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);
312 * Move the kernel virtual free pointer to the next
313 * 2MB. This is used to help improve performance
314 * by using a large (2MB) page for much of the kernel
315 * (.text, .data, .bss)
319 pmap_kmem_choose(vm_offset_t addr)
321 vm_offset_t newaddr = addr;
323 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
330 * Super fast pmap_pte routine best used when scanning the pv lists.
331 * This eliminates many course-grained invltlb calls. Note that many of
332 * the pv list scans are across different pmaps and it is very wasteful
333 * to do an entire invltlb when checking a single mapping.
335 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
339 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
341 return pmap_pte(pmap, va);
345 * Returns the pindex of a page table entry (representing a terminal page).
346 * There are NUPTE_TOTAL page table entries possible (a huge number)
348 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
349 * We want to properly translate negative KVAs.
353 pmap_pte_pindex(vm_offset_t va)
355 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
359 * Returns the pindex of a page table.
363 pmap_pt_pindex(vm_offset_t va)
365 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
369 * Returns the pindex of a page directory.
373 pmap_pd_pindex(vm_offset_t va)
375 return (NUPTE_TOTAL + NUPT_TOTAL +
376 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
381 pmap_pdp_pindex(vm_offset_t va)
383 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
384 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
389 pmap_pml4_pindex(void)
391 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
395 * Return various clipped indexes for a given VA
397 * Returns the index of a pte in a page table, representing a terminal
402 pmap_pte_index(vm_offset_t va)
404 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
408 * Returns the index of a pt in a page directory, representing a page
413 pmap_pt_index(vm_offset_t va)
415 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
419 * Returns the index of a pd in a page directory page, representing a page
424 pmap_pd_index(vm_offset_t va)
426 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
430 * Returns the index of a pdp in the pml4 table, representing a page
435 pmap_pdp_index(vm_offset_t va)
437 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
441 * Generic procedure to index a pte from a pt, pd, or pdp.
443 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
444 * a page table page index but is instead of PV lookup index.
448 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
452 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
453 return(&pte[pindex]);
457 * Return pointer to PDP slot in the PML4
461 pmap_pdp(pmap_t pmap, vm_offset_t va)
463 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
467 * Return pointer to PD slot in the PDP given a pointer to the PDP
471 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
475 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
476 return (&pd[pmap_pd_index(va)]);
480 * Return pointer to PD slot in the PDP.
484 pmap_pd(pmap_t pmap, vm_offset_t va)
488 pdp = pmap_pdp(pmap, va);
489 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
491 return (pmap_pdp_to_pd(*pdp, va));
495 * Return pointer to PT slot in the PD given a pointer to the PD
499 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
503 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
504 return (&pt[pmap_pt_index(va)]);
508 * Return pointer to PT slot in the PD
510 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
511 * so we cannot lookup the PD via the PDP. Instead we
512 * must look it up via the pmap.
516 pmap_pt(pmap_t pmap, vm_offset_t va)
520 vm_pindex_t pd_pindex;
522 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
523 pd_pindex = pmap_pd_pindex(va);
524 spin_lock(&pmap->pm_spin);
525 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
526 spin_unlock(&pmap->pm_spin);
527 if (pv == NULL || pv->pv_m == NULL)
529 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
531 pd = pmap_pd(pmap, va);
532 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
534 return (pmap_pd_to_pt(*pd, va));
539 * Return pointer to PTE slot in the PT given a pointer to the PT
543 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
547 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
548 return (&pte[pmap_pte_index(va)]);
552 * Return pointer to PTE slot in the PT
556 pmap_pte(pmap_t pmap, vm_offset_t va)
560 pt = pmap_pt(pmap, va);
561 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
563 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
564 return ((pt_entry_t *)pt);
565 return (pmap_pt_to_pte(*pt, va));
569 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
570 * the PT layer. This will speed up core pmap operations considerably.
572 * NOTE: Can be called with the pmap spin lock held shared. pm_pvhint
577 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
579 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
580 pv->pv_pmap->pm_pvhint = pv;
585 * KVM - return address of PT slot in PD
589 vtopt(vm_offset_t va)
591 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
592 NPML4EPGSHIFT)) - 1);
594 return (PDmap + ((va >> PDRSHIFT) & mask));
598 * KVM - return address of PTE slot in PT
602 vtopte(vm_offset_t va)
604 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
605 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
607 return (PTmap + ((va >> PAGE_SHIFT) & mask));
611 allocpages(vm_paddr_t *firstaddr, long n)
616 bzero((void *)ret, n * PAGE_SIZE);
617 *firstaddr += n * PAGE_SIZE;
623 create_pagetables(vm_paddr_t *firstaddr)
625 long i; /* must be 64 bits */
631 * We are running (mostly) V=P at this point
633 * Calculate NKPT - number of kernel page tables. We have to
634 * accomodoate prealloction of the vm_page_array, dump bitmap,
635 * MSGBUF_SIZE, and other stuff. Be generous.
637 * Maxmem is in pages.
639 * ndmpdp is the number of 1GB pages we wish to map.
641 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
642 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
644 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
647 * Starting at the beginning of kvm (not KERNBASE).
649 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
650 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
651 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
652 ndmpdp) + 511) / 512;
656 * Starting at KERNBASE - map 2G worth of page table pages.
657 * KERNBASE is offset -2G from the end of kvm.
659 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
664 KPTbase = allocpages(firstaddr, nkpt_base);
665 KPTphys = allocpages(firstaddr, nkpt_phys);
666 KPML4phys = allocpages(firstaddr, 1);
667 KPDPphys = allocpages(firstaddr, NKPML4E);
668 KPDphys = allocpages(firstaddr, NKPDPE);
671 * Calculate the page directory base for KERNBASE,
672 * that is where we start populating the page table pages.
673 * Basically this is the end - 2.
675 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
677 DMPDPphys = allocpages(firstaddr, NDMPML4E);
678 if ((amd_feature & AMDID_PAGE1GB) == 0)
679 DMPDphys = allocpages(firstaddr, ndmpdp);
680 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
683 * Fill in the underlying page table pages for the area around
684 * KERNBASE. This remaps low physical memory to KERNBASE.
686 * Read-only from zero to physfree
687 * XXX not fully used, underneath 2M pages
689 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
690 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
691 ((pt_entry_t *)KPTbase)[i] |=
692 pmap_bits_default[PG_RW_IDX] |
693 pmap_bits_default[PG_V_IDX] |
694 pmap_bits_default[PG_G_IDX];
698 * Now map the initial kernel page tables. One block of page
699 * tables is placed at the beginning of kernel virtual memory,
700 * and another block is placed at KERNBASE to map the kernel binary,
701 * data, bss, and initial pre-allocations.
703 for (i = 0; i < nkpt_base; i++) {
704 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
705 ((pd_entry_t *)KPDbase)[i] |=
706 pmap_bits_default[PG_RW_IDX] |
707 pmap_bits_default[PG_V_IDX];
709 for (i = 0; i < nkpt_phys; i++) {
710 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
711 ((pd_entry_t *)KPDphys)[i] |=
712 pmap_bits_default[PG_RW_IDX] |
713 pmap_bits_default[PG_V_IDX];
717 * Map from zero to end of allocations using 2M pages as an
718 * optimization. This will bypass some of the KPTBase pages
719 * above in the KERNBASE area.
721 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
722 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
723 ((pd_entry_t *)KPDbase)[i] |=
724 pmap_bits_default[PG_RW_IDX] |
725 pmap_bits_default[PG_V_IDX] |
726 pmap_bits_default[PG_PS_IDX] |
727 pmap_bits_default[PG_G_IDX];
731 * And connect up the PD to the PDP. The kernel pmap is expected
732 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
734 for (i = 0; i < NKPDPE; i++) {
735 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
736 KPDphys + (i << PAGE_SHIFT);
737 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
738 pmap_bits_default[PG_RW_IDX] |
739 pmap_bits_default[PG_V_IDX] |
740 pmap_bits_default[PG_U_IDX];
744 * Now set up the direct map space using either 2MB or 1GB pages
745 * Preset PG_M and PG_A because demotion expects it.
747 * When filling in entries in the PD pages make sure any excess
748 * entries are set to zero as we allocated enough PD pages
750 if ((amd_feature & AMDID_PAGE1GB) == 0) {
751 for (i = 0; i < NPDEPG * ndmpdp; i++) {
752 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
753 ((pd_entry_t *)DMPDphys)[i] |=
754 pmap_bits_default[PG_RW_IDX] |
755 pmap_bits_default[PG_V_IDX] |
756 pmap_bits_default[PG_PS_IDX] |
757 pmap_bits_default[PG_G_IDX] |
758 pmap_bits_default[PG_M_IDX] |
759 pmap_bits_default[PG_A_IDX];
763 * And the direct map space's PDP
765 for (i = 0; i < ndmpdp; i++) {
766 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
768 ((pdp_entry_t *)DMPDPphys)[i] |=
769 pmap_bits_default[PG_RW_IDX] |
770 pmap_bits_default[PG_V_IDX] |
771 pmap_bits_default[PG_U_IDX];
774 for (i = 0; i < ndmpdp; i++) {
775 ((pdp_entry_t *)DMPDPphys)[i] =
776 (vm_paddr_t)i << PDPSHIFT;
777 ((pdp_entry_t *)DMPDPphys)[i] |=
778 pmap_bits_default[PG_RW_IDX] |
779 pmap_bits_default[PG_V_IDX] |
780 pmap_bits_default[PG_PS_IDX] |
781 pmap_bits_default[PG_G_IDX] |
782 pmap_bits_default[PG_M_IDX] |
783 pmap_bits_default[PG_A_IDX];
787 /* And recursively map PML4 to itself in order to get PTmap */
788 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
789 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
790 pmap_bits_default[PG_RW_IDX] |
791 pmap_bits_default[PG_V_IDX] |
792 pmap_bits_default[PG_U_IDX];
795 * Connect the Direct Map slots up to the PML4
797 for (j = 0; j < NDMPML4E; ++j) {
798 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
799 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
800 pmap_bits_default[PG_RW_IDX] |
801 pmap_bits_default[PG_V_IDX] |
802 pmap_bits_default[PG_U_IDX];
806 * Connect the KVA slot up to the PML4
808 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
809 ((pdp_entry_t *)KPML4phys)[KPML4I] |=
810 pmap_bits_default[PG_RW_IDX] |
811 pmap_bits_default[PG_V_IDX] |
812 pmap_bits_default[PG_U_IDX];
816 * Bootstrap the system enough to run with virtual memory.
818 * On the i386 this is called after mapping has already been enabled
819 * and just syncs the pmap module with what has already been done.
820 * [We can't call it easily with mapping off since the kernel is not
821 * mapped with PA == VA, hence we would have to relocate every address
822 * from the linked base (virtual) address "KERNBASE" to the actual
823 * (physical) address starting relative to 0]
826 pmap_bootstrap(vm_paddr_t *firstaddr)
831 KvaStart = VM_MIN_KERNEL_ADDRESS;
832 KvaEnd = VM_MAX_KERNEL_ADDRESS;
833 KvaSize = KvaEnd - KvaStart;
835 avail_start = *firstaddr;
838 * Create an initial set of page tables to run the kernel in.
840 create_pagetables(firstaddr);
842 virtual2_start = KvaStart;
843 virtual2_end = PTOV_OFFSET;
845 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
846 virtual_start = pmap_kmem_choose(virtual_start);
848 virtual_end = VM_MAX_KERNEL_ADDRESS;
850 /* XXX do %cr0 as well */
851 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
855 * Initialize protection array.
857 i386_protection_init();
860 * The kernel's pmap is statically allocated so we don't have to use
861 * pmap_create, which is unlikely to work correctly at this part of
862 * the boot sequence (XXX and which no longer exists).
864 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
865 kernel_pmap.pm_count = 1;
866 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
867 RB_INIT(&kernel_pmap.pm_pvroot);
868 spin_init(&kernel_pmap.pm_spin);
869 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
872 * Reserve some special page table entries/VA space for temporary
875 #define SYSMAP(c, p, v, n) \
876 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
882 * CMAP1/CMAP2 are used for zeroing and copying pages.
884 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
889 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
892 * ptvmmap is used for reading arbitrary physical pages via
895 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
898 * msgbufp is used to map the system message buffer.
899 * XXX msgbufmap is not used.
