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, pv_entry_t pvp);
245 struct pmap_scan_info;
246 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
247 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
248 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
249 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
250 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
251 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
253 static void i386_protection_init (void);
254 static void create_pagetables(vm_paddr_t *firstaddr);
255 static void pmap_remove_all (vm_page_t m);
256 static boolean_t pmap_testbit (vm_page_t m, int bit);
258 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
259 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
261 static unsigned pdir4mb;
264 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
266 if (pv1->pv_pindex < pv2->pv_pindex)
268 if (pv1->pv_pindex > pv2->pv_pindex)
273 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
274 pv_entry_compare, vm_pindex_t, pv_pindex);
277 * Move the kernel virtual free pointer to the next
278 * 2MB. This is used to help improve performance
279 * by using a large (2MB) page for much of the kernel
280 * (.text, .data, .bss)
284 pmap_kmem_choose(vm_offset_t addr)
286 vm_offset_t newaddr = addr;
288 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
295 * Super fast pmap_pte routine best used when scanning the pv lists.
296 * This eliminates many course-grained invltlb calls. Note that many of
297 * the pv list scans are across different pmaps and it is very wasteful
298 * to do an entire invltlb when checking a single mapping.
300 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
304 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
306 return pmap_pte(pmap, va);
310 * Returns the pindex of a page table entry (representing a terminal page).
311 * There are NUPTE_TOTAL page table entries possible (a huge number)
313 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
314 * We want to properly translate negative KVAs.
318 pmap_pte_pindex(vm_offset_t va)
320 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
324 * Returns the pindex of a page table.
328 pmap_pt_pindex(vm_offset_t va)
330 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
334 * Returns the pindex of a page directory.
338 pmap_pd_pindex(vm_offset_t va)
340 return (NUPTE_TOTAL + NUPT_TOTAL +
341 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
346 pmap_pdp_pindex(vm_offset_t va)
348 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
349 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
354 pmap_pml4_pindex(void)
356 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
360 * Return various clipped indexes for a given VA
362 * Returns the index of a pte in a page table, representing a terminal
367 pmap_pte_index(vm_offset_t va)
369 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
373 * Returns the index of a pt in a page directory, representing a page
378 pmap_pt_index(vm_offset_t va)
380 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
384 * Returns the index of a pd in a page directory page, representing a page
389 pmap_pd_index(vm_offset_t va)
391 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
395 * Returns the index of a pdp in the pml4 table, representing a page
400 pmap_pdp_index(vm_offset_t va)
402 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
406 * Generic procedure to index a pte from a pt, pd, or pdp.
408 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
409 * a page table page index but is instead of PV lookup index.
413 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
417 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
418 return(&pte[pindex]);
422 * Return pointer to PDP slot in the PML4
426 pmap_pdp(pmap_t pmap, vm_offset_t va)
428 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
432 * Return pointer to PD slot in the PDP given a pointer to the PDP
436 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
440 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
441 return (&pd[pmap_pd_index(va)]);
445 * Return pointer to PD slot in the PDP.
449 pmap_pd(pmap_t pmap, vm_offset_t va)
453 pdp = pmap_pdp(pmap, va);
454 if ((*pdp & PG_V) == 0)
456 return (pmap_pdp_to_pd(*pdp, va));
460 * Return pointer to PT slot in the PD given a pointer to the PD
464 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
468 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
469 return (&pt[pmap_pt_index(va)]);
473 * Return pointer to PT slot in the PD
475 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
476 * so we cannot lookup the PD via the PDP. Instead we
477 * must look it up via the pmap.
481 pmap_pt(pmap_t pmap, vm_offset_t va)
485 vm_pindex_t pd_pindex;
487 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
488 pd_pindex = pmap_pd_pindex(va);
489 spin_lock(&pmap->pm_spin);
490 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
491 spin_unlock(&pmap->pm_spin);
492 if (pv == NULL || pv->pv_m == NULL)
494 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
496 pd = pmap_pd(pmap, va);
497 if (pd == NULL || (*pd & PG_V) == 0)
499 return (pmap_pd_to_pt(*pd, va));
504 * Return pointer to PTE slot in the PT given a pointer to the PT
508 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
512 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
513 return (&pte[pmap_pte_index(va)]);
517 * Return pointer to PTE slot in the PT
521 pmap_pte(pmap_t pmap, vm_offset_t va)
525 pt = pmap_pt(pmap, va);
526 if (pt == NULL || (*pt & PG_V) == 0)
528 if ((*pt & PG_PS) != 0)
529 return ((pt_entry_t *)pt);
530 return (pmap_pt_to_pte(*pt, va));
534 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
535 * the PT layer. This will speed up core pmap operations considerably.
539 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
541 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
542 pv->pv_pmap->pm_pvhint = pv;
547 * KVM - return address of PT slot in PD
551 vtopt(vm_offset_t va)
553 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
554 NPML4EPGSHIFT)) - 1);
556 return (PDmap + ((va >> PDRSHIFT) & mask));
560 * KVM - return address of PTE slot in PT
564 vtopte(vm_offset_t va)
566 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
567 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
569 return (PTmap + ((va >> PAGE_SHIFT) & mask));
573 allocpages(vm_paddr_t *firstaddr, long n)
578 bzero((void *)ret, n * PAGE_SIZE);
579 *firstaddr += n * PAGE_SIZE;
585 create_pagetables(vm_paddr_t *firstaddr)
587 long i; /* must be 64 bits */
593 * We are running (mostly) V=P at this point
595 * Calculate NKPT - number of kernel page tables. We have to
596 * accomodoate prealloction of the vm_page_array, dump bitmap,
597 * MSGBUF_SIZE, and other stuff. Be generous.
599 * Maxmem is in pages.
601 * ndmpdp is the number of 1GB pages we wish to map.
603 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
604 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
606 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
609 * Starting at the beginning of kvm (not KERNBASE).
611 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
612 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
613 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
614 ndmpdp) + 511) / 512;
618 * Starting at KERNBASE - map 2G worth of page table pages.
619 * KERNBASE is offset -2G from the end of kvm.
621 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
626 KPTbase = allocpages(firstaddr, nkpt_base);
627 KPTphys = allocpages(firstaddr, nkpt_phys);
628 KPML4phys = allocpages(firstaddr, 1);
629 KPDPphys = allocpages(firstaddr, NKPML4E);
630 KPDphys = allocpages(firstaddr, NKPDPE);
633 * Calculate the page directory base for KERNBASE,
634 * that is where we start populating the page table pages.
635 * Basically this is the end - 2.
637 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
639 DMPDPphys = allocpages(firstaddr, NDMPML4E);
640 if ((amd_feature & AMDID_PAGE1GB) == 0)
641 DMPDphys = allocpages(firstaddr, ndmpdp);
642 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
645 * Fill in the underlying page table pages for the area around
646 * KERNBASE. This remaps low physical memory to KERNBASE.
648 * Read-only from zero to physfree
649 * XXX not fully used, underneath 2M pages
651 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
652 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
653 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
657 * Now map the initial kernel page tables. One block of page
658 * tables is placed at the beginning of kernel virtual memory,
659 * and another block is placed at KERNBASE to map the kernel binary,
660 * data, bss, and initial pre-allocations.
662 for (i = 0; i < nkpt_base; i++) {
663 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
664 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
666 for (i = 0; i < nkpt_phys; i++) {
667 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
668 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
672 * Map from zero to end of allocations using 2M pages as an
673 * optimization. This will bypass some of the KPTBase pages
674 * above in the KERNBASE area.
676 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
677 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
678 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
682 * And connect up the PD to the PDP. The kernel pmap is expected
683 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
685 for (i = 0; i < NKPDPE; i++) {
686 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
687 KPDphys + (i << PAGE_SHIFT);
688 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
693 * Now set up the direct map space using either 2MB or 1GB pages
694 * Preset PG_M and PG_A because demotion expects it.
696 * When filling in entries in the PD pages make sure any excess
697 * entries are set to zero as we allocated enough PD pages
699 if ((amd_feature & AMDID_PAGE1GB) == 0) {
700 for (i = 0; i < NPDEPG * ndmpdp; i++) {
701 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
702 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
707 * And the direct map space's PDP
709 for (i = 0; i < ndmpdp; i++) {
710 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
712 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
715 for (i = 0; i < ndmpdp; i++) {
716 ((pdp_entry_t *)DMPDPphys)[i] =
717 (vm_paddr_t)i << PDPSHIFT;
718 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
723 /* And recursively map PML4 to itself in order to get PTmap */
724 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
725 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
728 * Connect the Direct Map slots up to the PML4
730 for (j = 0; j < NDMPML4E; ++j) {
731 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
732 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
737 * Connect the KVA slot up to the PML4
739 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
740 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
744 * Bootstrap the system enough to run with virtual memory.
746 * On the i386 this is called after mapping has already been enabled
747 * and just syncs the pmap module with what has already been done.
748 * [We can't call it easily with mapping off since the kernel is not
749 * mapped with PA == VA, hence we would have to relocate every address
750 * from the linked base (virtual) address "KERNBASE" to the actual
751 * (physical) address starting relative to 0]
754 pmap_bootstrap(vm_paddr_t *firstaddr)
759 KvaStart = VM_MIN_KERNEL_ADDRESS;
760 KvaEnd = VM_MAX_KERNEL_ADDRESS;
761 KvaSize = KvaEnd - KvaStart;
763 avail_start = *firstaddr;
766 * Create an initial set of page tables to run the kernel in.
768 create_pagetables(firstaddr);
770 virtual2_start = KvaStart;
771 virtual2_end = PTOV_OFFSET;
773 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
774 virtual_start = pmap_kmem_choose(virtual_start);
776 virtual_end = VM_MAX_KERNEL_ADDRESS;
778 /* XXX do %cr0 as well */
779 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
783 * Initialize protection array.
785 i386_protection_init();
788 * The kernel's pmap is statically allocated so we don't have to use
789 * pmap_create, which is unlikely to work correctly at this part of
790 * the boot sequence (XXX and which no longer exists).
792 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
793 kernel_pmap.pm_count = 1;
794 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
795 RB_INIT(&kernel_pmap.pm_pvroot);
796 spin_init(&kernel_pmap.pm_spin);
797 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
800 * Reserve some special page table entries/VA space for temporary
803 #define SYSMAP(c, p, v, n) \
804 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
810 * CMAP1/CMAP2 are used for zeroing and copying pages.
