2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
9 * Copyright (c) 2011-2012 Matthew Dillon
10 * All rights reserved.
12 * This code is derived from software contributed to Berkeley by
13 * the Systems Programming Group of the University of Utah Computer
14 * Science Department and William Jolitz of UUNET Technologies Inc.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * Manage physical address maps for x86-64 systems.
49 #include "opt_disable_pse.h"
52 #include "opt_msgbuf.h"
54 #include <sys/param.h>
55 #include <sys/kernel.h>
57 #include <sys/msgbuf.h>
58 #include <sys/vmmeter.h>
60 #include <sys/systm.h>
63 #include <vm/vm_param.h>
64 #include <sys/sysctl.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_pageout.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_zone.h>
76 #include <sys/thread2.h>
77 #include <sys/sysref2.h>
78 #include <sys/spinlock2.h>
79 #include <vm/vm_page2.h>
81 #include <machine/cputypes.h>
82 #include <machine/md_var.h>
83 #include <machine/specialreg.h>
84 #include <machine/smp.h>
85 #include <machine_base/apic/apicreg.h>
86 #include <machine/globaldata.h>
87 #include <machine/pmap.h>
88 #include <machine/pmap_inval.h>
89 #include <machine/inttypes.h>
93 #define PMAP_KEEP_PDIRS
94 #ifndef PMAP_SHPGPERPROC
95 #define PMAP_SHPGPERPROC 2000
98 #if defined(DIAGNOSTIC)
99 #define PMAP_DIAGNOSTIC
105 * pmap debugging will report who owns a pv lock when blocking.
109 #define PMAP_DEBUG_DECL ,const char *func, int lineno
110 #define PMAP_DEBUG_ARGS , __func__, __LINE__
111 #define PMAP_DEBUG_COPY , func, lineno
113 #define pv_get(pmap, pindex) _pv_get(pmap, pindex \
115 #define pv_lock(pv) _pv_lock(pv \
117 #define pv_hold_try(pv) _pv_hold_try(pv \
119 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \
124 #define PMAP_DEBUG_DECL
125 #define PMAP_DEBUG_ARGS
126 #define PMAP_DEBUG_COPY
128 #define pv_get(pmap, pindex) _pv_get(pmap, pindex)
129 #define pv_lock(pv) _pv_lock(pv)
130 #define pv_hold_try(pv) _pv_hold_try(pv)
131 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp)
136 * Get PDEs and PTEs for user/kernel address space
138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
140 #define pmap_pde_v(pmap, pte) ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
141 #define pmap_pte_w(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
142 #define pmap_pte_m(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
143 #define pmap_pte_u(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
144 #define pmap_pte_v(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
147 * Given a map and a machine independent protection code,
148 * convert to a vax protection code.
150 #define pte_prot(m, p) \
151 (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
152 static int protection_codes[PROTECTION_CODES_SIZE];
154 struct pmap kernel_pmap;
155 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
159 vm_paddr_t avail_start; /* PA of first available physical page */
160 vm_paddr_t avail_end; /* PA of last available physical page */
161 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
162 vm_offset_t virtual2_end;
163 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
164 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
165 vm_offset_t KvaStart; /* VA start of KVA space */
166 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
167 vm_offset_t KvaSize; /* max size of kernel virtual address space */
168 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
169 //static int pgeflag; /* PG_G or-in */
170 //static int pseflag; /* PG_PS or-in */
174 static vm_paddr_t dmaplimit;
176 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
178 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE]; /* PAT -> PG_ bits */
179 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
181 static uint64_t KPTbase;
182 static uint64_t KPTphys;
183 static uint64_t KPDphys; /* phys addr of kernel level 2 */
184 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
185 uint64_t KPDPphys; /* phys addr of kernel level 3 */
186 uint64_t KPML4phys; /* phys addr of kernel level 4 */
188 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
189 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
192 * Data for the pv entry allocation mechanism
194 static vm_zone_t pvzone;
195 static struct vm_zone pvzone_store;
196 static struct vm_object pvzone_obj;
197 static int pv_entry_max=0, pv_entry_high_water=0;
198 static int pmap_pagedaemon_waken = 0;
199 static struct pv_entry *pvinit;
202 * All those kernel PT submaps that BSD is so fond of
204 pt_entry_t *CMAP1 = NULL, *ptmmap;
205 caddr_t CADDR1 = NULL, ptvmmap = NULL;
206 static pt_entry_t *msgbufmap;
207 struct msgbuf *msgbufp=NULL;
210 * PMAP default PG_* bits. Needed to be able to add
211 * EPT/NPT pagetable pmap_bits for the VMM module
213 uint64_t pmap_bits_default[] = {
214 REGULAR_PMAP, /* TYPE_IDX 0 */
215 X86_PG_V, /* PG_V_IDX 1 */
216 X86_PG_RW, /* PG_RW_IDX 2 */
217 X86_PG_U, /* PG_U_IDX 3 */
218 X86_PG_A, /* PG_A_IDX 4 */
219 X86_PG_M, /* PG_M_IDX 5 */
220 X86_PG_PS, /* PG_PS_IDX3 6 */
221 X86_PG_G, /* PG_G_IDX 7 */
222 X86_PG_AVAIL1, /* PG_AVAIL1_IDX 8 */
223 X86_PG_AVAIL2, /* PG_AVAIL2_IDX 9 */
224 X86_PG_AVAIL3, /* PG_AVAIL3_IDX 10 */
225 X86_PG_NC_PWT | X86_PG_NC_PCD, /* PG_N_IDX 11 */
230 static pt_entry_t *pt_crashdumpmap;
231 static caddr_t crashdumpmap;
234 static int pmap_enter_debug = 0;
235 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
236 &pmap_enter_debug, 0, "Debug pmap_enter's");
238 static int pmap_yield_count = 64;
239 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
240 &pmap_yield_count, 0, "Yield during init_pt/release");
241 static int pmap_mmu_optimize = 0;
242 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
243 &pmap_mmu_optimize, 0, "Share page table pages when possible");
244 int pmap_fast_kernel_cpusync = 0;
245 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
246 &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
250 /* Standard user access funtions */
251 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
253 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
254 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
255 extern int std_fubyte (const void *base);
256 extern int std_subyte (void *base, int byte);
257 extern long std_fuword (const void *base);
258 extern int std_suword (void *base, long word);
259 extern int std_suword32 (void *base, int word);
261 static void pv_hold(pv_entry_t pv);
262 static int _pv_hold_try(pv_entry_t pv
264 static void pv_drop(pv_entry_t pv);
265 static void _pv_lock(pv_entry_t pv
267 static void pv_unlock(pv_entry_t pv);
268 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
270 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
272 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
273 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
274 static void pv_put(pv_entry_t pv);
275 static void pv_free(pv_entry_t pv);
276 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
277 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
279 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
280 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
281 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, int smp);
282 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
283 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, int issmp);
285 struct pmap_scan_info;
286 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
287 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
288 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
289 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
290 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
291 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
293 static void i386_protection_init (void);
294 static void create_pagetables(vm_paddr_t *firstaddr);
295 static void pmap_remove_all (vm_page_t m);
296 static boolean_t pmap_testbit (vm_page_t m, int bit);
298 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
299 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
301 static void pmap_pinit_defaults(struct pmap *pmap);
303 static unsigned pdir4mb;
306 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
308 if (pv1->pv_pindex < pv2->pv_pindex)
310 if (pv1->pv_pindex > pv2->pv_pindex)
315 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
316 pv_entry_compare, vm_pindex_t, pv_pindex);
320 pmap_page_stats_adding(vm_page_t m)
322 globaldata_t gd = mycpu;
324 if (TAILQ_EMPTY(&m->md.pv_list)) {
325 ++gd->gd_vmtotal.t_arm;
326 } else if (TAILQ_FIRST(&m->md.pv_list) ==
327 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
328 ++gd->gd_vmtotal.t_armshr;
329 ++gd->gd_vmtotal.t_avmshr;
331 ++gd->gd_vmtotal.t_avmshr;
337 pmap_page_stats_deleting(vm_page_t m)
339 globaldata_t gd = mycpu;
341 if (TAILQ_EMPTY(&m->md.pv_list)) {
342 --gd->gd_vmtotal.t_arm;
343 } else if (TAILQ_FIRST(&m->md.pv_list) ==
344 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
345 --gd->gd_vmtotal.t_armshr;
346 --gd->gd_vmtotal.t_avmshr;
348 --gd->gd_vmtotal.t_avmshr;
353 * Move the kernel virtual free pointer to the next
354 * 2MB. This is used to help improve performance
355 * by using a large (2MB) page for much of the kernel
356 * (.text, .data, .bss)
360 pmap_kmem_choose(vm_offset_t addr)
362 vm_offset_t newaddr = addr;
364 newaddr = roundup2(addr, NBPDR);
371 * Super fast pmap_pte routine best used when scanning the pv lists.
372 * This eliminates many course-grained invltlb calls. Note that many of
373 * the pv list scans are across different pmaps and it is very wasteful
374 * to do an entire invltlb when checking a single mapping.
376 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
380 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
382 return pmap_pte(pmap, va);
386 * Returns the pindex of a page table entry (representing a terminal page).
387 * There are NUPTE_TOTAL page table entries possible (a huge number)
389 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
390 * We want to properly translate negative KVAs.
394 pmap_pte_pindex(vm_offset_t va)
396 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
400 * Returns the pindex of a page table.
404 pmap_pt_pindex(vm_offset_t va)
406 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
410 * Returns the pindex of a page directory.
414 pmap_pd_pindex(vm_offset_t va)
416 return (NUPTE_TOTAL + NUPT_TOTAL +
417 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
422 pmap_pdp_pindex(vm_offset_t va)
424 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
425 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
430 pmap_pml4_pindex(void)
432 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
436 * Return various clipped indexes for a given VA
438 * Returns the index of a pte in a page table, representing a terminal
443 pmap_pte_index(vm_offset_t va)
445 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
449 * Returns the index of a pt in a page directory, representing a page
454 pmap_pt_index(vm_offset_t va)
456 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
460 * Returns the index of a pd in a page directory page, representing a page
465 pmap_pd_index(vm_offset_t va)
467 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
471 * Returns the index of a pdp in the pml4 table, representing a page
476 pmap_pdp_index(vm_offset_t va)
478 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
482 * Generic procedure to index a pte from a pt, pd, or pdp.
484 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
485 * a page table page index but is instead of PV lookup index.
489 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
493 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
494 return(&pte[pindex]);
498 * Return pointer to PDP slot in the PML4
502 pmap_pdp(pmap_t pmap, vm_offset_t va)
504 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
508 * Return pointer to PD slot in the PDP given a pointer to the PDP
512 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
516 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
517 return (&pd[pmap_pd_index(va)]);
521 * Return pointer to PD slot in the PDP.
525 pmap_pd(pmap_t pmap, vm_offset_t va)
529 pdp = pmap_pdp(pmap, va);
530 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
532 return (pmap_pdp_to_pd(*pdp, va));
536 * Return pointer to PT slot in the PD given a pointer to the PD
540 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
544 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
545 return (&pt[pmap_pt_index(va)]);
549 * Return pointer to PT slot in the PD
551 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
552 * so we cannot lookup the PD via the PDP. Instead we
553 * must look it up via the pmap.
557 pmap_pt(pmap_t pmap, vm_offset_t va)
561 vm_pindex_t pd_pindex;
563 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
564 pd_pindex = pmap_pd_pindex(va);
565 spin_lock(&pmap->pm_spin);
566 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
567 spin_unlock(&pmap->pm_spin);
568 if (pv == NULL || pv->pv_m == NULL)
570 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
572 pd = pmap_pd(pmap, va);
573 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
575 return (pmap_pd_to_pt(*pd, va));
580 * Return pointer to PTE slot in the PT given a pointer to the PT
584 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
588 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
589 return (&pte[pmap_pte_index(va)]);
593 * Return pointer to PTE slot in the PT
597 pmap_pte(pmap_t pmap, vm_offset_t va)
601 pt = pmap_pt(pmap, va);
602 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
604 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
605 return ((pt_entry_t *)pt);
606 return (pmap_pt_to_pte(*pt, va));
610 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
611 * the PT layer. This will speed up core pmap operations considerably.
613 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
614 * must be in a known associated state (typically by being locked when
615 * the pmap spinlock isn't held). We allow the race for that case.
619 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
621 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
622 pv->pv_pmap->pm_pvhint = pv;
627 * Return address of PT slot in PD (KVM only)
629 * Cannot be used for user page tables because it might interfere with
630 * the shared page-table-page optimization (pmap_mmu_optimize).
634 vtopt(vm_offset_t va)
636 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
637 NPML4EPGSHIFT)) - 1);
639 return (PDmap + ((va >> PDRSHIFT) & mask));
643 * KVM - return address of PTE slot in PT
647 vtopte(vm_offset_t va)
649 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
650 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
652 return (PTmap + ((va >> PAGE_SHIFT) & mask));
656 allocpages(vm_paddr_t *firstaddr, long n)
661 bzero((void *)ret, n * PAGE_SIZE);
662 *firstaddr += n * PAGE_SIZE;
668 create_pagetables(vm_paddr_t *firstaddr)
670 long i; /* must be 64 bits */
676 * We are running (mostly) V=P at this point
678 * Calculate NKPT - number of kernel page tables. We have to
679 * accomodoate prealloction of the vm_page_array, dump bitmap,
680 * MSGBUF_SIZE, and other stuff. Be generous.
682 * Maxmem is in pages.
684 * ndmpdp is the number of 1GB pages we wish to map.
686 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
687 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
689 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
692 * Starting at the beginning of kvm (not KERNBASE).
