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, pmarkp) _pv_get(pmap, pindex, pmarkp \
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, pmarkp) _pv_get(pmap, pindex, pmarkp)
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;
156 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
158 vm_paddr_t avail_start; /* PA of first available physical page */
159 vm_paddr_t avail_end; /* PA of last available physical page */
160 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
161 vm_offset_t virtual2_end;
162 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
163 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
164 vm_offset_t KvaStart; /* VA start of KVA space */
165 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
166 vm_offset_t KvaSize; /* max size of kernel virtual address space */
167 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
168 //static int pgeflag; /* PG_G or-in */
169 //static int pseflag; /* PG_PS or-in */
173 static vm_paddr_t dmaplimit;
175 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
177 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE]; /* PAT -> PG_ bits */
178 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
180 static uint64_t KPTbase;
181 static uint64_t KPTphys;
182 static uint64_t KPDphys; /* phys addr of kernel level 2 */
183 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
184 uint64_t KPDPphys; /* phys addr of kernel level 3 */
185 uint64_t KPML4phys; /* phys addr of kernel level 4 */
187 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
188 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
191 * Data for the pv entry allocation mechanism
193 static vm_zone_t pvzone;
194 static struct vm_zone pvzone_store;
195 static struct vm_object pvzone_obj;
196 static int pv_entry_max=0, pv_entry_high_water=0;
197 static int pmap_pagedaemon_waken = 0;
198 static struct pv_entry *pvinit;
201 * All those kernel PT submaps that BSD is so fond of
203 pt_entry_t *CMAP1 = NULL, *ptmmap;
204 caddr_t CADDR1 = NULL, ptvmmap = NULL;
205 static pt_entry_t *msgbufmap;
206 struct msgbuf *msgbufp=NULL;
209 * PMAP default PG_* bits. Needed to be able to add
210 * EPT/NPT pagetable pmap_bits for the VMM module
212 uint64_t pmap_bits_default[] = {
213 REGULAR_PMAP, /* TYPE_IDX 0 */
214 X86_PG_V, /* PG_V_IDX 1 */
215 X86_PG_RW, /* PG_RW_IDX 2 */
216 X86_PG_U, /* PG_U_IDX 3 */
217 X86_PG_A, /* PG_A_IDX 4 */
218 X86_PG_M, /* PG_M_IDX 5 */
219 X86_PG_PS, /* PG_PS_IDX3 6 */
220 X86_PG_G, /* PG_G_IDX 7 */
221 X86_PG_AVAIL1, /* PG_AVAIL1_IDX 8 */
222 X86_PG_AVAIL2, /* PG_AVAIL2_IDX 9 */
223 X86_PG_AVAIL3, /* PG_AVAIL3_IDX 10 */
224 X86_PG_NC_PWT | X86_PG_NC_PCD, /* PG_N_IDX 11 */
229 static pt_entry_t *pt_crashdumpmap;
230 static caddr_t crashdumpmap;
232 static int pmap_debug = 0;
233 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
234 &pmap_debug, 0, "Debug pmap's");
236 static int pmap_enter_debug = 0;
237 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
238 &pmap_enter_debug, 0, "Debug pmap_enter's");
240 static int pmap_yield_count = 64;
241 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
242 &pmap_yield_count, 0, "Yield during init_pt/release");
243 static int pmap_mmu_optimize = 0;
244 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
245 &pmap_mmu_optimize, 0, "Share page table pages when possible");
246 int pmap_fast_kernel_cpusync = 0;
247 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
248 &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
249 int pmap_dynamic_delete = -1;
250 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
251 &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
255 /* Standard user access funtions */
256 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
258 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
259 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
260 extern int std_fubyte (const void *base);
261 extern int std_subyte (void *base, int byte);
262 extern long std_fuword (const void *base);
263 extern int std_suword (void *base, long word);
264 extern int std_suword32 (void *base, int word);
266 static void pv_hold(pv_entry_t pv);
267 static int _pv_hold_try(pv_entry_t pv
269 static void pv_drop(pv_entry_t pv);
270 static void _pv_lock(pv_entry_t pv
272 static void pv_unlock(pv_entry_t pv);
273 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
275 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
277 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
278 vm_pindex_t **pmarkp, int *errorp);
279 static void pv_put(pv_entry_t pv);
280 static void pv_free(pv_entry_t pv, pv_entry_t pvp);
281 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
282 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
284 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
285 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
286 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
287 pmap_inval_bulk_t *bulk, int destroy);
288 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
289 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
290 pmap_inval_bulk_t *bulk);
292 struct pmap_scan_info;
293 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
294 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
295 pv_entry_t pt_pv, int sharept,
296 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
297 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
298 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
299 pv_entry_t pt_pv, int sharept,
300 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
302 static void i386_protection_init (void);
303 static void create_pagetables(vm_paddr_t *firstaddr);
304 static void pmap_remove_all (vm_page_t m);
305 static boolean_t pmap_testbit (vm_page_t m, int bit);
307 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
308 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
310 static void pmap_pinit_defaults(struct pmap *pmap);
311 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
312 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
314 static unsigned pdir4mb;
317 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
319 if (pv1->pv_pindex < pv2->pv_pindex)
321 if (pv1->pv_pindex > pv2->pv_pindex)
326 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
327 pv_entry_compare, vm_pindex_t, pv_pindex);
331 pmap_page_stats_adding(vm_page_t m)
333 globaldata_t gd = mycpu;
335 if (TAILQ_EMPTY(&m->md.pv_list)) {
336 ++gd->gd_vmtotal.t_arm;
337 } else if (TAILQ_FIRST(&m->md.pv_list) ==
338 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
339 ++gd->gd_vmtotal.t_armshr;
340 ++gd->gd_vmtotal.t_avmshr;
342 ++gd->gd_vmtotal.t_avmshr;
348 pmap_page_stats_deleting(vm_page_t m)
350 globaldata_t gd = mycpu;
352 if (TAILQ_EMPTY(&m->md.pv_list)) {
353 --gd->gd_vmtotal.t_arm;
354 } else if (TAILQ_FIRST(&m->md.pv_list) ==
355 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
356 --gd->gd_vmtotal.t_armshr;
357 --gd->gd_vmtotal.t_avmshr;
359 --gd->gd_vmtotal.t_avmshr;
364 * Move the kernel virtual free pointer to the next
365 * 2MB. This is used to help improve performance
366 * by using a large (2MB) page for much of the kernel
367 * (.text, .data, .bss)
371 pmap_kmem_choose(vm_offset_t addr)
373 vm_offset_t newaddr = addr;
375 newaddr = roundup2(addr, NBPDR);
382 * Super fast pmap_pte routine best used when scanning the pv lists.