812 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
817 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
820 * ptvmmap is used for reading arbitrary physical pages via
823 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
826 * msgbufp is used to map the system message buffer.
827 * XXX msgbufmap is not used.
829 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
830 atop(round_page(MSGBUF_SIZE)))
837 * PG_G is terribly broken on SMP because we IPI invltlb's in some
838 * cases rather then invl1pg. Actually, I don't even know why it
839 * works under UP because self-referential page table mappings
844 if (cpu_feature & CPUID_PGE)
849 * Initialize the 4MB page size flag
853 * The 4MB page version of the initial
854 * kernel page mapping.
858 #if !defined(DISABLE_PSE)
859 if (cpu_feature & CPUID_PSE) {
862 * Note that we have enabled PSE mode
865 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
866 ptditmp &= ~(NBPDR - 1);
867 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
872 * Enable the PSE mode. If we are SMP we can't do this
873 * now because the APs will not be able to use it when
876 load_cr4(rcr4() | CR4_PSE);
879 * We can do the mapping here for the single processor
880 * case. We simply ignore the old page table page from
884 * For SMP, we still need 4K pages to bootstrap APs,
885 * PSE will be enabled as soon as all APs are up.
887 PTD[KPTDI] = (pd_entry_t)ptditmp;
897 * Set 4mb pdir for mp startup
902 if (pseflag && (cpu_feature & CPUID_PSE)) {
903 load_cr4(rcr4() | CR4_PSE);
904 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
912 * Initialize the pmap module.
913 * Called by vm_init, to initialize any structures that the pmap
914 * system needs to map virtual memory.
915 * pmap_init has been enhanced to support in a fairly consistant
916 * way, discontiguous physical memory.
925 * Allocate memory for random pmap data structures. Includes the
929 for (i = 0; i < vm_page_array_size; i++) {
932 m = &vm_page_array[i];
933 TAILQ_INIT(&m->md.pv_list);
937 * init the pv free list
939 initial_pvs = vm_page_array_size;
940 if (initial_pvs < MINPV)
942 pvzone = &pvzone_store;
943 pvinit = (void *)kmem_alloc(&kernel_map,
944 initial_pvs * sizeof (struct pv_entry));
945 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
946 pvinit, initial_pvs);
949 * Now it is safe to enable pv_table recording.
951 pmap_initialized = TRUE;
955 * Initialize the address space (zone) for the pv_entries. Set a
956 * high water mark so that the system can recover from excessive
957 * numbers of pv entries.
962 int shpgperproc = PMAP_SHPGPERPROC;
965 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
966 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
967 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
968 pv_entry_high_water = 9 * (pv_entry_max / 10);
971 * Subtract out pages already installed in the zone (hack)
973 entry_max = pv_entry_max - vm_page_array_size;
977 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
981 /***************************************************
982 * Low level helper routines.....
983 ***************************************************/
986 * this routine defines the region(s) of memory that should
987 * not be tested for the modified bit.
991 pmap_track_modified(vm_pindex_t pindex)
993 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
994 if ((va < clean_sva) || (va >= clean_eva))
1001 * Extract the physical page address associated with the map/VA pair.
1002 * The page must be wired for this to work reliably.
1004 * XXX for the moment we're using pv_find() instead of pv_get(), as
1005 * callers might be expecting non-blocking operation.
1008 pmap_extract(pmap_t pmap, vm_offset_t va)
1015 if (va >= VM_MAX_USER_ADDRESS) {
1017 * Kernel page directories might be direct-mapped and
1018 * there is typically no PV tracking of pte's
1022 pt = pmap_pt(pmap, va);
1023 if (pt && (*pt & PG_V)) {
1025 rtval = *pt & PG_PS_FRAME;
1026 rtval |= va & PDRMASK;
1028 ptep = pmap_pt_to_pte(*pt, va);
1030 rtval = *ptep & PG_FRAME;
1031 rtval |= va & PAGE_MASK;
1037 * User pages currently do not direct-map the page directory
1038 * and some pages might not used managed PVs. But all PT's
1041 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1043 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1045 rtval = *ptep & PG_FRAME;
1046 rtval |= va & PAGE_MASK;
1055 * Extract the physical page address associated kernel virtual address.
1058 pmap_kextract(vm_offset_t va)
1060 pd_entry_t pt; /* pt entry in pd */
1063 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1064 pa = DMAP_TO_PHYS(va);
1068 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1071 * Beware of a concurrent promotion that changes the
1072 * PDE at this point! For example, vtopte() must not
1073 * be used to access the PTE because it would use the
1074 * new PDE. It is, however, safe to use the old PDE
1075 * because the page table page is preserved by the
1078 pa = *pmap_pt_to_pte(pt, va);
1079 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1085 /***************************************************
1086 * Low level mapping routines.....
1087 ***************************************************/
1090 * Routine: pmap_kenter
1092 * Add a wired page to the KVA
1093 * NOTE! note that in order for the mapping to take effect -- you
1094 * should do an invltlb after doing the pmap_kenter().
1097 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1101 pmap_inval_info info;
1103 pmap_inval_init(&info); /* XXX remove */
1104 npte = pa | PG_RW | PG_V | pgeflag;
1106 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1108 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1109 pmap_inval_done(&info); /* XXX remove */
1113 * Routine: pmap_kenter_quick
1115 * Similar to pmap_kenter(), except we only invalidate the
1116 * mapping on the current CPU.
1119 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1124 npte = pa | PG_RW | PG_V | pgeflag;
1127 cpu_invlpg((void *)va);
1131 pmap_kenter_sync(vm_offset_t va)
1133 pmap_inval_info info;
1135 pmap_inval_init(&info);
1136 pmap_inval_interlock(&info, &kernel_pmap, va);
1137 pmap_inval_deinterlock(&info, &kernel_pmap);
1138 pmap_inval_done(&info);
1142 pmap_kenter_sync_quick(vm_offset_t va)
1144 cpu_invlpg((void *)va);
1148 * remove a page from the kernel pagetables
1151 pmap_kremove(vm_offset_t va)
1154 pmap_inval_info info;
1156 pmap_inval_init(&info);
1158 pmap_inval_interlock(&info, &kernel_pmap, va);
1159 (void)pte_load_clear(pte);
1160 pmap_inval_deinterlock(&info, &kernel_pmap);
1161 pmap_inval_done(&info);
1165 pmap_kremove_quick(vm_offset_t va)
1169 (void)pte_load_clear(pte);
1170 cpu_invlpg((void *)va);
1174 * XXX these need to be recoded. They are not used in any critical path.
1177 pmap_kmodify_rw(vm_offset_t va)
1179 atomic_set_long(vtopte(va), PG_RW);
1180 cpu_invlpg((void *)va);
1184 pmap_kmodify_nc(vm_offset_t va)
1186 atomic_set_long(vtopte(va), PG_N);
1187 cpu_invlpg((void *)va);
1191 * Used to map a range of physical addresses into kernel virtual
1192 * address space during the low level boot, typically to map the
1193 * dump bitmap, message buffer, and vm_page_array.
1195 * These mappings are typically made at some pointer after the end of the
1198 * We could return PHYS_TO_DMAP(start) here and not allocate any
1199 * via (*virtp), but then kmem from userland and kernel dumps won't
1200 * have access to the related pointers.
1203 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1206 vm_offset_t va_start;
1208 /*return PHYS_TO_DMAP(start);*/
1213 while (start < end) {
1214 pmap_kenter_quick(va, start);
1224 * Add a list of wired pages to the kva
1225 * this routine is only used for temporary
1226 * kernel mappings that do not need to have
1227 * page modification or references recorded.
1228 * Note that old mappings are simply written
1229 * over. The page *must* be wired.
1232 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1236 end_va = va + count * PAGE_SIZE;
1238 while (va < end_va) {
1242 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1243 cpu_invlpg((void *)va);
1251 * This routine jerks page mappings from the
1252 * kernel -- it is meant only for temporary mappings.
1254 * MPSAFE, INTERRUPT SAFE (cluster callback)
1257 pmap_qremove(vm_offset_t va, int count)
1261 end_va = va + count * PAGE_SIZE;
1263 while (va < end_va) {
1267 (void)pte_load_clear(pte);
1268 cpu_invlpg((void *)va);
1275 * Create a new thread and optionally associate it with a (new) process.
1276 * NOTE! the new thread's cpu may not equal the current cpu.
1279 pmap_init_thread(thread_t td)
1281 /* enforce pcb placement & alignment */
1282 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1283 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1284 td->td_savefpu = &td->td_pcb->pcb_save;
1285 td->td_sp = (char *)td->td_pcb; /* no -16 */
1289 * This routine directly affects the fork perf for a process.
1292 pmap_init_proc(struct proc *p)
1297 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1298 * it, and IdlePTD, represents the template used to update all other pmaps.
1300 * On architectures where the kernel pmap is not integrated into the user
1301 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1302 * kernel_pmap should be used to directly access the kernel_pmap.
1305 pmap_pinit0(struct pmap *pmap)
1307 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1309 pmap->pm_active = 0;
1310 pmap->pm_pvhint = NULL;
1311 RB_INIT(&pmap->pm_pvroot);
1312 spin_init(&pmap->pm_spin);
1313 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1314 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1318 * Initialize a preallocated and zeroed pmap structure,
1319 * such as one in a vmspace structure.
1322 pmap_pinit_simple(struct pmap *pmap)
1325 * Misc initialization
1328 pmap->pm_active = 0;
1329 pmap->pm_pvhint = NULL;
1330 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1333 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1336 if (pmap->pm_pmlpv == NULL) {
1337 RB_INIT(&pmap->pm_pvroot);
1338 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1339 spin_init(&pmap->pm_spin);
1340 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1345 pmap_pinit(struct pmap *pmap)
1350 pmap_pinit_simple(pmap);
1351 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1354 * No need to allocate page table space yet but we do need a valid
1355 * page directory table.
1357 if (pmap->pm_pml4 == NULL) {
1359 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1363 * Allocate the page directory page, which wires it even though
1364 * it isn't being entered into some higher level page table (it
1365 * being the highest level). If one is already cached we don't
1366 * have to do anything.
1368 if ((pv = pmap->pm_pmlpv) == NULL) {
1369 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1370 pmap->pm_pmlpv = pv;
1371 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1372 VM_PAGE_TO_PHYS(pv->pv_m));
1376 * Install DMAP and KMAP.