694 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
695 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
696 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
697 ndmpdp) + 511) / 512;
701 * Starting at KERNBASE - map 2G worth of page table pages.
702 * KERNBASE is offset -2G from the end of kvm.
704 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
709 KPTbase = allocpages(firstaddr, nkpt_base);
710 KPTphys = allocpages(firstaddr, nkpt_phys);
711 KPML4phys = allocpages(firstaddr, 1);
712 KPDPphys = allocpages(firstaddr, NKPML4E);
713 KPDphys = allocpages(firstaddr, NKPDPE);
716 * Calculate the page directory base for KERNBASE,
717 * that is where we start populating the page table pages.
718 * Basically this is the end - 2.
720 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
722 DMPDPphys = allocpages(firstaddr, NDMPML4E);
723 if ((amd_feature & AMDID_PAGE1GB) == 0)
724 DMPDphys = allocpages(firstaddr, ndmpdp);
725 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
728 * Fill in the underlying page table pages for the area around
729 * KERNBASE. This remaps low physical memory to KERNBASE.
731 * Read-only from zero to physfree
732 * XXX not fully used, underneath 2M pages
734 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
735 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
736 ((pt_entry_t *)KPTbase)[i] |=
737 pmap_bits_default[PG_RW_IDX] |
738 pmap_bits_default[PG_V_IDX] |
739 pmap_bits_default[PG_G_IDX];
743 * Now map the initial kernel page tables. One block of page
744 * tables is placed at the beginning of kernel virtual memory,
745 * and another block is placed at KERNBASE to map the kernel binary,
746 * data, bss, and initial pre-allocations.
748 for (i = 0; i < nkpt_base; i++) {
749 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
750 ((pd_entry_t *)KPDbase)[i] |=
751 pmap_bits_default[PG_RW_IDX] |
752 pmap_bits_default[PG_V_IDX];
754 for (i = 0; i < nkpt_phys; i++) {
755 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
756 ((pd_entry_t *)KPDphys)[i] |=
757 pmap_bits_default[PG_RW_IDX] |
758 pmap_bits_default[PG_V_IDX];
762 * Map from zero to end of allocations using 2M pages as an
763 * optimization. This will bypass some of the KPTBase pages
764 * above in the KERNBASE area.
766 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
767 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
768 ((pd_entry_t *)KPDbase)[i] |=
769 pmap_bits_default[PG_RW_IDX] |
770 pmap_bits_default[PG_V_IDX] |
771 pmap_bits_default[PG_PS_IDX] |
772 pmap_bits_default[PG_G_IDX];
776 * And connect up the PD to the PDP. The kernel pmap is expected
777 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
779 for (i = 0; i < NKPDPE; i++) {
780 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
781 KPDphys + (i << PAGE_SHIFT);
782 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
783 pmap_bits_default[PG_RW_IDX] |
784 pmap_bits_default[PG_V_IDX] |
785 pmap_bits_default[PG_U_IDX];
789 * Now set up the direct map space using either 2MB or 1GB pages
790 * Preset PG_M and PG_A because demotion expects it.
792 * When filling in entries in the PD pages make sure any excess
793 * entries are set to zero as we allocated enough PD pages
795 if ((amd_feature & AMDID_PAGE1GB) == 0) {
796 for (i = 0; i < NPDEPG * ndmpdp; i++) {
797 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
798 ((pd_entry_t *)DMPDphys)[i] |=
799 pmap_bits_default[PG_RW_IDX] |
800 pmap_bits_default[PG_V_IDX] |
801 pmap_bits_default[PG_PS_IDX] |
802 pmap_bits_default[PG_G_IDX] |
803 pmap_bits_default[PG_M_IDX] |
804 pmap_bits_default[PG_A_IDX];
808 * And the direct map space's PDP
810 for (i = 0; i < ndmpdp; i++) {
811 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
813 ((pdp_entry_t *)DMPDPphys)[i] |=
814 pmap_bits_default[PG_RW_IDX] |
815 pmap_bits_default[PG_V_IDX] |
816 pmap_bits_default[PG_U_IDX];
819 for (i = 0; i < ndmpdp; i++) {
820 ((pdp_entry_t *)DMPDPphys)[i] =
821 (vm_paddr_t)i << PDPSHIFT;
822 ((pdp_entry_t *)DMPDPphys)[i] |=
823 pmap_bits_default[PG_RW_IDX] |
824 pmap_bits_default[PG_V_IDX] |
825 pmap_bits_default[PG_PS_IDX] |
826 pmap_bits_default[PG_G_IDX] |
827 pmap_bits_default[PG_M_IDX] |
828 pmap_bits_default[PG_A_IDX];
832 /* And recursively map PML4 to itself in order to get PTmap */
833 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
834 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
835 pmap_bits_default[PG_RW_IDX] |
836 pmap_bits_default[PG_V_IDX] |
837 pmap_bits_default[PG_U_IDX];
840 * Connect the Direct Map slots up to the PML4
842 for (j = 0; j < NDMPML4E; ++j) {
843 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
844 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
845 pmap_bits_default[PG_RW_IDX] |
846 pmap_bits_default[PG_V_IDX] |
847 pmap_bits_default[PG_U_IDX];
851 * Connect the KVA slot up to the PML4
853 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
854 ((pdp_entry_t *)KPML4phys)[KPML4I] |=
855 pmap_bits_default[PG_RW_IDX] |
856 pmap_bits_default[PG_V_IDX] |
857 pmap_bits_default[PG_U_IDX];
861 * Bootstrap the system enough to run with virtual memory.
863 * On the i386 this is called after mapping has already been enabled
864 * and just syncs the pmap module with what has already been done.
865 * [We can't call it easily with mapping off since the kernel is not
866 * mapped with PA == VA, hence we would have to relocate every address
867 * from the linked base (virtual) address "KERNBASE" to the actual
868 * (physical) address starting relative to 0]
871 pmap_bootstrap(vm_paddr_t *firstaddr)
876 KvaStart = VM_MIN_KERNEL_ADDRESS;
877 KvaEnd = VM_MAX_KERNEL_ADDRESS;
878 KvaSize = KvaEnd - KvaStart;
880 avail_start = *firstaddr;
883 * Create an initial set of page tables to run the kernel in.
885 create_pagetables(firstaddr);
887 virtual2_start = KvaStart;
888 virtual2_end = PTOV_OFFSET;
890 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
891 virtual_start = pmap_kmem_choose(virtual_start);
893 virtual_end = VM_MAX_KERNEL_ADDRESS;
895 /* XXX do %cr0 as well */
896 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
900 * Initialize protection array.
902 i386_protection_init();
905 * The kernel's pmap is statically allocated so we don't have to use
906 * pmap_create, which is unlikely to work correctly at this part of
907 * the boot sequence (XXX and which no longer exists).
909 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
910 kernel_pmap.pm_count = 1;
911 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
912 RB_INIT(&kernel_pmap.pm_pvroot);
913 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
914 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
917 * Reserve some special page table entries/VA space for temporary
920 #define SYSMAP(c, p, v, n) \
921 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
927 * CMAP1/CMAP2 are used for zeroing and copying pages.
929 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
934 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
937 * ptvmmap is used for reading arbitrary physical pages via
940 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
943 * msgbufp is used to map the system message buffer.
944 * XXX msgbufmap is not used.
946 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
947 atop(round_page(MSGBUF_SIZE)))
950 virtual_start = pmap_kmem_choose(virtual_start);
955 * PG_G is terribly broken on SMP because we IPI invltlb's in some
956 * cases rather then invl1pg. Actually, I don't even know why it
957 * works under UP because self-referential page table mappings
962 * Initialize the 4MB page size flag
966 * The 4MB page version of the initial
967 * kernel page mapping.
971 #if !defined(DISABLE_PSE)
972 if (cpu_feature & CPUID_PSE) {
975 * Note that we have enabled PSE mode
977 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
978 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
979 ptditmp &= ~(NBPDR - 1);
980 ptditmp |= pmap_bits_default[PG_V_IDX] |
981 pmap_bits_default[PG_RW_IDX] |
982 pmap_bits_default[PG_PS_IDX] |
983 pmap_bits_default[PG_U_IDX];
990 /* Initialize the PAT MSR */
992 pmap_pinit_defaults(&kernel_pmap);
994 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
995 &pmap_fast_kernel_cpusync);
1000 * Setup the PAT MSR.
1009 * Default values mapping PATi,PCD,PWT bits at system reset.
1010 * The default values effectively ignore the PATi bit by
1011 * repeating the encodings for 0-3 in 4-7, and map the PCD
1012 * and PWT bit combinations to the expected PAT types.
1014 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
1015 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
1016 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
1017 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
1018 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
1019 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
1020 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
1021 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
1022 pat_pte_index[PAT_WRITE_BACK] = 0;
1023 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
1024 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
1025 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
1026 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1027 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1029 if (cpu_feature & CPUID_PAT) {
1031 * If we support the PAT then set-up entries for
1032 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1035 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1036 PAT_VALUE(4, PAT_WRITE_PROTECTED);
1037 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1038 PAT_VALUE(5, PAT_WRITE_COMBINING);
1039 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1040 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1043 * Then enable the PAT
1048 load_cr4(cr4 & ~CR4_PGE);
1050 /* Disable caches (CD = 1, NW = 0). */
1052 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1054 /* Flushes caches and TLBs. */
1058 /* Update PAT and index table. */
1059 wrmsr(MSR_PAT, pat_msr);
1061 /* Flush caches and TLBs again. */
1065 /* Restore caches and PGE. */
1073 * Set 4mb pdir for mp startup
1078 if (cpu_feature & CPUID_PSE) {
1079 load_cr4(rcr4() | CR4_PSE);
1080 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
1087 * Initialize the pmap module.
1088 * Called by vm_init, to initialize any structures that the pmap
1089 * system needs to map virtual memory.
1090 * pmap_init has been enhanced to support in a fairly consistant
1091 * way, discontiguous physical memory.
1100 * Allocate memory for random pmap data structures. Includes the
1104 for (i = 0; i < vm_page_array_size; i++) {
1107 m = &vm_page_array[i];
1108 TAILQ_INIT(&m->md.pv_list);
1112 * init the pv free list
1114 initial_pvs = vm_page_array_size;
1115 if (initial_pvs < MINPV)
1116 initial_pvs = MINPV;
1117 pvzone = &pvzone_store;
1118 pvinit = (void *)kmem_alloc(&kernel_map,
1119 initial_pvs * sizeof (struct pv_entry));
1120 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1121 pvinit, initial_pvs);
1124 * Now it is safe to enable pv_table recording.
1126 pmap_initialized = TRUE;
1130 * Initialize the address space (zone) for the pv_entries. Set a
1131 * high water mark so that the system can recover from excessive
1132 * numbers of pv entries.
1137 int shpgperproc = PMAP_SHPGPERPROC;
1140 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1141 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1142 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1143 pv_entry_high_water = 9 * (pv_entry_max / 10);
1146 * Subtract out pages already installed in the zone (hack)
1148 entry_max = pv_entry_max - vm_page_array_size;
1152 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
1156 * Typically used to initialize a fictitious page by vm/device_pager.c
1159 pmap_page_init(struct vm_page *m)
1162 TAILQ_INIT(&m->md.pv_list);
1165 /***************************************************
1166 * Low level helper routines.....
1167 ***************************************************/
1170 * this routine defines the region(s) of memory that should
1171 * not be tested for the modified bit.
1175 pmap_track_modified(vm_pindex_t pindex)
1177 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1178 if ((va < clean_sva) || (va >= clean_eva))
1185 * Extract the physical page address associated with the map/VA pair.
1186 * The page must be wired for this to work reliably.
1188 * XXX for the moment we're using pv_find() instead of pv_get(), as
1189 * callers might be expecting non-blocking operation.
1192 pmap_extract(pmap_t pmap, vm_offset_t va)
1199 if (va >= VM_MAX_USER_ADDRESS) {
1201 * Kernel page directories might be direct-mapped and
1202 * there is typically no PV tracking of pte's
1206 pt = pmap_pt(pmap, va);
1207 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1208 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1209 rtval = *pt & PG_PS_FRAME;
1210 rtval |= va & PDRMASK;
1212 ptep = pmap_pt_to_pte(*pt, va);
1213 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1214 rtval = *ptep & PG_FRAME;
1215 rtval |= va & PAGE_MASK;
1221 * User pages currently do not direct-map the page directory
1222 * and some pages might not used managed PVs. But all PT's
1225 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1227 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1228 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1229 rtval = *ptep & PG_FRAME;
1230 rtval |= va & PAGE_MASK;
1239 * Similar to extract but checks protections, SMP-friendly short-cut for
1240 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1241 * fall-through to the real fault code.
1243 * The returned page, if not NULL, is held (and not busied).
1246 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1248 if (pmap && va < VM_MAX_USER_ADDRESS) {
1256 req = pmap->pmap_bits[PG_V_IDX] |
1257 pmap->pmap_bits[PG_U_IDX];
1258 if (prot & VM_PROT_WRITE)
1259 req |= pmap->pmap_bits[PG_RW_IDX];
1261 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1264 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1265 if ((*ptep & req) != req) {
1269 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1270 if (pte_pv && error == 0) {
1273 if (prot & VM_PROT_WRITE)
1276 } else if (pte_pv) {
1290 * Extract the physical page address associated kernel virtual address.
1293 pmap_kextract(vm_offset_t va)
1295 pd_entry_t pt; /* pt entry in pd */
1298 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1299 pa = DMAP_TO_PHYS(va);
1302 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1303 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1306 * Beware of a concurrent promotion that changes the
1307 * PDE at this point! For example, vtopte() must not
1308 * be used to access the PTE because it would use the
1309 * new PDE. It is, however, safe to use the old PDE
1310 * because the page table page is preserved by the
1313 pa = *pmap_pt_to_pte(pt, va);
1314 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1320 /***************************************************
1321 * Low level mapping routines.....
