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_debug = 0;
234 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
235 &pmap_debug, 0, "Debug pmap's");
237 static int pmap_enter_debug = 0;
238 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
239 &pmap_enter_debug, 0, "Debug pmap_enter's");
241 static int pmap_yield_count = 64;
242 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
243 &pmap_yield_count, 0, "Yield during init_pt/release");
244 static int pmap_mmu_optimize = 0;
245 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
246 &pmap_mmu_optimize, 0, "Share page table pages when possible");
247 int pmap_fast_kernel_cpusync = 0;
248 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
249 &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
250 int pmap_dynamic_delete = 1;
251 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
252 &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
256 /* Standard user access funtions */
257 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
259 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
260 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
261 extern int std_fubyte (const void *base);
262 extern int std_subyte (void *base, int byte);
263 extern long std_fuword (const void *base);
264 extern int std_suword (void *base, long word);
265 extern int std_suword32 (void *base, int word);
267 static void pv_hold(pv_entry_t pv);
268 static int _pv_hold_try(pv_entry_t pv
270 static void pv_drop(pv_entry_t pv);
271 static void _pv_lock(pv_entry_t pv
273 static void pv_unlock(pv_entry_t pv);
274 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
276 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
278 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
279 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
280 static void pv_put(pv_entry_t pv);
281 static void pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway);
282 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
283 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
285 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
286 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
287 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
288 pmap_inval_bulk_t *bulk, int destroy);
289 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
290 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
291 pmap_inval_bulk_t *bulk);
293 struct pmap_scan_info;
294 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
295 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
296 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
297 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
298 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
299 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
301 static void i386_protection_init (void);
302 static void create_pagetables(vm_paddr_t *firstaddr);
303 static void pmap_remove_all (vm_page_t m);
304 static boolean_t pmap_testbit (vm_page_t m, int bit);
306 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
307 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
309 static void pmap_pinit_defaults(struct pmap *pmap);
311 static unsigned pdir4mb;
314 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
316 if (pv1->pv_pindex < pv2->pv_pindex)
318 if (pv1->pv_pindex > pv2->pv_pindex)
323 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
324 pv_entry_compare, vm_pindex_t, pv_pindex);
328 pmap_page_stats_adding(vm_page_t m)
330 globaldata_t gd = mycpu;
332 if (TAILQ_EMPTY(&m->md.pv_list)) {
333 ++gd->gd_vmtotal.t_arm;
334 } else if (TAILQ_FIRST(&m->md.pv_list) ==
335 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
336 ++gd->gd_vmtotal.t_armshr;
337 ++gd->gd_vmtotal.t_avmshr;
339 ++gd->gd_vmtotal.t_avmshr;
345 pmap_page_stats_deleting(vm_page_t m)
347 globaldata_t gd = mycpu;
349 if (TAILQ_EMPTY(&m->md.pv_list)) {
350 --gd->gd_vmtotal.t_arm;
351 } else if (TAILQ_FIRST(&m->md.pv_list) ==
352 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
353 --gd->gd_vmtotal.t_armshr;
354 --gd->gd_vmtotal.t_avmshr;
356 --gd->gd_vmtotal.t_avmshr;
361 * Move the kernel virtual free pointer to the next
362 * 2MB. This is used to help improve performance
363 * by using a large (2MB) page for much of the kernel
364 * (.text, .data, .bss)
368 pmap_kmem_choose(vm_offset_t addr)
370 vm_offset_t newaddr = addr;
372 newaddr = roundup2(addr, NBPDR);
379 * Super fast pmap_pte routine best used when scanning the pv lists.
380 * This eliminates many course-grained invltlb calls. Note that many of
381 * the pv list scans are across different pmaps and it is very wasteful
382 * to do an entire invltlb when checking a single mapping.
384 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
388 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
390 return pmap_pte(pmap, va);
394 * Returns the pindex of a page table entry (representing a terminal page).
395 * There are NUPTE_TOTAL page table entries possible (a huge number)
397 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
398 * We want to properly translate negative KVAs.
402 pmap_pte_pindex(vm_offset_t va)
404 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
408 * Returns the pindex of a page table.
412 pmap_pt_pindex(vm_offset_t va)
414 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
418 * Returns the pindex of a page directory.
422 pmap_pd_pindex(vm_offset_t va)
424 return (NUPTE_TOTAL + NUPT_TOTAL +
425 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
430 pmap_pdp_pindex(vm_offset_t va)
432 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
433 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
438 pmap_pml4_pindex(void)
440 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
444 * Return various clipped indexes for a given VA
446 * Returns the index of a pt in a page directory, representing a page
451 pmap_pt_index(vm_offset_t va)
453 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
457 * Returns the index of a pd in a page directory page, representing a page
462 pmap_pd_index(vm_offset_t va)
464 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
468 * Returns the index of a pdp in the pml4 table, representing a page
473 pmap_pdp_index(vm_offset_t va)
475 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
479 * Generic procedure to index a pte from a pt, pd, or pdp.
481 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
482 * a page table page index but is instead of PV lookup index.
486 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
490 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
491 return(&pte[pindex]);
495 * Return pointer to PDP slot in the PML4
499 pmap_pdp(pmap_t pmap, vm_offset_t va)
501 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
505 * Return pointer to PD slot in the PDP given a pointer to the PDP
509 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
513 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
514 return (&pd[pmap_pd_index(va)]);
518 * Return pointer to PD slot in the PDP.
522 pmap_pd(pmap_t pmap, vm_offset_t va)
526 pdp = pmap_pdp(pmap, va);
527 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
529 return (pmap_pdp_to_pd(*pdp, va));
533 * Return pointer to PT slot in the PD given a pointer to the PD
537 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
541 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
542 return (&pt[pmap_pt_index(va)]);
546 * Return pointer to PT slot in the PD
548 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
549 * so we cannot lookup the PD via the PDP. Instead we
550 * must look it up via the pmap.
554 pmap_pt(pmap_t pmap, vm_offset_t va)
558 vm_pindex_t pd_pindex;
560 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
561 pd_pindex = pmap_pd_pindex(va);
562 spin_lock(&pmap->pm_spin);
563 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
564 spin_unlock(&pmap->pm_spin);
565 if (pv == NULL || pv->pv_m == NULL)
567 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
569 pd = pmap_pd(pmap, va);
570 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
572 return (pmap_pd_to_pt(*pd, va));
577 * Return pointer to PTE slot in the PT given a pointer to the PT
581 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
585 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
586 return (&pte[pmap_pte_index(va)]);
590 * Return pointer to PTE slot in the PT
594 pmap_pte(pmap_t pmap, vm_offset_t va)
598 pt = pmap_pt(pmap, va);
599 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
601 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
602 return ((pt_entry_t *)pt);
603 return (pmap_pt_to_pte(*pt, va));
607 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
608 * the PT layer. This will speed up core pmap operations considerably.
610 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
611 * must be in a known associated state (typically by being locked when
612 * the pmap spinlock isn't held). We allow the race for that case.
616 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
618 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
619 pv->pv_pmap->pm_pvhint = pv;
624 * Return address of PT slot in PD (KVM only)
626 * Cannot be used for user page tables because it might interfere with
627 * the shared page-table-page optimization (pmap_mmu_optimize).
631 vtopt(vm_offset_t va)
633 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
634 NPML4EPGSHIFT)) - 1);
636 return (PDmap + ((va >> PDRSHIFT) & mask));
640 * KVM - return address of PTE slot in PT
644 vtopte(vm_offset_t va)
646 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
647 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
649 return (PTmap + ((va >> PAGE_SHIFT) & mask));
653 allocpages(vm_paddr_t *firstaddr, long n)
658 bzero((void *)ret, n * PAGE_SIZE);
659 *firstaddr += n * PAGE_SIZE;
665 create_pagetables(vm_paddr_t *firstaddr)
667 long i; /* must be 64 bits */
673 * We are running (mostly) V=P at this point
675 * Calculate NKPT - number of kernel page tables. We have to
676 * accomodoate prealloction of the vm_page_array, dump bitmap,
677 * MSGBUF_SIZE, and other stuff. Be generous.
679 * Maxmem is in pages.
681 * ndmpdp is the number of 1GB pages we wish to map.
683 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
684 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
686 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
689 * Starting at the beginning of kvm (not KERNBASE).
691 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
692 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
693 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
694 ndmpdp) + 511) / 512;
698 * Starting at KERNBASE - map 2G worth of page table pages.
699 * KERNBASE is offset -2G from the end of kvm.
701 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
706 KPTbase = allocpages(firstaddr, nkpt_base);
707 KPTphys = allocpages(firstaddr, nkpt_phys);
708 KPML4phys = allocpages(firstaddr, 1);
709 KPDPphys = allocpages(firstaddr, NKPML4E);
710 KPDphys = allocpages(firstaddr, NKPDPE);
713 * Calculate the page directory base for KERNBASE,
714 * that is where we start populating the page table pages.
715 * Basically this is the end - 2.
717 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
719 DMPDPphys = allocpages(firstaddr, NDMPML4E);
720 if ((amd_feature & AMDID_PAGE1GB) == 0)
721 DMPDphys = allocpages(firstaddr, ndmpdp);
722 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
725 * Fill in the underlying page table pages for the area around
726 * KERNBASE. This remaps low physical memory to KERNBASE.
728 * Read-only from zero to physfree
729 * XXX not fully used, underneath 2M pages
731 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
732 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
733 ((pt_entry_t *)KPTbase)[i] |=
734 pmap_bits_default[PG_RW_IDX] |
735 pmap_bits_default[PG_V_IDX] |
736 pmap_bits_default[PG_G_IDX];
740 * Now map the initial kernel page tables. One block of page
741 * tables is placed at the beginning of kernel virtual memory,
742 * and another block is placed at KERNBASE to map the kernel binary,
743 * data, bss, and initial pre-allocations.
745 for (i = 0; i < nkpt_base; i++) {
746 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
747 ((pd_entry_t *)KPDbase)[i] |=
748 pmap_bits_default[PG_RW_IDX] |
749 pmap_bits_default[PG_V_IDX];
751 for (i = 0; i < nkpt_phys; i++) {
752 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
753 ((pd_entry_t *)KPDphys)[i] |=
754 pmap_bits_default[PG_RW_IDX] |
755 pmap_bits_default[PG_V_IDX];
759 * Map from zero to end of allocations using 2M pages as an
760 * optimization. This will bypass some of the KPTBase pages
761 * above in the KERNBASE area.
763 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
764 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
765 ((pd_entry_t *)KPDbase)[i] |=
766 pmap_bits_default[PG_RW_IDX] |
767 pmap_bits_default[PG_V_IDX] |
768 pmap_bits_default[PG_PS_IDX] |
769 pmap_bits_default[PG_G_IDX];
773 * And connect up the PD to the PDP. The kernel pmap is expected
774 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
776 for (i = 0; i < NKPDPE; i++) {
777 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
778 KPDphys + (i << PAGE_SHIFT);
779 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
780 pmap_bits_default[PG_RW_IDX] |
781 pmap_bits_default[PG_V_IDX] |
782 pmap_bits_default[PG_U_IDX];
786 * Now set up the direct map space using either 2MB or 1GB pages
787 * Preset PG_M and PG_A because demotion expects it.
789 * When filling in entries in the PD pages make sure any excess
790 * entries are set to zero as we allocated enough PD pages
792 if ((amd_feature & AMDID_PAGE1GB) == 0) {
793 for (i = 0; i < NPDEPG * ndmpdp; i++) {
794 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
795 ((pd_entry_t *)DMPDphys)[i] |=
796 pmap_bits_default[PG_RW_IDX] |
797 pmap_bits_default[PG_V_IDX] |
798 pmap_bits_default[PG_PS_IDX] |
799 pmap_bits_default[PG_G_IDX] |
800 pmap_bits_default[PG_M_IDX] |
801 pmap_bits_default[PG_A_IDX];
805 * And the direct map space's PDP
807 for (i = 0; i < ndmpdp; i++) {
808 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
810 ((pdp_entry_t *)DMPDPphys)[i] |=
811 pmap_bits_default[PG_RW_IDX] |
812 pmap_bits_default[PG_V_IDX] |
813 pmap_bits_default[PG_U_IDX];
816 for (i = 0; i < ndmpdp; i++) {
817 ((pdp_entry_t *)DMPDPphys)[i] =
818 (vm_paddr_t)i << PDPSHIFT;
819 ((pdp_entry_t *)DMPDPphys)[i] |=
820 pmap_bits_default[PG_RW_IDX] |
821 pmap_bits_default[PG_V_IDX] |
822 pmap_bits_default[PG_PS_IDX] |
823 pmap_bits_default[PG_G_IDX] |
824 pmap_bits_default[PG_M_IDX] |
825 pmap_bits_default[PG_A_IDX];
829 /* And recursively map PML4 to itself in order to get PTmap */
830 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
831 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
832 pmap_bits_default[PG_RW_IDX] |
833 pmap_bits_default[PG_V_IDX] |
834 pmap_bits_default[PG_U_IDX];
837 * Connect the Direct Map slots up to the PML4
839 for (j = 0; j < NDMPML4E; ++j) {
840 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
841 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
842 pmap_bits_default[PG_RW_IDX] |
843 pmap_bits_default[PG_V_IDX] |
844 pmap_bits_default[PG_U_IDX];
848 * Connect the KVA slot up to the PML4
850 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
851 ((pdp_entry_t *)KPML4phys)[KPML4I] |=
852 pmap_bits_default[PG_RW_IDX] |
853 pmap_bits_default[PG_V_IDX] |
854 pmap_bits_default[PG_U_IDX];
858 * Bootstrap the system enough to run with virtual memory.
860 * On the i386 this is called after mapping has already been enabled
861 * and just syncs the pmap module with what has already been done.