901 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
902 atop(round_page(MSGBUF_SIZE)))
909 * PG_G is terribly broken on SMP because we IPI invltlb's in some
910 * cases rather then invl1pg. Actually, I don't even know why it
911 * works under UP because self-referential page table mappings
916 * Initialize the 4MB page size flag
920 * The 4MB page version of the initial
921 * kernel page mapping.
925 #if !defined(DISABLE_PSE)
926 if (cpu_feature & CPUID_PSE) {
929 * Note that we have enabled PSE mode
931 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
932 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
933 ptditmp &= ~(NBPDR - 1);
934 ptditmp |= pmap_bits_default[PG_V_IDX] |
935 pmap_bits_default[PG_RW_IDX] |
936 pmap_bits_default[PG_PS_IDX] |
937 pmap_bits_default[PG_U_IDX];
944 /* Initialize the PAT MSR */
947 pmap_pinit_defaults(&kernel_pmap);
960 * Default values mapping PATi,PCD,PWT bits at system reset.
961 * The default values effectively ignore the PATi bit by
962 * repeating the encodings for 0-3 in 4-7, and map the PCD
963 * and PWT bit combinations to the expected PAT types.
965 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
966 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
967 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
968 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
969 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
970 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
971 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
972 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
973 pat_pte_index[PAT_WRITE_BACK] = 0;
974 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
975 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
976 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
977 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
978 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
980 if (cpu_feature & CPUID_PAT) {
982 * If we support the PAT then set-up entries for
983 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
986 pat_msr = (pat_msr & ~PAT_MASK(4)) |
987 PAT_VALUE(4, PAT_WRITE_PROTECTED);
988 pat_msr = (pat_msr & ~PAT_MASK(5)) |
989 PAT_VALUE(5, PAT_WRITE_COMBINING);
990 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
991 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
994 * Then enable the PAT
999 load_cr4(cr4 & ~CR4_PGE);
1001 /* Disable caches (CD = 1, NW = 0). */
1003 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1005 /* Flushes caches and TLBs. */
1009 /* Update PAT and index table. */
1010 wrmsr(MSR_PAT, pat_msr);
1012 /* Flush caches and TLBs again. */
1016 /* Restore caches and PGE. */
1024 * Set 4mb pdir for mp startup
1029 if (cpu_feature & CPUID_PSE) {
1030 load_cr4(rcr4() | CR4_PSE);
1031 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
1038 * Initialize the pmap module.
1039 * Called by vm_init, to initialize any structures that the pmap
1040 * system needs to map virtual memory.
1041 * pmap_init has been enhanced to support in a fairly consistant
1042 * way, discontiguous physical memory.
1051 * Allocate memory for random pmap data structures. Includes the
1055 for (i = 0; i < vm_page_array_size; i++) {
1058 m = &vm_page_array[i];
1059 TAILQ_INIT(&m->md.pv_list);
1063 * init the pv free list
1065 initial_pvs = vm_page_array_size;
1066 if (initial_pvs < MINPV)
1067 initial_pvs = MINPV;
1068 pvzone = &pvzone_store;
1069 pvinit = (void *)kmem_alloc(&kernel_map,
1070 initial_pvs * sizeof (struct pv_entry));
1071 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1072 pvinit, initial_pvs);
1075 * Now it is safe to enable pv_table recording.
1077 pmap_initialized = TRUE;
1081 * Initialize the address space (zone) for the pv_entries. Set a
1082 * high water mark so that the system can recover from excessive
1083 * numbers of pv entries.
1088 int shpgperproc = PMAP_SHPGPERPROC;
1091 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1092 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1093 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1094 pv_entry_high_water = 9 * (pv_entry_max / 10);
1097 * Subtract out pages already installed in the zone (hack)
1099 entry_max = pv_entry_max - vm_page_array_size;
1103 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
1107 * Typically used to initialize a fictitious page by vm/device_pager.c
1110 pmap_page_init(struct vm_page *m)
1113 TAILQ_INIT(&m->md.pv_list);
1116 /***************************************************
1117 * Low level helper routines.....
1118 ***************************************************/
1121 * this routine defines the region(s) of memory that should
1122 * not be tested for the modified bit.
1126 pmap_track_modified(vm_pindex_t pindex)
1128 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1129 if ((va < clean_sva) || (va >= clean_eva))
1136 * Extract the physical page address associated with the map/VA pair.
1137 * The page must be wired for this to work reliably.
1139 * XXX for the moment we're using pv_find() instead of pv_get(), as
1140 * callers might be expecting non-blocking operation.
1143 pmap_extract(pmap_t pmap, vm_offset_t va)
1150 if (va >= VM_MAX_USER_ADDRESS) {
1152 * Kernel page directories might be direct-mapped and
1153 * there is typically no PV tracking of pte's
1157 pt = pmap_pt(pmap, va);
1158 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1159 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1160 rtval = *pt & PG_PS_FRAME;
1161 rtval |= va & PDRMASK;
1163 ptep = pmap_pt_to_pte(*pt, va);
1164 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1165 rtval = *ptep & PG_FRAME;
1166 rtval |= va & PAGE_MASK;
1172 * User pages currently do not direct-map the page directory
1173 * and some pages might not used managed PVs. But all PT's
1176 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1178 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1179 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1180 rtval = *ptep & PG_FRAME;
1181 rtval |= va & PAGE_MASK;
1190 * Similar to extract but checks protections, SMP-friendly short-cut for
1191 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1192 * fall-through to the real fault code.
1194 * The returned page, if not NULL, is held (and not busied).
1197 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1199 if (pmap && va < VM_MAX_USER_ADDRESS) {
1207 req = pmap->pmap_bits[PG_V_IDX] |
1208 pmap->pmap_bits[PG_U_IDX];
1209 if (prot & VM_PROT_WRITE)
1210 req |= pmap->pmap_bits[PG_RW_IDX];
1212 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1215 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1216 if ((*ptep & req) != req) {
1220 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1221 if (pte_pv && error == 0) {
1224 if (prot & VM_PROT_WRITE)
1227 } else if (pte_pv) {
1241 * Extract the physical page address associated kernel virtual address.
1244 pmap_kextract(vm_offset_t va)
1246 pd_entry_t pt; /* pt entry in pd */
1249 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1250 pa = DMAP_TO_PHYS(va);
1253 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1254 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1257 * Beware of a concurrent promotion that changes the
1258 * PDE at this point! For example, vtopte() must not
1259 * be used to access the PTE because it would use the
1260 * new PDE. It is, however, safe to use the old PDE
1261 * because the page table page is preserved by the
1264 pa = *pmap_pt_to_pte(pt, va);
1265 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1271 /***************************************************
1272 * Low level mapping routines.....
1273 ***************************************************/
1276 * Routine: pmap_kenter
1278 * Add a wired page to the KVA
1279 * NOTE! note that in order for the mapping to take effect -- you
1280 * should do an invltlb after doing the pmap_kenter().
1283 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1287 pmap_inval_info info;
1289 pmap_inval_init(&info); /* XXX remove */
1291 kernel_pmap.pmap_bits[PG_RW_IDX] |
1292 kernel_pmap.pmap_bits[PG_V_IDX];
1295 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
1297 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1298 pmap_inval_done(&info); /* XXX remove */
1302 * Routine: pmap_kenter_quick
1304 * Similar to pmap_kenter(), except we only invalidate the
1305 * mapping on the current CPU.
1308 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1314 kernel_pmap.pmap_bits[PG_RW_IDX] |
1315 kernel_pmap.pmap_bits[PG_V_IDX];
1319 cpu_invlpg((void *)va);
1323 pmap_kenter_sync(vm_offset_t va)
1325 pmap_inval_info info;
1327 pmap_inval_init(&info);
1328 pmap_inval_interlock(&info, &kernel_pmap, va);
1329 pmap_inval_deinterlock(&info, &kernel_pmap);
1330 pmap_inval_done(&info);
1334 pmap_kenter_sync_quick(vm_offset_t va)
1336 cpu_invlpg((void *)va);
1340 * remove a page from the kernel pagetables
1343 pmap_kremove(vm_offset_t va)
1346 pmap_inval_info info;
1348 pmap_inval_init(&info);
1350 pmap_inval_interlock(&info, &kernel_pmap, va);
1351 (void)pte_load_clear(pte);
1352 pmap_inval_deinterlock(&info, &kernel_pmap);
1353 pmap_inval_done(&info);
1357 pmap_kremove_quick(vm_offset_t va)
1361 (void)pte_load_clear(pte);
1362 cpu_invlpg((void *)va);
1366 * XXX these need to be recoded. They are not used in any critical path.
1369 pmap_kmodify_rw(vm_offset_t va)
1371 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1372 cpu_invlpg((void *)va);
1377 pmap_kmodify_nc(vm_offset_t va)
1379 atomic_set_long(vtopte(va), PG_N);
1380 cpu_invlpg((void *)va);
1385 * Used to map a range of physical addresses into kernel virtual
1386 * address space during the low level boot, typically to map the
1387 * dump bitmap, message buffer, and vm_page_array.
1389 * These mappings are typically made at some pointer after the end of the
1392 * We could return PHYS_TO_DMAP(start) here and not allocate any
1393 * via (*virtp), but then kmem from userland and kernel dumps won't
1394 * have access to the related pointers.
1397 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1400 vm_offset_t va_start;
1402 /*return PHYS_TO_DMAP(start);*/
1407 while (start < end) {
1408 pmap_kenter_quick(va, start);
1416 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1419 * Remove the specified set of pages from the data and instruction caches.
1421 * In contrast to pmap_invalidate_cache_range(), this function does not
1422 * rely on the CPU's self-snoop feature, because it is intended for use
1423 * when moving pages into a different cache domain.
1426 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1428 vm_offset_t daddr, eva;
1431 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1432 (cpu_feature & CPUID_CLFSH) == 0)
1436 for (i = 0; i < count; i++) {
1437 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1438 eva = daddr + PAGE_SIZE;
1439 for (; daddr < eva; daddr += cpu_clflush_line_size)
1447 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1449 KASSERT((sva & PAGE_MASK) == 0,
1450 ("pmap_invalidate_cache_range: sva not page-aligned"));
1451 KASSERT((eva & PAGE_MASK) == 0,
1452 ("pmap_invalidate_cache_range: eva not page-aligned"));
1454 if (cpu_feature & CPUID_SS) {
1455 ; /* If "Self Snoop" is supported, do nothing. */
1457 /* Globally invalidate caches */
1458 cpu_wbinvd_on_all_cpus();
1462 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1464 smp_invlpg_range(pmap->pm_active, sva, eva);
1468 * Add a list of wired pages to the kva
1469 * this routine is only used for temporary
1470 * kernel mappings that do not need to have
1471 * page modification or references recorded.
1472 * Note that old mappings are simply written
1473 * over. The page *must* be wired.
1476 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1480 end_va = va + count * PAGE_SIZE;
1482 while (va < end_va) {
1486 *pte = VM_PAGE_TO_PHYS(*m) |
1487 kernel_pmap.pmap_bits[PG_RW_IDX] |
1488 kernel_pmap.pmap_bits[PG_V_IDX] |
1489 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1491 cpu_invlpg((void *)va);
1499 * This routine jerks page mappings from the
1500 * kernel -- it is meant only for temporary mappings.
1502 * MPSAFE, INTERRUPT SAFE (cluster callback)
1505 pmap_qremove(vm_offset_t va, int count)
1509 end_va = va + count * PAGE_SIZE;
1511 while (va < end_va) {
1515 (void)pte_load_clear(pte);
1516 cpu_invlpg((void *)va);
1523 * Create a new thread and optionally associate it with a (new) process.
1524 * NOTE! the new thread's cpu may not equal the current cpu.
1527 pmap_init_thread(thread_t td)
1529 /* enforce pcb placement & alignment */
1530 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1531 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1532 td->td_savefpu = &td->td_pcb->pcb_save;
1533 td->td_sp = (char *)td->td_pcb; /* no -16 */
1537 * This routine directly affects the fork perf for a process.
