1389 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1390 KKASSERT(pv->pv_entry.rbe_left == NULL);
1391 KKASSERT(pv->pv_entry.rbe_right == NULL);
1395 * Clean up a pmap structure so it can be physically freed. This routine
1396 * is called by the vmspace dtor function. A great deal of pmap data is
1397 * left passively mapped to improve vmspace management so we have a bit
1398 * of cleanup work to do here.
1401 pmap_puninit(pmap_t pmap)
1406 KKASSERT(pmap->pm_active == 0);
1407 if ((pv = pmap->pm_pmlpv) != NULL) {
1408 if (pv_hold_try(pv) == 0)
1410 p = pmap_remove_pv_page(pv);
1412 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1413 vm_page_busy_wait(p, FALSE, "pgpun");
1414 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1415 vm_page_unwire(p, 0);
1416 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1419 * XXX eventually clean out PML4 static entries and
1420 * use vm_page_free_zero()
1423 pmap->pm_pmlpv = NULL;
1425 if (pmap->pm_pml4) {
1426 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1427 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1428 pmap->pm_pml4 = NULL;
1430 KKASSERT(pmap->pm_stats.resident_count == 0);
1431 KKASSERT(pmap->pm_stats.wired_count == 0);
1435 * Wire in kernel global address entries. To avoid a race condition
1436 * between pmap initialization and pmap_growkernel, this procedure
1437 * adds the pmap to the master list (which growkernel scans to update),
1438 * then copies the template.
1441 pmap_pinit2(struct pmap *pmap)
1443 spin_lock(&pmap_spin);
1444 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1445 spin_unlock(&pmap_spin);
1449 * This routine is called when various levels in the page table need to
1450 * be populated. This routine cannot fail.
1452 * This function returns two locked pv_entry's, one representing the
1453 * requested pv and one representing the requested pv's parent pv. If
1454 * the pv did not previously exist it will be mapped into its parent
1455 * and wired, otherwise no additional wire count will be added.
1459 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1464 vm_pindex_t pt_pindex;
1470 * If the pv already exists and we aren't being asked for the
1471 * parent page table page we can just return it. A locked+held pv
1475 pv = pv_alloc(pmap, ptepindex, &isnew);
1476 if (isnew == 0 && pvpp == NULL)
1480 * This is a new PV, we have to resolve its parent page table and
1481 * add an additional wiring to the page if necessary.
1485 * Special case terminal PVs. These are not page table pages so
1486 * no vm_page is allocated (the caller supplied the vm_page). If
1487 * pvpp is non-NULL we are being asked to also removed the pt_pv
1490 * Note that pt_pv's are only returned for user VAs. We assert that
1491 * a pt_pv is not being requested for kernel VAs.
1493 if (ptepindex < pmap_pt_pindex(0)) {
1494 if (ptepindex >= NUPTE_USER)
1495 KKASSERT(pvpp == NULL);
1497 KKASSERT(pvpp != NULL);
1499 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1500 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1502 vm_page_wire_quick(pvp->pv_m);
1511 * Non-terminal PVs allocate a VM page to represent the page table,
1512 * so we have to resolve pvp and calculate ptepindex for the pvp
1513 * and then for the page table entry index in the pvp for
1516 if (ptepindex < pmap_pd_pindex(0)) {
1518 * pv is PT, pvp is PD
1520 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1521 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1522 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1529 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1530 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1532 } else if (ptepindex < pmap_pdp_pindex(0)) {
1534 * pv is PD, pvp is PDP
1536 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1539 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1540 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1542 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1543 KKASSERT(pvpp == NULL);
1546 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1554 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1555 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1556 } else if (ptepindex < pmap_pml4_pindex()) {
1558 * pv is PDP, pvp is the root pml4 table
1560 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1567 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1568 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1571 * pv represents the top-level PML4, there is no parent.
1579 * This code is only reached if isnew is TRUE and this is not a
1580 * terminal PV. We need to allocate a vm_page for the page table
1581 * at this level and enter it into the parent page table.
1583 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1586 m = vm_page_alloc(NULL, pv->pv_pindex,
1587 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1588 VM_ALLOC_INTERRUPT);
1593 vm_page_spin_lock(m);
1594 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1596 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1597 vm_page_spin_unlock(m);
1598 vm_page_unmanage(m); /* m must be spinunlocked */
1600 if ((m->flags & PG_ZERO) == 0) {
1601 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1605 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1608 m->valid = VM_PAGE_BITS_ALL;
1609 vm_page_flag_clear(m, PG_ZERO);
1610 vm_page_wire(m); /* wire for mapping in parent */
1613 * Wire the page into pvp, bump the wire-count for pvp's page table
1614 * page. Bump the resident_count for the pmap. There is no pvp
1615 * for the top level, address the pm_pml4[] array directly.
1617 * If the caller wants the parent we return it, otherwise
1618 * we just put it away.
1620 * No interlock is needed for pte 0 -> non-zero.
1622 * In the situation where *ptep is valid we might have an unmanaged
1623 * page table page shared from another page table which we need to
1624 * unshare before installing our private page table page.
1627 ptep = pv_pte_lookup(pvp, ptepindex);
1630 pmap_inval_info info;
1632 kprintf("pmap_allocpte: restate shared pg table pg\n");
1635 panic("pmap_allocpte: unexpected pte %p/%d",
1636 pvp, (int)ptepindex);
1638 pmap_inval_init(&info);
1639 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1640 pte = pte_load_clear(ptep);
1641 pmap_inval_deinterlock(&info, pmap);
1642 pmap_inval_done(&info);
1643 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
1644 panic("pmap_allocpte: shared pgtable pg bad wirecount");
1646 vm_page_wire_quick(pvp->pv_m);
1648 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1661 * This version of pmap_allocpte() checks for possible segment optimizations
1662 * that would allow page-table sharing. It can be called for terminal
1663 * page or page table page ptepindex's.
1665 * The function is called with page table page ptepindex's for fictitious
1666 * and unmanaged terminal pages. That is, we don't want to allocate a
1667 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
1670 * This function can return a pv and *pvpp associated with the passed in pmap
1671 * OR a pv and *pvpp associated with the shared pmap. In the latter case
1672 * an unmanaged page table page will be entered into the pass in pmap.
1676 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
1677 vm_map_entry_t entry, vm_offset_t va)
1679 struct pmap_inval_info info;
1684 pv_entry_t pte_pv; /* in original or shared pmap */
1685 pv_entry_t pt_pv; /* in original or shared pmap */
1686 pv_entry_t proc_pd_pv; /* in original pmap */
1687 pv_entry_t proc_pt_pv; /* in original pmap */
1688 pv_entry_t xpv; /* PT in shared pmap */
1689 pd_entry_t *pt; /* PT entry in PD of original pmap */
1690 pd_entry_t opte; /* contents of *pt */
1691 pd_entry_t npte; /* contents of *pt */
1695 * Basic tests, require a non-NULL vm_map_entry, require proper
1696 * alignment and type for the vm_map_entry, require that the
1697 * underlying object already be allocated.
1699 * We currently allow any type of object to use this optimization.
1700 * The object itself does NOT have to be sized to a multiple of the
1701 * segment size, but the memory mapping does.
1703 if (entry == NULL ||
1704 pmap_mmu_optimize == 0 || /* not enabled */
1705 ptepindex >= pmap_pd_pindex(0) || /* not terminal */
1706 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
1707 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
1708 entry->object.vm_object == NULL || /* needs VM object */
1709 (entry->offset & SEG_MASK) || /* must be aligned */
1710 (entry->start & SEG_MASK)) {
1711 return(pmap_allocpte(pmap, ptepindex, pvpp));
1715 * Make sure the full segment can be represented.
1717 b = va & ~(vm_offset_t)SEG_MASK;
1718 if (b < entry->start && b + SEG_SIZE > entry->end)
1719 return(pmap_allocpte(pmap, ptepindex, pvpp));
1722 * If the full segment can be represented dive the VM object's
1723 * shared pmap, allocating as required.
1725 object = entry->object.vm_object;
1727 if (entry->protection & VM_PROT_WRITE)
1728 obpmapp = &object->md.pmap_rw;
1730 obpmapp = &object->md.pmap_ro;
1733 * We allocate what appears to be a normal pmap but because portions
1734 * of this pmap are shared with other unrelated pmaps we have to
1735 * set pm_active to point to all cpus.
