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 = NULL, ptvmmap = NULL;
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_pte, vm_offset_t va)
439 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & 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_pte, vm_offset_t va)
467 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
468 return (&pt[pmap_pt_index(va)]);
472 * Return pointer to PT slot in the PD
474 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
475 * so we cannot lookup the PD via the PDP. Instead we
476 * must look it up via the pmap.
480 pmap_pt(pmap_t pmap, vm_offset_t va)
484 vm_pindex_t pd_pindex;
486 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
487 pd_pindex = pmap_pd_pindex(va);
488 spin_lock(&pmap->pm_spin);
489 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
490 spin_unlock(&pmap->pm_spin);
491 if (pv == NULL || pv->pv_m == NULL)
493 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
495 pd = pmap_pd(pmap, va);
496 if (pd == NULL || (*pd & PG_V) == 0)
498 return (pmap_pd_to_pt(*pd, va));
503 * Return pointer to PTE slot in the PT given a pointer to the PT
507 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
511 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
512 return (&pte[pmap_pte_index(va)]);
516 * Return pointer to PTE slot in the PT
520 pmap_pte(pmap_t pmap, vm_offset_t va)
524 pt = pmap_pt(pmap, va);
525 if (pt == NULL || (*pt & PG_V) == 0)
527 if ((*pt & PG_PS) != 0)
528 return ((pt_entry_t *)pt);
529 return (pmap_pt_to_pte(*pt, va));
533 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
534 * the PT layer. This will speed up core pmap operations considerably.
538 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
540 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
541 pv->pv_pmap->pm_pvhint = pv;
546 * KVM - return address of PT slot in PD
550 vtopt(vm_offset_t va)
552 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
553 NPML4EPGSHIFT)) - 1);
555 return (PDmap + ((va >> PDRSHIFT) & mask));
559 * KVM - return address of PTE slot in PT
563 vtopte(vm_offset_t va)
565 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
566 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
568 return (PTmap + ((va >> PAGE_SHIFT) & mask));
572 allocpages(vm_paddr_t *firstaddr, long n)
577 bzero((void *)ret, n * PAGE_SIZE);
578 *firstaddr += n * PAGE_SIZE;
584 create_pagetables(vm_paddr_t *firstaddr)
586 long i; /* must be 64 bits */
592 * We are running (mostly) V=P at this point
594 * Calculate NKPT - number of kernel page tables. We have to
595 * accomodoate prealloction of the vm_page_array, dump bitmap,
596 * MSGBUF_SIZE, and other stuff. Be generous.
598 * Maxmem is in pages.
600 * ndmpdp is the number of 1GB pages we wish to map.
602 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
603 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
605 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
608 * Starting at the beginning of kvm (not KERNBASE).
610 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
611 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
612 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
613 ndmpdp) + 511) / 512;
617 * Starting at KERNBASE - map 2G worth of page table pages.
618 * KERNBASE is offset -2G from the end of kvm.
620 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
625 KPTbase = allocpages(firstaddr, nkpt_base);
626 KPTphys = allocpages(firstaddr, nkpt_phys);
627 KPML4phys = allocpages(firstaddr, 1);
628 KPDPphys = allocpages(firstaddr, NKPML4E);
629 KPDphys = allocpages(firstaddr, NKPDPE);
632 * Calculate the page directory base for KERNBASE,
633 * that is where we start populating the page table pages.
634 * Basically this is the end - 2.
636 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
638 DMPDPphys = allocpages(firstaddr, NDMPML4E);
639 if ((amd_feature & AMDID_PAGE1GB) == 0)
640 DMPDphys = allocpages(firstaddr, ndmpdp);
641 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
644 * Fill in the underlying page table pages for the area around
645 * KERNBASE. This remaps low physical memory to KERNBASE.
647 * Read-only from zero to physfree
648 * XXX not fully used, underneath 2M pages
650 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
651 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
652 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
656 * Now map the initial kernel page tables. One block of page
657 * tables is placed at the beginning of kernel virtual memory,
658 * and another block is placed at KERNBASE to map the kernel binary,
659 * data, bss, and initial pre-allocations.
661 for (i = 0; i < nkpt_base; i++) {
662 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
663 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
665 for (i = 0; i < nkpt_phys; i++) {
666 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
667 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
671 * Map from zero to end of allocations using 2M pages as an
672 * optimization. This will bypass some of the KPTBase pages
673 * above in the KERNBASE area.
675 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
676 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
677 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
681 * And connect up the PD to the PDP. The kernel pmap is expected
682 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
684 for (i = 0; i < NKPDPE; i++) {
685 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
686 KPDphys + (i << PAGE_SHIFT);
687 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
692 * Now set up the direct map space using either 2MB or 1GB pages
693 * Preset PG_M and PG_A because demotion expects it.
695 * When filling in entries in the PD pages make sure any excess
696 * entries are set to zero as we allocated enough PD pages
698 if ((amd_feature & AMDID_PAGE1GB) == 0) {
699 for (i = 0; i < NPDEPG * ndmpdp; i++) {
700 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
701 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
706 * And the direct map space's PDP
708 for (i = 0; i < ndmpdp; i++) {
709 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
711 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
714 for (i = 0; i < ndmpdp; i++) {
715 ((pdp_entry_t *)DMPDPphys)[i] =
716 (vm_paddr_t)i << PDPSHIFT;
717 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
722 /* And recursively map PML4 to itself in order to get PTmap */
723 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
724 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
727 * Connect the Direct Map slots up to the PML4
729 for (j = 0; j < NDMPML4E; ++j) {
730 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
731 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
736 * Connect the KVA slot up to the PML4
738 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
739 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
743 * Bootstrap the system enough to run with virtual memory.
745 * On the i386 this is called after mapping has already been enabled
746 * and just syncs the pmap module with what has already been done.
747 * [We can't call it easily with mapping off since the kernel is not
748 * mapped with PA == VA, hence we would have to relocate every address
749 * from the linked base (virtual) address "KERNBASE" to the actual
750 * (physical) address starting relative to 0]
753 pmap_bootstrap(vm_paddr_t *firstaddr)
757 struct mdglobaldata *gd;
760 KvaStart = VM_MIN_KERNEL_ADDRESS;
761 KvaEnd = VM_MAX_KERNEL_ADDRESS;
762 KvaSize = KvaEnd - KvaStart;
764 avail_start = *firstaddr;
767 * Create an initial set of page tables to run the kernel in.
769 create_pagetables(firstaddr);
771 virtual2_start = KvaStart;
772 virtual2_end = PTOV_OFFSET;
774 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
775 virtual_start = pmap_kmem_choose(virtual_start);
777 virtual_end = VM_MAX_KERNEL_ADDRESS;
779 /* XXX do %cr0 as well */
780 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
784 * Initialize protection array.
786 i386_protection_init();
789 * The kernel's pmap is statically allocated so we don't have to use
790 * pmap_create, which is unlikely to work correctly at this part of
791 * the boot sequence (XXX and which no longer exists).
793 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
794 kernel_pmap.pm_count = 1;
795 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
796 RB_INIT(&kernel_pmap.pm_pvroot);
797 spin_init(&kernel_pmap.pm_spin);
798 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
801 * Reserve some special page table entries/VA space for temporary
804 #define SYSMAP(c, p, v, n) \
805 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
811 * CMAP1/CMAP2 are used for zeroing and copying pages.
813 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
818 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
821 * ptvmmap is used for reading arbitrary physical pages via
824 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
827 * msgbufp is used to map the system message buffer.
828 * XXX msgbufmap is not used.
830 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
831 atop(round_page(MSGBUF_SIZE)))
838 * PG_G is terribly broken on SMP because we IPI invltlb's in some
839 * cases rather then invl1pg. Actually, I don't even know why it
840 * works under UP because self-referential page table mappings
845 if (cpu_feature & CPUID_PGE)
850 * Initialize the 4MB page size flag
854 * The 4MB page version of the initial
855 * kernel page mapping.
859 #if !defined(DISABLE_PSE)
860 if (cpu_feature & CPUID_PSE) {
863 * Note that we have enabled PSE mode
866 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
867 ptditmp &= ~(NBPDR - 1);
868 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
873 * Enable the PSE mode. If we are SMP we can't do this
874 * now because the APs will not be able to use it when
877 load_cr4(rcr4() | CR4_PSE);
880 * We can do the mapping here for the single processor
881 * case. We simply ignore the old page table page from
885 * For SMP, we still need 4K pages to bootstrap APs,
886 * PSE will be enabled as soon as all APs are up.
888 PTD[KPTDI] = (pd_entry_t)ptditmp;
895 * We need to finish setting up the globaldata page for the BSP.
896 * locore has already populated the page table for the mdglobaldata
899 pg = MDGLOBALDATA_BASEALLOC_PAGES;
900 gd = &CPU_prvspace[0].mdglobaldata;
907 * Set 4mb pdir for mp startup
912 if (pseflag && (cpu_feature & CPUID_PSE)) {
913 load_cr4(rcr4() | CR4_PSE);
914 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
922 * Initialize the pmap module.
923 * Called by vm_init, to initialize any structures that the pmap
924 * system needs to map virtual memory.
925 * pmap_init has been enhanced to support in a fairly consistant
926 * way, discontiguous physical memory.
935 * Allocate memory for random pmap data structures. Includes the
939 for (i = 0; i < vm_page_array_size; i++) {
942 m = &vm_page_array[i];
943 TAILQ_INIT(&m->md.pv_list);
947 * init the pv free list
949 initial_pvs = vm_page_array_size;
950 if (initial_pvs < MINPV)
952 pvzone = &pvzone_store;
953 pvinit = (void *)kmem_alloc(&kernel_map,
954 initial_pvs * sizeof (struct pv_entry));
955 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
956 pvinit, initial_pvs);
959 * Now it is safe to enable pv_table recording.
961 pmap_initialized = TRUE;
965 * Initialize the address space (zone) for the pv_entries. Set a
966 * high water mark so that the system can recover from excessive
967 * numbers of pv entries.
972 int shpgperproc = PMAP_SHPGPERPROC;
975 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
976 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
977 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
978 pv_entry_high_water = 9 * (pv_entry_max / 10);
981 * Subtract out pages already installed in the zone (hack)
983 entry_max = pv_entry_max - vm_page_array_size;
987 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
991 /***************************************************
992 * Low level helper routines.....
993 ***************************************************/
996 * this routine defines the region(s) of memory that should
997 * not be tested for the modified bit.
1001 pmap_track_modified(vm_pindex_t pindex)
1003 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1004 if ((va < clean_sva) || (va >= clean_eva))
1011 * Extract the physical page address associated with the map/VA pair.