1540 pmap_init_proc(struct proc *p)
1545 pmap_pinit_defaults(struct pmap *pmap) {
1546 bcopy(pmap_bits_default, pmap->pmap_bits, sizeof(pmap_bits_default));
1547 bcopy(protection_codes, pmap->protection_codes, sizeof(protection_codes));
1548 bcopy(pat_pte_index, pmap->pmap_cache_bits, sizeof(pat_pte_index));
1549 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1550 pmap->copyinstr = std_copyinstr;
1551 pmap->copyin = std_copyin;
1552 pmap->copyout = std_copyout;
1553 pmap->fubyte = std_fubyte;
1554 pmap->subyte = std_subyte;
1555 pmap->fuword = std_fuword;
1556 pmap->suword = std_suword;
1557 pmap->suword32 = std_suword32;
1560 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1561 * it, and IdlePTD, represents the template used to update all other pmaps.
1563 * On architectures where the kernel pmap is not integrated into the user
1564 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1565 * kernel_pmap should be used to directly access the kernel_pmap.
1568 pmap_pinit0(struct pmap *pmap)
1570 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1572 pmap->pm_active = 0;
1573 pmap->pm_pvhint = NULL;
1574 RB_INIT(&pmap->pm_pvroot);
1575 spin_init(&pmap->pm_spin);
1576 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1577 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1578 pmap_pinit_defaults(pmap);
1582 * Initialize a preallocated and zeroed pmap structure,
1583 * such as one in a vmspace structure.
1586 pmap_pinit_simple(struct pmap *pmap)
1589 * Misc initialization
1592 pmap->pm_active = 0;
1593 pmap->pm_pvhint = NULL;
1594 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1596 pmap_pinit_defaults(pmap);
1599 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1602 if (pmap->pm_pmlpv == NULL) {
1603 RB_INIT(&pmap->pm_pvroot);
1604 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1605 spin_init(&pmap->pm_spin);
1606 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1611 pmap_pinit(struct pmap *pmap)
1616 if (pmap->pm_pmlpv) {
1617 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1622 pmap_pinit_simple(pmap);
1623 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1626 * No need to allocate page table space yet but we do need a valid
1627 * page directory table.
1629 if (pmap->pm_pml4 == NULL) {
1631 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1635 * Allocate the page directory page, which wires it even though
1636 * it isn't being entered into some higher level page table (it
1637 * being the highest level). If one is already cached we don't
1638 * have to do anything.
1640 if ((pv = pmap->pm_pmlpv) == NULL) {
1641 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1642 pmap->pm_pmlpv = pv;
1643 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1644 VM_PAGE_TO_PHYS(pv->pv_m));
1648 * Install DMAP and KMAP.
1650 for (j = 0; j < NDMPML4E; ++j) {
1651 pmap->pm_pml4[DMPML4I + j] =
1652 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1653 pmap->pmap_bits[PG_RW_IDX] |
1654 pmap->pmap_bits[PG_V_IDX] |
1655 pmap->pmap_bits[PG_U_IDX];
1657 pmap->pm_pml4[KPML4I] = KPDPphys |
1658 pmap->pmap_bits[PG_RW_IDX] |
1659 pmap->pmap_bits[PG_V_IDX] |
1660 pmap->pmap_bits[PG_U_IDX];
1663 * install self-referential address mapping entry
1665 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1666 pmap->pmap_bits[PG_V_IDX] |
1667 pmap->pmap_bits[PG_RW_IDX] |
1668 pmap->pmap_bits[PG_A_IDX] |
1669 pmap->pmap_bits[PG_M_IDX];
1671 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1672 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1674 KKASSERT(pmap->pm_pml4[255] == 0);
1675 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1676 KKASSERT(pv->pv_entry.rbe_left == NULL);
1677 KKASSERT(pv->pv_entry.rbe_right == NULL);
1681 * Clean up a pmap structure so it can be physically freed. This routine
1682 * is called by the vmspace dtor function. A great deal of pmap data is
1683 * left passively mapped to improve vmspace management so we have a bit
1684 * of cleanup work to do here.
1687 pmap_puninit(pmap_t pmap)
1692 KKASSERT(pmap->pm_active == 0);
1693 if ((pv = pmap->pm_pmlpv) != NULL) {
1694 if (pv_hold_try(pv) == 0)
1696 p = pmap_remove_pv_page(pv);
1698 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1699 vm_page_busy_wait(p, FALSE, "pgpun");
1700 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1701 vm_page_unwire(p, 0);
1702 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1705 * XXX eventually clean out PML4 static entries and
1706 * use vm_page_free_zero()
1709 pmap->pm_pmlpv = NULL;
1711 if (pmap->pm_pml4) {
1712 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1713 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1714 pmap->pm_pml4 = NULL;
1716 KKASSERT(pmap->pm_stats.resident_count == 0);
1717 KKASSERT(pmap->pm_stats.wired_count == 0);
1721 * Wire in kernel global address entries. To avoid a race condition
1722 * between pmap initialization and pmap_growkernel, this procedure
1723 * adds the pmap to the master list (which growkernel scans to update),
1724 * then copies the template.
1727 pmap_pinit2(struct pmap *pmap)
1729 spin_lock(&pmap_spin);
1730 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1731 spin_unlock(&pmap_spin);
1735 * This routine is called when various levels in the page table need to
1736 * be populated. This routine cannot fail.
1738 * This function returns two locked pv_entry's, one representing the
1739 * requested pv and one representing the requested pv's parent pv. If
1740 * the pv did not previously exist it will be mapped into its parent
1741 * and wired, otherwise no additional wire count will be added.
1745 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1750 vm_pindex_t pt_pindex;
1756 * If the pv already exists and we aren't being asked for the
1757 * parent page table page we can just return it. A locked+held pv
1761 pv = pv_alloc(pmap, ptepindex, &isnew);
1762 if (isnew == 0 && pvpp == NULL)
1766 * This is a new PV, we have to resolve its parent page table and
1767 * add an additional wiring to the page if necessary.
1771 * Special case terminal PVs. These are not page table pages so
1772 * no vm_page is allocated (the caller supplied the vm_page). If
1773 * pvpp is non-NULL we are being asked to also removed the pt_pv
1776 * Note that pt_pv's are only returned for user VAs. We assert that
1777 * a pt_pv is not being requested for kernel VAs.
1779 if (ptepindex < pmap_pt_pindex(0)) {
1780 if (ptepindex >= NUPTE_USER)
1781 KKASSERT(pvpp == NULL);
1783 KKASSERT(pvpp != NULL);
1785 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1786 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1788 vm_page_wire_quick(pvp->pv_m);
1797 * Non-terminal PVs allocate a VM page to represent the page table,
1798 * so we have to resolve pvp and calculate ptepindex for the pvp
1799 * and then for the page table entry index in the pvp for
1802 if (ptepindex < pmap_pd_pindex(0)) {
1804 * pv is PT, pvp is PD
1806 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1807 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1808 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1815 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1816 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1818 } else if (ptepindex < pmap_pdp_pindex(0)) {
1820 * pv is PD, pvp is PDP
1822 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1825 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1826 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1828 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1829 KKASSERT(pvpp == NULL);
1832 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1840 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1841 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1842 } else if (ptepindex < pmap_pml4_pindex()) {
1844 * pv is PDP, pvp is the root pml4 table
1846 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1853 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1854 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1857 * pv represents the top-level PML4, there is no parent.
1865 * This code is only reached if isnew is TRUE and this is not a
1866 * terminal PV. We need to allocate a vm_page for the page table
1867 * at this level and enter it into the parent page table.
1869 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1872 m = vm_page_alloc(NULL, pv->pv_pindex,
1873 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1874 VM_ALLOC_INTERRUPT);
1879 vm_page_spin_lock(m);
1880 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1882 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1883 vm_page_spin_unlock(m);
1884 vm_page_unmanage(m); /* m must be spinunlocked */
1886 if ((m->flags & PG_ZERO) == 0) {
1887 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1891 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1894 m->valid = VM_PAGE_BITS_ALL;
1895 vm_page_flag_clear(m, PG_ZERO);
1896 vm_page_wire(m); /* wire for mapping in parent */
1899 * Wire the page into pvp, bump the wire-count for pvp's page table
1900 * page. Bump the resident_count for the pmap. There is no pvp
1901 * for the top level, address the pm_pml4[] array directly.
1903 * If the caller wants the parent we return it, otherwise
1904 * we just put it away.
1906 * No interlock is needed for pte 0 -> non-zero.
1908 * In the situation where *ptep is valid we might have an unmanaged
1909 * page table page shared from another page table which we need to
1910 * unshare before installing our private page table page.
1913 ptep = pv_pte_lookup(pvp, ptepindex);
1914 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1916 pmap_inval_info info;
1919 panic("pmap_allocpte: unexpected pte %p/%d",
1920 pvp, (int)ptepindex);
1922 pmap_inval_init(&info);
1923 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1924 pte = pte_load_clear(ptep);
1925 pmap_inval_deinterlock(&info, pmap);
1926 pmap_inval_done(&info);
1927 if (vm_page_unwire_quick(
1928 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
1929 panic("pmap_allocpte: shared pgtable "
1930 "pg bad wirecount");
1932 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1934 vm_page_wire_quick(pvp->pv_m);
1936 *ptep = VM_PAGE_TO_PHYS(m) |
1937 (pmap->pmap_bits[PG_U_IDX] |
1938 pmap->pmap_bits[PG_RW_IDX] |
1939 pmap->pmap_bits[PG_V_IDX] |
1940 pmap->pmap_bits[PG_A_IDX] |
1941 pmap->pmap_bits[PG_M_IDX]);
1953 * This version of pmap_allocpte() checks for possible segment optimizations
1954 * that would allow page-table sharing. It can be called for terminal
1955 * page or page table page ptepindex's.
1957 * The function is called with page table page ptepindex's for fictitious
1958 * and unmanaged terminal pages. That is, we don't want to allocate a
1959 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
1962 * This function can return a pv and *pvpp associated with the passed in pmap
1963 * OR a pv and *pvpp associated with the shared pmap. In the latter case
1964 * an unmanaged page table page will be entered into the pass in pmap.
1968 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
1969 vm_map_entry_t entry, vm_offset_t va)
1971 struct pmap_inval_info info;
1976 pv_entry_t pte_pv; /* in original or shared pmap */
1977 pv_entry_t pt_pv; /* in original or shared pmap */
1978 pv_entry_t proc_pd_pv; /* in original pmap */
1979 pv_entry_t proc_pt_pv; /* in original pmap */
1980 pv_entry_t xpv; /* PT in shared pmap */
1981 pd_entry_t *pt; /* PT entry in PD of original pmap */
1982 pd_entry_t opte; /* contents of *pt */
1983 pd_entry_t npte; /* contents of *pt */
1988 * Basic tests, require a non-NULL vm_map_entry, require proper
1989 * alignment and type for the vm_map_entry, require that the
1990 * underlying object already be allocated.
1992 * We currently allow any type of object to use this optimization.
1993 * The object itself does NOT have to be sized to a multiple of the
1994 * segment size, but the memory mapping does.
1996 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
1997 * won't work as expected.
1999 if (entry == NULL ||
2000 pmap_mmu_optimize == 0 || /* not enabled */
2001 ptepindex >= pmap_pd_pindex(0) || /* not terminal */
2002 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2003 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2004 entry->object.vm_object == NULL || /* needs VM object */
2005 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2006 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2007 (entry->offset & SEG_MASK) || /* must be aligned */
2008 (entry->start & SEG_MASK)) {
2009 return(pmap_allocpte(pmap, ptepindex, pvpp));
2013 * Make sure the full segment can be represented.
2015 b = va & ~(vm_offset_t)SEG_MASK;
2016 if (b < entry->start && b + SEG_SIZE > entry->end)
2017 return(pmap_allocpte(pmap, ptepindex, pvpp));
2020 * If the full segment can be represented dive the VM object's
2021 * shared pmap, allocating as required.
2023 object = entry->object.vm_object;
2025 if (entry->protection & VM_PROT_WRITE)
2026 obpmapp = &object->md.pmap_rw;
2028 obpmapp = &object->md.pmap_ro;
2031 * We allocate what appears to be a normal pmap but because portions
2032 * of this pmap are shared with other unrelated pmaps we have to
2033 * set pm_active to point to all cpus.
2035 * XXX Currently using pmap_spin to interlock the update, can't use
2036 * vm_object_hold/drop because the token might already be held
2037 * shared OR exclusive and we don't know.