1322 ***************************************************/
1325 * Routine: pmap_kenter
1327 * Add a wired page to the KVA
1328 * NOTE! note that in order for the mapping to take effect -- you
1329 * should do an invltlb after doing the pmap_kenter().
1332 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1338 kernel_pmap.pmap_bits[PG_RW_IDX] |
1339 kernel_pmap.pmap_bits[PG_V_IDX];
1343 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1349 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1350 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1351 * (caller can conditionalize calling smp_invltlb()).
1354 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1361 kernel_pmap.pmap_bits[PG_RW_IDX] |
1362 kernel_pmap.pmap_bits[PG_V_IDX];
1367 cpu_invlpg((void *)va);
1373 * remove a page from the kernel pagetables
1376 pmap_kremove(vm_offset_t va)
1381 pmap_inval_smp(&kernel_pmap, va, ptep, 0);
1385 pmap_kremove_quick(vm_offset_t va)
1390 (void)pte_load_clear(ptep);
1391 cpu_invlpg((void *)va);
1395 * XXX these need to be recoded. They are not used in any critical path.
1398 pmap_kmodify_rw(vm_offset_t va)
1400 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1401 cpu_invlpg((void *)va);
1406 pmap_kmodify_nc(vm_offset_t va)
1408 atomic_set_long(vtopte(va), PG_N);
1409 cpu_invlpg((void *)va);
1414 * Used to map a range of physical addresses into kernel virtual
1415 * address space during the low level boot, typically to map the
1416 * dump bitmap, message buffer, and vm_page_array.
1418 * These mappings are typically made at some pointer after the end of the
1421 * We could return PHYS_TO_DMAP(start) here and not allocate any
1422 * via (*virtp), but then kmem from userland and kernel dumps won't
1423 * have access to the related pointers.
1426 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1429 vm_offset_t va_start;
1431 /*return PHYS_TO_DMAP(start);*/
1436 while (start < end) {
1437 pmap_kenter_quick(va, start);
1445 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1448 * Remove the specified set of pages from the data and instruction caches.
1450 * In contrast to pmap_invalidate_cache_range(), this function does not
1451 * rely on the CPU's self-snoop feature, because it is intended for use
1452 * when moving pages into a different cache domain.
1455 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1457 vm_offset_t daddr, eva;
1460 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1461 (cpu_feature & CPUID_CLFSH) == 0)
1465 for (i = 0; i < count; i++) {
1466 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1467 eva = daddr + PAGE_SIZE;
1468 for (; daddr < eva; daddr += cpu_clflush_line_size)
1476 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1478 KASSERT((sva & PAGE_MASK) == 0,
1479 ("pmap_invalidate_cache_range: sva not page-aligned"));
1480 KASSERT((eva & PAGE_MASK) == 0,
1481 ("pmap_invalidate_cache_range: eva not page-aligned"));
1483 if (cpu_feature & CPUID_SS) {
1484 ; /* If "Self Snoop" is supported, do nothing. */
1486 /* Globally invalidate caches */
1487 cpu_wbinvd_on_all_cpus();
1491 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1493 smp_invlpg_range(pmap->pm_active, sva, eva);
1497 * Add a list of wired pages to the kva
1498 * this routine is only used for temporary
1499 * kernel mappings that do not need to have
1500 * page modification or references recorded.
1501 * Note that old mappings are simply written
1502 * over. The page *must* be wired.
1505 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1508 int do_smpinvltlb = 0;
1510 end_va = va + count * PAGE_SIZE;
1512 while (va < end_va) {
1518 *pte = VM_PAGE_TO_PHYS(*m) |
1519 kernel_pmap.pmap_bits[PG_RW_IDX] |
1520 kernel_pmap.pmap_bits[PG_V_IDX] |
1521 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1523 cpu_invlpg((void *)va);
1532 * This routine jerks page mappings from the
1533 * kernel -- it is meant only for temporary mappings.
1535 * MPSAFE, INTERRUPT SAFE (cluster callback)
1538 pmap_qremove(vm_offset_t va, int count)
1542 end_va = va + count * PAGE_SIZE;
1544 while (va < end_va) {
1548 (void)pte_load_clear(pte);
1549 cpu_invlpg((void *)va);
1556 * Create a new thread and optionally associate it with a (new) process.
1557 * NOTE! the new thread's cpu may not equal the current cpu.
1560 pmap_init_thread(thread_t td)
1562 /* enforce pcb placement & alignment */
1563 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1564 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1565 td->td_savefpu = &td->td_pcb->pcb_save;
1566 td->td_sp = (char *)td->td_pcb; /* no -16 */
1570 * This routine directly affects the fork perf for a process.
1573 pmap_init_proc(struct proc *p)
1578 pmap_pinit_defaults(struct pmap *pmap)
1580 bcopy(pmap_bits_default, pmap->pmap_bits,
1581 sizeof(pmap_bits_default));
1582 bcopy(protection_codes, pmap->protection_codes,
1583 sizeof(protection_codes));
1584 bcopy(pat_pte_index, pmap->pmap_cache_bits,
1585 sizeof(pat_pte_index));
1586 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1587 pmap->copyinstr = std_copyinstr;
1588 pmap->copyin = std_copyin;
1589 pmap->copyout = std_copyout;
1590 pmap->fubyte = std_fubyte;
1591 pmap->subyte = std_subyte;
1592 pmap->fuword = std_fuword;
1593 pmap->suword = std_suword;
1594 pmap->suword32 = std_suword32;
1597 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1598 * it, and IdlePTD, represents the template used to update all other pmaps.
1600 * On architectures where the kernel pmap is not integrated into the user
1601 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1602 * kernel_pmap should be used to directly access the kernel_pmap.
1605 pmap_pinit0(struct pmap *pmap)
1607 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1609 CPUMASK_ASSZERO(pmap->pm_active);
1610 pmap->pm_pvhint = NULL;
1611 RB_INIT(&pmap->pm_pvroot);
1612 spin_init(&pmap->pm_spin, "pmapinit0");
1613 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1614 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1615 pmap_pinit_defaults(pmap);
1619 * Initialize a preallocated and zeroed pmap structure,
1620 * such as one in a vmspace structure.
1623 pmap_pinit_simple(struct pmap *pmap)
1626 * Misc initialization
1629 CPUMASK_ASSZERO(pmap->pm_active);
1630 pmap->pm_pvhint = NULL;
1631 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1633 pmap_pinit_defaults(pmap);
1636 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1639 if (pmap->pm_pmlpv == NULL) {
1640 RB_INIT(&pmap->pm_pvroot);
1641 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1642 spin_init(&pmap->pm_spin, "pmapinitsimple");
1643 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1648 pmap_pinit(struct pmap *pmap)
1653 if (pmap->pm_pmlpv) {
1654 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1659 pmap_pinit_simple(pmap);
1660 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1663 * No need to allocate page table space yet but we do need a valid
1664 * page directory table.
1666 if (pmap->pm_pml4 == NULL) {
1668 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1672 * Allocate the page directory page, which wires it even though
1673 * it isn't being entered into some higher level page table (it
1674 * being the highest level). If one is already cached we don't
1675 * have to do anything.
1677 if ((pv = pmap->pm_pmlpv) == NULL) {
1678 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1679 pmap->pm_pmlpv = pv;
1680 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1681 VM_PAGE_TO_PHYS(pv->pv_m));
1685 * Install DMAP and KMAP.
1687 for (j = 0; j < NDMPML4E; ++j) {
1688 pmap->pm_pml4[DMPML4I + j] =
1689 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1690 pmap->pmap_bits[PG_RW_IDX] |
1691 pmap->pmap_bits[PG_V_IDX] |
1692 pmap->pmap_bits[PG_U_IDX];
1694 pmap->pm_pml4[KPML4I] = KPDPphys |
1695 pmap->pmap_bits[PG_RW_IDX] |
1696 pmap->pmap_bits[PG_V_IDX] |
1697 pmap->pmap_bits[PG_U_IDX];
1700 * install self-referential address mapping entry
1702 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1703 pmap->pmap_bits[PG_V_IDX] |
1704 pmap->pmap_bits[PG_RW_IDX] |
1705 pmap->pmap_bits[PG_A_IDX] |
1706 pmap->pmap_bits[PG_M_IDX];
1708 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1709 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1711 KKASSERT(pmap->pm_pml4[255] == 0);
1712 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1713 KKASSERT(pv->pv_entry.rbe_left == NULL);
1714 KKASSERT(pv->pv_entry.rbe_right == NULL);
1718 * Clean up a pmap structure so it can be physically freed. This routine
1719 * is called by the vmspace dtor function. A great deal of pmap data is
1720 * left passively mapped to improve vmspace management so we have a bit
1721 * of cleanup work to do here.
1724 pmap_puninit(pmap_t pmap)
1729 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1730 if ((pv = pmap->pm_pmlpv) != NULL) {
1731 if (pv_hold_try(pv) == 0)
1733 KKASSERT(pv == pmap->pm_pmlpv);
1734 p = pmap_remove_pv_page(pv);
1736 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1737 vm_page_busy_wait(p, FALSE, "pgpun");
1738 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1739 vm_page_unwire(p, 0);
1740 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1743 * XXX eventually clean out PML4 static entries and
1744 * use vm_page_free_zero()
1747 pmap->pm_pmlpv = NULL;
1749 if (pmap->pm_pml4) {
1750 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1751 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1752 pmap->pm_pml4 = NULL;
1754 KKASSERT(pmap->pm_stats.resident_count == 0);
1755 KKASSERT(pmap->pm_stats.wired_count == 0);
1759 * Wire in kernel global address entries. To avoid a race condition
1760 * between pmap initialization and pmap_growkernel, this procedure
1761 * adds the pmap to the master list (which growkernel scans to update),
1762 * then copies the template.
1765 pmap_pinit2(struct pmap *pmap)
1767 spin_lock(&pmap_spin);
1768 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1769 spin_unlock(&pmap_spin);
1773 * This routine is called when various levels in the page table need to
1774 * be populated. This routine cannot fail.
1776 * This function returns two locked pv_entry's, one representing the
1777 * requested pv and one representing the requested pv's parent pv. If
1778 * the pv did not previously exist it will be mapped into its parent
1779 * and wired, otherwise no additional wire count will be added.
1783 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1788 vm_pindex_t pt_pindex;
1794 * If the pv already exists and we aren't being asked for the
1795 * parent page table page we can just return it. A locked+held pv
1796 * is returned. The pv will also have a second hold related to the
1797 * pmap association that we don't have to worry about.
1800 pv = pv_alloc(pmap, ptepindex, &isnew);
1801 if (isnew == 0 && pvpp == NULL)
1805 * Special case terminal PVs. These are not page table pages so
1806 * no vm_page is allocated (the caller supplied the vm_page). If
1807 * pvpp is non-NULL we are being asked to also removed the pt_pv
1810 * Note that pt_pv's are only returned for user VAs. We assert that
1811 * a pt_pv is not being requested for kernel VAs.
1813 if (ptepindex < pmap_pt_pindex(0)) {
1814 if (ptepindex >= NUPTE_USER)
1815 KKASSERT(pvpp == NULL);
1817 KKASSERT(pvpp != NULL);
1819 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1820 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1822 vm_page_wire_quick(pvp->pv_m);
1831 * Non-terminal PVs allocate a VM page to represent the page table,
1832 * so we have to resolve pvp and calculate ptepindex for the pvp
1833 * and then for the page table entry index in the pvp for
1836 if (ptepindex < pmap_pd_pindex(0)) {
1838 * pv is PT, pvp is PD
1840 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1841 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1842 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1849 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1850 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1852 } else if (ptepindex < pmap_pdp_pindex(0)) {
1854 * pv is PD, pvp is PDP
1856 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1859 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1860 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1862 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1863 KKASSERT(pvpp == NULL);
1866 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1874 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1875 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1876 } else if (ptepindex < pmap_pml4_pindex()) {
1878 * pv is PDP, pvp is the root pml4 table
1880 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1887 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1888 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1891 * pv represents the top-level PML4, there is no parent.
1899 * This code is only reached if isnew is TRUE and this is not a
1900 * terminal PV. We need to allocate a vm_page for the page table
1901 * at this level and enter it into the parent page table.
1903 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1906 m = vm_page_alloc(NULL, pv->pv_pindex,
1907 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1908 VM_ALLOC_INTERRUPT);
1913 vm_page_spin_lock(m);
1914 pmap_page_stats_adding(m);
1915 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1917 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1918 vm_page_spin_unlock(m);
1919 vm_page_unmanage(m); /* m must be spinunlocked */
1921 if ((m->flags & PG_ZERO) == 0) {
1922 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1926 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1929 m->valid = VM_PAGE_BITS_ALL;
1930 vm_page_flag_clear(m, PG_ZERO);
1931 vm_page_wire(m); /* wire for mapping in parent */
1934 * Wire the page into pvp, bump the wire-count for pvp's page table
1935 * page. Bump the resident_count for the pmap. There is no pvp
1936 * for the top level, address the pm_pml4[] array directly.
1938 * If the caller wants the parent we return it, otherwise
1939 * we just put it away.
1941 * No interlock is needed for pte 0 -> non-zero.
1943 * In the situation where *ptep is valid we might have an unmanaged
1944 * page table page shared from another page table which we need to
1945 * unshare before installing our private page table page.