1012 * The page must be wired for this to work reliably.
1014 * XXX for the moment we're using pv_find() instead of pv_get(), as
1015 * callers might be expecting non-blocking operation.
1018 pmap_extract(pmap_t pmap, vm_offset_t va)
1025 if (va >= VM_MAX_USER_ADDRESS) {
1027 * Kernel page directories might be direct-mapped and
1028 * there is typically no PV tracking of pte's
1032 pt = pmap_pt(pmap, va);
1033 if (pt && (*pt & PG_V)) {
1035 rtval = *pt & PG_PS_FRAME;
1036 rtval |= va & PDRMASK;
1038 ptep = pmap_pt_to_pte(*pt, va);
1040 rtval = *ptep & PG_FRAME;
1041 rtval |= va & PAGE_MASK;
1047 * User pages currently do not direct-map the page directory
1048 * and some pages might not used managed PVs. But all PT's
1051 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1053 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1055 rtval = *ptep & PG_FRAME;
1056 rtval |= va & PAGE_MASK;
1065 * Extract the physical page address associated kernel virtual address.
1068 pmap_kextract(vm_offset_t va)
1070 pd_entry_t pt; /* pt entry in pd */
1073 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1074 pa = DMAP_TO_PHYS(va);
1078 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1081 * Beware of a concurrent promotion that changes the
1082 * PDE at this point! For example, vtopte() must not
1083 * be used to access the PTE because it would use the
1084 * new PDE. It is, however, safe to use the old PDE
1085 * because the page table page is preserved by the
1088 pa = *pmap_pt_to_pte(pt, va);
1089 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1095 /***************************************************
1096 * Low level mapping routines.....
1097 ***************************************************/
1100 * Routine: pmap_kenter
1102 * Add a wired page to the KVA
1103 * NOTE! note that in order for the mapping to take effect -- you
1104 * should do an invltlb after doing the pmap_kenter().
1107 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1111 pmap_inval_info info;
1113 pmap_inval_init(&info); /* XXX remove */
1114 npte = pa | PG_RW | PG_V | pgeflag;
1116 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1118 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1119 pmap_inval_done(&info); /* XXX remove */
1123 * Routine: pmap_kenter_quick
1125 * Similar to pmap_kenter(), except we only invalidate the
1126 * mapping on the current CPU.
1129 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1134 npte = pa | PG_RW | PG_V | pgeflag;
1137 cpu_invlpg((void *)va);
1141 pmap_kenter_sync(vm_offset_t va)
1143 pmap_inval_info info;
1145 pmap_inval_init(&info);
1146 pmap_inval_interlock(&info, &kernel_pmap, va);
1147 pmap_inval_deinterlock(&info, &kernel_pmap);
1148 pmap_inval_done(&info);
1152 pmap_kenter_sync_quick(vm_offset_t va)
1154 cpu_invlpg((void *)va);
1158 * remove a page from the kernel pagetables
1161 pmap_kremove(vm_offset_t va)
1164 pmap_inval_info info;
1166 pmap_inval_init(&info);
1168 pmap_inval_interlock(&info, &kernel_pmap, va);
1169 (void)pte_load_clear(pte);
1170 pmap_inval_deinterlock(&info, &kernel_pmap);
1171 pmap_inval_done(&info);
1175 pmap_kremove_quick(vm_offset_t va)
1179 (void)pte_load_clear(pte);
1180 cpu_invlpg((void *)va);
1184 * XXX these need to be recoded. They are not used in any critical path.
1187 pmap_kmodify_rw(vm_offset_t va)
1189 atomic_set_long(vtopte(va), PG_RW);
1190 cpu_invlpg((void *)va);
1194 pmap_kmodify_nc(vm_offset_t va)
1196 atomic_set_long(vtopte(va), PG_N);
1197 cpu_invlpg((void *)va);
1201 * Used to map a range of physical addresses into kernel virtual
1202 * address space during the low level boot, typically to map the
1203 * dump bitmap, message buffer, and vm_page_array.
1205 * These mappings are typically made at some pointer after the end of the
1208 * We could return PHYS_TO_DMAP(start) here and not allocate any
1209 * via (*virtp), but then kmem from userland and kernel dumps won't
1210 * have access to the related pointers.
1213 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1216 vm_offset_t va_start;
1218 /*return PHYS_TO_DMAP(start);*/
1223 while (start < end) {
1224 pmap_kenter_quick(va, start);
1234 * Add a list of wired pages to the kva
1235 * this routine is only used for temporary
1236 * kernel mappings that do not need to have
1237 * page modification or references recorded.
1238 * Note that old mappings are simply written
1239 * over. The page *must* be wired.
1242 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1246 end_va = va + count * PAGE_SIZE;
1248 while (va < end_va) {
1252 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1253 cpu_invlpg((void *)va);
1261 * This routine jerks page mappings from the
1262 * kernel -- it is meant only for temporary mappings.
1264 * MPSAFE, INTERRUPT SAFE (cluster callback)
1267 pmap_qremove(vm_offset_t va, int count)
1271 end_va = va + count * PAGE_SIZE;
1273 while (va < end_va) {
1277 (void)pte_load_clear(pte);
1278 cpu_invlpg((void *)va);
1285 * Create a new thread and optionally associate it with a (new) process.
1286 * NOTE! the new thread's cpu may not equal the current cpu.
1289 pmap_init_thread(thread_t td)
1291 /* enforce pcb placement & alignment */
1292 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1293 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1294 td->td_savefpu = &td->td_pcb->pcb_save;
1295 td->td_sp = (char *)td->td_pcb; /* no -16 */
1299 * This routine directly affects the fork perf for a process.
1302 pmap_init_proc(struct proc *p)
1307 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1308 * it, and IdlePTD, represents the template used to update all other pmaps.
1310 * On architectures where the kernel pmap is not integrated into the user
1311 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1312 * kernel_pmap should be used to directly access the kernel_pmap.
1315 pmap_pinit0(struct pmap *pmap)
1317 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1319 pmap->pm_active = 0;
1320 pmap->pm_pvhint = NULL;
1321 RB_INIT(&pmap->pm_pvroot);
1322 spin_init(&pmap->pm_spin);
1323 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1324 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1328 * Initialize a preallocated and zeroed pmap structure,
1329 * such as one in a vmspace structure.
1332 pmap_pinit_simple(struct pmap *pmap)
1335 * Misc initialization
1338 pmap->pm_active = 0;
1339 pmap->pm_pvhint = NULL;
1340 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1343 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1346 if (pmap->pm_pmlpv == NULL) {
1347 RB_INIT(&pmap->pm_pvroot);
1348 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1349 spin_init(&pmap->pm_spin);
1350 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1355 pmap_pinit(struct pmap *pmap)
1360 pmap_pinit_simple(pmap);
1361 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1364 * No need to allocate page table space yet but we do need a valid
1365 * page directory table.
1367 if (pmap->pm_pml4 == NULL) {
1369 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1373 * Allocate the page directory page, which wires it even though
1374 * it isn't being entered into some higher level page table (it
1375 * being the highest level). If one is already cached we don't
1376 * have to do anything.
1378 if ((pv = pmap->pm_pmlpv) == NULL) {
1379 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1380 pmap->pm_pmlpv = pv;
1381 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1382 VM_PAGE_TO_PHYS(pv->pv_m));
1386 * Install DMAP and KMAP.
1388 for (j = 0; j < NDMPML4E; ++j) {
1389 pmap->pm_pml4[DMPML4I + j] =
1390 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1391 PG_RW | PG_V | PG_U;
1393 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1396 * install self-referential address mapping entry
1398 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1399 PG_V | PG_RW | PG_A | PG_M;
1401 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1402 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1404 KKASSERT(pmap->pm_pml4[255] == 0);
1405 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1406 KKASSERT(pv->pv_entry.rbe_left == NULL);
1407 KKASSERT(pv->pv_entry.rbe_right == NULL);
1411 * Clean up a pmap structure so it can be physically freed. This routine
1412 * is called by the vmspace dtor function. A great deal of pmap data is
1413 * left passively mapped to improve vmspace management so we have a bit
1414 * of cleanup work to do here.
1417 pmap_puninit(pmap_t pmap)
1422 KKASSERT(pmap->pm_active == 0);
1423 if ((pv = pmap->pm_pmlpv) != NULL) {
1424 if (pv_hold_try(pv) == 0)
1426 p = pmap_remove_pv_page(pv);
1428 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1429 vm_page_busy_wait(p, FALSE, "pgpun");
1430 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1431 vm_page_unwire(p, 0);
1432 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1435 * XXX eventually clean out PML4 static entries and
1436 * use vm_page_free_zero()
1439 pmap->pm_pmlpv = NULL;
1441 if (pmap->pm_pml4) {
1442 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1443 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1444 pmap->pm_pml4 = NULL;
1446 KKASSERT(pmap->pm_stats.resident_count == 0);
1447 KKASSERT(pmap->pm_stats.wired_count == 0);
1451 * Wire in kernel global address entries. To avoid a race condition
1452 * between pmap initialization and pmap_growkernel, this procedure
1453 * adds the pmap to the master list (which growkernel scans to update),
1454 * then copies the template.
1457 pmap_pinit2(struct pmap *pmap)
1459 spin_lock(&pmap_spin);
1460 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1461 spin_unlock(&pmap_spin);
1465 * This routine is called when various levels in the page table need to
1466 * be populated. This routine cannot fail.
1468 * This function returns two locked pv_entry's, one representing the
1469 * requested pv and one representing the requested pv's parent pv. If
1470 * the pv did not previously exist it will be mapped into its parent
1471 * and wired, otherwise no additional wire count will be added.
1475 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1480 vm_pindex_t pt_pindex;
1486 * If the pv already exists and we aren't being asked for the
1487 * parent page table page we can just return it. A locked+held pv
1491 pv = pv_alloc(pmap, ptepindex, &isnew);
1492 if (isnew == 0 && pvpp == NULL)
1496 * This is a new PV, we have to resolve its parent page table and
1497 * add an additional wiring to the page if necessary.
1501 * Special case terminal PVs. These are not page table pages so
1502 * no vm_page is allocated (the caller supplied the vm_page). If
1503 * pvpp is non-NULL we are being asked to also removed the pt_pv
1506 * Note that pt_pv's are only returned for user VAs. We assert that
1507 * a pt_pv is not being requested for kernel VAs.