2039 while ((obpmap = *obpmapp) == NULL) {
2040 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2041 pmap_pinit_simple(obpmap);
2042 pmap_pinit2(obpmap);
2043 spin_lock(&pmap_spin);
2044 if (*obpmapp != NULL) {
2048 spin_unlock(&pmap_spin);
2049 pmap_release(obpmap);
2050 pmap_puninit(obpmap);
2051 kfree(obpmap, M_OBJPMAP);
2053 obpmap->pm_active = smp_active_mask;
2055 spin_unlock(&pmap_spin);
2060 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2061 * pte/pt using the shared pmap from the object but also adjust
2062 * the process pmap's page table page as a side effect.
2066 * Resolve the terminal PTE and PT in the shared pmap. This is what
2067 * we will return. This is true if ptepindex represents a terminal
2068 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2072 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2073 if (ptepindex >= pmap_pt_pindex(0))
2079 * Resolve the PD in the process pmap so we can properly share the
2080 * page table page. Lock order is bottom-up (leaf first)!
2082 * NOTE: proc_pt_pv can be NULL.
2084 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2085 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2088 * xpv is the page table page pv from the shared object
2089 * (for convenience).
2091 * Calculate the pte value for the PT to load into the process PD.
2092 * If we have to change it we must properly dispose of the previous
2095 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2096 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2097 (pmap->pmap_bits[PG_U_IDX] |
2098 pmap->pmap_bits[PG_RW_IDX] |
2099 pmap->pmap_bits[PG_V_IDX] |
2100 pmap->pmap_bits[PG_A_IDX] |
2101 pmap->pmap_bits[PG_M_IDX]);
2104 * Dispose of previous page table page if it was local to the
2105 * process pmap. If the old pt is not empty we cannot dispose of it
2106 * until we clean it out. This case should not arise very often so
2107 * it is not optimized.
2110 if (proc_pt_pv->pv_m->wire_count != 1) {
2116 va & ~(vm_offset_t)SEG_MASK,
2117 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2120 pmap_release_pv(proc_pt_pv, proc_pd_pv);
2123 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2127 * Handle remaining cases.
2131 vm_page_wire_quick(xpv->pv_m);
2132 vm_page_wire_quick(proc_pd_pv->pv_m);
2133 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2134 } else if (*pt != npte) {
2135 pmap_inval_init(&info);
2136 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
2138 opte = pte_load_clear(pt);
2139 KKASSERT(opte && opte != npte);
2142 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
2145 * Clean up opte, bump the wire_count for the process
2146 * PD page representing the new entry if it was
2149 * If the entry was not previously empty and we have
2150 * a PT in the proc pmap then opte must match that
2151 * pt. The proc pt must be retired (this is done
2152 * later on in this procedure).
2154 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2157 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2158 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2159 if (vm_page_unwire_quick(m)) {
2160 panic("pmap_allocpte_seg: "
2161 "bad wire count %p",
2165 pmap_inval_deinterlock(&info, pmap);
2166 pmap_inval_done(&info);
2170 * The existing process page table was replaced and must be destroyed
2184 * Release any resources held by the given physical map.
2186 * Called when a pmap initialized by pmap_pinit is being released. Should
2187 * only be called if the map contains no valid mappings.
2189 * Caller must hold pmap->pm_token
2191 struct pmap_release_info {
2196 static int pmap_release_callback(pv_entry_t pv, void *data);
2199 pmap_release(struct pmap *pmap)
2201 struct pmap_release_info info;
2203 KASSERT(pmap->pm_active == 0,
2204 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
2206 spin_lock(&pmap_spin);
2207 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2208 spin_unlock(&pmap_spin);
2211 * Pull pv's off the RB tree in order from low to high and release
2217 spin_lock(&pmap->pm_spin);
2218 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2219 pmap_release_callback, &info);
2220 spin_unlock(&pmap->pm_spin);
2221 } while (info.retry);
2225 * One resident page (the pml4 page) should remain.
2226 * No wired pages should remain.
2228 KKASSERT(pmap->pm_stats.resident_count ==
2229 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2231 KKASSERT(pmap->pm_stats.wired_count == 0);
2235 pmap_release_callback(pv_entry_t pv, void *data)
2237 struct pmap_release_info *info = data;
2238 pmap_t pmap = info->pmap;
2241 if (pv_hold_try(pv)) {
2242 spin_unlock(&pmap->pm_spin);
2244 spin_unlock(&pmap->pm_spin);
2247 if (pv->pv_pmap != pmap) {
2249 spin_lock(&pmap->pm_spin);
2253 r = pmap_release_pv(pv, NULL);
2254 spin_lock(&pmap->pm_spin);
2259 * Called with held (i.e. also locked) pv. This function will dispose of
2260 * the lock along with the pv.
2262 * If the caller already holds the locked parent page table for pv it
2263 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2264 * pass NULL for pvp.
2267 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp)
2272 * The pmap is currently not spinlocked, pv is held+locked.
2273 * Remove the pv's page from its parent's page table. The
2274 * parent's page table page's wire_count will be decremented.
2276 pmap_remove_pv_pte(pv, pvp, NULL);
2279 * Terminal pvs are unhooked from their vm_pages. Because
2280 * terminal pages aren't page table pages they aren't wired
2281 * by us, so we have to be sure not to unwire them either.
2283 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2284 pmap_remove_pv_page(pv);
2289 * We leave the top-level page table page cached, wired, and
2290 * mapped in the pmap until the dtor function (pmap_puninit())
2293 * Since we are leaving the top-level pv intact we need
2294 * to break out of what would otherwise be an infinite loop.
2296 if (pv->pv_pindex == pmap_pml4_pindex()) {
2302 * For page table pages (other than the top-level page),
2303 * remove and free the vm_page. The representitive mapping
2304 * removed above by pmap_remove_pv_pte() did not undo the
2305 * last wire_count so we have to do that as well.
2307 p = pmap_remove_pv_page(pv);
2308 vm_page_busy_wait(p, FALSE, "pmaprl");
2309 if (p->wire_count != 1) {
2310 kprintf("p->wire_count was %016lx %d\n",
2311 pv->pv_pindex, p->wire_count);
2313 KKASSERT(p->wire_count == 1);
2314 KKASSERT(p->flags & PG_UNMANAGED);
2316 vm_page_unwire(p, 0);
2317 KKASSERT(p->wire_count == 0);
2320 * Theoretically this page, if not the pml4 page, should contain
2321 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2331 * This function will remove the pte associated with a pv from its parent.
2332 * Terminal pv's are supported. The removal will be interlocked if info
2333 * is non-NULL. The caller must dispose of pv instead of just unlocking
2336 * The wire count will be dropped on the parent page table. The wire
2337 * count on the page being removed (pv->pv_m) from the parent page table
2338 * is NOT touched. Note that terminal pages will not have any additional
2339 * wire counts while page table pages will have at least one representing
2340 * the mapping, plus others representing sub-mappings.
2342 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2343 * pages and user page table and terminal pages.
2345 * The pv must be locked.
2347 * XXX must lock parent pv's if they exist to remove pte XXX
2351 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2353 vm_pindex_t ptepindex = pv->pv_pindex;
2354 pmap_t pmap = pv->pv_pmap;
2360 if (ptepindex == pmap_pml4_pindex()) {
2362 * We are the top level pml4 table, there is no parent.
2364 p = pmap->pm_pmlpv->pv_m;
2365 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2367 * Remove a PDP page from the pml4e. This can only occur
2368 * with user page tables. We do not have to lock the
2369 * pml4 PV so just ignore pvp.
2371 vm_pindex_t pml4_pindex;
2372 vm_pindex_t pdp_index;
2375 pdp_index = ptepindex - pmap_pdp_pindex(0);
2377 pml4_pindex = pmap_pml4_pindex();
2378 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2382 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2383 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2384 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2386 KKASSERT(info == NULL);
2387 } else if (ptepindex >= pmap_pd_pindex(0)) {
2389 * Remove a PD page from the pdp
2391 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2392 * of a simple pmap because it stops at
2395 vm_pindex_t pdp_pindex;
2396 vm_pindex_t pd_index;
2399 pd_index = ptepindex - pmap_pd_pindex(0);
2402 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2403 (pd_index >> NPML4EPGSHIFT);
2404 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2409 pd = pv_pte_lookup(pvp, pd_index &
2410 ((1ul << NPDPEPGSHIFT) - 1));
2411 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2412 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2415 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2416 p = pv->pv_m; /* degenerate test later */
2418 KKASSERT(info == NULL);
2419 } else if (ptepindex >= pmap_pt_pindex(0)) {
2421 * Remove a PT page from the pd
2423 vm_pindex_t pd_pindex;
2424 vm_pindex_t pt_index;
2427 pt_index = ptepindex - pmap_pt_pindex(0);
2430 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2431 (pt_index >> NPDPEPGSHIFT);
2432 pvp = pv_get(pv->pv_pmap, pd_pindex);
2436 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2437 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2438 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2440 KKASSERT(info == NULL);
2443 * Remove a PTE from the PT page
2445 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2446 * pv is a pte_pv so we can safely lock pt_pv.
2448 * NOTE: FICTITIOUS pages may have multiple physical mappings
2449 * so PHYS_TO_VM_PAGE() will not necessarily work for
2452 vm_pindex_t pt_pindex;
2457 pt_pindex = ptepindex >> NPTEPGSHIFT;
2458 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2460 if (ptepindex >= NUPTE_USER) {
2461 ptep = vtopte(ptepindex << PAGE_SHIFT);
2462 KKASSERT(pvp == NULL);
2465 pt_pindex = NUPTE_TOTAL +
2466 (ptepindex >> NPDPEPGSHIFT);
2467 pvp = pv_get(pv->pv_pmap, pt_pindex);
2471 ptep = pv_pte_lookup(pvp, ptepindex &
2472 ((1ul << NPDPEPGSHIFT) - 1));
2476 pmap_inval_interlock(info, pmap, va);
2477 pte = pte_load_clear(ptep);
2479 pmap_inval_deinterlock(info, pmap);
2481 cpu_invlpg((void *)va);
2484 * Now update the vm_page_t
2486 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) !=
2487 (pmap->pmap_bits[PG_MANAGED_IDX]|pmap->pmap_bits[PG_V_IDX])) {
2488 kprintf("remove_pte badpte %016lx %016lx %d\n",
2490 pv->pv_pindex < pmap_pt_pindex(0));
2492 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2493 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2494 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2497 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2500 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2501 if (pmap_track_modified(ptepindex))
2504 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2505 vm_page_flag_set(p, PG_REFERENCED);
2507 if (pte & pmap->pmap_bits[PG_W_IDX])
2508 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2509 if (pte & pmap->pmap_bits[PG_G_IDX])
2510 cpu_invlpg((void *)va);
2514 * Unwire the parent page table page. The wire_count cannot go below
2515 * 1 here because the parent page table page is itself still mapped.
2517 * XXX remove the assertions later.
2519 KKASSERT(pv->pv_m == p);
2520 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2521 panic("pmap_remove_pv_pte: Insufficient wire_count");
2529 pmap_remove_pv_page(pv_entry_t pv)
2535 vm_page_spin_lock(m);
2537 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2540 atomic_add_int(&m->object->agg_pv_list_count, -1);
2542 if (TAILQ_EMPTY(&m->md.pv_list))
2543 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2544 vm_page_spin_unlock(m);
2549 * Grow the number of kernel page table entries, if needed.
2551 * This routine is always called to validate any address space
2552 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2553 * space below KERNBASE.
2556 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2559 vm_offset_t ptppaddr;
2561 pd_entry_t *pt, newpt;
2563 int update_kernel_vm_end;
2566 * bootstrap kernel_vm_end on first real VM use
2568 if (kernel_vm_end == 0) {
2569 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2571 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2572 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2573 ~(PAGE_SIZE * NPTEPG - 1);
2575 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2576 kernel_vm_end = kernel_map.max_offset;
2583 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2584 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2585 * do not want to force-fill 128G worth of page tables.