383 * This eliminates many course-grained invltlb calls. Note that many of
384 * the pv list scans are across different pmaps and it is very wasteful
385 * to do an entire invltlb when checking a single mapping.
387 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
391 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
393 return pmap_pte(pmap, va);
397 * Returns the pindex of a page table entry (representing a terminal page).
398 * There are NUPTE_TOTAL page table entries possible (a huge number)
400 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
401 * We want to properly translate negative KVAs.
405 pmap_pte_pindex(vm_offset_t va)
407 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
411 * Returns the pindex of a page table.
415 pmap_pt_pindex(vm_offset_t va)
417 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
421 * Returns the pindex of a page directory.
425 pmap_pd_pindex(vm_offset_t va)
427 return (NUPTE_TOTAL + NUPT_TOTAL +
428 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
433 pmap_pdp_pindex(vm_offset_t va)
435 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
436 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
441 pmap_pml4_pindex(void)
443 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
447 * Return various clipped indexes for a given VA
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 * The placemarker hash must be broken up into four zones so lock
483 * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
485 * Placemarkers are used to 'lock' page table indices that do not have
486 * a pv_entry. This allows the pmap to support managed and unmanaged
487 * pages and shared page tables.
489 #define PM_PLACE_BASE (PM_PLACEMARKS >> 2)
493 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
497 if (pindex < pmap_pt_pindex(0)) /* zone 0 - PTE */
499 else if (pindex < pmap_pd_pindex(0)) /* zone 1 - PT */
501 else if (pindex < pmap_pdp_pindex(0)) /* zone 2 - PD */
502 hi = PM_PLACE_BASE << 1;
503 else /* zone 3 - PDP (and PML4E) */
504 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
505 hi += pindex & (PM_PLACE_BASE - 1);
507 return (&pmap->pm_placemarks[hi]);
512 * Generic procedure to index a pte from a pt, pd, or pdp.
514 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
515 * a page table page index but is instead of PV lookup index.
519 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
523 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
524 return(&pte[pindex]);
528 * Return pointer to PDP slot in the PML4
532 pmap_pdp(pmap_t pmap, vm_offset_t va)
534 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
538 * Return pointer to PD slot in the PDP given a pointer to the PDP
542 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
546 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
547 return (&pd[pmap_pd_index(va)]);
551 * Return pointer to PD slot in the PDP.
555 pmap_pd(pmap_t pmap, vm_offset_t va)
559 pdp = pmap_pdp(pmap, va);
560 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
562 return (pmap_pdp_to_pd(*pdp, va));
566 * Return pointer to PT slot in the PD given a pointer to the PD
570 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
574 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
575 return (&pt[pmap_pt_index(va)]);
579 * Return pointer to PT slot in the PD
581 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
582 * so we cannot lookup the PD via the PDP. Instead we
583 * must look it up via the pmap.
587 pmap_pt(pmap_t pmap, vm_offset_t va)
591 vm_pindex_t pd_pindex;
593 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
594 pd_pindex = pmap_pd_pindex(va);
595 spin_lock(&pmap->pm_spin);
596 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
597 spin_unlock(&pmap->pm_spin);
598 if (pv == NULL || pv->pv_m == NULL)
600 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
602 pd = pmap_pd(pmap, va);
603 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
605 return (pmap_pd_to_pt(*pd, va));
610 * Return pointer to PTE slot in the PT given a pointer to the PT
614 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
618 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
619 return (&pte[pmap_pte_index(va)]);
623 * Return pointer to PTE slot in the PT
627 pmap_pte(pmap_t pmap, vm_offset_t va)
631 pt = pmap_pt(pmap, va);
632 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
634 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
635 return ((pt_entry_t *)pt);
636 return (pmap_pt_to_pte(*pt, va));
640 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
641 * the PT layer. This will speed up core pmap operations considerably.
643 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
644 * must be in a known associated state (typically by being locked when
645 * the pmap spinlock isn't held). We allow the race for that case.
647 * NOTE: pm_pvhint is only accessed (read) with the spin-lock held, using
648 * cpu_ccfence() to prevent compiler optimizations from reloading the
653 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
655 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0)) {
657 pv->pv_pmap->pm_pvhint = pv;
663 * Return address of PT slot in PD (KVM only)
665 * Cannot be used for user page tables because it might interfere with
666 * the shared page-table-page optimization (pmap_mmu_optimize).
670 vtopt(vm_offset_t va)
672 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
673 NPML4EPGSHIFT)) - 1);
675 return (PDmap + ((va >> PDRSHIFT) & mask));
679 * KVM - return address of PTE slot in PT
683 vtopte(vm_offset_t va)
685 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
686 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
688 return (PTmap + ((va >> PAGE_SHIFT) & mask));
692 allocpages(vm_paddr_t *firstaddr, long n)
697 bzero((void *)ret, n * PAGE_SIZE);
698 *firstaddr += n * PAGE_SIZE;
704 create_pagetables(vm_paddr_t *firstaddr)
706 long i; /* must be 64 bits */
712 * We are running (mostly) V=P at this point
714 * Calculate NKPT - number of kernel page tables. We have to
715 * accomodoate prealloction of the vm_page_array, dump bitmap,
716 * MSGBUF_SIZE, and other stuff. Be generous.
718 * Maxmem is in pages.