1509 if (ptepindex < pmap_pt_pindex(0)) {
1510 if (ptepindex >= NUPTE_USER)
1511 KKASSERT(pvpp == NULL);
1513 KKASSERT(pvpp != NULL);
1515 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1516 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1518 vm_page_wire_quick(pvp->pv_m);
1527 * Non-terminal PVs allocate a VM page to represent the page table,
1528 * so we have to resolve pvp and calculate ptepindex for the pvp
1529 * and then for the page table entry index in the pvp for
1532 if (ptepindex < pmap_pd_pindex(0)) {
1534 * pv is PT, pvp is PD
1536 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1537 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1538 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1545 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1546 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1548 } else if (ptepindex < pmap_pdp_pindex(0)) {
1550 * pv is PD, pvp is PDP
1552 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1555 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1556 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1558 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1559 KKASSERT(pvpp == NULL);
1562 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1570 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1571 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1572 } else if (ptepindex < pmap_pml4_pindex()) {
1574 * pv is PDP, pvp is the root pml4 table
1576 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1583 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1584 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1587 * pv represents the top-level PML4, there is no parent.
1595 * This code is only reached if isnew is TRUE and this is not a
1596 * terminal PV. We need to allocate a vm_page for the page table
1597 * at this level and enter it into the parent page table.
1599 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1602 m = vm_page_alloc(NULL, pv->pv_pindex,
1603 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1604 VM_ALLOC_INTERRUPT);
1609 vm_page_spin_lock(m);
1610 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1612 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1613 vm_page_spin_unlock(m);
1614 vm_page_unmanage(m); /* m must be spinunlocked */
1616 if ((m->flags & PG_ZERO) == 0) {
1617 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1621 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1624 m->valid = VM_PAGE_BITS_ALL;
1625 vm_page_flag_clear(m, PG_ZERO);
1626 vm_page_wire(m); /* wire for mapping in parent */
1629 * Wire the page into pvp, bump the wire-count for pvp's page table
1630 * page. Bump the resident_count for the pmap. There is no pvp
1631 * for the top level, address the pm_pml4[] array directly.
1633 * If the caller wants the parent we return it, otherwise
1634 * we just put it away.
1636 * No interlock is needed for pte 0 -> non-zero.
1638 * In the situation where *ptep is valid we might have an unmanaged
1639 * page table page shared from another page table which we need to
1640 * unshare before installing our private page table page.
1643 ptep = pv_pte_lookup(pvp, ptepindex);
1646 pmap_inval_info info;
1648 kprintf("pmap_allocpte: restate shared pg table pg\n");
1651 panic("pmap_allocpte: unexpected pte %p/%d",
1652 pvp, (int)ptepindex);
1654 pmap_inval_init(&info);
1655 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1656 pte = pte_load_clear(ptep);
1657 pmap_inval_deinterlock(&info, pmap);
1658 pmap_inval_done(&info);
1659 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
1660 panic("pmap_allocpte: shared pgtable pg bad wirecount");
1662 vm_page_wire_quick(pvp->pv_m);
1664 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1677 * This version of pmap_allocpte() checks for possible segment optimizations
1678 * that would allow page-table sharing. It can be called for terminal
1679 * page or page table page ptepindex's.
1681 * The function is called with page table page ptepindex's for fictitious
1682 * and unmanaged terminal pages. That is, we don't want to allocate a
1683 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
1686 * This function can return a pv and *pvpp associated with the passed in pmap
1687 * OR a pv and *pvpp associated with the shared pmap. In the latter case
1688 * an unmanaged page table page will be entered into the pass in pmap.
1692 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
1693 vm_map_entry_t entry, vm_offset_t va)
1695 struct pmap_inval_info info;
1700 pv_entry_t pte_pv; /* in original or shared pmap */
1701 pv_entry_t pt_pv; /* in original or shared pmap */
1702 pv_entry_t proc_pd_pv; /* in original pmap */
1703 pv_entry_t proc_pt_pv; /* in original pmap */
1704 pv_entry_t xpv; /* PT in shared pmap */
1705 pd_entry_t *pt; /* PT entry in PD of original pmap */
1706 pd_entry_t opte; /* contents of *pt */
1707 pd_entry_t npte; /* contents of *pt */
1711 * Basic tests, require a non-NULL vm_map_entry, require proper
1712 * alignment and type for the vm_map_entry, require that the
1713 * underlying object already be allocated.
1715 * We currently allow any type of object to use this optimization.
1716 * The object itself does NOT have to be sized to a multiple of the
1717 * segment size, but the memory mapping does.
1719 if (entry == NULL ||
1720 pmap_mmu_optimize == 0 || /* not enabled */
1721 ptepindex >= pmap_pd_pindex(0) || /* not terminal */
1722 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
1723 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
1724 entry->object.vm_object == NULL || /* needs VM object */
1725 (entry->offset & SEG_MASK) || /* must be aligned */
1726 (entry->start & SEG_MASK)) {
1727 return(pmap_allocpte(pmap, ptepindex, pvpp));
1731 * Make sure the full segment can be represented.
1733 b = va & ~(vm_offset_t)SEG_MASK;
1734 if (b < entry->start && b + SEG_SIZE > entry->end)
1735 return(pmap_allocpte(pmap, ptepindex, pvpp));
1738 * If the full segment can be represented dive the VM object's
1739 * shared pmap, allocating as required.
1741 object = entry->object.vm_object;
1743 if (entry->protection & VM_PROT_WRITE)
1744 obpmapp = &object->md.pmap_rw;
1746 obpmapp = &object->md.pmap_ro;
1749 * We allocate what appears to be a normal pmap but because portions
1750 * of this pmap are shared with other unrelated pmaps we have to
1751 * set pm_active to point to all cpus.
1753 * XXX Currently using pmap_spin to interlock the update, can't use
1754 * vm_object_hold/drop because the token might already be held
1755 * shared OR exclusive and we don't know.
1757 while ((obpmap = *obpmapp) == NULL) {
1758 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
1759 pmap_pinit_simple(obpmap);
1760 pmap_pinit2(obpmap);
1761 spin_lock(&pmap_spin);
1762 if (*obpmapp != NULL) {
1766 spin_unlock(&pmap_spin);
1767 pmap_release(obpmap);
1768 pmap_puninit(obpmap);
1769 kfree(obpmap, M_OBJPMAP);
1771 obpmap->pm_active = smp_active_mask;
1773 spin_unlock(&pmap_spin);
1778 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
1779 * pte/pt using the shared pmap from the object but also adjust
1780 * the process pmap's page table page as a side effect.
1784 * Resolve the terminal PTE and PT in the shared pmap. This is what
1785 * we will return. This is true if ptepindex represents a terminal
1786 * page, otherwise pte_pv is actually the PT and pt_pv is actually
1790 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
1791 if (ptepindex >= pmap_pt_pindex(0))
1797 * Resolve the PD in the process pmap so we can properly share the
1798 * page table page. Lock order is bottom-up (leaf first)!
1800 * NOTE: proc_pt_pv can be NULL.
1802 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
1803 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
1806 * xpv is the page table page pv from the shared object
1807 * (for convenience).
1809 * Calculate the pte value for the PT to load into the process PD.
1810 * If we have to change it we must properly dispose of the previous
1813 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1814 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
1815 (PG_U | PG_RW | PG_V | PG_A | PG_M);
1818 * Dispose of previous entry if it was local to the process pmap.
1819 * (This should zero-out *pt)
1822 pmap_release_pv(proc_pt_pv);
1825 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1829 * Handle remaining cases.
1833 vm_page_wire_quick(xpv->pv_m);
1834 vm_page_wire_quick(proc_pd_pv->pv_m);
1835 atomic_add_long(&pmap->pm_stats.resident_count, 1);
1836 } else if (*pt != npte) {
1837 pmap_inval_init(&info);
1838 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1840 opte = pte_load_clear(pt);
1841 KKASSERT(opte && opte != npte);
1844 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
1847 * Clean up opte, bump the wire_count for the process
1848 * PD page representing the new entry if it was
1851 * If the entry was not previously empty and we have
1852 * a PT in the proc pmap then opte must match that
1853 * pt. The proc pt must be retired (this is done
1854 * later on in this procedure).
1856 * NOTE: replacing valid pte, wire_count on proc_pd_pv
1859 KKASSERT(opte & PG_V);
1860 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
1861 if (vm_page_unwire_quick(m)) {
1862 panic("pmap_allocpte_seg: "
1863 "bad wire count %p",
1867 pmap_inval_deinterlock(&info, pmap);
1868 pmap_inval_done(&info);
1872 * The existing process page table was replaced and must be destroyed
1886 * Release any resources held by the given physical map.
1888 * Called when a pmap initialized by pmap_pinit is being released. Should
1889 * only be called if the map contains no valid mappings.
1891 * Caller must hold pmap->pm_token
1893 struct pmap_release_info {
1898 static int pmap_release_callback(pv_entry_t pv, void *data);
1901 pmap_release(struct pmap *pmap)
1903 struct pmap_release_info info;
1905 KASSERT(pmap->pm_active == 0,
1906 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1908 spin_lock(&pmap_spin);
1909 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1910 spin_unlock(&pmap_spin);
1913 * Pull pv's off the RB tree in order from low to high and release
1919 spin_lock(&pmap->pm_spin);
1920 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
1921 pmap_release_callback, &info);
1922 spin_unlock(&pmap->pm_spin);
1923 } while (info.retry);
1927 * One resident page (the pml4 page) should remain.
1928 * No wired pages should remain.
1930 KKASSERT(pmap->pm_stats.resident_count ==
1931 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
1933 KKASSERT(pmap->pm_stats.wired_count == 0);
1937 pmap_release_callback(pv_entry_t pv, void *data)
1939 struct pmap_release_info *info = data;
1940 pmap_t pmap = info->pmap;
1943 if (pv_hold_try(pv)) {
1944 spin_unlock(&pmap->pm_spin);
1946 spin_unlock(&pmap->pm_spin);
1948 if (pv->pv_pmap != pmap) {
1950 spin_lock(&pmap->pm_spin);
1955 r = pmap_release_pv(pv);
1956 spin_lock(&pmap->pm_spin);
1961 * Called with held (i.e. also locked) pv. This function will dispose of
1962 * the lock along with the pv.
1965 pmap_release_pv(pv_entry_t pv)
1970 * The pmap is currently not spinlocked, pv is held+locked.
1971 * Remove the pv's page from its parent's page table. The
1972 * parent's page table page's wire_count will be decremented.
1974 pmap_remove_pv_pte(pv, NULL, NULL);
1977 * Terminal pvs are unhooked from their vm_pages. Because
1978 * terminal pages aren't page table pages they aren't wired
1979 * by us, so we have to be sure not to unwire them either.
1981 if (pv->pv_pindex < pmap_pt_pindex(0)) {
1982 pmap_remove_pv_page(pv);
1987 * We leave the top-level page table page cached, wired, and
1988 * mapped in the pmap until the dtor function (pmap_puninit())
1991 * Since we are leaving the top-level pv intact we need
1992 * to break out of what would otherwise be an infinite loop.
