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;
233 static int pmap_yield_count = 64;
234 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
235 &pmap_yield_count, 0, "Yield during init_pt/release");
236 static int pmap_mmu_optimize = 0;
237 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
238 &pmap_mmu_optimize, 0, "Share page table pages when possible");
242 /* Standard user access funtions */
243 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
245 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
246 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
247 extern int std_fubyte (const void *base);
248 extern int std_subyte (void *base, int byte);
249 extern long std_fuword (const void *base);
250 extern int std_suword (void *base, long word);
251 extern int std_suword32 (void *base, int word);
253 static void pv_hold(pv_entry_t pv);
254 static int _pv_hold_try(pv_entry_t pv
256 static void pv_drop(pv_entry_t pv);
257 static void _pv_lock(pv_entry_t pv
259 static void pv_unlock(pv_entry_t pv);
260 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
262 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
264 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
265 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
266 static void pv_put(pv_entry_t pv);
267 static void pv_free(pv_entry_t pv);
268 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
269 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
271 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
272 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
273 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
274 struct pmap_inval_info *info);
275 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
276 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp);
278 struct pmap_scan_info;
279 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
280 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
281 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
282 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
283 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
284 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
286 static void i386_protection_init (void);
287 static void create_pagetables(vm_paddr_t *firstaddr);
288 static void pmap_remove_all (vm_page_t m);
289 static boolean_t pmap_testbit (vm_page_t m, int bit);
291 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
292 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
294 static void pmap_pinit_defaults(struct pmap *pmap);
296 static unsigned pdir4mb;
299 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
301 if (pv1->pv_pindex < pv2->pv_pindex)
303 if (pv1->pv_pindex > pv2->pv_pindex)
308 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
309 pv_entry_compare, vm_pindex_t, pv_pindex);
313 pmap_page_stats_adding(vm_page_t m)
315 globaldata_t gd = mycpu;
317 if (TAILQ_EMPTY(&m->md.pv_list)) {
318 ++gd->gd_vmtotal.t_arm;
319 } else if (TAILQ_FIRST(&m->md.pv_list) ==
320 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
321 ++gd->gd_vmtotal.t_armshr;
322 ++gd->gd_vmtotal.t_avmshr;
324 ++gd->gd_vmtotal.t_avmshr;
330 pmap_page_stats_deleting(vm_page_t m)
332 globaldata_t gd = mycpu;
334 if (TAILQ_EMPTY(&m->md.pv_list)) {
335 --gd->gd_vmtotal.t_arm;
336 } else if (TAILQ_FIRST(&m->md.pv_list) ==
337 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
338 --gd->gd_vmtotal.t_armshr;
339 --gd->gd_vmtotal.t_avmshr;
341 --gd->gd_vmtotal.t_avmshr;
346 * Move the kernel virtual free pointer to the next
347 * 2MB. This is used to help improve performance
348 * by using a large (2MB) page for much of the kernel
349 * (.text, .data, .bss)
353 pmap_kmem_choose(vm_offset_t addr)
355 vm_offset_t newaddr = addr;
357 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
364 * Super fast pmap_pte routine best used when scanning the pv lists.
365 * This eliminates many course-grained invltlb calls. Note that many of
366 * the pv list scans are across different pmaps and it is very wasteful
367 * to do an entire invltlb when checking a single mapping.
369 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
373 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
375 return pmap_pte(pmap, va);
379 * Returns the pindex of a page table entry (representing a terminal page).
380 * There are NUPTE_TOTAL page table entries possible (a huge number)
382 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
383 * We want to properly translate negative KVAs.
387 pmap_pte_pindex(vm_offset_t va)
389 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
393 * Returns the pindex of a page table.
397 pmap_pt_pindex(vm_offset_t va)
399 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
403 * Returns the pindex of a page directory.
407 pmap_pd_pindex(vm_offset_t va)
409 return (NUPTE_TOTAL + NUPT_TOTAL +
410 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
415 pmap_pdp_pindex(vm_offset_t va)
417 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
418 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
423 pmap_pml4_pindex(void)
425 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
429 * Return various clipped indexes for a given VA
431 * Returns the index of a pte in a page table, representing a terminal
436 pmap_pte_index(vm_offset_t va)
438 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
442 * Returns the index of a pt in a page directory, representing a page
447 pmap_pt_index(vm_offset_t va)
449 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
453 * Returns the index of a pd in a page directory page, representing a page
458 pmap_pd_index(vm_offset_t va)
460 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
464 * Returns the index of a pdp in the pml4 table, representing a page
469 pmap_pdp_index(vm_offset_t va)
471 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
475 * Generic procedure to index a pte from a pt, pd, or pdp.
477 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
478 * a page table page index but is instead of PV lookup index.
482 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
486 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
487 return(&pte[pindex]);
491 * Return pointer to PDP slot in the PML4
495 pmap_pdp(pmap_t pmap, vm_offset_t va)
497 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
501 * Return pointer to PD slot in the PDP given a pointer to the PDP
505 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
509 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
510 return (&pd[pmap_pd_index(va)]);
514 * Return pointer to PD slot in the PDP.
518 pmap_pd(pmap_t pmap, vm_offset_t va)
522 pdp = pmap_pdp(pmap, va);
523 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
525 return (pmap_pdp_to_pd(*pdp, va));
529 * Return pointer to PT slot in the PD given a pointer to the PD
533 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
537 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
538 return (&pt[pmap_pt_index(va)]);
542 * Return pointer to PT slot in the PD
544 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
545 * so we cannot lookup the PD via the PDP. Instead we
546 * must look it up via the pmap.
550 pmap_pt(pmap_t pmap, vm_offset_t va)
554 vm_pindex_t pd_pindex;
556 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
557 pd_pindex = pmap_pd_pindex(va);
558 spin_lock(&pmap->pm_spin);
559 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
560 spin_unlock(&pmap->pm_spin);
561 if (pv == NULL || pv->pv_m == NULL)
563 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
565 pd = pmap_pd(pmap, va);
566 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
568 return (pmap_pd_to_pt(*pd, va));
573 * Return pointer to PTE slot in the PT given a pointer to the PT
577 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
581 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
582 return (&pte[pmap_pte_index(va)]);
586 * Return pointer to PTE slot in the PT
590 pmap_pte(pmap_t pmap, vm_offset_t va)
594 pt = pmap_pt(pmap, va);
595 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
597 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
598 return ((pt_entry_t *)pt);
599 return (pmap_pt_to_pte(*pt, va));
603 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
604 * the PT layer. This will speed up core pmap operations considerably.
606 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
607 * must be in a known associated state (typically by being locked when
608 * the pmap spinlock isn't held). We allow the race for that case.
612 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
614 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
615 pv->pv_pmap->pm_pvhint = pv;
620 * KVM - return address of PT slot in PD
624 vtopt(vm_offset_t va)
626 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
627 NPML4EPGSHIFT)) - 1);
629 return (PDmap + ((va >> PDRSHIFT) & mask));
633 * KVM - return address of PTE slot in PT
637 vtopte(vm_offset_t va)
639 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
640 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
642 return (PTmap + ((va >> PAGE_SHIFT) & mask));
646 allocpages(vm_paddr_t *firstaddr, long n)
651 bzero((void *)ret, n * PAGE_SIZE);
652 *firstaddr += n * PAGE_SIZE;
658 create_pagetables(vm_paddr_t *firstaddr)
660 long i; /* must be 64 bits */
666 * We are running (mostly) V=P at this point
668 * Calculate NKPT - number of kernel page tables. We have to
669 * accomodoate prealloction of the vm_page_array, dump bitmap,
670 * MSGBUF_SIZE, and other stuff. Be generous.
672 * Maxmem is in pages.
674 * ndmpdp is the number of 1GB pages we wish to map.
676 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
677 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
679 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
682 * Starting at the beginning of kvm (not KERNBASE).
684 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
685 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
686 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
687 ndmpdp) + 511) / 512;
691 * Starting at KERNBASE - map 2G worth of page table pages.
692 * KERNBASE is offset -2G from the end of kvm.
694 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
699 KPTbase = allocpages(firstaddr, nkpt_base);
700 KPTphys = allocpages(firstaddr, nkpt_phys);
701 KPML4phys = allocpages(firstaddr, 1);
702 KPDPphys = allocpages(firstaddr, NKPML4E);
703 KPDphys = allocpages(firstaddr, NKPDPE);
706 * Calculate the page directory base for KERNBASE,
707 * that is where we start populating the page table pages.
708 * Basically this is the end - 2.
710 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
712 DMPDPphys = allocpages(firstaddr, NDMPML4E);
713 if ((amd_feature & AMDID_PAGE1GB) == 0)
714 DMPDphys = allocpages(firstaddr, ndmpdp);
715 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
718 * Fill in the underlying page table pages for the area around
719 * KERNBASE. This remaps low physical memory to KERNBASE.
721 * Read-only from zero to physfree
722 * XXX not fully used, underneath 2M pages
724 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
725 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
726 ((pt_entry_t *)KPTbase)[i] |=
727 pmap_bits_default[PG_RW_IDX] |
728 pmap_bits_default[PG_V_IDX] |
729 pmap_bits_default[PG_G_IDX];
733 * Now map the initial kernel page tables. One block of page
734 * tables is placed at the beginning of kernel virtual memory,
735 * and another block is placed at KERNBASE to map the kernel binary,
736 * data, bss, and initial pre-allocations.
738 for (i = 0; i < nkpt_base; i++) {
739 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
740 ((pd_entry_t *)KPDbase)[i] |=
741 pmap_bits_default[PG_RW_IDX] |
742 pmap_bits_default[PG_V_IDX];
744 for (i = 0; i < nkpt_phys; i++) {
745 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
746 ((pd_entry_t *)KPDphys)[i] |=
747 pmap_bits_default[PG_RW_IDX] |
748 pmap_bits_default[PG_V_IDX];
752 * Map from zero to end of allocations using 2M pages as an
753 * optimization. This will bypass some of the KPTBase pages
754 * above in the KERNBASE area.
756 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
757 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
758 ((pd_entry_t *)KPDbase)[i] |=
759 pmap_bits_default[PG_RW_IDX] |
760 pmap_bits_default[PG_V_IDX] |
761 pmap_bits_default[PG_PS_IDX] |
762 pmap_bits_default[PG_G_IDX];
766 * And connect up the PD to the PDP. The kernel pmap is expected
767 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
769 for (i = 0; i < NKPDPE; i++) {
770 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
771 KPDphys + (i << PAGE_SHIFT);
772 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
773 pmap_bits_default[PG_RW_IDX] |
774 pmap_bits_default[PG_V_IDX] |
775 pmap_bits_default[PG_U_IDX];
779 * Now set up the direct map space using either 2MB or 1GB pages
780 * Preset PG_M and PG_A because demotion expects it.
782 * When filling in entries in the PD pages make sure any excess
783 * entries are set to zero as we allocated enough PD pages
785 if ((amd_feature & AMDID_PAGE1GB) == 0) {
786 for (i = 0; i < NPDEPG * ndmpdp; i++) {
787 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
788 ((pd_entry_t *)DMPDphys)[i] |=
789 pmap_bits_default[PG_RW_IDX] |
790 pmap_bits_default[PG_V_IDX] |
791 pmap_bits_default[PG_PS_IDX] |
792 pmap_bits_default[PG_G_IDX] |
793 pmap_bits_default[PG_M_IDX] |
794 pmap_bits_default[PG_A_IDX];
798 * And the direct map space's PDP
800 for (i = 0; i < ndmpdp; i++) {
801 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
803 ((pdp_entry_t *)DMPDPphys)[i] |=
804 pmap_bits_default[PG_RW_IDX] |
805 pmap_bits_default[PG_V_IDX] |
806 pmap_bits_default[PG_U_IDX];
809 for (i = 0; i < ndmpdp; i++) {
810 ((pdp_entry_t *)DMPDPphys)[i] =
811 (vm_paddr_t)i << PDPSHIFT;
812 ((pdp_entry_t *)DMPDPphys)[i] |=
813 pmap_bits_default[PG_RW_IDX] |
814 pmap_bits_default[PG_V_IDX] |
815 pmap_bits_default[PG_PS_IDX] |
816 pmap_bits_default[PG_G_IDX] |
817 pmap_bits_default[PG_M_IDX] |
818 pmap_bits_default[PG_A_IDX];
822 /* And recursively map PML4 to itself in order to get PTmap */
823 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
824 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
825 pmap_bits_default[PG_RW_IDX] |
826 pmap_bits_default[PG_V_IDX] |
827 pmap_bits_default[PG_U_IDX];
830 * Connect the Direct Map slots up to the PML4
832 for (j = 0; j < NDMPML4E; ++j) {
833 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
834 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
835 pmap_bits_default[PG_RW_IDX] |
836 pmap_bits_default[PG_V_IDX] |
837 pmap_bits_default[PG_U_IDX];
841 * Connect the KVA slot up to the PML4
843 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
844 ((pdp_entry_t *)KPML4phys)[KPML4I] |=
845 pmap_bits_default[PG_RW_IDX] |
846 pmap_bits_default[PG_V_IDX] |
847 pmap_bits_default[PG_U_IDX];
851 * Bootstrap the system enough to run with virtual memory.