2587 if (kstart < KERNBASE) {
2588 if (kstart > kernel_vm_end)
2589 kstart = kernel_vm_end;
2590 KKASSERT(kend <= KERNBASE);
2591 update_kernel_vm_end = 1;
2593 update_kernel_vm_end = 0;
2596 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2597 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2599 if (kend - 1 >= kernel_map.max_offset)
2600 kend = kernel_map.max_offset;
2602 while (kstart < kend) {
2603 pt = pmap_pt(&kernel_pmap, kstart);
2605 /* We need a new PDP entry */
2606 nkpg = vm_page_alloc(NULL, nkpt,
2609 VM_ALLOC_INTERRUPT);
2611 panic("pmap_growkernel: no memory to grow "
2614 paddr = VM_PAGE_TO_PHYS(nkpg);
2615 if ((nkpg->flags & PG_ZERO) == 0)
2616 pmap_zero_page(paddr);
2617 vm_page_flag_clear(nkpg, PG_ZERO);
2618 newpd = (pdp_entry_t)
2620 kernel_pmap.pmap_bits[PG_V_IDX] |
2621 kernel_pmap.pmap_bits[PG_RW_IDX] |
2622 kernel_pmap.pmap_bits[PG_A_IDX] |
2623 kernel_pmap.pmap_bits[PG_M_IDX]);
2624 *pmap_pd(&kernel_pmap, kstart) = newpd;
2626 continue; /* try again */
2628 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2629 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2630 ~(PAGE_SIZE * NPTEPG - 1);
2631 if (kstart - 1 >= kernel_map.max_offset) {
2632 kstart = kernel_map.max_offset;
2639 * This index is bogus, but out of the way
2641 nkpg = vm_page_alloc(NULL, nkpt,
2644 VM_ALLOC_INTERRUPT);
2646 panic("pmap_growkernel: no memory to grow kernel");
2649 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2650 pmap_zero_page(ptppaddr);
2651 vm_page_flag_clear(nkpg, PG_ZERO);
2652 newpt = (pd_entry_t) (ptppaddr |
2653 kernel_pmap.pmap_bits[PG_V_IDX] |
2654 kernel_pmap.pmap_bits[PG_RW_IDX] |
2655 kernel_pmap.pmap_bits[PG_A_IDX] |
2656 kernel_pmap.pmap_bits[PG_M_IDX]);
2657 *pmap_pt(&kernel_pmap, kstart) = newpt;
2660 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2661 ~(PAGE_SIZE * NPTEPG - 1);
2663 if (kstart - 1 >= kernel_map.max_offset) {
2664 kstart = kernel_map.max_offset;
2670 * Only update kernel_vm_end for areas below KERNBASE.
2672 if (update_kernel_vm_end && kernel_vm_end < kstart)
2673 kernel_vm_end = kstart;
2677 * Add a reference to the specified pmap.
2680 pmap_reference(pmap_t pmap)
2683 lwkt_gettoken(&pmap->pm_token);
2685 lwkt_reltoken(&pmap->pm_token);
2689 /***************************************************
2690 * page management routines.
2691 ***************************************************/
2694 * Hold a pv without locking it
2697 pv_hold(pv_entry_t pv)
2701 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2705 count = pv->pv_hold;
2707 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2714 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2715 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2718 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2719 * pv list via its page) must be held by the caller.
2722 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2726 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2729 pv->pv_line = lineno;
2735 count = pv->pv_hold;
2737 if ((count & PV_HOLD_LOCKED) == 0) {
2738 if (atomic_cmpset_int(&pv->pv_hold, count,
2739 (count + 1) | PV_HOLD_LOCKED)) {
2742 pv->pv_line = lineno;
2747 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2755 * Drop a previously held pv_entry which could not be locked, allowing its
2758 * Must not be called with a spinlock held as we might zfree() the pv if it
2759 * is no longer associated with a pmap and this was the last hold count.
2762 pv_drop(pv_entry_t pv)
2766 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2767 if (pv->pv_pmap == NULL)
2773 count = pv->pv_hold;
2775 KKASSERT((count & PV_HOLD_MASK) > 0);
2776 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2777 (PV_HOLD_LOCKED | 1));
2778 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2779 if (count == 1 && pv->pv_pmap == NULL)
2788 * Find or allocate the requested PV entry, returning a locked pv
2792 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2795 pv_entry_t pnew = NULL;
2797 spin_lock(&pmap->pm_spin);
2799 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2800 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2805 spin_unlock(&pmap->pm_spin);
2806 pnew = zalloc(pvzone);
2807 spin_lock(&pmap->pm_spin);
2810 pnew->pv_pmap = pmap;
2811 pnew->pv_pindex = pindex;
2812 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2814 pnew->pv_func = func;
2815 pnew->pv_line = lineno;
2817 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2818 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2819 spin_unlock(&pmap->pm_spin);
2824 spin_unlock(&pmap->pm_spin);
2825 zfree(pvzone, pnew);
2827 spin_lock(&pmap->pm_spin);
2830 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2831 spin_unlock(&pmap->pm_spin);
2833 spin_unlock(&pmap->pm_spin);
2834 _pv_lock(pv PMAP_DEBUG_COPY);
2836 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2841 spin_lock(&pmap->pm_spin);
2846 * Find the requested PV entry, returning a locked+held pv or NULL
2850 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2854 spin_lock(&pmap->pm_spin);
2859 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2860 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2864 spin_unlock(&pmap->pm_spin);
2867 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2868 spin_unlock(&pmap->pm_spin);
2870 spin_unlock(&pmap->pm_spin);
2871 _pv_lock(pv PMAP_DEBUG_COPY);
2873 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2874 pv_cache(pv, pindex);
2878 spin_lock(&pmap->pm_spin);
2883 * Lookup, hold, and attempt to lock (pmap,pindex).
2885 * If the entry does not exist NULL is returned and *errorp is set to 0
2887 * If the entry exists and could be successfully locked it is returned and
2888 * errorp is set to 0.
2890 * If the entry exists but could NOT be successfully locked it is returned
2891 * held and *errorp is set to 1.
2895 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2899 spin_lock_shared(&pmap->pm_spin);
2900 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2901 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2903 spin_unlock_shared(&pmap->pm_spin);
2907 if (pv_hold_try(pv)) {
2908 pv_cache(pv, pindex);
2909 spin_unlock_shared(&pmap->pm_spin);
2911 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
2912 return(pv); /* lock succeeded */
2914 spin_unlock_shared(&pmap->pm_spin);
2916 return (pv); /* lock failed */
2920 * Find the requested PV entry, returning a held pv or NULL
2924 pv_find(pmap_t pmap, vm_pindex_t pindex)
2928 spin_lock_shared(&pmap->pm_spin);
2930 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2931 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2933 spin_unlock_shared(&pmap->pm_spin);
2937 pv_cache(pv, pindex);
2938 spin_unlock_shared(&pmap->pm_spin);
2943 * Lock a held pv, keeping the hold count
2947 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2952 count = pv->pv_hold;
2954 if ((count & PV_HOLD_LOCKED) == 0) {
2955 if (atomic_cmpset_int(&pv->pv_hold, count,
2956 count | PV_HOLD_LOCKED)) {
2959 pv->pv_line = lineno;
2965 tsleep_interlock(pv, 0);
2966 if (atomic_cmpset_int(&pv->pv_hold, count,
2967 count | PV_HOLD_WAITING)) {
2969 kprintf("pv waiting on %s:%d\n",
2970 pv->pv_func, pv->pv_line);
2972 tsleep(pv, PINTERLOCKED, "pvwait", hz);
2979 * Unlock a held and locked pv, keeping the hold count.
2983 pv_unlock(pv_entry_t pv)
2987 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2991 count = pv->pv_hold;
2993 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2994 (PV_HOLD_LOCKED | 1));
2995 if (atomic_cmpset_int(&pv->pv_hold, count,
2997 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2998 if (count & PV_HOLD_WAITING)
3006 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3007 * and the hold count drops to zero we will free it.
3009 * Caller should not hold any spin locks. We are protected from hold races
3010 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3011 * lock held. A pv cannot be located otherwise.
3015 pv_put(pv_entry_t pv)
3017 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
3018 if (pv->pv_pmap == NULL)
3027 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
3028 * pmap. Any pte operations must have already been completed.
3032 pv_free(pv_entry_t pv)
3036 KKASSERT(pv->pv_m == NULL);
3037 if ((pmap = pv->pv_pmap) != NULL) {
3038 spin_lock(&pmap->pm_spin);
3039 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3040 if (pmap->pm_pvhint == pv)
3041 pmap->pm_pvhint = NULL;
3042 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3045 spin_unlock(&pmap->pm_spin);
3051 * This routine is very drastic, but can save the system
3059 static int warningdone=0;
3061 if (pmap_pagedaemon_waken == 0)
3063 pmap_pagedaemon_waken = 0;
3064 if (warningdone < 5) {
3065 kprintf("pmap_collect: collecting pv entries -- "
3066 "suggest increasing PMAP_SHPGPERPROC\n");
3070 for (i = 0; i < vm_page_array_size; i++) {
3071 m = &vm_page_array[i];
3072 if (m->wire_count || m->hold_count)
3074 if (vm_page_busy_try(m, TRUE) == 0) {
3075 if (m->wire_count == 0 && m->hold_count == 0) {
3084 * Scan the pmap for active page table entries and issue a callback.
3085 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3086 * its parent page table.
3088 * pte_pv will be NULL if the page or page table is unmanaged.
3089 * pt_pv will point to the page table page containing the pte for the page.
3091 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3092 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3093 * process pmap's PD and page to the callback function. This can be
3094 * confusing because the pt_pv is really a pd_pv, and the target page
3095 * table page is simply aliased by the pmap and not owned by it.
3097 * It is assumed that the start and end are properly rounded to the page size.
3099 * It is assumed that PD pages and above are managed and thus in the RB tree,
3100 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3102 struct pmap_scan_info {
3106 vm_pindex_t sva_pd_pindex;
3107 vm_pindex_t eva_pd_pindex;
3108 void (*func)(pmap_t, struct pmap_scan_info *,
3109 pv_entry_t, pv_entry_t, int, vm_offset_t,
3110 pt_entry_t *, void *);
3113 struct pmap_inval_info inval;
3116 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3117 static int pmap_scan_callback(pv_entry_t pv, void *data);
3120 pmap_scan(struct pmap_scan_info *info)
3122 struct pmap *pmap = info->pmap;
3123 pv_entry_t pd_pv; /* A page directory PV */
3124 pv_entry_t pt_pv; /* A page table PV */
3125 pv_entry_t pte_pv; /* A page table entry PV */
3127 struct pv_entry dummy_pv;
3133 * Hold the token for stability; if the pmap is empty we have nothing
3136 lwkt_gettoken(&pmap->pm_token);
3138 if (pmap->pm_stats.resident_count == 0) {
3139 lwkt_reltoken(&pmap->pm_token);
3144 pmap_inval_init(&info->inval);
3147 * Special handling for scanning one page, which is a very common
3148 * operation (it is?).
3150 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3152 if (info->sva + PAGE_SIZE == info->eva) {
3153 if (info->sva >= VM_MAX_USER_ADDRESS) {
3155 * Kernel mappings do not track wire counts on
3156 * page table pages and only maintain pd_pv and
3157 * pte_pv levels so pmap_scan() works.
3160 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3161 ptep = vtopte(info->sva);
3164 * User pages which are unmanaged will not have a
3165 * pte_pv. User page table pages which are unmanaged
3166 * (shared from elsewhere) will also not have a pt_pv.
3167 * The func() callback will pass both pte_pv and pt_pv
3168 * as NULL in that case.
3170 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3171 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3172 if (pt_pv == NULL) {
3173 KKASSERT(pte_pv == NULL);
3174 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3176 ptep = pv_pte_lookup(pd_pv,
3177 pmap_pt_index(info->sva));
3179 info->func(pmap, info,
3188 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3192 * Unlike the pv_find() case below we actually
3193 * acquired a locked pv in this case so any
3194 * race should have been resolved. It is expected
3197 KKASSERT(pte_pv == NULL);
3198 } else if (pte_pv) {
3199 KASSERT((*ptep & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3200 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3201 ("bad *ptep %016lx sva %016lx pte_pv %p",
3202 *ptep, info->sva, pte_pv));
3203 info->func(pmap, info, pte_pv, pt_pv, 0,
3204 info->sva, ptep, info->arg);
3206 KASSERT((*ptep & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3207 pmap->pmap_bits[PG_V_IDX],
3208 ("bad *ptep %016lx sva %016lx pte_pv NULL",
3210 info->func(pmap, info, NULL, pt_pv, 0,
3211 info->sva, ptep, info->arg);
3216 pmap_inval_done(&info->inval);
3217 lwkt_reltoken(&pmap->pm_token);
3222 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3225 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3226 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3228 if (info->sva >= VM_MAX_USER_ADDRESS) {
3230 * The kernel does not currently maintain any pv_entry's for
3231 * higher-level page tables.