720 * ndmpdp is the number of 1GB pages we wish to map.
722 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
723 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
725 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
728 * Starting at the beginning of kvm (not KERNBASE).
730 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
731 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
732 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
733 ndmpdp) + 511) / 512;
737 * Starting at KERNBASE - map 2G worth of page table pages.
738 * KERNBASE is offset -2G from the end of kvm.
740 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
745 KPTbase = allocpages(firstaddr, nkpt_base);
746 KPTphys = allocpages(firstaddr, nkpt_phys);
747 KPML4phys = allocpages(firstaddr, 1);
748 KPDPphys = allocpages(firstaddr, NKPML4E);
749 KPDphys = allocpages(firstaddr, NKPDPE);
752 * Calculate the page directory base for KERNBASE,
753 * that is where we start populating the page table pages.
754 * Basically this is the end - 2.
756 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
758 DMPDPphys = allocpages(firstaddr, NDMPML4E);
759 if ((amd_feature & AMDID_PAGE1GB) == 0)
760 DMPDphys = allocpages(firstaddr, ndmpdp);
761 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
764 * Fill in the underlying page table pages for the area around
765 * KERNBASE. This remaps low physical memory to KERNBASE.
767 * Read-only from zero to physfree
768 * XXX not fully used, underneath 2M pages
770 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
771 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
772 ((pt_entry_t *)KPTbase)[i] |=
773 pmap_bits_default[PG_RW_IDX] |
774 pmap_bits_default[PG_V_IDX] |
775 pmap_bits_default[PG_G_IDX];
779 * Now map the initial kernel page tables. One block of page
780 * tables is placed at the beginning of kernel virtual memory,
781 * and another block is placed at KERNBASE to map the kernel binary,
782 * data, bss, and initial pre-allocations.
784 for (i = 0; i < nkpt_base; i++) {
785 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
786 ((pd_entry_t *)KPDbase)[i] |=
787 pmap_bits_default[PG_RW_IDX] |
788 pmap_bits_default[PG_V_IDX];
790 for (i = 0; i < nkpt_phys; i++) {
791 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
792 ((pd_entry_t *)KPDphys)[i] |=
793 pmap_bits_default[PG_RW_IDX] |
794 pmap_bits_default[PG_V_IDX];
798 * Map from zero to end of allocations using 2M pages as an
799 * optimization. This will bypass some of the KPTBase pages
800 * above in the KERNBASE area.
802 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
803 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
804 ((pd_entry_t *)KPDbase)[i] |=
805 pmap_bits_default[PG_RW_IDX] |
806 pmap_bits_default[PG_V_IDX] |
807 pmap_bits_default[PG_PS_IDX] |
808 pmap_bits_default[PG_G_IDX];
812 * And connect up the PD to the PDP. The kernel pmap is expected
813 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
815 for (i = 0; i < NKPDPE; i++) {
816 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
817 KPDphys + (i << PAGE_SHIFT);
818 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
819 pmap_bits_default[PG_RW_IDX] |
820 pmap_bits_default[PG_V_IDX] |
821 pmap_bits_default[PG_U_IDX];
825 * Now set up the direct map space using either 2MB or 1GB pages
826 * Preset PG_M and PG_A because demotion expects it.
828 * When filling in entries in the PD pages make sure any excess
829 * entries are set to zero as we allocated enough PD pages
831 if ((amd_feature & AMDID_PAGE1GB) == 0) {
832 for (i = 0; i < NPDEPG * ndmpdp; i++) {
833 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
834 ((pd_entry_t *)DMPDphys)[i] |=
835 pmap_bits_default[PG_RW_IDX] |
836 pmap_bits_default[PG_V_IDX] |
837 pmap_bits_default[PG_PS_IDX] |
838 pmap_bits_default[PG_G_IDX] |
839 pmap_bits_default[PG_M_IDX] |
840 pmap_bits_default[PG_A_IDX];
844 * And the direct map space's PDP
846 for (i = 0; i < ndmpdp; i++) {
847 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
849 ((pdp_entry_t *)DMPDPphys)[i] |=
850 pmap_bits_default[PG_RW_IDX] |
851 pmap_bits_default[PG_V_IDX] |
852 pmap_bits_default[PG_U_IDX];
855 for (i = 0; i < ndmpdp; i++) {
856 ((pdp_entry_t *)DMPDPphys)[i] =
857 (vm_paddr_t)i << PDPSHIFT;
858 ((pdp_entry_t *)DMPDPphys)[i] |=
859 pmap_bits_default[PG_RW_IDX] |
860 pmap_bits_default[PG_V_IDX] |
861 pmap_bits_default[PG_PS_IDX] |
862 pmap_bits_default[PG_G_IDX] |
863 pmap_bits_default[PG_M_IDX] |
864 pmap_bits_default[PG_A_IDX];
868 /* And recursively map PML4 to itself in order to get PTmap */
869 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
870 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
871 pmap_bits_default[PG_RW_IDX] |
872 pmap_bits_default[PG_V_IDX] |
873 pmap_bits_default[PG_U_IDX];
876 * Connect the Direct Map slots up to the PML4
878 for (j = 0; j < NDMPML4E; ++j) {
879 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
880 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
881 pmap_bits_default[PG_RW_IDX] |
882 pmap_bits_default[PG_V_IDX] |
883 pmap_bits_default[PG_U_IDX];
887 * Connect the KVA slot up to the PML4
889 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
890 ((pdp_entry_t *)KPML4phys)[KPML4I] |=
891 pmap_bits_default[PG_RW_IDX] |
892 pmap_bits_default[PG_V_IDX] |
893 pmap_bits_default[PG_U_IDX];
897 * Bootstrap the system enough to run with virtual memory.
899 * On the i386 this is called after mapping has already been enabled
900 * and just syncs the pmap module with what has already been done.