862 * [We can't call it easily with mapping off since the kernel is not
863 * mapped with PA == VA, hence we would have to relocate every address
864 * from the linked base (virtual) address "KERNBASE" to the actual
865 * (physical) address starting relative to 0]
868 pmap_bootstrap(vm_paddr_t *firstaddr)
873 KvaStart = VM_MIN_KERNEL_ADDRESS;
874 KvaEnd = VM_MAX_KERNEL_ADDRESS;
875 KvaSize = KvaEnd - KvaStart;
877 avail_start = *firstaddr;
880 * Create an initial set of page tables to run the kernel in.
882 create_pagetables(firstaddr);
884 virtual2_start = KvaStart;
885 virtual2_end = PTOV_OFFSET;
887 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
888 virtual_start = pmap_kmem_choose(virtual_start);
890 virtual_end = VM_MAX_KERNEL_ADDRESS;
892 /* XXX do %cr0 as well */
893 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
897 * Initialize protection array.
899 i386_protection_init();
902 * The kernel's pmap is statically allocated so we don't have to use
903 * pmap_create, which is unlikely to work correctly at this part of
904 * the boot sequence (XXX and which no longer exists).
906 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
907 kernel_pmap.pm_count = 1;
908 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
909 RB_INIT(&kernel_pmap.pm_pvroot);
910 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
911 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
914 * Reserve some special page table entries/VA space for temporary
917 #define SYSMAP(c, p, v, n) \
918 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
924 * CMAP1/CMAP2 are used for zeroing and copying pages.
926 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
931 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
934 * ptvmmap is used for reading arbitrary physical pages via
937 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
940 * msgbufp is used to map the system message buffer.
941 * XXX msgbufmap is not used.
943 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
944 atop(round_page(MSGBUF_SIZE)))
947 virtual_start = pmap_kmem_choose(virtual_start);
952 * PG_G is terribly broken on SMP because we IPI invltlb's in some
953 * cases rather then invl1pg. Actually, I don't even know why it
954 * works under UP because self-referential page table mappings
959 * Initialize the 4MB page size flag
963 * The 4MB page version of the initial
964 * kernel page mapping.
968 #if !defined(DISABLE_PSE)
969 if (cpu_feature & CPUID_PSE) {
972 * Note that we have enabled PSE mode
974 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
975 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
976 ptditmp &= ~(NBPDR - 1);
977 ptditmp |= pmap_bits_default[PG_V_IDX] |
978 pmap_bits_default[PG_RW_IDX] |
979 pmap_bits_default[PG_PS_IDX] |
980 pmap_bits_default[PG_U_IDX];
987 /* Initialize the PAT MSR */
989 pmap_pinit_defaults(&kernel_pmap);
991 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
992 &pmap_fast_kernel_cpusync);
1006 * Default values mapping PATi,PCD,PWT bits at system reset.
1007 * The default values effectively ignore the PATi bit by
1008 * repeating the encodings for 0-3 in 4-7, and map the PCD
1009 * and PWT bit combinations to the expected PAT types.
1011 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
1012 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
1013 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
1014 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
1015 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
1016 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
1017 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
1018 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
1019 pat_pte_index[PAT_WRITE_BACK] = 0;
1020 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
1021 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
1022 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
1023 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1024 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1026 if (cpu_feature & CPUID_PAT) {
1028 * If we support the PAT then set-up entries for
1029 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1032 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1033 PAT_VALUE(4, PAT_WRITE_PROTECTED);
1034 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1035 PAT_VALUE(5, PAT_WRITE_COMBINING);
1036 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1037 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1040 * Then enable the PAT
1045 load_cr4(cr4 & ~CR4_PGE);
1047 /* Disable caches (CD = 1, NW = 0). */
1049 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1051 /* Flushes caches and TLBs. */
1055 /* Update PAT and index table. */
1056 wrmsr(MSR_PAT, pat_msr);
1058 /* Flush caches and TLBs again. */
1062 /* Restore caches and PGE. */
1070 * Set 4mb pdir for mp startup
1075 if (cpu_feature & CPUID_PSE) {
1076 load_cr4(rcr4() | CR4_PSE);
1077 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
1084 * Initialize the pmap module.
1085 * Called by vm_init, to initialize any structures that the pmap
1086 * system needs to map virtual memory.
1087 * pmap_init has been enhanced to support in a fairly consistant
1088 * way, discontiguous physical memory.
1097 * Allocate memory for random pmap data structures. Includes the
1101 for (i = 0; i < vm_page_array_size; i++) {
1104 m = &vm_page_array[i];
1105 TAILQ_INIT(&m->md.pv_list);
1109 * init the pv free list
1111 initial_pvs = vm_page_array_size;
1112 if (initial_pvs < MINPV)
1113 initial_pvs = MINPV;
1114 pvzone = &pvzone_store;
1115 pvinit = (void *)kmem_alloc(&kernel_map,
1116 initial_pvs * sizeof (struct pv_entry),
1118 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1119 pvinit, initial_pvs);
1122 * Now it is safe to enable pv_table recording.
1124 pmap_initialized = TRUE;
1128 * Initialize the address space (zone) for the pv_entries. Set a
1129 * high water mark so that the system can recover from excessive
1130 * numbers of pv entries.
1135 int shpgperproc = PMAP_SHPGPERPROC;
1138 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1139 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1140 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1141 pv_entry_high_water = 9 * (pv_entry_max / 10);
1144 * Subtract out pages already installed in the zone (hack)
1146 entry_max = pv_entry_max - vm_page_array_size;
1150 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1154 * Typically used to initialize a fictitious page by vm/device_pager.c
1157 pmap_page_init(struct vm_page *m)
1160 TAILQ_INIT(&m->md.pv_list);
1163 /***************************************************
1164 * Low level helper routines.....
1165 ***************************************************/
1168 * this routine defines the region(s) of memory that should
1169 * not be tested for the modified bit.
1173 pmap_track_modified(vm_pindex_t pindex)
1175 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1176 if ((va < clean_sva) || (va >= clean_eva))
1183 * Extract the physical page address associated with the map/VA pair.
1184 * The page must be wired for this to work reliably.
1186 * XXX for the moment we're using pv_find() instead of pv_get(), as
1187 * callers might be expecting non-blocking operation.
1190 pmap_extract(pmap_t pmap, vm_offset_t va)
1197 if (va >= VM_MAX_USER_ADDRESS) {
1199 * Kernel page directories might be direct-mapped and
1200 * there is typically no PV tracking of pte's
1204 pt = pmap_pt(pmap, va);
1205 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1206 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1207 rtval = *pt & PG_PS_FRAME;
1208 rtval |= va & PDRMASK;
1210 ptep = pmap_pt_to_pte(*pt, va);
1211 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1212 rtval = *ptep & PG_FRAME;
1213 rtval |= va & PAGE_MASK;
1219 * User pages currently do not direct-map the page directory
1220 * and some pages might not used managed PVs. But all PT's
1223 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1225 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1226 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1227 rtval = *ptep & PG_FRAME;
1228 rtval |= va & PAGE_MASK;
1237 * Similar to extract but checks protections, SMP-friendly short-cut for
1238 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1239 * fall-through to the real fault code.
1241 * The returned page, if not NULL, is held (and not busied).
1244 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1246 if (pmap && va < VM_MAX_USER_ADDRESS) {
1254 req = pmap->pmap_bits[PG_V_IDX] |
1255 pmap->pmap_bits[PG_U_IDX];
1256 if (prot & VM_PROT_WRITE)
1257 req |= pmap->pmap_bits[PG_RW_IDX];
1259 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1262 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1263 if ((*ptep & req) != req) {
1267 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1268 if (pte_pv && error == 0) {
1271 if (prot & VM_PROT_WRITE)
1274 } else if (pte_pv) {
1288 * Extract the physical page address associated kernel virtual address.
1291 pmap_kextract(vm_offset_t va)
1293 pd_entry_t pt; /* pt entry in pd */
1296 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1297 pa = DMAP_TO_PHYS(va);
1300 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1301 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1304 * Beware of a concurrent promotion that changes the
1305 * PDE at this point! For example, vtopte() must not
1306 * be used to access the PTE because it would use the
1307 * new PDE. It is, however, safe to use the old PDE
1308 * because the page table page is preserved by the
1311 pa = *pmap_pt_to_pte(pt, va);
1312 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1318 /***************************************************
1319 * Low level mapping routines.....
1320 ***************************************************/
1323 * Routine: pmap_kenter
1325 * Add a wired page to the KVA
1326 * NOTE! note that in order for the mapping to take effect -- you
1327 * should do an invltlb after doing the pmap_kenter().
1330 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1336 kernel_pmap.pmap_bits[PG_RW_IDX] |
1337 kernel_pmap.pmap_bits[PG_V_IDX];
1341 pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1345 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1352 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1353 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1354 * (caller can conditionalize calling smp_invltlb()).
1357 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1364 kernel_pmap.pmap_bits[PG_RW_IDX] |
1365 kernel_pmap.pmap_bits[PG_V_IDX];
1375 cpu_invlpg((void *)va);
1381 * Enter addresses into the kernel pmap but don't bother
1382 * doing any tlb invalidations. Caller will do a rollup
1383 * invalidation via pmap_rollup_inval().
1386 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1393 kernel_pmap.pmap_bits[PG_RW_IDX] |
1394 kernel_pmap.pmap_bits[PG_V_IDX];
1404 cpu_invlpg((void *)va);
1410 * remove a page from the kernel pagetables
1413 pmap_kremove(vm_offset_t va)
1418 pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1422 pmap_kremove_quick(vm_offset_t va)
1427 (void)pte_load_clear(ptep);
1428 cpu_invlpg((void *)va);
1432 * Remove addresses from the kernel pmap but don't bother
1433 * doing any tlb invalidations. Caller will do a rollup
1434 * invalidation via pmap_rollup_inval().
1437 pmap_kremove_noinval(vm_offset_t va)
1442 (void)pte_load_clear(ptep);
1446 * XXX these need to be recoded. They are not used in any critical path.
1449 pmap_kmodify_rw(vm_offset_t va)
1451 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1452 cpu_invlpg((void *)va);
1457 pmap_kmodify_nc(vm_offset_t va)
1459 atomic_set_long(vtopte(va), PG_N);
1460 cpu_invlpg((void *)va);
1465 * Used to map a range of physical addresses into kernel virtual
1466 * address space during the low level boot, typically to map the
1467 * dump bitmap, message buffer, and vm_page_array.
1469 * These mappings are typically made at some pointer after the end of the
1472 * We could return PHYS_TO_DMAP(start) here and not allocate any
1473 * via (*virtp), but then kmem from userland and kernel dumps won't
1474 * have access to the related pointers.
1477 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1480 vm_offset_t va_start;
1482 /*return PHYS_TO_DMAP(start);*/
1487 while (start < end) {
1488 pmap_kenter_quick(va, start);
1496 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1499 * Remove the specified set of pages from the data and instruction caches.
1501 * In contrast to pmap_invalidate_cache_range(), this function does not
1502 * rely on the CPU's self-snoop feature, because it is intended for use
1503 * when moving pages into a different cache domain.
1506 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1508 vm_offset_t daddr, eva;
1511 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1512 (cpu_feature & CPUID_CLFSH) == 0)
1516 for (i = 0; i < count; i++) {
1517 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1518 eva = daddr + PAGE_SIZE;
1519 for (; daddr < eva; daddr += cpu_clflush_line_size)
1527 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1529 KASSERT((sva & PAGE_MASK) == 0,
1530 ("pmap_invalidate_cache_range: sva not page-aligned"));
1531 KASSERT((eva & PAGE_MASK) == 0,
1532 ("pmap_invalidate_cache_range: eva not page-aligned"));
1534 if (cpu_feature & CPUID_SS) {
1535 ; /* If "Self Snoop" is supported, do nothing. */
1537 /* Globally invalidate caches */
1538 cpu_wbinvd_on_all_cpus();
1543 * Invalidate the specified range of virtual memory on all cpus associated
1547 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1549 pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1553 * Add a list of wired pages to the kva. This routine is used for temporary
1554 * kernel mappings such as those found in buffer cache buffer. Page
1555 * modifications and accesses are not tracked or recorded.
1557 * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1558 * semantics as previous mappings may have been zerod without any
1561 * The page *must* be wired.
1564 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1569 end_va = beg_va + count * PAGE_SIZE;
1571 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1575 *pte = VM_PAGE_TO_PHYS(*m) |
1576 kernel_pmap.pmap_bits[PG_RW_IDX] |
1577 kernel_pmap.pmap_bits[PG_V_IDX] |
1578 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1582 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1586 * This routine jerks page mappings from the kernel -- it is meant only
1587 * for temporary mappings such as those found in buffer cache buffers.
1588 * No recording modified or access status occurs.
1590 * MPSAFE, INTERRUPT SAFE (cluster callback)
1593 pmap_qremove(vm_offset_t beg_va, int count)
1598 end_va = beg_va + count * PAGE_SIZE;
1600 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1604 (void)pte_load_clear(pte);
1605 cpu_invlpg((void *)va);
1607 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1611 * This routine removes temporary kernel mappings, only invalidating them
1612 * on the current cpu. It should only be used under carefully controlled
1616 pmap_qremove_quick(vm_offset_t beg_va, int count)
1621 end_va = beg_va + count * PAGE_SIZE;
1623 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1627 (void)pte_load_clear(pte);
1628 cpu_invlpg((void *)va);
1633 * This routine removes temporary kernel mappings *without* invalidating
1634 * the TLB. It can only be used on permanent kva reservations such as those
1635 * found in buffer cache buffers, under carefully controlled circumstances.
1637 * NOTE: Repopulating these KVAs requires unconditional invalidation.
1638 * (pmap_qenter() does unconditional invalidation).
1641 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1646 end_va = beg_va + count * PAGE_SIZE;
1648 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1652 (void)pte_load_clear(pte);
1657 * Create a new thread and optionally associate it with a (new) process.