853 * On the i386 this is called after mapping has already been enabled
854 * and just syncs the pmap module with what has already been done.
855 * [We can't call it easily with mapping off since the kernel is not
856 * mapped with PA == VA, hence we would have to relocate every address
857 * from the linked base (virtual) address "KERNBASE" to the actual
858 * (physical) address starting relative to 0]
861 pmap_bootstrap(vm_paddr_t *firstaddr)
866 KvaStart = VM_MIN_KERNEL_ADDRESS;
867 KvaEnd = VM_MAX_KERNEL_ADDRESS;
868 KvaSize = KvaEnd - KvaStart;
870 avail_start = *firstaddr;
873 * Create an initial set of page tables to run the kernel in.
875 create_pagetables(firstaddr);
877 virtual2_start = KvaStart;
878 virtual2_end = PTOV_OFFSET;
880 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
881 virtual_start = pmap_kmem_choose(virtual_start);
883 virtual_end = VM_MAX_KERNEL_ADDRESS;
885 /* XXX do %cr0 as well */
886 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
890 * Initialize protection array.
892 i386_protection_init();
895 * The kernel's pmap is statically allocated so we don't have to use
896 * pmap_create, which is unlikely to work correctly at this part of
897 * the boot sequence (XXX and which no longer exists).
899 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
900 kernel_pmap.pm_count = 1;
901 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
902 RB_INIT(&kernel_pmap.pm_pvroot);
903 spin_init(&kernel_pmap.pm_spin);
904 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
907 * Reserve some special page table entries/VA space for temporary
910 #define SYSMAP(c, p, v, n) \
911 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
917 * CMAP1/CMAP2 are used for zeroing and copying pages.
919 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
924 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
927 * ptvmmap is used for reading arbitrary physical pages via
930 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
933 * msgbufp is used to map the system message buffer.
934 * XXX msgbufmap is not used.
936 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
937 atop(round_page(MSGBUF_SIZE)))
944 * PG_G is terribly broken on SMP because we IPI invltlb's in some
945 * cases rather then invl1pg. Actually, I don't even know why it
946 * works under UP because self-referential page table mappings
951 * Initialize the 4MB page size flag
955 * The 4MB page version of the initial
956 * kernel page mapping.
960 #if !defined(DISABLE_PSE)
961 if (cpu_feature & CPUID_PSE) {
964 * Note that we have enabled PSE mode
966 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
967 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
968 ptditmp &= ~(NBPDR - 1);
969 ptditmp |= pmap_bits_default[PG_V_IDX] |
970 pmap_bits_default[PG_RW_IDX] |
971 pmap_bits_default[PG_PS_IDX] |
972 pmap_bits_default[PG_U_IDX];
979 /* Initialize the PAT MSR */
982 pmap_pinit_defaults(&kernel_pmap);
995 * Default values mapping PATi,PCD,PWT bits at system reset.
996 * The default values effectively ignore the PATi bit by
997 * repeating the encodings for 0-3 in 4-7, and map the PCD
998 * and PWT bit combinations to the expected PAT types.
1000 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
1001 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
1002 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
1003 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
1004 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
1005 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
1006 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
1007 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
1008 pat_pte_index[PAT_WRITE_BACK] = 0;
1009 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
1010 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
1011 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
1012 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1013 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1015 if (cpu_feature & CPUID_PAT) {
1017 * If we support the PAT then set-up entries for
1018 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1021 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1022 PAT_VALUE(4, PAT_WRITE_PROTECTED);
1023 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1024 PAT_VALUE(5, PAT_WRITE_COMBINING);
1025 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1026 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1029 * Then enable the PAT
1034 load_cr4(cr4 & ~CR4_PGE);
1036 /* Disable caches (CD = 1, NW = 0). */
1038 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1040 /* Flushes caches and TLBs. */
1044 /* Update PAT and index table. */
1045 wrmsr(MSR_PAT, pat_msr);
1047 /* Flush caches and TLBs again. */
1051 /* Restore caches and PGE. */
1059 * Set 4mb pdir for mp startup
1064 if (cpu_feature & CPUID_PSE) {
1065 load_cr4(rcr4() | CR4_PSE);
1066 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
1073 * Initialize the pmap module.
1074 * Called by vm_init, to initialize any structures that the pmap
1075 * system needs to map virtual memory.
1076 * pmap_init has been enhanced to support in a fairly consistant
1077 * way, discontiguous physical memory.
1086 * Allocate memory for random pmap data structures. Includes the
1090 for (i = 0; i < vm_page_array_size; i++) {
1093 m = &vm_page_array[i];
1094 TAILQ_INIT(&m->md.pv_list);
1098 * init the pv free list
1100 initial_pvs = vm_page_array_size;
1101 if (initial_pvs < MINPV)
1102 initial_pvs = MINPV;
1103 pvzone = &pvzone_store;
1104 pvinit = (void *)kmem_alloc(&kernel_map,
1105 initial_pvs * sizeof (struct pv_entry));
1106 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1107 pvinit, initial_pvs);
1110 * Now it is safe to enable pv_table recording.
1112 pmap_initialized = TRUE;
1116 * Initialize the address space (zone) for the pv_entries. Set a
1117 * high water mark so that the system can recover from excessive
1118 * numbers of pv entries.
1123 int shpgperproc = PMAP_SHPGPERPROC;
1126 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1127 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1128 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1129 pv_entry_high_water = 9 * (pv_entry_max / 10);
1132 * Subtract out pages already installed in the zone (hack)
1134 entry_max = pv_entry_max - vm_page_array_size;
1138 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
1142 * Typically used to initialize a fictitious page by vm/device_pager.c
1145 pmap_page_init(struct vm_page *m)
1148 TAILQ_INIT(&m->md.pv_list);
1151 /***************************************************
1152 * Low level helper routines.....
1153 ***************************************************/
1156 * this routine defines the region(s) of memory that should
1157 * not be tested for the modified bit.
1161 pmap_track_modified(vm_pindex_t pindex)
1163 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1164 if ((va < clean_sva) || (va >= clean_eva))
1171 * Extract the physical page address associated with the map/VA pair.
1172 * The page must be wired for this to work reliably.
1174 * XXX for the moment we're using pv_find() instead of pv_get(), as
1175 * callers might be expecting non-blocking operation.
1178 pmap_extract(pmap_t pmap, vm_offset_t va)
1185 if (va >= VM_MAX_USER_ADDRESS) {
1187 * Kernel page directories might be direct-mapped and
1188 * there is typically no PV tracking of pte's
1192 pt = pmap_pt(pmap, va);
1193 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1194 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1195 rtval = *pt & PG_PS_FRAME;
1196 rtval |= va & PDRMASK;
1198 ptep = pmap_pt_to_pte(*pt, va);
1199 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1200 rtval = *ptep & PG_FRAME;
1201 rtval |= va & PAGE_MASK;
1207 * User pages currently do not direct-map the page directory
1208 * and some pages might not used managed PVs. But all PT's
1211 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1213 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1214 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1215 rtval = *ptep & PG_FRAME;
1216 rtval |= va & PAGE_MASK;
1225 * Similar to extract but checks protections, SMP-friendly short-cut for
1226 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1227 * fall-through to the real fault code.
1229 * The returned page, if not NULL, is held (and not busied).
1232 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1234 if (pmap && va < VM_MAX_USER_ADDRESS) {
1242 req = pmap->pmap_bits[PG_V_IDX] |
1243 pmap->pmap_bits[PG_U_IDX];
1244 if (prot & VM_PROT_WRITE)
1245 req |= pmap->pmap_bits[PG_RW_IDX];
1247 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1250 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1251 if ((*ptep & req) != req) {
1255 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1256 if (pte_pv && error == 0) {
1259 if (prot & VM_PROT_WRITE)
1262 } else if (pte_pv) {
1276 * Extract the physical page address associated kernel virtual address.
1279 pmap_kextract(vm_offset_t va)
1281 pd_entry_t pt; /* pt entry in pd */
1284 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1285 pa = DMAP_TO_PHYS(va);
1288 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1289 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1292 * Beware of a concurrent promotion that changes the
1293 * PDE at this point! For example, vtopte() must not
1294 * be used to access the PTE because it would use the
1295 * new PDE. It is, however, safe to use the old PDE
1296 * because the page table page is preserved by the
1299 pa = *pmap_pt_to_pte(pt, va);
1300 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1306 /***************************************************
1307 * Low level mapping routines.....
1308 ***************************************************/
1311 * Routine: pmap_kenter
1313 * Add a wired page to the KVA
1314 * NOTE! note that in order for the mapping to take effect -- you
1315 * should do an invltlb after doing the pmap_kenter().
1318 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1322 pmap_inval_info info;
1324 pmap_inval_init(&info); /* XXX remove */
1326 kernel_pmap.pmap_bits[PG_RW_IDX] |
1327 kernel_pmap.pmap_bits[PG_V_IDX];
1330 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1332 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1333 pmap_inval_done(&info); /* XXX remove */
1337 * Routine: pmap_kenter_quick
1339 * Similar to pmap_kenter(), except we only invalidate the
1340 * mapping on the current CPU.
1343 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1349 kernel_pmap.pmap_bits[PG_RW_IDX] |
1350 kernel_pmap.pmap_bits[PG_V_IDX];
1354 cpu_invlpg((void *)va);
1358 pmap_kenter_sync(vm_offset_t va)
1360 pmap_inval_info info;
1362 pmap_inval_init(&info);
1363 pmap_inval_interlock(&info, &kernel_pmap, va);
1364 pmap_inval_deinterlock(&info, &kernel_pmap);
1365 pmap_inval_done(&info);
1369 pmap_kenter_sync_quick(vm_offset_t va)
1371 cpu_invlpg((void *)va);
1375 * remove a page from the kernel pagetables
1378 pmap_kremove(vm_offset_t va)
1381 pmap_inval_info info;
1383 pmap_inval_init(&info);
1385 pmap_inval_interlock(&info, &kernel_pmap, va);
1386 (void)pte_load_clear(pte);
1387 pmap_inval_deinterlock(&info, &kernel_pmap);
1388 pmap_inval_done(&info);
1392 pmap_kremove_quick(vm_offset_t va)
1396 (void)pte_load_clear(pte);
1397 cpu_invlpg((void *)va);
1401 * XXX these need to be recoded. They are not used in any critical path.
1404 pmap_kmodify_rw(vm_offset_t va)
1406 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1407 cpu_invlpg((void *)va);
1412 pmap_kmodify_nc(vm_offset_t va)
1414 atomic_set_long(vtopte(va), PG_N);
1415 cpu_invlpg((void *)va);
1420 * Used to map a range of physical addresses into kernel virtual
1421 * address space during the low level boot, typically to map the
1422 * dump bitmap, message buffer, and vm_page_array.
1424 * These mappings are typically made at some pointer after the end of the
1427 * We could return PHYS_TO_DMAP(start) here and not allocate any
1428 * via (*virtp), but then kmem from userland and kernel dumps won't
1429 * have access to the related pointers.
1432 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1435 vm_offset_t va_start;
1437 /*return PHYS_TO_DMAP(start);*/
1442 while (start < end) {
1443 pmap_kenter_quick(va, start);
1451 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1454 * Remove the specified set of pages from the data and instruction caches.
1456 * In contrast to pmap_invalidate_cache_range(), this function does not
1457 * rely on the CPU's self-snoop feature, because it is intended for use
1458 * when moving pages into a different cache domain.
1461 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1463 vm_offset_t daddr, eva;
1466 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1467 (cpu_feature & CPUID_CLFSH) == 0)
1471 for (i = 0; i < count; i++) {
1472 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1473 eva = daddr + PAGE_SIZE;
1474 for (; daddr < eva; daddr += cpu_clflush_line_size)
1482 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1484 KASSERT((sva & PAGE_MASK) == 0,
1485 ("pmap_invalidate_cache_range: sva not page-aligned"));
1486 KASSERT((eva & PAGE_MASK) == 0,
1487 ("pmap_invalidate_cache_range: eva not page-aligned"));
1489 if (cpu_feature & CPUID_SS) {
1490 ; /* If "Self Snoop" is supported, do nothing. */
1492 /* Globally invalidate caches */
1493 cpu_wbinvd_on_all_cpus();
1497 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1499 smp_invlpg_range(pmap->pm_active, sva, eva);
1503 * Add a list of wired pages to the kva
1504 * this routine is only used for temporary
1505 * kernel mappings that do not need to have
1506 * page modification or references recorded.
1507 * Note that old mappings are simply written
1508 * over. The page *must* be wired.
1511 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1515 end_va = va + count * PAGE_SIZE;
1517 while (va < end_va) {
1521 *pte = VM_PAGE_TO_PHYS(*m) |
1522 kernel_pmap.pmap_bits[PG_RW_IDX] |
1523 kernel_pmap.pmap_bits[PG_V_IDX] |
1524 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1526 cpu_invlpg((void *)va);
1534 * This routine jerks page mappings from the
1535 * kernel -- it is meant only for temporary mappings.