1737 * XXX Currently using pmap_spin to interlock the update, can't use
1738 * vm_object_hold/drop because the token might already be held
1739 * shared OR exclusive and we don't know.
1741 while ((obpmap = *obpmapp) == NULL) {
1742 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
1743 pmap_pinit_simple(obpmap);
1744 pmap_pinit2(obpmap);
1745 spin_lock(&pmap_spin);
1746 if (*obpmapp != NULL) {
1750 spin_unlock(&pmap_spin);
1751 pmap_release(obpmap);
1752 pmap_puninit(obpmap);
1753 kfree(obpmap, M_OBJPMAP);
1755 obpmap->pm_active = smp_active_mask;
1757 spin_unlock(&pmap_spin);
1762 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
1763 * pte/pt using the shared pmap from the object but also adjust
1764 * the process pmap's page table page as a side effect.
1768 * Resolve the terminal PTE and PT in the shared pmap. This is what
1769 * we will return. This is true if ptepindex represents a terminal
1770 * page, otherwise pte_pv is actually the PT and pt_pv is actually
1774 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
1775 if (ptepindex >= pmap_pt_pindex(0))
1781 * Resolve the PD in the process pmap so we can properly share the
1782 * page table page. Lock order is bottom-up (leaf first)!
1784 * NOTE: proc_pt_pv can be NULL.
1786 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
1787 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
1790 * xpv is the page table page pv from the shared object
1791 * (for convenience).
1793 * Calculate the pte value for the PT to load into the process PD.
1794 * If we have to change it we must properly dispose of the previous
1797 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1798 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
1799 (PG_U | PG_RW | PG_V | PG_A | PG_M);
1802 * Dispose of previous entry if it was local to the process pmap.
1803 * (This should zero-out *pt)
1806 pmap_release_pv(proc_pt_pv);
1809 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1813 * Handle remaining cases.
1817 vm_page_wire_quick(xpv->pv_m);
1818 vm_page_wire_quick(proc_pd_pv->pv_m);
1819 atomic_add_long(&pmap->pm_stats.resident_count, 1);
1820 } else if (*pt != npte) {
1821 pmap_inval_init(&info);
1822 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1824 opte = pte_load_clear(pt);
1825 KKASSERT(opte && opte != npte);
1828 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
1831 * Clean up opte, bump the wire_count for the process
1832 * PD page representing the new entry if it was
1835 * If the entry was not previously empty and we have
1836 * a PT in the proc pmap then opte must match that
1837 * pt. The proc pt must be retired (this is done
1838 * later on in this procedure).
1840 * NOTE: replacing valid pte, wire_count on proc_pd_pv
1843 KKASSERT(opte & PG_V);
1844 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
1845 if (vm_page_unwire_quick(m)) {
1846 panic("pmap_allocpte_seg: "
1847 "bad wire count %p",
1851 pmap_inval_deinterlock(&info, pmap);
1852 pmap_inval_done(&info);
1856 * The existing process page table was replaced and must be destroyed
1870 * Release any resources held by the given physical map.
1872 * Called when a pmap initialized by pmap_pinit is being released. Should
1873 * only be called if the map contains no valid mappings.
1875 * Caller must hold pmap->pm_token
1877 struct pmap_release_info {
1882 static int pmap_release_callback(pv_entry_t pv, void *data);
1885 pmap_release(struct pmap *pmap)
1887 struct pmap_release_info info;
1889 KASSERT(pmap->pm_active == 0,
1890 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1892 spin_lock(&pmap_spin);
1893 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1894 spin_unlock(&pmap_spin);
1897 * Pull pv's off the RB tree in order from low to high and release
1903 spin_lock(&pmap->pm_spin);
1904 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
1905 pmap_release_callback, &info);
1906 spin_unlock(&pmap->pm_spin);
1907 } while (info.retry);
1911 * One resident page (the pml4 page) should remain.
1912 * No wired pages should remain.
1914 KKASSERT(pmap->pm_stats.resident_count ==
1915 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
1917 KKASSERT(pmap->pm_stats.wired_count == 0);
1921 pmap_release_callback(pv_entry_t pv, void *data)
1923 struct pmap_release_info *info = data;
1924 pmap_t pmap = info->pmap;
1927 if (pv_hold_try(pv)) {
1928 spin_unlock(&pmap->pm_spin);
1930 spin_unlock(&pmap->pm_spin);
1932 if (pv->pv_pmap != pmap) {
1934 spin_lock(&pmap->pm_spin);
1939 r = pmap_release_pv(pv);
1940 spin_lock(&pmap->pm_spin);
1945 * Called with held (i.e. also locked) pv. This function will dispose of
1946 * the lock along with the pv.
1949 pmap_release_pv(pv_entry_t pv)
1954 * The pmap is currently not spinlocked, pv is held+locked.
1955 * Remove the pv's page from its parent's page table. The
1956 * parent's page table page's wire_count will be decremented.
1958 pmap_remove_pv_pte(pv, NULL, NULL);
1961 * Terminal pvs are unhooked from their vm_pages. Because
1962 * terminal pages aren't page table pages they aren't wired
1963 * by us, so we have to be sure not to unwire them either.
1965 if (pv->pv_pindex < pmap_pt_pindex(0)) {
1966 pmap_remove_pv_page(pv);
1971 * We leave the top-level page table page cached, wired, and
1972 * mapped in the pmap until the dtor function (pmap_puninit())
1975 * Since we are leaving the top-level pv intact we need
1976 * to break out of what would otherwise be an infinite loop.
1978 if (pv->pv_pindex == pmap_pml4_pindex()) {
1984 * For page table pages (other than the top-level page),
1985 * remove and free the vm_page. The representitive mapping
1986 * removed above by pmap_remove_pv_pte() did not undo the
1987 * last wire_count so we have to do that as well.
1989 p = pmap_remove_pv_page(pv);
1990 vm_page_busy_wait(p, FALSE, "pmaprl");
1991 if (p->wire_count != 1) {
1992 kprintf("p->wire_count was %016lx %d\n",
1993 pv->pv_pindex, p->wire_count);
1995 KKASSERT(p->wire_count == 1);
1996 KKASSERT(p->flags & PG_UNMANAGED);
1998 vm_page_unwire(p, 0);
1999 KKASSERT(p->wire_count == 0);
2002 * Theoretically this page, if not the pml4 page, should contain
2003 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2013 * This function will remove the pte associated with a pv from its parent.
2014 * Terminal pv's are supported. The removal will be interlocked if info
2015 * is non-NULL. The caller must dispose of pv instead of just unlocking
2018 * The wire count will be dropped on the parent page table. The wire
2019 * count on the page being removed (pv->pv_m) from the parent page table
2020 * is NOT touched. Note that terminal pages will not have any additional
2021 * wire counts while page table pages will have at least one representing
2022 * the mapping, plus others representing sub-mappings.
2024 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2025 * pages and user page table and terminal pages.
2027 * The pv must be locked.
2029 * XXX must lock parent pv's if they exist to remove pte XXX
2033 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2035 vm_pindex_t ptepindex = pv->pv_pindex;
2036 pmap_t pmap = pv->pv_pmap;
2042 if (ptepindex == pmap_pml4_pindex()) {
2044 * We are the top level pml4 table, there is no parent.
2046 p = pmap->pm_pmlpv->pv_m;
2047 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2049 * Remove a PDP page from the pml4e. This can only occur
2050 * with user page tables. We do not have to lock the
2051 * pml4 PV so just ignore pvp.
2053 vm_pindex_t pml4_pindex;
2054 vm_pindex_t pdp_index;
2057 pdp_index = ptepindex - pmap_pdp_pindex(0);
2059 pml4_pindex = pmap_pml4_pindex();
2060 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2064 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2065 KKASSERT((*pdp & PG_V) != 0);
2066 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2068 KKASSERT(info == NULL);
2069 } else if (ptepindex >= pmap_pd_pindex(0)) {
2071 * Remove a PD page from the pdp
2073 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2074 * of a simple pmap because it stops at
2077 vm_pindex_t pdp_pindex;
2078 vm_pindex_t pd_index;
2081 pd_index = ptepindex - pmap_pd_pindex(0);
2084 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2085 (pd_index >> NPML4EPGSHIFT);
2086 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2091 pd = pv_pte_lookup(pvp, pd_index &
2092 ((1ul << NPDPEPGSHIFT) - 1));
2093 KKASSERT((*pd & PG_V) != 0);
2094 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2097 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2098 p = pv->pv_m; /* degenerate test later */
2100 KKASSERT(info == NULL);
2101 } else if (ptepindex >= pmap_pt_pindex(0)) {
2103 * Remove a PT page from the pd
2105 vm_pindex_t pd_pindex;
2106 vm_pindex_t pt_index;
2109 pt_index = ptepindex - pmap_pt_pindex(0);
2112 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2113 (pt_index >> NPDPEPGSHIFT);
2114 pvp = pv_get(pv->pv_pmap, pd_pindex);
2118 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2119 KKASSERT((*pt & PG_V) != 0);
2120 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2122 KKASSERT(info == NULL);
2125 * Remove a PTE from the PT page
2127 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2128 * pv is a pte_pv so we can safely lock pt_pv.