1658 * NOTE! the new thread's cpu may not equal the current cpu.
1661 pmap_init_thread(thread_t td)
1663 /* enforce pcb placement & alignment */
1664 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1665 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1666 td->td_savefpu = &td->td_pcb->pcb_save;
1667 td->td_sp = (char *)td->td_pcb; /* no -16 */
1671 * This routine directly affects the fork perf for a process.
1674 pmap_init_proc(struct proc *p)
1679 pmap_pinit_defaults(struct pmap *pmap)
1681 bcopy(pmap_bits_default, pmap->pmap_bits,
1682 sizeof(pmap_bits_default));
1683 bcopy(protection_codes, pmap->protection_codes,
1684 sizeof(protection_codes));
1685 bcopy(pat_pte_index, pmap->pmap_cache_bits,
1686 sizeof(pat_pte_index));
1687 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1688 pmap->copyinstr = std_copyinstr;
1689 pmap->copyin = std_copyin;
1690 pmap->copyout = std_copyout;
1691 pmap->fubyte = std_fubyte;
1692 pmap->subyte = std_subyte;
1693 pmap->fuword = std_fuword;
1694 pmap->suword = std_suword;
1695 pmap->suword32 = std_suword32;
1698 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1699 * it, and IdlePTD, represents the template used to update all other pmaps.
1701 * On architectures where the kernel pmap is not integrated into the user
1702 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1703 * kernel_pmap should be used to directly access the kernel_pmap.
1706 pmap_pinit0(struct pmap *pmap)
1708 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1710 CPUMASK_ASSZERO(pmap->pm_active);
1711 pmap->pm_pvhint = NULL;
1712 RB_INIT(&pmap->pm_pvroot);
1713 spin_init(&pmap->pm_spin, "pmapinit0");
1714 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1715 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1716 pmap_pinit_defaults(pmap);
1720 * Initialize a preallocated and zeroed pmap structure,
1721 * such as one in a vmspace structure.
1724 pmap_pinit_simple(struct pmap *pmap)
1727 * Misc initialization
1730 CPUMASK_ASSZERO(pmap->pm_active);
1731 pmap->pm_pvhint = NULL;
1732 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1734 pmap_pinit_defaults(pmap);
1737 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1740 if (pmap->pm_pmlpv == NULL) {
1741 RB_INIT(&pmap->pm_pvroot);
1742 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1743 spin_init(&pmap->pm_spin, "pmapinitsimple");
1744 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1749 pmap_pinit(struct pmap *pmap)
1754 if (pmap->pm_pmlpv) {
1755 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1760 pmap_pinit_simple(pmap);
1761 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1764 * No need to allocate page table space yet but we do need a valid
1765 * page directory table.
1767 if (pmap->pm_pml4 == NULL) {
1769 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1775 * Allocate the page directory page, which wires it even though
1776 * it isn't being entered into some higher level page table (it
1777 * being the highest level). If one is already cached we don't
1778 * have to do anything.
1780 if ((pv = pmap->pm_pmlpv) == NULL) {
1781 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1782 pmap->pm_pmlpv = pv;
1783 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1784 VM_PAGE_TO_PHYS(pv->pv_m));
1788 * Install DMAP and KMAP.
1790 for (j = 0; j < NDMPML4E; ++j) {
1791 pmap->pm_pml4[DMPML4I + j] =
1792 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1793 pmap->pmap_bits[PG_RW_IDX] |
1794 pmap->pmap_bits[PG_V_IDX] |
1795 pmap->pmap_bits[PG_U_IDX];
1797 pmap->pm_pml4[KPML4I] = KPDPphys |
1798 pmap->pmap_bits[PG_RW_IDX] |
1799 pmap->pmap_bits[PG_V_IDX] |
1800 pmap->pmap_bits[PG_U_IDX];
1803 * install self-referential address mapping entry
1805 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1806 pmap->pmap_bits[PG_V_IDX] |
1807 pmap->pmap_bits[PG_RW_IDX] |
1808 pmap->pmap_bits[PG_A_IDX] |
1809 pmap->pmap_bits[PG_M_IDX];
1811 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1812 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1814 KKASSERT(pmap->pm_pml4[255] == 0);
1815 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1816 KKASSERT(pv->pv_entry.rbe_left == NULL);
1817 KKASSERT(pv->pv_entry.rbe_right == NULL);
1821 * Clean up a pmap structure so it can be physically freed. This routine
1822 * is called by the vmspace dtor function. A great deal of pmap data is
1823 * left passively mapped to improve vmspace management so we have a bit
1824 * of cleanup work to do here.
1827 pmap_puninit(pmap_t pmap)
1832 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1833 if ((pv = pmap->pm_pmlpv) != NULL) {
1834 if (pv_hold_try(pv) == 0)
1836 KKASSERT(pv == pmap->pm_pmlpv);
1837 p = pmap_remove_pv_page(pv);
1838 pv_free(pv, NULL, 1);
1839 pv = NULL; /* safety */
1840 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1841 vm_page_busy_wait(p, FALSE, "pgpun");
1842 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1843 vm_page_unwire(p, 0);
1844 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1847 * XXX eventually clean out PML4 static entries and
1848 * use vm_page_free_zero()
1851 pmap->pm_pmlpv = NULL;
1853 if (pmap->pm_pml4) {
1854 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1855 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1856 pmap->pm_pml4 = NULL;
1858 KKASSERT(pmap->pm_stats.resident_count == 0);
1859 KKASSERT(pmap->pm_stats.wired_count == 0);
1863 * Wire in kernel global address entries. To avoid a race condition
1864 * between pmap initialization and pmap_growkernel, this procedure
1865 * adds the pmap to the master list (which growkernel scans to update),
1866 * then copies the template.
1869 pmap_pinit2(struct pmap *pmap)
1871 spin_lock(&pmap_spin);
1872 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1873 spin_unlock(&pmap_spin);
1877 * This routine is called when various levels in the page table need to
1878 * be populated. This routine cannot fail.
1880 * This function returns two locked pv_entry's, one representing the
1881 * requested pv and one representing the requested pv's parent pv. If
1882 * an intermediate page table does not exist it will be created, mapped,
1883 * wired, and the parent page table will be given an additional hold
1884 * count representing the presence of the child pv_entry.
1888 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1893 vm_pindex_t pt_pindex;
1899 * If the pv already exists and we aren't being asked for the
1900 * parent page table page we can just return it. A locked+held pv
1901 * is returned. The pv will also have a second hold related to the
1902 * pmap association that we don't have to worry about.
1905 pv = pv_alloc(pmap, ptepindex, &isnew);
1906 if (isnew == 0 && pvpp == NULL)
1910 * Special case terminal PVs. These are not page table pages so
1911 * no vm_page is allocated (the caller supplied the vm_page). If
1912 * pvpp is non-NULL we are being asked to also removed the pt_pv
1915 * Note that pt_pv's are only returned for user VAs. We assert that
1916 * a pt_pv is not being requested for kernel VAs. The kernel
1917 * pre-wires all higher-level page tables so don't overload managed
1918 * higher-level page tables on top of it!
1920 if (ptepindex < pmap_pt_pindex(0)) {
1921 if (ptepindex >= NUPTE_USER) {
1922 /* kernel manages this manually for KVM */
1923 KKASSERT(pvpp == NULL);
1925 KKASSERT(pvpp != NULL);
1926 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1927 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1929 vm_page_wire_quick(pvp->pv_m);
1942 * The kernel never uses managed PT/PD/PDP pages.
1944 KKASSERT(pmap != &kernel_pmap);
1947 * Non-terminal PVs allocate a VM page to represent the page table,
1948 * so we have to resolve pvp and calculate ptepindex for the pvp
1949 * and then for the page table entry index in the pvp for
1952 if (ptepindex < pmap_pd_pindex(0)) {
1954 * pv is PT, pvp is PD
1956 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1957 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1958 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1965 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1966 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1968 } else if (ptepindex < pmap_pdp_pindex(0)) {
1970 * pv is PD, pvp is PDP
1972 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1975 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1976 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1978 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1979 KKASSERT(pvpp == NULL);
1982 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1990 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1991 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1992 } else if (ptepindex < pmap_pml4_pindex()) {
1994 * pv is PDP, pvp is the root pml4 table
1996 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2003 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2004 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2007 * pv represents the top-level PML4, there is no parent.
2015 * (isnew) is TRUE, pv is not terminal.
2017 * (1) Add a wire count to the parent page table (pvp).
2018 * (2) Allocate a VM page for the page table.
2019 * (3) Enter the VM page into the parent page table.
2021 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2024 vm_page_wire_quick(pvp->pv_m);
2027 m = vm_page_alloc(NULL, pv->pv_pindex,
2028 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2029 VM_ALLOC_INTERRUPT);
2034 vm_page_spin_lock(m);
2035 pmap_page_stats_adding(m);
2036 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2038 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2039 vm_page_spin_unlock(m);
2040 vm_page_unmanage(m); /* m must be spinunlocked */
2042 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2043 m->valid = VM_PAGE_BITS_ALL;
2044 vm_page_wire(m); /* wire for mapping in parent */
2047 * Wire the page into pvp. Bump the resident_count for the pmap.
2048 * There is no pvp for the top level, address the pm_pml4[] array
2051 * If the caller wants the parent we return it, otherwise
2052 * we just put it away.
2054 * No interlock is needed for pte 0 -> non-zero.
2056 * In the situation where *ptep is valid we might have an unmanaged
2057 * page table page shared from another page table which we need to
2058 * unshare before installing our private page table page.
2061 ptep = pv_pte_lookup(pvp, ptepindex);
2062 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2066 panic("pmap_allocpte: unexpected pte %p/%d",
2067 pvp, (int)ptepindex);
2069 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
2070 if (vm_page_unwire_quick(
2071 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2072 panic("pmap_allocpte: shared pgtable "
2073 "pg bad wirecount");
2075 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2077 *ptep = VM_PAGE_TO_PHYS(m) |
2078 (pmap->pmap_bits[PG_U_IDX] |
2079 pmap->pmap_bits[PG_RW_IDX] |
2080 pmap->pmap_bits[PG_V_IDX] |
2081 pmap->pmap_bits[PG_A_IDX] |
2082 pmap->pmap_bits[PG_M_IDX]);
2094 * This version of pmap_allocpte() checks for possible segment optimizations
2095 * that would allow page-table sharing. It can be called for terminal
2096 * page or page table page ptepindex's.
2098 * The function is called with page table page ptepindex's for fictitious
2099 * and unmanaged terminal pages. That is, we don't want to allocate a
2100 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2103 * This function can return a pv and *pvpp associated with the passed in pmap
2104 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2105 * an unmanaged page table page will be entered into the pass in pmap.
2109 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2110 vm_map_entry_t entry, vm_offset_t va)
2116 pv_entry_t pte_pv; /* in original or shared pmap */
2117 pv_entry_t pt_pv; /* in original or shared pmap */
2118 pv_entry_t proc_pd_pv; /* in original pmap */
2119 pv_entry_t proc_pt_pv; /* in original pmap */
2120 pv_entry_t xpv; /* PT in shared pmap */
2121 pd_entry_t *pt; /* PT entry in PD of original pmap */
2122 pd_entry_t opte; /* contents of *pt */
2123 pd_entry_t npte; /* contents of *pt */
2128 * Basic tests, require a non-NULL vm_map_entry, require proper
2129 * alignment and type for the vm_map_entry, require that the
2130 * underlying object already be allocated.
2132 * We allow almost any type of object to use this optimization.
2133 * The object itself does NOT have to be sized to a multiple of the
2134 * segment size, but the memory mapping does.
2136 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2137 * won't work as expected.
2139 if (entry == NULL ||
2140 pmap_mmu_optimize == 0 || /* not enabled */
2141 (pmap->pm_flags & PMAP_HVM) || /* special pmap */
2142 ptepindex >= pmap_pd_pindex(0) || /* not terminal or pt */
2143 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2144 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2145 entry->object.vm_object == NULL || /* needs VM object */
2146 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2147 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2148 (entry->offset & SEG_MASK) || /* must be aligned */
2149 (entry->start & SEG_MASK)) {
2150 return(pmap_allocpte(pmap, ptepindex, pvpp));
2154 * Make sure the full segment can be represented.
2156 b = va & ~(vm_offset_t)SEG_MASK;
2157 if (b < entry->start || b + SEG_SIZE > entry->end)
2158 return(pmap_allocpte(pmap, ptepindex, pvpp));
2161 * If the full segment can be represented dive the VM object's
2162 * shared pmap, allocating as required.
2164 object = entry->object.vm_object;
2166 if (entry->protection & VM_PROT_WRITE)
2167 obpmapp = &object->md.pmap_rw;
2169 obpmapp = &object->md.pmap_ro;
2172 if (pmap_enter_debug > 0) {
2174 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2176 va, entry->protection, object,
2178 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2179 entry, entry->start, entry->end);
2184 * We allocate what appears to be a normal pmap but because portions
2185 * of this pmap are shared with other unrelated pmaps we have to
2186 * set pm_active to point to all cpus.
2188 * XXX Currently using pmap_spin to interlock the update, can't use
2189 * vm_object_hold/drop because the token might already be held
2190 * shared OR exclusive and we don't know.
2192 while ((obpmap = *obpmapp) == NULL) {
2193 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2194 pmap_pinit_simple(obpmap);
2195 pmap_pinit2(obpmap);
2196 spin_lock(&pmap_spin);
2197 if (*obpmapp != NULL) {
2201 spin_unlock(&pmap_spin);
2202 pmap_release(obpmap);
2203 pmap_puninit(obpmap);
2204 kfree(obpmap, M_OBJPMAP);
2205 obpmap = *obpmapp; /* safety */
2207 obpmap->pm_active = smp_active_mask;
2208 obpmap->pm_flags |= PMAP_SEGSHARED;
2210 spin_unlock(&pmap_spin);
2215 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2216 * pte/pt using the shared pmap from the object but also adjust
2217 * the process pmap's page table page as a side effect.