1537 * MPSAFE, INTERRUPT SAFE (cluster callback)
1540 pmap_qremove(vm_offset_t va, int count)
1544 end_va = va + count * PAGE_SIZE;
1546 while (va < end_va) {
1550 (void)pte_load_clear(pte);
1551 cpu_invlpg((void *)va);
1558 * Create a new thread and optionally associate it with a (new) process.
1559 * NOTE! the new thread's cpu may not equal the current cpu.
1562 pmap_init_thread(thread_t td)
1564 /* enforce pcb placement & alignment */
1565 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1566 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1567 td->td_savefpu = &td->td_pcb->pcb_save;
1568 td->td_sp = (char *)td->td_pcb; /* no -16 */
1572 * This routine directly affects the fork perf for a process.
1575 pmap_init_proc(struct proc *p)
1580 pmap_pinit_defaults(struct pmap *pmap)
1582 bcopy(pmap_bits_default, pmap->pmap_bits,
1583 sizeof(pmap_bits_default));
1584 bcopy(protection_codes, pmap->protection_codes,
1585 sizeof(protection_codes));
1586 bcopy(pat_pte_index, pmap->pmap_cache_bits,
1587 sizeof(pat_pte_index));
1588 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1589 pmap->copyinstr = std_copyinstr;
1590 pmap->copyin = std_copyin;
1591 pmap->copyout = std_copyout;
1592 pmap->fubyte = std_fubyte;
1593 pmap->subyte = std_subyte;
1594 pmap->fuword = std_fuword;
1595 pmap->suword = std_suword;
1596 pmap->suword32 = std_suword32;
1599 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1600 * it, and IdlePTD, represents the template used to update all other pmaps.
1602 * On architectures where the kernel pmap is not integrated into the user
1603 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1604 * kernel_pmap should be used to directly access the kernel_pmap.
1607 pmap_pinit0(struct pmap *pmap)
1609 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1611 pmap->pm_active = 0;
1612 pmap->pm_pvhint = NULL;
1613 RB_INIT(&pmap->pm_pvroot);
1614 spin_init(&pmap->pm_spin);
1615 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1616 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1617 pmap_pinit_defaults(pmap);
1621 * Initialize a preallocated and zeroed pmap structure,
1622 * such as one in a vmspace structure.
1625 pmap_pinit_simple(struct pmap *pmap)
1628 * Misc initialization
1631 pmap->pm_active = 0;
1632 pmap->pm_pvhint = NULL;
1633 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1635 pmap_pinit_defaults(pmap);
1638 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1641 if (pmap->pm_pmlpv == NULL) {
1642 RB_INIT(&pmap->pm_pvroot);
1643 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1644 spin_init(&pmap->pm_spin);
1645 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1650 pmap_pinit(struct pmap *pmap)
1655 if (pmap->pm_pmlpv) {
1656 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1661 pmap_pinit_simple(pmap);
1662 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1665 * No need to allocate page table space yet but we do need a valid
1666 * page directory table.
1668 if (pmap->pm_pml4 == NULL) {
1670 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1674 * Allocate the page directory page, which wires it even though
1675 * it isn't being entered into some higher level page table (it
1676 * being the highest level). If one is already cached we don't
1677 * have to do anything.
1679 if ((pv = pmap->pm_pmlpv) == NULL) {
1680 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1681 pmap->pm_pmlpv = pv;
1682 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1683 VM_PAGE_TO_PHYS(pv->pv_m));
1687 * Install DMAP and KMAP.
1689 for (j = 0; j < NDMPML4E; ++j) {
1690 pmap->pm_pml4[DMPML4I + j] =
1691 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1692 pmap->pmap_bits[PG_RW_IDX] |
1693 pmap->pmap_bits[PG_V_IDX] |
1694 pmap->pmap_bits[PG_U_IDX];
1696 pmap->pm_pml4[KPML4I] = KPDPphys |
1697 pmap->pmap_bits[PG_RW_IDX] |
1698 pmap->pmap_bits[PG_V_IDX] |
1699 pmap->pmap_bits[PG_U_IDX];
1702 * install self-referential address mapping entry
1704 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1705 pmap->pmap_bits[PG_V_IDX] |
1706 pmap->pmap_bits[PG_RW_IDX] |
1707 pmap->pmap_bits[PG_A_IDX] |
1708 pmap->pmap_bits[PG_M_IDX];
1710 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1711 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1713 KKASSERT(pmap->pm_pml4[255] == 0);
1714 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1715 KKASSERT(pv->pv_entry.rbe_left == NULL);
1716 KKASSERT(pv->pv_entry.rbe_right == NULL);
1720 * Clean up a pmap structure so it can be physically freed. This routine
1721 * is called by the vmspace dtor function. A great deal of pmap data is
1722 * left passively mapped to improve vmspace management so we have a bit
1723 * of cleanup work to do here.
1726 pmap_puninit(pmap_t pmap)
1731 KKASSERT(pmap->pm_active == 0);
1732 if ((pv = pmap->pm_pmlpv) != NULL) {
1733 if (pv_hold_try(pv) == 0)
1735 KKASSERT(pv == pmap->pm_pmlpv);
1736 p = pmap_remove_pv_page(pv);
1738 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1739 vm_page_busy_wait(p, FALSE, "pgpun");
1740 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1741 vm_page_unwire(p, 0);
1742 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1745 * XXX eventually clean out PML4 static entries and
1746 * use vm_page_free_zero()
1749 pmap->pm_pmlpv = NULL;
1751 if (pmap->pm_pml4) {
1752 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1753 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1754 pmap->pm_pml4 = NULL;
1756 KKASSERT(pmap->pm_stats.resident_count == 0);
1757 KKASSERT(pmap->pm_stats.wired_count == 0);
1761 * Wire in kernel global address entries. To avoid a race condition
1762 * between pmap initialization and pmap_growkernel, this procedure
1763 * adds the pmap to the master list (which growkernel scans to update),
1764 * then copies the template.
1767 pmap_pinit2(struct pmap *pmap)
1769 spin_lock(&pmap_spin);
1770 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1771 spin_unlock(&pmap_spin);
1775 * This routine is called when various levels in the page table need to
1776 * be populated. This routine cannot fail.
1778 * This function returns two locked pv_entry's, one representing the
1779 * requested pv and one representing the requested pv's parent pv. If
1780 * the pv did not previously exist it will be mapped into its parent
1781 * and wired, otherwise no additional wire count will be added.
1785 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1790 vm_pindex_t pt_pindex;
1796 * If the pv already exists and we aren't being asked for the
1797 * parent page table page we can just return it. A locked+held pv
1798 * is returned. The pv will also have a second hold related to the
1799 * pmap association that we don't have to worry about.
1802 pv = pv_alloc(pmap, ptepindex, &isnew);
1803 if (isnew == 0 && pvpp == NULL)
1807 * Special case terminal PVs. These are not page table pages so
1808 * no vm_page is allocated (the caller supplied the vm_page). If
1809 * pvpp is non-NULL we are being asked to also removed the pt_pv
1812 * Note that pt_pv's are only returned for user VAs. We assert that
1813 * a pt_pv is not being requested for kernel VAs.
1815 if (ptepindex < pmap_pt_pindex(0)) {
1816 if (ptepindex >= NUPTE_USER)
1817 KKASSERT(pvpp == NULL);
1819 KKASSERT(pvpp != NULL);
1821 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1822 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1824 vm_page_wire_quick(pvp->pv_m);
1833 * Non-terminal PVs allocate a VM page to represent the page table,
1834 * so we have to resolve pvp and calculate ptepindex for the pvp
1835 * and then for the page table entry index in the pvp for
1838 if (ptepindex < pmap_pd_pindex(0)) {
1840 * pv is PT, pvp is PD
1842 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1843 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1844 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1851 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1852 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1854 } else if (ptepindex < pmap_pdp_pindex(0)) {
1856 * pv is PD, pvp is PDP
1858 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1861 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1862 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1864 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1865 KKASSERT(pvpp == NULL);
1868 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1876 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1877 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1878 } else if (ptepindex < pmap_pml4_pindex()) {
1880 * pv is PDP, pvp is the root pml4 table
1882 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1889 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1890 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1893 * pv represents the top-level PML4, there is no parent.
1901 * This code is only reached if isnew is TRUE and this is not a
1902 * terminal PV. We need to allocate a vm_page for the page table
1903 * at this level and enter it into the parent page table.
1905 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1908 m = vm_page_alloc(NULL, pv->pv_pindex,
1909 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1910 VM_ALLOC_INTERRUPT);
1915 vm_page_spin_lock(m);
1916 pmap_page_stats_adding(m);
1917 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1919 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1920 vm_page_spin_unlock(m);
1921 vm_page_unmanage(m); /* m must be spinunlocked */
1923 if ((m->flags & PG_ZERO) == 0) {
1924 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1928 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1931 m->valid = VM_PAGE_BITS_ALL;
1932 vm_page_flag_clear(m, PG_ZERO);
1933 vm_page_wire(m); /* wire for mapping in parent */
1936 * Wire the page into pvp, bump the wire-count for pvp's page table
1937 * page. Bump the resident_count for the pmap. There is no pvp
1938 * for the top level, address the pm_pml4[] array directly.
1940 * If the caller wants the parent we return it, otherwise
1941 * we just put it away.
1943 * No interlock is needed for pte 0 -> non-zero.
1945 * In the situation where *ptep is valid we might have an unmanaged
1946 * page table page shared from another page table which we need to
1947 * unshare before installing our private page table page.
1950 ptep = pv_pte_lookup(pvp, ptepindex);
1951 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1953 pmap_inval_info info;
1956 panic("pmap_allocpte: unexpected pte %p/%d",
1957 pvp, (int)ptepindex);
1959 pmap_inval_init(&info);
1960 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1961 pte = pte_load_clear(ptep);
1962 pmap_inval_deinterlock(&info, pmap);
1963 pmap_inval_done(&info);
1964 if (vm_page_unwire_quick(
1965 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
1966 panic("pmap_allocpte: shared pgtable "
1967 "pg bad wirecount");
1969 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1971 vm_page_wire_quick(pvp->pv_m);
1973 *ptep = VM_PAGE_TO_PHYS(m) |
1974 (pmap->pmap_bits[PG_U_IDX] |
1975 pmap->pmap_bits[PG_RW_IDX] |
1976 pmap->pmap_bits[PG_V_IDX] |
1977 pmap->pmap_bits[PG_A_IDX] |
1978 pmap->pmap_bits[PG_M_IDX]);
1990 * This version of pmap_allocpte() checks for possible segment optimizations
1991 * that would allow page-table sharing. It can be called for terminal
1992 * page or page table page ptepindex's.
1994 * The function is called with page table page ptepindex's for fictitious
1995 * and unmanaged terminal pages. That is, we don't want to allocate a
1996 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
1999 * This function can return a pv and *pvpp associated with the passed in pmap
2000 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2001 * an unmanaged page table page will be entered into the pass in pmap.
2005 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2006 vm_map_entry_t entry, vm_offset_t va)
2008 struct pmap_inval_info info;
2013 pv_entry_t pte_pv; /* in original or shared pmap */
2014 pv_entry_t pt_pv; /* in original or shared pmap */
2015 pv_entry_t proc_pd_pv; /* in original pmap */
2016 pv_entry_t proc_pt_pv; /* in original pmap */
2017 pv_entry_t xpv; /* PT in shared pmap */
2018 pd_entry_t *pt; /* PT entry in PD of original pmap */
2019 pd_entry_t opte; /* contents of *pt */
2020 pd_entry_t npte; /* contents of *pt */
2025 * Basic tests, require a non-NULL vm_map_entry, require proper
2026 * alignment and type for the vm_map_entry, require that the
2027 * underlying object already be allocated.
2029 * We currently allow any type of object to use this optimization.
2030 * The object itself does NOT have to be sized to a multiple of the
2031 * segment size, but the memory mapping does.
2033 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2034 * won't work as expected.
2036 if (entry == NULL ||
2037 pmap_mmu_optimize == 0 || /* not enabled */
2038 ptepindex >= pmap_pd_pindex(0) || /* not terminal */
2039 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2040 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2041 entry->object.vm_object == NULL || /* needs VM object */
2042 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2043 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2044 (entry->offset & SEG_MASK) || /* must be aligned */
2045 (entry->start & SEG_MASK)) {
2046 return(pmap_allocpte(pmap, ptepindex, pvpp));
2050 * Make sure the full segment can be represented.
2052 b = va & ~(vm_offset_t)SEG_MASK;
2053 if (b < entry->start && b + SEG_SIZE > entry->end)
2054 return(pmap_allocpte(pmap, ptepindex, pvpp));
2057 * If the full segment can be represented dive the VM object's
2058 * shared pmap, allocating as required.
2060 object = entry->object.vm_object;
2062 if (entry->protection & VM_PROT_WRITE)
2063 obpmapp = &object->md.pmap_rw;
2065 obpmapp = &object->md.pmap_ro;
2068 * We allocate what appears to be a normal pmap but because portions
2069 * of this pmap are shared with other unrelated pmaps we have to
2070 * set pm_active to point to all cpus.
