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/kernel.h>
57 #include <sys/msgbuf.h>
58 #include <sys/vmmeter.h>
60 #include <sys/systm.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(pmap, pte) ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
141 #define pmap_pte_w(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
142 #define pmap_pte_m(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
143 #define pmap_pte_u(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
144 #define pmap_pte_v(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 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 (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
152 static int protection_codes[PROTECTION_CODES_SIZE];
154 struct pmap kernel_pmap;
155 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
159 vm_paddr_t avail_start; /* PA of first available physical page */
160 vm_paddr_t avail_end; /* PA of last available physical page */
161 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
162 vm_offset_t virtual2_end;
163 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
164 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
165 vm_offset_t KvaStart; /* VA start of KVA space */
166 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
167 vm_offset_t KvaSize; /* max size of kernel virtual address space */
168 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
169 //static int pgeflag; /* PG_G or-in */
170 //static int pseflag; /* PG_PS or-in */
174 static vm_paddr_t dmaplimit;
176 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
178 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE]; /* PAT -> PG_ bits */
179 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
181 static uint64_t KPTbase;
182 static uint64_t KPTphys;
183 static uint64_t KPDphys; /* phys addr of kernel level 2 */
184 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
185 uint64_t KPDPphys; /* phys addr of kernel level 3 */
186 uint64_t KPML4phys; /* phys addr of kernel level 4 */
188 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
189 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
192 * Data for the pv entry allocation mechanism
194 static vm_zone_t pvzone;
195 static struct vm_zone pvzone_store;
196 static struct vm_object pvzone_obj;
197 static int pv_entry_max=0, pv_entry_high_water=0;
198 static int pmap_pagedaemon_waken = 0;
199 static struct pv_entry *pvinit;
202 * All those kernel PT submaps that BSD is so fond of
204 pt_entry_t *CMAP1 = NULL, *ptmmap;
205 caddr_t CADDR1 = NULL, ptvmmap = NULL;
206 static pt_entry_t *msgbufmap;
207 struct msgbuf *msgbufp=NULL;
210 * PMAP default PG_* bits. Needed to be able to add
211 * EPT/NPT pagetable pmap_bits for the VMM module
213 uint64_t pmap_bits_default[] = {
214 REGULAR_PMAP, /* TYPE_IDX 0 */
215 X86_PG_V, /* PG_V_IDX 1 */
216 X86_PG_RW, /* PG_RW_IDX 2 */
217 X86_PG_U, /* PG_U_IDX 3 */
218 X86_PG_A, /* PG_A_IDX 4 */
219 X86_PG_M, /* PG_M_IDX 5 */
220 X86_PG_PS, /* PG_PS_IDX3 6 */
221 X86_PG_G, /* PG_G_IDX 7 */
222 X86_PG_AVAIL1, /* PG_AVAIL1_IDX 8 */
223 X86_PG_AVAIL2, /* PG_AVAIL2_IDX 9 */
224 X86_PG_AVAIL3, /* PG_AVAIL3_IDX 10 */
225 X86_PG_NC_PWT | X86_PG_NC_PCD, /* PG_N_IDX 11 */
230 static pt_entry_t *pt_crashdumpmap;
231 static caddr_t crashdumpmap;
234 static int pmap_enter_debug = 0;
235 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
236 &pmap_enter_debug, 0, "Debug pmap_enter's");
238 static int pmap_yield_count = 64;
239 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
240 &pmap_yield_count, 0, "Yield during init_pt/release");
241 static int pmap_mmu_optimize = 0;
242 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
243 &pmap_mmu_optimize, 0, "Share page table pages when possible");
244 int pmap_fast_kernel_cpusync = 0;
245 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
246 &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
250 /* Standard user access funtions */
251 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
253 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
254 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
255 extern int std_fubyte (const void *base);
256 extern int std_subyte (void *base, int byte);
257 extern long std_fuword (const void *base);
258 extern int std_suword (void *base, long word);
259 extern int std_suword32 (void *base, int word);
261 static void pv_hold(pv_entry_t pv);
262 static int _pv_hold_try(pv_entry_t pv
264 static void pv_drop(pv_entry_t pv);
265 static void _pv_lock(pv_entry_t pv
267 static void pv_unlock(pv_entry_t pv);
268 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
270 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
272 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
273 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
274 static void pv_put(pv_entry_t pv);
275 static void pv_free(pv_entry_t pv);
276 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
277 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
279 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
280 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
281 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
282 struct pmap_inval_info *info);
283 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
284 static int pmap_release_pv( struct pmap_inval_info *info,
285 pv_entry_t pv, pv_entry_t pvp);
287 struct pmap_scan_info;
288 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
289 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
290 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
291 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
292 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
293 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
295 static void i386_protection_init (void);
296 static void create_pagetables(vm_paddr_t *firstaddr);
297 static void pmap_remove_all (vm_page_t m);
298 static boolean_t pmap_testbit (vm_page_t m, int bit);
300 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
301 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
303 static void pmap_pinit_defaults(struct pmap *pmap);
305 static unsigned pdir4mb;
308 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
310 if (pv1->pv_pindex < pv2->pv_pindex)
312 if (pv1->pv_pindex > pv2->pv_pindex)
317 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
318 pv_entry_compare, vm_pindex_t, pv_pindex);
322 pmap_page_stats_adding(vm_page_t m)
324 globaldata_t gd = mycpu;
326 if (TAILQ_EMPTY(&m->md.pv_list)) {
327 ++gd->gd_vmtotal.t_arm;
328 } else if (TAILQ_FIRST(&m->md.pv_list) ==
329 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
330 ++gd->gd_vmtotal.t_armshr;
331 ++gd->gd_vmtotal.t_avmshr;
333 ++gd->gd_vmtotal.t_avmshr;
339 pmap_page_stats_deleting(vm_page_t m)
341 globaldata_t gd = mycpu;
343 if (TAILQ_EMPTY(&m->md.pv_list)) {
344 --gd->gd_vmtotal.t_arm;
345 } else if (TAILQ_FIRST(&m->md.pv_list) ==
346 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
347 --gd->gd_vmtotal.t_armshr;
348 --gd->gd_vmtotal.t_avmshr;
350 --gd->gd_vmtotal.t_avmshr;
355 * Move the kernel virtual free pointer to the next
356 * 2MB. This is used to help improve performance
357 * by using a large (2MB) page for much of the kernel
358 * (.text, .data, .bss)
362 pmap_kmem_choose(vm_offset_t addr)
364 vm_offset_t newaddr = addr;
366 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
373 * Super fast pmap_pte routine best used when scanning the pv lists.
374 * This eliminates many course-grained invltlb calls. Note that many of
375 * the pv list scans are across different pmaps and it is very wasteful
376 * to do an entire invltlb when checking a single mapping.
378 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
382 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
384 return pmap_pte(pmap, va);
388 * Returns the pindex of a page table entry (representing a terminal page).
389 * There are NUPTE_TOTAL page table entries possible (a huge number)
391 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
392 * We want to properly translate negative KVAs.
396 pmap_pte_pindex(vm_offset_t va)
398 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
402 * Returns the pindex of a page table.
406 pmap_pt_pindex(vm_offset_t va)
408 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
412 * Returns the pindex of a page directory.
416 pmap_pd_pindex(vm_offset_t va)
418 return (NUPTE_TOTAL + NUPT_TOTAL +
419 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
424 pmap_pdp_pindex(vm_offset_t va)
426 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
427 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
432 pmap_pml4_pindex(void)
434 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
438 * Return various clipped indexes for a given VA
440 * Returns the index of a pte in a page table, representing a terminal
445 pmap_pte_index(vm_offset_t va)
447 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
451 * Returns the index of a pt in a page directory, representing a page
456 pmap_pt_index(vm_offset_t va)
458 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
462 * Returns the index of a pd in a page directory page, representing a page
467 pmap_pd_index(vm_offset_t va)
469 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
473 * Returns the index of a pdp in the pml4 table, representing a page
478 pmap_pdp_index(vm_offset_t va)
480 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
484 * Generic procedure to index a pte from a pt, pd, or pdp.
486 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
487 * a page table page index but is instead of PV lookup index.
491 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
495 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
496 return(&pte[pindex]);
500 * Return pointer to PDP slot in the PML4
504 pmap_pdp(pmap_t pmap, vm_offset_t va)
506 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
510 * Return pointer to PD slot in the PDP given a pointer to the PDP
514 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
518 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
519 return (&pd[pmap_pd_index(va)]);
523 * Return pointer to PD slot in the PDP.
527 pmap_pd(pmap_t pmap, vm_offset_t va)
531 pdp = pmap_pdp(pmap, va);
532 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
534 return (pmap_pdp_to_pd(*pdp, va));
538 * Return pointer to PT slot in the PD given a pointer to the PD
542 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
546 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
547 return (&pt[pmap_pt_index(va)]);
551 * Return pointer to PT slot in the PD
553 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
554 * so we cannot lookup the PD via the PDP. Instead we
555 * must look it up via the pmap.
559 pmap_pt(pmap_t pmap, vm_offset_t va)
563 vm_pindex_t pd_pindex;
565 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
566 pd_pindex = pmap_pd_pindex(va);
567 spin_lock(&pmap->pm_spin);
568 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
569 spin_unlock(&pmap->pm_spin);
570 if (pv == NULL || pv->pv_m == NULL)
572 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
574 pd = pmap_pd(pmap, va);
575 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
577 return (pmap_pd_to_pt(*pd, va));
582 * Return pointer to PTE slot in the PT given a pointer to the PT
586 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
590 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
591 return (&pte[pmap_pte_index(va)]);
595 * Return pointer to PTE slot in the PT
599 pmap_pte(pmap_t pmap, vm_offset_t va)
603 pt = pmap_pt(pmap, va);
604 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
606 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
607 return ((pt_entry_t *)pt);
608 return (pmap_pt_to_pte(*pt, va));
612 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
613 * the PT layer. This will speed up core pmap operations considerably.
615 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
616 * must be in a known associated state (typically by being locked when
617 * the pmap spinlock isn't held). We allow the race for that case.
621 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
623 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
624 pv->pv_pmap->pm_pvhint = pv;
629 * Return address of PT slot in PD (KVM only)
631 * Cannot be used for user page tables because it might interfere with
632 * the shared page-table-page optimization (pmap_mmu_optimize).
636 vtopt(vm_offset_t va)
638 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
639 NPML4EPGSHIFT)) - 1);
641 return (PDmap + ((va >> PDRSHIFT) & mask));
645 * KVM - return address of PTE slot in PT
649 vtopte(vm_offset_t va)
651 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
652 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
654 return (PTmap + ((va >> PAGE_SHIFT) & mask));
658 allocpages(vm_paddr_t *firstaddr, long n)
663 bzero((void *)ret, n * PAGE_SIZE);
664 *firstaddr += n * PAGE_SIZE;
670 create_pagetables(vm_paddr_t *firstaddr)
672 long i; /* must be 64 bits */
678 * We are running (mostly) V=P at this point
680 * Calculate NKPT - number of kernel page tables. We have to
681 * accomodoate prealloction of the vm_page_array, dump bitmap,
682 * MSGBUF_SIZE, and other stuff. Be generous.
684 * Maxmem is in pages.
686 * ndmpdp is the number of 1GB pages we wish to map.
688 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
689 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
691 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
694 * Starting at the beginning of kvm (not KERNBASE).
696 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
697 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
698 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
699 ndmpdp) + 511) / 512;
703 * Starting at KERNBASE - map 2G worth of page table pages.
704 * KERNBASE is offset -2G from the end of kvm.
706 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
711 KPTbase = allocpages(firstaddr, nkpt_base);
712 KPTphys = allocpages(firstaddr, nkpt_phys);
713 KPML4phys = allocpages(firstaddr, 1);
714 KPDPphys = allocpages(firstaddr, NKPML4E);
715 KPDphys = allocpages(firstaddr, NKPDPE);
718 * Calculate the page directory base for KERNBASE,
719 * that is where we start populating the page table pages.
720 * Basically this is the end - 2.
722 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
724 DMPDPphys = allocpages(firstaddr, NDMPML4E);
725 if ((amd_feature & AMDID_PAGE1GB) == 0)
726 DMPDphys = allocpages(firstaddr, ndmpdp);
727 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
730 * Fill in the underlying page table pages for the area around
731 * KERNBASE. This remaps low physical memory to KERNBASE.
733 * Read-only from zero to physfree
734 * XXX not fully used, underneath 2M pages
736 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
737 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
738 ((pt_entry_t *)KPTbase)[i] |=
739 pmap_bits_default[PG_RW_IDX] |
740 pmap_bits_default[PG_V_IDX] |
741 pmap_bits_default[PG_G_IDX];
745 * Now map the initial kernel page tables. One block of page
746 * tables is placed at the beginning of kernel virtual memory,
747 * and another block is placed at KERNBASE to map the kernel binary,
748 * data, bss, and initial pre-allocations.
750 for (i = 0; i < nkpt_base; i++) {
751 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
752 ((pd_entry_t *)KPDbase)[i] |=
753 pmap_bits_default[PG_RW_IDX] |
754 pmap_bits_default[PG_V_IDX];
756 for (i = 0; i < nkpt_phys; i++) {
757 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
758 ((pd_entry_t *)KPDphys)[i] |=
759 pmap_bits_default[PG_RW_IDX] |
760 pmap_bits_default[PG_V_IDX];
764 * Map from zero to end of allocations using 2M pages as an
765 * optimization. This will bypass some of the KPTBase pages
766 * above in the KERNBASE area.