1994 if (pv->pv_pindex == pmap_pml4_pindex()) {
2000 * For page table pages (other than the top-level page),
2001 * remove and free the vm_page. The representitive mapping
2002 * removed above by pmap_remove_pv_pte() did not undo the
2003 * last wire_count so we have to do that as well.
2005 p = pmap_remove_pv_page(pv);
2006 vm_page_busy_wait(p, FALSE, "pmaprl");
2007 if (p->wire_count != 1) {
2008 kprintf("p->wire_count was %016lx %d\n",
2009 pv->pv_pindex, p->wire_count);
2011 KKASSERT(p->wire_count == 1);
2012 KKASSERT(p->flags & PG_UNMANAGED);
2014 vm_page_unwire(p, 0);
2015 KKASSERT(p->wire_count == 0);
2018 * Theoretically this page, if not the pml4 page, should contain
2019 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2029 * This function will remove the pte associated with a pv from its parent.
2030 * Terminal pv's are supported. The removal will be interlocked if info
2031 * is non-NULL. The caller must dispose of pv instead of just unlocking
2034 * The wire count will be dropped on the parent page table. The wire
2035 * count on the page being removed (pv->pv_m) from the parent page table
2036 * is NOT touched. Note that terminal pages will not have any additional
2037 * wire counts while page table pages will have at least one representing
2038 * the mapping, plus others representing sub-mappings.
2040 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2041 * pages and user page table and terminal pages.
2043 * The pv must be locked.
2045 * XXX must lock parent pv's if they exist to remove pte XXX
2049 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2051 vm_pindex_t ptepindex = pv->pv_pindex;
2052 pmap_t pmap = pv->pv_pmap;
2058 if (ptepindex == pmap_pml4_pindex()) {
2060 * We are the top level pml4 table, there is no parent.
2062 p = pmap->pm_pmlpv->pv_m;
2063 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2065 * Remove a PDP page from the pml4e. This can only occur
2066 * with user page tables. We do not have to lock the
2067 * pml4 PV so just ignore pvp.
2069 vm_pindex_t pml4_pindex;
2070 vm_pindex_t pdp_index;
2073 pdp_index = ptepindex - pmap_pdp_pindex(0);
2075 pml4_pindex = pmap_pml4_pindex();
2076 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2080 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2081 KKASSERT((*pdp & PG_V) != 0);
2082 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2084 KKASSERT(info == NULL);
2085 } else if (ptepindex >= pmap_pd_pindex(0)) {
2087 * Remove a PD page from the pdp
2089 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2090 * of a simple pmap because it stops at
2093 vm_pindex_t pdp_pindex;
2094 vm_pindex_t pd_index;
2097 pd_index = ptepindex - pmap_pd_pindex(0);
2100 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2101 (pd_index >> NPML4EPGSHIFT);
2102 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2107 pd = pv_pte_lookup(pvp, pd_index &
2108 ((1ul << NPDPEPGSHIFT) - 1));
2109 KKASSERT((*pd & PG_V) != 0);
2110 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2113 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2114 p = pv->pv_m; /* degenerate test later */
2116 KKASSERT(info == NULL);
2117 } else if (ptepindex >= pmap_pt_pindex(0)) {
2119 * Remove a PT page from the pd
2121 vm_pindex_t pd_pindex;
2122 vm_pindex_t pt_index;
2125 pt_index = ptepindex - pmap_pt_pindex(0);
2128 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2129 (pt_index >> NPDPEPGSHIFT);
2130 pvp = pv_get(pv->pv_pmap, pd_pindex);
2134 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2135 KKASSERT((*pt & PG_V) != 0);
2136 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2138 KKASSERT(info == NULL);
2141 * Remove a PTE from the PT page
2143 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2144 * pv is a pte_pv so we can safely lock pt_pv.
2146 vm_pindex_t pt_pindex;
2151 pt_pindex = ptepindex >> NPTEPGSHIFT;
2152 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2154 if (ptepindex >= NUPTE_USER) {
2155 ptep = vtopte(ptepindex << PAGE_SHIFT);
2156 KKASSERT(pvp == NULL);
2159 pt_pindex = NUPTE_TOTAL +
2160 (ptepindex >> NPDPEPGSHIFT);
2161 pvp = pv_get(pv->pv_pmap, pt_pindex);
2165 ptep = pv_pte_lookup(pvp, ptepindex &
2166 ((1ul << NPDPEPGSHIFT) - 1));
2170 pmap_inval_interlock(info, pmap, va);
2171 pte = pte_load_clear(ptep);
2173 pmap_inval_deinterlock(info, pmap);
2175 cpu_invlpg((void *)va);
2178 * Now update the vm_page_t
2180 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
2181 kprintf("remove_pte badpte %016lx %016lx %d\n",
2183 pv->pv_pindex < pmap_pt_pindex(0));
2185 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2186 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2189 if (pmap_track_modified(ptepindex))
2193 vm_page_flag_set(p, PG_REFERENCED);
2196 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2198 cpu_invlpg((void *)va);
2202 * Unwire the parent page table page. The wire_count cannot go below
2203 * 1 here because the parent page table page is itself still mapped.
2205 * XXX remove the assertions later.
2207 KKASSERT(pv->pv_m == p);
2208 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2209 panic("pmap_remove_pv_pte: Insufficient wire_count");
2217 pmap_remove_pv_page(pv_entry_t pv)
2223 vm_page_spin_lock(m);
2225 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2228 atomic_add_int(&m->object->agg_pv_list_count, -1);
2230 if (TAILQ_EMPTY(&m->md.pv_list))
2231 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2232 vm_page_spin_unlock(m);
2237 * Grow the number of kernel page table entries, if needed.
2239 * This routine is always called to validate any address space
2240 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2241 * space below KERNBASE.
2244 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2247 vm_offset_t ptppaddr;
2249 pd_entry_t *pt, newpt;
2251 int update_kernel_vm_end;
2254 * bootstrap kernel_vm_end on first real VM use
2256 if (kernel_vm_end == 0) {
2257 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2259 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
2260 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2261 ~(PAGE_SIZE * NPTEPG - 1);
2263 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2264 kernel_vm_end = kernel_map.max_offset;
2271 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2272 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2273 * do not want to force-fill 128G worth of page tables.
2275 if (kstart < KERNBASE) {
2276 if (kstart > kernel_vm_end)
2277 kstart = kernel_vm_end;
2278 KKASSERT(kend <= KERNBASE);
2279 update_kernel_vm_end = 1;
2281 update_kernel_vm_end = 0;
2284 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2285 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2287 if (kend - 1 >= kernel_map.max_offset)
2288 kend = kernel_map.max_offset;
2290 while (kstart < kend) {
2291 pt = pmap_pt(&kernel_pmap, kstart);
2293 /* We need a new PDP entry */
2294 nkpg = vm_page_alloc(NULL, nkpt,
2297 VM_ALLOC_INTERRUPT);
2299 panic("pmap_growkernel: no memory to grow "
2302 paddr = VM_PAGE_TO_PHYS(nkpg);
2303 if ((nkpg->flags & PG_ZERO) == 0)
2304 pmap_zero_page(paddr);
2305 vm_page_flag_clear(nkpg, PG_ZERO);
2306 newpd = (pdp_entry_t)
2307 (paddr | PG_V | PG_RW | PG_A | PG_M);
2308 *pmap_pd(&kernel_pmap, kstart) = newpd;
2310 continue; /* try again */
2312 if ((*pt & PG_V) != 0) {
2313 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2314 ~(PAGE_SIZE * NPTEPG - 1);
2315 if (kstart - 1 >= kernel_map.max_offset) {
2316 kstart = kernel_map.max_offset;
2323 * This index is bogus, but out of the way
2325 nkpg = vm_page_alloc(NULL, nkpt,
2328 VM_ALLOC_INTERRUPT);
2330 panic("pmap_growkernel: no memory to grow kernel");
2333 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2334 pmap_zero_page(ptppaddr);
2335 vm_page_flag_clear(nkpg, PG_ZERO);
2336 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2337 *pmap_pt(&kernel_pmap, kstart) = newpt;
2340 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2341 ~(PAGE_SIZE * NPTEPG - 1);
2343 if (kstart - 1 >= kernel_map.max_offset) {
2344 kstart = kernel_map.max_offset;
2350 * Only update kernel_vm_end for areas below KERNBASE.
2352 if (update_kernel_vm_end && kernel_vm_end < kstart)
2353 kernel_vm_end = kstart;
2357 * Add a reference to the specified pmap.
2360 pmap_reference(pmap_t pmap)
2363 lwkt_gettoken(&pmap->pm_token);
2365 lwkt_reltoken(&pmap->pm_token);
2370 pmap_drop(pmap_t pmap)
2373 lwkt_gettoken(&pmap->pm_token);
2375 lwkt_reltoken(&pmap->pm_token);
2379 /***************************************************
2380 * page management routines.
2381 ***************************************************/
2384 * Hold a pv without locking it
2387 pv_hold(pv_entry_t pv)
2391 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2395 count = pv->pv_hold;
2397 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2404 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2405 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2408 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2409 * pv list via its page) must be held by the caller.
2412 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2416 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2419 pv->pv_line = lineno;
2425 count = pv->pv_hold;
2427 if ((count & PV_HOLD_LOCKED) == 0) {
2428 if (atomic_cmpset_int(&pv->pv_hold, count,
2429 (count + 1) | PV_HOLD_LOCKED)) {
2432 pv->pv_line = lineno;
2437 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2445 * Drop a previously held pv_entry which could not be locked, allowing its
2448 * Must not be called with a spinlock held as we might zfree() the pv if it
2449 * is no longer associated with a pmap and this was the last hold count.