1948 ptep = pv_pte_lookup(pvp, ptepindex);
1949 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1953 panic("pmap_allocpte: unexpected pte %p/%d",
1954 pvp, (int)ptepindex);
1956 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, ptep, 0);
1957 if (vm_page_unwire_quick(
1958 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
1959 panic("pmap_allocpte: shared pgtable "
1960 "pg bad wirecount");
1962 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1964 vm_page_wire_quick(pvp->pv_m);
1966 *ptep = VM_PAGE_TO_PHYS(m) |
1967 (pmap->pmap_bits[PG_U_IDX] |
1968 pmap->pmap_bits[PG_RW_IDX] |
1969 pmap->pmap_bits[PG_V_IDX] |
1970 pmap->pmap_bits[PG_A_IDX] |
1971 pmap->pmap_bits[PG_M_IDX]);
1983 * This version of pmap_allocpte() checks for possible segment optimizations
1984 * that would allow page-table sharing. It can be called for terminal
1985 * page or page table page ptepindex's.
1987 * The function is called with page table page ptepindex's for fictitious
1988 * and unmanaged terminal pages. That is, we don't want to allocate a
1989 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
1992 * This function can return a pv and *pvpp associated with the passed in pmap
1993 * OR a pv and *pvpp associated with the shared pmap. In the latter case
1994 * an unmanaged page table page will be entered into the pass in pmap.
1998 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
1999 vm_map_entry_t entry, vm_offset_t va)
2005 pv_entry_t pte_pv; /* in original or shared pmap */
2006 pv_entry_t pt_pv; /* in original or shared pmap */
2007 pv_entry_t proc_pd_pv; /* in original pmap */
2008 pv_entry_t proc_pt_pv; /* in original pmap */
2009 pv_entry_t xpv; /* PT in shared pmap */
2010 pd_entry_t *pt; /* PT entry in PD of original pmap */
2011 pd_entry_t opte; /* contents of *pt */
2012 pd_entry_t npte; /* contents of *pt */
2017 * Basic tests, require a non-NULL vm_map_entry, require proper
2018 * alignment and type for the vm_map_entry, require that the
2019 * underlying object already be allocated.
2021 * We allow almost any type of object to use this optimization.
2022 * The object itself does NOT have to be sized to a multiple of the
2023 * segment size, but the memory mapping does.
2025 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2026 * won't work as expected.
2028 if (entry == NULL ||
2029 pmap_mmu_optimize == 0 || /* not enabled */
2030 ptepindex >= pmap_pd_pindex(0) || /* not terminal or pt */
2031 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2032 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2033 entry->object.vm_object == NULL || /* needs VM object */
2034 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2035 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2036 (entry->offset & SEG_MASK) || /* must be aligned */
2037 (entry->start & SEG_MASK)) {
2038 return(pmap_allocpte(pmap, ptepindex, pvpp));
2042 * Make sure the full segment can be represented.
2044 b = va & ~(vm_offset_t)SEG_MASK;
2045 if (b < entry->start || b + SEG_SIZE > entry->end)
2046 return(pmap_allocpte(pmap, ptepindex, pvpp));
2049 * If the full segment can be represented dive the VM object's
2050 * shared pmap, allocating as required.
2052 object = entry->object.vm_object;
2054 if (entry->protection & VM_PROT_WRITE)
2055 obpmapp = &object->md.pmap_rw;
2057 obpmapp = &object->md.pmap_ro;
2060 if (pmap_enter_debug > 0) {
2062 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2064 va, entry->protection, object,
2066 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2067 entry, entry->start, entry->end);
2072 * We allocate what appears to be a normal pmap but because portions
2073 * of this pmap are shared with other unrelated pmaps we have to
2074 * set pm_active to point to all cpus.
2076 * XXX Currently using pmap_spin to interlock the update, can't use
2077 * vm_object_hold/drop because the token might already be held
2078 * shared OR exclusive and we don't know.
2080 while ((obpmap = *obpmapp) == NULL) {
2081 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2082 pmap_pinit_simple(obpmap);
2083 pmap_pinit2(obpmap);
2084 spin_lock(&pmap_spin);
2085 if (*obpmapp != NULL) {
2089 spin_unlock(&pmap_spin);
2090 pmap_release(obpmap);
2091 pmap_puninit(obpmap);
2092 kfree(obpmap, M_OBJPMAP);
2093 obpmap = *obpmapp; /* safety */
2095 obpmap->pm_active = smp_active_mask;
2097 spin_unlock(&pmap_spin);
2102 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2103 * pte/pt using the shared pmap from the object but also adjust
2104 * the process pmap's page table page as a side effect.
2108 * Resolve the terminal PTE and PT in the shared pmap. This is what
2109 * we will return. This is true if ptepindex represents a terminal
2110 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2114 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2115 if (ptepindex >= pmap_pt_pindex(0))
2121 * Resolve the PD in the process pmap so we can properly share the
2122 * page table page. Lock order is bottom-up (leaf first)!
2124 * NOTE: proc_pt_pv can be NULL.
2126 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2127 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2129 if (pmap_enter_debug > 0) {
2131 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2133 (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2140 * xpv is the page table page pv from the shared object
2141 * (for convenience), from above.
2143 * Calculate the pte value for the PT to load into the process PD.
2144 * If we have to change it we must properly dispose of the previous
2147 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2148 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2149 (pmap->pmap_bits[PG_U_IDX] |
2150 pmap->pmap_bits[PG_RW_IDX] |
2151 pmap->pmap_bits[PG_V_IDX] |
2152 pmap->pmap_bits[PG_A_IDX] |
2153 pmap->pmap_bits[PG_M_IDX]);
2156 * Dispose of previous page table page if it was local to the
2157 * process pmap. If the old pt is not empty we cannot dispose of it
2158 * until we clean it out. This case should not arise very often so
2159 * it is not optimized.
2162 if (proc_pt_pv->pv_m->wire_count != 1) {
2168 va & ~(vm_offset_t)SEG_MASK,
2169 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2174 * The release call will indirectly clean out *pt
2176 pmap_release_pv(proc_pt_pv, proc_pd_pv, 1);
2179 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2183 * Handle remaining cases.
2187 vm_page_wire_quick(xpv->pv_m);
2188 vm_page_wire_quick(proc_pd_pv->pv_m);
2189 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2190 } else if (*pt != npte) {
2191 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, pt, npte);
2194 opte = pte_load_clear(pt);
2195 KKASSERT(opte && opte != npte);
2199 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
2202 * Clean up opte, bump the wire_count for the process
2203 * PD page representing the new entry if it was
2206 * If the entry was not previously empty and we have
2207 * a PT in the proc pmap then opte must match that
2208 * pt. The proc pt must be retired (this is done
2209 * later on in this procedure).
2211 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2214 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2215 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2216 if (vm_page_unwire_quick(m)) {
2217 panic("pmap_allocpte_seg: "
2218 "bad wire count %p",
2224 * The existing process page table was replaced and must be destroyed
2238 * Release any resources held by the given physical map.
2240 * Called when a pmap initialized by pmap_pinit is being released. Should
2241 * only be called if the map contains no valid mappings.
2243 * Caller must hold pmap->pm_token
2245 struct pmap_release_info {
2250 static int pmap_release_callback(pv_entry_t pv, void *data);
2253 pmap_release(struct pmap *pmap)
2255 struct pmap_release_info info;
2257 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2258 ("pmap still active! %016jx",
2259 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2261 spin_lock(&pmap_spin);
2262 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2263 spin_unlock(&pmap_spin);
2266 * Pull pv's off the RB tree in order from low to high and release
2272 spin_lock(&pmap->pm_spin);
2273 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2274 pmap_release_callback, &info);
2275 spin_unlock(&pmap->pm_spin);
2276 } while (info.retry);
2280 * One resident page (the pml4 page) should remain.
2281 * No wired pages should remain.
2283 KKASSERT(pmap->pm_stats.resident_count ==
2284 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2286 KKASSERT(pmap->pm_stats.wired_count == 0);
2290 pmap_release_callback(pv_entry_t pv, void *data)
2292 struct pmap_release_info *info = data;
2293 pmap_t pmap = info->pmap;
2296 if (pv_hold_try(pv)) {
2297 spin_unlock(&pmap->pm_spin);
2299 spin_unlock(&pmap->pm_spin);
2302 if (pv->pv_pmap != pmap) {
2304 spin_lock(&pmap->pm_spin);
2308 r = pmap_release_pv(pv, NULL, 0);
2309 spin_lock(&pmap->pm_spin);
2314 * Called with held (i.e. also locked) pv. This function will dispose of
2315 * the lock along with the pv.
2317 * If the caller already holds the locked parent page table for pv it
2318 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2319 * pass NULL for pvp.
2322 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, int smp)
2327 * The pmap is currently not spinlocked, pv is held+locked.
2328 * Remove the pv's page from its parent's page table. The
2329 * parent's page table page's wire_count will be decremented.
2331 * This will clean out the pte at any level of the page table.
2332 * If smp != 0 all cpus are affected.
2334 pmap_remove_pv_pte(pv, pvp, smp);
2337 * Terminal pvs are unhooked from their vm_pages. Because
2338 * terminal pages aren't page table pages they aren't wired
2339 * by us, so we have to be sure not to unwire them either.
2341 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2342 pmap_remove_pv_page(pv);
2347 * We leave the top-level page table page cached, wired, and
2348 * mapped in the pmap until the dtor function (pmap_puninit())
2351 * Since we are leaving the top-level pv intact we need
2352 * to break out of what would otherwise be an infinite loop.
2354 if (pv->pv_pindex == pmap_pml4_pindex()) {
2360 * For page table pages (other than the top-level page),
2361 * remove and free the vm_page. The representitive mapping
2362 * removed above by pmap_remove_pv_pte() did not undo the
2363 * last wire_count so we have to do that as well.
2365 p = pmap_remove_pv_page(pv);
2366 vm_page_busy_wait(p, FALSE, "pmaprl");
2367 if (p->wire_count != 1) {
2368 kprintf("p->wire_count was %016lx %d\n",
2369 pv->pv_pindex, p->wire_count);
2371 KKASSERT(p->wire_count == 1);
2372 KKASSERT(p->flags & PG_UNMANAGED);
2374 vm_page_unwire(p, 0);
2375 KKASSERT(p->wire_count == 0);
2378 * Theoretically this page, if not the pml4 page, should contain
2379 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2389 * This function will remove the pte associated with a pv from its parent.
2390 * Terminal pv's are supported. All cpus are affected if smp != 0.
2392 * The wire count will be dropped on the parent page table. The wire
2393 * count on the page being removed (pv->pv_m) from the parent page table
2394 * is NOT touched. Note that terminal pages will not have any additional
2395 * wire counts while page table pages will have at least one representing
2396 * the mapping, plus others representing sub-mappings.
2398 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2399 * pages and user page table and terminal pages.
2401 * The pv must be locked.
2403 * XXX must lock parent pv's if they exist to remove pte XXX
2407 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, int smp)
2409 vm_pindex_t ptepindex = pv->pv_pindex;
2410 pmap_t pmap = pv->pv_pmap;
2416 if (ptepindex == pmap_pml4_pindex()) {
2418 * We are the top level pml4 table, there is no parent.
2420 p = pmap->pm_pmlpv->pv_m;
2421 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2423 * Remove a PDP page from the pml4e. This can only occur
2424 * with user page tables. We do not have to lock the
2425 * pml4 PV so just ignore pvp.
2427 vm_pindex_t pml4_pindex;
2428 vm_pindex_t pdp_index;
2431 pdp_index = ptepindex - pmap_pdp_pindex(0);
2433 pml4_pindex = pmap_pml4_pindex();
2434 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2438 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2439 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2440 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2442 pmap_inval_smp(pmap, (vm_offset_t)-1, pdp, 0);
2445 } else if (ptepindex >= pmap_pd_pindex(0)) {
2447 * Remove a PD page from the pdp
2449 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2450 * of a simple pmap because it stops at
2453 vm_pindex_t pdp_pindex;
2454 vm_pindex_t pd_index;
2457 pd_index = ptepindex - pmap_pd_pindex(0);
2460 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2461 (pd_index >> NPML4EPGSHIFT);
2462 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2467 pd = pv_pte_lookup(pvp, pd_index &
2468 ((1ul << NPDPEPGSHIFT) - 1));
2469 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2470 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2472 pmap_inval_smp(pmap, (vm_offset_t)-1, pd, 0);
2476 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2477 p = pv->pv_m; /* degenerate test later */
2479 } else if (ptepindex >= pmap_pt_pindex(0)) {
2481 * Remove a PT page from the pd
2483 vm_pindex_t pd_pindex;
2484 vm_pindex_t pt_index;
2487 pt_index = ptepindex - pmap_pt_pindex(0);
2490 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2491 (pt_index >> NPDPEPGSHIFT);
2492 pvp = pv_get(pv->pv_pmap, pd_pindex);
2496 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2497 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2498 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2500 pmap_inval_smp(pmap, (vm_offset_t)-1, pt, 0);
2505 * Remove a PTE from the PT page
2507 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2508 * pv is a pte_pv so we can safely lock pt_pv.
2510 * NOTE: FICTITIOUS pages may have multiple physical mappings
2511 * so PHYS_TO_VM_PAGE() will not necessarily work for
2514 vm_pindex_t pt_pindex;
2519 pt_pindex = ptepindex >> NPTEPGSHIFT;
2520 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2522 if (ptepindex >= NUPTE_USER) {
2523 ptep = vtopte(ptepindex << PAGE_SHIFT);
2524 KKASSERT(pvp == NULL);
2527 pt_pindex = NUPTE_TOTAL +
2528 (ptepindex >> NPDPEPGSHIFT);
2529 pvp = pv_get(pv->pv_pmap, pt_pindex);
2533 ptep = pv_pte_lookup(pvp, ptepindex &
2534 ((1ul << NPDPEPGSHIFT) - 1));
2537 pte = pmap_inval_smp(pmap, va, ptep, 0);
2539 pte = pte_load_clear(ptep);
2540 cpu_invlpg((void *)va);
2544 * Now update the vm_page_t
2546 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) !=
2547 (pmap->pmap_bits[PG_MANAGED_IDX]|pmap->pmap_bits[PG_V_IDX])) {
2548 kprintf("remove_pte badpte %016lx %016lx %d\n",
2550 pv->pv_pindex < pmap_pt_pindex(0));
2552 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2553 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2554 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2557 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2560 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2561 if (pmap_track_modified(ptepindex))
2564 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2565 vm_page_flag_set(p, PG_REFERENCED);
2567 if (pte & pmap->pmap_bits[PG_W_IDX])
2568 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2569 if (pte & pmap->pmap_bits[PG_G_IDX])
2570 cpu_invlpg((void *)va);
2574 * Unwire the parent page table page. The wire_count cannot go below
2575 * 1 here because the parent page table page is itself still mapped.