1378 for (j = 0; j < NDMPML4E; ++j) {
1379 pmap->pm_pml4[DMPML4I + j] =
1380 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1381 PG_RW | PG_V | PG_U;
1383 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1386 * install self-referential address mapping entry
1388 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1389 PG_V | PG_RW | PG_A | PG_M;
1391 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1392 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1394 KKASSERT(pmap->pm_pml4[255] == 0);
1395 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1396 KKASSERT(pv->pv_entry.rbe_left == NULL);
1397 KKASSERT(pv->pv_entry.rbe_right == NULL);
1401 * Clean up a pmap structure so it can be physically freed. This routine
1402 * is called by the vmspace dtor function. A great deal of pmap data is
1403 * left passively mapped to improve vmspace management so we have a bit
1404 * of cleanup work to do here.
1407 pmap_puninit(pmap_t pmap)
1412 KKASSERT(pmap->pm_active == 0);
1413 if ((pv = pmap->pm_pmlpv) != NULL) {
1414 if (pv_hold_try(pv) == 0)
1416 p = pmap_remove_pv_page(pv);
1418 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1419 vm_page_busy_wait(p, FALSE, "pgpun");
1420 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1421 vm_page_unwire(p, 0);
1422 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1425 * XXX eventually clean out PML4 static entries and
1426 * use vm_page_free_zero()
1429 pmap->pm_pmlpv = NULL;
1431 if (pmap->pm_pml4) {
1432 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1433 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1434 pmap->pm_pml4 = NULL;
1436 KKASSERT(pmap->pm_stats.resident_count == 0);
1437 KKASSERT(pmap->pm_stats.wired_count == 0);
1441 * Wire in kernel global address entries. To avoid a race condition
1442 * between pmap initialization and pmap_growkernel, this procedure
1443 * adds the pmap to the master list (which growkernel scans to update),
1444 * then copies the template.
1447 pmap_pinit2(struct pmap *pmap)
1449 spin_lock(&pmap_spin);
1450 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1451 spin_unlock(&pmap_spin);
1455 * This routine is called when various levels in the page table need to
1456 * be populated. This routine cannot fail.
1458 * This function returns two locked pv_entry's, one representing the
1459 * requested pv and one representing the requested pv's parent pv. If
1460 * the pv did not previously exist it will be mapped into its parent
1461 * and wired, otherwise no additional wire count will be added.
1465 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1470 vm_pindex_t pt_pindex;
1476 * If the pv already exists and we aren't being asked for the
1477 * parent page table page we can just return it. A locked+held pv
1481 pv = pv_alloc(pmap, ptepindex, &isnew);
1482 if (isnew == 0 && pvpp == NULL)
1486 * This is a new PV, we have to resolve its parent page table and
1487 * add an additional wiring to the page if necessary.
1491 * Special case terminal PVs. These are not page table pages so
1492 * no vm_page is allocated (the caller supplied the vm_page). If
1493 * pvpp is non-NULL we are being asked to also removed the pt_pv
1496 * Note that pt_pv's are only returned for user VAs. We assert that
1497 * a pt_pv is not being requested for kernel VAs.
1499 if (ptepindex < pmap_pt_pindex(0)) {
1500 if (ptepindex >= NUPTE_USER)
1501 KKASSERT(pvpp == NULL);
1503 KKASSERT(pvpp != NULL);
1505 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1506 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1508 vm_page_wire_quick(pvp->pv_m);
1517 * Non-terminal PVs allocate a VM page to represent the page table,
1518 * so we have to resolve pvp and calculate ptepindex for the pvp
1519 * and then for the page table entry index in the pvp for
1522 if (ptepindex < pmap_pd_pindex(0)) {
1524 * pv is PT, pvp is PD
1526 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1527 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1528 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1535 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1536 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1538 } else if (ptepindex < pmap_pdp_pindex(0)) {
1540 * pv is PD, pvp is PDP
1542 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1545 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1546 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1548 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1549 KKASSERT(pvpp == NULL);
1552 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1560 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1561 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1562 } else if (ptepindex < pmap_pml4_pindex()) {
1564 * pv is PDP, pvp is the root pml4 table
1566 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1573 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1574 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1577 * pv represents the top-level PML4, there is no parent.
1585 * This code is only reached if isnew is TRUE and this is not a
1586 * terminal PV. We need to allocate a vm_page for the page table
1587 * at this level and enter it into the parent page table.
1589 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1592 m = vm_page_alloc(NULL, pv->pv_pindex,
1593 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1594 VM_ALLOC_INTERRUPT);
1599 vm_page_spin_lock(m);
1600 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1602 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1603 vm_page_spin_unlock(m);
1604 vm_page_unmanage(m); /* m must be spinunlocked */
1606 if ((m->flags & PG_ZERO) == 0) {
1607 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1611 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1614 m->valid = VM_PAGE_BITS_ALL;
1615 vm_page_flag_clear(m, PG_ZERO);
1616 vm_page_wire(m); /* wire for mapping in parent */
1619 * Wire the page into pvp, bump the wire-count for pvp's page table
1620 * page. Bump the resident_count for the pmap. There is no pvp
1621 * for the top level, address the pm_pml4[] array directly.
1623 * If the caller wants the parent we return it, otherwise
1624 * we just put it away.
1626 * No interlock is needed for pte 0 -> non-zero.
1628 * In the situation where *ptep is valid we might have an unmanaged
1629 * page table page shared from another page table which we need to
1630 * unshare before installing our private page table page.
1633 ptep = pv_pte_lookup(pvp, ptepindex);
1636 pmap_inval_info info;
1639 panic("pmap_allocpte: unexpected pte %p/%d",
1640 pvp, (int)ptepindex);
1642 pmap_inval_init(&info);
1643 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1644 pte = pte_load_clear(ptep);
1645 pmap_inval_deinterlock(&info, pmap);
1646 pmap_inval_done(&info);
1647 if (vm_page_unwire_quick(
1648 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
1649 panic("pmap_allocpte: shared pgtable "
1650 "pg bad wirecount");
1652 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1654 vm_page_wire_quick(pvp->pv_m);
1656 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1669 * This version of pmap_allocpte() checks for possible segment optimizations
1670 * that would allow page-table sharing. It can be called for terminal
1671 * page or page table page ptepindex's.
1673 * The function is called with page table page ptepindex's for fictitious
1674 * and unmanaged terminal pages. That is, we don't want to allocate a
1675 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
1678 * This function can return a pv and *pvpp associated with the passed in pmap
1679 * OR a pv and *pvpp associated with the shared pmap. In the latter case
1680 * an unmanaged page table page will be entered into the pass in pmap.
1684 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
1685 vm_map_entry_t entry, vm_offset_t va)
1687 struct pmap_inval_info info;
1692 pv_entry_t pte_pv; /* in original or shared pmap */
1693 pv_entry_t pt_pv; /* in original or shared pmap */
1694 pv_entry_t proc_pd_pv; /* in original pmap */
1695 pv_entry_t proc_pt_pv; /* in original pmap */
1696 pv_entry_t xpv; /* PT in shared pmap */
1697 pd_entry_t *pt; /* PT entry in PD of original pmap */
1698 pd_entry_t opte; /* contents of *pt */
1699 pd_entry_t npte; /* contents of *pt */
1704 * Basic tests, require a non-NULL vm_map_entry, require proper
1705 * alignment and type for the vm_map_entry, require that the
1706 * underlying object already be allocated.
1708 * We currently allow any type of object to use this optimization.
1709 * The object itself does NOT have to be sized to a multiple of the
1710 * segment size, but the memory mapping does.
1712 if (entry == NULL ||
1713 pmap_mmu_optimize == 0 || /* not enabled */
1714 ptepindex >= pmap_pd_pindex(0) || /* not terminal */
1715 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
1716 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
1717 entry->object.vm_object == NULL || /* needs VM object */
1718 (entry->offset & SEG_MASK) || /* must be aligned */
1719 (entry->start & SEG_MASK)) {
1720 return(pmap_allocpte(pmap, ptepindex, pvpp));
1724 * Make sure the full segment can be represented.
1726 b = va & ~(vm_offset_t)SEG_MASK;
1727 if (b < entry->start && b + SEG_SIZE > entry->end)
1728 return(pmap_allocpte(pmap, ptepindex, pvpp));
1731 * If the full segment can be represented dive the VM object's
1732 * shared pmap, allocating as required.
1734 object = entry->object.vm_object;
1736 if (entry->protection & VM_PROT_WRITE)
1737 obpmapp = &object->md.pmap_rw;
1739 obpmapp = &object->md.pmap_ro;
1742 * We allocate what appears to be a normal pmap but because portions
1743 * of this pmap are shared with other unrelated pmaps we have to
1744 * set pm_active to point to all cpus.
1746 * XXX Currently using pmap_spin to interlock the update, can't use
1747 * vm_object_hold/drop because the token might already be held
1748 * shared OR exclusive and we don't know.
1750 while ((obpmap = *obpmapp) == NULL) {
1751 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
1752 pmap_pinit_simple(obpmap);
1753 pmap_pinit2(obpmap);
1754 spin_lock(&pmap_spin);
1755 if (*obpmapp != NULL) {
1759 spin_unlock(&pmap_spin);
1760 pmap_release(obpmap);
1761 pmap_puninit(obpmap);
1762 kfree(obpmap, M_OBJPMAP);
1764 obpmap->pm_active = smp_active_mask;
1766 spin_unlock(&pmap_spin);
1771 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
1772 * pte/pt using the shared pmap from the object but also adjust
1773 * the process pmap's page table page as a side effect.
1777 * Resolve the terminal PTE and PT in the shared pmap. This is what
1778 * we will return. This is true if ptepindex represents a terminal
1779 * page, otherwise pte_pv is actually the PT and pt_pv is actually
1783 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
1784 if (ptepindex >= pmap_pt_pindex(0))
1790 * Resolve the PD in the process pmap so we can properly share the
1791 * page table page. Lock order is bottom-up (leaf first)!
1793 * NOTE: proc_pt_pv can be NULL.
1795 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
1796 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
1799 * xpv is the page table page pv from the shared object
1800 * (for convenience).
1802 * Calculate the pte value for the PT to load into the process PD.