2452 pv_drop(pv_entry_t pv)
2456 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2457 if (pv->pv_pmap == NULL)
2463 count = pv->pv_hold;
2465 KKASSERT((count & PV_HOLD_MASK) > 0);
2466 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2467 (PV_HOLD_LOCKED | 1));
2468 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2469 if (count == 1 && pv->pv_pmap == NULL)
2478 * Find or allocate the requested PV entry, returning a locked pv
2482 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2485 pv_entry_t pnew = NULL;
2487 spin_lock(&pmap->pm_spin);
2489 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2490 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2495 spin_unlock(&pmap->pm_spin);
2496 pnew = zalloc(pvzone);
2497 spin_lock(&pmap->pm_spin);
2500 pnew->pv_pmap = pmap;
2501 pnew->pv_pindex = pindex;
2502 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2504 pnew->pv_func = func;
2505 pnew->pv_line = lineno;
2507 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2508 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2509 spin_unlock(&pmap->pm_spin);
2514 spin_unlock(&pmap->pm_spin);
2515 zfree(pvzone, pnew);
2517 spin_lock(&pmap->pm_spin);
2520 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2521 spin_unlock(&pmap->pm_spin);
2525 spin_unlock(&pmap->pm_spin);
2526 _pv_lock(pv PMAP_DEBUG_COPY);
2527 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2532 spin_lock(&pmap->pm_spin);
2539 * Find the requested PV entry, returning a locked+held pv or NULL
2543 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2547 spin_lock(&pmap->pm_spin);
2552 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2553 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2557 spin_unlock(&pmap->pm_spin);
2560 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2561 pv_cache(pv, pindex);
2562 spin_unlock(&pmap->pm_spin);
2565 spin_unlock(&pmap->pm_spin);
2566 _pv_lock(pv PMAP_DEBUG_COPY);
2567 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2570 spin_lock(&pmap->pm_spin);
2575 * Lookup, hold, and attempt to lock (pmap,pindex).
2577 * If the entry does not exist NULL is returned and *errorp is set to 0
2579 * If the entry exists and could be successfully locked it is returned and
2580 * errorp is set to 0.
2582 * If the entry exists but could NOT be successfully locked it is returned
2583 * held and *errorp is set to 1.
2587 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2591 spin_lock(&pmap->pm_spin);
2592 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2593 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2595 spin_unlock(&pmap->pm_spin);
2599 if (pv_hold_try(pv)) {
2600 pv_cache(pv, pindex);
2601 spin_unlock(&pmap->pm_spin);
2603 return(pv); /* lock succeeded */
2605 spin_unlock(&pmap->pm_spin);
2607 return (pv); /* lock failed */
2611 * Find the requested PV entry, returning a held pv or NULL
2615 pv_find(pmap_t pmap, vm_pindex_t pindex)
2619 spin_lock(&pmap->pm_spin);
2621 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2622 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2624 spin_unlock(&pmap->pm_spin);
2628 pv_cache(pv, pindex);
2629 spin_unlock(&pmap->pm_spin);
2634 * Lock a held pv, keeping the hold count
2638 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2643 count = pv->pv_hold;
2645 if ((count & PV_HOLD_LOCKED) == 0) {
2646 if (atomic_cmpset_int(&pv->pv_hold, count,
2647 count | PV_HOLD_LOCKED)) {
2650 pv->pv_line = lineno;
2656 tsleep_interlock(pv, 0);
2657 if (atomic_cmpset_int(&pv->pv_hold, count,
2658 count | PV_HOLD_WAITING)) {
2660 kprintf("pv waiting on %s:%d\n",
2661 pv->pv_func, pv->pv_line);
2663 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2670 * Unlock a held and locked pv, keeping the hold count.
2674 pv_unlock(pv_entry_t pv)
2678 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2682 count = pv->pv_hold;
2684 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2685 (PV_HOLD_LOCKED | 1));
2686 if (atomic_cmpset_int(&pv->pv_hold, count,
2688 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2689 if (count & PV_HOLD_WAITING)
2697 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2698 * and the hold count drops to zero we will free it.
2700 * Caller should not hold any spin locks. We are protected from hold races
2701 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2702 * lock held. A pv cannot be located otherwise.
2706 pv_put(pv_entry_t pv)
2708 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2709 if (pv->pv_pmap == NULL)
2718 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2719 * pmap. Any pte operations must have already been completed.
2723 pv_free(pv_entry_t pv)
2727 KKASSERT(pv->pv_m == NULL);
2728 if ((pmap = pv->pv_pmap) != NULL) {
2729 spin_lock(&pmap->pm_spin);
2730 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2731 if (pmap->pm_pvhint == pv)
2732 pmap->pm_pvhint = NULL;
2733 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2736 spin_unlock(&pmap->pm_spin);
2742 * This routine is very drastic, but can save the system
2750 static int warningdone=0;
2752 if (pmap_pagedaemon_waken == 0)
2754 pmap_pagedaemon_waken = 0;
2755 if (warningdone < 5) {
2756 kprintf("pmap_collect: collecting pv entries -- "
2757 "suggest increasing PMAP_SHPGPERPROC\n");
2761 for (i = 0; i < vm_page_array_size; i++) {
2762 m = &vm_page_array[i];
2763 if (m->wire_count || m->hold_count)
2765 if (vm_page_busy_try(m, TRUE) == 0) {
2766 if (m->wire_count == 0 && m->hold_count == 0) {
2775 * Scan the pmap for active page table entries and issue a callback.
2776 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
2777 * its parent page table.
2779 * pte_pv will be NULL if the page or page table is unmanaged.
2780 * pt_pv will point to the page table page containing the pte for the page.
2782 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
2783 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
2784 * process pmap's PD and page to the callback function. This can be
2785 * confusing because the pt_pv is really a pd_pv, and the target page
2786 * table page is simply aliased by the pmap and not owned by it.
2788 * It is assumed that the start and end are properly rounded to the page size.
2790 * It is assumed that PD pages and above are managed and thus in the RB tree,
2791 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
2793 struct pmap_scan_info {
2797 vm_pindex_t sva_pd_pindex;
2798 vm_pindex_t eva_pd_pindex;
2799 void (*func)(pmap_t, struct pmap_inval_info *,
2800 pv_entry_t, pv_entry_t, int, vm_offset_t,
2801 pt_entry_t *, void *);
2803 struct pmap_inval_info inval;
2806 static int pmap_scan_cmp(pv_entry_t pv, void *data);
2807 static int pmap_scan_callback(pv_entry_t pv, void *data);
2810 pmap_scan(struct pmap_scan_info *info)
2812 struct pmap *pmap = info->pmap;
2813 pv_entry_t pd_pv; /* A page directory PV */
2814 pv_entry_t pt_pv; /* A page table PV */
2815 pv_entry_t pte_pv; /* A page table entry PV */
2817 struct pv_entry dummy_pv;
2823 * Hold the token for stability; if the pmap is empty we have nothing
2826 lwkt_gettoken(&pmap->pm_token);
2828 if (pmap->pm_stats.resident_count == 0) {
2829 lwkt_reltoken(&pmap->pm_token);
2834 pmap_inval_init(&info->inval);
2837 * Special handling for scanning one page, which is a very common
2838 * operation (it is?).
2840 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
2842 if (info->sva + PAGE_SIZE == info->eva) {
2843 if (info->sva >= VM_MAX_USER_ADDRESS) {
2845 * Kernel mappings do not track wire counts on
2846 * page table pages and only maintain pd_pv and
2847 * pte_pv levels so pmap_scan() works.
2850 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
2851 ptep = vtopte(info->sva);
2854 * User pages which are unmanaged will not have a
2855 * pte_pv. User page table pages which are unmanaged
2856 * (shared from elsewhere) will also not have a pt_pv.
2857 * The func() callback will pass both pte_pv and pt_pv
2858 * as NULL in that case.
2860 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
2861 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
2862 if (pt_pv == NULL) {
2863 KKASSERT(pte_pv == NULL);
2864 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
2866 ptep = pv_pte_lookup(pd_pv,
2867 pmap_pt_index(info->sva));
2869 info->func(pmap, &info->inval,
2878 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
2882 * Unlike the pv_find() case below we actually
2883 * acquired a locked pv in this case so any
2884 * race should have been resolved. It is expected
2887 KKASSERT(pte_pv == NULL);
2888 } else if (pte_pv) {
2889 KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED|
2891 ("bad *ptep %016lx sva %016lx pte_pv %p",
2892 *ptep, info->sva, pte_pv));
2893 info->func(pmap, &info->inval, pte_pv, pt_pv, 0,
2894 info->sva, ptep, info->arg);
2896 KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V,
2897 ("bad *ptep %016lx sva %016lx pte_pv NULL",
2899 info->func(pmap, &info->inval, NULL, pt_pv, 0,
2900 info->sva, ptep, info->arg);
2905 pmap_inval_done(&info->inval);
2906 lwkt_reltoken(&pmap->pm_token);
2911 * Nominal scan case, RB_SCAN() for PD pages and iterate from
2914 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
2915 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
2917 if (info->sva >= VM_MAX_USER_ADDRESS) {
2919 * The kernel does not currently maintain any pv_entry's for
2920 * higher-level page tables.
2922 bzero(&dummy_pv, sizeof(dummy_pv));
2923 dummy_pv.pv_pindex = info->sva_pd_pindex;
2924 spin_lock(&pmap->pm_spin);
2925 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
2926 pmap_scan_callback(&dummy_pv, info);
2927 ++dummy_pv.pv_pindex;
2929 spin_unlock(&pmap->pm_spin);
2932 * User page tables maintain local PML4, PDP, and PD
2933 * pv_entry's at the very least. PT pv's might be
2934 * unmanaged and thus not exist. PTE pv's might be
2935 * unmanaged and thus not exist.
2937 spin_lock(&pmap->pm_spin);
2938 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
2939 pmap_scan_cmp, pmap_scan_callback, info);
2940 spin_unlock(&pmap->pm_spin);
2942 pmap_inval_done(&info->inval);
2943 lwkt_reltoken(&pmap->pm_token);
2947 * WARNING! pmap->pm_spin held
2950 pmap_scan_cmp(pv_entry_t pv, void *data)
2952 struct pmap_scan_info *info = data;
2953 if (pv->pv_pindex < info->sva_pd_pindex)
2955 if (pv->pv_pindex >= info->eva_pd_pindex)
2961 * WARNING! pmap->pm_spin held
2964 pmap_scan_callback(pv_entry_t pv, void *data)
2966 struct pmap_scan_info *info = data;
2967 struct pmap *pmap = info->pmap;
2968 pv_entry_t pd_pv; /* A page directory PV */
2969 pv_entry_t pt_pv; /* A page table PV */
2970 pv_entry_t pte_pv; /* A page table entry PV */
2974 vm_offset_t va_next;
2975 vm_pindex_t pd_pindex;
2979 * Pull the PD pindex from the pv before releasing the spinlock.
2981 * WARNING: pv is faked for kernel pmap scans.
2983 pd_pindex = pv->pv_pindex;
2984 spin_unlock(&pmap->pm_spin);
2985 pv = NULL; /* invalid after spinlock unlocked */
2988 * Calculate the page range within the PD. SIMPLE pmaps are
2989 * direct-mapped for the entire 2^64 address space. Normal pmaps
2990 * reflect the user and kernel address space which requires
2991 * cannonicalization w/regards to converting pd_pindex's back
2994 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
2995 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
2996 (sva & PML4_SIGNMASK)) {
2997 sva |= PML4_SIGNMASK;
2999 eva = sva + NBPDP; /* can overflow */
3000 if (sva < info->sva)
3002 if (eva < info->sva || eva > info->eva)
3006 * NOTE: kernel mappings do not track page table pages, only
3009 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3010 * However, for the scan to be efficient we try to
3011 * cache items top-down.