2072 * XXX Currently using pmap_spin to interlock the update, can't use
2073 * vm_object_hold/drop because the token might already be held
2074 * shared OR exclusive and we don't know.
2076 while ((obpmap = *obpmapp) == NULL) {
2077 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2078 pmap_pinit_simple(obpmap);
2079 pmap_pinit2(obpmap);
2080 spin_lock(&pmap_spin);
2081 if (*obpmapp != NULL) {
2085 spin_unlock(&pmap_spin);
2086 pmap_release(obpmap);
2087 pmap_puninit(obpmap);
2088 kfree(obpmap, M_OBJPMAP);
2090 obpmap->pm_active = smp_active_mask;
2092 spin_unlock(&pmap_spin);
2097 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2098 * pte/pt using the shared pmap from the object but also adjust
2099 * the process pmap's page table page as a side effect.
2103 * Resolve the terminal PTE and PT in the shared pmap. This is what
2104 * we will return. This is true if ptepindex represents a terminal
2105 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2109 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2110 if (ptepindex >= pmap_pt_pindex(0))
2116 * Resolve the PD in the process pmap so we can properly share the
2117 * page table page. Lock order is bottom-up (leaf first)!
2119 * NOTE: proc_pt_pv can be NULL.
2121 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2122 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2125 * xpv is the page table page pv from the shared object
2126 * (for convenience).
2128 * Calculate the pte value for the PT to load into the process PD.
2129 * If we have to change it we must properly dispose of the previous
2132 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2133 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2134 (pmap->pmap_bits[PG_U_IDX] |
2135 pmap->pmap_bits[PG_RW_IDX] |
2136 pmap->pmap_bits[PG_V_IDX] |
2137 pmap->pmap_bits[PG_A_IDX] |
2138 pmap->pmap_bits[PG_M_IDX]);
2141 * Dispose of previous page table page if it was local to the
2142 * process pmap. If the old pt is not empty we cannot dispose of it
2143 * until we clean it out. This case should not arise very often so
2144 * it is not optimized.
2147 if (proc_pt_pv->pv_m->wire_count != 1) {
2153 va & ~(vm_offset_t)SEG_MASK,
2154 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2157 pmap_release_pv(proc_pt_pv, proc_pd_pv);
2160 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2164 * Handle remaining cases.
2168 vm_page_wire_quick(xpv->pv_m);
2169 vm_page_wire_quick(proc_pd_pv->pv_m);
2170 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2171 } else if (*pt != npte) {
2172 pmap_inval_init(&info);
2173 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
2175 opte = pte_load_clear(pt);
2176 KKASSERT(opte && opte != npte);
2179 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
2182 * Clean up opte, bump the wire_count for the process
2183 * PD page representing the new entry if it was
2186 * If the entry was not previously empty and we have
2187 * a PT in the proc pmap then opte must match that
2188 * pt. The proc pt must be retired (this is done
2189 * later on in this procedure).
2191 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2194 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2195 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2196 if (vm_page_unwire_quick(m)) {
2197 panic("pmap_allocpte_seg: "
2198 "bad wire count %p",
2202 pmap_inval_deinterlock(&info, pmap);
2203 pmap_inval_done(&info);
2207 * The existing process page table was replaced and must be destroyed
2221 * Release any resources held by the given physical map.
2223 * Called when a pmap initialized by pmap_pinit is being released. Should
2224 * only be called if the map contains no valid mappings.
2226 * Caller must hold pmap->pm_token
2228 struct pmap_release_info {
2233 static int pmap_release_callback(pv_entry_t pv, void *data);
2236 pmap_release(struct pmap *pmap)
2238 struct pmap_release_info info;
2240 KASSERT(pmap->pm_active == 0,
2241 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
2243 spin_lock(&pmap_spin);
2244 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2245 spin_unlock(&pmap_spin);
2248 * Pull pv's off the RB tree in order from low to high and release
2254 spin_lock(&pmap->pm_spin);
2255 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2256 pmap_release_callback, &info);
2257 spin_unlock(&pmap->pm_spin);
2258 } while (info.retry);
2262 * One resident page (the pml4 page) should remain.
2263 * No wired pages should remain.
2265 KKASSERT(pmap->pm_stats.resident_count ==
2266 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2268 KKASSERT(pmap->pm_stats.wired_count == 0);
2272 pmap_release_callback(pv_entry_t pv, void *data)
2274 struct pmap_release_info *info = data;
2275 pmap_t pmap = info->pmap;
2278 if (pv_hold_try(pv)) {
2279 spin_unlock(&pmap->pm_spin);
2281 spin_unlock(&pmap->pm_spin);
2284 if (pv->pv_pmap != pmap) {
2286 spin_lock(&pmap->pm_spin);
2290 r = pmap_release_pv(pv, NULL);
2291 spin_lock(&pmap->pm_spin);
2296 * Called with held (i.e. also locked) pv. This function will dispose of
2297 * the lock along with the pv.
2299 * If the caller already holds the locked parent page table for pv it
2300 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2301 * pass NULL for pvp.
2304 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp)
2309 * The pmap is currently not spinlocked, pv is held+locked.
2310 * Remove the pv's page from its parent's page table. The
2311 * parent's page table page's wire_count will be decremented.
2313 pmap_remove_pv_pte(pv, pvp, NULL);
2316 * Terminal pvs are unhooked from their vm_pages. Because
2317 * terminal pages aren't page table pages they aren't wired
2318 * by us, so we have to be sure not to unwire them either.
2320 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2321 pmap_remove_pv_page(pv);
2326 * We leave the top-level page table page cached, wired, and
2327 * mapped in the pmap until the dtor function (pmap_puninit())
2330 * Since we are leaving the top-level pv intact we need
2331 * to break out of what would otherwise be an infinite loop.
2333 if (pv->pv_pindex == pmap_pml4_pindex()) {
2339 * For page table pages (other than the top-level page),
2340 * remove and free the vm_page. The representitive mapping
2341 * removed above by pmap_remove_pv_pte() did not undo the
2342 * last wire_count so we have to do that as well.
2344 p = pmap_remove_pv_page(pv);
2345 vm_page_busy_wait(p, FALSE, "pmaprl");
2346 if (p->wire_count != 1) {
2347 kprintf("p->wire_count was %016lx %d\n",
2348 pv->pv_pindex, p->wire_count);
2350 KKASSERT(p->wire_count == 1);
2351 KKASSERT(p->flags & PG_UNMANAGED);
2353 vm_page_unwire(p, 0);
2354 KKASSERT(p->wire_count == 0);
2357 * Theoretically this page, if not the pml4 page, should contain
2358 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2368 * This function will remove the pte associated with a pv from its parent.
2369 * Terminal pv's are supported. The removal will be interlocked if info
2370 * is non-NULL. The caller must dispose of pv instead of just unlocking
2373 * The wire count will be dropped on the parent page table. The wire
2374 * count on the page being removed (pv->pv_m) from the parent page table
2375 * is NOT touched. Note that terminal pages will not have any additional
2376 * wire counts while page table pages will have at least one representing
2377 * the mapping, plus others representing sub-mappings.
2379 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2380 * pages and user page table and terminal pages.
2382 * The pv must be locked.
2384 * XXX must lock parent pv's if they exist to remove pte XXX
2388 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2390 vm_pindex_t ptepindex = pv->pv_pindex;
2391 pmap_t pmap = pv->pv_pmap;
2397 if (ptepindex == pmap_pml4_pindex()) {
2399 * We are the top level pml4 table, there is no parent.
2401 p = pmap->pm_pmlpv->pv_m;
2402 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2404 * Remove a PDP page from the pml4e. This can only occur
2405 * with user page tables. We do not have to lock the
2406 * pml4 PV so just ignore pvp.
2408 vm_pindex_t pml4_pindex;
2409 vm_pindex_t pdp_index;
2412 pdp_index = ptepindex - pmap_pdp_pindex(0);
2414 pml4_pindex = pmap_pml4_pindex();
2415 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2419 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2420 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2421 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2423 KKASSERT(info == NULL);
2424 } else if (ptepindex >= pmap_pd_pindex(0)) {
2426 * Remove a PD page from the pdp
2428 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2429 * of a simple pmap because it stops at
2432 vm_pindex_t pdp_pindex;
2433 vm_pindex_t pd_index;
2436 pd_index = ptepindex - pmap_pd_pindex(0);
2439 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2440 (pd_index >> NPML4EPGSHIFT);
2441 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2446 pd = pv_pte_lookup(pvp, pd_index &
2447 ((1ul << NPDPEPGSHIFT) - 1));
2448 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2449 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2452 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2453 p = pv->pv_m; /* degenerate test later */
2455 KKASSERT(info == NULL);
2456 } else if (ptepindex >= pmap_pt_pindex(0)) {
2458 * Remove a PT page from the pd
2460 vm_pindex_t pd_pindex;
2461 vm_pindex_t pt_index;
2464 pt_index = ptepindex - pmap_pt_pindex(0);
2467 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2468 (pt_index >> NPDPEPGSHIFT);
2469 pvp = pv_get(pv->pv_pmap, pd_pindex);
2473 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2474 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2475 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2477 KKASSERT(info == NULL);
2480 * Remove a PTE from the PT page
2482 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2483 * pv is a pte_pv so we can safely lock pt_pv.
2485 * NOTE: FICTITIOUS pages may have multiple physical mappings
2486 * so PHYS_TO_VM_PAGE() will not necessarily work for
2489 vm_pindex_t pt_pindex;
2494 pt_pindex = ptepindex >> NPTEPGSHIFT;
2495 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2497 if (ptepindex >= NUPTE_USER) {
2498 ptep = vtopte(ptepindex << PAGE_SHIFT);
2499 KKASSERT(pvp == NULL);
2502 pt_pindex = NUPTE_TOTAL +
2503 (ptepindex >> NPDPEPGSHIFT);
2504 pvp = pv_get(pv->pv_pmap, pt_pindex);
2508 ptep = pv_pte_lookup(pvp, ptepindex &
2509 ((1ul << NPDPEPGSHIFT) - 1));
2513 pmap_inval_interlock(info, pmap, va);
2514 pte = pte_load_clear(ptep);
2516 pmap_inval_deinterlock(info, pmap);
2518 cpu_invlpg((void *)va);
2521 * Now update the vm_page_t
2523 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) !=
2524 (pmap->pmap_bits[PG_MANAGED_IDX]|pmap->pmap_bits[PG_V_IDX])) {
2525 kprintf("remove_pte badpte %016lx %016lx %d\n",
2527 pv->pv_pindex < pmap_pt_pindex(0));
2529 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2530 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2531 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2534 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2537 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2538 if (pmap_track_modified(ptepindex))
2541 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2542 vm_page_flag_set(p, PG_REFERENCED);
2544 if (pte & pmap->pmap_bits[PG_W_IDX])
2545 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2546 if (pte & pmap->pmap_bits[PG_G_IDX])
2547 cpu_invlpg((void *)va);
2551 * Unwire the parent page table page. The wire_count cannot go below
2552 * 1 here because the parent page table page is itself still mapped.
2554 * XXX remove the assertions later.
2556 KKASSERT(pv->pv_m == p);
2557 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2558 panic("pmap_remove_pv_pte: Insufficient wire_count");
2565 * Remove the vm_page association to a pv. The pv must be locked.
2569 pmap_remove_pv_page(pv_entry_t pv)
2575 vm_page_spin_lock(m);
2577 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2578 pmap_page_stats_deleting(m);
2581 atomic_add_int(&m->object->agg_pv_list_count, -1);
2583 if (TAILQ_EMPTY(&m->md.pv_list))
2584 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2585 vm_page_spin_unlock(m);
2590 * Grow the number of kernel page table entries, if needed.
2592 * This routine is always called to validate any address space
2593 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2594 * space below KERNBASE.
2597 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2600 vm_offset_t ptppaddr;
2602 pd_entry_t *pt, newpt;
2604 int update_kernel_vm_end;
2607 * bootstrap kernel_vm_end on first real VM use
2609 if (kernel_vm_end == 0) {
2610 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2612 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2613 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2614 ~(PAGE_SIZE * NPTEPG - 1);
2616 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2617 kernel_vm_end = kernel_map.max_offset;
2624 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2625 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2626 * do not want to force-fill 128G worth of page tables.