2577 * XXX remove the assertions later.
2579 KKASSERT(pv->pv_m == p);
2580 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2581 panic("pmap_remove_pv_pte: Insufficient wire_count");
2588 * Remove the vm_page association to a pv. The pv must be locked.
2592 pmap_remove_pv_page(pv_entry_t pv)
2598 vm_page_spin_lock(m);
2600 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2601 pmap_page_stats_deleting(m);
2604 atomic_add_int(&m->object->agg_pv_list_count, -1);
2606 if (TAILQ_EMPTY(&m->md.pv_list))
2607 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2608 vm_page_spin_unlock(m);
2613 * Grow the number of kernel page table entries, if needed.
2615 * This routine is always called to validate any address space
2616 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2617 * space below KERNBASE.
2620 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2623 vm_offset_t ptppaddr;
2625 pd_entry_t *pt, newpt;
2627 int update_kernel_vm_end;
2630 * bootstrap kernel_vm_end on first real VM use
2632 if (kernel_vm_end == 0) {
2633 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2635 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2636 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2637 ~(PAGE_SIZE * NPTEPG - 1);
2639 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2640 kernel_vm_end = kernel_map.max_offset;
2647 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2648 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2649 * do not want to force-fill 128G worth of page tables.
2651 if (kstart < KERNBASE) {
2652 if (kstart > kernel_vm_end)
2653 kstart = kernel_vm_end;
2654 KKASSERT(kend <= KERNBASE);
2655 update_kernel_vm_end = 1;
2657 update_kernel_vm_end = 0;
2660 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2661 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2663 if (kend - 1 >= kernel_map.max_offset)
2664 kend = kernel_map.max_offset;
2666 while (kstart < kend) {
2667 pt = pmap_pt(&kernel_pmap, kstart);
2669 /* We need a new PDP entry */
2670 nkpg = vm_page_alloc(NULL, nkpt,
2673 VM_ALLOC_INTERRUPT);
2675 panic("pmap_growkernel: no memory to grow "
2678 paddr = VM_PAGE_TO_PHYS(nkpg);
2679 if ((nkpg->flags & PG_ZERO) == 0)
2680 pmap_zero_page(paddr);
2681 vm_page_flag_clear(nkpg, PG_ZERO);
2682 newpd = (pdp_entry_t)
2684 kernel_pmap.pmap_bits[PG_V_IDX] |
2685 kernel_pmap.pmap_bits[PG_RW_IDX] |
2686 kernel_pmap.pmap_bits[PG_A_IDX] |
2687 kernel_pmap.pmap_bits[PG_M_IDX]);
2688 *pmap_pd(&kernel_pmap, kstart) = newpd;
2690 continue; /* try again */
2692 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2693 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2694 ~(PAGE_SIZE * NPTEPG - 1);
2695 if (kstart - 1 >= kernel_map.max_offset) {
2696 kstart = kernel_map.max_offset;
2703 * This index is bogus, but out of the way
2705 nkpg = vm_page_alloc(NULL, nkpt,
2708 VM_ALLOC_INTERRUPT);
2710 panic("pmap_growkernel: no memory to grow kernel");
2713 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2714 pmap_zero_page(ptppaddr);
2715 vm_page_flag_clear(nkpg, PG_ZERO);
2716 newpt = (pd_entry_t) (ptppaddr |
2717 kernel_pmap.pmap_bits[PG_V_IDX] |
2718 kernel_pmap.pmap_bits[PG_RW_IDX] |
2719 kernel_pmap.pmap_bits[PG_A_IDX] |
2720 kernel_pmap.pmap_bits[PG_M_IDX]);
2721 *pmap_pt(&kernel_pmap, kstart) = newpt;
2724 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2725 ~(PAGE_SIZE * NPTEPG - 1);
2727 if (kstart - 1 >= kernel_map.max_offset) {
2728 kstart = kernel_map.max_offset;
2734 * Only update kernel_vm_end for areas below KERNBASE.
2736 if (update_kernel_vm_end && kernel_vm_end < kstart)
2737 kernel_vm_end = kstart;
2741 * Add a reference to the specified pmap.
2744 pmap_reference(pmap_t pmap)
2747 lwkt_gettoken(&pmap->pm_token);
2749 lwkt_reltoken(&pmap->pm_token);
2753 /***************************************************
2754 * page management routines.
2755 ***************************************************/
2758 * Hold a pv without locking it
2761 pv_hold(pv_entry_t pv)
2763 atomic_add_int(&pv->pv_hold, 1);
2767 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2768 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2771 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2772 * pv list via its page) must be held by the caller.
2775 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2780 * Critical path shortcut expects pv to already have one ref
2781 * (for the pv->pv_pmap).
2783 if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
2786 pv->pv_line = lineno;
2792 count = pv->pv_hold;
2794 if ((count & PV_HOLD_LOCKED) == 0) {
2795 if (atomic_cmpset_int(&pv->pv_hold, count,
2796 (count + 1) | PV_HOLD_LOCKED)) {
2799 pv->pv_line = lineno;
2804 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2812 * Drop a previously held pv_entry which could not be locked, allowing its
2815 * Must not be called with a spinlock held as we might zfree() the pv if it
2816 * is no longer associated with a pmap and this was the last hold count.
2819 pv_drop(pv_entry_t pv)
2824 count = pv->pv_hold;
2826 KKASSERT((count & PV_HOLD_MASK) > 0);
2827 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2828 (PV_HOLD_LOCKED | 1));
2829 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2830 if ((count & PV_HOLD_MASK) == 1) {
2832 if (pmap_enter_debug > 0) {
2834 kprintf("pv_drop: free pv %p\n", pv);
2837 KKASSERT(count == 1);
2838 KKASSERT(pv->pv_pmap == NULL);
2848 * Find or allocate the requested PV entry, returning a locked, held pv.
2850 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
2851 * for the caller and one representing the pmap and vm_page association.
2853 * If (*isnew) is zero, the returned pv will have only one hold count.
2855 * Since both associations can only be adjusted while the pv is locked,
2856 * together they represent just one additional hold.
2860 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2863 pv_entry_t pnew = NULL;
2865 spin_lock(&pmap->pm_spin);
2867 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2868 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2873 spin_unlock(&pmap->pm_spin);
2874 pnew = zalloc(pvzone);
2875 spin_lock(&pmap->pm_spin);
2878 pnew->pv_pmap = pmap;
2879 pnew->pv_pindex = pindex;
2880 pnew->pv_hold = PV_HOLD_LOCKED | 2;
2882 pnew->pv_func = func;
2883 pnew->pv_line = lineno;
2885 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2886 ++pmap->pm_generation;
2887 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2888 spin_unlock(&pmap->pm_spin);
2893 spin_unlock(&pmap->pm_spin);
2894 zfree(pvzone, pnew);
2896 spin_lock(&pmap->pm_spin);
2899 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2900 spin_unlock(&pmap->pm_spin);
2902 spin_unlock(&pmap->pm_spin);
2903 _pv_lock(pv PMAP_DEBUG_COPY);
2905 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2910 spin_lock(&pmap->pm_spin);
2915 * Find the requested PV entry, returning a locked+held pv or NULL
2919 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2923 spin_lock(&pmap->pm_spin);
2928 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2929 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2933 spin_unlock(&pmap->pm_spin);
2936 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2937 spin_unlock(&pmap->pm_spin);
2939 spin_unlock(&pmap->pm_spin);
2940 _pv_lock(pv PMAP_DEBUG_COPY);
2942 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2943 pv_cache(pv, pindex);
2947 spin_lock(&pmap->pm_spin);
2952 * Lookup, hold, and attempt to lock (pmap,pindex).
2954 * If the entry does not exist NULL is returned and *errorp is set to 0
2956 * If the entry exists and could be successfully locked it is returned and
2957 * errorp is set to 0.
2959 * If the entry exists but could NOT be successfully locked it is returned
2960 * held and *errorp is set to 1.
2964 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2968 spin_lock_shared(&pmap->pm_spin);
2969 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2970 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2972 spin_unlock_shared(&pmap->pm_spin);
2976 if (pv_hold_try(pv)) {
2977 pv_cache(pv, pindex);
2978 spin_unlock_shared(&pmap->pm_spin);
2980 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
2981 return(pv); /* lock succeeded */
2983 spin_unlock_shared(&pmap->pm_spin);
2985 return (pv); /* lock failed */
2989 * Find the requested PV entry, returning a held pv or NULL
2993 pv_find(pmap_t pmap, vm_pindex_t pindex)
2997 spin_lock_shared(&pmap->pm_spin);
2999 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3000 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3002 spin_unlock_shared(&pmap->pm_spin);
3006 pv_cache(pv, pindex);
3007 spin_unlock_shared(&pmap->pm_spin);
3012 * Lock a held pv, keeping the hold count
3016 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3021 count = pv->pv_hold;
3023 if ((count & PV_HOLD_LOCKED) == 0) {
3024 if (atomic_cmpset_int(&pv->pv_hold, count,
3025 count | PV_HOLD_LOCKED)) {
3028 pv->pv_line = lineno;
3034 tsleep_interlock(pv, 0);
3035 if (atomic_cmpset_int(&pv->pv_hold, count,
3036 count | PV_HOLD_WAITING)) {
3038 kprintf("pv waiting on %s:%d\n",
3039 pv->pv_func, pv->pv_line);
3041 tsleep(pv, PINTERLOCKED, "pvwait", hz);
3048 * Unlock a held and locked pv, keeping the hold count.
3052 pv_unlock(pv_entry_t pv)
3057 count = pv->pv_hold;
3059 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3060 (PV_HOLD_LOCKED | 1));
3061 if (atomic_cmpset_int(&pv->pv_hold, count,
3063 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3064 if (count & PV_HOLD_WAITING)
3072 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3073 * and the hold count drops to zero we will free it.
3075 * Caller should not hold any spin locks. We are protected from hold races
3076 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3077 * lock held. A pv cannot be located otherwise.
3081 pv_put(pv_entry_t pv)
3084 if (pmap_enter_debug > 0) {
3086 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3091 * Fast - shortcut most common condition
3093 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3104 * Remove the pmap association from a pv, require that pv_m already be removed,
3105 * then unlock and drop the pv. Any pte operations must have already been
3106 * completed. This call may result in a last-drop which will physically free
3109 * Removing the pmap association entails an additional drop.
3111 * pv must be exclusively locked on call and will be disposed of on return.
3115 pv_free(pv_entry_t pv)
3119 KKASSERT(pv->pv_m == NULL);
3120 KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3121 if ((pmap = pv->pv_pmap) != NULL) {
3122 spin_lock(&pmap->pm_spin);
3123 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3124 ++pmap->pm_generation;
3125 if (pmap->pm_pvhint == pv)
3126 pmap->pm_pvhint = NULL;
3127 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3130 spin_unlock(&pmap->pm_spin);
3133 * Try to shortcut three atomic ops, otherwise fall through
3134 * and do it normally. Drop two refs and the lock all in
3137 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3139 if (pmap_enter_debug > 0) {
3141 kprintf("pv_free: free pv %p\n", pv);
3147 pv_drop(pv); /* ref for pv_pmap */
3153 * This routine is very drastic, but can save the system
3161 static int warningdone=0;
3163 if (pmap_pagedaemon_waken == 0)
3165 pmap_pagedaemon_waken = 0;
3166 if (warningdone < 5) {
3167 kprintf("pmap_collect: collecting pv entries -- "
3168 "suggest increasing PMAP_SHPGPERPROC\n");
3172 for (i = 0; i < vm_page_array_size; i++) {
3173 m = &vm_page_array[i];
3174 if (m->wire_count || m->hold_count)
3176 if (vm_page_busy_try(m, TRUE) == 0) {
3177 if (m->wire_count == 0 && m->hold_count == 0) {
3186 * Scan the pmap for active page table entries and issue a callback.
3187 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3188 * its parent page table.
3190 * pte_pv will be NULL if the page or page table is unmanaged.
3191 * pt_pv will point to the page table page containing the pte for the page.
3193 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3194 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3195 * process pmap's PD and page to the callback function. This can be
3196 * confusing because the pt_pv is really a pd_pv, and the target page
3197 * table page is simply aliased by the pmap and not owned by it.
3199 * It is assumed that the start and end are properly rounded to the page size.
3201 * It is assumed that PD pages and above are managed and thus in the RB tree,
3202 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3204 struct pmap_scan_info {
3208 vm_pindex_t sva_pd_pindex;
3209 vm_pindex_t eva_pd_pindex;
3210 void (*func)(pmap_t, struct pmap_scan_info *,
3211 pv_entry_t, pv_entry_t, int, vm_offset_t,
3212 pt_entry_t *, void *);
3218 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3219 static int pmap_scan_callback(pv_entry_t pv, void *data);
3222 pmap_scan(struct pmap_scan_info *info)
3224 struct pmap *pmap = info->pmap;
3225 pv_entry_t pd_pv; /* A page directory PV */
3226 pv_entry_t pt_pv; /* A page table PV */
3227 pv_entry_t pte_pv; /* A page table entry PV */
3230 struct pv_entry dummy_pv;
3237 * Hold the token for stability; if the pmap is empty we have nothing
3240 lwkt_gettoken(&pmap->pm_token);
3242 if (pmap->pm_stats.resident_count == 0) {
3243 lwkt_reltoken(&pmap->pm_token);
3252 * Special handling for scanning one page, which is a very common
3253 * operation (it is?).