3016 for (; sva < eva; sva = va_next) {
3018 if (sva >= VM_MAX_USER_ADDRESS) {
3027 * PD cache (degenerate case if we skip). It is possible
3028 * for the PD to not exist due to races. This is ok.
3030 if (pd_pv == NULL) {
3031 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3032 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3034 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3036 if (pd_pv == NULL) {
3037 va_next = (sva + NBPDP) & ~PDPMASK;
3046 if (pt_pv == NULL) {
3051 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3052 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3058 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3062 * If pt_pv is NULL we either have an shared page table
3063 * page and must issue a callback specific to that case,
3064 * or there is no page table page.
3066 * Either way we can skip the page table page.
3068 if (pt_pv == NULL) {
3070 * Possible unmanaged (shared from another pmap)
3074 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3075 KKASSERT(pd_pv != NULL);
3076 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3078 info->func(pmap, &info->inval, NULL, pd_pv, 1,
3079 sva, ptep, info->arg);
3083 * Done, move to next page table page.
3085 va_next = (sva + NBPDR) & ~PDRMASK;
3092 * From this point in the loop testing pt_pv for non-NULL
3093 * means we are in UVM, else if it is NULL we are in KVM.
3095 * Limit our scan to either the end of the va represented
3096 * by the current page table page, or to the end of the
3097 * range being removed.
3100 va_next = (sva + NBPDR) & ~PDRMASK;
3107 * Scan the page table for pages. Some pages may not be
3108 * managed (might not have a pv_entry).
3110 * There is no page table management for kernel pages so
3111 * pt_pv will be NULL in that case, but otherwise pt_pv
3112 * is non-NULL, locked, and referenced.
3116 * At this point a non-NULL pt_pv means a UVA, and a NULL
3117 * pt_pv means a KVA.
3120 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3124 while (sva < va_next) {
3126 * Acquire the related pte_pv, if any. If *ptep == 0
3127 * the related pte_pv should not exist, but if *ptep
3128 * is not zero the pte_pv may or may not exist (e.g.
3129 * will not exist for an unmanaged page).
3131 * However a multitude of races are possible here.
3133 * In addition, the (pt_pv, pte_pv) lock order is
3134 * backwards, so we have to be careful in aquiring
3135 * a properly locked pte_pv.
3139 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3146 pv_put(pt_pv); /* must be non-NULL */
3148 pv_lock(pte_pv); /* safe to block now */
3151 pt_pv = pv_get(pmap,
3152 pmap_pt_pindex(sva));
3154 * pt_pv reloaded, need new ptep
3156 KKASSERT(pt_pv != NULL);
3157 ptep = pv_pte_lookup(pt_pv,
3158 pmap_pte_index(sva));
3162 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3166 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3170 kprintf("Unexpected non-NULL pte_pv "
3171 "%p pt_pv %p *ptep = %016lx\n",
3172 pte_pv, pt_pv, *ptep);
3173 panic("Unexpected non-NULL pte_pv");
3181 * Ready for the callback. The locked pte_pv (if any)
3182 * is consumed by the callback. pte_pv will exist if
3183 * the page is managed, and will not exist if it
3187 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3189 ("bad *ptep %016lx sva %016lx "
3191 *ptep, sva, pte_pv));
3192 info->func(pmap, &info->inval, pte_pv, pt_pv, 0,
3193 sva, ptep, info->arg);
3195 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3197 ("bad *ptep %016lx sva %016lx "
3200 info->func(pmap, &info->inval, NULL, pt_pv, 0,
3201 sva, ptep, info->arg);
3218 * Relock before returning.
3220 spin_lock(&pmap->pm_spin);
3225 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3227 struct pmap_scan_info info;
3232 info.func = pmap_remove_callback;
3238 pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
3239 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3240 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3246 * This will also drop pt_pv's wire_count. Note that
3247 * terminal pages are not wired based on mmu presence.
3249 pmap_remove_pv_pte(pte_pv, pt_pv, info);
3250 pmap_remove_pv_page(pte_pv);
3252 } else if (sharept == 0) {
3256 * pt_pv's wire_count is still bumped by unmanaged pages
3257 * so we must decrement it manually.
3259 pmap_inval_interlock(info, pmap, va);
3260 pte = pte_load_clear(ptep);
3261 pmap_inval_deinterlock(info, pmap);
3263 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3264 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3265 if (vm_page_unwire_quick(pt_pv->pv_m))
3266 panic("pmap_remove: insufficient wirecount");
3269 * Unmanaged page table, pt_pv is actually the pd_pv
3270 * for our pmap (not the share object pmap).
3272 * We have to unwire the target page table page and we
3273 * have to unwire our page directory page.
3275 pmap_inval_interlock(info, pmap, va);
3276 pte = pte_load_clear(ptep);
3277 pmap_inval_deinterlock(info, pmap);
3278 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3279 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3280 panic("pmap_remove: shared pgtable1 bad wirecount");
3281 if (vm_page_unwire_quick(pt_pv->pv_m))
3282 panic("pmap_remove: shared pgtable2 bad wirecount");
3287 * Removes this physical page from all physical maps in which it resides.
3288 * Reflects back modify bits to the pager.
3290 * This routine may not be called from an interrupt.
3294 pmap_remove_all(vm_page_t m)
3296 struct pmap_inval_info info;
3299 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3302 pmap_inval_init(&info);
3303 vm_page_spin_lock(m);
3304 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3305 KKASSERT(pv->pv_m == m);
3306 if (pv_hold_try(pv)) {
3307 vm_page_spin_unlock(m);
3309 vm_page_spin_unlock(m);
3311 if (pv->pv_m != m) {
3313 vm_page_spin_lock(m);
3318 * Holding no spinlocks, pv is locked.
3320 pmap_remove_pv_pte(pv, NULL, &info);
3321 pmap_remove_pv_page(pv);
3323 vm_page_spin_lock(m);
3325 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3326 vm_page_spin_unlock(m);
3327 pmap_inval_done(&info);
3331 * Set the physical protection on the specified range of this map
3332 * as requested. This function is typically only used for debug watchpoints
3335 * This function may not be called from an interrupt if the map is
3336 * not the kernel_pmap.
3338 * NOTE! For shared page table pages we just unmap the page.
3341 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3343 struct pmap_scan_info info;
3344 /* JG review for NX */
3348 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3349 pmap_remove(pmap, sva, eva);
3352 if (prot & VM_PROT_WRITE)
3357 info.func = pmap_protect_callback;
3364 pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
3365 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3366 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3376 pmap_inval_interlock(info, pmap, va);
3383 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3384 KKASSERT(m == pte_pv->pv_m);
3385 vm_page_flag_set(m, PG_REFERENCED);
3389 if (pmap_track_modified(pte_pv->pv_pindex)) {
3391 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3396 } else if (sharept) {
3398 * Unmanaged page table, pt_pv is actually the pd_pv
3399 * for our pmap (not the share object pmap).
3401 * When asked to protect something in a shared page table
3402 * page we just unmap the page table page. We have to
3403 * invalidate the tlb in this situation.
3405 pte = pte_load_clear(ptep);
3406 pmap_inval_invltlb(info);
3407 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3408 panic("pmap_protect: pgtable1 pg bad wirecount");
3409 if (vm_page_unwire_quick(pt_pv->pv_m))
3410 panic("pmap_protect: pgtable2 pg bad wirecount");
3413 /* else unmanaged page, adjust bits, no wire changes */
3417 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
3421 pmap_inval_deinterlock(info, pmap);
3427 * Insert the vm_page (m) at the virtual address (va), replacing any prior
3428 * mapping at that address. Set protection and wiring as requested.
3430 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
3431 * possible. If it is we enter the page into the appropriate shared pmap
3432 * hanging off the related VM object instead of the passed pmap, then we
3433 * share the page table page from the VM object's pmap into the current pmap.
3435 * NOTE: This routine MUST insert the page into the pmap now, it cannot
3439 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3440 boolean_t wired, vm_map_entry_t entry __unused)
3442 pmap_inval_info info;
3443 pv_entry_t pt_pv; /* page table */
3444 pv_entry_t pte_pv; /* page table entry */
3447 pt_entry_t origpte, newpte;
3452 va = trunc_page(va);
3453 #ifdef PMAP_DIAGNOSTIC
3455 panic("pmap_enter: toobig");
3456 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
3457 panic("pmap_enter: invalid to pmap_enter page table "
3458 "pages (va: 0x%lx)", va);
3460 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
3461 kprintf("Warning: pmap_enter called on UVA with "
3464 db_print_backtrace();
3467 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
3468 kprintf("Warning: pmap_enter called on KVA without"
3471 db_print_backtrace();
3476 * Get locked PV entries for our new page table entry (pte_pv)
3477 * and for its parent page table (pt_pv). We need the parent
3478 * so we can resolve the location of the ptep.
3480 * Only hardware MMU actions can modify the ptep out from
3483 * if (m) is fictitious or unmanaged we do not create a managing
3484 * pte_pv for it. Any pre-existing page's management state must
3485 * match (avoiding code complexity).
3487 * If the pmap is still being initialized we assume existing
3490 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3491 * pmap_allocpte() checks the
3493 if (pmap_initialized == FALSE) {
3497 } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) { /* XXX */
3499 if (va >= VM_MAX_USER_ADDRESS) {
3503 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
3505 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3507 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0);
3509 if (va >= VM_MAX_USER_ADDRESS) {
3511 * Kernel map, pv_entry-tracked.
3514 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3520 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
3522 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3524 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED));
3527 pa = VM_PAGE_TO_PHYS(m);
3529 opa = origpte & PG_FRAME;
3531 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V | PG_A);
3534 if (va < VM_MAX_USER_ADDRESS)
3537 newpte |= PG_MANAGED;
3538 if (pmap == &kernel_pmap)
3542 * It is possible for multiple faults to occur in threaded
3543 * environments, the existing pte might be correct.
3545 if (((origpte ^ newpte) & ~(pt_entry_t)(PG_M|PG_A)) == 0)
3548 if ((prot & VM_PROT_NOSYNC) == 0)
3549 pmap_inval_init(&info);
3552 * Ok, either the address changed or the protection or wiring
3555 * Clear the current entry, interlocking the removal. For managed
3556 * pte's this will also flush the modified state to the vm_page.