901 * [We can't call it easily with mapping off since the kernel is not
902 * mapped with PA == VA, hence we would have to relocate every address
903 * from the linked base (virtual) address "KERNBASE" to the actual
904 * (physical) address starting relative to 0]
907 pmap_bootstrap(vm_paddr_t *firstaddr)
913 KvaStart = VM_MIN_KERNEL_ADDRESS;
914 KvaEnd = VM_MAX_KERNEL_ADDRESS;
915 KvaSize = KvaEnd - KvaStart;
917 avail_start = *firstaddr;
920 * Create an initial set of page tables to run the kernel in.
922 create_pagetables(firstaddr);
924 virtual2_start = KvaStart;
925 virtual2_end = PTOV_OFFSET;
927 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
928 virtual_start = pmap_kmem_choose(virtual_start);
930 virtual_end = VM_MAX_KERNEL_ADDRESS;
932 /* XXX do %cr0 as well */
933 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
937 * Initialize protection array.
939 i386_protection_init();
942 * The kernel's pmap is statically allocated so we don't have to use
943 * pmap_create, which is unlikely to work correctly at this part of
944 * the boot sequence (XXX and which no longer exists).
946 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
947 kernel_pmap.pm_count = 1;
948 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
949 RB_INIT(&kernel_pmap.pm_pvroot);
950 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
951 for (i = 0; i < PM_PLACEMARKS; ++i)
952 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
955 * Reserve some special page table entries/VA space for temporary
958 #define SYSMAP(c, p, v, n) \
959 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
965 * CMAP1/CMAP2 are used for zeroing and copying pages.
967 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
972 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
975 * ptvmmap is used for reading arbitrary physical pages via
978 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
981 * msgbufp is used to map the system message buffer.
982 * XXX msgbufmap is not used.
984 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
985 atop(round_page(MSGBUF_SIZE)))
988 virtual_start = pmap_kmem_choose(virtual_start);
993 * PG_G is terribly broken on SMP because we IPI invltlb's in some
994 * cases rather then invl1pg. Actually, I don't even know why it
995 * works under UP because self-referential page table mappings
1000 * Initialize the 4MB page size flag
1004 * The 4MB page version of the initial
1005 * kernel page mapping.
1009 #if !defined(DISABLE_PSE)
1010 if (cpu_feature & CPUID_PSE) {
1013 * Note that we have enabled PSE mode
1015 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1016 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
1017 ptditmp &= ~(NBPDR - 1);
1018 ptditmp |= pmap_bits_default[PG_V_IDX] |
1019 pmap_bits_default[PG_RW_IDX] |
1020 pmap_bits_default[PG_PS_IDX] |
1021 pmap_bits_default[PG_U_IDX];
1028 /* Initialize the PAT MSR */
1030 pmap_pinit_defaults(&kernel_pmap);
1032 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1033 &pmap_fast_kernel_cpusync);
1038 * Setup the PAT MSR.
1047 * Default values mapping PATi,PCD,PWT bits at system reset.
1048 * The default values effectively ignore the PATi bit by
1049 * repeating the encodings for 0-3 in 4-7, and map the PCD
1050 * and PWT bit combinations to the expected PAT types.
1052 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
1053 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
1054 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
1055 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
1056 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
1057 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
1058 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
1059 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
1060 pat_pte_index[PAT_WRITE_BACK] = 0;
1061 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
1062 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
1063 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
1064 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1065 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1067 if (cpu_feature & CPUID_PAT) {
1069 * If we support the PAT then set-up entries for
1070 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1073 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1074 PAT_VALUE(4, PAT_WRITE_PROTECTED);
1075 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1076 PAT_VALUE(5, PAT_WRITE_COMBINING);
1077 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1078 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1081 * Then enable the PAT
1086 load_cr4(cr4 & ~CR4_PGE);
1088 /* Disable caches (CD = 1, NW = 0). */
1090 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1092 /* Flushes caches and TLBs. */
1096 /* Update PAT and index table. */
1097 wrmsr(MSR_PAT, pat_msr);
1099 /* Flush caches and TLBs again. */
1103 /* Restore caches and PGE. */
1111 * Set 4mb pdir for mp startup
1116 if (cpu_feature & CPUID_PSE) {
1117 load_cr4(rcr4() | CR4_PSE);
1118 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
1125 * Initialize the pmap module.
1126 * Called by vm_init, to initialize any structures that the pmap
1127 * system needs to map virtual memory.
1128 * pmap_init has been enhanced to support in a fairly consistant
1129 * way, discontiguous physical memory.
1138 * Allocate memory for random pmap data structures. Includes the
1142 for (i = 0; i < vm_page_array_size; i++) {
1145 m = &vm_page_array[i];
1146 TAILQ_INIT(&m->md.pv_list);
1150 * init the pv free list
1152 initial_pvs = vm_page_array_size;
1153 if (initial_pvs < MINPV)
1154 initial_pvs = MINPV;
1155 pvzone = &pvzone_store;
1156 pvinit = (void *)kmem_alloc(&kernel_map,
1157 initial_pvs * sizeof (struct pv_entry),
1159 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1160 pvinit, initial_pvs);
1163 * Now it is safe to enable pv_table recording.
1165 pmap_initialized = TRUE;
1169 * Initialize the address space (zone) for the pv_entries. Set a
1170 * high water mark so that the system can recover from excessive
1171 * numbers of pv entries.
1176 int shpgperproc = PMAP_SHPGPERPROC;
1179 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1180 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1181 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1182 pv_entry_high_water = 9 * (pv_entry_max / 10);
1185 * Subtract out pages already installed in the zone (hack)
1187 entry_max = pv_entry_max - vm_page_array_size;
1191 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1194 * Enable dynamic deletion of empty higher-level page table pages
1195 * by default only if system memory is < 8GB (use 7GB for slop).
1196 * This can save a little memory, but imposes significant
1197 * performance overhead for things like bulk builds, and for programs
1198 * which do a lot of memory mapping and memory unmapping.