768 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
769 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
770 ((pd_entry_t *)KPDbase)[i] |=
771 pmap_bits_default[PG_RW_IDX] |
772 pmap_bits_default[PG_V_IDX] |
773 pmap_bits_default[PG_PS_IDX] |
774 pmap_bits_default[PG_G_IDX];
778 * And connect up the PD to the PDP. The kernel pmap is expected
779 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
781 for (i = 0; i < NKPDPE; i++) {
782 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
783 KPDphys + (i << PAGE_SHIFT);
784 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
785 pmap_bits_default[PG_RW_IDX] |
786 pmap_bits_default[PG_V_IDX] |
787 pmap_bits_default[PG_U_IDX];
791 * Now set up the direct map space using either 2MB or 1GB pages
792 * Preset PG_M and PG_A because demotion expects it.
794 * When filling in entries in the PD pages make sure any excess
795 * entries are set to zero as we allocated enough PD pages
797 if ((amd_feature & AMDID_PAGE1GB) == 0) {
798 for (i = 0; i < NPDEPG * ndmpdp; i++) {
799 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
800 ((pd_entry_t *)DMPDphys)[i] |=
801 pmap_bits_default[PG_RW_IDX] |
802 pmap_bits_default[PG_V_IDX] |
803 pmap_bits_default[PG_PS_IDX] |
804 pmap_bits_default[PG_G_IDX] |
805 pmap_bits_default[PG_M_IDX] |
806 pmap_bits_default[PG_A_IDX];
810 * And the direct map space's PDP
812 for (i = 0; i < ndmpdp; i++) {
813 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
815 ((pdp_entry_t *)DMPDPphys)[i] |=
816 pmap_bits_default[PG_RW_IDX] |
817 pmap_bits_default[PG_V_IDX] |
818 pmap_bits_default[PG_U_IDX];
821 for (i = 0; i < ndmpdp; i++) {
822 ((pdp_entry_t *)DMPDPphys)[i] =
823 (vm_paddr_t)i << PDPSHIFT;
824 ((pdp_entry_t *)DMPDPphys)[i] |=
825 pmap_bits_default[PG_RW_IDX] |
826 pmap_bits_default[PG_V_IDX] |
827 pmap_bits_default[PG_PS_IDX] |
828 pmap_bits_default[PG_G_IDX] |
829 pmap_bits_default[PG_M_IDX] |
830 pmap_bits_default[PG_A_IDX];
834 /* And recursively map PML4 to itself in order to get PTmap */
835 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
836 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
837 pmap_bits_default[PG_RW_IDX] |
838 pmap_bits_default[PG_V_IDX] |
839 pmap_bits_default[PG_U_IDX];
842 * Connect the Direct Map slots up to the PML4
844 for (j = 0; j < NDMPML4E; ++j) {
845 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
846 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
847 pmap_bits_default[PG_RW_IDX] |
848 pmap_bits_default[PG_V_IDX] |
849 pmap_bits_default[PG_U_IDX];
853 * Connect the KVA slot up to the PML4
855 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
856 ((pdp_entry_t *)KPML4phys)[KPML4I] |=
857 pmap_bits_default[PG_RW_IDX] |
858 pmap_bits_default[PG_V_IDX] |
859 pmap_bits_default[PG_U_IDX];
863 * Bootstrap the system enough to run with virtual memory.
865 * On the i386 this is called after mapping has already been enabled
866 * and just syncs the pmap module with what has already been done.
867 * [We can't call it easily with mapping off since the kernel is not
868 * mapped with PA == VA, hence we would have to relocate every address
869 * from the linked base (virtual) address "KERNBASE" to the actual
870 * (physical) address starting relative to 0]
873 pmap_bootstrap(vm_paddr_t *firstaddr)
878 KvaStart = VM_MIN_KERNEL_ADDRESS;
879 KvaEnd = VM_MAX_KERNEL_ADDRESS;
880 KvaSize = KvaEnd - KvaStart;
882 avail_start = *firstaddr;
885 * Create an initial set of page tables to run the kernel in.
887 create_pagetables(firstaddr);
889 virtual2_start = KvaStart;
890 virtual2_end = PTOV_OFFSET;
892 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
893 virtual_start = pmap_kmem_choose(virtual_start);
895 virtual_end = VM_MAX_KERNEL_ADDRESS;
897 /* XXX do %cr0 as well */
898 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
902 * Initialize protection array.
904 i386_protection_init();
907 * The kernel's pmap is statically allocated so we don't have to use
908 * pmap_create, which is unlikely to work correctly at this part of
909 * the boot sequence (XXX and which no longer exists).
911 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
912 kernel_pmap.pm_count = 1;
913 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
914 RB_INIT(&kernel_pmap.pm_pvroot);
915 spin_init(&kernel_pmap.pm_spin);
916 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
919 * Reserve some special page table entries/VA space for temporary
922 #define SYSMAP(c, p, v, n) \
923 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
929 * CMAP1/CMAP2 are used for zeroing and copying pages.
931 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
936 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
939 * ptvmmap is used for reading arbitrary physical pages via
942 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
945 * msgbufp is used to map the system message buffer.
946 * XXX msgbufmap is not used.
948 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
949 atop(round_page(MSGBUF_SIZE)))
956 * PG_G is terribly broken on SMP because we IPI invltlb's in some
957 * cases rather then invl1pg. Actually, I don't even know why it
958 * works under UP because self-referential page table mappings
963 * Initialize the 4MB page size flag
967 * The 4MB page version of the initial
968 * kernel page mapping.
972 #if !defined(DISABLE_PSE)
973 if (cpu_feature & CPUID_PSE) {
976 * Note that we have enabled PSE mode
978 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
979 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
980 ptditmp &= ~(NBPDR - 1);
981 ptditmp |= pmap_bits_default[PG_V_IDX] |
982 pmap_bits_default[PG_RW_IDX] |
983 pmap_bits_default[PG_PS_IDX] |
984 pmap_bits_default[PG_U_IDX];
991 /* Initialize the PAT MSR */
993 pmap_pinit_defaults(&kernel_pmap);
995 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
996 &pmap_fast_kernel_cpusync);
1001 * Setup the PAT MSR.
1010 * Default values mapping PATi,PCD,PWT bits at system reset.
1011 * The default values effectively ignore the PATi bit by
1012 * repeating the encodings for 0-3 in 4-7, and map the PCD
1013 * and PWT bit combinations to the expected PAT types.
1015 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
1016 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
1017 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
1018 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
1019 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
1020 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
1021 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
1022 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
1023 pat_pte_index[PAT_WRITE_BACK] = 0;
1024 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
1025 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
1026 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
1027 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1028 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1030 if (cpu_feature & CPUID_PAT) {
1032 * If we support the PAT then set-up entries for
1033 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1036 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1037 PAT_VALUE(4, PAT_WRITE_PROTECTED);
1038 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1039 PAT_VALUE(5, PAT_WRITE_COMBINING);
1040 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1041 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1044 * Then enable the PAT
1049 load_cr4(cr4 & ~CR4_PGE);
1051 /* Disable caches (CD = 1, NW = 0). */
1053 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1055 /* Flushes caches and TLBs. */
1059 /* Update PAT and index table. */
1060 wrmsr(MSR_PAT, pat_msr);
1062 /* Flush caches and TLBs again. */
1066 /* Restore caches and PGE. */
1074 * Set 4mb pdir for mp startup
1079 if (cpu_feature & CPUID_PSE) {
1080 load_cr4(rcr4() | CR4_PSE);
1081 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
1088 * Initialize the pmap module.
1089 * Called by vm_init, to initialize any structures that the pmap
1090 * system needs to map virtual memory.
1091 * pmap_init has been enhanced to support in a fairly consistant
1092 * way, discontiguous physical memory.
1101 * Allocate memory for random pmap data structures. Includes the
1105 for (i = 0; i < vm_page_array_size; i++) {
1108 m = &vm_page_array[i];
1109 TAILQ_INIT(&m->md.pv_list);
1113 * init the pv free list
1115 initial_pvs = vm_page_array_size;
1116 if (initial_pvs < MINPV)
1117 initial_pvs = MINPV;
1118 pvzone = &pvzone_store;
1119 pvinit = (void *)kmem_alloc(&kernel_map,
1120 initial_pvs * sizeof (struct pv_entry));
1121 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1122 pvinit, initial_pvs);
1125 * Now it is safe to enable pv_table recording.
1127 pmap_initialized = TRUE;
1131 * Initialize the address space (zone) for the pv_entries. Set a
1132 * high water mark so that the system can recover from excessive
1133 * numbers of pv entries.
1138 int shpgperproc = PMAP_SHPGPERPROC;
1141 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1142 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1143 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1144 pv_entry_high_water = 9 * (pv_entry_max / 10);
1147 * Subtract out pages already installed in the zone (hack)
1149 entry_max = pv_entry_max - vm_page_array_size;
1153 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
1157 * Typically used to initialize a fictitious page by vm/device_pager.c
1160 pmap_page_init(struct vm_page *m)
1163 TAILQ_INIT(&m->md.pv_list);
1166 /***************************************************
1167 * Low level helper routines.....
1168 ***************************************************/
1171 * this routine defines the region(s) of memory that should
1172 * not be tested for the modified bit.
1176 pmap_track_modified(vm_pindex_t pindex)
1178 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1179 if ((va < clean_sva) || (va >= clean_eva))
1186 * Extract the physical page address associated with the map/VA pair.
1187 * The page must be wired for this to work reliably.
1189 * XXX for the moment we're using pv_find() instead of pv_get(), as
1190 * callers might be expecting non-blocking operation.
1193 pmap_extract(pmap_t pmap, vm_offset_t va)
1200 if (va >= VM_MAX_USER_ADDRESS) {
1202 * Kernel page directories might be direct-mapped and
1203 * there is typically no PV tracking of pte's
1207 pt = pmap_pt(pmap, va);
1208 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1209 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1210 rtval = *pt & PG_PS_FRAME;
1211 rtval |= va & PDRMASK;
1213 ptep = pmap_pt_to_pte(*pt, va);
1214 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1215 rtval = *ptep & PG_FRAME;
1216 rtval |= va & PAGE_MASK;
1222 * User pages currently do not direct-map the page directory
1223 * and some pages might not used managed PVs. But all PT's
1226 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1228 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1229 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1230 rtval = *ptep & PG_FRAME;
1231 rtval |= va & PAGE_MASK;
1240 * Similar to extract but checks protections, SMP-friendly short-cut for
1241 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1242 * fall-through to the real fault code.
1244 * The returned page, if not NULL, is held (and not busied).
1247 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1249 if (pmap && va < VM_MAX_USER_ADDRESS) {
1257 req = pmap->pmap_bits[PG_V_IDX] |
1258 pmap->pmap_bits[PG_U_IDX];
1259 if (prot & VM_PROT_WRITE)
1260 req |= pmap->pmap_bits[PG_RW_IDX];
1262 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1265 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1266 if ((*ptep & req) != req) {
1270 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1271 if (pte_pv && error == 0) {
1274 if (prot & VM_PROT_WRITE)
1277 } else if (pte_pv) {
1291 * Extract the physical page address associated kernel virtual address.
1294 pmap_kextract(vm_offset_t va)
1296 pd_entry_t pt; /* pt entry in pd */
1299 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1300 pa = DMAP_TO_PHYS(va);
1303 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1304 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1307 * Beware of a concurrent promotion that changes the
1308 * PDE at this point! For example, vtopte() must not
1309 * be used to access the PTE because it would use the
1310 * new PDE. It is, however, safe to use the old PDE
1311 * because the page table page is preserved by the
1314 pa = *pmap_pt_to_pte(pt, va);
1315 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1321 /***************************************************
1322 * Low level mapping routines.....
1323 ***************************************************/
1326 * Routine: pmap_kenter
1328 * Add a wired page to the KVA
1329 * NOTE! note that in order for the mapping to take effect -- you
1330 * should do an invltlb after doing the pmap_kenter().
1333 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1337 pmap_inval_info info;
1339 pmap_inval_init(&info); /* XXX remove */
1341 kernel_pmap.pmap_bits[PG_RW_IDX] |
1342 kernel_pmap.pmap_bits[PG_V_IDX];
1345 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1347 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1348 pmap_inval_done(&info); /* XXX remove */
1352 * Routine: pmap_kenter_quick
1354 * Similar to pmap_kenter(), except we only invalidate the
1355 * mapping on the current CPU.
1358 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1364 kernel_pmap.pmap_bits[PG_RW_IDX] |
1365 kernel_pmap.pmap_bits[PG_V_IDX];
1369 cpu_invlpg((void *)va);
1373 pmap_kenter_sync(vm_offset_t va)
1375 pmap_inval_info info;
1377 pmap_inval_init(&info);
1378 pmap_inval_interlock(&info, &kernel_pmap, va);
1379 pmap_inval_deinterlock(&info, &kernel_pmap);
1380 pmap_inval_done(&info);
1384 pmap_kenter_sync_quick(vm_offset_t va)
1386 cpu_invlpg((void *)va);
1390 * remove a page from the kernel pagetables
1393 pmap_kremove(vm_offset_t va)
1396 pmap_inval_info info;
1398 pmap_inval_init(&info);
1400 pmap_inval_interlock(&info, &kernel_pmap, va);
1401 (void)pte_load_clear(pte);
1402 pmap_inval_deinterlock(&info, &kernel_pmap);
1403 pmap_inval_done(&info);
1407 pmap_kremove_quick(vm_offset_t va)
1411 (void)pte_load_clear(pte);
1412 cpu_invlpg((void *)va);
1416 * XXX these need to be recoded. They are not used in any critical path.
1419 pmap_kmodify_rw(vm_offset_t va)
1421 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1422 cpu_invlpg((void *)va);
1427 pmap_kmodify_nc(vm_offset_t va)
1429 atomic_set_long(vtopte(va), PG_N);
1430 cpu_invlpg((void *)va);
1435 * Used to map a range of physical addresses into kernel virtual
1436 * address space during the low level boot, typically to map the
1437 * dump bitmap, message buffer, and vm_page_array.