2221 * Resolve the terminal PTE and PT in the shared pmap. This is what
2222 * we will return. This is true if ptepindex represents a terminal
2223 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2227 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2228 if (ptepindex >= pmap_pt_pindex(0))
2234 * Resolve the PD in the process pmap so we can properly share the
2235 * page table page. Lock order is bottom-up (leaf first)!
2237 * NOTE: proc_pt_pv can be NULL.
2239 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2240 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2242 if (pmap_enter_debug > 0) {
2244 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2246 (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2253 * xpv is the page table page pv from the shared object
2254 * (for convenience), from above.
2256 * Calculate the pte value for the PT to load into the process PD.
2257 * If we have to change it we must properly dispose of the previous
2260 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2261 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2262 (pmap->pmap_bits[PG_U_IDX] |
2263 pmap->pmap_bits[PG_RW_IDX] |
2264 pmap->pmap_bits[PG_V_IDX] |
2265 pmap->pmap_bits[PG_A_IDX] |
2266 pmap->pmap_bits[PG_M_IDX]);
2269 * Dispose of previous page table page if it was local to the
2270 * process pmap. If the old pt is not empty we cannot dispose of it
2271 * until we clean it out. This case should not arise very often so
2272 * it is not optimized.
2275 pmap_inval_bulk_t bulk;
2277 if (proc_pt_pv->pv_m->wire_count != 1) {
2283 va & ~(vm_offset_t)SEG_MASK,
2284 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2289 * The release call will indirectly clean out *pt
2291 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2292 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2293 pmap_inval_bulk_flush(&bulk);
2296 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2300 * Handle remaining cases.
2304 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2305 vm_page_wire_quick(proc_pd_pv->pv_m); /* proc pd for sh pt */
2306 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2307 } else if (*pt != npte) {
2308 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2311 opte = pte_load_clear(pt);
2312 KKASSERT(opte && opte != npte);
2316 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2319 * Clean up opte, bump the wire_count for the process
2320 * PD page representing the new entry if it was
2323 * If the entry was not previously empty and we have
2324 * a PT in the proc pmap then opte must match that
2325 * pt. The proc pt must be retired (this is done
2326 * later on in this procedure).
2328 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2331 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2332 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2333 if (vm_page_unwire_quick(m)) {
2334 panic("pmap_allocpte_seg: "
2335 "bad wire count %p",
2341 * The existing process page table was replaced and must be destroyed
2355 * Release any resources held by the given physical map.
2357 * Called when a pmap initialized by pmap_pinit is being released. Should
2358 * only be called if the map contains no valid mappings.
2360 * Caller must hold pmap->pm_token
2362 struct pmap_release_info {
2368 static int pmap_release_callback(pv_entry_t pv, void *data);
2371 pmap_release(struct pmap *pmap)
2373 struct pmap_release_info info;
2375 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2376 ("pmap still active! %016jx",
2377 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2379 spin_lock(&pmap_spin);
2380 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2381 spin_unlock(&pmap_spin);
2384 * Pull pv's off the RB tree in order from low to high and release
2392 spin_lock(&pmap->pm_spin);
2393 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2394 pmap_release_callback, &info);
2395 spin_unlock(&pmap->pm_spin);
2399 } while (info.retry);
2403 * One resident page (the pml4 page) should remain.
2404 * No wired pages should remain.
2406 KKASSERT(pmap->pm_stats.resident_count ==
2407 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2409 KKASSERT(pmap->pm_stats.wired_count == 0);
2413 * Called from low to high. We must cache the proper parent pv so we
2414 * can adjust its wired count.
2417 pmap_release_callback(pv_entry_t pv, void *data)
2419 struct pmap_release_info *info = data;
2420 pmap_t pmap = info->pmap;
2424 if (info->pvp == pv) {
2425 spin_unlock(&pmap->pm_spin);
2427 } else if (pv_hold_try(pv)) {
2428 spin_unlock(&pmap->pm_spin);
2430 spin_unlock(&pmap->pm_spin);
2433 if (pv->pv_pmap != pmap) {
2435 spin_lock(&pmap->pm_spin);
2440 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2444 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2445 pindex += NUPTE_TOTAL;
2446 } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2450 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2451 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2452 } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2456 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2458 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2459 } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2461 * parent is PML4 (there's only one)
2464 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2465 NUPD_TOTAL) >> NPML4EPGSHIFT;
2466 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2468 pindex = pmap_pml4_pindex();
2480 if (info->pvp && info->pvp->pv_pindex != pindex) {
2484 if (info->pvp == NULL)
2485 info->pvp = pv_get(pmap, pindex);
2492 r = pmap_release_pv(pv, info->pvp, NULL);
2493 spin_lock(&pmap->pm_spin);
2498 * Called with held (i.e. also locked) pv. This function will dispose of
2499 * the lock along with the pv.
2501 * If the caller already holds the locked parent page table for pv it
2502 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2503 * pass NULL for pvp.
2506 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2511 * The pmap is currently not spinlocked, pv is held+locked.
2512 * Remove the pv's page from its parent's page table. The
2513 * parent's page table page's wire_count will be decremented.
2515 * This will clean out the pte at any level of the page table.
2516 * If smp != 0 all cpus are affected.
2518 * Do not tear-down recursively, its faster to just let the
2519 * release run its course.
2521 pmap_remove_pv_pte(pv, pvp, bulk, 0);
2524 * Terminal pvs are unhooked from their vm_pages. Because
2525 * terminal pages aren't page table pages they aren't wired
2526 * by us, so we have to be sure not to unwire them either.
2528 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2529 pmap_remove_pv_page(pv);
2534 * We leave the top-level page table page cached, wired, and
2535 * mapped in the pmap until the dtor function (pmap_puninit())
2538 * Since we are leaving the top-level pv intact we need
2539 * to break out of what would otherwise be an infinite loop.
2541 if (pv->pv_pindex == pmap_pml4_pindex()) {
2547 * For page table pages (other than the top-level page),
2548 * remove and free the vm_page. The representitive mapping
2549 * removed above by pmap_remove_pv_pte() did not undo the
2550 * last wire_count so we have to do that as well.
2552 p = pmap_remove_pv_page(pv);
2553 vm_page_busy_wait(p, FALSE, "pmaprl");
2554 if (p->wire_count != 1) {
2555 kprintf("p->wire_count was %016lx %d\n",
2556 pv->pv_pindex, p->wire_count);
2558 KKASSERT(p->wire_count == 1);
2559 KKASSERT(p->flags & PG_UNMANAGED);
2561 vm_page_unwire(p, 0);
2562 KKASSERT(p->wire_count == 0);
2566 pv_free(pv, pvp, 1);
2572 * This function will remove the pte associated with a pv from its parent.
2573 * Terminal pv's are supported. All cpus specified by (bulk) are properly
2576 * The wire count will be dropped on the parent page table. The wire
2577 * count on the page being removed (pv->pv_m) from the parent page table
2578 * is NOT touched. Note that terminal pages will not have any additional
2579 * wire counts while page table pages will have at least one representing
2580 * the mapping, plus others representing sub-mappings.
2582 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2583 * pages and user page table and terminal pages.
2585 * The pv must be locked. The pvp, if supplied, must be locked. All
2586 * supplied pv's will remain locked on return.
2588 * XXX must lock parent pv's if they exist to remove pte XXX
2592 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2595 vm_pindex_t ptepindex = pv->pv_pindex;
2596 pmap_t pmap = pv->pv_pmap;
2602 if (ptepindex == pmap_pml4_pindex()) {
2604 * We are the top level pml4 table, there is no parent.
2606 p = pmap->pm_pmlpv->pv_m;
2607 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2609 * Remove a PDP page from the pml4e. This can only occur
2610 * with user page tables. We do not have to lock the
2611 * pml4 PV so just ignore pvp.
2613 vm_pindex_t pml4_pindex;
2614 vm_pindex_t pdp_index;
2617 pdp_index = ptepindex - pmap_pdp_pindex(0);
2619 pml4_pindex = pmap_pml4_pindex();
2620 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2624 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2625 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2626 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2627 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2628 } else if (ptepindex >= pmap_pd_pindex(0)) {
2630 * Remove a PD page from the pdp
2632 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2633 * of a simple pmap because it stops at
2636 vm_pindex_t pdp_pindex;
2637 vm_pindex_t pd_index;
2640 pd_index = ptepindex - pmap_pd_pindex(0);
2643 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2644 (pd_index >> NPML4EPGSHIFT);
2645 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2649 pd = pv_pte_lookup(pvp, pd_index &
2650 ((1ul << NPDPEPGSHIFT) - 1));
2651 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2652 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2653 pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2655 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2656 p = pv->pv_m; /* degenerate test later */
2658 } else if (ptepindex >= pmap_pt_pindex(0)) {
2660 * Remove a PT page from the pd
2662 vm_pindex_t pd_pindex;
2663 vm_pindex_t pt_index;
2666 pt_index = ptepindex - pmap_pt_pindex(0);
2669 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2670 (pt_index >> NPDPEPGSHIFT);
2671 pvp = pv_get(pv->pv_pmap, pd_pindex);
2675 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2676 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2677 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2678 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2681 * Remove a managed PTE from the PT page. Userland pmaps
2682 * manage PT/PD/PDP page tables pages but the kernel_pmap
2685 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2686 * pv is a pte_pv so we can safely lock pt_pv.
2688 * NOTE: FICTITIOUS pages may have multiple physical mappings
2689 * so PHYS_TO_VM_PAGE() will not necessarily work for
2692 vm_pindex_t pt_pindex;
2697 pt_pindex = ptepindex >> NPTEPGSHIFT;
2698 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2700 if (ptepindex >= NUPTE_USER) {
2701 ptep = vtopte(ptepindex << PAGE_SHIFT);
2702 KKASSERT(pvp == NULL);
2705 pt_pindex = NUPTE_TOTAL +
2706 (ptepindex >> NPDPEPGSHIFT);
2707 pvp = pv_get(pv->pv_pmap, pt_pindex);
2711 ptep = pv_pte_lookup(pvp, ptepindex &
2712 ((1ul << NPDPEPGSHIFT) - 1));
2714 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2715 if (bulk == NULL) /* XXX */
2716 cpu_invlpg((void *)va); /* XXX */
2719 * Now update the vm_page_t
2721 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] |
2722 pmap->pmap_bits[PG_V_IDX])) !=
2723 (pmap->pmap_bits[PG_MANAGED_IDX] |
2724 pmap->pmap_bits[PG_V_IDX])) {
2725 kprintf("remove_pte badpte %016lx %016lx %d\n",
2727 pv->pv_pindex < pmap_pt_pindex(0));
2730 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2731 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2732 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2735 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2738 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2739 if (pmap_track_modified(ptepindex))
2742 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2743 vm_page_flag_set(p, PG_REFERENCED);
2745 if (pte & pmap->pmap_bits[PG_W_IDX])
2746 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2747 if (pte & pmap->pmap_bits[PG_G_IDX])
2748 cpu_invlpg((void *)va);
2750 KKASSERT(pv->pv_m == p); /* XXX remove me later */
2753 * If requested, scrap the underlying pv->pv_m and the underlying
2754 * pv. If this is a page-table-page we must also free the page.
2756 * pvp must be returned locked.
2760 * page table page (PT, PD, PDP, PML4), caller was responsible
2761 * for testing wired_count.
2765 KKASSERT(pv->pv_m->wire_count == 1);
2766 p = pmap_remove_pv_page(pv);
2767 pv_free(pv, pvp, 1);
2770 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2771 vm_page_busy_wait(p, FALSE, "pgpun");
2772 vm_page_unwire(p, 0);
2773 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2775 } else if (destroy == 2) {
2777 * Normal page (leave page untouched)
2779 pmap_remove_pv_page(pv);
2780 pv_free(pv, pvp, 1);
2781 pv = NULL; /* safety */
2785 * If we acquired pvp ourselves then we are responsible for
2786 * recursively deleting it.
2788 if (pvp && gotpvp) {
2790 * Recursively destroy higher-level page tables.
2792 * This is optional. If we do not, they will still
2793 * be destroyed when the process exits.
2795 if (pmap_dynamic_delete &&
2797 pvp->pv_m->wire_count == 1 &&
2798 pvp->pv_pindex != pmap_pml4_pindex()) {
2799 if (pmap != &kernel_pmap) {
2800 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2801 pvp = NULL; /* safety */
2803 kprintf("Attempt to remove kernel_pmap pindex "
2804 "%jd\n", pvp->pv_pindex);
2814 * Remove the vm_page association to a pv. The pv must be locked.
2818 pmap_remove_pv_page(pv_entry_t pv)
2824 vm_page_spin_lock(m);
2826 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2827 pmap_page_stats_deleting(m);
2830 atomic_add_int(&m->object->agg_pv_list_count, -1);
2832 if (TAILQ_EMPTY(&m->md.pv_list))
2833 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2834 vm_page_spin_unlock(m);
2840 * Grow the number of kernel page table entries, if needed.
2842 * This routine is always called to validate any address space
2843 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2844 * space below KERNBASE.
2847 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2850 vm_offset_t ptppaddr;
2852 pd_entry_t *pt, newpt;
2854 int update_kernel_vm_end;
2857 * bootstrap kernel_vm_end on first real VM use
2859 if (kernel_vm_end == 0) {
2860 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2862 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2863 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2864 ~(PAGE_SIZE * NPTEPG - 1);
2866 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2867 kernel_vm_end = kernel_map.max_offset;
2874 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2875 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2876 * do not want to force-fill 128G worth of page tables.