1200 if (pmap_dynamic_delete < 0) {
1201 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1202 pmap_dynamic_delete = 1;
1204 pmap_dynamic_delete = 0;
1209 * Typically used to initialize a fictitious page by vm/device_pager.c
1212 pmap_page_init(struct vm_page *m)
1215 TAILQ_INIT(&m->md.pv_list);
1218 /***************************************************
1219 * Low level helper routines.....
1220 ***************************************************/
1223 * this routine defines the region(s) of memory that should
1224 * not be tested for the modified bit.
1228 pmap_track_modified(vm_pindex_t pindex)
1230 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1231 if ((va < clean_sva) || (va >= clean_eva))
1238 * Extract the physical page address associated with the map/VA pair.
1239 * The page must be wired for this to work reliably.
1242 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1249 if (va >= VM_MAX_USER_ADDRESS) {
1251 * Kernel page directories might be direct-mapped and
1252 * there is typically no PV tracking of pte's
1256 pt = pmap_pt(pmap, va);
1257 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1258 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1259 rtval = *pt & PG_PS_FRAME;
1260 rtval |= va & PDRMASK;
1262 ptep = pmap_pt_to_pte(*pt, va);
1263 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1264 rtval = *ptep & PG_FRAME;
1265 rtval |= va & PAGE_MASK;
1273 * User pages currently do not direct-map the page directory
1274 * and some pages might not used managed PVs. But all PT's
1277 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1279 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1280 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1281 rtval = *ptep & PG_FRAME;
1282 rtval |= va & PAGE_MASK;
1285 *handlep = pt_pv; /* locked until done */
1288 } else if (handlep) {
1296 pmap_extract_done(void *handle)
1299 pv_put((pv_entry_t)handle);
1303 * Similar to extract but checks protections, SMP-friendly short-cut for
1304 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1305 * fall-through to the real fault code.
1307 * The returned page, if not NULL, is held (and not busied).
1310 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1312 if (pmap && va < VM_MAX_USER_ADDRESS) {
1320 req = pmap->pmap_bits[PG_V_IDX] |
1321 pmap->pmap_bits[PG_U_IDX];
1322 if (prot & VM_PROT_WRITE)
1323 req |= pmap->pmap_bits[PG_RW_IDX];
1325 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1328 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1329 if ((*ptep & req) != req) {
1333 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1334 if (pte_pv && error == 0) {
1337 if (prot & VM_PROT_WRITE)
1340 } else if (pte_pv) {
1344 /* error can be 0 or 1 */
1355 * Extract the physical page address associated kernel virtual address.
1358 pmap_kextract(vm_offset_t va)
1360 pd_entry_t pt; /* pt entry in pd */
1363 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1364 pa = DMAP_TO_PHYS(va);
1367 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1368 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1371 * Beware of a concurrent promotion that changes the
1372 * PDE at this point! For example, vtopte() must not
1373 * be used to access the PTE because it would use the
1374 * new PDE. It is, however, safe to use the old PDE
1375 * because the page table page is preserved by the
1378 pa = *pmap_pt_to_pte(pt, va);
1379 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1385 /***************************************************
1386 * Low level mapping routines.....
1387 ***************************************************/
1390 * Routine: pmap_kenter
1392 * Add a wired page to the KVA
1393 * NOTE! note that in order for the mapping to take effect -- you
1394 * should do an invltlb after doing the pmap_kenter().
1397 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1403 kernel_pmap.pmap_bits[PG_RW_IDX] |
1404 kernel_pmap.pmap_bits[PG_V_IDX];
1408 pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1412 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1419 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1420 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1421 * (caller can conditionalize calling smp_invltlb()).
1424 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1430 npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1431 kernel_pmap.pmap_bits[PG_V_IDX];
1440 atomic_swap_long(ptep, npte);
1441 cpu_invlpg((void *)va);
1447 * Enter addresses into the kernel pmap but don't bother
1448 * doing any tlb invalidations. Caller will do a rollup
1449 * invalidation via pmap_rollup_inval().
1452 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1459 kernel_pmap.pmap_bits[PG_RW_IDX] |
1460 kernel_pmap.pmap_bits[PG_V_IDX];
1469 atomic_swap_long(ptep, npte);
1470 cpu_invlpg((void *)va);
1476 * remove a page from the kernel pagetables
1479 pmap_kremove(vm_offset_t va)
1484 pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1488 pmap_kremove_quick(vm_offset_t va)
1493 (void)pte_load_clear(ptep);
1494 cpu_invlpg((void *)va);
1498 * Remove addresses from the kernel pmap but don't bother
1499 * doing any tlb invalidations. Caller will do a rollup
1500 * invalidation via pmap_rollup_inval().
1503 pmap_kremove_noinval(vm_offset_t va)
1508 (void)pte_load_clear(ptep);
1512 * XXX these need to be recoded. They are not used in any critical path.
1515 pmap_kmodify_rw(vm_offset_t va)
1517 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1518 cpu_invlpg((void *)va);
1523 pmap_kmodify_nc(vm_offset_t va)
1525 atomic_set_long(vtopte(va), PG_N);
1526 cpu_invlpg((void *)va);
1531 * Used to map a range of physical addresses into kernel virtual
1532 * address space during the low level boot, typically to map the
1533 * dump bitmap, message buffer, and vm_page_array.
1535 * These mappings are typically made at some pointer after the end of the
1538 * We could return PHYS_TO_DMAP(start) here and not allocate any
1539 * via (*virtp), but then kmem from userland and kernel dumps won't
1540 * have access to the related pointers.
1543 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1546 vm_offset_t va_start;
1548 /*return PHYS_TO_DMAP(start);*/
1553 while (start < end) {
1554 pmap_kenter_quick(va, start);
1562 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1565 * Remove the specified set of pages from the data and instruction caches.
1567 * In contrast to pmap_invalidate_cache_range(), this function does not
1568 * rely on the CPU's self-snoop feature, because it is intended for use
1569 * when moving pages into a different cache domain.