1439 * These mappings are typically made at some pointer after the end of the
1442 * We could return PHYS_TO_DMAP(start) here and not allocate any
1443 * via (*virtp), but then kmem from userland and kernel dumps won't
1444 * have access to the related pointers.
1447 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1450 vm_offset_t va_start;
1452 /*return PHYS_TO_DMAP(start);*/
1457 while (start < end) {
1458 pmap_kenter_quick(va, start);
1466 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1469 * Remove the specified set of pages from the data and instruction caches.
1471 * In contrast to pmap_invalidate_cache_range(), this function does not
1472 * rely on the CPU's self-snoop feature, because it is intended for use
1473 * when moving pages into a different cache domain.
1476 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1478 vm_offset_t daddr, eva;
1481 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1482 (cpu_feature & CPUID_CLFSH) == 0)
1486 for (i = 0; i < count; i++) {
1487 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1488 eva = daddr + PAGE_SIZE;
1489 for (; daddr < eva; daddr += cpu_clflush_line_size)
1497 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1499 KASSERT((sva & PAGE_MASK) == 0,
1500 ("pmap_invalidate_cache_range: sva not page-aligned"));
1501 KASSERT((eva & PAGE_MASK) == 0,
1502 ("pmap_invalidate_cache_range: eva not page-aligned"));
1504 if (cpu_feature & CPUID_SS) {
1505 ; /* If "Self Snoop" is supported, do nothing. */
1507 /* Globally invalidate caches */
1508 cpu_wbinvd_on_all_cpus();
1512 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1514 smp_invlpg_range(pmap->pm_active, sva, eva);
1518 * Add a list of wired pages to the kva
1519 * this routine is only used for temporary
1520 * kernel mappings that do not need to have
1521 * page modification or references recorded.
1522 * Note that old mappings are simply written
1523 * over. The page *must* be wired.
1526 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1530 end_va = va + count * PAGE_SIZE;
1532 while (va < end_va) {
1536 *pte = VM_PAGE_TO_PHYS(*m) |
1537 kernel_pmap.pmap_bits[PG_RW_IDX] |
1538 kernel_pmap.pmap_bits[PG_V_IDX] |
1539 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1541 cpu_invlpg((void *)va);
1549 * This routine jerks page mappings from the
1550 * kernel -- it is meant only for temporary mappings.
1552 * MPSAFE, INTERRUPT SAFE (cluster callback)
1555 pmap_qremove(vm_offset_t va, int count)
1559 end_va = va + count * PAGE_SIZE;
1561 while (va < end_va) {
1565 (void)pte_load_clear(pte);
1566 cpu_invlpg((void *)va);
1573 * Create a new thread and optionally associate it with a (new) process.
1574 * NOTE! the new thread's cpu may not equal the current cpu.
1577 pmap_init_thread(thread_t td)
1579 /* enforce pcb placement & alignment */
1580 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1581 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1582 td->td_savefpu = &td->td_pcb->pcb_save;
1583 td->td_sp = (char *)td->td_pcb; /* no -16 */
1587 * This routine directly affects the fork perf for a process.
1590 pmap_init_proc(struct proc *p)
1595 pmap_pinit_defaults(struct pmap *pmap)
1597 bcopy(pmap_bits_default, pmap->pmap_bits,
1598 sizeof(pmap_bits_default));
1599 bcopy(protection_codes, pmap->protection_codes,
1600 sizeof(protection_codes));
1601 bcopy(pat_pte_index, pmap->pmap_cache_bits,
1602 sizeof(pat_pte_index));
1603 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1604 pmap->copyinstr = std_copyinstr;
1605 pmap->copyin = std_copyin;
1606 pmap->copyout = std_copyout;
1607 pmap->fubyte = std_fubyte;
1608 pmap->subyte = std_subyte;
1609 pmap->fuword = std_fuword;
1610 pmap->suword = std_suword;
1611 pmap->suword32 = std_suword32;
1614 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1615 * it, and IdlePTD, represents the template used to update all other pmaps.
1617 * On architectures where the kernel pmap is not integrated into the user
1618 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1619 * kernel_pmap should be used to directly access the kernel_pmap.
1622 pmap_pinit0(struct pmap *pmap)
1624 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1626 CPUMASK_ASSZERO(pmap->pm_active);
1627 pmap->pm_pvhint = NULL;
1628 RB_INIT(&pmap->pm_pvroot);
1629 spin_init(&pmap->pm_spin);
1630 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1631 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1632 pmap_pinit_defaults(pmap);
1636 * Initialize a preallocated and zeroed pmap structure,
1637 * such as one in a vmspace structure.
1640 pmap_pinit_simple(struct pmap *pmap)
1643 * Misc initialization
1646 CPUMASK_ASSZERO(pmap->pm_active);
1647 pmap->pm_pvhint = NULL;
1648 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1650 pmap_pinit_defaults(pmap);
1653 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1656 if (pmap->pm_pmlpv == NULL) {
1657 RB_INIT(&pmap->pm_pvroot);
1658 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1659 spin_init(&pmap->pm_spin);
1660 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1665 pmap_pinit(struct pmap *pmap)
1670 if (pmap->pm_pmlpv) {
1671 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1676 pmap_pinit_simple(pmap);
1677 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1680 * No need to allocate page table space yet but we do need a valid
1681 * page directory table.
1683 if (pmap->pm_pml4 == NULL) {
1685 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1689 * Allocate the page directory page, which wires it even though
1690 * it isn't being entered into some higher level page table (it
1691 * being the highest level). If one is already cached we don't
1692 * have to do anything.
1694 if ((pv = pmap->pm_pmlpv) == NULL) {
1695 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1696 pmap->pm_pmlpv = pv;
1697 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1698 VM_PAGE_TO_PHYS(pv->pv_m));
1702 * Install DMAP and KMAP.
1704 for (j = 0; j < NDMPML4E; ++j) {
1705 pmap->pm_pml4[DMPML4I + j] =
1706 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1707 pmap->pmap_bits[PG_RW_IDX] |
1708 pmap->pmap_bits[PG_V_IDX] |
1709 pmap->pmap_bits[PG_U_IDX];
1711 pmap->pm_pml4[KPML4I] = KPDPphys |
1712 pmap->pmap_bits[PG_RW_IDX] |
1713 pmap->pmap_bits[PG_V_IDX] |
1714 pmap->pmap_bits[PG_U_IDX];
1717 * install self-referential address mapping entry
1719 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1720 pmap->pmap_bits[PG_V_IDX] |
1721 pmap->pmap_bits[PG_RW_IDX] |
1722 pmap->pmap_bits[PG_A_IDX] |
1723 pmap->pmap_bits[PG_M_IDX];
1725 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1726 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1728 KKASSERT(pmap->pm_pml4[255] == 0);
1729 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1730 KKASSERT(pv->pv_entry.rbe_left == NULL);
1731 KKASSERT(pv->pv_entry.rbe_right == NULL);
1735 * Clean up a pmap structure so it can be physically freed. This routine
1736 * is called by the vmspace dtor function. A great deal of pmap data is
1737 * left passively mapped to improve vmspace management so we have a bit
1738 * of cleanup work to do here.
1741 pmap_puninit(pmap_t pmap)
1746 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1747 if ((pv = pmap->pm_pmlpv) != NULL) {
1748 if (pv_hold_try(pv) == 0)
1750 KKASSERT(pv == pmap->pm_pmlpv);
1751 p = pmap_remove_pv_page(pv);
1753 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1754 vm_page_busy_wait(p, FALSE, "pgpun");
1755 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1756 vm_page_unwire(p, 0);
1757 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1760 * XXX eventually clean out PML4 static entries and
1761 * use vm_page_free_zero()
1764 pmap->pm_pmlpv = NULL;
1766 if (pmap->pm_pml4) {
1767 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1768 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1769 pmap->pm_pml4 = NULL;
1771 KKASSERT(pmap->pm_stats.resident_count == 0);
1772 KKASSERT(pmap->pm_stats.wired_count == 0);
1776 * Wire in kernel global address entries. To avoid a race condition
1777 * between pmap initialization and pmap_growkernel, this procedure
1778 * adds the pmap to the master list (which growkernel scans to update),
1779 * then copies the template.
1782 pmap_pinit2(struct pmap *pmap)
1784 spin_lock(&pmap_spin);
1785 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1786 spin_unlock(&pmap_spin);
1790 * This routine is called when various levels in the page table need to
1791 * be populated. This routine cannot fail.
1793 * This function returns two locked pv_entry's, one representing the
1794 * requested pv and one representing the requested pv's parent pv. If
1795 * the pv did not previously exist it will be mapped into its parent
1796 * and wired, otherwise no additional wire count will be added.
1800 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1805 vm_pindex_t pt_pindex;
1811 * If the pv already exists and we aren't being asked for the
1812 * parent page table page we can just return it. A locked+held pv
1813 * is returned. The pv will also have a second hold related to the
1814 * pmap association that we don't have to worry about.
1817 pv = pv_alloc(pmap, ptepindex, &isnew);
1818 if (isnew == 0 && pvpp == NULL)
1822 * Special case terminal PVs. These are not page table pages so
1823 * no vm_page is allocated (the caller supplied the vm_page). If
1824 * pvpp is non-NULL we are being asked to also removed the pt_pv
1827 * Note that pt_pv's are only returned for user VAs. We assert that
1828 * a pt_pv is not being requested for kernel VAs.
1830 if (ptepindex < pmap_pt_pindex(0)) {
1831 if (ptepindex >= NUPTE_USER)
1832 KKASSERT(pvpp == NULL);
1834 KKASSERT(pvpp != NULL);
1836 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1837 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1839 vm_page_wire_quick(pvp->pv_m);
1848 * Non-terminal PVs allocate a VM page to represent the page table,
1849 * so we have to resolve pvp and calculate ptepindex for the pvp
1850 * and then for the page table entry index in the pvp for
1853 if (ptepindex < pmap_pd_pindex(0)) {
1855 * pv is PT, pvp is PD
1857 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1858 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1859 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1866 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1867 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1869 } else if (ptepindex < pmap_pdp_pindex(0)) {
1871 * pv is PD, pvp is PDP
1873 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1876 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1877 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1879 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1880 KKASSERT(pvpp == NULL);
1883 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1891 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1892 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1893 } else if (ptepindex < pmap_pml4_pindex()) {
1895 * pv is PDP, pvp is the root pml4 table
1897 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1904 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1905 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1908 * pv represents the top-level PML4, there is no parent.
1916 * This code is only reached if isnew is TRUE and this is not a
1917 * terminal PV. We need to allocate a vm_page for the page table
1918 * at this level and enter it into the parent page table.
1920 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1923 m = vm_page_alloc(NULL, pv->pv_pindex,
1924 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1925 VM_ALLOC_INTERRUPT);
1930 vm_page_spin_lock(m);
1931 pmap_page_stats_adding(m);
1932 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1934 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1935 vm_page_spin_unlock(m);
1936 vm_page_unmanage(m); /* m must be spinunlocked */
1938 if ((m->flags & PG_ZERO) == 0) {
1939 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1943 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1946 m->valid = VM_PAGE_BITS_ALL;
1947 vm_page_flag_clear(m, PG_ZERO);
1948 vm_page_wire(m); /* wire for mapping in parent */
1951 * Wire the page into pvp, bump the wire-count for pvp's page table
1952 * page. Bump the resident_count for the pmap. There is no pvp
1953 * for the top level, address the pm_pml4[] array directly.
1955 * If the caller wants the parent we return it, otherwise
1956 * we just put it away.
1958 * No interlock is needed for pte 0 -> non-zero.
1960 * In the situation where *ptep is valid we might have an unmanaged
1961 * page table page shared from another page table which we need to
1962 * unshare before installing our private page table page.
1965 ptep = pv_pte_lookup(pvp, ptepindex);
1966 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1968 pmap_inval_info info;
1971 panic("pmap_allocpte: unexpected pte %p/%d",
1972 pvp, (int)ptepindex);
1974 pmap_inval_init(&info);
1975 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1976 pte = pte_load_clear(ptep);
1977 pmap_inval_deinterlock(&info, pmap);
1978 pmap_inval_done(&info);
1979 if (vm_page_unwire_quick(
1980 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
1981 panic("pmap_allocpte: shared pgtable "
1982 "pg bad wirecount");
1984 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1986 vm_page_wire_quick(pvp->pv_m);
1988 *ptep = VM_PAGE_TO_PHYS(m) |
1989 (pmap->pmap_bits[PG_U_IDX] |
1990 pmap->pmap_bits[PG_RW_IDX] |
1991 pmap->pmap_bits[PG_V_IDX] |
1992 pmap->pmap_bits[PG_A_IDX] |
1993 pmap->pmap_bits[PG_M_IDX]);
2005 * This version of pmap_allocpte() checks for possible segment optimizations
2006 * that would allow page-table sharing. It can be called for terminal
2007 * page or page table page ptepindex's.
2009 * The function is called with page table page ptepindex's for fictitious
2010 * and unmanaged terminal pages. That is, we don't want to allocate a
2011 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2014 * This function can return a pv and *pvpp associated with the passed in pmap
2015 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2016 * an unmanaged page table page will be entered into the pass in pmap.
2020 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2021 vm_map_entry_t entry, vm_offset_t va)
2023 struct pmap_inval_info info;
2028 pv_entry_t pte_pv; /* in original or shared pmap */
2029 pv_entry_t pt_pv; /* in original or shared pmap */
2030 pv_entry_t proc_pd_pv; /* in original pmap */
2031 pv_entry_t proc_pt_pv; /* in original pmap */
2032 pv_entry_t xpv; /* PT in shared pmap */
2033 pd_entry_t *pt; /* PT entry in PD of original pmap */
2034 pd_entry_t opte; /* contents of *pt */
2035 pd_entry_t npte; /* contents of *pt */
2040 * Basic tests, require a non-NULL vm_map_entry, require proper
2041 * alignment and type for the vm_map_entry, require that the
2042 * underlying object already be allocated.