2130 vm_pindex_t pt_pindex;
2135 pt_pindex = ptepindex >> NPTEPGSHIFT;
2136 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2138 if (ptepindex >= NUPTE_USER) {
2139 ptep = vtopte(ptepindex << PAGE_SHIFT);
2140 KKASSERT(pvp == NULL);
2143 pt_pindex = NUPTE_TOTAL +
2144 (ptepindex >> NPDPEPGSHIFT);
2145 pvp = pv_get(pv->pv_pmap, pt_pindex);
2149 ptep = pv_pte_lookup(pvp, ptepindex &
2150 ((1ul << NPDPEPGSHIFT) - 1));
2154 pmap_inval_interlock(info, pmap, va);
2155 pte = pte_load_clear(ptep);
2157 pmap_inval_deinterlock(info, pmap);
2159 cpu_invlpg((void *)va);
2162 * Now update the vm_page_t
2164 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
2165 kprintf("remove_pte badpte %016lx %016lx %d\n",
2167 pv->pv_pindex < pmap_pt_pindex(0));
2169 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2170 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2173 if (pmap_track_modified(ptepindex))
2177 vm_page_flag_set(p, PG_REFERENCED);
2180 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2182 cpu_invlpg((void *)va);
2186 * Unwire the parent page table page. The wire_count cannot go below
2187 * 1 here because the parent page table page is itself still mapped.
2189 * XXX remove the assertions later.
2191 KKASSERT(pv->pv_m == p);
2192 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2193 panic("pmap_remove_pv_pte: Insufficient wire_count");
2201 pmap_remove_pv_page(pv_entry_t pv)
2207 vm_page_spin_lock(m);
2209 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2212 atomic_add_int(&m->object->agg_pv_list_count, -1);
2214 if (TAILQ_EMPTY(&m->md.pv_list))
2215 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2216 vm_page_spin_unlock(m);
2221 * Grow the number of kernel page table entries, if needed.
2223 * This routine is always called to validate any address space
2224 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2225 * space below KERNBASE.
2228 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2231 vm_offset_t ptppaddr;
2233 pd_entry_t *pt, newpt;
2235 int update_kernel_vm_end;
2238 * bootstrap kernel_vm_end on first real VM use
2240 if (kernel_vm_end == 0) {
2241 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2243 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
2244 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2245 ~(PAGE_SIZE * NPTEPG - 1);
2247 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2248 kernel_vm_end = kernel_map.max_offset;
2255 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2256 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2257 * do not want to force-fill 128G worth of page tables.
2259 if (kstart < KERNBASE) {
2260 if (kstart > kernel_vm_end)
2261 kstart = kernel_vm_end;
2262 KKASSERT(kend <= KERNBASE);
2263 update_kernel_vm_end = 1;
2265 update_kernel_vm_end = 0;
2268 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2269 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2271 if (kend - 1 >= kernel_map.max_offset)
2272 kend = kernel_map.max_offset;
2274 while (kstart < kend) {
2275 pt = pmap_pt(&kernel_pmap, kstart);
2277 /* We need a new PDP entry */
2278 nkpg = vm_page_alloc(NULL, nkpt,
2281 VM_ALLOC_INTERRUPT);
2283 panic("pmap_growkernel: no memory to grow "
2286 paddr = VM_PAGE_TO_PHYS(nkpg);
2287 if ((nkpg->flags & PG_ZERO) == 0)
2288 pmap_zero_page(paddr);
2289 vm_page_flag_clear(nkpg, PG_ZERO);
2290 newpd = (pdp_entry_t)
2291 (paddr | PG_V | PG_RW | PG_A | PG_M);
2292 *pmap_pd(&kernel_pmap, kstart) = newpd;
2294 continue; /* try again */
2296 if ((*pt & PG_V) != 0) {
2297 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2298 ~(PAGE_SIZE * NPTEPG - 1);
2299 if (kstart - 1 >= kernel_map.max_offset) {
2300 kstart = kernel_map.max_offset;
2307 * This index is bogus, but out of the way
2309 nkpg = vm_page_alloc(NULL, nkpt,
2312 VM_ALLOC_INTERRUPT);
2314 panic("pmap_growkernel: no memory to grow kernel");
2317 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2318 pmap_zero_page(ptppaddr);
2319 vm_page_flag_clear(nkpg, PG_ZERO);
2320 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2321 *pmap_pt(&kernel_pmap, kstart) = newpt;
2324 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2325 ~(PAGE_SIZE * NPTEPG - 1);
2327 if (kstart - 1 >= kernel_map.max_offset) {
2328 kstart = kernel_map.max_offset;
2334 * Only update kernel_vm_end for areas below KERNBASE.
2336 if (update_kernel_vm_end && kernel_vm_end < kstart)
2337 kernel_vm_end = kstart;
2341 * Add a reference to the specified pmap.
2344 pmap_reference(pmap_t pmap)
2347 lwkt_gettoken(&pmap->pm_token);
2349 lwkt_reltoken(&pmap->pm_token);
2354 pmap_drop(pmap_t pmap)
2357 lwkt_gettoken(&pmap->pm_token);
2359 lwkt_reltoken(&pmap->pm_token);
2363 /***************************************************
2364 * page management routines.
2365 ***************************************************/
2368 * Hold a pv without locking it
2371 pv_hold(pv_entry_t pv)
2375 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2379 count = pv->pv_hold;
2381 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2388 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2389 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2392 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2393 * pv list via its page) must be held by the caller.
2396 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2400 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2403 pv->pv_line = lineno;
2409 count = pv->pv_hold;
2411 if ((count & PV_HOLD_LOCKED) == 0) {
2412 if (atomic_cmpset_int(&pv->pv_hold, count,
2413 (count + 1) | PV_HOLD_LOCKED)) {
2416 pv->pv_line = lineno;
2421 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2429 * Drop a previously held pv_entry which could not be locked, allowing its
2432 * Must not be called with a spinlock held as we might zfree() the pv if it
2433 * is no longer associated with a pmap and this was the last hold count.
2436 pv_drop(pv_entry_t pv)
2440 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2441 if (pv->pv_pmap == NULL)
2447 count = pv->pv_hold;
2449 KKASSERT((count & PV_HOLD_MASK) > 0);
2450 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2451 (PV_HOLD_LOCKED | 1));
2452 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2453 if (count == 1 && pv->pv_pmap == NULL)
2462 * Find or allocate the requested PV entry, returning a locked pv
2466 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2469 pv_entry_t pnew = NULL;
2471 spin_lock(&pmap->pm_spin);
2473 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2474 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2479 spin_unlock(&pmap->pm_spin);
2480 pnew = zalloc(pvzone);
2481 spin_lock(&pmap->pm_spin);
2484 pnew->pv_pmap = pmap;
2485 pnew->pv_pindex = pindex;
2486 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2488 pnew->pv_func = func;
2489 pnew->pv_line = lineno;
2491 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2492 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2493 spin_unlock(&pmap->pm_spin);
2498 spin_unlock(&pmap->pm_spin);
2499 zfree(pvzone, pnew);
2501 spin_lock(&pmap->pm_spin);
2504 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2505 spin_unlock(&pmap->pm_spin);
2509 spin_unlock(&pmap->pm_spin);
2510 _pv_lock(pv PMAP_DEBUG_COPY);
2511 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2516 spin_lock(&pmap->pm_spin);
2523 * Find the requested PV entry, returning a locked+held pv or NULL
2527 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2531 spin_lock(&pmap->pm_spin);
2536 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2537 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2541 spin_unlock(&pmap->pm_spin);
2544 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2545 pv_cache(pv, pindex);
2546 spin_unlock(&pmap->pm_spin);
2549 spin_unlock(&pmap->pm_spin);
2550 _pv_lock(pv PMAP_DEBUG_COPY);
2551 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2554 spin_lock(&pmap->pm_spin);
2559 * Lookup, hold, and attempt to lock (pmap,pindex).