3233 bzero(&dummy_pv, sizeof(dummy_pv));
3234 dummy_pv.pv_pindex = info->sva_pd_pindex;
3235 spin_lock(&pmap->pm_spin);
3236 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3237 pmap_scan_callback(&dummy_pv, info);
3238 ++dummy_pv.pv_pindex;
3240 spin_unlock(&pmap->pm_spin);
3243 * User page tables maintain local PML4, PDP, and PD
3244 * pv_entry's at the very least. PT pv's might be
3245 * unmanaged and thus not exist. PTE pv's might be
3246 * unmanaged and thus not exist.
3248 spin_lock(&pmap->pm_spin);
3249 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3250 pmap_scan_cmp, pmap_scan_callback, info);
3251 spin_unlock(&pmap->pm_spin);
3253 pmap_inval_done(&info->inval);
3254 lwkt_reltoken(&pmap->pm_token);
3258 * WARNING! pmap->pm_spin held
3261 pmap_scan_cmp(pv_entry_t pv, void *data)
3263 struct pmap_scan_info *info = data;
3264 if (pv->pv_pindex < info->sva_pd_pindex)
3266 if (pv->pv_pindex >= info->eva_pd_pindex)
3272 * WARNING! pmap->pm_spin held
3275 pmap_scan_callback(pv_entry_t pv, void *data)
3277 struct pmap_scan_info *info = data;
3278 struct pmap *pmap = info->pmap;
3279 pv_entry_t pd_pv; /* A page directory PV */
3280 pv_entry_t pt_pv; /* A page table PV */
3281 pv_entry_t pte_pv; /* A page table entry PV */
3285 vm_offset_t va_next;
3286 vm_pindex_t pd_pindex;
3290 * Pull the PD pindex from the pv before releasing the spinlock.
3292 * WARNING: pv is faked for kernel pmap scans.
3294 pd_pindex = pv->pv_pindex;
3295 spin_unlock(&pmap->pm_spin);
3296 pv = NULL; /* invalid after spinlock unlocked */
3299 * Calculate the page range within the PD. SIMPLE pmaps are
3300 * direct-mapped for the entire 2^64 address space. Normal pmaps
3301 * reflect the user and kernel address space which requires
3302 * cannonicalization w/regards to converting pd_pindex's back
3305 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3306 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3307 (sva & PML4_SIGNMASK)) {
3308 sva |= PML4_SIGNMASK;
3310 eva = sva + NBPDP; /* can overflow */
3311 if (sva < info->sva)
3313 if (eva < info->sva || eva > info->eva)
3317 * NOTE: kernel mappings do not track page table pages, only
3320 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3321 * However, for the scan to be efficient we try to
3322 * cache items top-down.
3327 for (; sva < eva; sva = va_next) {
3328 if (sva >= VM_MAX_USER_ADDRESS) {
3337 * PD cache (degenerate case if we skip). It is possible
3338 * for the PD to not exist due to races. This is ok.
3340 if (pd_pv == NULL) {
3341 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3342 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3344 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3346 if (pd_pv == NULL) {
3347 va_next = (sva + NBPDP) & ~PDPMASK;
3356 if (pt_pv == NULL) {
3361 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3362 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3368 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3372 * If pt_pv is NULL we either have an shared page table
3373 * page and must issue a callback specific to that case,
3374 * or there is no page table page.
3376 * Either way we can skip the page table page.
3378 if (pt_pv == NULL) {
3380 * Possible unmanaged (shared from another pmap)
3384 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3385 KKASSERT(pd_pv != NULL);
3386 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3387 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3388 info->func(pmap, info, NULL, pd_pv, 1,
3389 sva, ptep, info->arg);
3393 * Done, move to next page table page.
3395 va_next = (sva + NBPDR) & ~PDRMASK;
3402 * From this point in the loop testing pt_pv for non-NULL
3403 * means we are in UVM, else if it is NULL we are in KVM.
3405 * Limit our scan to either the end of the va represented
3406 * by the current page table page, or to the end of the
3407 * range being removed.
3410 va_next = (sva + NBPDR) & ~PDRMASK;
3417 * Scan the page table for pages. Some pages may not be
3418 * managed (might not have a pv_entry).
3420 * There is no page table management for kernel pages so
3421 * pt_pv will be NULL in that case, but otherwise pt_pv
3422 * is non-NULL, locked, and referenced.
3426 * At this point a non-NULL pt_pv means a UVA, and a NULL
3427 * pt_pv means a KVA.
3430 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3434 while (sva < va_next) {
3436 * Acquire the related pte_pv, if any. If *ptep == 0
3437 * the related pte_pv should not exist, but if *ptep
3438 * is not zero the pte_pv may or may not exist (e.g.
3439 * will not exist for an unmanaged page).
3441 * However a multitude of races are possible here.
3443 * In addition, the (pt_pv, pte_pv) lock order is
3444 * backwards, so we have to be careful in aquiring
3445 * a properly locked pte_pv.
3448 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3455 pv_put(pt_pv); /* must be non-NULL */
3457 pv_lock(pte_pv); /* safe to block now */
3460 pt_pv = pv_get(pmap,
3461 pmap_pt_pindex(sva));
3463 * pt_pv reloaded, need new ptep
3465 KKASSERT(pt_pv != NULL);
3466 ptep = pv_pte_lookup(pt_pv,
3467 pmap_pte_index(sva));
3471 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3475 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3479 kprintf("Unexpected non-NULL pte_pv "
3480 "%p pt_pv %p *ptep = %016lx\n",
3481 pte_pv, pt_pv, *ptep);
3482 panic("Unexpected non-NULL pte_pv");
3490 * Ready for the callback. The locked pte_pv (if any)
3491 * is consumed by the callback. pte_pv will exist if
3492 * the page is managed, and will not exist if it
3496 KASSERT((*ptep & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3497 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3498 ("bad *ptep %016lx sva %016lx "
3500 *ptep, sva, pte_pv));
3501 info->func(pmap, info, pte_pv, pt_pv, 0,
3502 sva, ptep, info->arg);
3504 KASSERT((*ptep & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3505 pmap->pmap_bits[PG_V_IDX],
3506 ("bad *ptep %016lx sva %016lx "
3509 info->func(pmap, info, NULL, pt_pv, 0,
3510 sva, ptep, info->arg);
3529 * Relock before returning.
3531 spin_lock(&pmap->pm_spin);
3536 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3538 struct pmap_scan_info info;
3543 info.func = pmap_remove_callback;
3545 info.doinval = 1; /* normal remove requires pmap inval */
3550 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3552 struct pmap_scan_info info;
3557 info.func = pmap_remove_callback;
3559 info.doinval = 0; /* normal remove requires pmap inval */
3564 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
3565 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3566 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3572 * This will also drop pt_pv's wire_count. Note that
3573 * terminal pages are not wired based on mmu presence.
3576 pmap_remove_pv_pte(pte_pv, pt_pv, &info->inval);
3578 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3579 pmap_remove_pv_page(pte_pv);
3581 } else if (sharept == 0) {
3585 * pt_pv's wire_count is still bumped by unmanaged pages
3586 * so we must decrement it manually.
3589 pmap_inval_interlock(&info->inval, pmap, va);
3590 pte = pte_load_clear(ptep);
3592 pmap_inval_deinterlock(&info->inval, pmap);
3593 if (pte & pmap->pmap_bits[PG_W_IDX])
3594 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3595 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3596 if (vm_page_unwire_quick(pt_pv->pv_m))
3597 panic("pmap_remove: insufficient wirecount");
3600 * Unmanaged page table, pt_pv is actually the pd_pv
3601 * for our pmap (not the share object pmap).
3603 * We have to unwire the target page table page and we
3604 * have to unwire our page directory page.
3607 pmap_inval_interlock(&info->inval, pmap, va);
3608 pte = pte_load_clear(ptep);
3610 pmap_inval_deinterlock(&info->inval, pmap);
3611 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3612 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
3613 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3614 panic("pmap_remove: shared pgtable1 bad wirecount");
3615 if (vm_page_unwire_quick(pt_pv->pv_m))
3616 panic("pmap_remove: shared pgtable2 bad wirecount");
3621 * Removes this physical page from all physical maps in which it resides.
3622 * Reflects back modify bits to the pager.
3624 * This routine may not be called from an interrupt.
3628 pmap_remove_all(vm_page_t m)
3630 struct pmap_inval_info info;
3633 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
3636 pmap_inval_init(&info);
3637 vm_page_spin_lock(m);
3638 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3639 KKASSERT(pv->pv_m == m);
3640 if (pv_hold_try(pv)) {
3641 vm_page_spin_unlock(m);
3643 vm_page_spin_unlock(m);
3646 if (pv->pv_m != m) {
3648 vm_page_spin_lock(m);
3652 * Holding no spinlocks, pv is locked.
3654 pmap_remove_pv_pte(pv, NULL, &info);
3655 pmap_remove_pv_page(pv);
3657 vm_page_spin_lock(m);
3659 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3660 vm_page_spin_unlock(m);
3661 pmap_inval_done(&info);
3665 * Set the physical protection on the specified range of this map
3666 * as requested. This function is typically only used for debug watchpoints
3669 * This function may not be called from an interrupt if the map is
3670 * not the kernel_pmap.
3672 * NOTE! For shared page table pages we just unmap the page.
3675 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3677 struct pmap_scan_info info;
3678 /* JG review for NX */
3682 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3683 pmap_remove(pmap, sva, eva);
3686 if (prot & VM_PROT_WRITE)
3691 info.func = pmap_protect_callback;
3699 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
3700 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3701 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3711 pmap_inval_interlock(&info->inval, pmap, va);
3717 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
3718 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3719 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3720 KKASSERT(m == pte_pv->pv_m);
3721 vm_page_flag_set(m, PG_REFERENCED);
3723 cbits &= ~pmap->pmap_bits[PG_A_IDX];
3725 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
3726 if (pmap_track_modified(pte_pv->pv_pindex)) {
3727 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3729 m = PHYS_TO_VM_PAGE(pbits &
3734 cbits &= ~pmap->pmap_bits[PG_M_IDX];
3737 } else if (sharept) {
3739 * Unmanaged page table, pt_pv is actually the pd_pv
3740 * for our pmap (not the share object pmap).
3742 * When asked to protect something in a shared page table
3743 * page we just unmap the page table page. We have to
3744 * invalidate the tlb in this situation.
3746 * XXX Warning, shared page tables will not be used for
3747 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
3748 * so PHYS_TO_VM_PAGE() should be safe here.
3750 pte = pte_load_clear(ptep);
3751 pmap_inval_invltlb(&info->inval);
3752 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3753 panic("pmap_protect: pgtable1 pg bad wirecount");
3754 if (vm_page_unwire_quick(pt_pv->pv_m))
3755 panic("pmap_protect: pgtable2 pg bad wirecount");
3758 /* else unmanaged page, adjust bits, no wire changes */
3761 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
3762 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
3766 pmap_inval_deinterlock(&info->inval, pmap);
3772 * Insert the vm_page (m) at the virtual address (va), replacing any prior
3773 * mapping at that address. Set protection and wiring as requested.
3775 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
3776 * possible. If it is we enter the page into the appropriate shared pmap
3777 * hanging off the related VM object instead of the passed pmap, then we
3778 * share the page table page from the VM object's pmap into the current pmap.
3780 * NOTE: This routine MUST insert the page into the pmap now, it cannot
3784 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3785 boolean_t wired, vm_map_entry_t entry)
3787 pmap_inval_info info;
3788 pv_entry_t pt_pv; /* page table */
3789 pv_entry_t pte_pv; /* page table entry */
3792 pt_entry_t origpte, newpte;
3797 va = trunc_page(va);
3798 #ifdef PMAP_DIAGNOSTIC
3800 panic("pmap_enter: toobig");
3801 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
3802 panic("pmap_enter: invalid to pmap_enter page table "
3803 "pages (va: 0x%lx)", va);
3805 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
3806 kprintf("Warning: pmap_enter called on UVA with "
3809 db_print_backtrace();
3812 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
3813 kprintf("Warning: pmap_enter called on KVA without"
3816 db_print_backtrace();
3821 * Get locked PV entries for our new page table entry (pte_pv)
3822 * and for its parent page table (pt_pv). We need the parent
3823 * so we can resolve the location of the ptep.