2044 * We allow almost any type of object to use this optimization.
2045 * The object itself does NOT have to be sized to a multiple of the
2046 * segment size, but the memory mapping does.
2048 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2049 * won't work as expected.
2051 if (entry == NULL ||
2052 pmap_mmu_optimize == 0 || /* not enabled */
2053 ptepindex >= pmap_pd_pindex(0) || /* not terminal or pt */
2054 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2055 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2056 entry->object.vm_object == NULL || /* needs VM object */
2057 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2058 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2059 (entry->offset & SEG_MASK) || /* must be aligned */
2060 (entry->start & SEG_MASK)) {
2061 return(pmap_allocpte(pmap, ptepindex, pvpp));
2065 * Make sure the full segment can be represented.
2067 b = va & ~(vm_offset_t)SEG_MASK;
2068 if (b < entry->start || b + SEG_SIZE > entry->end)
2069 return(pmap_allocpte(pmap, ptepindex, pvpp));
2072 * If the full segment can be represented dive the VM object's
2073 * shared pmap, allocating as required.
2075 object = entry->object.vm_object;
2077 if (entry->protection & VM_PROT_WRITE)
2078 obpmapp = &object->md.pmap_rw;
2080 obpmapp = &object->md.pmap_ro;
2083 if (pmap_enter_debug > 0) {
2085 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2087 va, entry->protection, object,
2089 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2090 entry, entry->start, entry->end);
2095 * We allocate what appears to be a normal pmap but because portions
2096 * of this pmap are shared with other unrelated pmaps we have to
2097 * set pm_active to point to all cpus.
2099 * XXX Currently using pmap_spin to interlock the update, can't use
2100 * vm_object_hold/drop because the token might already be held
2101 * shared OR exclusive and we don't know.
2103 while ((obpmap = *obpmapp) == NULL) {
2104 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2105 pmap_pinit_simple(obpmap);
2106 pmap_pinit2(obpmap);
2107 spin_lock(&pmap_spin);
2108 if (*obpmapp != NULL) {
2112 spin_unlock(&pmap_spin);
2113 pmap_release(obpmap);
2114 pmap_puninit(obpmap);
2115 kfree(obpmap, M_OBJPMAP);
2116 obpmap = *obpmapp; /* safety */
2118 obpmap->pm_active = smp_active_mask;
2120 spin_unlock(&pmap_spin);
2125 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2126 * pte/pt using the shared pmap from the object but also adjust
2127 * the process pmap's page table page as a side effect.
2131 * Resolve the terminal PTE and PT in the shared pmap. This is what
2132 * we will return. This is true if ptepindex represents a terminal
2133 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2137 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2138 if (ptepindex >= pmap_pt_pindex(0))
2144 * Resolve the PD in the process pmap so we can properly share the
2145 * page table page. Lock order is bottom-up (leaf first)!
2147 * NOTE: proc_pt_pv can be NULL.
2149 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2150 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2152 if (pmap_enter_debug > 0) {
2154 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2156 (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2163 * xpv is the page table page pv from the shared object
2164 * (for convenience), from above.
2166 * Calculate the pte value for the PT to load into the process PD.
2167 * If we have to change it we must properly dispose of the previous
2170 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2171 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2172 (pmap->pmap_bits[PG_U_IDX] |
2173 pmap->pmap_bits[PG_RW_IDX] |
2174 pmap->pmap_bits[PG_V_IDX] |
2175 pmap->pmap_bits[PG_A_IDX] |
2176 pmap->pmap_bits[PG_M_IDX]);
2179 * Dispose of previous page table page if it was local to the
2180 * process pmap. If the old pt is not empty we cannot dispose of it
2181 * until we clean it out. This case should not arise very often so
2182 * it is not optimized.
2185 if (proc_pt_pv->pv_m->wire_count != 1) {
2191 va & ~(vm_offset_t)SEG_MASK,
2192 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2197 * The release call will indirectly clean out *pt
2199 pmap_inval_init(&info);
2200 pmap_release_pv(&info, proc_pt_pv, proc_pd_pv);
2201 pmap_inval_done(&info);
2204 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2208 * Handle remaining cases.
2212 vm_page_wire_quick(xpv->pv_m);
2213 vm_page_wire_quick(proc_pd_pv->pv_m);
2214 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2215 } else if (*pt != npte) {
2216 pmap_inval_init(&info);
2217 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
2219 opte = pte_load_clear(pt);
2220 KKASSERT(opte && opte != npte);
2223 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
2226 * Clean up opte, bump the wire_count for the process
2227 * PD page representing the new entry if it was
2230 * If the entry was not previously empty and we have
2231 * a PT in the proc pmap then opte must match that
2232 * pt. The proc pt must be retired (this is done
2233 * later on in this procedure).
2235 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2238 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2239 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2240 if (vm_page_unwire_quick(m)) {
2241 panic("pmap_allocpte_seg: "
2242 "bad wire count %p",
2246 pmap_inval_deinterlock(&info, pmap);
2247 pmap_inval_done(&info);
2251 * The existing process page table was replaced and must be destroyed
2265 * Release any resources held by the given physical map.
2267 * Called when a pmap initialized by pmap_pinit is being released. Should
2268 * only be called if the map contains no valid mappings.
2270 * Caller must hold pmap->pm_token
2272 struct pmap_release_info {
2277 static int pmap_release_callback(pv_entry_t pv, void *data);
2280 pmap_release(struct pmap *pmap)
2282 struct pmap_release_info info;
2284 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2285 ("pmap still active! %016jx",
2286 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2288 spin_lock(&pmap_spin);
2289 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2290 spin_unlock(&pmap_spin);
2293 * Pull pv's off the RB tree in order from low to high and release
2299 spin_lock(&pmap->pm_spin);
2300 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2301 pmap_release_callback, &info);
2302 spin_unlock(&pmap->pm_spin);
2303 } while (info.retry);
2307 * One resident page (the pml4 page) should remain.
2308 * No wired pages should remain.
2310 KKASSERT(pmap->pm_stats.resident_count ==
2311 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2313 KKASSERT(pmap->pm_stats.wired_count == 0);
2317 pmap_release_callback(pv_entry_t pv, void *data)
2319 struct pmap_release_info *info = data;
2320 pmap_t pmap = info->pmap;
2323 if (pv_hold_try(pv)) {
2324 spin_unlock(&pmap->pm_spin);
2326 spin_unlock(&pmap->pm_spin);
2329 if (pv->pv_pmap != pmap) {
2331 spin_lock(&pmap->pm_spin);
2335 r = pmap_release_pv(NULL, pv, NULL);
2336 spin_lock(&pmap->pm_spin);
2341 * Called with held (i.e. also locked) pv. This function will dispose of
2342 * the lock along with the pv.
2344 * If the caller already holds the locked parent page table for pv it
2345 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2346 * pass NULL for pvp.
2349 pmap_release_pv(struct pmap_inval_info *info, pv_entry_t pv, pv_entry_t pvp)
2354 * The pmap is currently not spinlocked, pv is held+locked.
2355 * Remove the pv's page from its parent's page table. The
2356 * parent's page table page's wire_count will be decremented.
2358 * This will clean out the pte at any level of the page table.
2359 * If info is not NULL the appropriate invlpg/invltlb/smp
2360 * invalidation will be made.
2362 pmap_remove_pv_pte(pv, pvp, info);
2365 * Terminal pvs are unhooked from their vm_pages. Because
2366 * terminal pages aren't page table pages they aren't wired
2367 * by us, so we have to be sure not to unwire them either.
2369 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2370 pmap_remove_pv_page(pv);
2375 * We leave the top-level page table page cached, wired, and
2376 * mapped in the pmap until the dtor function (pmap_puninit())
2379 * Since we are leaving the top-level pv intact we need
2380 * to break out of what would otherwise be an infinite loop.
2382 if (pv->pv_pindex == pmap_pml4_pindex()) {
2388 * For page table pages (other than the top-level page),
2389 * remove and free the vm_page. The representitive mapping
2390 * removed above by pmap_remove_pv_pte() did not undo the
2391 * last wire_count so we have to do that as well.
2393 p = pmap_remove_pv_page(pv);
2394 vm_page_busy_wait(p, FALSE, "pmaprl");
2395 if (p->wire_count != 1) {
2396 kprintf("p->wire_count was %016lx %d\n",
2397 pv->pv_pindex, p->wire_count);
2399 KKASSERT(p->wire_count == 1);
2400 KKASSERT(p->flags & PG_UNMANAGED);
2402 vm_page_unwire(p, 0);
2403 KKASSERT(p->wire_count == 0);
2406 * Theoretically this page, if not the pml4 page, should contain
2407 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2417 * This function will remove the pte associated with a pv from its parent.
2418 * Terminal pv's are supported. The removal will be interlocked if info
2419 * is non-NULL. The caller must dispose of pv instead of just unlocking
2422 * The wire count will be dropped on the parent page table. The wire
2423 * count on the page being removed (pv->pv_m) from the parent page table
2424 * is NOT touched. Note that terminal pages will not have any additional
2425 * wire counts while page table pages will have at least one representing
2426 * the mapping, plus others representing sub-mappings.
2428 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2429 * pages and user page table and terminal pages.
2431 * The pv must be locked.
2433 * XXX must lock parent pv's if they exist to remove pte XXX
2437 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2439 vm_pindex_t ptepindex = pv->pv_pindex;
2440 pmap_t pmap = pv->pv_pmap;
2446 if (ptepindex == pmap_pml4_pindex()) {
2448 * We are the top level pml4 table, there is no parent.
2450 p = pmap->pm_pmlpv->pv_m;
2451 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2453 * Remove a PDP page from the pml4e. This can only occur
2454 * with user page tables. We do not have to lock the
2455 * pml4 PV so just ignore pvp.
2457 vm_pindex_t pml4_pindex;
2458 vm_pindex_t pdp_index;
2461 pdp_index = ptepindex - pmap_pdp_pindex(0);
2463 pml4_pindex = pmap_pml4_pindex();
2464 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2468 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2469 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2470 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2472 pmap_inval_interlock(info, pmap, (vm_offset_t)-1);
2473 pte_load_clear(pdp);
2474 pmap_inval_deinterlock(info, pmap);
2478 } else if (ptepindex >= pmap_pd_pindex(0)) {
2480 * Remove a PD page from the pdp
2482 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2483 * of a simple pmap because it stops at
2486 vm_pindex_t pdp_pindex;
2487 vm_pindex_t pd_index;
2490 pd_index = ptepindex - pmap_pd_pindex(0);
2493 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2494 (pd_index >> NPML4EPGSHIFT);
2495 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2500 pd = pv_pte_lookup(pvp, pd_index &
2501 ((1ul << NPDPEPGSHIFT) - 1));
2502 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2503 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2505 pmap_inval_interlock(info, pmap,
2508 pmap_inval_deinterlock(info, pmap);
2513 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2514 p = pv->pv_m; /* degenerate test later */
2516 } else if (ptepindex >= pmap_pt_pindex(0)) {
2518 * Remove a PT page from the pd
2520 vm_pindex_t pd_pindex;
2521 vm_pindex_t pt_index;
2524 pt_index = ptepindex - pmap_pt_pindex(0);
2527 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2528 (pt_index >> NPDPEPGSHIFT);
2529 pvp = pv_get(pv->pv_pmap, pd_pindex);
2533 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2534 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2535 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2537 pmap_inval_interlock(info, pmap, (vm_offset_t)-1);
2539 pmap_inval_deinterlock(info, pmap);
2545 * Remove a PTE from the PT page
2547 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2548 * pv is a pte_pv so we can safely lock pt_pv.
2550 * NOTE: FICTITIOUS pages may have multiple physical mappings
2551 * so PHYS_TO_VM_PAGE() will not necessarily work for
2554 vm_pindex_t pt_pindex;
2559 pt_pindex = ptepindex >> NPTEPGSHIFT;
2560 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2562 if (ptepindex >= NUPTE_USER) {
2563 ptep = vtopte(ptepindex << PAGE_SHIFT);
2564 KKASSERT(pvp == NULL);
2567 pt_pindex = NUPTE_TOTAL +
2568 (ptepindex >> NPDPEPGSHIFT);
2569 pvp = pv_get(pv->pv_pmap, pt_pindex);
2573 ptep = pv_pte_lookup(pvp, ptepindex &
2574 ((1ul << NPDPEPGSHIFT) - 1));
2578 pmap_inval_interlock(info, pmap, va);
2579 pte = pte_load_clear(ptep);
2581 pmap_inval_deinterlock(info, pmap);
2583 cpu_invlpg((void *)va);
2586 * Now update the vm_page_t
2588 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) !=
2589 (pmap->pmap_bits[PG_MANAGED_IDX]|pmap->pmap_bits[PG_V_IDX])) {
2590 kprintf("remove_pte badpte %016lx %016lx %d\n",
2592 pv->pv_pindex < pmap_pt_pindex(0));
2594 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2595 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2596 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2599 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2602 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2603 if (pmap_track_modified(ptepindex))
2606 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2607 vm_page_flag_set(p, PG_REFERENCED);
2609 if (pte & pmap->pmap_bits[PG_W_IDX])
2610 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2611 if (pte & pmap->pmap_bits[PG_G_IDX])
2612 cpu_invlpg((void *)va);
2616 * Unwire the parent page table page. The wire_count cannot go below
2617 * 1 here because the parent page table page is itself still mapped.