2628 if (kstart < KERNBASE) {
2629 if (kstart > kernel_vm_end)
2630 kstart = kernel_vm_end;
2631 KKASSERT(kend <= KERNBASE);
2632 update_kernel_vm_end = 1;
2634 update_kernel_vm_end = 0;
2637 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2638 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2640 if (kend - 1 >= kernel_map.max_offset)
2641 kend = kernel_map.max_offset;
2643 while (kstart < kend) {
2644 pt = pmap_pt(&kernel_pmap, kstart);
2646 /* We need a new PDP entry */
2647 nkpg = vm_page_alloc(NULL, nkpt,
2650 VM_ALLOC_INTERRUPT);
2652 panic("pmap_growkernel: no memory to grow "
2655 paddr = VM_PAGE_TO_PHYS(nkpg);
2656 if ((nkpg->flags & PG_ZERO) == 0)
2657 pmap_zero_page(paddr);
2658 vm_page_flag_clear(nkpg, PG_ZERO);
2659 newpd = (pdp_entry_t)
2661 kernel_pmap.pmap_bits[PG_V_IDX] |
2662 kernel_pmap.pmap_bits[PG_RW_IDX] |
2663 kernel_pmap.pmap_bits[PG_A_IDX] |
2664 kernel_pmap.pmap_bits[PG_M_IDX]);
2665 *pmap_pd(&kernel_pmap, kstart) = newpd;
2667 continue; /* try again */
2669 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2670 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2671 ~(PAGE_SIZE * NPTEPG - 1);
2672 if (kstart - 1 >= kernel_map.max_offset) {
2673 kstart = kernel_map.max_offset;
2680 * This index is bogus, but out of the way
2682 nkpg = vm_page_alloc(NULL, nkpt,
2685 VM_ALLOC_INTERRUPT);
2687 panic("pmap_growkernel: no memory to grow kernel");
2690 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2691 pmap_zero_page(ptppaddr);
2692 vm_page_flag_clear(nkpg, PG_ZERO);
2693 newpt = (pd_entry_t) (ptppaddr |
2694 kernel_pmap.pmap_bits[PG_V_IDX] |
2695 kernel_pmap.pmap_bits[PG_RW_IDX] |
2696 kernel_pmap.pmap_bits[PG_A_IDX] |
2697 kernel_pmap.pmap_bits[PG_M_IDX]);
2698 *pmap_pt(&kernel_pmap, kstart) = newpt;
2701 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2702 ~(PAGE_SIZE * NPTEPG - 1);
2704 if (kstart - 1 >= kernel_map.max_offset) {
2705 kstart = kernel_map.max_offset;
2711 * Only update kernel_vm_end for areas below KERNBASE.
2713 if (update_kernel_vm_end && kernel_vm_end < kstart)
2714 kernel_vm_end = kstart;
2718 * Add a reference to the specified pmap.
2721 pmap_reference(pmap_t pmap)
2724 lwkt_gettoken(&pmap->pm_token);
2726 lwkt_reltoken(&pmap->pm_token);
2730 /***************************************************
2731 * page management routines.
2732 ***************************************************/
2735 * Hold a pv without locking it
2738 pv_hold(pv_entry_t pv)
2742 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2746 count = pv->pv_hold;
2748 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2755 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2756 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2759 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2760 * pv list via its page) must be held by the caller.
2763 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2768 * Critical path shortcut expects pv to already have one ref
2769 * (for the pv->pv_pmap).
2771 if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
2774 pv->pv_line = lineno;
2780 count = pv->pv_hold;
2782 if ((count & PV_HOLD_LOCKED) == 0) {
2783 if (atomic_cmpset_int(&pv->pv_hold, count,
2784 (count + 1) | PV_HOLD_LOCKED)) {
2787 pv->pv_line = lineno;
2792 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2800 * Drop a previously held pv_entry which could not be locked, allowing its
2803 * Must not be called with a spinlock held as we might zfree() the pv if it
2804 * is no longer associated with a pmap and this was the last hold count.
2807 pv_drop(pv_entry_t pv)
2812 count = pv->pv_hold;
2814 KKASSERT((count & PV_HOLD_MASK) > 0);
2815 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2816 (PV_HOLD_LOCKED | 1));
2817 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2818 if ((count & PV_HOLD_MASK) == 1) {
2819 KKASSERT(count == 1);
2820 KKASSERT(pv->pv_pmap == NULL);
2830 * Find or allocate the requested PV entry, returning a locked, held pv.
2832 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
2833 * for the caller and one representing the pmap and vm_page association.
2835 * If (*isnew) is zero, the returned pv will have only one hold count.
2837 * Since both associations can only be adjusted while the pv is locked,
2838 * together they represent just one additional hold.
2842 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2845 pv_entry_t pnew = NULL;
2847 spin_lock(&pmap->pm_spin);
2849 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2850 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2855 spin_unlock(&pmap->pm_spin);
2856 pnew = zalloc(pvzone);
2857 spin_lock(&pmap->pm_spin);
2860 pnew->pv_pmap = pmap;
2861 pnew->pv_pindex = pindex;
2862 pnew->pv_hold = PV_HOLD_LOCKED | 2;
2864 pnew->pv_func = func;
2865 pnew->pv_line = lineno;
2867 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2868 ++pmap->pm_generation;
2869 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2870 spin_unlock(&pmap->pm_spin);
2875 spin_unlock(&pmap->pm_spin);
2876 zfree(pvzone, pnew);
2878 spin_lock(&pmap->pm_spin);
2881 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2882 spin_unlock(&pmap->pm_spin);
2884 spin_unlock(&pmap->pm_spin);
2885 _pv_lock(pv PMAP_DEBUG_COPY);
2887 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2892 spin_lock(&pmap->pm_spin);
2897 * Find the requested PV entry, returning a locked+held pv or NULL
2901 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2905 spin_lock(&pmap->pm_spin);
2910 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2911 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2915 spin_unlock(&pmap->pm_spin);
2918 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2919 spin_unlock(&pmap->pm_spin);
2921 spin_unlock(&pmap->pm_spin);
2922 _pv_lock(pv PMAP_DEBUG_COPY);
2924 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2925 pv_cache(pv, pindex);
2929 spin_lock(&pmap->pm_spin);
2934 * Lookup, hold, and attempt to lock (pmap,pindex).
2936 * If the entry does not exist NULL is returned and *errorp is set to 0
2938 * If the entry exists and could be successfully locked it is returned and
2939 * errorp is set to 0.
2941 * If the entry exists but could NOT be successfully locked it is returned
2942 * held and *errorp is set to 1.
2946 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2950 spin_lock_shared(&pmap->pm_spin);
2951 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2952 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2954 spin_unlock_shared(&pmap->pm_spin);
2958 if (pv_hold_try(pv)) {
2959 pv_cache(pv, pindex);
2960 spin_unlock_shared(&pmap->pm_spin);
2962 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
2963 return(pv); /* lock succeeded */
2965 spin_unlock_shared(&pmap->pm_spin);
2967 return (pv); /* lock failed */
2971 * Find the requested PV entry, returning a held pv or NULL
2975 pv_find(pmap_t pmap, vm_pindex_t pindex)
2979 spin_lock_shared(&pmap->pm_spin);
2981 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2982 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2984 spin_unlock_shared(&pmap->pm_spin);
2988 pv_cache(pv, pindex);
2989 spin_unlock_shared(&pmap->pm_spin);
2994 * Lock a held pv, keeping the hold count
2998 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3003 count = pv->pv_hold;
3005 if ((count & PV_HOLD_LOCKED) == 0) {
3006 if (atomic_cmpset_int(&pv->pv_hold, count,
3007 count | PV_HOLD_LOCKED)) {
3010 pv->pv_line = lineno;
3016 tsleep_interlock(pv, 0);
3017 if (atomic_cmpset_int(&pv->pv_hold, count,
3018 count | PV_HOLD_WAITING)) {
3020 kprintf("pv waiting on %s:%d\n",
3021 pv->pv_func, pv->pv_line);
3023 tsleep(pv, PINTERLOCKED, "pvwait", hz);
3030 * Unlock a held and locked pv, keeping the hold count.
3034 pv_unlock(pv_entry_t pv)
3039 count = pv->pv_hold;
3041 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3042 (PV_HOLD_LOCKED | 1));
3043 if (atomic_cmpset_int(&pv->pv_hold, count,
3045 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3046 if (count & PV_HOLD_WAITING)
3054 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3055 * and the hold count drops to zero we will free it.
3057 * Caller should not hold any spin locks. We are protected from hold races
3058 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3059 * lock held. A pv cannot be located otherwise.
3063 pv_put(pv_entry_t pv)
3066 * Fast - shortcut most common condition
3068 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3079 * Remove the pmap association from a pv, require that pv_m already be removed,
3080 * then unlock and drop the pv. Any pte operations must have already been
3081 * completed. This call may result in a last-drop which will physically free
3084 * Removing the pmap association entails an additional drop.
3086 * pv must be exclusively locked on call and will be disposed of on return.
3090 pv_free(pv_entry_t pv)
3094 KKASSERT(pv->pv_m == NULL);
3095 KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3096 if ((pmap = pv->pv_pmap) != NULL) {
3097 spin_lock(&pmap->pm_spin);
3098 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3099 ++pmap->pm_generation;
3100 if (pmap->pm_pvhint == pv)
3101 pmap->pm_pvhint = NULL;
3102 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3105 spin_unlock(&pmap->pm_spin);
3108 * Try to shortcut three atomic ops, otherwise fall through
3109 * and do it normally. Drop two refs and the lock all in
3112 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3116 pv_drop(pv); /* ref for pv_pmap */
3122 * This routine is very drastic, but can save the system
3130 static int warningdone=0;
3132 if (pmap_pagedaemon_waken == 0)
3134 pmap_pagedaemon_waken = 0;
3135 if (warningdone < 5) {
3136 kprintf("pmap_collect: collecting pv entries -- "
3137 "suggest increasing PMAP_SHPGPERPROC\n");
3141 for (i = 0; i < vm_page_array_size; i++) {
3142 m = &vm_page_array[i];
3143 if (m->wire_count || m->hold_count)
3145 if (vm_page_busy_try(m, TRUE) == 0) {
3146 if (m->wire_count == 0 && m->hold_count == 0) {
3155 * Scan the pmap for active page table entries and issue a callback.
3156 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3157 * its parent page table.
3159 * pte_pv will be NULL if the page or page table is unmanaged.
3160 * pt_pv will point to the page table page containing the pte for the page.
3162 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3163 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3164 * process pmap's PD and page to the callback function. This can be
3165 * confusing because the pt_pv is really a pd_pv, and the target page
3166 * table page is simply aliased by the pmap and not owned by it.
3168 * It is assumed that the start and end are properly rounded to the page size.
3170 * It is assumed that PD pages and above are managed and thus in the RB tree,
3171 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3173 struct pmap_scan_info {
3177 vm_pindex_t sva_pd_pindex;
3178 vm_pindex_t eva_pd_pindex;
3179 void (*func)(pmap_t, struct pmap_scan_info *,
3180 pv_entry_t, pv_entry_t, int, vm_offset_t,
3181 pt_entry_t *, void *);
3184 struct pmap_inval_info inval;
3187 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3188 static int pmap_scan_callback(pv_entry_t pv, void *data);
3191 pmap_scan(struct pmap_scan_info *info)
3193 struct pmap *pmap = info->pmap;
3194 pv_entry_t pd_pv; /* A page directory PV */
3195 pv_entry_t pt_pv; /* A page table PV */
3196 pv_entry_t pte_pv; /* A page table entry PV */
3199 struct pv_entry dummy_pv;
3206 * Hold the token for stability; if the pmap is empty we have nothing
3209 lwkt_gettoken(&pmap->pm_token);
3211 if (pmap->pm_stats.resident_count == 0) {
3212 lwkt_reltoken(&pmap->pm_token);
3217 pmap_inval_init(&info->inval);
3221 * Special handling for scanning one page, which is a very common
3222 * operation (it is?).
3224 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3226 if (info->sva + PAGE_SIZE == info->eva) {
3227 generation = pmap->pm_generation;
3228 if (info->sva >= VM_MAX_USER_ADDRESS) {
3230 * Kernel mappings do not track wire counts on
3231 * page table pages and only maintain pd_pv and
3232 * pte_pv levels so pmap_scan() works.
3235 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3236 ptep = vtopte(info->sva);
3239 * User pages which are unmanaged will not have a
3240 * pte_pv. User page table pages which are unmanaged
3241 * (shared from elsewhere) will also not have a pt_pv.
3242 * The func() callback will pass both pte_pv and pt_pv
3243 * as NULL in that case.
3245 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3246 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3247 if (pt_pv == NULL) {
3248 KKASSERT(pte_pv == NULL);
3249 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3251 ptep = pv_pte_lookup(pd_pv,
3252 pmap_pt_index(info->sva));
3254 info->func(pmap, info,
3263 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3267 * NOTE: *ptep can't be ripped out from under us if we hold
3268 * pte_pv locked, but bits can change. However, there is
3269 * a race where another thread may be inserting pte_pv
3270 * and setting *ptep just after our pte_pv lookup fails.
3272 * In this situation we can end up with a NULL pte_pv
3273 * but find that we have a managed *ptep. We explicitly
3274 * check for this race.