1803 * If we have to change it we must properly dispose of the previous
1806 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1807 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
1808 (PG_U | PG_RW | PG_V | PG_A | PG_M);
1811 * Dispose of previous page table page if it was local to the
1812 * process pmap. If the old pt is not empty we cannot dispose of it
1813 * until we clean it out. This case should not arise very often so
1814 * it is not optimized.
1817 if (proc_pt_pv->pv_m->wire_count != 1) {
1823 va & ~(vm_offset_t)SEG_MASK,
1824 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
1827 pmap_release_pv(proc_pt_pv, proc_pd_pv);
1830 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1834 * Handle remaining cases.
1838 vm_page_wire_quick(xpv->pv_m);
1839 vm_page_wire_quick(proc_pd_pv->pv_m);
1840 atomic_add_long(&pmap->pm_stats.resident_count, 1);
1841 } else if (*pt != npte) {
1842 pmap_inval_init(&info);
1843 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1845 opte = pte_load_clear(pt);
1846 KKASSERT(opte && opte != npte);
1849 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
1852 * Clean up opte, bump the wire_count for the process
1853 * PD page representing the new entry if it was
1856 * If the entry was not previously empty and we have
1857 * a PT in the proc pmap then opte must match that
1858 * pt. The proc pt must be retired (this is done
1859 * later on in this procedure).
1861 * NOTE: replacing valid pte, wire_count on proc_pd_pv
1864 KKASSERT(opte & PG_V);
1865 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
1866 if (vm_page_unwire_quick(m)) {
1867 panic("pmap_allocpte_seg: "
1868 "bad wire count %p",
1872 pmap_inval_deinterlock(&info, pmap);
1873 pmap_inval_done(&info);
1877 * The existing process page table was replaced and must be destroyed
1891 * Release any resources held by the given physical map.
1893 * Called when a pmap initialized by pmap_pinit is being released. Should
1894 * only be called if the map contains no valid mappings.
1896 * Caller must hold pmap->pm_token
1898 struct pmap_release_info {
1903 static int pmap_release_callback(pv_entry_t pv, void *data);
1906 pmap_release(struct pmap *pmap)
1908 struct pmap_release_info info;
1910 KASSERT(pmap->pm_active == 0,
1911 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1913 spin_lock(&pmap_spin);
1914 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1915 spin_unlock(&pmap_spin);
1918 * Pull pv's off the RB tree in order from low to high and release
1924 spin_lock(&pmap->pm_spin);
1925 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
1926 pmap_release_callback, &info);
1927 spin_unlock(&pmap->pm_spin);
1928 } while (info.retry);
1932 * One resident page (the pml4 page) should remain.
1933 * No wired pages should remain.
1935 KKASSERT(pmap->pm_stats.resident_count ==
1936 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
1938 KKASSERT(pmap->pm_stats.wired_count == 0);
1942 pmap_release_callback(pv_entry_t pv, void *data)
1944 struct pmap_release_info *info = data;
1945 pmap_t pmap = info->pmap;
1948 if (pv_hold_try(pv)) {
1949 spin_unlock(&pmap->pm_spin);
1951 spin_unlock(&pmap->pm_spin);
1953 if (pv->pv_pmap != pmap) {
1955 spin_lock(&pmap->pm_spin);
1960 r = pmap_release_pv(pv, NULL);
1961 spin_lock(&pmap->pm_spin);
1966 * Called with held (i.e. also locked) pv. This function will dispose of
1967 * the lock along with the pv.
1969 * If the caller already holds the locked parent page table for pv it
1970 * must pass it as pvp, allowing us to avoid a deadlock, else it can
1971 * pass NULL for pvp.
1974 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp)
1979 * The pmap is currently not spinlocked, pv is held+locked.
1980 * Remove the pv's page from its parent's page table. The
1981 * parent's page table page's wire_count will be decremented.
1983 pmap_remove_pv_pte(pv, pvp, NULL);
1986 * Terminal pvs are unhooked from their vm_pages. Because
1987 * terminal pages aren't page table pages they aren't wired
1988 * by us, so we have to be sure not to unwire them either.
1990 if (pv->pv_pindex < pmap_pt_pindex(0)) {
1991 pmap_remove_pv_page(pv);
1996 * We leave the top-level page table page cached, wired, and
1997 * mapped in the pmap until the dtor function (pmap_puninit())
2000 * Since we are leaving the top-level pv intact we need
2001 * to break out of what would otherwise be an infinite loop.
2003 if (pv->pv_pindex == pmap_pml4_pindex()) {
2009 * For page table pages (other than the top-level page),
2010 * remove and free the vm_page. The representitive mapping
2011 * removed above by pmap_remove_pv_pte() did not undo the
2012 * last wire_count so we have to do that as well.
2014 p = pmap_remove_pv_page(pv);
2015 vm_page_busy_wait(p, FALSE, "pmaprl");
2016 if (p->wire_count != 1) {
2017 kprintf("p->wire_count was %016lx %d\n",
2018 pv->pv_pindex, p->wire_count);
2020 KKASSERT(p->wire_count == 1);
2021 KKASSERT(p->flags & PG_UNMANAGED);
2023 vm_page_unwire(p, 0);
2024 KKASSERT(p->wire_count == 0);
2027 * Theoretically this page, if not the pml4 page, should contain
2028 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2038 * This function will remove the pte associated with a pv from its parent.
2039 * Terminal pv's are supported. The removal will be interlocked if info
2040 * is non-NULL. The caller must dispose of pv instead of just unlocking
2043 * The wire count will be dropped on the parent page table. The wire
2044 * count on the page being removed (pv->pv_m) from the parent page table
2045 * is NOT touched. Note that terminal pages will not have any additional
2046 * wire counts while page table pages will have at least one representing
2047 * the mapping, plus others representing sub-mappings.
2049 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2050 * pages and user page table and terminal pages.
2052 * The pv must be locked.
2054 * XXX must lock parent pv's if they exist to remove pte XXX
2058 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2060 vm_pindex_t ptepindex = pv->pv_pindex;
2061 pmap_t pmap = pv->pv_pmap;
2067 if (ptepindex == pmap_pml4_pindex()) {
2069 * We are the top level pml4 table, there is no parent.
2071 p = pmap->pm_pmlpv->pv_m;
2072 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2074 * Remove a PDP page from the pml4e. This can only occur
2075 * with user page tables. We do not have to lock the
2076 * pml4 PV so just ignore pvp.
2078 vm_pindex_t pml4_pindex;
2079 vm_pindex_t pdp_index;
2082 pdp_index = ptepindex - pmap_pdp_pindex(0);
2084 pml4_pindex = pmap_pml4_pindex();
2085 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2089 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2090 KKASSERT((*pdp & PG_V) != 0);
2091 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2093 KKASSERT(info == NULL);
2094 } else if (ptepindex >= pmap_pd_pindex(0)) {
2096 * Remove a PD page from the pdp
2098 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2099 * of a simple pmap because it stops at
2102 vm_pindex_t pdp_pindex;
2103 vm_pindex_t pd_index;
2106 pd_index = ptepindex - pmap_pd_pindex(0);
2109 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2110 (pd_index >> NPML4EPGSHIFT);
2111 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2116 pd = pv_pte_lookup(pvp, pd_index &
2117 ((1ul << NPDPEPGSHIFT) - 1));
2118 KKASSERT((*pd & PG_V) != 0);
2119 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2122 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2123 p = pv->pv_m; /* degenerate test later */
2125 KKASSERT(info == NULL);
2126 } else if (ptepindex >= pmap_pt_pindex(0)) {
2128 * Remove a PT page from the pd
2130 vm_pindex_t pd_pindex;
2131 vm_pindex_t pt_index;
2134 pt_index = ptepindex - pmap_pt_pindex(0);
2137 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2138 (pt_index >> NPDPEPGSHIFT);
2139 pvp = pv_get(pv->pv_pmap, pd_pindex);
2143 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2144 KKASSERT((*pt & PG_V) != 0);
2145 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2147 KKASSERT(info == NULL);
2150 * Remove a PTE from the PT page
2152 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2153 * pv is a pte_pv so we can safely lock pt_pv.
2155 vm_pindex_t pt_pindex;
2160 pt_pindex = ptepindex >> NPTEPGSHIFT;
2161 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2163 if (ptepindex >= NUPTE_USER) {
2164 ptep = vtopte(ptepindex << PAGE_SHIFT);
2165 KKASSERT(pvp == NULL);
2168 pt_pindex = NUPTE_TOTAL +
2169 (ptepindex >> NPDPEPGSHIFT);
2170 pvp = pv_get(pv->pv_pmap, pt_pindex);
2174 ptep = pv_pte_lookup(pvp, ptepindex &
2175 ((1ul << NPDPEPGSHIFT) - 1));
2179 pmap_inval_interlock(info, pmap, va);
2180 pte = pte_load_clear(ptep);
2182 pmap_inval_deinterlock(info, pmap);
2184 cpu_invlpg((void *)va);
2187 * Now update the vm_page_t
2189 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
2190 kprintf("remove_pte badpte %016lx %016lx %d\n",
2192 pv->pv_pindex < pmap_pt_pindex(0));
2194 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2195 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2198 if (pmap_track_modified(ptepindex))
2202 vm_page_flag_set(p, PG_REFERENCED);
2205 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2207 cpu_invlpg((void *)va);
2211 * Unwire the parent page table page. The wire_count cannot go below
2212 * 1 here because the parent page table page is itself still mapped.
2214 * XXX remove the assertions later.
2216 KKASSERT(pv->pv_m == p);
2217 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2218 panic("pmap_remove_pv_pte: Insufficient wire_count");
2226 pmap_remove_pv_page(pv_entry_t pv)
2232 vm_page_spin_lock(m);
2234 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2237 atomic_add_int(&m->object->agg_pv_list_count, -1);
2239 if (TAILQ_EMPTY(&m->md.pv_list))
2240 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2241 vm_page_spin_unlock(m);
2246 * Grow the number of kernel page table entries, if needed.
2248 * This routine is always called to validate any address space
2249 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2250 * space below KERNBASE.
2253 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2256 vm_offset_t ptppaddr;
2258 pd_entry_t *pt, newpt;
2260 int update_kernel_vm_end;
2263 * bootstrap kernel_vm_end on first real VM use
2265 if (kernel_vm_end == 0) {
2266 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2268 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
2269 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2270 ~(PAGE_SIZE * NPTEPG - 1);
2272 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2273 kernel_vm_end = kernel_map.max_offset;
2280 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2281 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2282 * do not want to force-fill 128G worth of page tables.