2619 * XXX remove the assertions later.
2621 KKASSERT(pv->pv_m == p);
2622 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2623 panic("pmap_remove_pv_pte: Insufficient wire_count");
2630 * Remove the vm_page association to a pv. The pv must be locked.
2634 pmap_remove_pv_page(pv_entry_t pv)
2640 vm_page_spin_lock(m);
2642 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2643 pmap_page_stats_deleting(m);
2646 atomic_add_int(&m->object->agg_pv_list_count, -1);
2648 if (TAILQ_EMPTY(&m->md.pv_list))
2649 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2650 vm_page_spin_unlock(m);
2655 * Grow the number of kernel page table entries, if needed.
2657 * This routine is always called to validate any address space
2658 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2659 * space below KERNBASE.
2662 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2665 vm_offset_t ptppaddr;
2667 pd_entry_t *pt, newpt;
2669 int update_kernel_vm_end;
2672 * bootstrap kernel_vm_end on first real VM use
2674 if (kernel_vm_end == 0) {
2675 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2677 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2678 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2679 ~(PAGE_SIZE * NPTEPG - 1);
2681 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2682 kernel_vm_end = kernel_map.max_offset;
2689 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2690 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2691 * do not want to force-fill 128G worth of page tables.
2693 if (kstart < KERNBASE) {
2694 if (kstart > kernel_vm_end)
2695 kstart = kernel_vm_end;
2696 KKASSERT(kend <= KERNBASE);
2697 update_kernel_vm_end = 1;
2699 update_kernel_vm_end = 0;
2702 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2703 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2705 if (kend - 1 >= kernel_map.max_offset)
2706 kend = kernel_map.max_offset;
2708 while (kstart < kend) {
2709 pt = pmap_pt(&kernel_pmap, kstart);
2711 /* We need a new PDP entry */
2712 nkpg = vm_page_alloc(NULL, nkpt,
2715 VM_ALLOC_INTERRUPT);
2717 panic("pmap_growkernel: no memory to grow "
2720 paddr = VM_PAGE_TO_PHYS(nkpg);
2721 if ((nkpg->flags & PG_ZERO) == 0)
2722 pmap_zero_page(paddr);
2723 vm_page_flag_clear(nkpg, PG_ZERO);
2724 newpd = (pdp_entry_t)
2726 kernel_pmap.pmap_bits[PG_V_IDX] |
2727 kernel_pmap.pmap_bits[PG_RW_IDX] |
2728 kernel_pmap.pmap_bits[PG_A_IDX] |
2729 kernel_pmap.pmap_bits[PG_M_IDX]);
2730 *pmap_pd(&kernel_pmap, kstart) = newpd;
2732 continue; /* try again */
2734 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2735 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2736 ~(PAGE_SIZE * NPTEPG - 1);
2737 if (kstart - 1 >= kernel_map.max_offset) {
2738 kstart = kernel_map.max_offset;
2745 * This index is bogus, but out of the way
2747 nkpg = vm_page_alloc(NULL, nkpt,
2750 VM_ALLOC_INTERRUPT);
2752 panic("pmap_growkernel: no memory to grow kernel");
2755 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2756 pmap_zero_page(ptppaddr);
2757 vm_page_flag_clear(nkpg, PG_ZERO);
2758 newpt = (pd_entry_t) (ptppaddr |
2759 kernel_pmap.pmap_bits[PG_V_IDX] |
2760 kernel_pmap.pmap_bits[PG_RW_IDX] |
2761 kernel_pmap.pmap_bits[PG_A_IDX] |
2762 kernel_pmap.pmap_bits[PG_M_IDX]);
2763 *pmap_pt(&kernel_pmap, kstart) = newpt;
2766 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2767 ~(PAGE_SIZE * NPTEPG - 1);
2769 if (kstart - 1 >= kernel_map.max_offset) {
2770 kstart = kernel_map.max_offset;
2776 * Only update kernel_vm_end for areas below KERNBASE.
2778 if (update_kernel_vm_end && kernel_vm_end < kstart)
2779 kernel_vm_end = kstart;
2783 * Add a reference to the specified pmap.
2786 pmap_reference(pmap_t pmap)
2789 lwkt_gettoken(&pmap->pm_token);
2791 lwkt_reltoken(&pmap->pm_token);
2795 /***************************************************
2796 * page management routines.
2797 ***************************************************/
2800 * Hold a pv without locking it
2803 pv_hold(pv_entry_t pv)
2805 atomic_add_int(&pv->pv_hold, 1);
2809 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2810 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2813 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2814 * pv list via its page) must be held by the caller.
2817 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2822 * Critical path shortcut expects pv to already have one ref
2823 * (for the pv->pv_pmap).
2825 if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
2828 pv->pv_line = lineno;
2834 count = pv->pv_hold;
2836 if ((count & PV_HOLD_LOCKED) == 0) {
2837 if (atomic_cmpset_int(&pv->pv_hold, count,
2838 (count + 1) | PV_HOLD_LOCKED)) {
2841 pv->pv_line = lineno;
2846 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2854 * Drop a previously held pv_entry which could not be locked, allowing its
2857 * Must not be called with a spinlock held as we might zfree() the pv if it
2858 * is no longer associated with a pmap and this was the last hold count.
2861 pv_drop(pv_entry_t pv)
2866 count = pv->pv_hold;
2868 KKASSERT((count & PV_HOLD_MASK) > 0);
2869 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2870 (PV_HOLD_LOCKED | 1));
2871 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2872 if ((count & PV_HOLD_MASK) == 1) {
2874 if (pmap_enter_debug > 0) {
2876 kprintf("pv_drop: free pv %p\n", pv);
2879 KKASSERT(count == 1);
2880 KKASSERT(pv->pv_pmap == NULL);
2890 * Find or allocate the requested PV entry, returning a locked, held pv.
2892 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
2893 * for the caller and one representing the pmap and vm_page association.
2895 * If (*isnew) is zero, the returned pv will have only one hold count.
2897 * Since both associations can only be adjusted while the pv is locked,
2898 * together they represent just one additional hold.
2902 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2905 pv_entry_t pnew = NULL;
2907 spin_lock(&pmap->pm_spin);
2909 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2910 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2915 spin_unlock(&pmap->pm_spin);
2916 pnew = zalloc(pvzone);
2917 spin_lock(&pmap->pm_spin);
2920 pnew->pv_pmap = pmap;
2921 pnew->pv_pindex = pindex;
2922 pnew->pv_hold = PV_HOLD_LOCKED | 2;
2924 pnew->pv_func = func;
2925 pnew->pv_line = lineno;
2927 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2928 ++pmap->pm_generation;
2929 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2930 spin_unlock(&pmap->pm_spin);
2935 spin_unlock(&pmap->pm_spin);
2936 zfree(pvzone, pnew);
2938 spin_lock(&pmap->pm_spin);
2941 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2942 spin_unlock(&pmap->pm_spin);
2944 spin_unlock(&pmap->pm_spin);
2945 _pv_lock(pv PMAP_DEBUG_COPY);
2947 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2952 spin_lock(&pmap->pm_spin);
2957 * Find the requested PV entry, returning a locked+held pv or NULL
2961 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2965 spin_lock(&pmap->pm_spin);
2970 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2971 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2975 spin_unlock(&pmap->pm_spin);
2978 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2979 spin_unlock(&pmap->pm_spin);
2981 spin_unlock(&pmap->pm_spin);
2982 _pv_lock(pv PMAP_DEBUG_COPY);
2984 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2985 pv_cache(pv, pindex);
2989 spin_lock(&pmap->pm_spin);
2994 * Lookup, hold, and attempt to lock (pmap,pindex).
2996 * If the entry does not exist NULL is returned and *errorp is set to 0
2998 * If the entry exists and could be successfully locked it is returned and
2999 * errorp is set to 0.
3001 * If the entry exists but could NOT be successfully locked it is returned
3002 * held and *errorp is set to 1.
3006 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
3010 spin_lock_shared(&pmap->pm_spin);
3011 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3012 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3014 spin_unlock_shared(&pmap->pm_spin);
3018 if (pv_hold_try(pv)) {
3019 pv_cache(pv, pindex);
3020 spin_unlock_shared(&pmap->pm_spin);
3022 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3023 return(pv); /* lock succeeded */
3025 spin_unlock_shared(&pmap->pm_spin);
3027 return (pv); /* lock failed */
3031 * Find the requested PV entry, returning a held pv or NULL
3035 pv_find(pmap_t pmap, vm_pindex_t pindex)
3039 spin_lock_shared(&pmap->pm_spin);
3041 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3042 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3044 spin_unlock_shared(&pmap->pm_spin);
3048 pv_cache(pv, pindex);
3049 spin_unlock_shared(&pmap->pm_spin);
3054 * Lock a held pv, keeping the hold count
3058 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3063 count = pv->pv_hold;
3065 if ((count & PV_HOLD_LOCKED) == 0) {
3066 if (atomic_cmpset_int(&pv->pv_hold, count,
3067 count | PV_HOLD_LOCKED)) {
3070 pv->pv_line = lineno;
3076 tsleep_interlock(pv, 0);
3077 if (atomic_cmpset_int(&pv->pv_hold, count,
3078 count | PV_HOLD_WAITING)) {
3080 kprintf("pv waiting on %s:%d\n",
3081 pv->pv_func, pv->pv_line);
3083 tsleep(pv, PINTERLOCKED, "pvwait", hz);
3090 * Unlock a held and locked pv, keeping the hold count.
3094 pv_unlock(pv_entry_t pv)
3099 count = pv->pv_hold;
3101 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3102 (PV_HOLD_LOCKED | 1));
3103 if (atomic_cmpset_int(&pv->pv_hold, count,
3105 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3106 if (count & PV_HOLD_WAITING)
3114 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3115 * and the hold count drops to zero we will free it.
3117 * Caller should not hold any spin locks. We are protected from hold races
3118 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3119 * lock held. A pv cannot be located otherwise.
3123 pv_put(pv_entry_t pv)
3126 if (pmap_enter_debug > 0) {
3128 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3133 * Fast - shortcut most common condition
3135 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3146 * Remove the pmap association from a pv, require that pv_m already be removed,
3147 * then unlock and drop the pv. Any pte operations must have already been
3148 * completed. This call may result in a last-drop which will physically free
3151 * Removing the pmap association entails an additional drop.
3153 * pv must be exclusively locked on call and will be disposed of on return.
3157 pv_free(pv_entry_t pv)
3161 KKASSERT(pv->pv_m == NULL);
3162 KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3163 if ((pmap = pv->pv_pmap) != NULL) {
3164 spin_lock(&pmap->pm_spin);
3165 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3166 ++pmap->pm_generation;
3167 if (pmap->pm_pvhint == pv)
3168 pmap->pm_pvhint = NULL;
3169 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3172 spin_unlock(&pmap->pm_spin);
3175 * Try to shortcut three atomic ops, otherwise fall through
3176 * and do it normally. Drop two refs and the lock all in
3179 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3181 if (pmap_enter_debug > 0) {
3183 kprintf("pv_free: free pv %p\n", pv);
3189 pv_drop(pv); /* ref for pv_pmap */
3195 * This routine is very drastic, but can save the system
3203 static int warningdone=0;
3205 if (pmap_pagedaemon_waken == 0)
3207 pmap_pagedaemon_waken = 0;
3208 if (warningdone < 5) {
3209 kprintf("pmap_collect: collecting pv entries -- "
3210 "suggest increasing PMAP_SHPGPERPROC\n");
3214 for (i = 0; i < vm_page_array_size; i++) {
3215 m = &vm_page_array[i];
3216 if (m->wire_count || m->hold_count)
3218 if (vm_page_busy_try(m, TRUE) == 0) {
3219 if (m->wire_count == 0 && m->hold_count == 0) {
3228 * Scan the pmap for active page table entries and issue a callback.
3229 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3230 * its parent page table.
3232 * pte_pv will be NULL if the page or page table is unmanaged.
3233 * pt_pv will point to the page table page containing the pte for the page.
3235 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3236 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3237 * process pmap's PD and page to the callback function. This can be
3238 * confusing because the pt_pv is really a pd_pv, and the target page
3239 * table page is simply aliased by the pmap and not owned by it.
3241 * It is assumed that the start and end are properly rounded to the page size.
3243 * It is assumed that PD pages and above are managed and thus in the RB tree,
3244 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3246 struct pmap_scan_info {
3250 vm_pindex_t sva_pd_pindex;
3251 vm_pindex_t eva_pd_pindex;
3252 void (*func)(pmap_t, struct pmap_scan_info *,
3253 pv_entry_t, pv_entry_t, int, vm_offset_t,
3254 pt_entry_t *, void *);
3257 struct pmap_inval_info inval;
3260 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3261 static int pmap_scan_callback(pv_entry_t pv, void *data);
3264 pmap_scan(struct pmap_scan_info *info)
3266 struct pmap *pmap = info->pmap;
3267 pv_entry_t pd_pv; /* A page directory PV */
3268 pv_entry_t pt_pv; /* A page table PV */
3269 pv_entry_t pte_pv; /* A page table entry PV */
3272 struct pv_entry dummy_pv;
3279 * Hold the token for stability; if the pmap is empty we have nothing
3282 lwkt_gettoken(&pmap->pm_token);
3284 if (pmap->pm_stats.resident_count == 0) {
3285 lwkt_reltoken(&pmap->pm_token);
3290 pmap_inval_init(&info->inval);
3294 * Special handling for scanning one page, which is a very common
3295 * operation (it is?).
3297 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3299 if (info->sva + PAGE_SIZE == info->eva) {
3300 generation = pmap->pm_generation;
3301 if (info->sva >= VM_MAX_USER_ADDRESS) {
3303 * Kernel mappings do not track wire counts on
3304 * page table pages and only maintain pd_pv and
3305 * pte_pv levels so pmap_scan() works.