2878 if (kstart < KERNBASE) {
2879 if (kstart > kernel_vm_end)
2880 kstart = kernel_vm_end;
2881 KKASSERT(kend <= KERNBASE);
2882 update_kernel_vm_end = 1;
2884 update_kernel_vm_end = 0;
2887 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2888 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2890 if (kend - 1 >= kernel_map.max_offset)
2891 kend = kernel_map.max_offset;
2893 while (kstart < kend) {
2894 pt = pmap_pt(&kernel_pmap, kstart);
2896 /* We need a new PD entry */
2897 nkpg = vm_page_alloc(NULL, nkpt,
2900 VM_ALLOC_INTERRUPT);
2902 panic("pmap_growkernel: no memory to grow "
2905 paddr = VM_PAGE_TO_PHYS(nkpg);
2906 pmap_zero_page(paddr);
2907 newpd = (pdp_entry_t)
2909 kernel_pmap.pmap_bits[PG_V_IDX] |
2910 kernel_pmap.pmap_bits[PG_RW_IDX] |
2911 kernel_pmap.pmap_bits[PG_A_IDX] |
2912 kernel_pmap.pmap_bits[PG_M_IDX]);
2913 *pmap_pd(&kernel_pmap, kstart) = newpd;
2915 continue; /* try again */
2917 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2918 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2919 ~(PAGE_SIZE * NPTEPG - 1);
2920 if (kstart - 1 >= kernel_map.max_offset) {
2921 kstart = kernel_map.max_offset;
2930 * This index is bogus, but out of the way
2932 nkpg = vm_page_alloc(NULL, nkpt,
2935 VM_ALLOC_INTERRUPT);
2937 panic("pmap_growkernel: no memory to grow kernel");
2940 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2941 pmap_zero_page(ptppaddr);
2942 newpt = (pd_entry_t) (ptppaddr |
2943 kernel_pmap.pmap_bits[PG_V_IDX] |
2944 kernel_pmap.pmap_bits[PG_RW_IDX] |
2945 kernel_pmap.pmap_bits[PG_A_IDX] |
2946 kernel_pmap.pmap_bits[PG_M_IDX]);
2947 *pmap_pt(&kernel_pmap, kstart) = newpt;
2950 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2951 ~(PAGE_SIZE * NPTEPG - 1);
2953 if (kstart - 1 >= kernel_map.max_offset) {
2954 kstart = kernel_map.max_offset;
2960 * Only update kernel_vm_end for areas below KERNBASE.
2962 if (update_kernel_vm_end && kernel_vm_end < kstart)
2963 kernel_vm_end = kstart;
2967 * Add a reference to the specified pmap.
2970 pmap_reference(pmap_t pmap)
2973 lwkt_gettoken(&pmap->pm_token);
2975 lwkt_reltoken(&pmap->pm_token);
2979 /***************************************************
2980 * page management routines.
2981 ***************************************************/
2984 * Hold a pv without locking it
2987 pv_hold(pv_entry_t pv)
2989 atomic_add_int(&pv->pv_hold, 1);
2993 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2994 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2997 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2998 * pv list via its page) must be held by the caller.
3001 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3006 * Critical path shortcut expects pv to already have one ref
3007 * (for the pv->pv_pmap).
3009 if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3012 pv->pv_line = lineno;
3018 count = pv->pv_hold;
3020 if ((count & PV_HOLD_LOCKED) == 0) {
3021 if (atomic_cmpset_int(&pv->pv_hold, count,
3022 (count + 1) | PV_HOLD_LOCKED)) {
3025 pv->pv_line = lineno;
3030 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3038 * Drop a previously held pv_entry which could not be locked, allowing its
3041 * Must not be called with a spinlock held as we might zfree() the pv if it
3042 * is no longer associated with a pmap and this was the last hold count.
3045 pv_drop(pv_entry_t pv)
3050 count = pv->pv_hold;
3052 KKASSERT((count & PV_HOLD_MASK) > 0);
3053 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3054 (PV_HOLD_LOCKED | 1));
3055 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3056 if ((count & PV_HOLD_MASK) == 1) {
3058 if (pmap_enter_debug > 0) {
3060 kprintf("pv_drop: free pv %p\n", pv);
3063 KKASSERT(count == 1);
3064 KKASSERT(pv->pv_pmap == NULL);
3074 * Find or allocate the requested PV entry, returning a locked, held pv.
3076 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3077 * for the caller and one representing the pmap and vm_page association.
3079 * If (*isnew) is zero, the returned pv will have only one hold count.
3081 * Since both associations can only be adjusted while the pv is locked,
3082 * together they represent just one additional hold.
3086 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3089 pv_entry_t pnew = NULL;
3091 spin_lock(&pmap->pm_spin);
3093 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3094 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3099 spin_unlock(&pmap->pm_spin);
3100 pnew = zalloc(pvzone);
3101 spin_lock(&pmap->pm_spin);
3104 pnew->pv_pmap = pmap;
3105 pnew->pv_pindex = pindex;
3106 pnew->pv_hold = PV_HOLD_LOCKED | 2;
3108 pnew->pv_func = func;
3109 pnew->pv_line = lineno;
3111 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3112 ++pmap->pm_generation;
3113 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3114 spin_unlock(&pmap->pm_spin);
3119 spin_unlock(&pmap->pm_spin);
3120 zfree(pvzone, pnew);
3122 spin_lock(&pmap->pm_spin);
3125 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3126 spin_unlock(&pmap->pm_spin);
3128 spin_unlock(&pmap->pm_spin);
3129 _pv_lock(pv PMAP_DEBUG_COPY);
3131 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3136 spin_lock(&pmap->pm_spin);
3141 * Find the requested PV entry, returning a locked+held pv or NULL
3145 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
3149 spin_lock(&pmap->pm_spin);
3154 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3155 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3159 spin_unlock(&pmap->pm_spin);
3162 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3163 spin_unlock(&pmap->pm_spin);
3165 spin_unlock(&pmap->pm_spin);
3166 _pv_lock(pv PMAP_DEBUG_COPY);
3168 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3169 pv_cache(pv, pindex);
3173 spin_lock(&pmap->pm_spin);
3178 * Lookup, hold, and attempt to lock (pmap,pindex).
3180 * If the entry does not exist NULL is returned and *errorp is set to 0
3182 * If the entry exists and could be successfully locked it is returned and
3183 * errorp is set to 0.
3185 * If the entry exists but could NOT be successfully locked it is returned
3186 * held and *errorp is set to 1.
3190 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
3194 spin_lock_shared(&pmap->pm_spin);
3195 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3196 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3198 spin_unlock_shared(&pmap->pm_spin);
3202 if (pv_hold_try(pv)) {
3203 pv_cache(pv, pindex);
3204 spin_unlock_shared(&pmap->pm_spin);
3206 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3207 return(pv); /* lock succeeded */
3209 spin_unlock_shared(&pmap->pm_spin);
3211 return (pv); /* lock failed */
3215 * Find the requested PV entry, returning a held pv or NULL
3219 pv_find(pmap_t pmap, vm_pindex_t pindex)
3223 spin_lock_shared(&pmap->pm_spin);
3225 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3226 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3228 spin_unlock_shared(&pmap->pm_spin);
3232 pv_cache(pv, pindex);
3233 spin_unlock_shared(&pmap->pm_spin);
3238 * Lock a held pv, keeping the hold count
3242 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3247 count = pv->pv_hold;
3249 if ((count & PV_HOLD_LOCKED) == 0) {
3250 if (atomic_cmpset_int(&pv->pv_hold, count,
3251 count | PV_HOLD_LOCKED)) {
3254 pv->pv_line = lineno;
3260 tsleep_interlock(pv, 0);
3261 if (atomic_cmpset_int(&pv->pv_hold, count,
3262 count | PV_HOLD_WAITING)) {
3264 kprintf("pv waiting on %s:%d\n",
3265 pv->pv_func, pv->pv_line);
3267 tsleep(pv, PINTERLOCKED, "pvwait", hz);
3274 * Unlock a held and locked pv, keeping the hold count.
3278 pv_unlock(pv_entry_t pv)
3283 count = pv->pv_hold;
3285 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3286 (PV_HOLD_LOCKED | 1));
3287 if (atomic_cmpset_int(&pv->pv_hold, count,
3289 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3290 if (count & PV_HOLD_WAITING)
3298 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3299 * and the hold count drops to zero we will free it.
3301 * Caller should not hold any spin locks. We are protected from hold races
3302 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3303 * lock held. A pv cannot be located otherwise.
3307 pv_put(pv_entry_t pv)
3310 if (pmap_enter_debug > 0) {
3312 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3317 * Fast - shortcut most common condition
3319 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3330 * Remove the pmap association from a pv, require that pv_m already be removed,
3331 * then unlock and drop the pv. Any pte operations must have already been
3332 * completed. This call may result in a last-drop which will physically free
3335 * Removing the pmap association entails an additional drop.
3337 * pv must be exclusively locked on call and will be disposed of on return.
3341 pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway)
3345 KKASSERT(pv->pv_m == NULL);
3346 KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3347 if ((pmap = pv->pv_pmap) != NULL) {
3348 spin_lock(&pmap->pm_spin);
3349 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3350 ++pmap->pm_generation;
3351 if (pmap->pm_pvhint == pv)
3352 pmap->pm_pvhint = NULL;
3353 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3356 spin_unlock(&pmap->pm_spin);
3359 * Try to shortcut three atomic ops, otherwise fall through
3360 * and do it normally. Drop two refs and the lock all in
3364 atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3366 if (pmap_enter_debug > 0) {
3368 kprintf("pv_free: free pv %p\n", pv);
3373 vm_page_unwire_quick(pvp->pv_m);
3376 pv_drop(pv); /* ref for pv_pmap */
3381 vm_page_unwire_quick(pvp->pv_m);
3385 * This routine is very drastic, but can save the system
3393 static int warningdone=0;
3395 if (pmap_pagedaemon_waken == 0)
3397 pmap_pagedaemon_waken = 0;
3398 if (warningdone < 5) {
3399 kprintf("pmap_collect: collecting pv entries -- "
3400 "suggest increasing PMAP_SHPGPERPROC\n");
3404 for (i = 0; i < vm_page_array_size; i++) {
3405 m = &vm_page_array[i];
3406 if (m->wire_count || m->hold_count)
3408 if (vm_page_busy_try(m, TRUE) == 0) {
3409 if (m->wire_count == 0 && m->hold_count == 0) {
3418 * Scan the pmap for active page table entries and issue a callback.
3419 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3420 * its parent page table.
3422 * pte_pv will be NULL if the page or page table is unmanaged.
3423 * pt_pv will point to the page table page containing the pte for the page.
3425 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3426 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3427 * process pmap's PD and page to the callback function. This can be
3428 * confusing because the pt_pv is really a pd_pv, and the target page
3429 * table page is simply aliased by the pmap and not owned by it.
3431 * It is assumed that the start and end are properly rounded to the page size.
3433 * It is assumed that PD pages and above are managed and thus in the RB tree,
3434 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3436 struct pmap_scan_info {
3440 vm_pindex_t sva_pd_pindex;
3441 vm_pindex_t eva_pd_pindex;
3442 void (*func)(pmap_t, struct pmap_scan_info *,
3443 pv_entry_t, pv_entry_t, int, vm_offset_t,
3444 pt_entry_t *, void *);
3446 pmap_inval_bulk_t bulk_core;
3447 pmap_inval_bulk_t *bulk;
3452 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3453 static int pmap_scan_callback(pv_entry_t pv, void *data);
3456 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3458 struct pmap *pmap = info->pmap;
3459 pv_entry_t pd_pv; /* A page directory PV */
3460 pv_entry_t pt_pv; /* A page table PV */
3461 pv_entry_t pte_pv; /* A page table entry PV */
3464 struct pv_entry dummy_pv;
3471 info->bulk = &info->bulk_core;
3472 pmap_inval_bulk_init(&info->bulk_core, pmap);
3478 * Hold the token for stability; if the pmap is empty we have nothing
3481 lwkt_gettoken(&pmap->pm_token);
3483 if (pmap->pm_stats.resident_count == 0) {
3484 lwkt_reltoken(&pmap->pm_token);
3493 * Special handling for scanning one page, which is a very common
3494 * operation (it is?).
3496 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3498 if (info->sva + PAGE_SIZE == info->eva) {
3499 generation = pmap->pm_generation;
3500 if (info->sva >= VM_MAX_USER_ADDRESS) {
3502 * Kernel mappings do not track wire counts on
3503 * page table pages and only maintain pd_pv and
3504 * pte_pv levels so pmap_scan() works.
3507 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3508 ptep = vtopte(info->sva);
3511 * User pages which are unmanaged will not have a
3512 * pte_pv. User page table pages which are unmanaged
3513 * (shared from elsewhere) will also not have a pt_pv.
3514 * The func() callback will pass both pte_pv and pt_pv
3515 * as NULL in that case.
3517 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3518 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3519 if (pt_pv == NULL) {
3520 KKASSERT(pte_pv == NULL);
3521 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3523 ptep = pv_pte_lookup(pd_pv,
3524 pmap_pt_index(info->sva));
3526 info->func(pmap, info,
3535 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3539 * NOTE: *ptep can't be ripped out from under us if we hold
3540 * pte_pv locked, but bits can change. However, there is
3541 * a race where another thread may be inserting pte_pv
3542 * and setting *ptep just after our pte_pv lookup fails.
3544 * In this situation we can end up with a NULL pte_pv
3545 * but find that we have a managed *ptep. We explicitly
3546 * check for this race.