2284 if (kstart < KERNBASE) {
2285 if (kstart > kernel_vm_end)
2286 kstart = kernel_vm_end;
2287 KKASSERT(kend <= KERNBASE);
2288 update_kernel_vm_end = 1;
2290 update_kernel_vm_end = 0;
2293 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2294 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2296 if (kend - 1 >= kernel_map.max_offset)
2297 kend = kernel_map.max_offset;
2299 while (kstart < kend) {
2300 pt = pmap_pt(&kernel_pmap, kstart);
2302 /* We need a new PDP entry */
2303 nkpg = vm_page_alloc(NULL, nkpt,
2306 VM_ALLOC_INTERRUPT);
2308 panic("pmap_growkernel: no memory to grow "
2311 paddr = VM_PAGE_TO_PHYS(nkpg);
2312 if ((nkpg->flags & PG_ZERO) == 0)
2313 pmap_zero_page(paddr);
2314 vm_page_flag_clear(nkpg, PG_ZERO);
2315 newpd = (pdp_entry_t)
2316 (paddr | PG_V | PG_RW | PG_A | PG_M);
2317 *pmap_pd(&kernel_pmap, kstart) = newpd;
2319 continue; /* try again */
2321 if ((*pt & PG_V) != 0) {
2322 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2323 ~(PAGE_SIZE * NPTEPG - 1);
2324 if (kstart - 1 >= kernel_map.max_offset) {
2325 kstart = kernel_map.max_offset;
2332 * This index is bogus, but out of the way
2334 nkpg = vm_page_alloc(NULL, nkpt,
2337 VM_ALLOC_INTERRUPT);
2339 panic("pmap_growkernel: no memory to grow kernel");
2342 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2343 pmap_zero_page(ptppaddr);
2344 vm_page_flag_clear(nkpg, PG_ZERO);
2345 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2346 *pmap_pt(&kernel_pmap, kstart) = newpt;
2349 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2350 ~(PAGE_SIZE * NPTEPG - 1);
2352 if (kstart - 1 >= kernel_map.max_offset) {
2353 kstart = kernel_map.max_offset;
2359 * Only update kernel_vm_end for areas below KERNBASE.
2361 if (update_kernel_vm_end && kernel_vm_end < kstart)
2362 kernel_vm_end = kstart;
2366 * Add a reference to the specified pmap.
2369 pmap_reference(pmap_t pmap)
2372 lwkt_gettoken(&pmap->pm_token);
2374 lwkt_reltoken(&pmap->pm_token);
2378 /***************************************************
2379 * page management routines.
2380 ***************************************************/
2383 * Hold a pv without locking it
2386 pv_hold(pv_entry_t pv)
2390 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2394 count = pv->pv_hold;
2396 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2403 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2404 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2407 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2408 * pv list via its page) must be held by the caller.
2411 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2415 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2418 pv->pv_line = lineno;
2424 count = pv->pv_hold;
2426 if ((count & PV_HOLD_LOCKED) == 0) {
2427 if (atomic_cmpset_int(&pv->pv_hold, count,
2428 (count + 1) | PV_HOLD_LOCKED)) {
2431 pv->pv_line = lineno;
2436 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2444 * Drop a previously held pv_entry which could not be locked, allowing its
2447 * Must not be called with a spinlock held as we might zfree() the pv if it
2448 * is no longer associated with a pmap and this was the last hold count.
2451 pv_drop(pv_entry_t pv)
2455 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2456 if (pv->pv_pmap == NULL)
2462 count = pv->pv_hold;
2464 KKASSERT((count & PV_HOLD_MASK) > 0);
2465 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2466 (PV_HOLD_LOCKED | 1));
2467 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2468 if (count == 1 && pv->pv_pmap == NULL)
2477 * Find or allocate the requested PV entry, returning a locked pv
2481 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2484 pv_entry_t pnew = NULL;
2486 spin_lock(&pmap->pm_spin);
2488 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2489 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2494 spin_unlock(&pmap->pm_spin);
2495 pnew = zalloc(pvzone);
2496 spin_lock(&pmap->pm_spin);
2499 pnew->pv_pmap = pmap;
2500 pnew->pv_pindex = pindex;
2501 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2503 pnew->pv_func = func;
2504 pnew->pv_line = lineno;
2506 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2507 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2508 spin_unlock(&pmap->pm_spin);
2513 spin_unlock(&pmap->pm_spin);
2514 zfree(pvzone, pnew);
2516 spin_lock(&pmap->pm_spin);
2519 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2520 spin_unlock(&pmap->pm_spin);
2524 spin_unlock(&pmap->pm_spin);
2525 _pv_lock(pv PMAP_DEBUG_COPY);
2526 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2531 spin_lock(&pmap->pm_spin);
2538 * Find the requested PV entry, returning a locked+held pv or NULL
2542 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2546 spin_lock(&pmap->pm_spin);
2551 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2552 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2556 spin_unlock(&pmap->pm_spin);
2559 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2560 pv_cache(pv, pindex);
2561 spin_unlock(&pmap->pm_spin);
2564 spin_unlock(&pmap->pm_spin);
2565 _pv_lock(pv PMAP_DEBUG_COPY);
2566 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2569 spin_lock(&pmap->pm_spin);
2574 * Lookup, hold, and attempt to lock (pmap,pindex).
2576 * If the entry does not exist NULL is returned and *errorp is set to 0
2578 * If the entry exists and could be successfully locked it is returned and
2579 * errorp is set to 0.
2581 * If the entry exists but could NOT be successfully locked it is returned
2582 * held and *errorp is set to 1.
2586 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2590 spin_lock(&pmap->pm_spin);
2591 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2592 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2594 spin_unlock(&pmap->pm_spin);
2598 if (pv_hold_try(pv)) {
2599 pv_cache(pv, pindex);
2600 spin_unlock(&pmap->pm_spin);
2602 return(pv); /* lock succeeded */
2604 spin_unlock(&pmap->pm_spin);
2606 return (pv); /* lock failed */
2610 * Find the requested PV entry, returning a held pv or NULL
2614 pv_find(pmap_t pmap, vm_pindex_t pindex)
2618 spin_lock(&pmap->pm_spin);
2620 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2621 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2623 spin_unlock(&pmap->pm_spin);
2627 pv_cache(pv, pindex);
2628 spin_unlock(&pmap->pm_spin);
2633 * Lock a held pv, keeping the hold count
2637 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2642 count = pv->pv_hold;
2644 if ((count & PV_HOLD_LOCKED) == 0) {
2645 if (atomic_cmpset_int(&pv->pv_hold, count,
2646 count | PV_HOLD_LOCKED)) {
2649 pv->pv_line = lineno;
2655 tsleep_interlock(pv, 0);
2656 if (atomic_cmpset_int(&pv->pv_hold, count,
2657 count | PV_HOLD_WAITING)) {
2659 kprintf("pv waiting on %s:%d\n",
2660 pv->pv_func, pv->pv_line);
2662 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2669 * Unlock a held and locked pv, keeping the hold count.
2673 pv_unlock(pv_entry_t pv)
2677 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2681 count = pv->pv_hold;
2683 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2684 (PV_HOLD_LOCKED | 1));
2685 if (atomic_cmpset_int(&pv->pv_hold, count,
2687 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2688 if (count & PV_HOLD_WAITING)
2696 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2697 * and the hold count drops to zero we will free it.
2699 * Caller should not hold any spin locks. We are protected from hold races
2700 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2701 * lock held. A pv cannot be located otherwise.
2705 pv_put(pv_entry_t pv)
2707 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2708 if (pv->pv_pmap == NULL)
2717 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2718 * pmap. Any pte operations must have already been completed.
2722 pv_free(pv_entry_t pv)
2726 KKASSERT(pv->pv_m == NULL);
2727 if ((pmap = pv->pv_pmap) != NULL) {
2728 spin_lock(&pmap->pm_spin);
2729 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2730 if (pmap->pm_pvhint == pv)
2731 pmap->pm_pvhint = NULL;
2732 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2735 spin_unlock(&pmap->pm_spin);
2741 * This routine is very drastic, but can save the system
2749 static int warningdone=0;
2751 if (pmap_pagedaemon_waken == 0)
2753 pmap_pagedaemon_waken = 0;
2754 if (warningdone < 5) {
2755 kprintf("pmap_collect: collecting pv entries -- "
2756 "suggest increasing PMAP_SHPGPERPROC\n");
2760 for (i = 0; i < vm_page_array_size; i++) {
2761 m = &vm_page_array[i];
2762 if (m->wire_count || m->hold_count)
2764 if (vm_page_busy_try(m, TRUE) == 0) {
2765 if (m->wire_count == 0 && m->hold_count == 0) {
2774 * Scan the pmap for active page table entries and issue a callback.
2775 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
2776 * its parent page table.
2778 * pte_pv will be NULL if the page or page table is unmanaged.
2779 * pt_pv will point to the page table page containing the pte for the page.
2781 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
2782 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
2783 * process pmap's PD and page to the callback function. This can be
2784 * confusing because the pt_pv is really a pd_pv, and the target page
2785 * table page is simply aliased by the pmap and not owned by it.
2787 * It is assumed that the start and end are properly rounded to the page size.
2789 * It is assumed that PD pages and above are managed and thus in the RB tree,
2790 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
2792 struct pmap_scan_info {
2796 vm_pindex_t sva_pd_pindex;
2797 vm_pindex_t eva_pd_pindex;
2798 void (*func)(pmap_t, struct pmap_scan_info *,
2799 pv_entry_t, pv_entry_t, int, vm_offset_t,
2800 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,
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, 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, 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) {
3017 if (sva >= VM_MAX_USER_ADDRESS) {
3026 * PD cache (degenerate case if we skip). It is possible
3027 * for the PD to not exist due to races. This is ok.
3029 if (pd_pv == NULL) {
3030 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3031 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3033 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3035 if (pd_pv == NULL) {
3036 va_next = (sva + NBPDP) & ~PDPMASK;
3045 if (pt_pv == NULL) {
3050 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3051 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3057 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3061 * If pt_pv is NULL we either have an shared page table
3062 * page and must issue a callback specific to that case,
3063 * or there is no page table page.
3065 * Either way we can skip the page table page.