3255 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3257 if (info->sva + PAGE_SIZE == info->eva) {
3258 generation = pmap->pm_generation;
3259 if (info->sva >= VM_MAX_USER_ADDRESS) {
3261 * Kernel mappings do not track wire counts on
3262 * page table pages and only maintain pd_pv and
3263 * pte_pv levels so pmap_scan() works.
3266 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3267 ptep = vtopte(info->sva);
3270 * User pages which are unmanaged will not have a
3271 * pte_pv. User page table pages which are unmanaged
3272 * (shared from elsewhere) will also not have a pt_pv.
3273 * The func() callback will pass both pte_pv and pt_pv
3274 * as NULL in that case.
3276 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3277 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3278 if (pt_pv == NULL) {
3279 KKASSERT(pte_pv == NULL);
3280 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3282 ptep = pv_pte_lookup(pd_pv,
3283 pmap_pt_index(info->sva));
3285 info->func(pmap, info,
3294 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3298 * NOTE: *ptep can't be ripped out from under us if we hold
3299 * pte_pv locked, but bits can change. However, there is
3300 * a race where another thread may be inserting pte_pv
3301 * and setting *ptep just after our pte_pv lookup fails.
3303 * In this situation we can end up with a NULL pte_pv
3304 * but find that we have a managed *ptep. We explicitly
3305 * check for this race.
3311 * Unlike the pv_find() case below we actually
3312 * acquired a locked pv in this case so any
3313 * race should have been resolved. It is expected
3316 KKASSERT(pte_pv == NULL);
3317 } else if (pte_pv) {
3318 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3319 pmap->pmap_bits[PG_V_IDX])) ==
3320 (pmap->pmap_bits[PG_MANAGED_IDX] |
3321 pmap->pmap_bits[PG_V_IDX]),
3322 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3324 *ptep, oldpte, info->sva, pte_pv,
3325 generation, pmap->pm_generation));
3326 info->func(pmap, info, pte_pv, pt_pv, 0,
3327 info->sva, ptep, info->arg);
3330 * Check for insertion race
3332 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3334 pte_pv = pv_find(pmap,
3335 pmap_pte_pindex(info->sva));
3339 kprintf("pmap_scan: RACE1 "
3349 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3350 pmap->pmap_bits[PG_V_IDX])) ==
3351 pmap->pmap_bits[PG_V_IDX],
3352 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3354 *ptep, oldpte, info->sva,
3355 generation, pmap->pm_generation));
3356 info->func(pmap, info, NULL, pt_pv, 0,
3357 info->sva, ptep, info->arg);
3362 lwkt_reltoken(&pmap->pm_token);
3367 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3370 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3371 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3373 if (info->sva >= VM_MAX_USER_ADDRESS) {
3375 * The kernel does not currently maintain any pv_entry's for
3376 * higher-level page tables.
3378 bzero(&dummy_pv, sizeof(dummy_pv));
3379 dummy_pv.pv_pindex = info->sva_pd_pindex;
3380 spin_lock(&pmap->pm_spin);
3381 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3382 pmap_scan_callback(&dummy_pv, info);
3383 ++dummy_pv.pv_pindex;
3385 spin_unlock(&pmap->pm_spin);
3388 * User page tables maintain local PML4, PDP, and PD
3389 * pv_entry's at the very least. PT pv's might be
3390 * unmanaged and thus not exist. PTE pv's might be
3391 * unmanaged and thus not exist.
3393 spin_lock(&pmap->pm_spin);
3394 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3395 pmap_scan_cmp, pmap_scan_callback, info);
3396 spin_unlock(&pmap->pm_spin);
3398 lwkt_reltoken(&pmap->pm_token);
3402 * WARNING! pmap->pm_spin held
3405 pmap_scan_cmp(pv_entry_t pv, void *data)
3407 struct pmap_scan_info *info = data;
3408 if (pv->pv_pindex < info->sva_pd_pindex)
3410 if (pv->pv_pindex >= info->eva_pd_pindex)
3416 * WARNING! pmap->pm_spin held
3419 pmap_scan_callback(pv_entry_t pv, void *data)
3421 struct pmap_scan_info *info = data;
3422 struct pmap *pmap = info->pmap;
3423 pv_entry_t pd_pv; /* A page directory PV */
3424 pv_entry_t pt_pv; /* A page table PV */
3425 pv_entry_t pte_pv; /* A page table entry PV */
3430 vm_offset_t va_next;
3431 vm_pindex_t pd_pindex;
3436 * Pull the PD pindex from the pv before releasing the spinlock.
3438 * WARNING: pv is faked for kernel pmap scans.
3440 pd_pindex = pv->pv_pindex;
3441 spin_unlock(&pmap->pm_spin);
3442 pv = NULL; /* invalid after spinlock unlocked */
3445 * Calculate the page range within the PD. SIMPLE pmaps are
3446 * direct-mapped for the entire 2^64 address space. Normal pmaps
3447 * reflect the user and kernel address space which requires
3448 * cannonicalization w/regards to converting pd_pindex's back
3451 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3452 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3453 (sva & PML4_SIGNMASK)) {
3454 sva |= PML4_SIGNMASK;
3456 eva = sva + NBPDP; /* can overflow */
3457 if (sva < info->sva)
3459 if (eva < info->sva || eva > info->eva)
3463 * NOTE: kernel mappings do not track page table pages, only
3466 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3467 * However, for the scan to be efficient we try to
3468 * cache items top-down.
3473 for (; sva < eva; sva = va_next) {
3474 if (sva >= VM_MAX_USER_ADDRESS) {
3483 * PD cache (degenerate case if we skip). It is possible
3484 * for the PD to not exist due to races. This is ok.
3486 if (pd_pv == NULL) {
3487 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3488 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3490 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3492 if (pd_pv == NULL) {
3493 va_next = (sva + NBPDP) & ~PDPMASK;
3502 if (pt_pv == NULL) {
3507 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3508 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3514 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3518 * If pt_pv is NULL we either have an shared page table
3519 * page and must issue a callback specific to that case,
3520 * or there is no page table page.
3522 * Either way we can skip the page table page.
3524 if (pt_pv == NULL) {
3526 * Possible unmanaged (shared from another pmap)
3530 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3531 KKASSERT(pd_pv != NULL);
3532 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3533 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3534 info->func(pmap, info, NULL, pd_pv, 1,
3535 sva, ptep, info->arg);
3539 * Done, move to next page table page.
3541 va_next = (sva + NBPDR) & ~PDRMASK;
3548 * From this point in the loop testing pt_pv for non-NULL
3549 * means we are in UVM, else if it is NULL we are in KVM.
3551 * Limit our scan to either the end of the va represented
3552 * by the current page table page, or to the end of the
3553 * range being removed.
3556 va_next = (sva + NBPDR) & ~PDRMASK;
3563 * Scan the page table for pages. Some pages may not be
3564 * managed (might not have a pv_entry).
3566 * There is no page table management for kernel pages so
3567 * pt_pv will be NULL in that case, but otherwise pt_pv
3568 * is non-NULL, locked, and referenced.
3572 * At this point a non-NULL pt_pv means a UVA, and a NULL
3573 * pt_pv means a KVA.
3576 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3580 while (sva < va_next) {
3582 * Yield every 64 pages.
3584 if ((++info->count & 63) == 0)
3588 * Acquire the related pte_pv, if any. If *ptep == 0
3589 * the related pte_pv should not exist, but if *ptep
3590 * is not zero the pte_pv may or may not exist (e.g.
3591 * will not exist for an unmanaged page).
3593 * However a multitude of races are possible here.
3595 * In addition, the (pt_pv, pte_pv) lock order is
3596 * backwards, so we have to be careful in aquiring
3597 * a properly locked pte_pv.
3599 generation = pmap->pm_generation;
3601 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3608 pv_put(pt_pv); /* must be non-NULL */
3610 pv_lock(pte_pv); /* safe to block now */
3613 pt_pv = pv_get(pmap,
3614 pmap_pt_pindex(sva));
3616 * pt_pv reloaded, need new ptep
3618 KKASSERT(pt_pv != NULL);
3619 ptep = pv_pte_lookup(pt_pv,
3620 pmap_pte_index(sva));
3624 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3628 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3633 kprintf("Unexpected non-NULL pte_pv "
3635 "*ptep = %016lx/%016lx\n",
3636 pte_pv, pt_pv, *ptep, oldpte);
3637 panic("Unexpected non-NULL pte_pv");
3645 * Ready for the callback. The locked pte_pv (if any)
3646 * is consumed by the callback. pte_pv will exist if
3647 * the page is managed, and will not exist if it
3651 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3652 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3653 ("badC *ptep %016lx/%016lx sva %016lx "
3654 "pte_pv %p pm_generation %d/%d",
3655 *ptep, oldpte, sva, pte_pv,
3656 generation, pmap->pm_generation));
3657 info->func(pmap, info, pte_pv, pt_pv, 0,
3658 sva, ptep, info->arg);
3661 * Check for insertion race. Since there is no
3662 * pte_pv to guard us it is possible for us
3663 * to race another thread doing an insertion.
3664 * Our lookup misses the pte_pv but our *ptep
3665 * check sees the inserted pte.
3667 * XXX panic case seems to occur within a
3668 * vm_fork() of /bin/sh, which frankly
3669 * shouldn't happen since no other threads
3670 * should be inserting to our pmap in that
3671 * situation. Removing, possibly. Inserting,
3674 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3676 pte_pv = pv_find(pmap,
3677 pmap_pte_pindex(sva));
3680 kprintf("pmap_scan: RACE2 "
3690 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3691 pmap->pmap_bits[PG_V_IDX],
3692 ("badD *ptep %016lx/%016lx sva %016lx "
3693 "pte_pv NULL pm_generation %d/%d",
3695 generation, pmap->pm_generation));
3696 info->func(pmap, info, NULL, pt_pv, 0,
3697 sva, ptep, info->arg);
3712 if ((++info->count & 7) == 0)
3716 * Relock before returning.
3718 spin_lock(&pmap->pm_spin);
3723 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3725 struct pmap_scan_info info;
3730 info.func = pmap_remove_callback;
3732 info.dosmp = 1; /* normal remove requires pmap inval */
3737 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3739 struct pmap_scan_info info;
3744 info.func = pmap_remove_callback;
3746 info.dosmp = 0; /* do not synchronize w/other cpus */
3751 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
3752 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3753 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3759 * This will also drop pt_pv's wire_count. Note that
3760 * terminal pages are not wired based on mmu presence.
3762 pmap_remove_pv_pte(pte_pv, pt_pv, info->dosmp);
3763 pmap_remove_pv_page(pte_pv);
3765 } else if (sharept == 0) {
3767 * Unmanaged page table (pt, pd, or pdp. Not pte).
3769 * pt_pv's wire_count is still bumped by unmanaged pages
3770 * so we must decrement it manually.
3772 * We have to unwire the target page table page.
3774 * It is unclear how we can invalidate a segment so we
3775 * invalidate -1 which invlidates the tlb.
3778 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, ptep, 0);
3780 pte = pte_load_clear(ptep);
3781 if (pte & pmap->pmap_bits[PG_W_IDX])
3782 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3783 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3784 if (vm_page_unwire_quick(pt_pv->pv_m))
3785 panic("pmap_remove: insufficient wirecount");
3788 * Unmanaged page table (pt, pd, or pdp. Not pte) for
3789 * a shared page table.
3791 * pt_pv is actually the pd_pv for our pmap (not the shared
3794 * We have to unwire the target page table page and we
3795 * have to unwire our page directory page.
3797 * It is unclear how we can invalidate a segment so we
3798 * invalidate -1 which invlidates the tlb.
3801 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, ptep, 0);
3803 pte = pte_load_clear(ptep);
3804 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3805 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
3806 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3807 panic("pmap_remove: shared pgtable1 bad wirecount");
3808 if (vm_page_unwire_quick(pt_pv->pv_m))
3809 panic("pmap_remove: shared pgtable2 bad wirecount");
3814 * Removes this physical page from all physical maps in which it resides.
3815 * Reflects back modify bits to the pager.
3817 * This routine may not be called from an interrupt.
3821 pmap_remove_all(vm_page_t m)
3825 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
3828 vm_page_spin_lock(m);
3829 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3830 KKASSERT(pv->pv_m == m);
3831 if (pv_hold_try(pv)) {
3832 vm_page_spin_unlock(m);
3834 vm_page_spin_unlock(m);
3837 if (pv->pv_m != m) {
3839 vm_page_spin_lock(m);
3844 * Holding no spinlocks, pv is locked.
3846 pmap_remove_pv_pte(pv, NULL, 1);
3847 pmap_remove_pv_page(pv);
3849 vm_page_spin_lock(m);
3851 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3852 vm_page_spin_unlock(m);
3856 * Set the physical protection on the specified range of this map
3857 * as requested. This function is typically only used for debug watchpoints
3860 * This function may not be called from an interrupt if the map is
3861 * not the kernel_pmap.
3863 * NOTE! For shared page table pages we just unmap the page.