3552 * Unlike the pv_find() case below we actually
3553 * acquired a locked pv in this case so any
3554 * race should have been resolved. It is expected
3557 KKASSERT(pte_pv == NULL);
3558 } else if (pte_pv) {
3559 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3560 pmap->pmap_bits[PG_V_IDX])) ==
3561 (pmap->pmap_bits[PG_MANAGED_IDX] |
3562 pmap->pmap_bits[PG_V_IDX]),
3563 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3565 *ptep, oldpte, info->sva, pte_pv,
3566 generation, pmap->pm_generation));
3567 info->func(pmap, info, pte_pv, pt_pv, 0,
3568 info->sva, ptep, info->arg);
3571 * Check for insertion race
3573 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3575 pte_pv = pv_find(pmap,
3576 pmap_pte_pindex(info->sva));
3580 kprintf("pmap_scan: RACE1 "
3590 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3591 pmap->pmap_bits[PG_V_IDX])) ==
3592 pmap->pmap_bits[PG_V_IDX],
3593 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3595 *ptep, oldpte, info->sva,
3596 generation, pmap->pm_generation));
3597 info->func(pmap, info, NULL, pt_pv, 0,
3598 info->sva, ptep, info->arg);
3603 pmap_inval_bulk_flush(info->bulk);
3604 lwkt_reltoken(&pmap->pm_token);
3609 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3612 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3613 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3615 if (info->sva >= VM_MAX_USER_ADDRESS) {
3617 * The kernel does not currently maintain any pv_entry's for
3618 * higher-level page tables.
3620 bzero(&dummy_pv, sizeof(dummy_pv));
3621 dummy_pv.pv_pindex = info->sva_pd_pindex;
3622 spin_lock(&pmap->pm_spin);
3623 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3624 pmap_scan_callback(&dummy_pv, info);
3625 ++dummy_pv.pv_pindex;
3627 spin_unlock(&pmap->pm_spin);
3630 * User page tables maintain local PML4, PDP, and PD
3631 * pv_entry's at the very least. PT pv's might be
3632 * unmanaged and thus not exist. PTE pv's might be
3633 * unmanaged and thus not exist.
3635 spin_lock(&pmap->pm_spin);
3636 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3637 pmap_scan_cmp, pmap_scan_callback, info);
3638 spin_unlock(&pmap->pm_spin);
3640 pmap_inval_bulk_flush(info->bulk);
3641 lwkt_reltoken(&pmap->pm_token);
3645 * WARNING! pmap->pm_spin held
3648 pmap_scan_cmp(pv_entry_t pv, void *data)
3650 struct pmap_scan_info *info = data;
3651 if (pv->pv_pindex < info->sva_pd_pindex)
3653 if (pv->pv_pindex >= info->eva_pd_pindex)
3659 * WARNING! pmap->pm_spin held
3662 pmap_scan_callback(pv_entry_t pv, void *data)
3664 struct pmap_scan_info *info = data;
3665 struct pmap *pmap = info->pmap;
3666 pv_entry_t pd_pv; /* A page directory PV */
3667 pv_entry_t pt_pv; /* A page table PV */
3668 pv_entry_t pte_pv; /* A page table entry PV */
3673 vm_offset_t va_next;
3674 vm_pindex_t pd_pindex;
3685 * Pull the PD pindex from the pv before releasing the spinlock.
3687 * WARNING: pv is faked for kernel pmap scans.
3689 pd_pindex = pv->pv_pindex;
3690 spin_unlock(&pmap->pm_spin);
3691 pv = NULL; /* invalid after spinlock unlocked */
3694 * Calculate the page range within the PD. SIMPLE pmaps are
3695 * direct-mapped for the entire 2^64 address space. Normal pmaps
3696 * reflect the user and kernel address space which requires
3697 * cannonicalization w/regards to converting pd_pindex's back
3700 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3701 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3702 (sva & PML4_SIGNMASK)) {
3703 sva |= PML4_SIGNMASK;
3705 eva = sva + NBPDP; /* can overflow */
3706 if (sva < info->sva)
3708 if (eva < info->sva || eva > info->eva)
3712 * NOTE: kernel mappings do not track page table pages, only
3715 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3716 * However, for the scan to be efficient we try to
3717 * cache items top-down.
3722 for (; sva < eva; sva = va_next) {
3725 if (sva >= VM_MAX_USER_ADDRESS) {
3734 * PD cache (degenerate case if we skip). It is possible
3735 * for the PD to not exist due to races. This is ok.
3737 if (pd_pv == NULL) {
3738 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3739 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3741 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3743 if (pd_pv == NULL) {
3744 va_next = (sva + NBPDP) & ~PDPMASK;
3753 if (pt_pv == NULL) {
3754 vm_page_wire_quick(pd_pv->pv_m);
3756 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3758 vm_page_unwire_quick(pd_pv->pv_m);
3759 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3760 vm_page_wire_quick(pd_pv->pv_m);
3763 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3765 vm_page_unwire_quick(pd_pv->pv_m);
3769 * If pt_pv is NULL we either have an shared page table
3770 * page and must issue a callback specific to that case,
3771 * or there is no page table page.
3773 * Either way we can skip the page table page.
3775 if (pt_pv == NULL) {
3777 * Possible unmanaged (shared from another pmap)
3780 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3781 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3782 info->func(pmap, info, NULL, pd_pv, 1,
3783 sva, ptep, info->arg);
3787 * Done, move to next page table page.
3789 va_next = (sva + NBPDR) & ~PDRMASK;
3796 * From this point in the loop testing pt_pv for non-NULL
3797 * means we are in UVM, else if it is NULL we are in KVM.
3799 * Limit our scan to either the end of the va represented
3800 * by the current page table page, or to the end of the
3801 * range being removed.
3804 va_next = (sva + NBPDR) & ~PDRMASK;
3811 * Scan the page table for pages. Some pages may not be
3812 * managed (might not have a pv_entry).
3814 * There is no page table management for kernel pages so
3815 * pt_pv will be NULL in that case, but otherwise pt_pv
3816 * is non-NULL, locked, and referenced.
3820 * At this point a non-NULL pt_pv means a UVA, and a NULL
3821 * pt_pv means a KVA.
3824 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3828 while (sva < va_next) {
3830 * Yield every 64 pages, stop if requested.
3832 if ((++info->count & 63) == 0)
3838 * Check if pt_pv has been lost (probably due to
3839 * a remove of the underlying pages).
3841 if (pt_pv && pt_pv->pv_pmap == NULL)
3845 * Acquire the related pte_pv, if any. If *ptep == 0
3846 * the related pte_pv should not exist, but if *ptep
3847 * is not zero the pte_pv may or may not exist (e.g.
3848 * will not exist for an unmanaged page).
3850 * However a multitude of races are possible here.
3852 * In addition, the (pt_pv, pte_pv) lock order is
3853 * backwards, so we have to be careful in aquiring
3854 * a properly locked pte_pv.
3856 generation = pmap->pm_generation;
3858 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3862 vm_page_wire_quick(pd_pv->pv_m);
3865 vm_page_wire_quick(pt_pv->pv_m);
3866 pv_unlock(pt_pv);/* must be non-NULL */
3867 pv_lock(pte_pv); /* safe to block now */
3871 vm_page_unwire_quick(pt_pv->pv_m);
3874 * pt_pv reloaded, need new ptep
3876 KKASSERT(pt_pv != NULL);
3877 ptep = pv_pte_lookup(pt_pv,
3878 pmap_pte_index(sva));
3881 vm_page_unwire_quick(pd_pv->pv_m);
3886 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3890 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3895 kprintf("Unexpected non-NULL pte_pv "
3897 "*ptep = %016lx/%016lx\n",
3898 pte_pv, pt_pv, *ptep, oldpte);
3899 panic("Unexpected non-NULL pte_pv");
3907 * Ready for the callback. The locked pte_pv (if any)
3908 * is consumed by the callback. pte_pv will exist if
3909 * the page is managed, and will not exist if it
3913 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3914 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3915 ("badC *ptep %016lx/%016lx sva %016lx "
3916 "pte_pv %p pm_generation %d/%d",
3917 *ptep, oldpte, sva, pte_pv,
3918 generation, pmap->pm_generation));
3920 * We must unlock pd_pv across the callback
3921 * to avoid deadlocks on any recursive
3922 * disposal. Re-check that it still exists
3927 info->func(pmap, info, pte_pv, pt_pv, 0,
3928 sva, ptep, info->arg);
3931 if (pd_pv->pv_pmap == NULL) {
3938 * Check for insertion race. Since there is no
3939 * pte_pv to guard us it is possible for us
3940 * to race another thread doing an insertion.
3941 * Our lookup misses the pte_pv but our *ptep
3942 * check sees the inserted pte.
3944 * XXX panic case seems to occur within a
3945 * vm_fork() of /bin/sh, which frankly
3946 * shouldn't happen since no other threads
3947 * should be inserting to our pmap in that
3948 * situation. Removing, possibly. Inserting,
3951 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3953 pte_pv = pv_find(pmap,
3954 pmap_pte_pindex(sva));
3957 kprintf("pmap_scan: RACE2 "
3967 * We must unlock pd_pv across the callback
3968 * to avoid deadlocks on any recursive
3969 * disposal. Re-check that it still exists
3972 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3973 pmap->pmap_bits[PG_V_IDX],
3974 ("badD *ptep %016lx/%016lx sva %016lx "
3975 "pte_pv NULL pm_generation %d/%d",
3977 generation, pmap->pm_generation));
3980 info->func(pmap, info, NULL, pt_pv, 0,
3981 sva, ptep, info->arg);
3984 if (pd_pv->pv_pmap == NULL) {
4003 if ((++info->count & 7) == 0)
4007 * Relock before returning.
4009 spin_lock(&pmap->pm_spin);
4014 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4016 struct pmap_scan_info info;
4021 info.func = pmap_remove_callback;
4023 pmap_scan(&info, 1);
4027 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4029 struct pmap_scan_info info;
4034 info.func = pmap_remove_callback;
4036 pmap_scan(&info, 0);
4040 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4041 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4042 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4048 * This will also drop pt_pv's wire_count. Note that
4049 * terminal pages are not wired based on mmu presence.
4051 * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4053 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4054 pte_pv = NULL; /* safety */
4057 * Recursively destroy higher-level page tables.
4059 * This is optional. If we do not, they will still
4060 * be destroyed when the process exits.
4062 if (pmap_dynamic_delete &&
4065 pt_pv->pv_m->wire_count == 1 &&
4066 pt_pv->pv_pindex != pmap_pml4_pindex()) {
4068 pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4071 } else if (sharept == 0) {
4073 * Unmanaged page table (pt, pd, or pdp. Not pte).
4075 * pt_pv's wire_count is still bumped by unmanaged pages
4076 * so we must decrement it manually.
4078 * We have to unwire the target page table page.
4080 * It is unclear how we can invalidate a segment so we
4081 * invalidate -1 which invlidates the tlb.
4083 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4084 if (pte & pmap->pmap_bits[PG_W_IDX])
4085 atomic_add_long(&pmap->pm_stats.wired_count, -1);
4086 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4087 if (vm_page_unwire_quick(pt_pv->pv_m))
4088 panic("pmap_remove: insufficient wirecount");
4091 * Unmanaged page table (pt, pd, or pdp. Not pte) for
4092 * a shared page table.
4094 * pt_pv is actually the pd_pv for our pmap (not the shared
4097 * We have to unwire the target page table page and we
4098 * have to unwire our page directory page.
4100 * It is unclear how we can invalidate a segment so we
4101 * invalidate -1 which invlidates the tlb.
4103 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4104 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4105 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4106 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4107 panic("pmap_remove: shared pgtable1 bad wirecount");
4108 if (vm_page_unwire_quick(pt_pv->pv_m))
4109 panic("pmap_remove: shared pgtable2 bad wirecount");
4114 * Removes this physical page from all physical maps in which it resides.
4115 * Reflects back modify bits to the pager.
4117 * This routine may not be called from an interrupt.
4121 pmap_remove_all(vm_page_t m)
4124 pmap_inval_bulk_t bulk;
4126 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4129 vm_page_spin_lock(m);
4130 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4131 KKASSERT(pv->pv_m == m);
4132 if (pv_hold_try(pv)) {
4133 vm_page_spin_unlock(m);
4135 vm_page_spin_unlock(m);
4138 if (pv->pv_m != m) {
4140 vm_page_spin_lock(m);
4145 * Holding no spinlocks, pv is locked.
4147 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4148 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4149 pv = NULL; /* safety */
4150 pmap_inval_bulk_flush(&bulk);
4152 pmap_remove_pv_page(pv);
4155 vm_page_spin_lock(m);
4157 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4158 vm_page_spin_unlock(m);
4162 * Removes the page from a particular pmap
4165 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4168 pmap_inval_bulk_t bulk;
4170 if (!pmap_initialized)
4174 vm_page_spin_lock(m);
4175 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4176 if (pv->pv_pmap != pmap)
4178 KKASSERT(pv->pv_m == m);
4179 if (pv_hold_try(pv)) {
4180 vm_page_spin_unlock(m);
4182 vm_page_spin_unlock(m);
4185 if (pv->pv_m != m) {
4191 * Holding no spinlocks, pv is locked.
4193 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4194 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4195 pv = NULL; /* safety */
4196 pmap_inval_bulk_flush(&bulk);
4198 pmap_remove_pv_page(pv);
4203 vm_page_spin_unlock(m);
4207 * Set the physical protection on the specified range of this map
4208 * as requested. This function is typically only used for debug watchpoints
4211 * This function may not be called from an interrupt if the map is
4212 * not the kernel_pmap.
4214 * NOTE! For shared page table pages we just unmap the page.
4217 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4219 struct pmap_scan_info info;
4220 /* JG review for NX */
4224 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4225 pmap_remove(pmap, sva, eva);
4228 if (prot & VM_PROT_WRITE)
4233 info.func = pmap_protect_callback;
4235 pmap_scan(&info, 1);
4240 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4241 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4242 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4254 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4255 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4256 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4257 KKASSERT(m == pte_pv->pv_m);
4258 vm_page_flag_set(m, PG_REFERENCED);
4260 cbits &= ~pmap->pmap_bits[PG_A_IDX];
4262 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4263 if (pmap_track_modified(pte_pv->pv_pindex)) {
4264 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4266 m = PHYS_TO_VM_PAGE(pbits &
4271 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4274 } else if (sharept) {
4276 * Unmanaged page table, pt_pv is actually the pd_pv
4277 * for our pmap (not the object's shared pmap).