2561 * If the entry does not exist NULL is returned and *errorp is set to 0
2563 * If the entry exists and could be successfully locked it is returned and
2564 * errorp is set to 0.
2566 * If the entry exists but could NOT be successfully locked it is returned
2567 * held and *errorp is set to 1.
2571 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2575 spin_lock(&pmap->pm_spin);
2576 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2577 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2579 spin_unlock(&pmap->pm_spin);
2583 if (pv_hold_try(pv)) {
2584 pv_cache(pv, pindex);
2585 spin_unlock(&pmap->pm_spin);
2587 return(pv); /* lock succeeded */
2589 spin_unlock(&pmap->pm_spin);
2591 return (pv); /* lock failed */
2595 * Find the requested PV entry, returning a held pv or NULL
2599 pv_find(pmap_t pmap, vm_pindex_t pindex)
2603 spin_lock(&pmap->pm_spin);
2605 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2606 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2608 spin_unlock(&pmap->pm_spin);
2612 pv_cache(pv, pindex);
2613 spin_unlock(&pmap->pm_spin);
2618 * Lock a held pv, keeping the hold count
2622 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2627 count = pv->pv_hold;
2629 if ((count & PV_HOLD_LOCKED) == 0) {
2630 if (atomic_cmpset_int(&pv->pv_hold, count,
2631 count | PV_HOLD_LOCKED)) {
2634 pv->pv_line = lineno;
2640 tsleep_interlock(pv, 0);
2641 if (atomic_cmpset_int(&pv->pv_hold, count,
2642 count | PV_HOLD_WAITING)) {
2644 kprintf("pv waiting on %s:%d\n",
2645 pv->pv_func, pv->pv_line);
2647 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2654 * Unlock a held and locked pv, keeping the hold count.
2658 pv_unlock(pv_entry_t pv)
2662 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2666 count = pv->pv_hold;
2668 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2669 (PV_HOLD_LOCKED | 1));
2670 if (atomic_cmpset_int(&pv->pv_hold, count,
2672 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2673 if (count & PV_HOLD_WAITING)
2681 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2682 * and the hold count drops to zero we will free it.
2684 * Caller should not hold any spin locks. We are protected from hold races
2685 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2686 * lock held. A pv cannot be located otherwise.
2690 pv_put(pv_entry_t pv)
2692 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2693 if (pv->pv_pmap == NULL)
2702 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2703 * pmap. Any pte operations must have already been completed.
2707 pv_free(pv_entry_t pv)
2711 KKASSERT(pv->pv_m == NULL);
2712 if ((pmap = pv->pv_pmap) != NULL) {
2713 spin_lock(&pmap->pm_spin);
2714 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2715 if (pmap->pm_pvhint == pv)
2716 pmap->pm_pvhint = NULL;
2717 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2720 spin_unlock(&pmap->pm_spin);
2726 * This routine is very drastic, but can save the system
2734 static int warningdone=0;
2736 if (pmap_pagedaemon_waken == 0)
2738 pmap_pagedaemon_waken = 0;
2739 if (warningdone < 5) {
2740 kprintf("pmap_collect: collecting pv entries -- "
2741 "suggest increasing PMAP_SHPGPERPROC\n");
2745 for (i = 0; i < vm_page_array_size; i++) {
2746 m = &vm_page_array[i];
2747 if (m->wire_count || m->hold_count)
2749 if (vm_page_busy_try(m, TRUE) == 0) {
2750 if (m->wire_count == 0 && m->hold_count == 0) {
2759 * Scan the pmap for active page table entries and issue a callback.
2760 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
2761 * its parent page table.
2763 * pte_pv will be NULL if the page or page table is unmanaged.
2764 * pt_pv will point to the page table page containing the pte for the page.
2766 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
2767 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
2768 * process pmap's PD and page to the callback function. This can be
2769 * confusing because the pt_pv is really a pd_pv, and the target page
2770 * table page is simply aliased by the pmap and not owned by it.
2772 * It is assumed that the start and end are properly rounded to the page size.
2774 * It is assumed that PD pages and above are managed and thus in the RB tree,
2775 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
2777 struct pmap_scan_info {
2781 vm_pindex_t sva_pd_pindex;
2782 vm_pindex_t eva_pd_pindex;
2783 void (*func)(pmap_t, struct pmap_inval_info *,
2784 pv_entry_t, pv_entry_t, int, vm_offset_t,
2785 pt_entry_t *, void *);
2787 struct pmap_inval_info inval;
2790 static int pmap_scan_cmp(pv_entry_t pv, void *data);
2791 static int pmap_scan_callback(pv_entry_t pv, void *data);
2794 pmap_scan(struct pmap_scan_info *info)
2796 struct pmap *pmap = info->pmap;
2797 pv_entry_t pd_pv; /* A page directory PV */
2798 pv_entry_t pt_pv; /* A page table PV */
2799 pv_entry_t pte_pv; /* A page table entry PV */
2801 struct pv_entry dummy_pv;
2807 * Hold the token for stability; if the pmap is empty we have nothing
2810 lwkt_gettoken(&pmap->pm_token);
2812 if (pmap->pm_stats.resident_count == 0) {
2813 lwkt_reltoken(&pmap->pm_token);
2818 pmap_inval_init(&info->inval);
2821 * Special handling for scanning one page, which is a very common
2822 * operation (it is?).
2824 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
2826 if (info->sva + PAGE_SIZE == info->eva) {
2827 if (info->sva >= VM_MAX_USER_ADDRESS) {
2829 * Kernel mappings do not track wire counts on
2830 * page table pages and only maintain pd_pv and
2831 * pte_pv levels so pmap_scan() works.
2834 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
2835 ptep = vtopte(info->sva);
2838 * User pages which are unmanaged will not have a
2839 * pte_pv. User page table pages which are unmanaged
2840 * (shared from elsewhere) will also not have a pt_pv.
2841 * The func() callback will pass both pte_pv and pt_pv
2842 * as NULL in that case.
2844 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
2845 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
2846 if (pt_pv == NULL) {
2847 KKASSERT(pte_pv == NULL);
2848 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
2850 ptep = pv_pte_lookup(pd_pv,
2851 pmap_pt_index(info->sva));
2853 info->func(pmap, &info->inval,
2862 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
2866 * Unlike the pv_find() case below we actually
2867 * acquired a locked pv in this case so any
2868 * race should have been resolved. It is expected
2871 KKASSERT(pte_pv == NULL);
2872 } else if (pte_pv) {
2873 KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED|
2875 ("bad *ptep %016lx sva %016lx pte_pv %p",
2876 *ptep, info->sva, pte_pv));
2877 info->func(pmap, &info->inval, pte_pv, pt_pv, 0,
2878 info->sva, ptep, info->arg);
2880 KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V,
2881 ("bad *ptep %016lx sva %016lx pte_pv NULL",
2883 info->func(pmap, &info->inval, NULL, pt_pv, 0,
2884 info->sva, ptep, info->arg);
2889 pmap_inval_done(&info->inval);
2890 lwkt_reltoken(&pmap->pm_token);
2895 * Nominal scan case, RB_SCAN() for PD pages and iterate from
2898 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
2899 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
2901 if (info->sva >= VM_MAX_USER_ADDRESS) {
2903 * The kernel does not currently maintain any pv_entry's for
2904 * higher-level page tables.
2906 bzero(&dummy_pv, sizeof(dummy_pv));
2907 dummy_pv.pv_pindex = info->sva_pd_pindex;
2908 spin_lock(&pmap->pm_spin);
2909 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
2910 pmap_scan_callback(&dummy_pv, info);
2911 ++dummy_pv.pv_pindex;
2913 spin_unlock(&pmap->pm_spin);
2916 * User page tables maintain local PML4, PDP, and PD
2917 * pv_entry's at the very least. PT pv's might be
2918 * unmanaged and thus not exist. PTE pv's might be
2919 * unmanaged and thus not exist.