3825 * Only hardware MMU actions can modify the ptep out from
3828 * if (m) is fictitious or unmanaged we do not create a managing
3829 * pte_pv for it. Any pre-existing page's management state must
3830 * match (avoiding code complexity).
3832 * If the pmap is still being initialized we assume existing
3835 * Kernel mapppings do not track page table pages (i.e. pt_pv).
3836 * pmap_allocpte() checks the
3838 if (pmap_initialized == FALSE) {
3842 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
3844 if (va >= VM_MAX_USER_ADDRESS) {
3848 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
3850 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3852 KKASSERT(*ptep == 0 || (*ptep & pmap->pmap_bits[PG_MANAGED_IDX]) == 0);
3854 if (va >= VM_MAX_USER_ADDRESS) {
3856 * Kernel map, pv_entry-tracked.
3859 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3865 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
3867 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3869 KKASSERT(*ptep == 0 || (*ptep & pmap->pmap_bits[PG_MANAGED_IDX]));
3872 pa = VM_PAGE_TO_PHYS(m);
3874 opa = origpte & PG_FRAME;
3876 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
3877 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
3879 newpte |= pmap->pmap_bits[PG_W_IDX];
3880 if (va < VM_MAX_USER_ADDRESS)
3881 newpte |= pmap->pmap_bits[PG_U_IDX];
3883 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
3884 // if (pmap == &kernel_pmap)
3885 // newpte |= pgeflag;
3886 newpte |= pmap->pmap_cache_bits[m->pat_mode];
3887 if (m->flags & PG_FICTITIOUS)
3888 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
3891 * It is possible for multiple faults to occur in threaded
3892 * environments, the existing pte might be correct.
3894 if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
3895 pmap->pmap_bits[PG_A_IDX])) == 0)
3898 if ((prot & VM_PROT_NOSYNC) == 0)
3899 pmap_inval_init(&info);
3902 * Ok, either the address changed or the protection or wiring
3905 * Clear the current entry, interlocking the removal. For managed
3906 * pte's this will also flush the modified state to the vm_page.
3907 * Atomic ops are mandatory in order to ensure that PG_M events are
3908 * not lost during any transition.
3913 * pmap_remove_pv_pte() unwires pt_pv and assumes
3914 * we will free pte_pv, but since we are reusing
3915 * pte_pv we want to retain the wire count.
3917 * pt_pv won't exist for a kernel page (managed or
3921 vm_page_wire_quick(pt_pv->pv_m);
3922 if (prot & VM_PROT_NOSYNC)
3923 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3925 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
3927 pmap_remove_pv_page(pte_pv);
3928 } else if (prot & VM_PROT_NOSYNC) {
3930 * Unmanaged page, NOSYNC (no mmu sync) requested.
3932 * Leave wire count on PT page intact.
3934 (void)pte_load_clear(ptep);
3935 cpu_invlpg((void *)va);
3936 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3939 * Unmanaged page, normal enter.
3941 * Leave wire count on PT page intact.
3943 pmap_inval_interlock(&info, pmap, va);
3944 (void)pte_load_clear(ptep);
3945 pmap_inval_deinterlock(&info, pmap);
3946 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3948 KKASSERT(*ptep == 0);
3953 * Enter on the PV list if part of our managed memory.
3954 * Wiring of the PT page is already handled.
3956 KKASSERT(pte_pv->pv_m == NULL);
3957 vm_page_spin_lock(m);
3959 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
3962 atomic_add_int(&m->object->agg_pv_list_count, 1);
3964 vm_page_flag_set(m, PG_MAPPED);
3965 vm_page_spin_unlock(m);
3966 } else if (pt_pv && opa == 0) {
3968 * We have to adjust the wire count on the PT page ourselves
3969 * for unmanaged entries. If opa was non-zero we retained
3970 * the existing wire count from the removal.
3972 vm_page_wire_quick(pt_pv->pv_m);
3976 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
3978 * User VMAs do not because those will be zero->non-zero, so no
3979 * stale entries to worry about at this point.
3981 * For KVM there appear to still be issues. Theoretically we
3982 * should be able to scrap the interlocks entirely but we
3985 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3986 pmap_inval_interlock(&info, pmap, va);
3991 *(volatile pt_entry_t *)ptep = newpte;
3993 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
3994 pmap_inval_deinterlock(&info, pmap);
3995 else if (pt_pv == NULL)
3996 cpu_invlpg((void *)va);
4000 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4003 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4006 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4007 vm_page_flag_set(m, PG_WRITEABLE);
4010 * Unmanaged pages need manual resident_count tracking.
4012 if (pte_pv == NULL && pt_pv)
4013 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4018 if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
4019 pmap_inval_done(&info);
4021 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 || (m->flags & PG_MAPPED));
4024 * Cleanup the pv entry, allowing other accessors.
4033 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4034 * This code also assumes that the pmap has no pre-existing entry for this
4037 * This code currently may only be used on user pmaps, not kernel_pmap.
4040 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4042 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4046 * Make a temporary mapping for a physical address. This is only intended
4047 * to be used for panic dumps.
4049 * The caller is responsible for calling smp_invltlb().
4052 pmap_kenter_temporary(vm_paddr_t pa, long i)
4054 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4055 return ((void *)crashdumpmap);
4058 #define MAX_INIT_PT (96)
4061 * This routine preloads the ptes for a given object into the specified pmap.
4062 * This eliminates the blast of soft faults on process startup and
4063 * immediately after an mmap.
4065 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4068 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4069 vm_object_t object, vm_pindex_t pindex,
4070 vm_size_t size, int limit)
4072 struct rb_vm_page_scan_info info;
4077 * We can't preinit if read access isn't set or there is no pmap
4080 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4084 * We can't preinit if the pmap is not the current pmap
4086 lp = curthread->td_lwp;
4087 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4091 * Misc additional checks
4093 psize = x86_64_btop(size);
4095 if ((object->type != OBJT_VNODE) ||
4096 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4097 (object->resident_page_count > MAX_INIT_PT))) {
4101 if (pindex + psize > object->size) {
4102 if (object->size < pindex)
4104 psize = object->size - pindex;
4111 * If everything is segment-aligned do not pre-init here. Instead
4112 * allow the normal vm_fault path to pass a segment hint to
4113 * pmap_enter() which will then use an object-referenced shared
4116 if ((addr & SEG_MASK) == 0 &&
4117 (ctob(psize) & SEG_MASK) == 0 &&
4118 (ctob(pindex) & SEG_MASK) == 0) {
4123 * Use a red-black scan to traverse the requested range and load
4124 * any valid pages found into the pmap.
4126 * We cannot safely scan the object's memq without holding the
4129 info.start_pindex = pindex;
4130 info.end_pindex = pindex + psize - 1;
4136 vm_object_hold_shared(object);
4137 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4138 pmap_object_init_pt_callback, &info);
4139 vm_object_drop(object);
4144 pmap_object_init_pt_callback(vm_page_t p, void *data)
4146 struct rb_vm_page_scan_info *info = data;
4147 vm_pindex_t rel_index;
4150 * don't allow an madvise to blow away our really
4151 * free pages allocating pv entries.
4153 if ((info->limit & MAP_PREFAULT_MADVISE) &&
4154 vmstats.v_free_count < vmstats.v_free_reserved) {
4159 * Ignore list markers and ignore pages we cannot instantly
4160 * busy (while holding the object token).
4162 if (p->flags & PG_MARKER)
4164 if (vm_page_busy_try(p, TRUE))
4166 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4167 (p->flags & PG_FICTITIOUS) == 0) {
4168 if ((p->queue - p->pc) == PQ_CACHE)
4169 vm_page_deactivate(p);
4170 rel_index = p->pindex - info->start_pindex;
4171 pmap_enter_quick(info->pmap,
4172 info->addr + x86_64_ptob(rel_index), p);
4180 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4183 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4186 * XXX This is safe only because page table pages are not freed.
4189 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4193 /*spin_lock(&pmap->pm_spin);*/
4194 if ((pte = pmap_pte(pmap, addr)) != NULL) {
4195 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4196 /*spin_unlock(&pmap->pm_spin);*/
4200 /*spin_unlock(&pmap->pm_spin);*/
4205 * Change the wiring attribute for a pmap/va pair. The mapping must already
4206 * exist in the pmap. The mapping may or may not be managed.
4209 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4210 vm_map_entry_t entry)
4217 lwkt_gettoken(&pmap->pm_token);
4218 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
4219 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4221 if (wired && !pmap_pte_w(pmap, ptep))
4222 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4223 else if (!wired && pmap_pte_w(pmap, ptep))
4224 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4227 * Wiring is not a hardware characteristic so there is no need to
4228 * invalidate TLB. However, in an SMP environment we must use
4229 * a locked bus cycle to update the pte (if we are not using
4230 * the pmap_inval_*() API that is)... it's ok to do this for simple
4234 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4236 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4238 lwkt_reltoken(&pmap->pm_token);
4244 * Copy the range specified by src_addr/len from the source map to
4245 * the range dst_addr/len in the destination map.
4247 * This routine is only advisory and need not do anything.
4250 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4251 vm_size_t len, vm_offset_t src_addr)
4258 * Zero the specified physical page.
4260 * This function may be called from an interrupt and no locking is
4264 pmap_zero_page(vm_paddr_t phys)
4266 vm_offset_t va = PHYS_TO_DMAP(phys);
4268 pagezero((void *)va);
4272 * pmap_page_assertzero:
4274 * Assert that a page is empty, panic if it isn't.
4277 pmap_page_assertzero(vm_paddr_t phys)
4279 vm_offset_t va = PHYS_TO_DMAP(phys);
4282 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
4283 if (*(long *)((char *)va + i) != 0) {
4284 panic("pmap_page_assertzero() @ %p not zero!",
4285 (void *)(intptr_t)va);
4293 * Zero part of a physical page by mapping it into memory and clearing
4294 * its contents with bzero.
4296 * off and size may not cover an area beyond a single hardware page.
4299 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4301 vm_offset_t virt = PHYS_TO_DMAP(phys);
4303 bzero((char *)virt + off, size);
4309 * Copy the physical page from the source PA to the target PA.
4310 * This function may be called from an interrupt. No locking
4314 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4316 vm_offset_t src_virt, dst_virt;
4318 src_virt = PHYS_TO_DMAP(src);
4319 dst_virt = PHYS_TO_DMAP(dst);
4320 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4324 * pmap_copy_page_frag:
4326 * Copy the physical page from the source PA to the target PA.
4327 * This function may be called from an interrupt. No locking
4331 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4333 vm_offset_t src_virt, dst_virt;
4335 src_virt = PHYS_TO_DMAP(src);
4336 dst_virt = PHYS_TO_DMAP(dst);
4338 bcopy((char *)src_virt + (src & PAGE_MASK),
4339 (char *)dst_virt + (dst & PAGE_MASK),
4344 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4345 * this page. This count may be changed upwards or downwards in the future;
4346 * it is only necessary that true be returned for a small subset of pmaps
4347 * for proper page aging.
4350 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4355 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4358 vm_page_spin_lock(m);
4359 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4360 if (pv->pv_pmap == pmap) {
4361 vm_page_spin_unlock(m);
4368 vm_page_spin_unlock(m);
4373 * Remove all pages from specified address space this aids process exit
4374 * speeds. Also, this code may be special cased for the current process
4378 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4380 pmap_remove_noinval(pmap, sva, eva);
4385 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4386 * routines are inline, and a lot of things compile-time evaluate.
4390 pmap_testbit(vm_page_t m, int bit)
4396 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4399 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4401 vm_page_spin_lock(m);
4402 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4403 vm_page_spin_unlock(m);
4407 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4409 #if defined(PMAP_DIAGNOSTIC)
4410 if (pv->pv_pmap == NULL) {
4411 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4419 * if the bit being tested is the modified bit, then
4420 * mark clean_map and ptes as never
4423 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4424 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4425 if (!pmap_track_modified(pv->pv_pindex))
4429 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4430 if (*pte & pmap->pmap_bits[bit]) {
4431 vm_page_spin_unlock(m);
4435 vm_page_spin_unlock(m);
4440 * This routine is used to modify bits in ptes. Only one bit should be
4441 * specified. PG_RW requires special handling.