4279 * When asked to protect something in a shared page table
4280 * page we just unmap the page table page. We have to
4281 * invalidate the tlb in this situation.
4283 * XXX Warning, shared page tables will not be used for
4284 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4285 * so PHYS_TO_VM_PAGE() should be safe here.
4287 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4288 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4289 panic("pmap_protect: pgtable1 pg bad wirecount");
4290 if (vm_page_unwire_quick(pt_pv->pv_m))
4291 panic("pmap_protect: pgtable2 pg bad wirecount");
4294 /* else unmanaged page, adjust bits, no wire changes */
4297 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4299 if (pmap_enter_debug > 0) {
4301 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4302 "pt_pv=%p cbits=%08lx\n",
4308 if (pbits != cbits) {
4309 if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4310 ptep, pbits, cbits)) {
4320 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4321 * mapping at that address. Set protection and wiring as requested.
4323 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4324 * possible. If it is we enter the page into the appropriate shared pmap
4325 * hanging off the related VM object instead of the passed pmap, then we
4326 * share the page table page from the VM object's pmap into the current pmap.
4328 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4332 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4333 boolean_t wired, vm_map_entry_t entry)
4335 pv_entry_t pt_pv; /* page table */
4336 pv_entry_t pte_pv; /* page table entry */
4339 pt_entry_t origpte, newpte;
4344 va = trunc_page(va);
4345 #ifdef PMAP_DIAGNOSTIC
4347 panic("pmap_enter: toobig");
4348 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4349 panic("pmap_enter: invalid to pmap_enter page table "
4350 "pages (va: 0x%lx)", va);
4352 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4353 kprintf("Warning: pmap_enter called on UVA with "
4356 db_print_backtrace();
4359 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4360 kprintf("Warning: pmap_enter called on KVA without"
4363 db_print_backtrace();
4368 * Get locked PV entries for our new page table entry (pte_pv)
4369 * and for its parent page table (pt_pv). We need the parent
4370 * so we can resolve the location of the ptep.
4372 * Only hardware MMU actions can modify the ptep out from
4375 * if (m) is fictitious or unmanaged we do not create a managing
4376 * pte_pv for it. Any pre-existing page's management state must
4377 * match (avoiding code complexity).
4379 * If the pmap is still being initialized we assume existing
4382 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4384 if (pmap_initialized == FALSE) {
4389 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4391 if (va >= VM_MAX_USER_ADDRESS) {
4395 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4397 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4401 KASSERT(origpte == 0 ||
4402 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4403 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4405 if (va >= VM_MAX_USER_ADDRESS) {
4407 * Kernel map, pv_entry-tracked.
4410 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4416 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4418 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4422 KASSERT(origpte == 0 ||
4423 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4424 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4427 pa = VM_PAGE_TO_PHYS(m);
4428 opa = origpte & PG_FRAME;
4430 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4431 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4433 newpte |= pmap->pmap_bits[PG_W_IDX];
4434 if (va < VM_MAX_USER_ADDRESS)
4435 newpte |= pmap->pmap_bits[PG_U_IDX];
4437 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4438 // if (pmap == &kernel_pmap)
4439 // newpte |= pgeflag;
4440 newpte |= pmap->pmap_cache_bits[m->pat_mode];
4441 if (m->flags & PG_FICTITIOUS)
4442 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4445 * It is possible for multiple faults to occur in threaded
4446 * environments, the existing pte might be correct.
4448 if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4449 pmap->pmap_bits[PG_A_IDX])) == 0)
4453 * Ok, either the address changed or the protection or wiring
4456 * Clear the current entry, interlocking the removal. For managed
4457 * pte's this will also flush the modified state to the vm_page.
4458 * Atomic ops are mandatory in order to ensure that PG_M events are
4459 * not lost during any transition.
4461 * WARNING: The caller has busied the new page but not the original
4462 * vm_page which we are trying to replace. Because we hold
4463 * the pte_pv lock, but have not busied the page, PG bits
4464 * can be cleared out from under us.
4469 * pt_pv won't exist for a kernel page (managed or
4472 if (prot & VM_PROT_NOSYNC) {
4473 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4475 pmap_inval_bulk_t bulk;
4477 pmap_inval_bulk_init(&bulk, pmap);
4478 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4479 pmap_inval_bulk_flush(&bulk);
4482 pmap_remove_pv_page(pte_pv);
4483 } else if (prot & VM_PROT_NOSYNC) {
4485 * Unmanaged page, NOSYNC (no mmu sync) requested.
4487 * Leave wire count on PT page intact.
4489 (void)pte_load_clear(ptep);
4490 cpu_invlpg((void *)va);
4491 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4494 * Unmanaged page, normal enter.
4496 * Leave wire count on PT page intact.
4498 pmap_inval_smp(pmap, va, 1, ptep, 0);
4499 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4501 KKASSERT(*ptep == 0);
4505 if (pmap_enter_debug > 0) {
4507 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4508 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4510 origpte, newpte, ptep,
4511 pte_pv, pt_pv, opa, prot);
4517 * Enter on the PV list if part of our managed memory.
4518 * Wiring of the PT page is already handled.
4520 KKASSERT(pte_pv->pv_m == NULL);
4521 vm_page_spin_lock(m);
4523 pmap_page_stats_adding(m);
4524 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4525 vm_page_flag_set(m, PG_MAPPED);
4526 vm_page_spin_unlock(m);
4527 } else if (pt_pv && opa == 0) {
4529 * We have to adjust the wire count on the PT page ourselves
4530 * for unmanaged entries. If opa was non-zero we retained
4531 * the existing wire count from the removal.
4533 vm_page_wire_quick(pt_pv->pv_m);
4537 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4539 * User VMAs do not because those will be zero->non-zero, so no
4540 * stale entries to worry about at this point.
4542 * For KVM there appear to still be issues. Theoretically we
4543 * should be able to scrap the interlocks entirely but we
4546 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4547 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4549 *(volatile pt_entry_t *)ptep = newpte;
4551 cpu_invlpg((void *)va);
4556 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4559 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4562 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4563 vm_page_flag_set(m, PG_WRITEABLE);
4566 * Unmanaged pages need manual resident_count tracking.
4568 if (pte_pv == NULL && pt_pv)
4569 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4575 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4576 (m->flags & PG_MAPPED));
4579 * Cleanup the pv entry, allowing other accessors.
4588 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4589 * This code also assumes that the pmap has no pre-existing entry for this
4592 * This code currently may only be used on user pmaps, not kernel_pmap.
4595 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4597 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4601 * Make a temporary mapping for a physical address. This is only intended
4602 * to be used for panic dumps.
4604 * The caller is responsible for calling smp_invltlb().
4607 pmap_kenter_temporary(vm_paddr_t pa, long i)
4609 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4610 return ((void *)crashdumpmap);
4613 #define MAX_INIT_PT (96)
4616 * This routine preloads the ptes for a given object into the specified pmap.
4617 * This eliminates the blast of soft faults on process startup and
4618 * immediately after an mmap.
4620 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4623 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4624 vm_object_t object, vm_pindex_t pindex,
4625 vm_size_t size, int limit)
4627 struct rb_vm_page_scan_info info;
4632 * We can't preinit if read access isn't set or there is no pmap
4635 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4639 * We can't preinit if the pmap is not the current pmap
4641 lp = curthread->td_lwp;
4642 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4646 * Misc additional checks
4648 psize = x86_64_btop(size);
4650 if ((object->type != OBJT_VNODE) ||
4651 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4652 (object->resident_page_count > MAX_INIT_PT))) {
4656 if (pindex + psize > object->size) {
4657 if (object->size < pindex)
4659 psize = object->size - pindex;
4666 * If everything is segment-aligned do not pre-init here. Instead
4667 * allow the normal vm_fault path to pass a segment hint to
4668 * pmap_enter() which will then use an object-referenced shared
4671 if ((addr & SEG_MASK) == 0 &&
4672 (ctob(psize) & SEG_MASK) == 0 &&
4673 (ctob(pindex) & SEG_MASK) == 0) {
4678 * Use a red-black scan to traverse the requested range and load
4679 * any valid pages found into the pmap.
4681 * We cannot safely scan the object's memq without holding the
4684 info.start_pindex = pindex;
4685 info.end_pindex = pindex + psize - 1;
4691 vm_object_hold_shared(object);
4692 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4693 pmap_object_init_pt_callback, &info);
4694 vm_object_drop(object);
4699 pmap_object_init_pt_callback(vm_page_t p, void *data)
4701 struct rb_vm_page_scan_info *info = data;
4702 vm_pindex_t rel_index;
4705 * don't allow an madvise to blow away our really
4706 * free pages allocating pv entries.
4708 if ((info->limit & MAP_PREFAULT_MADVISE) &&
4709 vmstats.v_free_count < vmstats.v_free_reserved) {
4714 * Ignore list markers and ignore pages we cannot instantly
4715 * busy (while holding the object token).
4717 if (p->flags & PG_MARKER)
4719 if (vm_page_busy_try(p, TRUE))
4721 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4722 (p->flags & PG_FICTITIOUS) == 0) {
4723 if ((p->queue - p->pc) == PQ_CACHE)
4724 vm_page_deactivate(p);
4725 rel_index = p->pindex - info->start_pindex;
4726 pmap_enter_quick(info->pmap,
4727 info->addr + x86_64_ptob(rel_index), p);
4735 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4738 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4741 * XXX This is safe only because page table pages are not freed.
4744 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4748 /*spin_lock(&pmap->pm_spin);*/
4749 if ((pte = pmap_pte(pmap, addr)) != NULL) {
4750 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4751 /*spin_unlock(&pmap->pm_spin);*/
4755 /*spin_unlock(&pmap->pm_spin);*/
4760 * Change the wiring attribute for a pmap/va pair. The mapping must already
4761 * exist in the pmap. The mapping may or may not be managed.
4763 * Wiring is not a hardware characteristic so there is no need to invalidate
4764 * TLB. However, in an SMP environment we must use a locked bus cycle to
4765 * update the pte (if we are not using the pmap_inval_*() API that is)...
4766 * it's ok to do this for simple wiring changes.
4769 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4770 vm_map_entry_t entry)
4778 lwkt_gettoken(&pmap->pm_token);
4779 if (pmap == &kernel_pmap) {
4781 * The kernel may have managed pages, but not managed
4784 ptep = pmap_pte_quick(pmap, va);
4786 if (wired && !pmap_pte_w(pmap, ptep))
4787 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4788 else if (!wired && pmap_pte_w(pmap, ptep))
4789 atomic_add_long(&pmap->pm_stats.wired_count, -1);
4792 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4794 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4797 * Userland, the pmap of the possibly shared segment might
4800 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL,
4802 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4804 if (wired && !pmap_pte_w(pmap, ptep))
4805 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4806 else if (!wired && pmap_pte_w(pmap, ptep))
4807 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4810 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4812 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4815 lwkt_reltoken(&pmap->pm_token);
4821 * Copy the range specified by src_addr/len from the source map to
4822 * the range dst_addr/len in the destination map.
4824 * This routine is only advisory and need not do anything.
4827 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4828 vm_size_t len, vm_offset_t src_addr)
4835 * Zero the specified physical page.
4837 * This function may be called from an interrupt and no locking is
4841 pmap_zero_page(vm_paddr_t phys)
4843 vm_offset_t va = PHYS_TO_DMAP(phys);
4845 pagezero((void *)va);
4851 * Zero part of a physical page by mapping it into memory and clearing
4852 * its contents with bzero.
4854 * off and size may not cover an area beyond a single hardware page.
4857 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4859 vm_offset_t virt = PHYS_TO_DMAP(phys);
4861 bzero((char *)virt + off, size);
4867 * Copy the physical page from the source PA to the target PA.
4868 * This function may be called from an interrupt. No locking
4872 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4874 vm_offset_t src_virt, dst_virt;
4876 src_virt = PHYS_TO_DMAP(src);
4877 dst_virt = PHYS_TO_DMAP(dst);
4878 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4882 * pmap_copy_page_frag:
4884 * Copy the physical page from the source PA to the target PA.
4885 * This function may be called from an interrupt. No locking
4889 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4891 vm_offset_t src_virt, dst_virt;
4893 src_virt = PHYS_TO_DMAP(src);
4894 dst_virt = PHYS_TO_DMAP(dst);
4896 bcopy((char *)src_virt + (src & PAGE_MASK),
4897 (char *)dst_virt + (dst & PAGE_MASK),
4902 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4903 * this page. This count may be changed upwards or downwards in the future;
4904 * it is only necessary that true be returned for a small subset of pmaps
4905 * for proper page aging.
4908 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4913 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4916 vm_page_spin_lock(m);
4917 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4918 if (pv->pv_pmap == pmap) {
4919 vm_page_spin_unlock(m);
4926 vm_page_spin_unlock(m);
4931 * Remove all pages from specified address space this aids process exit
4932 * speeds. Also, this code may be special cased for the current process
4936 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4938 pmap_remove_noinval(pmap, sva, eva);
4943 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4944 * routines are inline, and a lot of things compile-time evaluate.
4948 pmap_testbit(vm_page_t m, int bit)
4954 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4957 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4959 vm_page_spin_lock(m);
4960 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4961 vm_page_spin_unlock(m);
4965 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4967 #if defined(PMAP_DIAGNOSTIC)
4968 if (pv->pv_pmap == NULL) {
4969 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4977 * If the bit being tested is the modified bit, then
4978 * mark clean_map and ptes as never
4981 * WARNING! Because we do not lock the pv, *pte can be in a
4982 * state of flux. Despite this the value of *pte
4983 * will still be related to the vm_page in some way
4984 * because the pv cannot be destroyed as long as we
4985 * hold the vm_page spin lock.