2921 spin_lock(&pmap->pm_spin);
2922 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
2923 pmap_scan_cmp, pmap_scan_callback, info);
2924 spin_unlock(&pmap->pm_spin);
2926 pmap_inval_done(&info->inval);
2927 lwkt_reltoken(&pmap->pm_token);
2931 * WARNING! pmap->pm_spin held
2934 pmap_scan_cmp(pv_entry_t pv, void *data)
2936 struct pmap_scan_info *info = data;
2937 if (pv->pv_pindex < info->sva_pd_pindex)
2939 if (pv->pv_pindex >= info->eva_pd_pindex)
2945 * WARNING! pmap->pm_spin held
2948 pmap_scan_callback(pv_entry_t pv, void *data)
2950 struct pmap_scan_info *info = data;
2951 struct pmap *pmap = info->pmap;
2952 pv_entry_t pd_pv; /* A page directory PV */
2953 pv_entry_t pt_pv; /* A page table PV */
2954 pv_entry_t pte_pv; /* A page table entry PV */
2958 vm_offset_t va_next;
2959 vm_pindex_t pd_pindex;
2963 * Pull the PD pindex from the pv before releasing the spinlock.
2965 * WARNING: pv is faked for kernel pmap scans.
2967 pd_pindex = pv->pv_pindex;
2968 spin_unlock(&pmap->pm_spin);
2969 pv = NULL; /* invalid after spinlock unlocked */
2972 * Calculate the page range within the PD. SIMPLE pmaps are
2973 * direct-mapped for the entire 2^64 address space. Normal pmaps
2974 * reflect the user and kernel address space which requires
2975 * cannonicalization w/regards to converting pd_pindex's back
2978 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
2979 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
2980 (sva & PML4_SIGNMASK)) {
2981 sva |= PML4_SIGNMASK;
2983 eva = sva + NBPDP; /* can overflow */
2984 if (sva < info->sva)
2986 if (eva < info->sva || eva > info->eva)
2990 * NOTE: kernel mappings do not track page table pages, only
2993 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
2994 * However, for the scan to be efficient we try to
2995 * cache items top-down.
3000 for (; sva < eva; sva = va_next) {
3002 if (sva >= VM_MAX_USER_ADDRESS) {
3011 * PD cache (degenerate case if we skip). It is possible
3012 * for the PD to not exist due to races. This is ok.
3014 if (pd_pv == NULL) {
3015 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3016 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3018 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3020 if (pd_pv == NULL) {
3021 va_next = (sva + NBPDP) & ~PDPMASK;
3030 if (pt_pv == NULL) {
3035 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3036 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3042 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3046 * If pt_pv is NULL we either have an shared page table
3047 * page and must issue a callback specific to that case,
3048 * or there is no page table page.
3050 * Either way we can skip the page table page.
3052 if (pt_pv == NULL) {
3054 * Possible unmanaged (shared from another pmap)
3058 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3059 KKASSERT(pd_pv != NULL);
3060 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3062 info->func(pmap, &info->inval, NULL, pd_pv, 1,
3063 sva, ptep, info->arg);
3067 * Done, move to next page table page.
3069 va_next = (sva + NBPDR) & ~PDRMASK;
3076 * From this point in the loop testing pt_pv for non-NULL
3077 * means we are in UVM, else if it is NULL we are in KVM.
3079 * Limit our scan to either the end of the va represented
3080 * by the current page table page, or to the end of the
3081 * range being removed.
3084 va_next = (sva + NBPDR) & ~PDRMASK;
3091 * Scan the page table for pages. Some pages may not be
3092 * managed (might not have a pv_entry).
3094 * There is no page table management for kernel pages so
3095 * pt_pv will be NULL in that case, but otherwise pt_pv
3096 * is non-NULL, locked, and referenced.
3100 * At this point a non-NULL pt_pv means a UVA, and a NULL
3101 * pt_pv means a KVA.
3104 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3108 while (sva < va_next) {
3110 * Acquire the related pte_pv, if any. If *ptep == 0
3111 * the related pte_pv should not exist, but if *ptep
3112 * is not zero the pte_pv may or may not exist (e.g.
3113 * will not exist for an unmanaged page).
3115 * However a multitude of races are possible here.
3117 * In addition, the (pt_pv, pte_pv) lock order is
3118 * backwards, so we have to be careful in aquiring
3119 * a properly locked pte_pv.
3123 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3130 pv_put(pt_pv); /* must be non-NULL */
3132 pv_lock(pte_pv); /* safe to block now */
3135 pt_pv = pv_get(pmap,
3136 pmap_pt_pindex(sva));
3138 * pt_pv reloaded, need new ptep
3140 KKASSERT(pt_pv != NULL);
3141 ptep = pv_pte_lookup(pt_pv,
3142 pmap_pte_index(sva));
3146 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3150 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3154 kprintf("Unexpected non-NULL pte_pv "
3155 "%p pt_pv %p *ptep = %016lx\n",
3156 pte_pv, pt_pv, *ptep);
3157 panic("Unexpected non-NULL pte_pv");
3165 * Ready for the callback. The locked pte_pv (if any)
3166 * is consumed by the callback. pte_pv will exist if
3167 * the page is managed, and will not exist if it
3171 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3173 ("bad *ptep %016lx sva %016lx "
3175 *ptep, sva, pte_pv));
3176 info->func(pmap, &info->inval, pte_pv, pt_pv, 0,
3177 sva, ptep, info->arg);
3179 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3181 ("bad *ptep %016lx sva %016lx "
3184 info->func(pmap, &info->inval, NULL, pt_pv, 0,
3185 sva, ptep, info->arg);
3202 * Relock before returning.
3204 spin_lock(&pmap->pm_spin);
3209 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3211 struct pmap_scan_info info;
3216 info.func = pmap_remove_callback;
3222 pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
3223 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3224 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3230 * This will also drop pt_pv's wire_count. Note that
3231 * terminal pages are not wired based on mmu presence.
3233 pmap_remove_pv_pte(pte_pv, pt_pv, info);
3234 pmap_remove_pv_page(pte_pv);
3236 } else if (sharept == 0) {
3240 * pt_pv's wire_count is still bumped by unmanaged pages
3241 * so we must decrement it manually.
3243 pmap_inval_interlock(info, pmap, va);
3244 pte = pte_load_clear(ptep);
3245 pmap_inval_deinterlock(info, pmap);
3247 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3248 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3249 if (vm_page_unwire_quick(pt_pv->pv_m))
3250 panic("pmap_remove: insufficient wirecount");
3253 * Unmanaged page table, pt_pv is actually the pd_pv
3254 * for our pmap (not the share object pmap).
3256 * We have to unwire the target page table page and we
3257 * have to unwire our page directory page.
3259 pmap_inval_interlock(info, pmap, va);
3260 pte = pte_load_clear(ptep);
3261 pmap_inval_deinterlock(info, pmap);
3262 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3263 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3264 panic("pmap_remove: shared pgtable1 bad wirecount");
3265 if (vm_page_unwire_quick(pt_pv->pv_m))
3266 panic("pmap_remove: shared pgtable2 bad wirecount");
3271 * Removes this physical page from all physical maps in which it resides.
3272 * Reflects back modify bits to the pager.
3274 * This routine may not be called from an interrupt.
3278 pmap_remove_all(vm_page_t m)
3280 struct pmap_inval_info info;
3283 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3286 pmap_inval_init(&info);
3287 vm_page_spin_lock(m);
3288 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3289 KKASSERT(pv->pv_m == m);
3290 if (pv_hold_try(pv)) {
3291 vm_page_spin_unlock(m);
3293 vm_page_spin_unlock(m);
3295 if (pv->pv_m != m) {
3297 vm_page_spin_lock(m);
3302 * Holding no spinlocks, pv is locked.
3304 pmap_remove_pv_pte(pv, NULL, &info);
3305 pmap_remove_pv_page(pv);
3307 vm_page_spin_lock(m);
3309 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3310 vm_page_spin_unlock(m);
3311 pmap_inval_done(&info);
3315 * Set the physical protection on the specified range of this map
3316 * as requested. This function is typically only used for debug watchpoints
3319 * This function may not be called from an interrupt if the map is
3320 * not the kernel_pmap.
3322 * NOTE! For shared page table pages we just unmap the page.
3325 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3327 struct pmap_scan_info info;
3328 /* JG review for NX */
3332 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3333 pmap_remove(pmap, sva, eva);
3336 if (prot & VM_PROT_WRITE)
3341 info.func = pmap_protect_callback;
3348 pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
3349 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3350 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3360 pmap_inval_interlock(info, pmap, va);
3367 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3368 KKASSERT(m == pte_pv->pv_m);
3369 vm_page_flag_set(m, PG_REFERENCED);
3373 if (pmap_track_modified(pte_pv->pv_pindex)) {
3375 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3380 } else if (sharept) {
3382 * Unmanaged page table, pt_pv is actually the pd_pv
3383 * for our pmap (not the share object pmap).