3067 if (pt_pv == NULL) {
3069 * Possible unmanaged (shared from another pmap)
3073 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3074 KKASSERT(pd_pv != NULL);
3075 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3077 info->func(pmap, info, NULL, pd_pv, 1,
3078 sva, ptep, info->arg);
3082 * Done, move to next page table page.
3084 va_next = (sva + NBPDR) & ~PDRMASK;
3091 * From this point in the loop testing pt_pv for non-NULL
3092 * means we are in UVM, else if it is NULL we are in KVM.
3094 * Limit our scan to either the end of the va represented
3095 * by the current page table page, or to the end of the
3096 * range being removed.
3099 va_next = (sva + NBPDR) & ~PDRMASK;
3106 * Scan the page table for pages. Some pages may not be
3107 * managed (might not have a pv_entry).
3109 * There is no page table management for kernel pages so
3110 * pt_pv will be NULL in that case, but otherwise pt_pv
3111 * is non-NULL, locked, and referenced.
3115 * At this point a non-NULL pt_pv means a UVA, and a NULL
3116 * pt_pv means a KVA.
3119 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3123 while (sva < va_next) {
3125 * Acquire the related pte_pv, if any. If *ptep == 0
3126 * the related pte_pv should not exist, but if *ptep
3127 * is not zero the pte_pv may or may not exist (e.g.
3128 * will not exist for an unmanaged page).
3130 * However a multitude of races are possible here.
3132 * In addition, the (pt_pv, pte_pv) lock order is
3133 * backwards, so we have to be careful in aquiring
3134 * a properly locked pte_pv.
3137 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3144 pv_put(pt_pv); /* must be non-NULL */
3146 pv_lock(pte_pv); /* safe to block now */
3149 pt_pv = pv_get(pmap,
3150 pmap_pt_pindex(sva));
3152 * pt_pv reloaded, need new ptep
3154 KKASSERT(pt_pv != NULL);
3155 ptep = pv_pte_lookup(pt_pv,
3156 pmap_pte_index(sva));
3160 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3164 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3168 kprintf("Unexpected non-NULL pte_pv "
3169 "%p pt_pv %p *ptep = %016lx\n",
3170 pte_pv, pt_pv, *ptep);
3171 panic("Unexpected non-NULL pte_pv");
3179 * Ready for the callback. The locked pte_pv (if any)
3180 * is consumed by the callback. pte_pv will exist if
3181 * the page is managed, and will not exist if it
3185 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3187 ("bad *ptep %016lx sva %016lx "
3189 *ptep, sva, pte_pv));
3190 info->func(pmap, info, pte_pv, pt_pv, 0,
3191 sva, ptep, info->arg);
3193 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3195 ("bad *ptep %016lx sva %016lx "
3198 info->func(pmap, info, NULL, pt_pv, 0,
3199 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;
3234 info.doinval = 1; /* normal remove requires pmap inval */
3239 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3241 struct pmap_scan_info info;
3246 info.func = pmap_remove_callback;
3248 info.doinval = 0; /* normal remove requires pmap inval */
3253 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
3254 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3255 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3261 * This will also drop pt_pv's wire_count. Note that
3262 * terminal pages are not wired based on mmu presence.
3265 pmap_remove_pv_pte(pte_pv, pt_pv, &info->inval);
3267 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3268 pmap_remove_pv_page(pte_pv);
3270 } else if (sharept == 0) {
3274 * pt_pv's wire_count is still bumped by unmanaged pages
3275 * so we must decrement it manually.
3278 pmap_inval_interlock(&info->inval, pmap, va);
3279 pte = pte_load_clear(ptep);
3281 pmap_inval_deinterlock(&info->inval, pmap);
3283 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3284 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3285 if (vm_page_unwire_quick(pt_pv->pv_m))
3286 panic("pmap_remove: insufficient wirecount");
3289 * Unmanaged page table, pt_pv is actually the pd_pv
3290 * for our pmap (not the share object pmap).
3292 * We have to unwire the target page table page and we
3293 * have to unwire our page directory page.
3296 pmap_inval_interlock(&info->inval, pmap, va);
3297 pte = pte_load_clear(ptep);
3299 pmap_inval_deinterlock(&info->inval, pmap);
3300 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3301 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3302 panic("pmap_remove: shared pgtable1 bad wirecount");
3303 if (vm_page_unwire_quick(pt_pv->pv_m))
3304 panic("pmap_remove: shared pgtable2 bad wirecount");
3309 * Removes this physical page from all physical maps in which it resides.
3310 * Reflects back modify bits to the pager.
3312 * This routine may not be called from an interrupt.
3316 pmap_remove_all(vm_page_t m)
3318 struct pmap_inval_info info;
3321 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3324 pmap_inval_init(&info);
3325 vm_page_spin_lock(m);
3326 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3327 KKASSERT(pv->pv_m == m);
3328 if (pv_hold_try(pv)) {
3329 vm_page_spin_unlock(m);
3331 vm_page_spin_unlock(m);
3333 if (pv->pv_m != m) {
3335 vm_page_spin_lock(m);
3340 * Holding no spinlocks, pv is locked.
3342 pmap_remove_pv_pte(pv, NULL, &info);
3343 pmap_remove_pv_page(pv);
3345 vm_page_spin_lock(m);
3347 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3348 vm_page_spin_unlock(m);
3349 pmap_inval_done(&info);
3353 * Set the physical protection on the specified range of this map
3354 * as requested. This function is typically only used for debug watchpoints
3357 * This function may not be called from an interrupt if the map is
3358 * not the kernel_pmap.
3360 * NOTE! For shared page table pages we just unmap the page.
3363 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3365 struct pmap_scan_info info;
3366 /* JG review for NX */
3370 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3371 pmap_remove(pmap, sva, eva);
3374 if (prot & VM_PROT_WRITE)
3379 info.func = pmap_protect_callback;
3387 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
3388 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3389 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3399 pmap_inval_interlock(&info->inval, pmap, va);
3406 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3407 KKASSERT(m == pte_pv->pv_m);
3408 vm_page_flag_set(m, PG_REFERENCED);
3412 if (pmap_track_modified(pte_pv->pv_pindex)) {
3414 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3419 } else if (sharept) {
3421 * Unmanaged page table, pt_pv is actually the pd_pv
3422 * for our pmap (not the share object pmap).
3424 * When asked to protect something in a shared page table
3425 * page we just unmap the page table page. We have to
3426 * invalidate the tlb in this situation.
3428 pte = pte_load_clear(ptep);
3429 pmap_inval_invltlb(&info->inval);
3430 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3431 panic("pmap_protect: pgtable1 pg bad wirecount");
3432 if (vm_page_unwire_quick(pt_pv->pv_m))
3433 panic("pmap_protect: pgtable2 pg bad wirecount");
3436 /* else unmanaged page, adjust bits, no wire changes */
3440 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
3444 pmap_inval_deinterlock(&info->inval, pmap);
3450 * Insert the vm_page (m) at the virtual address (va), replacing any prior
3451 * mapping at that address. Set protection and wiring as requested.
3453 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
3454 * possible. If it is we enter the page into the appropriate shared pmap
3455 * hanging off the related VM object instead of the passed pmap, then we
3456 * share the page table page from the VM object's pmap into the current pmap.
3458 * NOTE: This routine MUST insert the page into the pmap now, it cannot
3462 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3463 boolean_t wired, vm_map_entry_t entry __unused)
3465 pmap_inval_info info;
3466 pv_entry_t pt_pv; /* page table */
3467 pv_entry_t pte_pv; /* page table entry */
3470 pt_entry_t origpte, newpte;
3475 va = trunc_page(va);
3476 #ifdef PMAP_DIAGNOSTIC
3478 panic("pmap_enter: toobig");
3479 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
3480 panic("pmap_enter: invalid to pmap_enter page table "
3481 "pages (va: 0x%lx)", va);
3483 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
3484 kprintf("Warning: pmap_enter called on UVA with "
3487 db_print_backtrace();
3490 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
3491 kprintf("Warning: pmap_enter called on KVA without"
3494 db_print_backtrace();
3499 * Get locked PV entries for our new page table entry (pte_pv)
3500 * and for its parent page table (pt_pv). We need the parent
3501 * so we can resolve the location of the ptep.
3503 * Only hardware MMU actions can modify the ptep out from
3506 * if (m) is fictitious or unmanaged we do not create a managing
3507 * pte_pv for it. Any pre-existing page's management state must
3508 * match (avoiding code complexity).
3510 * If the pmap is still being initialized we assume existing
3513 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3514 * pmap_allocpte() checks the
3516 if (pmap_initialized == FALSE) {
3520 } else if (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) { /* XXX */
3522 if (va >= VM_MAX_USER_ADDRESS) {
3526 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
3528 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3530 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED) == 0);
3532 if (va >= VM_MAX_USER_ADDRESS) {
3534 * Kernel map, pv_entry-tracked.
3537 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3543 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
3545 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3547 KKASSERT(*ptep == 0 || (*ptep & PG_MANAGED));
3550 pa = VM_PAGE_TO_PHYS(m);
3552 opa = origpte & PG_FRAME;
3554 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V | PG_A);
3557 if (va < VM_MAX_USER_ADDRESS)
3560 newpte |= PG_MANAGED;
3561 if (pmap == &kernel_pmap)
3565 * It is possible for multiple faults to occur in threaded
3566 * environments, the existing pte might be correct.
3568 if (((origpte ^ newpte) & ~(pt_entry_t)(PG_M|PG_A)) == 0)
3571 if ((prot & VM_PROT_NOSYNC) == 0)
3572 pmap_inval_init(&info);
3575 * Ok, either the address changed or the protection or wiring
3578 * Clear the current entry, interlocking the removal. For managed
3579 * pte's this will also flush the modified state to the vm_page.
3580 * Atomic ops are mandatory in order to ensure that PG_M events are
3581 * not lost during any transition.
3586 * pmap_remove_pv_pte() unwires pt_pv and assumes
3587 * we will free pte_pv, but since we are reusing
3588 * pte_pv we want to retain the wire count.
3590 * pt_pv won't exist for a kernel page (managed or
3594 vm_page_wire_quick(pt_pv->pv_m);
3595 if (prot & VM_PROT_NOSYNC)
3596 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3598 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3600 pmap_remove_pv_page(pte_pv);
3601 } else if (prot & VM_PROT_NOSYNC) {
3603 * Unmanaged page, NOSYNC (no mmu sync) requested.