3308 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3309 ptep = vtopte(info->sva);
3312 * User pages which are unmanaged will not have a
3313 * pte_pv. User page table pages which are unmanaged
3314 * (shared from elsewhere) will also not have a pt_pv.
3315 * The func() callback will pass both pte_pv and pt_pv
3316 * as NULL in that case.
3318 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3319 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3320 if (pt_pv == NULL) {
3321 KKASSERT(pte_pv == NULL);
3322 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3324 ptep = pv_pte_lookup(pd_pv,
3325 pmap_pt_index(info->sva));
3327 info->func(pmap, info,
3336 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3340 * NOTE: *ptep can't be ripped out from under us if we hold
3341 * pte_pv locked, but bits can change. However, there is
3342 * a race where another thread may be inserting pte_pv
3343 * and setting *ptep just after our pte_pv lookup fails.
3345 * In this situation we can end up with a NULL pte_pv
3346 * but find that we have a managed *ptep. We explicitly
3347 * check for this race.
3353 * Unlike the pv_find() case below we actually
3354 * acquired a locked pv in this case so any
3355 * race should have been resolved. It is expected
3358 KKASSERT(pte_pv == NULL);
3359 } else if (pte_pv) {
3360 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3361 pmap->pmap_bits[PG_V_IDX])) ==
3362 (pmap->pmap_bits[PG_MANAGED_IDX] |
3363 pmap->pmap_bits[PG_V_IDX]),
3364 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3366 *ptep, oldpte, info->sva, pte_pv,
3367 generation, pmap->pm_generation));
3368 info->func(pmap, info, pte_pv, pt_pv, 0,
3369 info->sva, ptep, info->arg);
3372 * Check for insertion race
3374 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3376 pte_pv = pv_find(pmap,
3377 pmap_pte_pindex(info->sva));
3381 kprintf("pmap_scan: RACE1 "
3391 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3392 pmap->pmap_bits[PG_V_IDX])) ==
3393 pmap->pmap_bits[PG_V_IDX],
3394 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3396 *ptep, oldpte, info->sva,
3397 generation, pmap->pm_generation));
3398 info->func(pmap, info, NULL, pt_pv, 0,
3399 info->sva, ptep, info->arg);
3404 pmap_inval_done(&info->inval);
3405 lwkt_reltoken(&pmap->pm_token);
3410 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3413 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3414 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3416 if (info->sva >= VM_MAX_USER_ADDRESS) {
3418 * The kernel does not currently maintain any pv_entry's for
3419 * higher-level page tables.
3421 bzero(&dummy_pv, sizeof(dummy_pv));
3422 dummy_pv.pv_pindex = info->sva_pd_pindex;
3423 spin_lock(&pmap->pm_spin);
3424 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3425 pmap_scan_callback(&dummy_pv, info);
3426 ++dummy_pv.pv_pindex;
3428 spin_unlock(&pmap->pm_spin);
3431 * User page tables maintain local PML4, PDP, and PD
3432 * pv_entry's at the very least. PT pv's might be
3433 * unmanaged and thus not exist. PTE pv's might be
3434 * unmanaged and thus not exist.
3436 spin_lock(&pmap->pm_spin);
3437 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3438 pmap_scan_cmp, pmap_scan_callback, info);
3439 spin_unlock(&pmap->pm_spin);
3441 pmap_inval_done(&info->inval);
3442 lwkt_reltoken(&pmap->pm_token);
3446 * WARNING! pmap->pm_spin held
3449 pmap_scan_cmp(pv_entry_t pv, void *data)
3451 struct pmap_scan_info *info = data;
3452 if (pv->pv_pindex < info->sva_pd_pindex)
3454 if (pv->pv_pindex >= info->eva_pd_pindex)
3460 * WARNING! pmap->pm_spin held
3463 pmap_scan_callback(pv_entry_t pv, void *data)
3465 struct pmap_scan_info *info = data;
3466 struct pmap *pmap = info->pmap;
3467 pv_entry_t pd_pv; /* A page directory PV */
3468 pv_entry_t pt_pv; /* A page table PV */
3469 pv_entry_t pte_pv; /* A page table entry PV */
3474 vm_offset_t va_next;
3475 vm_pindex_t pd_pindex;
3480 * Pull the PD pindex from the pv before releasing the spinlock.
3482 * WARNING: pv is faked for kernel pmap scans.
3484 pd_pindex = pv->pv_pindex;
3485 spin_unlock(&pmap->pm_spin);
3486 pv = NULL; /* invalid after spinlock unlocked */
3489 * Calculate the page range within the PD. SIMPLE pmaps are
3490 * direct-mapped for the entire 2^64 address space. Normal pmaps
3491 * reflect the user and kernel address space which requires
3492 * cannonicalization w/regards to converting pd_pindex's back
3495 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3496 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3497 (sva & PML4_SIGNMASK)) {
3498 sva |= PML4_SIGNMASK;
3500 eva = sva + NBPDP; /* can overflow */
3501 if (sva < info->sva)
3503 if (eva < info->sva || eva > info->eva)
3507 * NOTE: kernel mappings do not track page table pages, only
3510 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3511 * However, for the scan to be efficient we try to
3512 * cache items top-down.
3517 for (; sva < eva; sva = va_next) {
3518 if (sva >= VM_MAX_USER_ADDRESS) {
3527 * PD cache (degenerate case if we skip). It is possible
3528 * for the PD to not exist due to races. This is ok.
3530 if (pd_pv == NULL) {
3531 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3532 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3534 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3536 if (pd_pv == NULL) {
3537 va_next = (sva + NBPDP) & ~PDPMASK;
3546 if (pt_pv == NULL) {
3551 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3552 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3558 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3562 * If pt_pv is NULL we either have an shared page table
3563 * page and must issue a callback specific to that case,
3564 * or there is no page table page.
3566 * Either way we can skip the page table page.
3568 if (pt_pv == NULL) {
3570 * Possible unmanaged (shared from another pmap)
3574 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3575 KKASSERT(pd_pv != NULL);
3576 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3577 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3578 info->func(pmap, info, NULL, pd_pv, 1,
3579 sva, ptep, info->arg);
3583 * Done, move to next page table page.
3585 va_next = (sva + NBPDR) & ~PDRMASK;
3592 * From this point in the loop testing pt_pv for non-NULL
3593 * means we are in UVM, else if it is NULL we are in KVM.
3595 * Limit our scan to either the end of the va represented
3596 * by the current page table page, or to the end of the
3597 * range being removed.
3600 va_next = (sva + NBPDR) & ~PDRMASK;
3607 * Scan the page table for pages. Some pages may not be
3608 * managed (might not have a pv_entry).
3610 * There is no page table management for kernel pages so
3611 * pt_pv will be NULL in that case, but otherwise pt_pv
3612 * is non-NULL, locked, and referenced.
3616 * At this point a non-NULL pt_pv means a UVA, and a NULL
3617 * pt_pv means a KVA.
3620 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3624 while (sva < va_next) {
3626 * Acquire the related pte_pv, if any. If *ptep == 0
3627 * the related pte_pv should not exist, but if *ptep
3628 * is not zero the pte_pv may or may not exist (e.g.
3629 * will not exist for an unmanaged page).
3631 * However a multitude of races are possible here.
3633 * In addition, the (pt_pv, pte_pv) lock order is
3634 * backwards, so we have to be careful in aquiring
3635 * a properly locked pte_pv.
3637 generation = pmap->pm_generation;
3639 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3646 pv_put(pt_pv); /* must be non-NULL */
3648 pv_lock(pte_pv); /* safe to block now */
3651 pt_pv = pv_get(pmap,
3652 pmap_pt_pindex(sva));
3654 * pt_pv reloaded, need new ptep
3656 KKASSERT(pt_pv != NULL);
3657 ptep = pv_pte_lookup(pt_pv,
3658 pmap_pte_index(sva));
3662 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3666 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3671 kprintf("Unexpected non-NULL pte_pv "
3673 "*ptep = %016lx/%016lx\n",
3674 pte_pv, pt_pv, *ptep, oldpte);
3675 panic("Unexpected non-NULL pte_pv");
3683 * Ready for the callback. The locked pte_pv (if any)
3684 * is consumed by the callback. pte_pv will exist if
3685 * the page is managed, and will not exist if it
3689 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3690 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3691 ("badC *ptep %016lx/%016lx sva %016lx "
3692 "pte_pv %p pm_generation %d/%d",
3693 *ptep, oldpte, sva, pte_pv,
3694 generation, pmap->pm_generation));
3695 info->func(pmap, info, pte_pv, pt_pv, 0,
3696 sva, ptep, info->arg);
3699 * Check for insertion race. Since there is no
3700 * pte_pv to guard us it is possible for us
3701 * to race another thread doing an insertion.
3702 * Our lookup misses the pte_pv but our *ptep
3703 * check sees the inserted pte.
3705 * XXX panic case seems to occur within a
3706 * vm_fork() of /bin/sh, which frankly
3707 * shouldn't happen since no other threads
3708 * should be inserting to our pmap in that
3709 * situation. Removing, possibly. Inserting,
3712 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3714 pte_pv = pv_find(pmap,
3715 pmap_pte_pindex(sva));
3718 kprintf("pmap_scan: RACE2 "
3728 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3729 pmap->pmap_bits[PG_V_IDX],
3730 ("badD *ptep %016lx/%016lx sva %016lx "
3731 "pte_pv NULL pm_generation %d/%d",
3733 generation, pmap->pm_generation));
3734 info->func(pmap, info, NULL, pt_pv, 0,
3735 sva, ptep, info->arg);
3754 * Relock before returning.
3756 spin_lock(&pmap->pm_spin);
3761 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3763 struct pmap_scan_info info;
3768 info.func = pmap_remove_callback;
3770 info.doinval = 1; /* normal remove requires pmap inval */
3775 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3777 struct pmap_scan_info info;
3782 info.func = pmap_remove_callback;
3784 info.doinval = 0; /* normal remove requires pmap inval */
3789 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
3790 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3791 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3797 * This will also drop pt_pv's wire_count. Note that
3798 * terminal pages are not wired based on mmu presence.
3801 pmap_remove_pv_pte(pte_pv, pt_pv, &info->inval);
3803 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3804 pmap_remove_pv_page(pte_pv);
3806 } else if (sharept == 0) {
3808 * Unmanaged page table (pt, pd, or pdp. Not pte).
3810 * pt_pv's wire_count is still bumped by unmanaged pages
3811 * so we must decrement it manually.
3813 * We have to unwire the target page table page.
3815 * It is unclear how we can invalidate a segment so we
3816 * invalidate -1 which invlidates the tlb.
3819 pmap_inval_interlock(&info->inval, pmap, -1);
3820 pte = pte_load_clear(ptep);
3822 pmap_inval_deinterlock(&info->inval, pmap);
3823 if (pte & pmap->pmap_bits[PG_W_IDX])
3824 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3825 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3826 if (vm_page_unwire_quick(pt_pv->pv_m))
3827 panic("pmap_remove: insufficient wirecount");
3830 * Unmanaged page table (pt, pd, or pdp. Not pte) for
3831 * a shared page table.
3833 * pt_pv is actually the pd_pv for our pmap (not the shared
3836 * We have to unwire the target page table page and we
3837 * have to unwire our page directory page.
3839 * It is unclear how we can invalidate a segment so we
3840 * invalidate -1 which invlidates the tlb.
3843 pmap_inval_interlock(&info->inval, pmap, -1);
3844 pte = pte_load_clear(ptep);
3846 pmap_inval_deinterlock(&info->inval, pmap);
3847 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3848 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
3849 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3850 panic("pmap_remove: shared pgtable1 bad wirecount");
3851 if (vm_page_unwire_quick(pt_pv->pv_m))
3852 panic("pmap_remove: shared pgtable2 bad wirecount");
3857 * Removes this physical page from all physical maps in which it resides.
3858 * Reflects back modify bits to the pager.
3860 * This routine may not be called from an interrupt.
3864 pmap_remove_all(vm_page_t m)
3866 struct pmap_inval_info info;
3869 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
3872 pmap_inval_init(&info);
3873 vm_page_spin_lock(m);
3874 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3875 KKASSERT(pv->pv_m == m);
3876 if (pv_hold_try(pv)) {
3877 vm_page_spin_unlock(m);
3879 vm_page_spin_unlock(m);
3882 if (pv->pv_m != m) {
3884 vm_page_spin_lock(m);
3889 * Holding no spinlocks, pv is locked.
3891 pmap_remove_pv_pte(pv, NULL, &info);
3892 pmap_remove_pv_page(pv);
3894 vm_page_spin_lock(m);
3896 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3897 vm_page_spin_unlock(m);
3898 pmap_inval_done(&info);
3902 * Set the physical protection on the specified range of this map
3903 * as requested. This function is typically only used for debug watchpoints
3906 * This function may not be called from an interrupt if the map is
3907 * not the kernel_pmap.
3909 * NOTE! For shared page table pages we just unmap the page.
3912 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3914 struct pmap_scan_info info;
3915 /* JG review for NX */
3919 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3920 pmap_remove(pmap, sva, eva);
3923 if (prot & VM_PROT_WRITE)
3928 info.func = pmap_protect_callback;
3936 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
3937 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3938 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3948 pmap_inval_interlock(&info->inval, pmap, va);
3954 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
3955 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3956 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3957 KKASSERT(m == pte_pv->pv_m);
3958 vm_page_flag_set(m, PG_REFERENCED);
3960 cbits &= ~pmap->pmap_bits[PG_A_IDX];
3962 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
3963 if (pmap_track_modified(pte_pv->pv_pindex)) {
3964 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3966 m = PHYS_TO_VM_PAGE(pbits &
3971 cbits &= ~pmap->pmap_bits[PG_M_IDX];
3974 } else if (sharept) {
3976 * Unmanaged page table, pt_pv is actually the pd_pv
3977 * for our pmap (not the object's shared pmap).