3557 * Atomic ops are mandatory in order to ensure that PG_M events are
3558 * not lost during any transition.
3563 * pmap_remove_pv_pte() unwires pt_pv and assumes
3564 * we will free pte_pv, but since we are reusing
3565 * pte_pv we want to retain the wire count.
3567 * pt_pv won't exist for a kernel page (managed or
3571 vm_page_wire_quick(pt_pv->pv_m);
3572 if (prot & VM_PROT_NOSYNC)
3573 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3575 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3577 pmap_remove_pv_page(pte_pv);
3578 } else if (prot & VM_PROT_NOSYNC) {
3580 * Unmanaged page, NOSYNC (no mmu sync) requested.
3582 * Leave wire count on PT page intact.
3584 (void)pte_load_clear(ptep);
3585 cpu_invlpg((void *)va);
3586 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3589 * Unmanaged page, normal enter.
3591 * Leave wire count on PT page intact.
3593 pmap_inval_interlock(&info, pmap, va);
3594 (void)pte_load_clear(ptep);
3595 pmap_inval_deinterlock(&info, pmap);
3596 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3598 KKASSERT(*ptep == 0);
3603 * Enter on the PV list if part of our managed memory.
3604 * Wiring of the PT page is already handled.
3606 KKASSERT(pte_pv->pv_m == NULL);
3607 vm_page_spin_lock(m);
3609 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3612 atomic_add_int(&m->object->agg_pv_list_count, 1);
3614 vm_page_flag_set(m, PG_MAPPED);
3615 vm_page_spin_unlock(m);
3616 } else if (pt_pv && opa == 0) {
3618 * We have to adjust the wire count on the PT page ourselves
3619 * for unmanaged entries. If opa was non-zero we retained
3620 * the existing wire count from the removal.
3622 vm_page_wire_quick(pt_pv->pv_m);
3626 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
3628 * User VMAs do not because those will be zero->non-zero, so no
3629 * stale entries to worry about at this point.
3631 * For KVM there appear to still be issues. Theoretically we
3632 * should be able to scrap the interlocks entirely but we
3635 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3636 pmap_inval_interlock(&info, pmap, va);
3641 *(volatile pt_entry_t *)ptep = newpte;
3643 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3644 pmap_inval_deinterlock(&info, pmap);
3645 else if (pt_pv == NULL)
3646 cpu_invlpg((void *)va);
3650 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
3653 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3657 vm_page_flag_set(m, PG_WRITEABLE);
3660 * Unmanaged pages need manual resident_count tracking.
3662 if (pte_pv == NULL && pt_pv)
3663 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
3668 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
3669 pmap_inval_done(&info);
3671 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
3674 * Cleanup the pv entry, allowing other accessors.
3683 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
3684 * This code also assumes that the pmap has no pre-existing entry for this
3687 * This code currently may only be used on user pmaps, not kernel_pmap.
3690 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
3692 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
3696 * Make a temporary mapping for a physical address. This is only intended
3697 * to be used for panic dumps.
3699 * The caller is responsible for calling smp_invltlb().
3702 pmap_kenter_temporary(vm_paddr_t pa, long i)
3704 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
3705 return ((void *)crashdumpmap);
3708 #define MAX_INIT_PT (96)
3711 * This routine preloads the ptes for a given object into the specified pmap.
3712 * This eliminates the blast of soft faults on process startup and
3713 * immediately after an mmap.
3715 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
3718 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
3719 vm_object_t object, vm_pindex_t pindex,
3720 vm_size_t size, int limit)
3722 struct rb_vm_page_scan_info info;
3727 * We can't preinit if read access isn't set or there is no pmap
3730 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
3734 * We can't preinit if the pmap is not the current pmap
3736 lp = curthread->td_lwp;
3737 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
3741 * Misc additional checks
3743 psize = x86_64_btop(size);
3745 if ((object->type != OBJT_VNODE) ||
3746 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
3747 (object->resident_page_count > MAX_INIT_PT))) {
3751 if (pindex + psize > object->size) {
3752 if (object->size < pindex)
3754 psize = object->size - pindex;
3761 * If everything is segment-aligned do not pre-init here. Instead
3762 * allow the normal vm_fault path to pass a segment hint to
3763 * pmap_enter() which will then use an object-referenced shared
3766 if ((addr & SEG_MASK) == 0 &&
3767 (ctob(psize) & SEG_MASK) == 0 &&
3768 (ctob(pindex) & SEG_MASK) == 0) {
3773 * Use a red-black scan to traverse the requested range and load
3774 * any valid pages found into the pmap.
3776 * We cannot safely scan the object's memq without holding the
3779 info.start_pindex = pindex;
3780 info.end_pindex = pindex + psize - 1;
3786 vm_object_hold_shared(object);
3787 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3788 pmap_object_init_pt_callback, &info);
3789 vm_object_drop(object);
3794 pmap_object_init_pt_callback(vm_page_t p, void *data)
3796 struct rb_vm_page_scan_info *info = data;
3797 vm_pindex_t rel_index;
3800 * don't allow an madvise to blow away our really
3801 * free pages allocating pv entries.
3803 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3804 vmstats.v_free_count < vmstats.v_free_reserved) {
3809 * Ignore list markers and ignore pages we cannot instantly
3810 * busy (while holding the object token).
3812 if (p->flags & PG_MARKER)
3814 if (vm_page_busy_try(p, TRUE))
3816 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3817 (p->flags & PG_FICTITIOUS) == 0) {
3818 if ((p->queue - p->pc) == PQ_CACHE)
3819 vm_page_deactivate(p);
3820 rel_index = p->pindex - info->start_pindex;
3821 pmap_enter_quick(info->pmap,
3822 info->addr + x86_64_ptob(rel_index), p);
3830 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
3833 * Returns FALSE if it would be non-trivial or if a pte is already loaded
3836 * XXX This is safe only because page table pages are not freed.
3839 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3843 /*spin_lock(&pmap->pm_spin);*/
3844 if ((pte = pmap_pte(pmap, addr)) != NULL) {
3846 /*spin_unlock(&pmap->pm_spin);*/
3850 /*spin_unlock(&pmap->pm_spin);*/
3855 * Change the wiring attribute for a pmap/va pair. The mapping must already
3856 * exist in the pmap. The mapping may or may not be managed.
3859 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
3860 vm_map_entry_t entry)
3867 lwkt_gettoken(&pmap->pm_token);
3868 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
3869 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
3871 if (wired && !pmap_pte_w(ptep))
3872 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
3873 else if (!wired && pmap_pte_w(ptep))
3874 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
3877 * Wiring is not a hardware characteristic so there is no need to
3878 * invalidate TLB. However, in an SMP environment we must use
3879 * a locked bus cycle to update the pte (if we are not using
3880 * the pmap_inval_*() API that is)... it's ok to do this for simple
3885 atomic_set_long(ptep, PG_W);
3887 atomic_clear_long(ptep, PG_W);
3890 atomic_set_long_nonlocked(ptep, PG_W);
3892 atomic_clear_long_nonlocked(ptep, PG_W);
3895 lwkt_reltoken(&pmap->pm_token);
3901 * Copy the range specified by src_addr/len from the source map to
3902 * the range dst_addr/len in the destination map.
3904 * This routine is only advisory and need not do anything.
3907 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3908 vm_size_t len, vm_offset_t src_addr)
3915 * Zero the specified physical page.
3917 * This function may be called from an interrupt and no locking is
3921 pmap_zero_page(vm_paddr_t phys)
3923 vm_offset_t va = PHYS_TO_DMAP(phys);
3925 pagezero((void *)va);
3929 * pmap_page_assertzero:
3931 * Assert that a page is empty, panic if it isn't.
3934 pmap_page_assertzero(vm_paddr_t phys)
3936 vm_offset_t va = PHYS_TO_DMAP(phys);
3939 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3940 if (*(long *)((char *)va + i) != 0) {
3941 panic("pmap_page_assertzero() @ %p not zero!",
3942 (void *)(intptr_t)va);
3950 * Zero part of a physical page by mapping it into memory and clearing
3951 * its contents with bzero.
3953 * off and size may not cover an area beyond a single hardware page.
3956 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3958 vm_offset_t virt = PHYS_TO_DMAP(phys);
3960 bzero((char *)virt + off, size);
3966 * Copy the physical page from the source PA to the target PA.
3967 * This function may be called from an interrupt. No locking
3971 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3973 vm_offset_t src_virt, dst_virt;
3975 src_virt = PHYS_TO_DMAP(src);
3976 dst_virt = PHYS_TO_DMAP(dst);
3977 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3981 * pmap_copy_page_frag:
3983 * Copy the physical page from the source PA to the target PA.
3984 * This function may be called from an interrupt. No locking
3988 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3990 vm_offset_t src_virt, dst_virt;
3992 src_virt = PHYS_TO_DMAP(src);
3993 dst_virt = PHYS_TO_DMAP(dst);
3995 bcopy((char *)src_virt + (src & PAGE_MASK),
3996 (char *)dst_virt + (dst & PAGE_MASK),
4001 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4002 * this page. This count may be changed upwards or downwards in the future;
4003 * it is only necessary that true be returned for a small subset of pmaps
4004 * for proper page aging.
4007 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4012 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4015 vm_page_spin_lock(m);
4016 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4017 if (pv->pv_pmap == pmap) {
4018 vm_page_spin_unlock(m);
4025 vm_page_spin_unlock(m);
4030 * Remove all pages from specified address space this aids process exit
4031 * speeds. Also, this code may be special cased for the current process
4035 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4037 pmap_remove(pmap, sva, eva);
4041 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4042 * routines are inline, and a lot of things compile-time evaluate.
4046 pmap_testbit(vm_page_t m, int bit)
4051 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4054 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4056 vm_page_spin_lock(m);
4057 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4058 vm_page_spin_unlock(m);
4062 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4064 * if the bit being tested is the modified bit, then
4065 * mark clean_map and ptes as never
4068 if (bit & (PG_A|PG_M)) {
4069 if (!pmap_track_modified(pv->pv_pindex))
4073 #if defined(PMAP_DIAGNOSTIC)
4074 if (pv->pv_pmap == NULL) {
4075 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4080 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4082 vm_page_spin_unlock(m);
4086 vm_page_spin_unlock(m);
4091 * This routine is used to modify bits in ptes. Only one bit should be
4092 * specified. PG_RW requires special handling.
4094 * Caller must NOT hold any spin locks
4098 pmap_clearbit(vm_page_t m, int bit)
4100 struct pmap_inval_info info;
4107 vm_page_flag_clear(m, PG_WRITEABLE);
4108 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4115 * Loop over all current mappings setting/clearing as appropos If
4116 * setting RO do we need to clear the VAC?