3605 * Leave wire count on PT page intact.
3607 (void)pte_load_clear(ptep);
3608 cpu_invlpg((void *)va);
3609 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3612 * Unmanaged page, normal enter.
3614 * Leave wire count on PT page intact.
3616 pmap_inval_interlock(&info, pmap, va);
3617 (void)pte_load_clear(ptep);
3618 pmap_inval_deinterlock(&info, pmap);
3619 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3621 KKASSERT(*ptep == 0);
3626 * Enter on the PV list if part of our managed memory.
3627 * Wiring of the PT page is already handled.
3629 KKASSERT(pte_pv->pv_m == NULL);
3630 vm_page_spin_lock(m);
3632 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3635 atomic_add_int(&m->object->agg_pv_list_count, 1);
3637 vm_page_flag_set(m, PG_MAPPED);
3638 vm_page_spin_unlock(m);
3639 } else if (pt_pv && opa == 0) {
3641 * We have to adjust the wire count on the PT page ourselves
3642 * for unmanaged entries. If opa was non-zero we retained
3643 * the existing wire count from the removal.
3645 vm_page_wire_quick(pt_pv->pv_m);
3649 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
3651 * User VMAs do not because those will be zero->non-zero, so no
3652 * stale entries to worry about at this point.
3654 * For KVM there appear to still be issues. Theoretically we
3655 * should be able to scrap the interlocks entirely but we
3658 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3659 pmap_inval_interlock(&info, pmap, va);
3664 *(volatile pt_entry_t *)ptep = newpte;
3666 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3667 pmap_inval_deinterlock(&info, pmap);
3668 else if (pt_pv == NULL)
3669 cpu_invlpg((void *)va);
3673 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
3676 atomic_add_long(&pmap->pm_stats.wired_count, 1);
3680 vm_page_flag_set(m, PG_WRITEABLE);
3683 * Unmanaged pages need manual resident_count tracking.
3685 if (pte_pv == NULL && pt_pv)
3686 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
3691 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
3692 pmap_inval_done(&info);
3694 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
3697 * Cleanup the pv entry, allowing other accessors.
3706 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
3707 * This code also assumes that the pmap has no pre-existing entry for this
3710 * This code currently may only be used on user pmaps, not kernel_pmap.
3713 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
3715 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
3719 * Make a temporary mapping for a physical address. This is only intended
3720 * to be used for panic dumps.
3722 * The caller is responsible for calling smp_invltlb().
3725 pmap_kenter_temporary(vm_paddr_t pa, long i)
3727 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
3728 return ((void *)crashdumpmap);
3731 #define MAX_INIT_PT (96)
3734 * This routine preloads the ptes for a given object into the specified pmap.
3735 * This eliminates the blast of soft faults on process startup and
3736 * immediately after an mmap.
3738 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
3741 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
3742 vm_object_t object, vm_pindex_t pindex,
3743 vm_size_t size, int limit)
3745 struct rb_vm_page_scan_info info;
3750 * We can't preinit if read access isn't set or there is no pmap
3753 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
3757 * We can't preinit if the pmap is not the current pmap
3759 lp = curthread->td_lwp;
3760 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
3764 * Misc additional checks
3766 psize = x86_64_btop(size);
3768 if ((object->type != OBJT_VNODE) ||
3769 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
3770 (object->resident_page_count > MAX_INIT_PT))) {
3774 if (pindex + psize > object->size) {
3775 if (object->size < pindex)
3777 psize = object->size - pindex;
3784 * If everything is segment-aligned do not pre-init here. Instead
3785 * allow the normal vm_fault path to pass a segment hint to
3786 * pmap_enter() which will then use an object-referenced shared
3789 if ((addr & SEG_MASK) == 0 &&
3790 (ctob(psize) & SEG_MASK) == 0 &&
3791 (ctob(pindex) & SEG_MASK) == 0) {
3796 * Use a red-black scan to traverse the requested range and load
3797 * any valid pages found into the pmap.
3799 * We cannot safely scan the object's memq without holding the
3802 info.start_pindex = pindex;
3803 info.end_pindex = pindex + psize - 1;
3809 vm_object_hold_shared(object);
3810 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
3811 pmap_object_init_pt_callback, &info);
3812 vm_object_drop(object);
3817 pmap_object_init_pt_callback(vm_page_t p, void *data)
3819 struct rb_vm_page_scan_info *info = data;
3820 vm_pindex_t rel_index;
3823 * don't allow an madvise to blow away our really
3824 * free pages allocating pv entries.
3826 if ((info->limit & MAP_PREFAULT_MADVISE) &&
3827 vmstats.v_free_count < vmstats.v_free_reserved) {
3832 * Ignore list markers and ignore pages we cannot instantly
3833 * busy (while holding the object token).
3835 if (p->flags & PG_MARKER)
3837 if (vm_page_busy_try(p, TRUE))
3839 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3840 (p->flags & PG_FICTITIOUS) == 0) {
3841 if ((p->queue - p->pc) == PQ_CACHE)
3842 vm_page_deactivate(p);
3843 rel_index = p->pindex - info->start_pindex;
3844 pmap_enter_quick(info->pmap,
3845 info->addr + x86_64_ptob(rel_index), p);
3853 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
3856 * Returns FALSE if it would be non-trivial or if a pte is already loaded
3859 * XXX This is safe only because page table pages are not freed.
3862 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
3866 /*spin_lock(&pmap->pm_spin);*/
3867 if ((pte = pmap_pte(pmap, addr)) != NULL) {
3869 /*spin_unlock(&pmap->pm_spin);*/
3873 /*spin_unlock(&pmap->pm_spin);*/
3878 * Change the wiring attribute for a pmap/va pair. The mapping must already
3879 * exist in the pmap. The mapping may or may not be managed.
3882 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
3883 vm_map_entry_t entry)
3890 lwkt_gettoken(&pmap->pm_token);
3891 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
3892 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
3894 if (wired && !pmap_pte_w(ptep))
3895 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
3896 else if (!wired && pmap_pte_w(ptep))
3897 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
3900 * Wiring is not a hardware characteristic so there is no need to
3901 * invalidate TLB. However, in an SMP environment we must use
3902 * a locked bus cycle to update the pte (if we are not using
3903 * the pmap_inval_*() API that is)... it's ok to do this for simple
3908 atomic_set_long(ptep, PG_W);
3910 atomic_clear_long(ptep, PG_W);
3913 atomic_set_long_nonlocked(ptep, PG_W);
3915 atomic_clear_long_nonlocked(ptep, PG_W);
3918 lwkt_reltoken(&pmap->pm_token);
3924 * Copy the range specified by src_addr/len from the source map to
3925 * the range dst_addr/len in the destination map.
3927 * This routine is only advisory and need not do anything.
3930 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3931 vm_size_t len, vm_offset_t src_addr)
3938 * Zero the specified physical page.
3940 * This function may be called from an interrupt and no locking is
3944 pmap_zero_page(vm_paddr_t phys)
3946 vm_offset_t va = PHYS_TO_DMAP(phys);
3948 pagezero((void *)va);
3952 * pmap_page_assertzero:
3954 * Assert that a page is empty, panic if it isn't.
3957 pmap_page_assertzero(vm_paddr_t phys)
3959 vm_offset_t va = PHYS_TO_DMAP(phys);
3962 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3963 if (*(long *)((char *)va + i) != 0) {
3964 panic("pmap_page_assertzero() @ %p not zero!",
3965 (void *)(intptr_t)va);
3973 * Zero part of a physical page by mapping it into memory and clearing
3974 * its contents with bzero.
3976 * off and size may not cover an area beyond a single hardware page.
3979 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3981 vm_offset_t virt = PHYS_TO_DMAP(phys);
3983 bzero((char *)virt + off, size);
3989 * Copy the physical page from the source PA to the target PA.
3990 * This function may be called from an interrupt. No locking
3994 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3996 vm_offset_t src_virt, dst_virt;
3998 src_virt = PHYS_TO_DMAP(src);
3999 dst_virt = PHYS_TO_DMAP(dst);
4000 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4004 * pmap_copy_page_frag:
4006 * Copy the physical page from the source PA to the target PA.
4007 * This function may be called from an interrupt. No locking
4011 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4013 vm_offset_t src_virt, dst_virt;
4015 src_virt = PHYS_TO_DMAP(src);
4016 dst_virt = PHYS_TO_DMAP(dst);
4018 bcopy((char *)src_virt + (src & PAGE_MASK),
4019 (char *)dst_virt + (dst & PAGE_MASK),
4024 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4025 * this page. This count may be changed upwards or downwards in the future;
4026 * it is only necessary that true be returned for a small subset of pmaps
4027 * for proper page aging.
4030 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4035 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4038 vm_page_spin_lock(m);
4039 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4040 if (pv->pv_pmap == pmap) {
4041 vm_page_spin_unlock(m);
4048 vm_page_spin_unlock(m);
4053 * Remove all pages from specified address space this aids process exit
4054 * speeds. Also, this code may be special cased for the current process
4058 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4060 pmap_remove_noinval(pmap, sva, eva);
4065 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4066 * routines are inline, and a lot of things compile-time evaluate.
4070 pmap_testbit(vm_page_t m, int bit)
4075 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4078 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4080 vm_page_spin_lock(m);
4081 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4082 vm_page_spin_unlock(m);
4086 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4088 * if the bit being tested is the modified bit, then
4089 * mark clean_map and ptes as never
4092 if (bit & (PG_A|PG_M)) {
4093 if (!pmap_track_modified(pv->pv_pindex))
4097 #if defined(PMAP_DIAGNOSTIC)
4098 if (pv->pv_pmap == NULL) {
4099 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4104 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4106 vm_page_spin_unlock(m);
4110 vm_page_spin_unlock(m);
4115 * This routine is used to modify bits in ptes. Only one bit should be
4116 * specified. PG_RW requires special handling.
4118 * Caller must NOT hold any spin locks
4122 pmap_clearbit(vm_page_t m, int bit)
4124 struct pmap_inval_info info;
4131 vm_page_flag_clear(m, PG_WRITEABLE);
4132 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4139 * Loop over all current mappings setting/clearing as appropos If
4140 * setting RO do we need to clear the VAC?