3979 * When asked to protect something in a shared page table
3980 * page we just unmap the page table page. We have to
3981 * invalidate the tlb in this situation.
3983 * XXX Warning, shared page tables will not be used for
3984 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
3985 * so PHYS_TO_VM_PAGE() should be safe here.
3987 pte = pte_load_clear(ptep);
3988 pmap_inval_invltlb(&info->inval);
3989 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3990 panic("pmap_protect: pgtable1 pg bad wirecount");
3991 if (vm_page_unwire_quick(pt_pv->pv_m))
3992 panic("pmap_protect: pgtable2 pg bad wirecount");
3995 /* else unmanaged page, adjust bits, no wire changes */
3998 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4000 if (pmap_enter_debug > 0) {
4002 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4003 "pt_pv=%p cbits=%08lx\n",
4009 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
4013 pmap_inval_deinterlock(&info->inval, pmap);
4019 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4020 * mapping at that address. Set protection and wiring as requested.
4022 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4023 * possible. If it is we enter the page into the appropriate shared pmap
4024 * hanging off the related VM object instead of the passed pmap, then we
4025 * share the page table page from the VM object's pmap into the current pmap.
4027 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4031 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4032 boolean_t wired, vm_map_entry_t entry)
4034 pmap_inval_info info;
4035 pv_entry_t pt_pv; /* page table */
4036 pv_entry_t pte_pv; /* page table entry */
4039 pt_entry_t origpte, newpte;
4044 va = trunc_page(va);
4045 #ifdef PMAP_DIAGNOSTIC
4047 panic("pmap_enter: toobig");
4048 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4049 panic("pmap_enter: invalid to pmap_enter page table "
4050 "pages (va: 0x%lx)", va);
4052 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4053 kprintf("Warning: pmap_enter called on UVA with "
4056 db_print_backtrace();
4059 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4060 kprintf("Warning: pmap_enter called on KVA without"
4063 db_print_backtrace();
4068 * Get locked PV entries for our new page table entry (pte_pv)
4069 * and for its parent page table (pt_pv). We need the parent
4070 * so we can resolve the location of the ptep.
4072 * Only hardware MMU actions can modify the ptep out from
4075 * if (m) is fictitious or unmanaged we do not create a managing
4076 * pte_pv for it. Any pre-existing page's management state must
4077 * match (avoiding code complexity).
4079 * If the pmap is still being initialized we assume existing
4082 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4084 if (pmap_initialized == FALSE) {
4089 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4091 if (va >= VM_MAX_USER_ADDRESS) {
4095 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4097 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4101 KKASSERT(origpte == 0 ||
4102 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0);
4104 if (va >= VM_MAX_USER_ADDRESS) {
4106 * Kernel map, pv_entry-tracked.
4109 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4115 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4117 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4121 KKASSERT(origpte == 0 ||
4122 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]));
4125 pa = VM_PAGE_TO_PHYS(m);
4126 opa = origpte & PG_FRAME;
4128 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4129 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4131 newpte |= pmap->pmap_bits[PG_W_IDX];
4132 if (va < VM_MAX_USER_ADDRESS)
4133 newpte |= pmap->pmap_bits[PG_U_IDX];
4135 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4136 // if (pmap == &kernel_pmap)
4137 // newpte |= pgeflag;
4138 newpte |= pmap->pmap_cache_bits[m->pat_mode];
4139 if (m->flags & PG_FICTITIOUS)
4140 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4143 * It is possible for multiple faults to occur in threaded
4144 * environments, the existing pte might be correct.
4146 if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4147 pmap->pmap_bits[PG_A_IDX])) == 0)
4150 if ((prot & VM_PROT_NOSYNC) == 0)
4151 pmap_inval_init(&info);
4154 * Ok, either the address changed or the protection or wiring
4157 * Clear the current entry, interlocking the removal. For managed
4158 * pte's this will also flush the modified state to the vm_page.
4159 * Atomic ops are mandatory in order to ensure that PG_M events are
4160 * not lost during any transition.
4162 * WARNING: The caller has busied the new page but not the original
4163 * vm_page which we are trying to replace. Because we hold
4164 * the pte_pv lock, but have not busied the page, PG bits
4165 * can be cleared out from under us.
4170 * pmap_remove_pv_pte() unwires pt_pv and assumes
4171 * we will free pte_pv, but since we are reusing
4172 * pte_pv we want to retain the wire count.
4174 * pt_pv won't exist for a kernel page (managed or
4178 vm_page_wire_quick(pt_pv->pv_m);
4179 if (prot & VM_PROT_NOSYNC)
4180 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
4182 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
4184 pmap_remove_pv_page(pte_pv);
4185 } else if (prot & VM_PROT_NOSYNC) {
4187 * Unmanaged page, NOSYNC (no mmu sync) requested.
4189 * Leave wire count on PT page intact.
4191 (void)pte_load_clear(ptep);
4192 cpu_invlpg((void *)va);
4193 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4196 * Unmanaged page, normal enter.
4198 * Leave wire count on PT page intact.
4200 pmap_inval_interlock(&info, pmap, va);
4201 (void)pte_load_clear(ptep);
4202 pmap_inval_deinterlock(&info, pmap);
4203 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4205 KKASSERT(*ptep == 0);
4209 if (pmap_enter_debug > 0) {
4211 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4212 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4214 origpte, newpte, ptep,
4215 pte_pv, pt_pv, opa, prot);
4221 * Enter on the PV list if part of our managed memory.
4222 * Wiring of the PT page is already handled.
4224 KKASSERT(pte_pv->pv_m == NULL);
4225 vm_page_spin_lock(m);
4227 pmap_page_stats_adding(m);
4228 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4229 vm_page_flag_set(m, PG_MAPPED);
4230 vm_page_spin_unlock(m);
4231 } else if (pt_pv && opa == 0) {
4233 * We have to adjust the wire count on the PT page ourselves
4234 * for unmanaged entries. If opa was non-zero we retained
4235 * the existing wire count from the removal.
4237 vm_page_wire_quick(pt_pv->pv_m);
4241 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4243 * User VMAs do not because those will be zero->non-zero, so no
4244 * stale entries to worry about at this point.
4246 * For KVM there appear to still be issues. Theoretically we
4247 * should be able to scrap the interlocks entirely but we
4250 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
4251 pmap_inval_interlock(&info, pmap, va);
4256 *(volatile pt_entry_t *)ptep = newpte;
4258 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
4259 pmap_inval_deinterlock(&info, pmap);
4260 else if (pt_pv == NULL)
4261 cpu_invlpg((void *)va);
4265 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4268 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4271 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4272 vm_page_flag_set(m, PG_WRITEABLE);
4275 * Unmanaged pages need manual resident_count tracking.
4277 if (pte_pv == NULL && pt_pv)
4278 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4283 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
4284 pmap_inval_done(&info);
4286 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4287 (m->flags & PG_MAPPED));
4290 * Cleanup the pv entry, allowing other accessors.
4299 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4300 * This code also assumes that the pmap has no pre-existing entry for this
4303 * This code currently may only be used on user pmaps, not kernel_pmap.
4306 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4308 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4312 * Make a temporary mapping for a physical address. This is only intended
4313 * to be used for panic dumps.
4315 * The caller is responsible for calling smp_invltlb().
4318 pmap_kenter_temporary(vm_paddr_t pa, long i)
4320 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4321 return ((void *)crashdumpmap);
4324 #define MAX_INIT_PT (96)
4327 * This routine preloads the ptes for a given object into the specified pmap.
4328 * This eliminates the blast of soft faults on process startup and
4329 * immediately after an mmap.
4331 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4334 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4335 vm_object_t object, vm_pindex_t pindex,
4336 vm_size_t size, int limit)
4338 struct rb_vm_page_scan_info info;
4343 * We can't preinit if read access isn't set or there is no pmap
4346 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4350 * We can't preinit if the pmap is not the current pmap
4352 lp = curthread->td_lwp;
4353 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4357 * Misc additional checks
4359 psize = x86_64_btop(size);
4361 if ((object->type != OBJT_VNODE) ||
4362 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4363 (object->resident_page_count > MAX_INIT_PT))) {
4367 if (pindex + psize > object->size) {
4368 if (object->size < pindex)
4370 psize = object->size - pindex;
4377 * If everything is segment-aligned do not pre-init here. Instead
4378 * allow the normal vm_fault path to pass a segment hint to
4379 * pmap_enter() which will then use an object-referenced shared
4382 if ((addr & SEG_MASK) == 0 &&
4383 (ctob(psize) & SEG_MASK) == 0 &&
4384 (ctob(pindex) & SEG_MASK) == 0) {
4389 * Use a red-black scan to traverse the requested range and load
4390 * any valid pages found into the pmap.
4392 * We cannot safely scan the object's memq without holding the
4395 info.start_pindex = pindex;
4396 info.end_pindex = pindex + psize - 1;
4402 vm_object_hold_shared(object);
4403 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4404 pmap_object_init_pt_callback, &info);
4405 vm_object_drop(object);
4410 pmap_object_init_pt_callback(vm_page_t p, void *data)
4412 struct rb_vm_page_scan_info *info = data;
4413 vm_pindex_t rel_index;
4416 * don't allow an madvise to blow away our really
4417 * free pages allocating pv entries.
4419 if ((info->limit & MAP_PREFAULT_MADVISE) &&
4420 vmstats.v_free_count < vmstats.v_free_reserved) {
4425 * Ignore list markers and ignore pages we cannot instantly
4426 * busy (while holding the object token).
4428 if (p->flags & PG_MARKER)
4430 if (vm_page_busy_try(p, TRUE))
4432 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4433 (p->flags & PG_FICTITIOUS) == 0) {
4434 if ((p->queue - p->pc) == PQ_CACHE)
4435 vm_page_deactivate(p);
4436 rel_index = p->pindex - info->start_pindex;
4437 pmap_enter_quick(info->pmap,
4438 info->addr + x86_64_ptob(rel_index), p);
4446 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4449 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4452 * XXX This is safe only because page table pages are not freed.
4455 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4459 /*spin_lock(&pmap->pm_spin);*/
4460 if ((pte = pmap_pte(pmap, addr)) != NULL) {
4461 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4462 /*spin_unlock(&pmap->pm_spin);*/
4466 /*spin_unlock(&pmap->pm_spin);*/
4471 * Change the wiring attribute for a pmap/va pair. The mapping must already
4472 * exist in the pmap. The mapping may or may not be managed.
4475 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4476 vm_map_entry_t entry)
4483 lwkt_gettoken(&pmap->pm_token);
4484 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
4485 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4487 if (wired && !pmap_pte_w(pmap, ptep))
4488 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4489 else if (!wired && pmap_pte_w(pmap, ptep))
4490 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4493 * Wiring is not a hardware characteristic so there is no need to
4494 * invalidate TLB. However, in an SMP environment we must use
4495 * a locked bus cycle to update the pte (if we are not using
4496 * the pmap_inval_*() API that is)... it's ok to do this for simple
4500 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4502 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4504 lwkt_reltoken(&pmap->pm_token);
4510 * Copy the range specified by src_addr/len from the source map to
4511 * the range dst_addr/len in the destination map.
4513 * This routine is only advisory and need not do anything.
4516 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4517 vm_size_t len, vm_offset_t src_addr)
4524 * Zero the specified physical page.
4526 * This function may be called from an interrupt and no locking is
4530 pmap_zero_page(vm_paddr_t phys)
4532 vm_offset_t va = PHYS_TO_DMAP(phys);
4534 pagezero((void *)va);
4538 * pmap_page_assertzero:
4540 * Assert that a page is empty, panic if it isn't.
4543 pmap_page_assertzero(vm_paddr_t phys)
4545 vm_offset_t va = PHYS_TO_DMAP(phys);
4548 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
4549 if (*(long *)((char *)va + i) != 0) {
4550 panic("pmap_page_assertzero() @ %p not zero!",
4551 (void *)(intptr_t)va);
4559 * Zero part of a physical page by mapping it into memory and clearing
4560 * its contents with bzero.
4562 * off and size may not cover an area beyond a single hardware page.
4565 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4567 vm_offset_t virt = PHYS_TO_DMAP(phys);
4569 bzero((char *)virt + off, size);
4575 * Copy the physical page from the source PA to the target PA.
4576 * This function may be called from an interrupt. No locking
4580 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4582 vm_offset_t src_virt, dst_virt;
4584 src_virt = PHYS_TO_DMAP(src);
4585 dst_virt = PHYS_TO_DMAP(dst);
4586 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4590 * pmap_copy_page_frag:
4592 * Copy the physical page from the source PA to the target PA.
4593 * This function may be called from an interrupt. No locking
4597 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4599 vm_offset_t src_virt, dst_virt;
4601 src_virt = PHYS_TO_DMAP(src);
4602 dst_virt = PHYS_TO_DMAP(dst);
4604 bcopy((char *)src_virt + (src & PAGE_MASK),
4605 (char *)dst_virt + (dst & PAGE_MASK),
4610 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4611 * this page. This count may be changed upwards or downwards in the future;
4612 * it is only necessary that true be returned for a small subset of pmaps
4613 * for proper page aging.
4616 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4621 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4624 vm_page_spin_lock(m);
4625 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4626 if (pv->pv_pmap == pmap) {
4627 vm_page_spin_unlock(m);
4634 vm_page_spin_unlock(m);
4639 * Remove all pages from specified address space this aids process exit
4640 * speeds. Also, this code may be special cased for the current process
4644 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4646 pmap_remove_noinval(pmap, sva, eva);
4651 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4652 * routines are inline, and a lot of things compile-time evaluate.