4118 * NOTE: When clearing PG_M we could also (not implemented) drop
4119 * through to the PG_RW code and clear PG_RW too, forcing
4120 * a fault on write to redetect PG_M for virtual kernels, but
4121 * it isn't necessary since virtual kernels invalidate the
4122 * pte when they clear the VPTE_M bit in their virtual page
4125 * NOTE: Does not re-dirty the page when clearing only PG_M.
4127 if ((bit & PG_RW) == 0) {
4128 vm_page_spin_lock(m);
4129 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4130 #if defined(PMAP_DIAGNOSTIC)
4131 if (pv->pv_pmap == NULL) {
4132 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4137 pte = pmap_pte_quick(pv->pv_pmap,
4138 pv->pv_pindex << PAGE_SHIFT);
4141 atomic_clear_long(pte, bit);
4143 vm_page_spin_unlock(m);
4148 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4151 pmap_inval_init(&info);
4154 vm_page_spin_lock(m);
4155 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4157 * don't write protect pager mappings
4159 if (!pmap_track_modified(pv->pv_pindex))
4162 #if defined(PMAP_DIAGNOSTIC)
4163 if (pv->pv_pmap == NULL) {
4164 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4170 * Skip pages which do not have PG_RW set.
4172 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4173 if ((*pte & PG_RW) == 0)
4179 if (pv_hold_try(pv) == 0) {
4180 vm_page_spin_unlock(m);
4181 pv_lock(pv); /* held, now do a blocking lock */
4182 pv_put(pv); /* and release */
4183 goto restart; /* anything could have happened */
4186 save_pmap = pv->pv_pmap;
4187 vm_page_spin_unlock(m);
4188 pmap_inval_interlock(&info, save_pmap,
4189 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
4190 KKASSERT(pv->pv_pmap == save_pmap);
4194 if (atomic_cmpset_long(pte, pbits,
4195 pbits & ~(PG_RW|PG_M))) {
4199 pmap_inval_deinterlock(&info, save_pmap);
4200 vm_page_spin_lock(m);
4203 * If PG_M was found to be set while we were clearing PG_RW
4204 * we also clear PG_M (done above) and mark the page dirty.
4205 * Callers expect this behavior.
4211 vm_page_spin_unlock(m);
4212 pmap_inval_done(&info);
4216 * Lower the permission for all mappings to a given page.
4218 * Page must be busied by caller.
4221 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4223 /* JG NX support? */
4224 if ((prot & VM_PROT_WRITE) == 0) {
4225 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4227 * NOTE: pmap_clearbit(.. PG_RW) also clears
4228 * the PG_WRITEABLE flag in (m).
4230 pmap_clearbit(m, PG_RW);
4238 pmap_phys_address(vm_pindex_t ppn)
4240 return (x86_64_ptob(ppn));
4244 * Return a count of reference bits for a page, clearing those bits.
4245 * It is not necessary for every reference bit to be cleared, but it
4246 * is necessary that 0 only be returned when there are truly no
4247 * reference bits set.
4249 * XXX: The exact number of bits to check and clear is a matter that
4250 * should be tested and standardized at some point in the future for
4251 * optimal aging of shared pages.
4253 * This routine may not block.
4256 pmap_ts_referenced(vm_page_t m)
4262 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4265 vm_page_spin_lock(m);
4266 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4267 if (!pmap_track_modified(pv->pv_pindex))
4269 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4270 if (pte && (*pte & PG_A)) {
4272 atomic_clear_long(pte, PG_A);
4274 atomic_clear_long_nonlocked(pte, PG_A);
4281 vm_page_spin_unlock(m);
4288 * Return whether or not the specified physical page was modified
4289 * in any physical maps.
4292 pmap_is_modified(vm_page_t m)
4296 res = pmap_testbit(m, PG_M);
4301 * Clear the modify bits on the specified physical page.
4304 pmap_clear_modify(vm_page_t m)
4306 pmap_clearbit(m, PG_M);
4310 * pmap_clear_reference:
4312 * Clear the reference bit on the specified physical page.
4315 pmap_clear_reference(vm_page_t m)
4317 pmap_clearbit(m, PG_A);
4321 * Miscellaneous support routines follow
4326 i386_protection_init(void)
4330 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4331 kp = protection_codes;
4332 for (prot = 0; prot < 8; prot++) {
4334 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4336 * Read access is also 0. There isn't any execute bit,
4337 * so just make it readable.
4339 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4340 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4341 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4344 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4345 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4346 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4347 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4355 * Map a set of physical memory pages into the kernel virtual
4356 * address space. Return a pointer to where it is mapped. This
4357 * routine is intended to be used for mapping device memory,
4360 * NOTE: we can't use pgeflag unless we invalidate the pages one at
4364 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
4366 vm_offset_t va, tmpva, offset;
4369 offset = pa & PAGE_MASK;
4370 size = roundup(offset + size, PAGE_SIZE);
4372 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4374 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4376 pa = pa & ~PAGE_MASK;
4377 for (tmpva = va; size > 0;) {
4378 pte = vtopte(tmpva);
4379 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
4387 return ((void *)(va + offset));
4391 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
4393 vm_offset_t va, tmpva, offset;
4396 offset = pa & PAGE_MASK;
4397 size = roundup(offset + size, PAGE_SIZE);
4399 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4401 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4403 pa = pa & ~PAGE_MASK;
4404 for (tmpva = va; size > 0;) {
4405 pte = vtopte(tmpva);
4406 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
4414 return ((void *)(va + offset));
4418 pmap_unmapdev(vm_offset_t va, vm_size_t size)
4420 vm_offset_t base, offset;
4422 base = va & ~PAGE_MASK;
4423 offset = va & PAGE_MASK;
4424 size = roundup(offset + size, PAGE_SIZE);
4425 pmap_qremove(va, size >> PAGE_SHIFT);
4426 kmem_free(&kernel_map, base, size);
4430 * perform the pmap work for mincore
4433 pmap_mincore(pmap_t pmap, vm_offset_t addr)
4435 pt_entry_t *ptep, pte;
4439 lwkt_gettoken(&pmap->pm_token);
4440 ptep = pmap_pte(pmap, addr);
4442 if (ptep && (pte = *ptep) != 0) {
4445 val = MINCORE_INCORE;
4446 if ((pte & PG_MANAGED) == 0)
4449 pa = pte & PG_FRAME;
4451 m = PHYS_TO_VM_PAGE(pa);
4457 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
4459 * Modified by someone
4461 else if (m->dirty || pmap_is_modified(m))
4462 val |= MINCORE_MODIFIED_OTHER;
4467 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
4470 * Referenced by someone
4472 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
4473 val |= MINCORE_REFERENCED_OTHER;
4474 vm_page_flag_set(m, PG_REFERENCED);
4478 lwkt_reltoken(&pmap->pm_token);
4484 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
4485 * vmspace will be ref'd and the old one will be deref'd.
4487 * The vmspace for all lwps associated with the process will be adjusted
4488 * and cr3 will be reloaded if any lwp is the current lwp.
4490 * The process must hold the vmspace->vm_map.token for oldvm and newvm
4493 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
4495 struct vmspace *oldvm;
4498 oldvm = p->p_vmspace;
4499 if (oldvm != newvm) {
4501 sysref_get(&newvm->vm_sysref);
4502 p->p_vmspace = newvm;
4503 KKASSERT(p->p_nthreads == 1);
4504 lp = RB_ROOT(&p->p_lwp_tree);
4505 pmap_setlwpvm(lp, newvm);
4507 sysref_put(&oldvm->vm_sysref);
4512 * Set the vmspace for a LWP. The vmspace is almost universally set the
4513 * same as the process vmspace, but virtual kernels need to swap out contexts
4514 * on a per-lwp basis.
4516 * Caller does not necessarily hold any vmspace tokens. Caller must control
4517 * the lwp (typically be in the context of the lwp). We use a critical
4518 * section to protect against statclock and hardclock (statistics collection).
4521 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
4523 struct vmspace *oldvm;
4526 oldvm = lp->lwp_vmspace;
4528 if (oldvm != newvm) {
4530 lp->lwp_vmspace = newvm;
4531 if (curthread->td_lwp == lp) {
4532 pmap = vmspace_pmap(newvm);
4534 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4535 if (pmap->pm_active & CPUMASK_LOCK)
4536 pmap_interlock_wait(newvm);
4538 pmap->pm_active |= 1;
4540 #if defined(SWTCH_OPTIM_STATS)
4543 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4544 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
4545 load_cr3(curthread->td_pcb->pcb_cr3);
4546 pmap = vmspace_pmap(oldvm);
4548 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4550 pmap->pm_active &= ~(cpumask_t)1;
4560 * Called when switching to a locked pmap, used to interlock against pmaps
4561 * undergoing modifications to prevent us from activating the MMU for the
4562 * target pmap until all such modifications have completed. We have to do
4563 * this because the thread making the modifications has already set up its
4564 * SMP synchronization mask.
4566 * This function cannot sleep!
4571 pmap_interlock_wait(struct vmspace *vm)
4573 struct pmap *pmap = &vm->vm_pmap;
4575 if (pmap->pm_active & CPUMASK_LOCK) {
4577 KKASSERT(curthread->td_critcount >= 2);
4578 DEBUG_PUSH_INFO("pmap_interlock_wait");
4579 while (pmap->pm_active & CPUMASK_LOCK) {
4581 lwkt_process_ipiq();
4591 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4594 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4598 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4603 * Used by kmalloc/kfree, page already exists at va
4606 pmap_kvtom(vm_offset_t va)
4608 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));
4612 * Initialize machine-specific shared page directory support. This
4613 * is executed when a VM object is created.
4616 pmap_object_init(vm_object_t object)
4618 object->md.pmap_rw = NULL;
4619 object->md.pmap_ro = NULL;
4623 * Clean up machine-specific shared page directory support. This
4624 * is executed when a VM object is destroyed.
4627 pmap_object_free(vm_object_t object)
4631 if ((pmap = object->md.pmap_rw) != NULL) {
4632 object->md.pmap_rw = NULL;
4633 kprintf("pmap_object_free: destroying pmap %p in obj %p\n",
4635 pmap_remove_pages(pmap,
4636 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
4637 pmap->pm_active = 0;
4640 kfree(pmap, M_OBJPMAP);
4642 if ((pmap = object->md.pmap_ro) != NULL) {
4643 object->md.pmap_ro = NULL;
4644 kprintf("pmap_object_free: destroying pmap %p in obj %p\n",
4646 pmap_remove_pages(pmap,
4647 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
4648 pmap->pm_active = 0;
4651 kfree(pmap, M_OBJPMAP);