1572 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1574 vm_offset_t daddr, eva;
1577 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1578 (cpu_feature & CPUID_CLFSH) == 0)
1582 for (i = 0; i < count; i++) {
1583 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1584 eva = daddr + PAGE_SIZE;
1585 for (; daddr < eva; daddr += cpu_clflush_line_size)
1593 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1595 KASSERT((sva & PAGE_MASK) == 0,
1596 ("pmap_invalidate_cache_range: sva not page-aligned"));
1597 KASSERT((eva & PAGE_MASK) == 0,
1598 ("pmap_invalidate_cache_range: eva not page-aligned"));
1600 if (cpu_feature & CPUID_SS) {
1601 ; /* If "Self Snoop" is supported, do nothing. */
1603 /* Globally invalidate caches */
1604 cpu_wbinvd_on_all_cpus();
1609 * Invalidate the specified range of virtual memory on all cpus associated
1613 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1615 pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1619 * Add a list of wired pages to the kva. This routine is used for temporary
1620 * kernel mappings such as those found in buffer cache buffer. Page
1621 * modifications and accesses are not tracked or recorded.
1623 * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1624 * semantics as previous mappings may have been zerod without any
1627 * The page *must* be wired.
1630 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1635 end_va = beg_va + count * PAGE_SIZE;
1637 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1642 pte = VM_PAGE_TO_PHYS(*m) |
1643 kernel_pmap.pmap_bits[PG_RW_IDX] |
1644 kernel_pmap.pmap_bits[PG_V_IDX] |
1645 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1647 atomic_swap_long(ptep, pte);
1650 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1654 * This routine jerks page mappings from the kernel -- it is meant only
1655 * for temporary mappings such as those found in buffer cache buffers.
1656 * No recording modified or access status occurs.
1658 * MPSAFE, INTERRUPT SAFE (cluster callback)
1661 pmap_qremove(vm_offset_t beg_va, int count)
1666 end_va = beg_va + count * PAGE_SIZE;
1668 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1672 (void)pte_load_clear(pte);
1673 cpu_invlpg((void *)va);
1675 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1679 * This routine removes temporary kernel mappings, only invalidating them
1680 * on the current cpu. It should only be used under carefully controlled
1684 pmap_qremove_quick(vm_offset_t beg_va, int count)
1689 end_va = beg_va + count * PAGE_SIZE;
1691 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1695 (void)pte_load_clear(pte);
1696 cpu_invlpg((void *)va);
1701 * This routine removes temporary kernel mappings *without* invalidating
1702 * the TLB. It can only be used on permanent kva reservations such as those
1703 * found in buffer cache buffers, under carefully controlled circumstances.
1705 * NOTE: Repopulating these KVAs requires unconditional invalidation.
1706 * (pmap_qenter() does unconditional invalidation).
1709 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1714 end_va = beg_va + count * PAGE_SIZE;
1716 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1720 (void)pte_load_clear(pte);
1725 * Create a new thread and optionally associate it with a (new) process.
1726 * NOTE! the new thread's cpu may not equal the current cpu.
1729 pmap_init_thread(thread_t td)
1731 /* enforce pcb placement & alignment */
1732 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1733 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1734 td->td_savefpu = &td->td_pcb->pcb_save;
1735 td->td_sp = (char *)td->td_pcb; /* no -16 */
1739 * This routine directly affects the fork perf for a process.
1742 pmap_init_proc(struct proc *p)
1747 pmap_pinit_defaults(struct pmap *pmap)
1749 bcopy(pmap_bits_default, pmap->pmap_bits,
1750 sizeof(pmap_bits_default));
1751 bcopy(protection_codes, pmap->protection_codes,
1752 sizeof(protection_codes));
1753 bcopy(pat_pte_index, pmap->pmap_cache_bits,
1754 sizeof(pat_pte_index));
1755 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1756 pmap->copyinstr = std_copyinstr;
1757 pmap->copyin = std_copyin;
1758 pmap->copyout = std_copyout;
1759 pmap->fubyte = std_fubyte;
1760 pmap->subyte = std_subyte;
1761 pmap->fuword = std_fuword;
1762 pmap->suword = std_suword;
1763 pmap->suword32 = std_suword32;
1766 * Initialize pmap0/vmspace0.
1768 * On architectures where the kernel pmap is not integrated into the user
1769 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1770 * kernel_pmap should be used to directly access the kernel_pmap.
1773 pmap_pinit0(struct pmap *pmap)
1777 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1779 CPUMASK_ASSZERO(pmap->pm_active);
1780 pmap->pm_pvhint = NULL;
1781 RB_INIT(&pmap->pm_pvroot);
1782 spin_init(&pmap->pm_spin, "pmapinit0");
1783 for (i = 0; i < PM_PLACEMARKS; ++i)
1784 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1785 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1786 pmap_pinit_defaults(pmap);
1790 * Initialize a preallocated and zeroed pmap structure,
1791 * such as one in a vmspace structure.
1794 pmap_pinit_simple(struct pmap *pmap)
1799 * Misc initialization
1802 CPUMASK_ASSZERO(pmap->pm_active);
1803 pmap->pm_pvhint = NULL;
1804 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1806 pmap_pinit_defaults(pmap);
1809 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1812 if (pmap->pm_pmlpv == NULL) {
1813 RB_INIT(&pmap->pm_pvroot);
1814 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1815 spin_init(&pmap->pm_spin, "pmapinitsimple");
1816 for (i = 0; i < PM_PLACEMARKS; ++i)
1817 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1822 pmap_pinit(struct pmap *pmap)
1827 if (pmap->pm_pmlpv) {
1828 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1833 pmap_pinit_simple(pmap);
1834 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1837 * No need to allocate page table space yet but we do need a valid
1838 * page directory table.
1840 if (pmap->pm_pml4 == NULL) {
1842 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1848 * Allocate the page directory page, which wires it even though
1849 * it isn't being entered into some higher level page table (it
1850 * being the highest level). If one is already cached we don't
1851 * have to do anything.
1853 if ((pv = pmap->pm_pmlpv) == NULL) {
1854 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1855 pmap->pm_pmlpv = pv;
1856 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1857 VM_PAGE_TO_PHYS(pv->pv_m));
1861 * Install DMAP and KMAP.