4656 pmap_testbit(vm_page_t m, int bit)
4662 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4665 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4667 vm_page_spin_lock(m);
4668 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4669 vm_page_spin_unlock(m);
4673 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4675 #if defined(PMAP_DIAGNOSTIC)
4676 if (pv->pv_pmap == NULL) {
4677 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4685 * If the bit being tested is the modified bit, then
4686 * mark clean_map and ptes as never
4689 * WARNING! Because we do not lock the pv, *pte can be in a
4690 * state of flux. Despite this the value of *pte
4691 * will still be related to the vm_page in some way
4692 * because the pv cannot be destroyed as long as we
4693 * hold the vm_page spin lock.
4695 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4696 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4697 if (!pmap_track_modified(pv->pv_pindex))
4701 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4702 if (*pte & pmap->pmap_bits[bit]) {
4703 vm_page_spin_unlock(m);
4707 vm_page_spin_unlock(m);
4712 * This routine is used to modify bits in ptes. Only one bit should be
4713 * specified. PG_RW requires special handling.
4715 * Caller must NOT hold any spin locks
4719 pmap_clearbit(vm_page_t m, int bit_index)
4721 struct pmap_inval_info info;
4727 if (bit_index == PG_RW_IDX)
4728 vm_page_flag_clear(m, PG_WRITEABLE);
4729 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4736 * Loop over all current mappings setting/clearing as appropos If
4737 * setting RO do we need to clear the VAC?
4739 * NOTE: When clearing PG_M we could also (not implemented) drop
4740 * through to the PG_RW code and clear PG_RW too, forcing
4741 * a fault on write to redetect PG_M for virtual kernels, but
4742 * it isn't necessary since virtual kernels invalidate the
4743 * pte when they clear the VPTE_M bit in their virtual page
4746 * NOTE: Does not re-dirty the page when clearing only PG_M.
4748 * NOTE: Because we do not lock the pv, *pte can be in a state of
4749 * flux. Despite this the value of *pte is still somewhat
4750 * related while we hold the vm_page spin lock.
4752 * *pte can be zero due to this race. Since we are clearing
4753 * bits we basically do no harm when this race ccurs.
4755 if (bit_index != PG_RW_IDX) {
4756 vm_page_spin_lock(m);
4757 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4758 #if defined(PMAP_DIAGNOSTIC)
4759 if (pv->pv_pmap == NULL) {
4760 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4766 pte = pmap_pte_quick(pv->pv_pmap,
4767 pv->pv_pindex << PAGE_SHIFT);
4769 if (pbits & pmap->pmap_bits[bit_index])
4770 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
4772 vm_page_spin_unlock(m);
4777 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4780 pmap_inval_init(&info);
4783 vm_page_spin_lock(m);
4784 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4786 * don't write protect pager mappings
4788 if (!pmap_track_modified(pv->pv_pindex))
4791 #if defined(PMAP_DIAGNOSTIC)
4792 if (pv->pv_pmap == NULL) {
4793 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4800 * Skip pages which do not have PG_RW set.
4802 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4803 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
4809 if (pv_hold_try(pv)) {
4810 vm_page_spin_unlock(m);
4812 vm_page_spin_unlock(m);
4813 pv_lock(pv); /* held, now do a blocking lock */
4815 if (pv->pv_pmap != pmap || pv->pv_m != m) {
4816 pv_put(pv); /* and release */
4817 goto restart; /* anything could have happened */
4819 pmap_inval_interlock(&info, pmap,
4820 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
4821 KKASSERT(pv->pv_pmap == pmap);
4825 if (atomic_cmpset_long(pte, pbits, pbits &
4826 ~(pmap->pmap_bits[PG_RW_IDX] |
4827 pmap->pmap_bits[PG_M_IDX]))) {
4831 pmap_inval_deinterlock(&info, pmap);
4832 vm_page_spin_lock(m);
4835 * If PG_M was found to be set while we were clearing PG_RW
4836 * we also clear PG_M (done above) and mark the page dirty.
4837 * Callers expect this behavior.
4839 if (pbits & pmap->pmap_bits[PG_M_IDX])
4843 vm_page_spin_unlock(m);
4844 pmap_inval_done(&info);
4848 * Lower the permission for all mappings to a given page.
4850 * Page must be busied by caller. Because page is busied by caller this
4851 * should not be able to race a pmap_enter().
4854 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4856 /* JG NX support? */
4857 if ((prot & VM_PROT_WRITE) == 0) {
4858 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4860 * NOTE: pmap_clearbit(.. PG_RW) also clears
4861 * the PG_WRITEABLE flag in (m).
4863 pmap_clearbit(m, PG_RW_IDX);
4871 pmap_phys_address(vm_pindex_t ppn)
4873 return (x86_64_ptob(ppn));
4877 * Return a count of reference bits for a page, clearing those bits.
4878 * It is not necessary for every reference bit to be cleared, but it
4879 * is necessary that 0 only be returned when there are truly no
4880 * reference bits set.
4882 * XXX: The exact number of bits to check and clear is a matter that
4883 * should be tested and standardized at some point in the future for
4884 * optimal aging of shared pages.
4886 * This routine may not block.
4889 pmap_ts_referenced(vm_page_t m)
4896 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4899 vm_page_spin_lock(m);
4900 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4901 if (!pmap_track_modified(pv->pv_pindex))
4904 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4905 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
4906 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
4912 vm_page_spin_unlock(m);
4919 * Return whether or not the specified physical page was modified
4920 * in any physical maps.
4923 pmap_is_modified(vm_page_t m)
4927 res = pmap_testbit(m, PG_M_IDX);
4932 * Clear the modify bits on the specified physical page.
4935 pmap_clear_modify(vm_page_t m)
4937 pmap_clearbit(m, PG_M_IDX);
4941 * pmap_clear_reference:
4943 * Clear the reference bit on the specified physical page.
4946 pmap_clear_reference(vm_page_t m)
4948 pmap_clearbit(m, PG_A_IDX);
4952 * Miscellaneous support routines follow
4957 i386_protection_init(void)
4961 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4962 kp = protection_codes;
4963 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
4965 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4967 * Read access is also 0. There isn't any execute bit,
4968 * so just make it readable.
4970 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4971 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4972 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4975 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4976 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4977 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4978 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4979 *kp++ = pmap_bits_default[PG_RW_IDX];
4986 * Map a set of physical memory pages into the kernel virtual
4987 * address space. Return a pointer to where it is mapped. This
4988 * routine is intended to be used for mapping device memory,
4991 * NOTE: We can't use pgeflag unless we invalidate the pages one at
4994 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
4995 * work whether the cpu supports PAT or not. The remaining PAT
4996 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5000 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5002 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5006 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5008 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5012 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5014 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5018 * Map a set of physical memory pages into the kernel virtual
5019 * address space. Return a pointer to where it is mapped. This
5020 * routine is intended to be used for mapping device memory,
5024 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5026 vm_offset_t va, tmpva, offset;
5030 offset = pa & PAGE_MASK;
5031 size = roundup(offset + size, PAGE_SIZE);
5033 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
5035 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5037 pa = pa & ~PAGE_MASK;
5038 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5039 pte = vtopte(tmpva);
5041 kernel_pmap.pmap_bits[PG_RW_IDX] |
5042 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5043 kernel_pmap.pmap_cache_bits[mode];
5044 tmpsize -= PAGE_SIZE;
5048 pmap_invalidate_range(&kernel_pmap, va, va + size);
5049 pmap_invalidate_cache_range(va, va + size);
5051 return ((void *)(va + offset));
5055 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5057 vm_offset_t base, offset;
5059 base = va & ~PAGE_MASK;
5060 offset = va & PAGE_MASK;
5061 size = roundup(offset + size, PAGE_SIZE);
5062 pmap_qremove(va, size >> PAGE_SHIFT);
5063 kmem_free(&kernel_map, base, size);
5067 * Sets the memory attribute for the specified page.
5070 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5076 * If "m" is a normal page, update its direct mapping. This update
5077 * can be relied upon to perform any cache operations that are
5078 * required for data coherence.
5080 if ((m->flags & PG_FICTITIOUS) == 0)
5081 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE,
5086 * Change the PAT attribute on an existing kernel memory map. Caller
5087 * must ensure that the virtual memory in question is not accessed
5088 * during the adjustment.
5091 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5098 panic("pmap_change_attr: va is NULL");
5099 base = trunc_page(va);
5103 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5104 kernel_pmap.pmap_cache_bits[mode];
5109 changed = 1; /* XXX: not optimal */
5112 * Flush CPU caches if required to make sure any data isn't cached that
5113 * shouldn't be, etc.
5116 pmap_invalidate_range(&kernel_pmap, base, va);
5117 pmap_invalidate_cache_range(base, va);
5122 * perform the pmap work for mincore
5125 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5127 pt_entry_t *ptep, pte;
5131 lwkt_gettoken(&pmap->pm_token);
5132 ptep = pmap_pte(pmap, addr);
5134 if (ptep && (pte = *ptep) != 0) {
5137 val = MINCORE_INCORE;
5138 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5141 pa = pte & PG_FRAME;
5143 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5146 m = PHYS_TO_VM_PAGE(pa);
5151 if (pte & pmap->pmap_bits[PG_M_IDX])
5152 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5154 * Modified by someone
5156 else if (m && (m->dirty || pmap_is_modified(m)))
5157 val |= MINCORE_MODIFIED_OTHER;
5161 if (pte & pmap->pmap_bits[PG_A_IDX])
5162 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5165 * Referenced by someone
5167 else if (m && ((m->flags & PG_REFERENCED) ||
5168 pmap_ts_referenced(m))) {
5169 val |= MINCORE_REFERENCED_OTHER;
5170 vm_page_flag_set(m, PG_REFERENCED);
5174 lwkt_reltoken(&pmap->pm_token);
5180 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5181 * vmspace will be ref'd and the old one will be deref'd.
5183 * The vmspace for all lwps associated with the process will be adjusted
5184 * and cr3 will be reloaded if any lwp is the current lwp.
5186 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5189 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5191 struct vmspace *oldvm;
5194 oldvm = p->p_vmspace;
5195 if (oldvm != newvm) {
5197 sysref_get(&newvm->vm_sysref);
5198 p->p_vmspace = newvm;
5199 KKASSERT(p->p_nthreads == 1);
5200 lp = RB_ROOT(&p->p_lwp_tree);
5201 pmap_setlwpvm(lp, newvm);
5203 sysref_put(&oldvm->vm_sysref);
5208 * Set the vmspace for a LWP. The vmspace is almost universally set the
5209 * same as the process vmspace, but virtual kernels need to swap out contexts
5210 * on a per-lwp basis.
5212 * Caller does not necessarily hold any vmspace tokens. Caller must control
5213 * the lwp (typically be in the context of the lwp). We use a critical
5214 * section to protect against statclock and hardclock (statistics collection).
5217 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5219 struct vmspace *oldvm;
5222 oldvm = lp->lwp_vmspace;
5224 if (oldvm != newvm) {
5226 lp->lwp_vmspace = newvm;
5227 if (curthread->td_lwp == lp) {
5228 pmap = vmspace_pmap(newvm);
5229 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5230 if (pmap->pm_active_lock & CPULOCK_EXCL)
5231 pmap_interlock_wait(newvm);
5232 #if defined(SWTCH_OPTIM_STATS)
5235 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5236 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5237 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5238 curthread->td_pcb->pcb_cr3 = KPML4phys;
5240 panic("pmap_setlwpvm: unknown pmap type\n");
5242 load_cr3(curthread->td_pcb->pcb_cr3);
5243 pmap = vmspace_pmap(oldvm);
5244 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5252 * Called when switching to a locked pmap, used to interlock against pmaps
5253 * undergoing modifications to prevent us from activating the MMU for the
5254 * target pmap until all such modifications have completed. We have to do
5255 * this because the thread making the modifications has already set up its
5256 * SMP synchronization mask.
5258 * This function cannot sleep!
5263 pmap_interlock_wait(struct vmspace *vm)
5265 struct pmap *pmap = &vm->vm_pmap;
5267 if (pmap->pm_active_lock & CPULOCK_EXCL) {
5269 KKASSERT(curthread->td_critcount >= 2);
5270 DEBUG_PUSH_INFO("pmap_interlock_wait");
5271 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5273 lwkt_process_ipiq();
5281 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5284 if ((obj == NULL) || (size < NBPDR) ||
5285 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5289 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
5294 * Used by kmalloc/kfree, page already exists at va
5297 pmap_kvtom(vm_offset_t va)
5299 pt_entry_t *ptep = vtopte(va);
5301 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5302 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5306 * Initialize machine-specific shared page directory support. This
5307 * is executed when a VM object is created.
5310 pmap_object_init(vm_object_t object)
5312 object->md.pmap_rw = NULL;
5313 object->md.pmap_ro = NULL;
5317 * Clean up machine-specific shared page directory support. This
5318 * is executed when a VM object is destroyed.
5321 pmap_object_free(vm_object_t object)
5325 if ((pmap = object->md.pmap_rw) != NULL) {
5326 object->md.pmap_rw = NULL;
5327 pmap_remove_noinval(pmap,
5328 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5329 CPUMASK_ASSZERO(pmap->pm_active);
5332 kfree(pmap, M_OBJPMAP);
5334 if ((pmap = object->md.pmap_ro) != NULL) {
5335 object->md.pmap_ro = NULL;
5336 pmap_remove_noinval(pmap,
5337 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5338 CPUMASK_ASSZERO(pmap->pm_active);
5341 kfree(pmap, M_OBJPMAP);