3866 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3868 struct pmap_scan_info info;
3869 /* JG review for NX */
3873 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3874 pmap_remove(pmap, sva, eva);
3877 if (prot & VM_PROT_WRITE)
3882 info.func = pmap_protect_callback;
3890 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
3891 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3892 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3904 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
3905 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3906 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3907 KKASSERT(m == pte_pv->pv_m);
3908 vm_page_flag_set(m, PG_REFERENCED);
3910 cbits &= ~pmap->pmap_bits[PG_A_IDX];
3912 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
3913 if (pmap_track_modified(pte_pv->pv_pindex)) {
3914 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3916 m = PHYS_TO_VM_PAGE(pbits &
3921 cbits &= ~pmap->pmap_bits[PG_M_IDX];
3924 } else if (sharept) {
3926 * Unmanaged page table, pt_pv is actually the pd_pv
3927 * for our pmap (not the object's shared pmap).
3929 * When asked to protect something in a shared page table
3930 * page we just unmap the page table page. We have to
3931 * invalidate the tlb in this situation.
3933 * XXX Warning, shared page tables will not be used for
3934 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
3935 * so PHYS_TO_VM_PAGE() should be safe here.
3937 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, ptep, 0);
3938 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3939 panic("pmap_protect: pgtable1 pg bad wirecount");
3940 if (vm_page_unwire_quick(pt_pv->pv_m))
3941 panic("pmap_protect: pgtable2 pg bad wirecount");
3944 /* else unmanaged page, adjust bits, no wire changes */
3947 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
3949 if (pmap_enter_debug > 0) {
3951 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
3952 "pt_pv=%p cbits=%08lx\n",
3958 if (pbits != cbits) {
3959 if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
3960 ptep, pbits, cbits)) {
3970 * Insert the vm_page (m) at the virtual address (va), replacing any prior
3971 * mapping at that address. Set protection and wiring as requested.
3973 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
3974 * possible. If it is we enter the page into the appropriate shared pmap
3975 * hanging off the related VM object instead of the passed pmap, then we
3976 * share the page table page from the VM object's pmap into the current pmap.
3978 * NOTE: This routine MUST insert the page into the pmap now, it cannot
3982 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3983 boolean_t wired, vm_map_entry_t entry)
3985 pv_entry_t pt_pv; /* page table */
3986 pv_entry_t pte_pv; /* page table entry */
3989 pt_entry_t origpte, newpte;
3994 va = trunc_page(va);
3995 #ifdef PMAP_DIAGNOSTIC
3997 panic("pmap_enter: toobig");
3998 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
3999 panic("pmap_enter: invalid to pmap_enter page table "
4000 "pages (va: 0x%lx)", va);
4002 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4003 kprintf("Warning: pmap_enter called on UVA with "
4006 db_print_backtrace();
4009 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4010 kprintf("Warning: pmap_enter called on KVA without"
4013 db_print_backtrace();
4018 * Get locked PV entries for our new page table entry (pte_pv)
4019 * and for its parent page table (pt_pv). We need the parent
4020 * so we can resolve the location of the ptep.
4022 * Only hardware MMU actions can modify the ptep out from
4025 * if (m) is fictitious or unmanaged we do not create a managing
4026 * pte_pv for it. Any pre-existing page's management state must
4027 * match (avoiding code complexity).
4029 * If the pmap is still being initialized we assume existing
4032 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4034 if (pmap_initialized == FALSE) {
4039 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4041 if (va >= VM_MAX_USER_ADDRESS) {
4045 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4047 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4051 KASSERT(origpte == 0 ||
4052 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4053 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4055 if (va >= VM_MAX_USER_ADDRESS) {
4057 * Kernel map, pv_entry-tracked.
4060 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4066 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4068 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4072 KASSERT(origpte == 0 ||
4073 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4074 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4077 pa = VM_PAGE_TO_PHYS(m);
4078 opa = origpte & PG_FRAME;
4080 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4081 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4083 newpte |= pmap->pmap_bits[PG_W_IDX];
4084 if (va < VM_MAX_USER_ADDRESS)
4085 newpte |= pmap->pmap_bits[PG_U_IDX];
4087 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4088 // if (pmap == &kernel_pmap)
4089 // newpte |= pgeflag;
4090 newpte |= pmap->pmap_cache_bits[m->pat_mode];
4091 if (m->flags & PG_FICTITIOUS)
4092 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4095 * It is possible for multiple faults to occur in threaded
4096 * environments, the existing pte might be correct.
4098 if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4099 pmap->pmap_bits[PG_A_IDX])) == 0)
4103 * Ok, either the address changed or the protection or wiring
4106 * Clear the current entry, interlocking the removal. For managed
4107 * pte's this will also flush the modified state to the vm_page.
4108 * Atomic ops are mandatory in order to ensure that PG_M events are
4109 * not lost during any transition.
4111 * WARNING: The caller has busied the new page but not the original
4112 * vm_page which we are trying to replace. Because we hold
4113 * the pte_pv lock, but have not busied the page, PG bits
4114 * can be cleared out from under us.
4119 * pmap_remove_pv_pte() unwires pt_pv and assumes
4120 * we will free pte_pv, but since we are reusing
4121 * pte_pv we want to retain the wire count.
4123 * pt_pv won't exist for a kernel page (managed or
4127 vm_page_wire_quick(pt_pv->pv_m);
4128 if (prot & VM_PROT_NOSYNC)
4129 pmap_remove_pv_pte(pte_pv, pt_pv, 0);
4131 pmap_remove_pv_pte(pte_pv, pt_pv, 1);
4133 pmap_remove_pv_page(pte_pv);
4134 } else if (prot & VM_PROT_NOSYNC) {
4136 * Unmanaged page, NOSYNC (no mmu sync) requested.
4138 * Leave wire count on PT page intact.
4140 (void)pte_load_clear(ptep);
4141 cpu_invlpg((void *)va);
4142 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4145 * Unmanaged page, normal enter.
4147 * Leave wire count on PT page intact.
4149 pmap_inval_smp(pmap, va, ptep, 0);
4150 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4152 KKASSERT(*ptep == 0);
4156 if (pmap_enter_debug > 0) {
4158 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4159 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4161 origpte, newpte, ptep,
4162 pte_pv, pt_pv, opa, prot);
4168 * Enter on the PV list if part of our managed memory.
4169 * Wiring of the PT page is already handled.
4171 KKASSERT(pte_pv->pv_m == NULL);
4172 vm_page_spin_lock(m);
4174 pmap_page_stats_adding(m);
4175 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4176 vm_page_flag_set(m, PG_MAPPED);
4177 vm_page_spin_unlock(m);
4178 } else if (pt_pv && opa == 0) {
4180 * We have to adjust the wire count on the PT page ourselves
4181 * for unmanaged entries. If opa was non-zero we retained
4182 * the existing wire count from the removal.
4184 vm_page_wire_quick(pt_pv->pv_m);
4188 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4190 * User VMAs do not because those will be zero->non-zero, so no
4191 * stale entries to worry about at this point.
4193 * For KVM there appear to still be issues. Theoretically we
4194 * should be able to scrap the interlocks entirely but we
4197 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4198 pmap_inval_smp(pmap, va, ptep, newpte);
4200 *(volatile pt_entry_t *)ptep = newpte;
4202 cpu_invlpg((void *)va);
4207 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4210 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4213 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4214 vm_page_flag_set(m, PG_WRITEABLE);
4217 * Unmanaged pages need manual resident_count tracking.
4219 if (pte_pv == NULL && pt_pv)
4220 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4226 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4227 (m->flags & PG_MAPPED));
4230 * Cleanup the pv entry, allowing other accessors.
4239 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4240 * This code also assumes that the pmap has no pre-existing entry for this
4243 * This code currently may only be used on user pmaps, not kernel_pmap.
4246 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4248 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4252 * Make a temporary mapping for a physical address. This is only intended
4253 * to be used for panic dumps.
4255 * The caller is responsible for calling smp_invltlb().
4258 pmap_kenter_temporary(vm_paddr_t pa, long i)
4260 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4261 return ((void *)crashdumpmap);
4264 #define MAX_INIT_PT (96)
4267 * This routine preloads the ptes for a given object into the specified pmap.
4268 * This eliminates the blast of soft faults on process startup and
4269 * immediately after an mmap.
4271 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4274 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4275 vm_object_t object, vm_pindex_t pindex,
4276 vm_size_t size, int limit)
4278 struct rb_vm_page_scan_info info;
4283 * We can't preinit if read access isn't set or there is no pmap
4286 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4290 * We can't preinit if the pmap is not the current pmap
4292 lp = curthread->td_lwp;
4293 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4297 * Misc additional checks
4299 psize = x86_64_btop(size);
4301 if ((object->type != OBJT_VNODE) ||
4302 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4303 (object->resident_page_count > MAX_INIT_PT))) {
4307 if (pindex + psize > object->size) {
4308 if (object->size < pindex)
4310 psize = object->size - pindex;
4317 * If everything is segment-aligned do not pre-init here. Instead
4318 * allow the normal vm_fault path to pass a segment hint to
4319 * pmap_enter() which will then use an object-referenced shared
4322 if ((addr & SEG_MASK) == 0 &&
4323 (ctob(psize) & SEG_MASK) == 0 &&
4324 (ctob(pindex) & SEG_MASK) == 0) {
4329 * Use a red-black scan to traverse the requested range and load
4330 * any valid pages found into the pmap.
4332 * We cannot safely scan the object's memq without holding the
4335 info.start_pindex = pindex;
4336 info.end_pindex = pindex + psize - 1;
4342 vm_object_hold_shared(object);
4343 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4344 pmap_object_init_pt_callback, &info);
4345 vm_object_drop(object);
4350 pmap_object_init_pt_callback(vm_page_t p, void *data)
4352 struct rb_vm_page_scan_info *info = data;
4353 vm_pindex_t rel_index;
4356 * don't allow an madvise to blow away our really
4357 * free pages allocating pv entries.
4359 if ((info->limit & MAP_PREFAULT_MADVISE) &&
4360 vmstats.v_free_count < vmstats.v_free_reserved) {
4365 * Ignore list markers and ignore pages we cannot instantly
4366 * busy (while holding the object token).
4368 if (p->flags & PG_MARKER)
4370 if (vm_page_busy_try(p, TRUE))
4372 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4373 (p->flags & PG_FICTITIOUS) == 0) {
4374 if ((p->queue - p->pc) == PQ_CACHE)
4375 vm_page_deactivate(p);
4376 rel_index = p->pindex - info->start_pindex;
4377 pmap_enter_quick(info->pmap,
4378 info->addr + x86_64_ptob(rel_index), p);
4386 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4389 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4392 * XXX This is safe only because page table pages are not freed.
4395 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4399 /*spin_lock(&pmap->pm_spin);*/
4400 if ((pte = pmap_pte(pmap, addr)) != NULL) {
4401 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4402 /*spin_unlock(&pmap->pm_spin);*/
4406 /*spin_unlock(&pmap->pm_spin);*/
4411 * Change the wiring attribute for a pmap/va pair. The mapping must already
4412 * exist in the pmap. The mapping may or may not be managed.
4415 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4416 vm_map_entry_t entry)
4423 lwkt_gettoken(&pmap->pm_token);
4424 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
4425 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4427 if (wired && !pmap_pte_w(pmap, ptep))
4428 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4429 else if (!wired && pmap_pte_w(pmap, ptep))
4430 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4433 * Wiring is not a hardware characteristic so there is no need to
4434 * invalidate TLB. However, in an SMP environment we must use
4435 * a locked bus cycle to update the pte (if we are not using
4436 * the pmap_inval_*() API that is)... it's ok to do this for simple
4440 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4442 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4444 lwkt_reltoken(&pmap->pm_token);
4450 * Copy the range specified by src_addr/len from the source map to
4451 * the range dst_addr/len in the destination map.
4453 * This routine is only advisory and need not do anything.
4456 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4457 vm_size_t len, vm_offset_t src_addr)
4464 * Zero the specified physical page.
4466 * This function may be called from an interrupt and no locking is
4470 pmap_zero_page(vm_paddr_t phys)
4472 vm_offset_t va = PHYS_TO_DMAP(phys);
4474 pagezero((void *)va);
4478 * pmap_page_assertzero:
4480 * Assert that a page is empty, panic if it isn't.
4483 pmap_page_assertzero(vm_paddr_t phys)
4485 vm_offset_t va = PHYS_TO_DMAP(phys);
4488 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
4489 if (*(long *)((char *)va + i) != 0) {
4490 panic("pmap_page_assertzero() @ %p not zero!",
4491 (void *)(intptr_t)va);
4499 * Zero part of a physical page by mapping it into memory and clearing
4500 * its contents with bzero.
4502 * off and size may not cover an area beyond a single hardware page.
4505 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4507 vm_offset_t virt = PHYS_TO_DMAP(phys);
4509 bzero((char *)virt + off, size);
4515 * Copy the physical page from the source PA to the target PA.
4516 * This function may be called from an interrupt. No locking
4520 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4522 vm_offset_t src_virt, dst_virt;
4524 src_virt = PHYS_TO_DMAP(src);
4525 dst_virt = PHYS_TO_DMAP(dst);
4526 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4530 * pmap_copy_page_frag:
4532 * Copy the physical page from the source PA to the target PA.
4533 * This function may be called from an interrupt. No locking
4537 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4539 vm_offset_t src_virt, dst_virt;
4541 src_virt = PHYS_TO_DMAP(src);
4542 dst_virt = PHYS_TO_DMAP(dst);
4544 bcopy((char *)src_virt + (src & PAGE_MASK),
4545 (char *)dst_virt + (dst & PAGE_MASK),
4550 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4551 * this page. This count may be changed upwards or downwards in the future;
4552 * it is only necessary that true be returned for a small subset of pmaps
4553 * for proper page aging.