1863 for (j = 0; j < NDMPML4E; ++j) {
1864 pmap->pm_pml4[DMPML4I + j] =
1865 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1866 pmap->pmap_bits[PG_RW_IDX] |
1867 pmap->pmap_bits[PG_V_IDX] |
1868 pmap->pmap_bits[PG_U_IDX];
1870 pmap->pm_pml4[KPML4I] = KPDPphys |
1871 pmap->pmap_bits[PG_RW_IDX] |
1872 pmap->pmap_bits[PG_V_IDX] |
1873 pmap->pmap_bits[PG_U_IDX];
1876 * install self-referential address mapping entry
1878 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1879 pmap->pmap_bits[PG_V_IDX] |
1880 pmap->pmap_bits[PG_RW_IDX] |
1881 pmap->pmap_bits[PG_A_IDX] |
1882 pmap->pmap_bits[PG_M_IDX];
1884 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1885 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1887 KKASSERT(pmap->pm_pml4[255] == 0);
1888 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1889 KKASSERT(pv->pv_entry.rbe_left == NULL);
1890 KKASSERT(pv->pv_entry.rbe_right == NULL);
1894 * Clean up a pmap structure so it can be physically freed. This routine
1895 * is called by the vmspace dtor function. A great deal of pmap data is
1896 * left passively mapped to improve vmspace management so we have a bit
1897 * of cleanup work to do here.
1900 pmap_puninit(pmap_t pmap)
1905 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1906 if ((pv = pmap->pm_pmlpv) != NULL) {
1907 if (pv_hold_try(pv) == 0)
1909 KKASSERT(pv == pmap->pm_pmlpv);
1910 p = pmap_remove_pv_page(pv);
1912 pv = NULL; /* safety */
1913 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1914 vm_page_busy_wait(p, FALSE, "pgpun");
1915 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1916 vm_page_unwire(p, 0);
1917 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1920 * XXX eventually clean out PML4 static entries and
1921 * use vm_page_free_zero()
1924 pmap->pm_pmlpv = NULL;
1926 if (pmap->pm_pml4) {
1927 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1928 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1929 pmap->pm_pml4 = NULL;
1931 KKASSERT(pmap->pm_stats.resident_count == 0);
1932 KKASSERT(pmap->pm_stats.wired_count == 0);
1936 * This function is now unused (used to add the pmap to the pmap_list)
1939 pmap_pinit2(struct pmap *pmap)
1944 * This routine is called when various levels in the page table need to
1945 * be populated. This routine cannot fail.
1947 * This function returns two locked pv_entry's, one representing the
1948 * requested pv and one representing the requested pv's parent pv. If
1949 * an intermediate page table does not exist it will be created, mapped,
1950 * wired, and the parent page table will be given an additional hold
1951 * count representing the presence of the child pv_entry.
1955 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1961 vm_pindex_t pt_pindex;
1967 * If the pv already exists and we aren't being asked for the
1968 * parent page table page we can just return it. A locked+held pv
1969 * is returned. The pv will also have a second hold related to the
1970 * pmap association that we don't have to worry about.
1973 pv = pv_alloc(pmap, ptepindex, &isnew);
1974 if (isnew == 0 && pvpp == NULL)
1978 * Special case terminal PVs. These are not page table pages so
1979 * no vm_page is allocated (the caller supplied the vm_page). If
1980 * pvpp is non-NULL we are being asked to also removed the pt_pv
1983 * Note that pt_pv's are only returned for user VAs. We assert that
1984 * a pt_pv is not being requested for kernel VAs. The kernel
1985 * pre-wires all higher-level page tables so don't overload managed
1986 * higher-level page tables on top of it!
1988 if (ptepindex < pmap_pt_pindex(0)) {
1989 if (ptepindex >= NUPTE_USER) {
1990 /* kernel manages this manually for KVM */
1991 KKASSERT(pvpp == NULL);
1993 KKASSERT(pvpp != NULL);
1994 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1995 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1997 vm_page_wire_quick(pvp->pv_m);
2004 * The kernel never uses managed PT/PD/PDP pages.
2006 KKASSERT(pmap != &kernel_pmap);
2009 * Non-terminal PVs allocate a VM page to represent the page table,
2010 * so we have to resolve pvp and calculate ptepindex for the pvp
2011 * and then for the page table entry index in the pvp for
2014 if (ptepindex < pmap_pd_pindex(0)) {
2016 * pv is PT, pvp is PD
2018 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2019 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2020 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2025 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2026 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2028 } else if (ptepindex < pmap_pdp_pindex(0)) {
2030 * pv is PD, pvp is PDP
2032 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2035 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2036 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2038 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2039 KKASSERT(pvpp == NULL);
2042 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2048 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2049 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2050 } else if (ptepindex < pmap_pml4_pindex()) {
2052 * pv is PDP, pvp is the root pml4 table
2054 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2059 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2060 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2063 * pv represents the top-level PML4, there is no parent.
2072 * (isnew) is TRUE, pv is not terminal.
2074 * (1) Add a wire count to the parent page table (pvp).
2075 * (2) Allocate a VM page for the page table.
2076 * (3) Enter the VM page into the parent page table.
2078 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2081 vm_page_wire_quick(pvp->pv_m);
2084 m = vm_page_alloc(NULL, pv->pv_pindex,
2085 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2086 VM_ALLOC_INTERRUPT);
2091 vm_page_wire(m); /* wire for mapping in parent */
2092 vm_page_unmanage(m); /* m must be spinunlocked */
2093 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2094 m->valid = VM_PAGE_BITS_ALL;
2096 vm_page_spin_lock(m);
2097 pmap_page_stats_adding(m);
2098 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2100 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2101 vm_page_spin_unlock(m);
2104 * (isnew) is TRUE, pv is not terminal.
2106 * Wire the page into pvp. Bump the resident_count for the pmap.