3280 * Unlike the pv_find() case below we actually
3281 * acquired a locked pv in this case so any
3282 * race should have been resolved. It is expected
3285 KKASSERT(pte_pv == NULL);
3286 } else if (pte_pv) {
3287 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3288 pmap->pmap_bits[PG_V_IDX])) ==
3289 (pmap->pmap_bits[PG_MANAGED_IDX] |
3290 pmap->pmap_bits[PG_V_IDX]),
3291 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3293 *ptep, oldpte, info->sva, pte_pv,
3294 generation, pmap->pm_generation));
3295 info->func(pmap, info, pte_pv, pt_pv, 0,
3296 info->sva, ptep, info->arg);
3299 * Check for insertion race
3301 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3303 pte_pv = pv_find(pmap,
3304 pmap_pte_pindex(info->sva));
3308 kprintf("pmap_scan: RACE1 "
3318 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3319 pmap->pmap_bits[PG_V_IDX])) ==
3320 pmap->pmap_bits[PG_V_IDX],
3321 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3323 *ptep, oldpte, info->sva,
3324 generation, pmap->pm_generation));
3325 info->func(pmap, info, NULL, pt_pv, 0,
3326 info->sva, ptep, info->arg);
3331 pmap_inval_done(&info->inval);
3332 lwkt_reltoken(&pmap->pm_token);
3337 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3340 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3341 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3343 if (info->sva >= VM_MAX_USER_ADDRESS) {
3345 * The kernel does not currently maintain any pv_entry's for
3346 * higher-level page tables.
3348 bzero(&dummy_pv, sizeof(dummy_pv));
3349 dummy_pv.pv_pindex = info->sva_pd_pindex;
3350 spin_lock(&pmap->pm_spin);
3351 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3352 pmap_scan_callback(&dummy_pv, info);
3353 ++dummy_pv.pv_pindex;
3355 spin_unlock(&pmap->pm_spin);
3358 * User page tables maintain local PML4, PDP, and PD
3359 * pv_entry's at the very least. PT pv's might be
3360 * unmanaged and thus not exist. PTE pv's might be
3361 * unmanaged and thus not exist.
3363 spin_lock(&pmap->pm_spin);
3364 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3365 pmap_scan_cmp, pmap_scan_callback, info);
3366 spin_unlock(&pmap->pm_spin);
3368 pmap_inval_done(&info->inval);
3369 lwkt_reltoken(&pmap->pm_token);
3373 * WARNING! pmap->pm_spin held
3376 pmap_scan_cmp(pv_entry_t pv, void *data)
3378 struct pmap_scan_info *info = data;
3379 if (pv->pv_pindex < info->sva_pd_pindex)
3381 if (pv->pv_pindex >= info->eva_pd_pindex)
3387 * WARNING! pmap->pm_spin held
3390 pmap_scan_callback(pv_entry_t pv, void *data)
3392 struct pmap_scan_info *info = data;
3393 struct pmap *pmap = info->pmap;
3394 pv_entry_t pd_pv; /* A page directory PV */
3395 pv_entry_t pt_pv; /* A page table PV */
3396 pv_entry_t pte_pv; /* A page table entry PV */
3401 vm_offset_t va_next;
3402 vm_pindex_t pd_pindex;
3407 * Pull the PD pindex from the pv before releasing the spinlock.
3409 * WARNING: pv is faked for kernel pmap scans.
3411 pd_pindex = pv->pv_pindex;
3412 spin_unlock(&pmap->pm_spin);
3413 pv = NULL; /* invalid after spinlock unlocked */
3416 * Calculate the page range within the PD. SIMPLE pmaps are
3417 * direct-mapped for the entire 2^64 address space. Normal pmaps
3418 * reflect the user and kernel address space which requires
3419 * cannonicalization w/regards to converting pd_pindex's back
3422 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3423 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3424 (sva & PML4_SIGNMASK)) {
3425 sva |= PML4_SIGNMASK;
3427 eva = sva + NBPDP; /* can overflow */
3428 if (sva < info->sva)
3430 if (eva < info->sva || eva > info->eva)
3434 * NOTE: kernel mappings do not track page table pages, only
3437 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3438 * However, for the scan to be efficient we try to
3439 * cache items top-down.
3444 for (; sva < eva; sva = va_next) {
3445 if (sva >= VM_MAX_USER_ADDRESS) {
3454 * PD cache (degenerate case if we skip). It is possible
3455 * for the PD to not exist due to races. This is ok.
3457 if (pd_pv == NULL) {
3458 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3459 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3461 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3463 if (pd_pv == NULL) {
3464 va_next = (sva + NBPDP) & ~PDPMASK;
3473 if (pt_pv == NULL) {
3478 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3479 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3485 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3489 * If pt_pv is NULL we either have an shared page table
3490 * page and must issue a callback specific to that case,
3491 * or there is no page table page.
3493 * Either way we can skip the page table page.
3495 if (pt_pv == NULL) {
3497 * Possible unmanaged (shared from another pmap)
3501 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3502 KKASSERT(pd_pv != NULL);
3503 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3504 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3505 info->func(pmap, info, NULL, pd_pv, 1,
3506 sva, ptep, info->arg);
3510 * Done, move to next page table page.
3512 va_next = (sva + NBPDR) & ~PDRMASK;
3519 * From this point in the loop testing pt_pv for non-NULL
3520 * means we are in UVM, else if it is NULL we are in KVM.
3522 * Limit our scan to either the end of the va represented
3523 * by the current page table page, or to the end of the
3524 * range being removed.
3527 va_next = (sva + NBPDR) & ~PDRMASK;
3534 * Scan the page table for pages. Some pages may not be
3535 * managed (might not have a pv_entry).
3537 * There is no page table management for kernel pages so
3538 * pt_pv will be NULL in that case, but otherwise pt_pv
3539 * is non-NULL, locked, and referenced.
3543 * At this point a non-NULL pt_pv means a UVA, and a NULL
3544 * pt_pv means a KVA.
3547 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3551 while (sva < va_next) {
3553 * Acquire the related pte_pv, if any. If *ptep == 0
3554 * the related pte_pv should not exist, but if *ptep
3555 * is not zero the pte_pv may or may not exist (e.g.
3556 * will not exist for an unmanaged page).
3558 * However a multitude of races are possible here.
3560 * In addition, the (pt_pv, pte_pv) lock order is
3561 * backwards, so we have to be careful in aquiring
3562 * a properly locked pte_pv.
3564 generation = pmap->pm_generation;
3566 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3573 pv_put(pt_pv); /* must be non-NULL */
3575 pv_lock(pte_pv); /* safe to block now */
3578 pt_pv = pv_get(pmap,
3579 pmap_pt_pindex(sva));
3581 * pt_pv reloaded, need new ptep
3583 KKASSERT(pt_pv != NULL);
3584 ptep = pv_pte_lookup(pt_pv,
3585 pmap_pte_index(sva));
3589 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3593 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3598 kprintf("Unexpected non-NULL pte_pv "
3600 "*ptep = %016lx/%016lx\n",
3601 pte_pv, pt_pv, *ptep, oldpte);
3602 panic("Unexpected non-NULL pte_pv");
3610 * Ready for the callback. The locked pte_pv (if any)
3611 * is consumed by the callback. pte_pv will exist if
3612 * the page is managed, and will not exist if it
3616 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3617 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3618 ("badC *ptep %016lx/%016lx sva %016lx "
3619 "pte_pv %p pm_generation %d/%d",
3620 *ptep, oldpte, sva, pte_pv,
3621 generation, pmap->pm_generation));
3622 info->func(pmap, info, pte_pv, pt_pv, 0,
3623 sva, ptep, info->arg);
3626 * Check for insertion race. Since there is no
3627 * pte_pv to guard us it is possible for us
3628 * to race another thread doing an insertion.
3629 * Our lookup misses the pte_pv but our *ptep
3630 * check sees the inserted pte.
3632 * XXX panic case seems to occur within a
3633 * vm_fork() of /bin/sh, which frankly
3634 * shouldn't happen since no other threads
3635 * should be inserting to our pmap in that
3636 * situation. Removing, possibly. Inserting,
3639 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3641 pte_pv = pv_find(pmap,
3642 pmap_pte_pindex(sva));
3645 kprintf("pmap_scan: RACE2 "
3655 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3656 pmap->pmap_bits[PG_V_IDX],
3657 ("badD *ptep %016lx/%016lx sva %016lx "
3658 "pte_pv NULL pm_generation %d/%d",
3660 generation, pmap->pm_generation));
3661 info->func(pmap, info, NULL, pt_pv, 0,
3662 sva, ptep, info->arg);
3681 * Relock before returning.
3683 spin_lock(&pmap->pm_spin);
3688 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3690 struct pmap_scan_info info;
3695 info.func = pmap_remove_callback;
3697 info.doinval = 1; /* normal remove requires pmap inval */
3702 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3704 struct pmap_scan_info info;
3709 info.func = pmap_remove_callback;
3711 info.doinval = 0; /* normal remove requires pmap inval */
3716 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
3717 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3718 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3724 * This will also drop pt_pv's wire_count. Note that
3725 * terminal pages are not wired based on mmu presence.
3728 pmap_remove_pv_pte(pte_pv, pt_pv, &info->inval);
3730 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3731 pmap_remove_pv_page(pte_pv);
3733 } else if (sharept == 0) {
3737 * pt_pv's wire_count is still bumped by unmanaged pages
3738 * so we must decrement it manually.
3741 pmap_inval_interlock(&info->inval, pmap, va);
3742 pte = pte_load_clear(ptep);
3744 pmap_inval_deinterlock(&info->inval, pmap);
3745 if (pte & pmap->pmap_bits[PG_W_IDX])
3746 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3747 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3748 if (vm_page_unwire_quick(pt_pv->pv_m))
3749 panic("pmap_remove: insufficient wirecount");
3752 * Unmanaged page table, pt_pv is actually the pd_pv
3753 * for our pmap (not the share object pmap).
3755 * We have to unwire the target page table page and we
3756 * have to unwire our page directory page.
3759 pmap_inval_interlock(&info->inval, pmap, va);
3760 pte = pte_load_clear(ptep);
3762 pmap_inval_deinterlock(&info->inval, pmap);
3763 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3764 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
3765 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3766 panic("pmap_remove: shared pgtable1 bad wirecount");
3767 if (vm_page_unwire_quick(pt_pv->pv_m))
3768 panic("pmap_remove: shared pgtable2 bad wirecount");
3773 * Removes this physical page from all physical maps in which it resides.
3774 * Reflects back modify bits to the pager.
3776 * This routine may not be called from an interrupt.
3780 pmap_remove_all(vm_page_t m)
3782 struct pmap_inval_info info;
3785 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
3788 pmap_inval_init(&info);
3789 vm_page_spin_lock(m);
3790 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3791 KKASSERT(pv->pv_m == m);
3792 if (pv_hold_try(pv)) {
3793 vm_page_spin_unlock(m);
3795 vm_page_spin_unlock(m);
3798 if (pv->pv_m != m) {
3800 vm_page_spin_lock(m);
3804 * Holding no spinlocks, pv is locked.
3806 pmap_remove_pv_pte(pv, NULL, &info);
3807 pmap_remove_pv_page(pv);
3809 vm_page_spin_lock(m);
3811 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3812 vm_page_spin_unlock(m);
3813 pmap_inval_done(&info);
3817 * Set the physical protection on the specified range of this map
3818 * as requested. This function is typically only used for debug watchpoints
3821 * This function may not be called from an interrupt if the map is
3822 * not the kernel_pmap.
3824 * NOTE! For shared page table pages we just unmap the page.
3827 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3829 struct pmap_scan_info info;
3830 /* JG review for NX */
3834 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3835 pmap_remove(pmap, sva, eva);
3838 if (prot & VM_PROT_WRITE)
3843 info.func = pmap_protect_callback;
3851 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
3852 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3853 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3863 pmap_inval_interlock(&info->inval, pmap, va);
3869 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
3870 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3871 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3872 KKASSERT(m == pte_pv->pv_m);
3873 vm_page_flag_set(m, PG_REFERENCED);
3875 cbits &= ~pmap->pmap_bits[PG_A_IDX];
3877 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
3878 if (pmap_track_modified(pte_pv->pv_pindex)) {
3879 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3881 m = PHYS_TO_VM_PAGE(pbits &
3886 cbits &= ~pmap->pmap_bits[PG_M_IDX];
3889 } else if (sharept) {
3891 * Unmanaged page table, pt_pv is actually the pd_pv
3892 * for our pmap (not the share object pmap).
3894 * When asked to protect something in a shared page table
3895 * page we just unmap the page table page. We have to
3896 * invalidate the tlb in this situation.
3898 * XXX Warning, shared page tables will not be used for
3899 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
3900 * so PHYS_TO_VM_PAGE() should be safe here.
3902 pte = pte_load_clear(ptep);
3903 pmap_inval_invltlb(&info->inval);
3904 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3905 panic("pmap_protect: pgtable1 pg bad wirecount");
3906 if (vm_page_unwire_quick(pt_pv->pv_m))
3907 panic("pmap_protect: pgtable2 pg bad wirecount");
3910 /* else unmanaged page, adjust bits, no wire changes */
3913 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
3914 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
3918 pmap_inval_deinterlock(&info->inval, pmap);
3924 * Insert the vm_page (m) at the virtual address (va), replacing any prior
3925 * mapping at that address. Set protection and wiring as requested.