3385 * When asked to protect something in a shared page table
3386 * page we just unmap the page table page. We have to
3387 * invalidate the tlb in this situation.
3389 pte = pte_load_clear(ptep);
3390 pmap_inval_invltlb(info);
3391 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3392 panic("pmap_protect: pgtable1 pg bad wirecount");
3393 if (vm_page_unwire_quick(pt_pv->pv_m))
3394 panic("pmap_protect: pgtable2 pg bad wirecount");
3397 /* else unmanaged page, adjust bits, no wire changes */
3401 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
3405 pmap_inval_deinterlock(info, pmap);
3411 * Insert the vm_page (m) at the virtual address (va), replacing any prior
3412 * mapping at that address. Set protection and wiring as requested.
3414 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
3415 * possible. If it is we enter the page into the appropriate shared pmap
3416 * hanging off the related VM object instead of the passed pmap, then we
3417 * share the page table page from the VM object's pmap into the current pmap.
3419 * NOTE: This routine MUST insert the page into the pmap now, it cannot
3423 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3424 boolean_t wired, vm_map_entry_t entry __unused)
3426 pmap_inval_info info;
3427 pv_entry_t pt_pv; /* page table */
3428 pv_entry_t pte_pv; /* page table entry */
3431 pt_entry_t origpte, newpte;
3436 va = trunc_page(va);
3437 #ifdef PMAP_DIAGNOSTIC
3439 panic("pmap_enter: toobig");
3440 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
3441 panic("pmap_enter: invalid to pmap_enter page table "
3442 "pages (va: 0x%lx)", va);
3444 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
3445 kprintf("Warning: pmap_enter called on UVA with "
3448 db_print_backtrace();
3451 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
3452 kprintf("Warning: pmap_enter called on KVA without"
3455 db_print_backtrace();
3460 * Get locked PV entries for our new page table entry (pte_pv)
3461 * and for its parent page table (pt_pv). We need the parent
3462 * so we can resolve the location of the ptep.
3464 * Only hardware MMU actions can modify the ptep out from
3467 * if (m) is fictitious or unmanaged we do not create a managing
3468 * pte_pv for it. Any pre-existing page's management state must
3469 * match (avoiding code complexity).
3471 * If the pmap is still being initialized we assume existing
3474 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3475 * pmap_allocpte() checks the
3477 if (pmap_initialized == FALSE) {
3481 } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) { /* XXX */
3483 if (va >= VM_MAX_USER_ADDRESS) {
3487 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
3489 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3491 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0);
3493 if (va >= VM_MAX_USER_ADDRESS) {
3495 * Kernel map, pv_entry-tracked.
3498 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3504 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
3506 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3508 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED));
3511 pa = VM_PAGE_TO_PHYS(m);
3513 opa = origpte & PG_FRAME;
3515 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V | PG_A);
3518 if (va < VM_MAX_USER_ADDRESS)
3521 newpte |= PG_MANAGED;
3522 if (pmap == &kernel_pmap)
3526 * It is possible for multiple faults to occur in threaded
3527 * environments, the existing pte might be correct.
3529 if (((origpte ^ newpte) & ~(pt_entry_t)(PG_M|PG_A)) == 0)
3532 if ((prot & VM_PROT_NOSYNC) == 0)
3533 pmap_inval_init(&info);
3536 * Ok, either the address changed or the protection or wiring
3539 * Clear the current entry, interlocking the removal. For managed
3540 * pte's this will also flush the modified state to the vm_page.
3541 * Atomic ops are mandatory in order to ensure that PG_M events are
3542 * not lost during any transition.
3547 * pmap_remove_pv_pte() unwires pt_pv and assumes
3548 * we will free pte_pv, but since we are reusing
3549 * pte_pv we want to retain the wire count.
3551 * pt_pv won't exist for a kernel page (managed or
3555 vm_page_wire_quick(pt_pv->pv_m);
3556 if (prot & VM_PROT_NOSYNC)
3557 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3559 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3561 pmap_remove_pv_page(pte_pv);
3562 } else if (prot & VM_PROT_NOSYNC) {
3564 * Unmanaged page, NOSYNC (no mmu sync) requested.
3566 * Leave wire count on PT page intact.
3568 (void)pte_load_clear(ptep);
3569 cpu_invlpg((void *)va);
3570 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3573 * Unmanaged page, normal enter.
3575 * Leave wire count on PT page intact.
3577 pmap_inval_interlock(&info, pmap, va);
3578 (void)pte_load_clear(ptep);
3579 pmap_inval_deinterlock(&info, pmap);
3580 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3582 KKASSERT(*ptep == 0);
3587 * Enter on the PV list if part of our managed memory.
3588 * Wiring of the PT page is already handled.
3590 KKASSERT(pte_pv->pv_m == NULL);
3591 vm_page_spin_lock(m);
3593 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3596 atomic_add_int(&m->object->agg_pv_list_count, 1);
3598 vm_page_flag_set(m, PG_MAPPED);
3599 vm_page_spin_unlock(m);
3600 } else if (pt_pv && opa == 0) {
3602 * We have to adjust the wire count on the PT page ourselves
3603 * for unmanaged entries. If opa was non-zero we retained
3604 * the existing wire count from the removal.
3606 vm_page_wire_quick(pt_pv->pv_m);
3610 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
3612 * User VMAs do not because those will be zero->non-zero, so no
3613 * stale entries to worry about at this point.
3615 * For KVM there appear to still be issues. Theoretically we
3616 * should be able to scrap the interlocks entirely but we
3619 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3620 pmap_inval_interlock(&info, pmap, va);
3625 *(volatile pt_entry_t *)ptep = newpte;
3627 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3628 pmap_inval_deinterlock(&info, pmap);
3629 else if (pt_pv == NULL)
3630 cpu_invlpg((void *)va);
3634 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
3637 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3641 vm_page_flag_set(m, PG_WRITEABLE);
3644 * Unmanaged pages need manual resident_count tracking.
3646 if (pte_pv == NULL && pt_pv)
3647 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
3652 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
3653 pmap_inval_done(&info);
3655 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
3658 * Cleanup the pv entry, allowing other accessors.
3667 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
3668 * This code also assumes that the pmap has no pre-existing entry for this
3671 * This code currently may only be used on user pmaps, not kernel_pmap.
3674 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
3676 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
3680 * Make a temporary mapping for a physical address. This is only intended
3681 * to be used for panic dumps.
3683 * The caller is responsible for calling smp_invltlb().
3686 pmap_kenter_temporary(vm_paddr_t pa, long i)
3688 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
3689 return ((void *)crashdumpmap);
3692 #define MAX_INIT_PT (96)
3695 * This routine preloads the ptes for a given object into the specified pmap.
3696 * This eliminates the blast of soft faults on process startup and
3697 * immediately after an mmap.
3699 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
3702 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
3703 vm_object_t object, vm_pindex_t pindex,
3704 vm_size_t size, int limit)
3706 struct rb_vm_page_scan_info info;
3711 * We can't preinit if read access isn't set or there is no pmap
3714 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
3718 * We can't preinit if the pmap is not the current pmap
3720 lp = curthread->td_lwp;
3721 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
3725 * Misc additional checks
3727 psize = x86_64_btop(size);
3729 if ((object->type != OBJT_VNODE) ||
3730 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
3731 (object->resident_page_count > MAX_INIT_PT))) {
3735 if (pindex + psize > object->size) {
3736 if (object->size < pindex)
3738 psize = object->size - pindex;
3745 * If everything is segment-aligned do not pre-init here. Instead
3746 * allow the normal vm_fault path to pass a segment hint to
3747 * pmap_enter() which will then use an object-referenced shared
3750 if ((addr & SEG_MASK) == 0 &&
3751 (ctob(psize) & SEG_MASK) == 0 &&
3752 (ctob(pindex) & SEG_MASK) == 0) {
3757 * Use a red-black scan to traverse the requested range and load
3758 * any valid pages found into the pmap.