4443 * Caller must NOT hold any spin locks
4447 pmap_clearbit(vm_page_t m, int bit_index)
4449 struct pmap_inval_info info;
4455 if (bit_index == PG_RW_IDX)
4456 vm_page_flag_clear(m, PG_WRITEABLE);
4457 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4464 * Loop over all current mappings setting/clearing as appropos If
4465 * setting RO do we need to clear the VAC?
4467 * NOTE: When clearing PG_M we could also (not implemented) drop
4468 * through to the PG_RW code and clear PG_RW too, forcing
4469 * a fault on write to redetect PG_M for virtual kernels, but
4470 * it isn't necessary since virtual kernels invalidate the
4471 * pte when they clear the VPTE_M bit in their virtual page
4474 * NOTE: Does not re-dirty the page when clearing only PG_M.
4476 if (bit_index != PG_RW_IDX) {
4477 vm_page_spin_lock(m);
4478 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4479 #if defined(PMAP_DIAGNOSTIC)
4480 if (pv->pv_pmap == NULL) {
4481 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4487 pte = pmap_pte_quick(pv->pv_pmap,
4488 pv->pv_pindex << PAGE_SHIFT);
4490 if (pbits & pmap->pmap_bits[bit_index])
4491 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
4493 vm_page_spin_unlock(m);
4498 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4501 pmap_inval_init(&info);
4504 vm_page_spin_lock(m);
4505 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4507 * don't write protect pager mappings
4509 if (!pmap_track_modified(pv->pv_pindex))
4512 #if defined(PMAP_DIAGNOSTIC)
4513 if (pv->pv_pmap == NULL) {
4514 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4521 * Skip pages which do not have PG_RW set.
4523 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4524 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
4530 if (pv_hold_try(pv)) {
4531 vm_page_spin_unlock(m);
4533 vm_page_spin_unlock(m);
4534 pv_lock(pv); /* held, now do a blocking lock */
4536 if (pv->pv_pmap != pmap || pv->pv_m != m) {
4537 pv_put(pv); /* and release */
4538 goto restart; /* anything could have happened */
4540 pmap_inval_interlock(&info, pmap,
4541 (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
4542 KKASSERT(pv->pv_pmap == pmap);
4546 if (atomic_cmpset_long(pte, pbits, pbits &
4547 ~(pmap->pmap_bits[PG_RW_IDX] |
4548 pmap->pmap_bits[PG_M_IDX]))) {
4552 pmap_inval_deinterlock(&info, pmap);
4553 vm_page_spin_lock(m);
4556 * If PG_M was found to be set while we were clearing PG_RW
4557 * we also clear PG_M (done above) and mark the page dirty.
4558 * Callers expect this behavior.
4560 if (pbits & pmap->pmap_bits[PG_M_IDX])
4564 vm_page_spin_unlock(m);
4565 pmap_inval_done(&info);
4569 * Lower the permission for all mappings to a given page.
4571 * Page must be busied by caller.
4574 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4576 /* JG NX support? */
4577 if ((prot & VM_PROT_WRITE) == 0) {
4578 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4580 * NOTE: pmap_clearbit(.. PG_RW) also clears
4581 * the PG_WRITEABLE flag in (m).
4583 pmap_clearbit(m, PG_RW_IDX);
4591 pmap_phys_address(vm_pindex_t ppn)
4593 return (x86_64_ptob(ppn));
4597 * Return a count of reference bits for a page, clearing those bits.
4598 * It is not necessary for every reference bit to be cleared, but it
4599 * is necessary that 0 only be returned when there are truly no
4600 * reference bits set.
4602 * XXX: The exact number of bits to check and clear is a matter that
4603 * should be tested and standardized at some point in the future for
4604 * optimal aging of shared pages.
4606 * This routine may not block.
4609 pmap_ts_referenced(vm_page_t m)
4616 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4619 vm_page_spin_lock(m);
4620 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4621 if (!pmap_track_modified(pv->pv_pindex))
4624 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4625 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
4626 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
4632 vm_page_spin_unlock(m);
4639 * Return whether or not the specified physical page was modified
4640 * in any physical maps.
4643 pmap_is_modified(vm_page_t m)
4647 res = pmap_testbit(m, PG_M_IDX);
4652 * Clear the modify bits on the specified physical page.
4655 pmap_clear_modify(vm_page_t m)
4657 pmap_clearbit(m, PG_M_IDX);
4661 * pmap_clear_reference:
4663 * Clear the reference bit on the specified physical page.
4666 pmap_clear_reference(vm_page_t m)
4668 pmap_clearbit(m, PG_A_IDX);
4672 * Miscellaneous support routines follow
4677 i386_protection_init(void)
4681 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4682 kp = protection_codes;
4683 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
4685 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4687 * Read access is also 0. There isn't any execute bit,
4688 * so just make it readable.
4690 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4691 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4692 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4695 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4696 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4697 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4698 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4699 *kp++ = pmap_bits_default[PG_RW_IDX];
4706 * Map a set of physical memory pages into the kernel virtual
4707 * address space. Return a pointer to where it is mapped. This
4708 * routine is intended to be used for mapping device memory,
4711 * NOTE: We can't use pgeflag unless we invalidate the pages one at
4714 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
4715 * work whether the cpu supports PAT or not. The remaining PAT
4716 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
4720 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
4722 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4726 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
4728 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
4732 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
4734 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4738 * Map a set of physical memory pages into the kernel virtual
4739 * address space. Return a pointer to where it is mapped. This
4740 * routine is intended to be used for mapping device memory,
4744 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
4746 vm_offset_t va, tmpva, offset;
4750 offset = pa & PAGE_MASK;
4751 size = roundup(offset + size, PAGE_SIZE);
4753 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4755 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4757 pa = pa & ~PAGE_MASK;
4758 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
4759 pte = vtopte(tmpva);
4761 kernel_pmap.pmap_bits[PG_RW_IDX] |
4762 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
4763 kernel_pmap.pmap_cache_bits[mode];
4764 tmpsize -= PAGE_SIZE;
4768 pmap_invalidate_range(&kernel_pmap, va, va + size);
4769 pmap_invalidate_cache_range(va, va + size);
4771 return ((void *)(va + offset));
4775 pmap_unmapdev(vm_offset_t va, vm_size_t size)
4777 vm_offset_t base, offset;
4779 base = va & ~PAGE_MASK;
4780 offset = va & PAGE_MASK;
4781 size = roundup(offset + size, PAGE_SIZE);
4782 pmap_qremove(va, size >> PAGE_SHIFT);
4783 kmem_free(&kernel_map, base, size);
4787 * Sets the memory attribute for the specified page.
4790 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
4796 * If "m" is a normal page, update its direct mapping. This update
4797 * can be relied upon to perform any cache operations that are
4798 * required for data coherence.
4800 if ((m->flags & PG_FICTITIOUS) == 0)
4801 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE,
4806 * Change the PAT attribute on an existing kernel memory map. Caller
4807 * must ensure that the virtual memory in question is not accessed
4808 * during the adjustment.
4811 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
4818 panic("pmap_change_attr: va is NULL");
4819 base = trunc_page(va);
4823 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
4824 kernel_pmap.pmap_cache_bits[mode];
4829 changed = 1; /* XXX: not optimal */
4832 * Flush CPU caches if required to make sure any data isn't cached that
4833 * shouldn't be, etc.
4836 pmap_invalidate_range(&kernel_pmap, base, va);
4837 pmap_invalidate_cache_range(base, va);
4842 * perform the pmap work for mincore
4845 pmap_mincore(pmap_t pmap, vm_offset_t addr)
4847 pt_entry_t *ptep, pte;
4851 lwkt_gettoken(&pmap->pm_token);
4852 ptep = pmap_pte(pmap, addr);
4854 if (ptep && (pte = *ptep) != 0) {
4857 val = MINCORE_INCORE;
4858 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
4861 pa = pte & PG_FRAME;
4863 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
4866 m = PHYS_TO_VM_PAGE(pa);
4871 if (pte & pmap->pmap_bits[PG_M_IDX])
4872 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
4874 * Modified by someone
4876 else if (m && (m->dirty || pmap_is_modified(m)))
4877 val |= MINCORE_MODIFIED_OTHER;
4881 if (pte & pmap->pmap_bits[PG_A_IDX])
4882 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
4885 * Referenced by someone
4887 else if (m && ((m->flags & PG_REFERENCED) ||
4888 pmap_ts_referenced(m))) {
4889 val |= MINCORE_REFERENCED_OTHER;
4890 vm_page_flag_set(m, PG_REFERENCED);
4894 lwkt_reltoken(&pmap->pm_token);
4900 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
4901 * vmspace will be ref'd and the old one will be deref'd.
4903 * The vmspace for all lwps associated with the process will be adjusted
4904 * and cr3 will be reloaded if any lwp is the current lwp.
4906 * The process must hold the vmspace->vm_map.token for oldvm and newvm
4909 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
4911 struct vmspace *oldvm;
4914 oldvm = p->p_vmspace;
4915 if (oldvm != newvm) {
4917 sysref_get(&newvm->vm_sysref);
4918 p->p_vmspace = newvm;
4919 KKASSERT(p->p_nthreads == 1);
4920 lp = RB_ROOT(&p->p_lwp_tree);
4921 pmap_setlwpvm(lp, newvm);
4923 sysref_put(&oldvm->vm_sysref);
4928 * Set the vmspace for a LWP. The vmspace is almost universally set the
4929 * same as the process vmspace, but virtual kernels need to swap out contexts
4930 * on a per-lwp basis.
4932 * Caller does not necessarily hold any vmspace tokens. Caller must control
4933 * the lwp (typically be in the context of the lwp). We use a critical
4934 * section to protect against statclock and hardclock (statistics collection).
4937 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
4939 struct vmspace *oldvm;
4942 oldvm = lp->lwp_vmspace;
4944 if (oldvm != newvm) {
4946 lp->lwp_vmspace = newvm;
4947 if (curthread->td_lwp == lp) {
4948 pmap = vmspace_pmap(newvm);
4949 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4950 if (pmap->pm_active & CPUMASK_LOCK)
4951 pmap_interlock_wait(newvm);
4952 #if defined(SWTCH_OPTIM_STATS)
4955 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
4956 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
4957 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
4958 curthread->td_pcb->pcb_cr3 = KPML4phys;
4960 panic("pmap_setlwpvm: unknown pmap type\n");
4962 load_cr3(curthread->td_pcb->pcb_cr3);
4963 pmap = vmspace_pmap(oldvm);
4964 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
4971 * Called when switching to a locked pmap, used to interlock against pmaps
4972 * undergoing modifications to prevent us from activating the MMU for the
4973 * target pmap until all such modifications have completed. We have to do
4974 * this because the thread making the modifications has already set up its
4975 * SMP synchronization mask.
4977 * This function cannot sleep!
4982 pmap_interlock_wait(struct vmspace *vm)
4984 struct pmap *pmap = &vm->vm_pmap;
4986 if (pmap->pm_active & CPUMASK_LOCK) {
4988 KKASSERT(curthread->td_critcount >= 2);
4989 DEBUG_PUSH_INFO("pmap_interlock_wait");
4990 while (pmap->pm_active & CPUMASK_LOCK) {
4992 lwkt_process_ipiq();
5000 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5003 if ((obj == NULL) || (size < NBPDR) ||
5004 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5008 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
5013 * Used by kmalloc/kfree, page already exists at va
5016 pmap_kvtom(vm_offset_t va)
5018 pt_entry_t *ptep = vtopte(va);
5020 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5021 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5025 * Initialize machine-specific shared page directory support. This
5026 * is executed when a VM object is created.
5029 pmap_object_init(vm_object_t object)
5031 object->md.pmap_rw = NULL;
5032 object->md.pmap_ro = NULL;
5036 * Clean up machine-specific shared page directory support. This
5037 * is executed when a VM object is destroyed.
5040 pmap_object_free(vm_object_t object)
5044 if ((pmap = object->md.pmap_rw) != NULL) {
5045 object->md.pmap_rw = NULL;
5046 pmap_remove_noinval(pmap,
5047 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5048 pmap->pm_active = 0;
5051 kfree(pmap, M_OBJPMAP);
5053 if ((pmap = object->md.pmap_ro) != NULL) {
5054 object->md.pmap_ro = NULL;
5055 pmap_remove_noinval(pmap,
5056 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5057 pmap->pm_active = 0;
5060 kfree(pmap, M_OBJPMAP);