3927 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
3928 * possible. If it is we enter the page into the appropriate shared pmap
3929 * hanging off the related VM object instead of the passed pmap, then we
3930 * share the page table page from the VM object's pmap into the current pmap.
3932 * NOTE: This routine MUST insert the page into the pmap now, it cannot
3936 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3937 boolean_t wired, vm_map_entry_t entry)
3939 pmap_inval_info info;
3940 pv_entry_t pt_pv; /* page table */
3941 pv_entry_t pte_pv; /* page table entry */
3944 pt_entry_t origpte, newpte;
3949 va = trunc_page(va);
3950 #ifdef PMAP_DIAGNOSTIC
3952 panic("pmap_enter: toobig");
3953 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
3954 panic("pmap_enter: invalid to pmap_enter page table "
3955 "pages (va: 0x%lx)", va);
3957 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
3958 kprintf("Warning: pmap_enter called on UVA with "
3961 db_print_backtrace();
3964 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
3965 kprintf("Warning: pmap_enter called on KVA without"
3968 db_print_backtrace();
3973 * Get locked PV entries for our new page table entry (pte_pv)
3974 * and for its parent page table (pt_pv). We need the parent
3975 * so we can resolve the location of the ptep.
3977 * Only hardware MMU actions can modify the ptep out from
3980 * if (m) is fictitious or unmanaged we do not create a managing
3981 * pte_pv for it. Any pre-existing page's management state must
3982 * match (avoiding code complexity).
3984 * If the pmap is still being initialized we assume existing
3987 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3989 if (pmap_initialized == FALSE) {
3994 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
3996 if (va >= VM_MAX_USER_ADDRESS) {
4000 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4002 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4006 KKASSERT(origpte == 0 ||
4007 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0);
4009 if (va >= VM_MAX_USER_ADDRESS) {
4011 * Kernel map, pv_entry-tracked.
4014 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4020 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4022 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4026 KKASSERT(origpte == 0 ||
4027 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]));
4030 pa = VM_PAGE_TO_PHYS(m);
4031 opa = origpte & PG_FRAME;
4033 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4034 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4036 newpte |= pmap->pmap_bits[PG_W_IDX];
4037 if (va < VM_MAX_USER_ADDRESS)
4038 newpte |= pmap->pmap_bits[PG_U_IDX];
4040 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4041 // if (pmap == &kernel_pmap)
4042 // newpte |= pgeflag;
4043 newpte |= pmap->pmap_cache_bits[m->pat_mode];
4044 if (m->flags & PG_FICTITIOUS)
4045 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4048 * It is possible for multiple faults to occur in threaded
4049 * environments, the existing pte might be correct.
4051 if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4052 pmap->pmap_bits[PG_A_IDX])) == 0)
4055 if ((prot & VM_PROT_NOSYNC) == 0)
4056 pmap_inval_init(&info);
4059 * Ok, either the address changed or the protection or wiring
4062 * Clear the current entry, interlocking the removal. For managed
4063 * pte's this will also flush the modified state to the vm_page.
4064 * Atomic ops are mandatory in order to ensure that PG_M events are
4065 * not lost during any transition.
4067 * WARNING: The caller has busied the new page but not the original
4068 * vm_page which we are trying to replace. Because we hold
4069 * the pte_pv lock, but have not busied the page, PG bits
4070 * can be cleared out from under us.
4075 * pmap_remove_pv_pte() unwires pt_pv and assumes
4076 * we will free pte_pv, but since we are reusing
4077 * pte_pv we want to retain the wire count.
4079 * pt_pv won't exist for a kernel page (managed or
4083 vm_page_wire_quick(pt_pv->pv_m);
4084 if (prot & VM_PROT_NOSYNC)
4085 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
4087 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
4089 pmap_remove_pv_page(pte_pv);
4090 } else if (prot & VM_PROT_NOSYNC) {
4092 * Unmanaged page, NOSYNC (no mmu sync) requested.
4094 * Leave wire count on PT page intact.
4096 (void)pte_load_clear(ptep);
4097 cpu_invlpg((void *)va);
4098 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4101 * Unmanaged page, normal enter.
4103 * Leave wire count on PT page intact.
4105 pmap_inval_interlock(&info, pmap, va);
4106 (void)pte_load_clear(ptep);
4107 pmap_inval_deinterlock(&info, pmap);
4108 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4110 KKASSERT(*ptep == 0);
4115 * Enter on the PV list if part of our managed memory.
4116 * Wiring of the PT page is already handled.
4118 KKASSERT(pte_pv->pv_m == NULL);
4119 vm_page_spin_lock(m);
4121 pmap_page_stats_adding(m);
4122 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4123 vm_page_flag_set(m, PG_MAPPED);
4124 vm_page_spin_unlock(m);
4125 } else if (pt_pv && opa == 0) {
4127 * We have to adjust the wire count on the PT page ourselves
4128 * for unmanaged entries. If opa was non-zero we retained
4129 * the existing wire count from the removal.
4131 vm_page_wire_quick(pt_pv->pv_m);
4135 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4137 * User VMAs do not because those will be zero->non-zero, so no
4138 * stale entries to worry about at this point.
4140 * For KVM there appear to still be issues. Theoretically we
4141 * should be able to scrap the interlocks entirely but we
4144 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
4145 pmap_inval_interlock(&info, pmap, va);
4150 *(volatile pt_entry_t *)ptep = newpte;
4152 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
4153 pmap_inval_deinterlock(&info, pmap);
4154 else if (pt_pv == NULL)
4155 cpu_invlpg((void *)va);
4159 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4162 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4165 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4166 vm_page_flag_set(m, PG_WRITEABLE);
4169 * Unmanaged pages need manual resident_count tracking.
4171 if (pte_pv == NULL && pt_pv)
4172 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4177 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
4178 pmap_inval_done(&info);
4180 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4181 (m->flags & PG_MAPPED));
4184 * Cleanup the pv entry, allowing other accessors.
4193 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4194 * This code also assumes that the pmap has no pre-existing entry for this
4197 * This code currently may only be used on user pmaps, not kernel_pmap.
4200 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4202 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4206 * Make a temporary mapping for a physical address. This is only intended
4207 * to be used for panic dumps.
4209 * The caller is responsible for calling smp_invltlb().
4212 pmap_kenter_temporary(vm_paddr_t pa, long i)
4214 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4215 return ((void *)crashdumpmap);
4218 #define MAX_INIT_PT (96)
4221 * This routine preloads the ptes for a given object into the specified pmap.
4222 * This eliminates the blast of soft faults on process startup and
4223 * immediately after an mmap.
4225 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4228 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4229 vm_object_t object, vm_pindex_t pindex,
4230 vm_size_t size, int limit)
4232 struct rb_vm_page_scan_info info;
4237 * We can't preinit if read access isn't set or there is no pmap
4240 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4244 * We can't preinit if the pmap is not the current pmap
4246 lp = curthread->td_lwp;
4247 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4251 * Misc additional checks
4253 psize = x86_64_btop(size);
4255 if ((object->type != OBJT_VNODE) ||
4256 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4257 (object->resident_page_count > MAX_INIT_PT))) {
4261 if (pindex + psize > object->size) {
4262 if (object->size < pindex)
4264 psize = object->size - pindex;
4271 * If everything is segment-aligned do not pre-init here. Instead
4272 * allow the normal vm_fault path to pass a segment hint to
4273 * pmap_enter() which will then use an object-referenced shared
4276 if ((addr & SEG_MASK) == 0 &&
4277 (ctob(psize) & SEG_MASK) == 0 &&
4278 (ctob(pindex) & SEG_MASK) == 0) {
4283 * Use a red-black scan to traverse the requested range and load
4284 * any valid pages found into the pmap.
4286 * We cannot safely scan the object's memq without holding the
4289 info.start_pindex = pindex;
4290 info.end_pindex = pindex + psize - 1;
4296 vm_object_hold_shared(object);
4297 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4298 pmap_object_init_pt_callback, &info);
4299 vm_object_drop(object);
4304 pmap_object_init_pt_callback(vm_page_t p, void *data)
4306 struct rb_vm_page_scan_info *info = data;
4307 vm_pindex_t rel_index;
4310 * don't allow an madvise to blow away our really
4311 * free pages allocating pv entries.
4313 if ((info->limit & MAP_PREFAULT_MADVISE) &&
4314 vmstats.v_free_count < vmstats.v_free_reserved) {
4319 * Ignore list markers and ignore pages we cannot instantly
4320 * busy (while holding the object token).
4322 if (p->flags & PG_MARKER)
4324 if (vm_page_busy_try(p, TRUE))
4326 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4327 (p->flags & PG_FICTITIOUS) == 0) {
4328 if ((p->queue - p->pc) == PQ_CACHE)
4329 vm_page_deactivate(p);
4330 rel_index = p->pindex - info->start_pindex;
4331 pmap_enter_quick(info->pmap,
4332 info->addr + x86_64_ptob(rel_index), p);
4340 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4343 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4346 * XXX This is safe only because page table pages are not freed.
4349 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4353 /*spin_lock(&pmap->pm_spin);*/
4354 if ((pte = pmap_pte(pmap, addr)) != NULL) {
4355 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4356 /*spin_unlock(&pmap->pm_spin);*/
4360 /*spin_unlock(&pmap->pm_spin);*/
4365 * Change the wiring attribute for a pmap/va pair. The mapping must already
4366 * exist in the pmap. The mapping may or may not be managed.
4369 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4370 vm_map_entry_t entry)
4377 lwkt_gettoken(&pmap->pm_token);
4378 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
4379 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4381 if (wired && !pmap_pte_w(pmap, ptep))
4382 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4383 else if (!wired && pmap_pte_w(pmap, ptep))
4384 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4387 * Wiring is not a hardware characteristic so there is no need to
4388 * invalidate TLB. However, in an SMP environment we must use
4389 * a locked bus cycle to update the pte (if we are not using
4390 * the pmap_inval_*() API that is)... it's ok to do this for simple
4394 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4396 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4398 lwkt_reltoken(&pmap->pm_token);
4404 * Copy the range specified by src_addr/len from the source map to
4405 * the range dst_addr/len in the destination map.
4407 * This routine is only advisory and need not do anything.
4410 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4411 vm_size_t len, vm_offset_t src_addr)
4418 * Zero the specified physical page.
4420 * This function may be called from an interrupt and no locking is
4424 pmap_zero_page(vm_paddr_t phys)
4426 vm_offset_t va = PHYS_TO_DMAP(phys);
4428 pagezero((void *)va);
4432 * pmap_page_assertzero:
4434 * Assert that a page is empty, panic if it isn't.
4437 pmap_page_assertzero(vm_paddr_t phys)
4439 vm_offset_t va = PHYS_TO_DMAP(phys);
4442 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
4443 if (*(long *)((char *)va + i) != 0) {
4444 panic("pmap_page_assertzero() @ %p not zero!",
4445 (void *)(intptr_t)va);
4453 * Zero part of a physical page by mapping it into memory and clearing
4454 * its contents with bzero.
4456 * off and size may not cover an area beyond a single hardware page.
4459 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4461 vm_offset_t virt = PHYS_TO_DMAP(phys);
4463 bzero((char *)virt + off, size);
4469 * Copy the physical page from the source PA to the target PA.
4470 * This function may be called from an interrupt. No locking
4474 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4476 vm_offset_t src_virt, dst_virt;
4478 src_virt = PHYS_TO_DMAP(src);
4479 dst_virt = PHYS_TO_DMAP(dst);
4480 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4484 * pmap_copy_page_frag:
4486 * Copy the physical page from the source PA to the target PA.
4487 * This function may be called from an interrupt. No locking
4491 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4493 vm_offset_t src_virt, dst_virt;
4495 src_virt = PHYS_TO_DMAP(src);
4496 dst_virt = PHYS_TO_DMAP(dst);
4498 bcopy((char *)src_virt + (src & PAGE_MASK),
4499 (char *)dst_virt + (dst & PAGE_MASK),
4504 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4505 * this page. This count may be changed upwards or downwards in the future;
4506 * it is only necessary that true be returned for a small subset of pmaps
4507 * for proper page aging.
4510 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4515 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4518 vm_page_spin_lock(m);
4519 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4520 if (pv->pv_pmap == pmap) {
4521 vm_page_spin_unlock(m);
4528 vm_page_spin_unlock(m);
4533 * Remove all pages from specified address space this aids process exit
4534 * speeds. Also, this code may be special cased for the current process
4538 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4540 pmap_remove_noinval(pmap, sva, eva);
4545 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4546 * routines are inline, and a lot of things compile-time evaluate.
4550 pmap_testbit(vm_page_t m, int bit)
4556 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4559 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4561 vm_page_spin_lock(m);
4562 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4563 vm_page_spin_unlock(m);
4567 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4569 #if defined(PMAP_DIAGNOSTIC)
4570 if (pv->pv_pmap == NULL) {
4571 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4579 * If the bit being tested is the modified bit, then
4580 * mark clean_map and ptes as never
4583 * WARNING! Because we do not lock the pv, *pte can be in a
4584 * state of flux. Despite this the value of *pte
4585 * will still be related to the vm_page in some way
4586 * because the pv cannot be destroyed as long as we
4587 * hold the vm_page spin lock.