3760 * We cannot safely scan the object's memq without holding the
3763 info.start_pindex = pindex;
3764 info.end_pindex = pindex + psize - 1;
3770 vm_object_hold_shared(object);
3771 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3772 pmap_object_init_pt_callback, &info);
3773 vm_object_drop(object);
3778 pmap_object_init_pt_callback(vm_page_t p, void *data)
3780 struct rb_vm_page_scan_info *info = data;
3781 vm_pindex_t rel_index;
3784 * don't allow an madvise to blow away our really
3785 * free pages allocating pv entries.
3787 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3788 vmstats.v_free_count < vmstats.v_free_reserved) {
3793 * Ignore list markers and ignore pages we cannot instantly
3794 * busy (while holding the object token).
3796 if (p->flags & PG_MARKER)
3798 if (vm_page_busy_try(p, TRUE))
3800 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3801 (p->flags & PG_FICTITIOUS) == 0) {
3802 if ((p->queue - p->pc) == PQ_CACHE)
3803 vm_page_deactivate(p);
3804 rel_index = p->pindex - info->start_pindex;
3805 pmap_enter_quick(info->pmap,
3806 info->addr + x86_64_ptob(rel_index), p);
3814 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
3817 * Returns FALSE if it would be non-trivial or if a pte is already loaded
3820 * XXX This is safe only because page table pages are not freed.
3823 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3827 /*spin_lock(&pmap->pm_spin);*/
3828 if ((pte = pmap_pte(pmap, addr)) != NULL) {
3830 /*spin_unlock(&pmap->pm_spin);*/
3834 /*spin_unlock(&pmap->pm_spin);*/
3839 * Change the wiring attribute for a pmap/va pair. The mapping must already
3840 * exist in the pmap. The mapping may or may not be managed.
3843 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
3844 vm_map_entry_t entry)
3851 lwkt_gettoken(&pmap->pm_token);
3852 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
3853 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
3855 if (wired && !pmap_pte_w(ptep))
3856 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
3857 else if (!wired && pmap_pte_w(ptep))
3858 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
3861 * Wiring is not a hardware characteristic so there is no need to
3862 * invalidate TLB. However, in an SMP environment we must use
3863 * a locked bus cycle to update the pte (if we are not using
3864 * the pmap_inval_*() API that is)... it's ok to do this for simple
3869 atomic_set_long(ptep, PG_W);
3871 atomic_clear_long(ptep, PG_W);
3874 atomic_set_long_nonlocked(ptep, PG_W);
3876 atomic_clear_long_nonlocked(ptep, PG_W);
3879 lwkt_reltoken(&pmap->pm_token);
3885 * Copy the range specified by src_addr/len from the source map to
3886 * the range dst_addr/len in the destination map.
3888 * This routine is only advisory and need not do anything.
3891 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3892 vm_size_t len, vm_offset_t src_addr)
3899 * Zero the specified physical page.
3901 * This function may be called from an interrupt and no locking is
3905 pmap_zero_page(vm_paddr_t phys)
3907 vm_offset_t va = PHYS_TO_DMAP(phys);
3909 pagezero((void *)va);
3913 * pmap_page_assertzero:
3915 * Assert that a page is empty, panic if it isn't.
3918 pmap_page_assertzero(vm_paddr_t phys)
3920 vm_offset_t va = PHYS_TO_DMAP(phys);
3923 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3924 if (*(long *)((char *)va + i) != 0) {
3925 panic("pmap_page_assertzero() @ %p not zero!",
3926 (void *)(intptr_t)va);
3934 * Zero part of a physical page by mapping it into memory and clearing
3935 * its contents with bzero.
3937 * off and size may not cover an area beyond a single hardware page.
3940 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3942 vm_offset_t virt = PHYS_TO_DMAP(phys);
3944 bzero((char *)virt + off, size);
3950 * Copy the physical page from the source PA to the target PA.
3951 * This function may be called from an interrupt. No locking
3955 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3957 vm_offset_t src_virt, dst_virt;
3959 src_virt = PHYS_TO_DMAP(src);
3960 dst_virt = PHYS_TO_DMAP(dst);
3961 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3965 * pmap_copy_page_frag:
3967 * Copy the physical page from the source PA to the target PA.
3968 * This function may be called from an interrupt. No locking
3972 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3974 vm_offset_t src_virt, dst_virt;
3976 src_virt = PHYS_TO_DMAP(src);
3977 dst_virt = PHYS_TO_DMAP(dst);
3979 bcopy((char *)src_virt + (src & PAGE_MASK),
3980 (char *)dst_virt + (dst & PAGE_MASK),
3985 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
3986 * this page. This count may be changed upwards or downwards in the future;
3987 * it is only necessary that true be returned for a small subset of pmaps
3988 * for proper page aging.
3991 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3996 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3999 vm_page_spin_lock(m);
4000 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4001 if (pv->pv_pmap == pmap) {
4002 vm_page_spin_unlock(m);
4009 vm_page_spin_unlock(m);
4014 * Remove all pages from specified address space this aids process exit
4015 * speeds. Also, this code may be special cased for the current process
4019 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4021 pmap_remove(pmap, sva, eva);
4025 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4026 * routines are inline, and a lot of things compile-time evaluate.
4030 pmap_testbit(vm_page_t m, int bit)
4035 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4038 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4040 vm_page_spin_lock(m);
4041 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4042 vm_page_spin_unlock(m);
4046 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4048 * if the bit being tested is the modified bit, then
4049 * mark clean_map and ptes as never
4052 if (bit & (PG_A|PG_M)) {
4053 if (!pmap_track_modified(pv->pv_pindex))
4057 #if defined(PMAP_DIAGNOSTIC)
4058 if (pv->pv_pmap == NULL) {
4059 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4064 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4066 vm_page_spin_unlock(m);
4070 vm_page_spin_unlock(m);
4075 * This routine is used to modify bits in ptes. Only one bit should be
4076 * specified. PG_RW requires special handling.
4078 * Caller must NOT hold any spin locks
4082 pmap_clearbit(vm_page_t m, int bit)
4084 struct pmap_inval_info info;
4091 vm_page_flag_clear(m, PG_WRITEABLE);
4092 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4099 * Loop over all current mappings setting/clearing as appropos If
4100 * setting RO do we need to clear the VAC?
4102 * NOTE: When clearing PG_M we could also (not implemented) drop
4103 * through to the PG_RW code and clear PG_RW too, forcing
4104 * a fault on write to redetect PG_M for virtual kernels, but
4105 * it isn't necessary since virtual kernels invalidate the
4106 * pte when they clear the VPTE_M bit in their virtual page
4109 * NOTE: Does not re-dirty the page when clearing only PG_M.
4111 if ((bit & PG_RW) == 0) {
4112 vm_page_spin_lock(m);
4113 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4114 #if defined(PMAP_DIAGNOSTIC)
4115 if (pv->pv_pmap == NULL) {
4116 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4121 pte = pmap_pte_quick(pv->pv_pmap,
4122 pv->pv_pindex << PAGE_SHIFT);
4125 atomic_clear_long(pte, bit);
4127 vm_page_spin_unlock(m);
4132 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4135 pmap_inval_init(&info);
4138 vm_page_spin_lock(m);
4139 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4141 * don't write protect pager mappings
4143 if (!pmap_track_modified(pv->pv_pindex))
4146 #if defined(PMAP_DIAGNOSTIC)
4147 if (pv->pv_pmap == NULL) {
4148 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4154 * Skip pages which do not have PG_RW set.
4156 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4157 if ((*pte & PG_RW) == 0)
4163 if (pv_hold_try(pv) == 0) {
4164 vm_page_spin_unlock(m);
4165 pv_lock(pv); /* held, now do a blocking lock */
4166 pv_put(pv); /* and release */
4167 goto restart; /* anything could have happened */
4170 save_pmap = pv->pv_pmap;
4171 vm_page_spin_unlock(m);
4172 pmap_inval_interlock(&info, save_pmap,
4173 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
4174 KKASSERT(pv->pv_pmap == save_pmap);
4178 if (atomic_cmpset_long(pte, pbits,
4179 pbits & ~(PG_RW|PG_M))) {
4183 pmap_inval_deinterlock(&info, save_pmap);
4184 vm_page_spin_lock(m);
4187 * If PG_M was found to be set while we were clearing PG_RW
4188 * we also clear PG_M (done above) and mark the page dirty.
4189 * Callers expect this behavior.
4195 vm_page_spin_unlock(m);
4196 pmap_inval_done(&info);
4200 * Lower the permission for all mappings to a given page.