4142 * NOTE: When clearing PG_M we could also (not implemented) drop
4143 * through to the PG_RW code and clear PG_RW too, forcing
4144 * a fault on write to redetect PG_M for virtual kernels, but
4145 * it isn't necessary since virtual kernels invalidate the
4146 * pte when they clear the VPTE_M bit in their virtual page
4149 * NOTE: Does not re-dirty the page when clearing only PG_M.
4151 if ((bit & PG_RW) == 0) {
4152 vm_page_spin_lock(m);
4153 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4154 #if defined(PMAP_DIAGNOSTIC)
4155 if (pv->pv_pmap == NULL) {
4156 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4161 pte = pmap_pte_quick(pv->pv_pmap,
4162 pv->pv_pindex << PAGE_SHIFT);
4165 atomic_clear_long(pte, bit);
4167 vm_page_spin_unlock(m);
4172 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4175 pmap_inval_init(&info);
4178 vm_page_spin_lock(m);
4179 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4181 * don't write protect pager mappings
4183 if (!pmap_track_modified(pv->pv_pindex))
4186 #if defined(PMAP_DIAGNOSTIC)
4187 if (pv->pv_pmap == NULL) {
4188 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4194 * Skip pages which do not have PG_RW set.
4196 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4197 if ((*pte & PG_RW) == 0)
4203 if (pv_hold_try(pv) == 0) {
4204 vm_page_spin_unlock(m);
4205 pv_lock(pv); /* held, now do a blocking lock */
4206 pv_put(pv); /* and release */
4207 goto restart; /* anything could have happened */
4210 save_pmap = pv->pv_pmap;
4211 vm_page_spin_unlock(m);
4212 pmap_inval_interlock(&info, save_pmap,
4213 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
4214 KKASSERT(pv->pv_pmap == save_pmap);
4218 if (atomic_cmpset_long(pte, pbits,
4219 pbits & ~(PG_RW|PG_M))) {
4223 pmap_inval_deinterlock(&info, save_pmap);
4224 vm_page_spin_lock(m);
4227 * If PG_M was found to be set while we were clearing PG_RW
4228 * we also clear PG_M (done above) and mark the page dirty.
4229 * Callers expect this behavior.
4235 vm_page_spin_unlock(m);
4236 pmap_inval_done(&info);
4240 * Lower the permission for all mappings to a given page.
4242 * Page must be busied by caller.
4245 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4247 /* JG NX support? */
4248 if ((prot & VM_PROT_WRITE) == 0) {
4249 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4251 * NOTE: pmap_clearbit(.. PG_RW) also clears
4252 * the PG_WRITEABLE flag in (m).
4254 pmap_clearbit(m, PG_RW);
4262 pmap_phys_address(vm_pindex_t ppn)
4264 return (x86_64_ptob(ppn));
4268 * Return a count of reference bits for a page, clearing those bits.
4269 * It is not necessary for every reference bit to be cleared, but it
4270 * is necessary that 0 only be returned when there are truly no
4271 * reference bits set.
4273 * XXX: The exact number of bits to check and clear is a matter that
4274 * should be tested and standardized at some point in the future for
4275 * optimal aging of shared pages.
4277 * This routine may not block.
4280 pmap_ts_referenced(vm_page_t m)
4286 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4289 vm_page_spin_lock(m);
4290 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4291 if (!pmap_track_modified(pv->pv_pindex))
4293 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4294 if (pte && (*pte & PG_A)) {
4296 atomic_clear_long(pte, PG_A);
4298 atomic_clear_long_nonlocked(pte, PG_A);
4305 vm_page_spin_unlock(m);
4312 * Return whether or not the specified physical page was modified
4313 * in any physical maps.
4316 pmap_is_modified(vm_page_t m)
4320 res = pmap_testbit(m, PG_M);
4325 * Clear the modify bits on the specified physical page.
4328 pmap_clear_modify(vm_page_t m)
4330 pmap_clearbit(m, PG_M);
4334 * pmap_clear_reference:
4336 * Clear the reference bit on the specified physical page.
4339 pmap_clear_reference(vm_page_t m)
4341 pmap_clearbit(m, PG_A);
4345 * Miscellaneous support routines follow
4350 i386_protection_init(void)
4354 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4355 kp = protection_codes;
4356 for (prot = 0; prot < 8; prot++) {
4358 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4360 * Read access is also 0. There isn't any execute bit,
4361 * so just make it readable.
4363 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4364 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4365 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4368 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4369 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4370 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4371 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4379 * Map a set of physical memory pages into the kernel virtual
4380 * address space. Return a pointer to where it is mapped. This
4381 * routine is intended to be used for mapping device memory,
4384 * NOTE: we can't use pgeflag unless we invalidate the pages one at
4388 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
4390 vm_offset_t va, tmpva, offset;
4393 offset = pa & PAGE_MASK;
4394 size = roundup(offset + size, PAGE_SIZE);
4396 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4398 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4400 pa = pa & ~PAGE_MASK;
4401 for (tmpva = va; size > 0;) {
4402 pte = vtopte(tmpva);
4403 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
4411 return ((void *)(va + offset));
4415 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
4417 vm_offset_t va, tmpva, offset;
4420 offset = pa & PAGE_MASK;
4421 size = roundup(offset + size, PAGE_SIZE);
4423 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4425 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4427 pa = pa & ~PAGE_MASK;
4428 for (tmpva = va; size > 0;) {
4429 pte = vtopte(tmpva);
4430 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
4438 return ((void *)(va + offset));
4442 pmap_unmapdev(vm_offset_t va, vm_size_t size)
4444 vm_offset_t base, offset;
4446 base = va & ~PAGE_MASK;
4447 offset = va & PAGE_MASK;
4448 size = roundup(offset + size, PAGE_SIZE);
4449 pmap_qremove(va, size >> PAGE_SHIFT);
4450 kmem_free(&kernel_map, base, size);
4454 * perform the pmap work for mincore
4457 pmap_mincore(pmap_t pmap, vm_offset_t addr)
4459 pt_entry_t *ptep, pte;
4463 lwkt_gettoken(&pmap->pm_token);
4464 ptep = pmap_pte(pmap, addr);
4466 if (ptep && (pte = *ptep) != 0) {
4469 val = MINCORE_INCORE;
4470 if ((pte & PG_MANAGED) == 0)
4473 pa = pte & PG_FRAME;
4475 m = PHYS_TO_VM_PAGE(pa);
4481 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
4483 * Modified by someone
4485 else if (m->dirty || pmap_is_modified(m))
4486 val |= MINCORE_MODIFIED_OTHER;
4491 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
4494 * Referenced by someone
4496 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
4497 val |= MINCORE_REFERENCED_OTHER;
4498 vm_page_flag_set(m, PG_REFERENCED);
4502 lwkt_reltoken(&pmap->pm_token);
4508 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
4509 * vmspace will be ref'd and the old one will be deref'd.
4511 * The vmspace for all lwps associated with the process will be adjusted
4512 * and cr3 will be reloaded if any lwp is the current lwp.
4514 * The process must hold the vmspace->vm_map.token for oldvm and newvm
4517 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
4519 struct vmspace *oldvm;
4522 oldvm = p->p_vmspace;
4523 if (oldvm != newvm) {
4525 sysref_get(&newvm->vm_sysref);
4526 p->p_vmspace = newvm;
4527 KKASSERT(p->p_nthreads == 1);
4528 lp = RB_ROOT(&p->p_lwp_tree);
4529 pmap_setlwpvm(lp, newvm);
4531 sysref_put(&oldvm->vm_sysref);
4536 * Set the vmspace for a LWP. The vmspace is almost universally set the
4537 * same as the process vmspace, but virtual kernels need to swap out contexts
4538 * on a per-lwp basis.
4540 * Caller does not necessarily hold any vmspace tokens. Caller must control
4541 * the lwp (typically be in the context of the lwp). We use a critical
4542 * section to protect against statclock and hardclock (statistics collection).
4545 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
4547 struct vmspace *oldvm;
4550 oldvm = lp->lwp_vmspace;
4552 if (oldvm != newvm) {
4554 lp->lwp_vmspace = newvm;
4555 if (curthread->td_lwp == lp) {
4556 pmap = vmspace_pmap(newvm);
4558 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4559 if (pmap->pm_active & CPUMASK_LOCK)
4560 pmap_interlock_wait(newvm);
4562 pmap->pm_active |= 1;
4564 #if defined(SWTCH_OPTIM_STATS)
4567 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4568 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
4569 load_cr3(curthread->td_pcb->pcb_cr3);
4570 pmap = vmspace_pmap(oldvm);
4572 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4574 pmap->pm_active &= ~(cpumask_t)1;
4584 * Called when switching to a locked pmap, used to interlock against pmaps
4585 * undergoing modifications to prevent us from activating the MMU for the
4586 * target pmap until all such modifications have completed. We have to do
4587 * this because the thread making the modifications has already set up its
4588 * SMP synchronization mask.
4590 * This function cannot sleep!
4595 pmap_interlock_wait(struct vmspace *vm)
4597 struct pmap *pmap = &vm->vm_pmap;
4599 if (pmap->pm_active & CPUMASK_LOCK) {
4601 KKASSERT(curthread->td_critcount >= 2);
4602 DEBUG_PUSH_INFO("pmap_interlock_wait");
4603 while (pmap->pm_active & CPUMASK_LOCK) {
4605 lwkt_process_ipiq();
4615 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
4618 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
4622 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
4627 * Used by kmalloc/kfree, page already exists at va
4630 pmap_kvtom(vm_offset_t va)
4632 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));
4636 * Initialize machine-specific shared page directory support. This
4637 * is executed when a VM object is created.
4640 pmap_object_init(vm_object_t object)
4642 object->md.pmap_rw = NULL;
4643 object->md.pmap_ro = NULL;
4647 * Clean up machine-specific shared page directory support. This
4648 * is executed when a VM object is destroyed.
4651 pmap_object_free(vm_object_t object)
4655 if ((pmap = object->md.pmap_rw) != NULL) {
4656 object->md.pmap_rw = NULL;
4657 pmap_remove_noinval(pmap,
4658 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
4659 pmap->pm_active = 0;
4662 kfree(pmap, M_OBJPMAP);
4664 if ((pmap = object->md.pmap_ro) != NULL) {
4665 object->md.pmap_ro = NULL;
4666 pmap_remove_noinval(pmap,
4667 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
4668 pmap->pm_active = 0;
4671 kfree(pmap, M_OBJPMAP);