4987 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4988 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4989 if (!pmap_track_modified(pv->pv_pindex))
4993 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4994 if (*pte & pmap->pmap_bits[bit]) {
4995 vm_page_spin_unlock(m);
4999 vm_page_spin_unlock(m);
5004 * This routine is used to modify bits in ptes. Only one bit should be
5005 * specified. PG_RW requires special handling.
5007 * Caller must NOT hold any spin locks
5011 pmap_clearbit(vm_page_t m, int bit_index)
5018 if (bit_index == PG_RW_IDX)
5019 vm_page_flag_clear(m, PG_WRITEABLE);
5020 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5027 * Loop over all current mappings setting/clearing as appropos If
5028 * setting RO do we need to clear the VAC?
5030 * NOTE: When clearing PG_M we could also (not implemented) drop
5031 * through to the PG_RW code and clear PG_RW too, forcing
5032 * a fault on write to redetect PG_M for virtual kernels, but
5033 * it isn't necessary since virtual kernels invalidate the
5034 * pte when they clear the VPTE_M bit in their virtual page
5037 * NOTE: Does not re-dirty the page when clearing only PG_M.
5039 * NOTE: Because we do not lock the pv, *pte can be in a state of
5040 * flux. Despite this the value of *pte is still somewhat
5041 * related while we hold the vm_page spin lock.
5043 * *pte can be zero due to this race. Since we are clearing
5044 * bits we basically do no harm when this race ccurs.
5046 if (bit_index != PG_RW_IDX) {
5047 vm_page_spin_lock(m);
5048 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5049 #if defined(PMAP_DIAGNOSTIC)
5050 if (pv->pv_pmap == NULL) {
5051 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5057 pte = pmap_pte_quick(pv->pv_pmap,
5058 pv->pv_pindex << PAGE_SHIFT);
5060 if (pbits & pmap->pmap_bits[bit_index])
5061 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5063 vm_page_spin_unlock(m);
5068 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
5072 vm_page_spin_lock(m);
5073 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5075 * don't write protect pager mappings
5077 if (!pmap_track_modified(pv->pv_pindex))
5080 #if defined(PMAP_DIAGNOSTIC)
5081 if (pv->pv_pmap == NULL) {
5082 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5089 * Skip pages which do not have PG_RW set.
5091 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5092 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5098 if (pv_hold_try(pv)) {
5099 vm_page_spin_unlock(m);
5101 vm_page_spin_unlock(m);
5102 pv_lock(pv); /* held, now do a blocking lock */
5104 if (pv->pv_pmap != pmap || pv->pv_m != m) {
5105 pv_put(pv); /* and release */
5106 goto restart; /* anything could have happened */
5108 KKASSERT(pv->pv_pmap == pmap);
5114 nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5115 pmap->pmap_bits[PG_M_IDX]);
5116 if (pmap_inval_smp_cmpset(pmap,
5117 ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5118 pte, pbits, nbits)) {
5123 vm_page_spin_lock(m);
5126 * If PG_M was found to be set while we were clearing PG_RW
5127 * we also clear PG_M (done above) and mark the page dirty.
5128 * Callers expect this behavior.
5130 if (pbits & pmap->pmap_bits[PG_M_IDX])
5134 vm_page_spin_unlock(m);
5138 * Lower the permission for all mappings to a given page.
5140 * Page must be busied by caller. Because page is busied by caller this
5141 * should not be able to race a pmap_enter().
5144 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5146 /* JG NX support? */
5147 if ((prot & VM_PROT_WRITE) == 0) {
5148 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5150 * NOTE: pmap_clearbit(.. PG_RW) also clears
5151 * the PG_WRITEABLE flag in (m).
5153 pmap_clearbit(m, PG_RW_IDX);
5161 pmap_phys_address(vm_pindex_t ppn)
5163 return (x86_64_ptob(ppn));
5167 * Return a count of reference bits for a page, clearing those bits.
5168 * It is not necessary for every reference bit to be cleared, but it
5169 * is necessary that 0 only be returned when there are truly no
5170 * reference bits set.
5172 * XXX: The exact number of bits to check and clear is a matter that
5173 * should be tested and standardized at some point in the future for
5174 * optimal aging of shared pages.
5176 * This routine may not block.
5179 pmap_ts_referenced(vm_page_t m)
5186 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5189 vm_page_spin_lock(m);
5190 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5191 if (!pmap_track_modified(pv->pv_pindex))
5194 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5195 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5196 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5202 vm_page_spin_unlock(m);
5209 * Return whether or not the specified physical page was modified
5210 * in any physical maps.
5213 pmap_is_modified(vm_page_t m)
5217 res = pmap_testbit(m, PG_M_IDX);
5222 * Clear the modify bits on the specified physical page.
5225 pmap_clear_modify(vm_page_t m)
5227 pmap_clearbit(m, PG_M_IDX);
5231 * pmap_clear_reference:
5233 * Clear the reference bit on the specified physical page.
5236 pmap_clear_reference(vm_page_t m)
5238 pmap_clearbit(m, PG_A_IDX);
5242 * Miscellaneous support routines follow
5247 i386_protection_init(void)
5251 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
5252 kp = protection_codes;
5253 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5255 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5257 * Read access is also 0. There isn't any execute bit,
5258 * so just make it readable.
5260 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5261 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5262 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5265 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5266 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5267 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5268 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5269 *kp++ = pmap_bits_default[PG_RW_IDX];
5276 * Map a set of physical memory pages into the kernel virtual
5277 * address space. Return a pointer to where it is mapped. This
5278 * routine is intended to be used for mapping device memory,
5281 * NOTE: We can't use pgeflag unless we invalidate the pages one at
5284 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5285 * work whether the cpu supports PAT or not. The remaining PAT
5286 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5290 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5292 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5296 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5298 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5302 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5304 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5308 * Map a set of physical memory pages into the kernel virtual
5309 * address space. Return a pointer to where it is mapped. This
5310 * routine is intended to be used for mapping device memory,
5314 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5316 vm_offset_t va, tmpva, offset;
5320 offset = pa & PAGE_MASK;
5321 size = roundup(offset + size, PAGE_SIZE);
5323 va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5325 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5327 pa = pa & ~PAGE_MASK;
5328 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5329 pte = vtopte(tmpva);
5331 kernel_pmap.pmap_bits[PG_RW_IDX] |
5332 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5333 kernel_pmap.pmap_cache_bits[mode];
5334 tmpsize -= PAGE_SIZE;
5338 pmap_invalidate_range(&kernel_pmap, va, va + size);
5339 pmap_invalidate_cache_range(va, va + size);
5341 return ((void *)(va + offset));
5345 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5347 vm_offset_t base, offset;
5349 base = va & ~PAGE_MASK;
5350 offset = va & PAGE_MASK;
5351 size = roundup(offset + size, PAGE_SIZE);
5352 pmap_qremove(va, size >> PAGE_SHIFT);
5353 kmem_free(&kernel_map, base, size);
5357 * Sets the memory attribute for the specified page.
5360 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5366 * If "m" is a normal page, update its direct mapping. This update
5367 * can be relied upon to perform any cache operations that are
5368 * required for data coherence.
5370 if ((m->flags & PG_FICTITIOUS) == 0)
5371 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5375 * Change the PAT attribute on an existing kernel memory map. Caller
5376 * must ensure that the virtual memory in question is not accessed
5377 * during the adjustment.
5380 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5387 panic("pmap_change_attr: va is NULL");
5388 base = trunc_page(va);
5392 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5393 kernel_pmap.pmap_cache_bits[mode];
5398 changed = 1; /* XXX: not optimal */
5401 * Flush CPU caches if required to make sure any data isn't cached that
5402 * shouldn't be, etc.
5405 pmap_invalidate_range(&kernel_pmap, base, va);
5406 pmap_invalidate_cache_range(base, va);
5411 * perform the pmap work for mincore
5414 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5416 pt_entry_t *ptep, pte;
5420 lwkt_gettoken(&pmap->pm_token);
5421 ptep = pmap_pte(pmap, addr);
5423 if (ptep && (pte = *ptep) != 0) {
5426 val = MINCORE_INCORE;
5427 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5430 pa = pte & PG_FRAME;
5432 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5435 m = PHYS_TO_VM_PAGE(pa);
5440 if (pte & pmap->pmap_bits[PG_M_IDX])
5441 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5443 * Modified by someone
5445 else if (m && (m->dirty || pmap_is_modified(m)))
5446 val |= MINCORE_MODIFIED_OTHER;
5450 if (pte & pmap->pmap_bits[PG_A_IDX])
5451 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5454 * Referenced by someone
5456 else if (m && ((m->flags & PG_REFERENCED) ||
5457 pmap_ts_referenced(m))) {
5458 val |= MINCORE_REFERENCED_OTHER;
5459 vm_page_flag_set(m, PG_REFERENCED);
5463 lwkt_reltoken(&pmap->pm_token);
5469 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5470 * vmspace will be ref'd and the old one will be deref'd.
5472 * The vmspace for all lwps associated with the process will be adjusted
5473 * and cr3 will be reloaded if any lwp is the current lwp.
5475 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5478 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5480 struct vmspace *oldvm;
5483 oldvm = p->p_vmspace;
5484 if (oldvm != newvm) {
5487 p->p_vmspace = newvm;
5488 KKASSERT(p->p_nthreads == 1);
5489 lp = RB_ROOT(&p->p_lwp_tree);
5490 pmap_setlwpvm(lp, newvm);
5497 * Set the vmspace for a LWP. The vmspace is almost universally set the
5498 * same as the process vmspace, but virtual kernels need to swap out contexts
5499 * on a per-lwp basis.
5501 * Caller does not necessarily hold any vmspace tokens. Caller must control
5502 * the lwp (typically be in the context of the lwp). We use a critical
5503 * section to protect against statclock and hardclock (statistics collection).
5506 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5508 struct vmspace *oldvm;
5511 oldvm = lp->lwp_vmspace;
5513 if (oldvm != newvm) {
5515 lp->lwp_vmspace = newvm;
5516 if (curthread->td_lwp == lp) {
5517 pmap = vmspace_pmap(newvm);
5518 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5519 if (pmap->pm_active_lock & CPULOCK_EXCL)
5520 pmap_interlock_wait(newvm);
5521 #if defined(SWTCH_OPTIM_STATS)
5524 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5525 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5526 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5527 curthread->td_pcb->pcb_cr3 = KPML4phys;
5529 panic("pmap_setlwpvm: unknown pmap type\n");
5531 load_cr3(curthread->td_pcb->pcb_cr3);
5532 pmap = vmspace_pmap(oldvm);
5533 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5541 * Called when switching to a locked pmap, used to interlock against pmaps
5542 * undergoing modifications to prevent us from activating the MMU for the
5543 * target pmap until all such modifications have completed. We have to do
5544 * this because the thread making the modifications has already set up its
5545 * SMP synchronization mask.
5547 * This function cannot sleep!
5552 pmap_interlock_wait(struct vmspace *vm)
5554 struct pmap *pmap = &vm->vm_pmap;
5556 if (pmap->pm_active_lock & CPULOCK_EXCL) {
5558 KKASSERT(curthread->td_critcount >= 2);
5559 DEBUG_PUSH_INFO("pmap_interlock_wait");
5560 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5562 lwkt_process_ipiq();
5570 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5573 if ((obj == NULL) || (size < NBPDR) ||
5574 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5578 addr = roundup2(addr, NBPDR);
5583 * Used by kmalloc/kfree, page already exists at va
5586 pmap_kvtom(vm_offset_t va)
5588 pt_entry_t *ptep = vtopte(va);
5590 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5591 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5595 * Initialize machine-specific shared page directory support. This
5596 * is executed when a VM object is created.
5599 pmap_object_init(vm_object_t object)
5601 object->md.pmap_rw = NULL;
5602 object->md.pmap_ro = NULL;
5606 * Clean up machine-specific shared page directory support. This
5607 * is executed when a VM object is destroyed.
5610 pmap_object_free(vm_object_t object)
5614 if ((pmap = object->md.pmap_rw) != NULL) {
5615 object->md.pmap_rw = NULL;
5616 pmap_remove_noinval(pmap,
5617 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5618 CPUMASK_ASSZERO(pmap->pm_active);
5621 kfree(pmap, M_OBJPMAP);
5623 if ((pmap = object->md.pmap_ro) != NULL) {
5624 object->md.pmap_ro = NULL;
5625 pmap_remove_noinval(pmap,
5626 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5627 CPUMASK_ASSZERO(pmap->pm_active);
5630 kfree(pmap, M_OBJPMAP);
5635 * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5636 * VM page and issue a pginfo->callback.
5638 * We are expected to dispose of any non-NULL pte_pv.
5642 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5643 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
5644 vm_offset_t va, pt_entry_t *ptep, void *arg)
5646 struct pmap_pgscan_info *pginfo = arg;
5651 * Try to busy the page while we hold the pte_pv locked.
5653 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
5654 if (vm_page_busy_try(m, TRUE) == 0) {
5655 if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
5657 * The callback is issued with the pte_pv
5658 * unlocked and put away, and the pt_pv
5664 if (pginfo->callback(pginfo, va, m) < 0)
5673 ++pginfo->busycount;
5676 } else if (sharept) {
5677 /* shared page table */
5679 /* else unmanaged page */
5684 pmap_pgscan(struct pmap_pgscan_info *pginfo)
5686 struct pmap_scan_info info;
5688 pginfo->offset = pginfo->beg_addr;
5689 info.pmap = pginfo->pmap;
5690 info.sva = pginfo->beg_addr;
5691 info.eva = pginfo->end_addr;
5692 info.func = pmap_pgscan_callback;
5694 pmap_scan(&info, 0);
5696 pginfo->offset = pginfo->end_addr;