4202 * Page must be busied by caller.
4205 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4207 /* JG NX support? */
4208 if ((prot & VM_PROT_WRITE) == 0) {
4209 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4211 * NOTE: pmap_clearbit(.. PG_RW) also clears
4212 * the PG_WRITEABLE flag in (m).
4214 pmap_clearbit(m, PG_RW);
4222 pmap_phys_address(vm_pindex_t ppn)
4224 return (x86_64_ptob(ppn));
4228 * Return a count of reference bits for a page, clearing those bits.
4229 * It is not necessary for every reference bit to be cleared, but it
4230 * is necessary that 0 only be returned when there are truly no
4231 * reference bits set.
4233 * XXX: The exact number of bits to check and clear is a matter that
4234 * should be tested and standardized at some point in the future for
4235 * optimal aging of shared pages.
4237 * This routine may not block.
4240 pmap_ts_referenced(vm_page_t m)
4246 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4249 vm_page_spin_lock(m);
4250 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4251 if (!pmap_track_modified(pv->pv_pindex))
4253 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4254 if (pte && (*pte & PG_A)) {
4256 atomic_clear_long(pte, PG_A);
4258 atomic_clear_long_nonlocked(pte, PG_A);
4265 vm_page_spin_unlock(m);
4272 * Return whether or not the specified physical page was modified
4273 * in any physical maps.
4276 pmap_is_modified(vm_page_t m)
4280 res = pmap_testbit(m, PG_M);
4285 * Clear the modify bits on the specified physical page.
4288 pmap_clear_modify(vm_page_t m)
4290 pmap_clearbit(m, PG_M);
4294 * pmap_clear_reference:
4296 * Clear the reference bit on the specified physical page.
4299 pmap_clear_reference(vm_page_t m)
4301 pmap_clearbit(m, PG_A);
4305 * Miscellaneous support routines follow
4310 i386_protection_init(void)
4314 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4315 kp = protection_codes;
4316 for (prot = 0; prot < 8; prot++) {
4318 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4320 * Read access is also 0. There isn't any execute bit,
4321 * so just make it readable.
4323 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4324 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4325 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4328 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4329 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4330 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4331 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4339 * Map a set of physical memory pages into the kernel virtual
4340 * address space. Return a pointer to where it is mapped. This
4341 * routine is intended to be used for mapping device memory,
4344 * NOTE: we can't use pgeflag unless we invalidate the pages one at
4348 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
4350 vm_offset_t va, tmpva, offset;
4353 offset = pa & PAGE_MASK;
4354 size = roundup(offset + size, PAGE_SIZE);
4356 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4358 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4360 pa = pa & ~PAGE_MASK;
4361 for (tmpva = va; size > 0;) {
4362 pte = vtopte(tmpva);
4363 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
4371 return ((void *)(va + offset));
4375 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
4377 vm_offset_t va, tmpva, offset;
4380 offset = pa & PAGE_MASK;
4381 size = roundup(offset + size, PAGE_SIZE);
4383 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4385 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4387 pa = pa & ~PAGE_MASK;
4388 for (tmpva = va; size > 0;) {
4389 pte = vtopte(tmpva);
4390 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
4398 return ((void *)(va + offset));
4402 pmap_unmapdev(vm_offset_t va, vm_size_t size)
4404 vm_offset_t base, offset;
4406 base = va & ~PAGE_MASK;
4407 offset = va & PAGE_MASK;
4408 size = roundup(offset + size, PAGE_SIZE);
4409 pmap_qremove(va, size >> PAGE_SHIFT);
4410 kmem_free(&kernel_map, base, size);
4414 * perform the pmap work for mincore
4417 pmap_mincore(pmap_t pmap, vm_offset_t addr)
4419 pt_entry_t *ptep, pte;
4423 lwkt_gettoken(&pmap->pm_token);
4424 ptep = pmap_pte(pmap, addr);
4426 if (ptep && (pte = *ptep) != 0) {
4429 val = MINCORE_INCORE;
4430 if ((pte & PG_MANAGED) == 0)
4433 pa = pte & PG_FRAME;
4435 m = PHYS_TO_VM_PAGE(pa);
4441 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
4443 * Modified by someone
4445 else if (m->dirty || pmap_is_modified(m))
4446 val |= MINCORE_MODIFIED_OTHER;
4451 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
4454 * Referenced by someone
4456 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
4457 val |= MINCORE_REFERENCED_OTHER;
4458 vm_page_flag_set(m, PG_REFERENCED);
4462 lwkt_reltoken(&pmap->pm_token);
4468 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
4469 * vmspace will be ref'd and the old one will be deref'd.
4471 * The vmspace for all lwps associated with the process will be adjusted
4472 * and cr3 will be reloaded if any lwp is the current lwp.
4474 * The process must hold the vmspace->vm_map.token for oldvm and newvm
4477 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
4479 struct vmspace *oldvm;
4482 oldvm = p->p_vmspace;
4483 if (oldvm != newvm) {
4485 sysref_get(&newvm->vm_sysref);
4486 p->p_vmspace = newvm;
4487 KKASSERT(p->p_nthreads == 1);
4488 lp = RB_ROOT(&p->p_lwp_tree);
4489 pmap_setlwpvm(lp, newvm);
4491 sysref_put(&oldvm->vm_sysref);
4496 * Set the vmspace for a LWP. The vmspace is almost universally set the
4497 * same as the process vmspace, but virtual kernels need to swap out contexts
4498 * on a per-lwp basis.
4500 * Caller does not necessarily hold any vmspace tokens. Caller must control
4501 * the lwp (typically be in the context of the lwp). We use a critical
4502 * section to protect against statclock and hardclock (statistics collection).
4505 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
4507 struct vmspace *oldvm;
4510 oldvm = lp->lwp_vmspace;
4512 if (oldvm != newvm) {
4514 lp->lwp_vmspace = newvm;
4515 if (curthread->td_lwp == lp) {
4516 pmap = vmspace_pmap(newvm);
4518 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4519 if (pmap->pm_active & CPUMASK_LOCK)
4520 pmap_interlock_wait(newvm);
4522 pmap->pm_active |= 1;
4524 #if defined(SWTCH_OPTIM_STATS)
4527 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4528 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
4529 load_cr3(curthread->td_pcb->pcb_cr3);
4530 pmap = vmspace_pmap(oldvm);
4532 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4534 pmap->pm_active &= ~(cpumask_t)1;
4544 * Called when switching to a locked pmap, used to interlock against pmaps
4545 * undergoing modifications to prevent us from activating the MMU for the
4546 * target pmap until all such modifications have completed. We have to do
4547 * this because the thread making the modifications has already set up its
4548 * SMP synchronization mask.
4550 * This function cannot sleep!
4555 pmap_interlock_wait(struct vmspace *vm)
4557 struct pmap *pmap = &vm->vm_pmap;
4559 if (pmap->pm_active & CPUMASK_LOCK) {
4561 KKASSERT(curthread->td_critcount >= 2);
4562 DEBUG_PUSH_INFO("pmap_interlock_wait");
4563 while (pmap->pm_active & CPUMASK_LOCK) {
4565 lwkt_process_ipiq();
4575 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4578 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4582 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4587 * Used by kmalloc/kfree, page already exists at va
4590 pmap_kvtom(vm_offset_t va)
4592 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));
4596 * Initialize machine-specific shared page directory support. This
4597 * is executed when a VM object is created.
4600 pmap_object_init(vm_object_t object)
4602 object->md.pmap_rw = NULL;
4603 object->md.pmap_ro = NULL;
4607 * Clean up machine-specific shared page directory support. This
4608 * is executed when a VM object is destroyed.
4611 pmap_object_free(vm_object_t object)
4615 if ((pmap = object->md.pmap_rw) != NULL) {
4616 object->md.pmap_rw = NULL;
4617 kprintf("pmap_object_free: destroying pmap %p in obj %p\n",
4619 pmap_remove_pages(pmap,
4620 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
4621 pmap->pm_active = 0;
4624 kfree(pmap, M_OBJPMAP);
4626 if ((pmap = object->md.pmap_ro) != NULL) {
4627 object->md.pmap_ro = NULL;
4628 kprintf("pmap_object_free: destroying pmap %p in obj %p\n",
4630 pmap_remove_pages(pmap,
4631 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
4632 pmap->pm_active = 0;
4635 kfree(pmap, M_OBJPMAP);