4556 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4561 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4564 vm_page_spin_lock(m);
4565 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4566 if (pv->pv_pmap == pmap) {
4567 vm_page_spin_unlock(m);
4574 vm_page_spin_unlock(m);
4579 * Remove all pages from specified address space this aids process exit
4580 * speeds. Also, this code may be special cased for the current process
4584 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4586 pmap_remove_noinval(pmap, sva, eva);
4591 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4592 * routines are inline, and a lot of things compile-time evaluate.
4596 pmap_testbit(vm_page_t m, int bit)
4602 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4605 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4607 vm_page_spin_lock(m);
4608 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4609 vm_page_spin_unlock(m);
4613 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4615 #if defined(PMAP_DIAGNOSTIC)
4616 if (pv->pv_pmap == NULL) {
4617 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4625 * If the bit being tested is the modified bit, then
4626 * mark clean_map and ptes as never
4629 * WARNING! Because we do not lock the pv, *pte can be in a
4630 * state of flux. Despite this the value of *pte
4631 * will still be related to the vm_page in some way
4632 * because the pv cannot be destroyed as long as we
4633 * hold the vm_page spin lock.
4635 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4636 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4637 if (!pmap_track_modified(pv->pv_pindex))
4641 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4642 if (*pte & pmap->pmap_bits[bit]) {
4643 vm_page_spin_unlock(m);
4647 vm_page_spin_unlock(m);
4652 * This routine is used to modify bits in ptes. Only one bit should be
4653 * specified. PG_RW requires special handling.
4655 * Caller must NOT hold any spin locks
4659 pmap_clearbit(vm_page_t m, int bit_index)
4666 if (bit_index == PG_RW_IDX)
4667 vm_page_flag_clear(m, PG_WRITEABLE);
4668 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4675 * Loop over all current mappings setting/clearing as appropos If
4676 * setting RO do we need to clear the VAC?
4678 * NOTE: When clearing PG_M we could also (not implemented) drop
4679 * through to the PG_RW code and clear PG_RW too, forcing
4680 * a fault on write to redetect PG_M for virtual kernels, but
4681 * it isn't necessary since virtual kernels invalidate the
4682 * pte when they clear the VPTE_M bit in their virtual page
4685 * NOTE: Does not re-dirty the page when clearing only PG_M.
4687 * NOTE: Because we do not lock the pv, *pte can be in a state of
4688 * flux. Despite this the value of *pte is still somewhat
4689 * related while we hold the vm_page spin lock.
4691 * *pte can be zero due to this race. Since we are clearing
4692 * bits we basically do no harm when this race ccurs.
4694 if (bit_index != PG_RW_IDX) {
4695 vm_page_spin_lock(m);
4696 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4697 #if defined(PMAP_DIAGNOSTIC)
4698 if (pv->pv_pmap == NULL) {
4699 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4705 pte = pmap_pte_quick(pv->pv_pmap,
4706 pv->pv_pindex << PAGE_SHIFT);
4708 if (pbits & pmap->pmap_bits[bit_index])
4709 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
4711 vm_page_spin_unlock(m);
4716 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4720 vm_page_spin_lock(m);
4721 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4723 * don't write protect pager mappings
4725 if (!pmap_track_modified(pv->pv_pindex))
4728 #if defined(PMAP_DIAGNOSTIC)
4729 if (pv->pv_pmap == NULL) {
4730 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4737 * Skip pages which do not have PG_RW set.
4739 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4740 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
4746 if (pv_hold_try(pv)) {
4747 vm_page_spin_unlock(m);
4749 vm_page_spin_unlock(m);
4750 pv_lock(pv); /* held, now do a blocking lock */
4752 if (pv->pv_pmap != pmap || pv->pv_m != m) {
4753 pv_put(pv); /* and release */
4754 goto restart; /* anything could have happened */
4756 KKASSERT(pv->pv_pmap == pmap);
4762 nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
4763 pmap->pmap_bits[PG_M_IDX]);
4764 if (pmap_inval_smp_cmpset(pmap,
4765 ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
4766 pte, pbits, nbits)) {
4771 vm_page_spin_lock(m);
4774 * If PG_M was found to be set while we were clearing PG_RW
4775 * we also clear PG_M (done above) and mark the page dirty.
4776 * Callers expect this behavior.
4778 if (pbits & pmap->pmap_bits[PG_M_IDX])
4782 vm_page_spin_unlock(m);
4786 * Lower the permission for all mappings to a given page.
4788 * Page must be busied by caller. Because page is busied by caller this
4789 * should not be able to race a pmap_enter().
4792 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4794 /* JG NX support? */
4795 if ((prot & VM_PROT_WRITE) == 0) {
4796 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4798 * NOTE: pmap_clearbit(.. PG_RW) also clears
4799 * the PG_WRITEABLE flag in (m).
4801 pmap_clearbit(m, PG_RW_IDX);
4809 pmap_phys_address(vm_pindex_t ppn)
4811 return (x86_64_ptob(ppn));
4815 * Return a count of reference bits for a page, clearing those bits.
4816 * It is not necessary for every reference bit to be cleared, but it
4817 * is necessary that 0 only be returned when there are truly no
4818 * reference bits set.
4820 * XXX: The exact number of bits to check and clear is a matter that
4821 * should be tested and standardized at some point in the future for
4822 * optimal aging of shared pages.
4824 * This routine may not block.
4827 pmap_ts_referenced(vm_page_t m)
4834 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4837 vm_page_spin_lock(m);
4838 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4839 if (!pmap_track_modified(pv->pv_pindex))
4842 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4843 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
4844 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
4850 vm_page_spin_unlock(m);
4857 * Return whether or not the specified physical page was modified
4858 * in any physical maps.
4861 pmap_is_modified(vm_page_t m)
4865 res = pmap_testbit(m, PG_M_IDX);
4870 * Clear the modify bits on the specified physical page.
4873 pmap_clear_modify(vm_page_t m)
4875 pmap_clearbit(m, PG_M_IDX);
4879 * pmap_clear_reference:
4881 * Clear the reference bit on the specified physical page.
4884 pmap_clear_reference(vm_page_t m)
4886 pmap_clearbit(m, PG_A_IDX);
4890 * Miscellaneous support routines follow
4895 i386_protection_init(void)
4899 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4900 kp = protection_codes;
4901 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
4903 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4905 * Read access is also 0. There isn't any execute bit,
4906 * so just make it readable.
4908 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4909 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4910 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4913 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4914 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4915 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4916 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4917 *kp++ = pmap_bits_default[PG_RW_IDX];
4924 * Map a set of physical memory pages into the kernel virtual
4925 * address space. Return a pointer to where it is mapped. This
4926 * routine is intended to be used for mapping device memory,
4929 * NOTE: We can't use pgeflag unless we invalidate the pages one at
4932 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
4933 * work whether the cpu supports PAT or not. The remaining PAT
4934 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
4938 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
4940 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4944 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
4946 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
4950 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
4952 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4956 * Map a set of physical memory pages into the kernel virtual
4957 * address space. Return a pointer to where it is mapped. This
4958 * routine is intended to be used for mapping device memory,
4962 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
4964 vm_offset_t va, tmpva, offset;
4968 offset = pa & PAGE_MASK;
4969 size = roundup(offset + size, PAGE_SIZE);
4971 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4973 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4975 pa = pa & ~PAGE_MASK;
4976 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
4977 pte = vtopte(tmpva);
4979 kernel_pmap.pmap_bits[PG_RW_IDX] |
4980 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
4981 kernel_pmap.pmap_cache_bits[mode];
4982 tmpsize -= PAGE_SIZE;
4986 pmap_invalidate_range(&kernel_pmap, va, va + size);
4987 pmap_invalidate_cache_range(va, va + size);
4989 return ((void *)(va + offset));
4993 pmap_unmapdev(vm_offset_t va, vm_size_t size)
4995 vm_offset_t base, offset;
4997 base = va & ~PAGE_MASK;
4998 offset = va & PAGE_MASK;
4999 size = roundup(offset + size, PAGE_SIZE);
5000 pmap_qremove(va, size >> PAGE_SHIFT);
5001 kmem_free(&kernel_map, base, size);
5005 * Sets the memory attribute for the specified page.
5008 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5014 * If "m" is a normal page, update its direct mapping. This update
5015 * can be relied upon to perform any cache operations that are
5016 * required for data coherence.
5018 if ((m->flags & PG_FICTITIOUS) == 0)
5019 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5023 * Change the PAT attribute on an existing kernel memory map. Caller
5024 * must ensure that the virtual memory in question is not accessed
5025 * during the adjustment.
5028 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5035 panic("pmap_change_attr: va is NULL");
5036 base = trunc_page(va);
5040 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5041 kernel_pmap.pmap_cache_bits[mode];
5046 changed = 1; /* XXX: not optimal */
5049 * Flush CPU caches if required to make sure any data isn't cached that
5050 * shouldn't be, etc.
5053 pmap_invalidate_range(&kernel_pmap, base, va);
5054 pmap_invalidate_cache_range(base, va);
5059 * perform the pmap work for mincore
5062 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5064 pt_entry_t *ptep, pte;
5068 lwkt_gettoken(&pmap->pm_token);
5069 ptep = pmap_pte(pmap, addr);
5071 if (ptep && (pte = *ptep) != 0) {
5074 val = MINCORE_INCORE;
5075 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5078 pa = pte & PG_FRAME;
5080 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5083 m = PHYS_TO_VM_PAGE(pa);
5088 if (pte & pmap->pmap_bits[PG_M_IDX])
5089 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5091 * Modified by someone
5093 else if (m && (m->dirty || pmap_is_modified(m)))
5094 val |= MINCORE_MODIFIED_OTHER;
5098 if (pte & pmap->pmap_bits[PG_A_IDX])
5099 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5102 * Referenced by someone
5104 else if (m && ((m->flags & PG_REFERENCED) ||
5105 pmap_ts_referenced(m))) {
5106 val |= MINCORE_REFERENCED_OTHER;
5107 vm_page_flag_set(m, PG_REFERENCED);
5111 lwkt_reltoken(&pmap->pm_token);
5117 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5118 * vmspace will be ref'd and the old one will be deref'd.
5120 * The vmspace for all lwps associated with the process will be adjusted
5121 * and cr3 will be reloaded if any lwp is the current lwp.
5123 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5126 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5128 struct vmspace *oldvm;
5131 oldvm = p->p_vmspace;
5132 if (oldvm != newvm) {
5135 p->p_vmspace = newvm;
5136 KKASSERT(p->p_nthreads == 1);
5137 lp = RB_ROOT(&p->p_lwp_tree);
5138 pmap_setlwpvm(lp, newvm);
5145 * Set the vmspace for a LWP. The vmspace is almost universally set the
5146 * same as the process vmspace, but virtual kernels need to swap out contexts
5147 * on a per-lwp basis.
5149 * Caller does not necessarily hold any vmspace tokens. Caller must control
5150 * the lwp (typically be in the context of the lwp). We use a critical
5151 * section to protect against statclock and hardclock (statistics collection).
5154 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5156 struct vmspace *oldvm;
5159 oldvm = lp->lwp_vmspace;
5161 if (oldvm != newvm) {
5163 lp->lwp_vmspace = newvm;
5164 if (curthread->td_lwp == lp) {
5165 pmap = vmspace_pmap(newvm);
5166 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5167 if (pmap->pm_active_lock & CPULOCK_EXCL)
5168 pmap_interlock_wait(newvm);
5169 #if defined(SWTCH_OPTIM_STATS)
5172 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5173 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5174 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5175 curthread->td_pcb->pcb_cr3 = KPML4phys;
5177 panic("pmap_setlwpvm: unknown pmap type\n");
5179 load_cr3(curthread->td_pcb->pcb_cr3);
5180 pmap = vmspace_pmap(oldvm);
5181 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5189 * Called when switching to a locked pmap, used to interlock against pmaps
5190 * undergoing modifications to prevent us from activating the MMU for the
5191 * target pmap until all such modifications have completed. We have to do
5192 * this because the thread making the modifications has already set up its
5193 * SMP synchronization mask.
5195 * This function cannot sleep!
5200 pmap_interlock_wait(struct vmspace *vm)
5202 struct pmap *pmap = &vm->vm_pmap;
5204 if (pmap->pm_active_lock & CPULOCK_EXCL) {
5206 KKASSERT(curthread->td_critcount >= 2);
5207 DEBUG_PUSH_INFO("pmap_interlock_wait");
5208 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5210 lwkt_process_ipiq();
5218 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5221 if ((obj == NULL) || (size < NBPDR) ||
5222 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5226 addr = roundup2(addr, NBPDR);
5231 * Used by kmalloc/kfree, page already exists at va
5234 pmap_kvtom(vm_offset_t va)
5236 pt_entry_t *ptep = vtopte(va);
5238 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5239 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5243 * Initialize machine-specific shared page directory support. This
5244 * is executed when a VM object is created.
5247 pmap_object_init(vm_object_t object)
5249 object->md.pmap_rw = NULL;
5250 object->md.pmap_ro = NULL;
5254 * Clean up machine-specific shared page directory support. This
5255 * is executed when a VM object is destroyed.
5258 pmap_object_free(vm_object_t object)
5262 if ((pmap = object->md.pmap_rw) != NULL) {
5263 object->md.pmap_rw = NULL;
5264 pmap_remove_noinval(pmap,
5265 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5266 CPUMASK_ASSZERO(pmap->pm_active);
5269 kfree(pmap, M_OBJPMAP);
5271 if ((pmap = object->md.pmap_ro) != NULL) {
5272 object->md.pmap_ro = NULL;
5273 pmap_remove_noinval(pmap,
5274 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5275 CPUMASK_ASSZERO(pmap->pm_active);
5278 kfree(pmap, M_OBJPMAP);