4589 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4590 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4591 if (!pmap_track_modified(pv->pv_pindex))
4595 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4596 if (*pte & pmap->pmap_bits[bit]) {
4597 vm_page_spin_unlock(m);
4601 vm_page_spin_unlock(m);
4606 * This routine is used to modify bits in ptes. Only one bit should be
4607 * specified. PG_RW requires special handling.
4609 * Caller must NOT hold any spin locks
4613 pmap_clearbit(vm_page_t m, int bit_index)
4615 struct pmap_inval_info info;
4621 if (bit_index == PG_RW_IDX)
4622 vm_page_flag_clear(m, PG_WRITEABLE);
4623 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4630 * Loop over all current mappings setting/clearing as appropos If
4631 * setting RO do we need to clear the VAC?
4633 * NOTE: When clearing PG_M we could also (not implemented) drop
4634 * through to the PG_RW code and clear PG_RW too, forcing
4635 * a fault on write to redetect PG_M for virtual kernels, but
4636 * it isn't necessary since virtual kernels invalidate the
4637 * pte when they clear the VPTE_M bit in their virtual page
4640 * NOTE: Does not re-dirty the page when clearing only PG_M.
4642 * NOTE: Because we do not lock the pv, *pte can be in a state of
4643 * flux. Despite this the value of *pte is still somewhat
4644 * related while we hold the vm_page spin lock.
4646 * *pte can be zero due to this race. Since we are clearing
4647 * bits we basically do no harm when this race ccurs.
4649 if (bit_index != PG_RW_IDX) {
4650 vm_page_spin_lock(m);
4651 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4652 #if defined(PMAP_DIAGNOSTIC)
4653 if (pv->pv_pmap == NULL) {
4654 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4660 pte = pmap_pte_quick(pv->pv_pmap,
4661 pv->pv_pindex << PAGE_SHIFT);
4663 if (pbits & pmap->pmap_bits[bit_index])
4664 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
4666 vm_page_spin_unlock(m);
4671 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4674 pmap_inval_init(&info);
4677 vm_page_spin_lock(m);
4678 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4680 * don't write protect pager mappings
4682 if (!pmap_track_modified(pv->pv_pindex))
4685 #if defined(PMAP_DIAGNOSTIC)
4686 if (pv->pv_pmap == NULL) {
4687 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4694 * Skip pages which do not have PG_RW set.
4696 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4697 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
4703 if (pv_hold_try(pv)) {
4704 vm_page_spin_unlock(m);
4706 vm_page_spin_unlock(m);
4707 pv_lock(pv); /* held, now do a blocking lock */
4709 if (pv->pv_pmap != pmap || pv->pv_m != m) {
4710 pv_put(pv); /* and release */
4711 goto restart; /* anything could have happened */
4713 pmap_inval_interlock(&info, pmap,
4714 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
4715 KKASSERT(pv->pv_pmap == pmap);
4719 if (atomic_cmpset_long(pte, pbits, pbits &
4720 ~(pmap->pmap_bits[PG_RW_IDX] |
4721 pmap->pmap_bits[PG_M_IDX]))) {
4725 pmap_inval_deinterlock(&info, pmap);
4726 vm_page_spin_lock(m);
4729 * If PG_M was found to be set while we were clearing PG_RW
4730 * we also clear PG_M (done above) and mark the page dirty.
4731 * Callers expect this behavior.
4733 if (pbits & pmap->pmap_bits[PG_M_IDX])
4737 vm_page_spin_unlock(m);
4738 pmap_inval_done(&info);
4742 * Lower the permission for all mappings to a given page.
4744 * Page must be busied by caller. Because page is busied by caller this
4745 * should not be able to race a pmap_enter().
4748 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4750 /* JG NX support? */
4751 if ((prot & VM_PROT_WRITE) == 0) {
4752 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4754 * NOTE: pmap_clearbit(.. PG_RW) also clears
4755 * the PG_WRITEABLE flag in (m).
4757 pmap_clearbit(m, PG_RW_IDX);
4765 pmap_phys_address(vm_pindex_t ppn)
4767 return (x86_64_ptob(ppn));
4771 * Return a count of reference bits for a page, clearing those bits.
4772 * It is not necessary for every reference bit to be cleared, but it
4773 * is necessary that 0 only be returned when there are truly no
4774 * reference bits set.
4776 * XXX: The exact number of bits to check and clear is a matter that
4777 * should be tested and standardized at some point in the future for
4778 * optimal aging of shared pages.
4780 * This routine may not block.
4783 pmap_ts_referenced(vm_page_t m)
4790 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4793 vm_page_spin_lock(m);
4794 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4795 if (!pmap_track_modified(pv->pv_pindex))
4798 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4799 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
4800 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
4806 vm_page_spin_unlock(m);
4813 * Return whether or not the specified physical page was modified
4814 * in any physical maps.
4817 pmap_is_modified(vm_page_t m)
4821 res = pmap_testbit(m, PG_M_IDX);
4826 * Clear the modify bits on the specified physical page.
4829 pmap_clear_modify(vm_page_t m)
4831 pmap_clearbit(m, PG_M_IDX);
4835 * pmap_clear_reference:
4837 * Clear the reference bit on the specified physical page.
4840 pmap_clear_reference(vm_page_t m)
4842 pmap_clearbit(m, PG_A_IDX);
4846 * Miscellaneous support routines follow
4851 i386_protection_init(void)
4855 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4856 kp = protection_codes;
4857 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
4859 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4861 * Read access is also 0. There isn't any execute bit,
4862 * so just make it readable.
4864 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4865 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4866 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4869 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4870 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4871 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4872 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4873 *kp++ = pmap_bits_default[PG_RW_IDX];
4880 * Map a set of physical memory pages into the kernel virtual
4881 * address space. Return a pointer to where it is mapped. This
4882 * routine is intended to be used for mapping device memory,
4885 * NOTE: We can't use pgeflag unless we invalidate the pages one at
4888 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
4889 * work whether the cpu supports PAT or not. The remaining PAT
4890 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
4894 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
4896 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4900 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
4902 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
4906 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
4908 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4912 * Map a set of physical memory pages into the kernel virtual
4913 * address space. Return a pointer to where it is mapped. This
4914 * routine is intended to be used for mapping device memory,
4918 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
4920 vm_offset_t va, tmpva, offset;
4924 offset = pa & PAGE_MASK;
4925 size = roundup(offset + size, PAGE_SIZE);
4927 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4929 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4931 pa = pa & ~PAGE_MASK;
4932 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
4933 pte = vtopte(tmpva);
4935 kernel_pmap.pmap_bits[PG_RW_IDX] |
4936 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
4937 kernel_pmap.pmap_cache_bits[mode];
4938 tmpsize -= PAGE_SIZE;
4942 pmap_invalidate_range(&kernel_pmap, va, va + size);
4943 pmap_invalidate_cache_range(va, va + size);
4945 return ((void *)(va + offset));
4949 pmap_unmapdev(vm_offset_t va, vm_size_t size)
4951 vm_offset_t base, offset;
4953 base = va & ~PAGE_MASK;
4954 offset = va & PAGE_MASK;
4955 size = roundup(offset + size, PAGE_SIZE);
4956 pmap_qremove(va, size >> PAGE_SHIFT);
4957 kmem_free(&kernel_map, base, size);
4961 * Sets the memory attribute for the specified page.
4964 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
4970 * If "m" is a normal page, update its direct mapping. This update
4971 * can be relied upon to perform any cache operations that are
4972 * required for data coherence.
4974 if ((m->flags & PG_FICTITIOUS) == 0)
4975 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE,
4980 * Change the PAT attribute on an existing kernel memory map. Caller
4981 * must ensure that the virtual memory in question is not accessed
4982 * during the adjustment.
4985 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
4992 panic("pmap_change_attr: va is NULL");
4993 base = trunc_page(va);
4997 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
4998 kernel_pmap.pmap_cache_bits[mode];
5003 changed = 1; /* XXX: not optimal */
5006 * Flush CPU caches if required to make sure any data isn't cached that
5007 * shouldn't be, etc.
5010 pmap_invalidate_range(&kernel_pmap, base, va);
5011 pmap_invalidate_cache_range(base, va);
5016 * perform the pmap work for mincore
5019 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5021 pt_entry_t *ptep, pte;
5025 lwkt_gettoken(&pmap->pm_token);
5026 ptep = pmap_pte(pmap, addr);
5028 if (ptep && (pte = *ptep) != 0) {
5031 val = MINCORE_INCORE;
5032 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5035 pa = pte & PG_FRAME;
5037 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5040 m = PHYS_TO_VM_PAGE(pa);
5045 if (pte & pmap->pmap_bits[PG_M_IDX])
5046 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5048 * Modified by someone
5050 else if (m && (m->dirty || pmap_is_modified(m)))
5051 val |= MINCORE_MODIFIED_OTHER;
5055 if (pte & pmap->pmap_bits[PG_A_IDX])
5056 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5059 * Referenced by someone
5061 else if (m && ((m->flags & PG_REFERENCED) ||
5062 pmap_ts_referenced(m))) {
5063 val |= MINCORE_REFERENCED_OTHER;
5064 vm_page_flag_set(m, PG_REFERENCED);
5068 lwkt_reltoken(&pmap->pm_token);
5074 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5075 * vmspace will be ref'd and the old one will be deref'd.
5077 * The vmspace for all lwps associated with the process will be adjusted
5078 * and cr3 will be reloaded if any lwp is the current lwp.
5080 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5083 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5085 struct vmspace *oldvm;
5088 oldvm = p->p_vmspace;
5089 if (oldvm != newvm) {
5091 sysref_get(&newvm->vm_sysref);
5092 p->p_vmspace = newvm;
5093 KKASSERT(p->p_nthreads == 1);
5094 lp = RB_ROOT(&p->p_lwp_tree);
5095 pmap_setlwpvm(lp, newvm);
5097 sysref_put(&oldvm->vm_sysref);
5102 * Set the vmspace for a LWP. The vmspace is almost universally set the
5103 * same as the process vmspace, but virtual kernels need to swap out contexts
5104 * on a per-lwp basis.
5106 * Caller does not necessarily hold any vmspace tokens. Caller must control
5107 * the lwp (typically be in the context of the lwp). We use a critical
5108 * section to protect against statclock and hardclock (statistics collection).
5111 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5113 struct vmspace *oldvm;
5116 oldvm = lp->lwp_vmspace;
5118 if (oldvm != newvm) {
5120 lp->lwp_vmspace = newvm;
5121 if (curthread->td_lwp == lp) {
5122 pmap = vmspace_pmap(newvm);
5123 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
5124 if (pmap->pm_active & CPUMASK_LOCK)
5125 pmap_interlock_wait(newvm);
5126 #if defined(SWTCH_OPTIM_STATS)
5129 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5130 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5131 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5132 curthread->td_pcb->pcb_cr3 = KPML4phys;
5134 panic("pmap_setlwpvm: unknown pmap type\n");
5136 load_cr3(curthread->td_pcb->pcb_cr3);
5137 pmap = vmspace_pmap(oldvm);
5138 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
5145 * Called when switching to a locked pmap, used to interlock against pmaps
5146 * undergoing modifications to prevent us from activating the MMU for the
5147 * target pmap until all such modifications have completed. We have to do
5148 * this because the thread making the modifications has already set up its
5149 * SMP synchronization mask.
5151 * This function cannot sleep!
5156 pmap_interlock_wait(struct vmspace *vm)
5158 struct pmap *pmap = &vm->vm_pmap;
5160 if (pmap->pm_active & CPUMASK_LOCK) {
5162 KKASSERT(curthread->td_critcount >= 2);
5163 DEBUG_PUSH_INFO("pmap_interlock_wait");
5164 while (pmap->pm_active & CPUMASK_LOCK) {
5166 lwkt_process_ipiq();
5174 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5177 if ((obj == NULL) || (size < NBPDR) ||
5178 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5182 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
5187 * Used by kmalloc/kfree, page already exists at va
5190 pmap_kvtom(vm_offset_t va)
5192 pt_entry_t *ptep = vtopte(va);
5194 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5195 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5199 * Initialize machine-specific shared page directory support. This
5200 * is executed when a VM object is created.
5203 pmap_object_init(vm_object_t object)
5205 object->md.pmap_rw = NULL;
5206 object->md.pmap_ro = NULL;
5210 * Clean up machine-specific shared page directory support. This
5211 * is executed when a VM object is destroyed.
5214 pmap_object_free(vm_object_t object)
5218 if ((pmap = object->md.pmap_rw) != NULL) {
5219 object->md.pmap_rw = NULL;
5220 pmap_remove_noinval(pmap,
5221 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5222 pmap->pm_active = 0;
5225 kfree(pmap, M_OBJPMAP);
5227 if ((pmap = object->md.pmap_ro) != NULL) {
5228 object->md.pmap_ro = NULL;
5229 pmap_remove_noinval(pmap,
5230 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5231 pmap->pm_active = 0;
5234 kfree(pmap, M_OBJPMAP);