2107 * There is no pvp for the top level, address the pm_pml4[] array
2110 * If the caller wants the parent we return it, otherwise
2111 * we just put it away.
2113 * No interlock is needed for pte 0 -> non-zero.
2115 * In the situation where *ptep is valid we might have an unmanaged
2116 * page table page shared from another page table which we need to
2117 * unshare before installing our private page table page.
2120 v = VM_PAGE_TO_PHYS(m) |
2121 (pmap->pmap_bits[PG_U_IDX] |
2122 pmap->pmap_bits[PG_RW_IDX] |
2123 pmap->pmap_bits[PG_V_IDX] |
2124 pmap->pmap_bits[PG_A_IDX] |
2125 pmap->pmap_bits[PG_M_IDX]);
2126 ptep = pv_pte_lookup(pvp, ptepindex);
2127 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2131 panic("pmap_allocpte: unexpected pte %p/%d",
2132 pvp, (int)ptepindex);
2134 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2135 if (vm_page_unwire_quick(
2136 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2137 panic("pmap_allocpte: shared pgtable "
2138 "pg bad wirecount");
2143 pte = atomic_swap_long(ptep, v);
2145 kprintf("install pgtbl mixup 0x%016jx "
2146 "old/new 0x%016jx/0x%016jx\n",
2147 (intmax_t)ptepindex, pte, v);
2154 * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2158 KKASSERT(pvp->pv_m != NULL);
2159 ptep = pv_pte_lookup(pvp, ptepindex);
2160 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2161 (pmap->pmap_bits[PG_U_IDX] |
2162 pmap->pmap_bits[PG_RW_IDX] |
2163 pmap->pmap_bits[PG_V_IDX] |
2164 pmap->pmap_bits[PG_A_IDX] |
2165 pmap->pmap_bits[PG_M_IDX]);
2167 kprintf("mismatched upper level pt %016jx/%016jx\n",
2179 * This version of pmap_allocpte() checks for possible segment optimizations
2180 * that would allow page-table sharing. It can be called for terminal
2181 * page or page table page ptepindex's.
2183 * The function is called with page table page ptepindex's for fictitious
2184 * and unmanaged terminal pages. That is, we don't want to allocate a
2185 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2188 * This function can return a pv and *pvpp associated with the passed in pmap
2189 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2190 * an unmanaged page table page will be entered into the pass in pmap.
2194 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2195 vm_map_entry_t entry, vm_offset_t va)
2201 pv_entry_t pte_pv; /* in original or shared pmap */
2202 pv_entry_t pt_pv; /* in original or shared pmap */
2203 pv_entry_t proc_pd_pv; /* in original pmap */
2204 pv_entry_t proc_pt_pv; /* in original pmap */
2205 pv_entry_t xpv; /* PT in shared pmap */
2206 pd_entry_t *pt; /* PT entry in PD of original pmap */
2207 pd_entry_t opte; /* contents of *pt */
2208 pd_entry_t npte; /* contents of *pt */
2213 * Basic tests, require a non-NULL vm_map_entry, require proper
2214 * alignment and type for the vm_map_entry, require that the
2215 * underlying object already be allocated.
2217 * We allow almost any type of object to use this optimization.
2218 * The object itself does NOT have to be sized to a multiple of the
2219 * segment size, but the memory mapping does.
2221 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2222 * won't work as expected.
2224 if (entry == NULL ||
2225 pmap_mmu_optimize == 0 || /* not enabled */
2226 (pmap->pm_flags & PMAP_HVM) || /* special pmap */
2227 ptepindex >= pmap_pd_pindex(0) || /* not terminal or pt */
2228 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2229 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2230 entry->object.vm_object == NULL || /* needs VM object */
2231 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2232 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2233 (entry->offset & SEG_MASK) || /* must be aligned */
2234 (entry->start & SEG_MASK)) {
2235 return(pmap_allocpte(pmap, ptepindex, pvpp));
2239 * Make sure the full segment can be represented.
2241 b = va & ~(vm_offset_t)SEG_MASK;
2242 if (b < entry->start || b + SEG_SIZE > entry->end)
2243 return(pmap_allocpte(pmap, ptepindex, pvpp));
2246 * If the full segment can be represented dive the VM object's
2247 * shared pmap, allocating as required.
2249 object = entry->object.vm_object;
2251 if (entry->protection & VM_PROT_WRITE)
2252 obpmapp = &object->md.pmap_rw;
2254 obpmapp = &object->md.pmap_ro;
2257 if (pmap_enter_debug > 0) {
2259 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2261 va, entry->protection, object,
2263 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2264 entry, entry->start, entry->end);
2269 * We allocate what appears to be a normal pmap but because portions
2270 * of this pmap are shared with other unrelated pmaps we have to
2271 * set pm_active to point to all cpus.
2273 * XXX Currently using pmap_spin to interlock the update, can't use
2274 * vm_object_hold/drop because the token might already be held
2275 * shared OR exclusive and we don't know.
2277 while ((obpmap = *obpmapp) == NULL) {
2278 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2279 pmap_pinit_simple(obpmap);
2280 pmap_pinit2(obpmap);
2281 spin_lock(&pmap_spin);
2282 if (*obpmapp != NULL) {
2286 spin_unlock(&pmap_spin);
2287 pmap_release(obpmap);
2288 pmap_puninit(obpmap);
2289 kfree(obpmap, M_OBJPMAP);
2290 obpmap = *obpmapp; /* safety */
2292 obpmap->pm_active = smp_active_mask;
2293 obpmap->pm_flags |= PMAP_SEGSHARED;
2295 spin_unlock(&pmap_spin);
2300 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2301 * pte/pt using the shared pmap from the object but also adjust
2302 * the process pmap's page table page as a side effect.
2306 * Resolve the terminal PTE and PT in the shared pmap. This is what
2307 * we will return. This is true if ptepindex represents a terminal
2308 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2312 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2313 if (ptepindex >= pmap_pt_pindex(0))
2319 * Resolve the PD in the process pmap so we can properly share the
2320 * page table page. Lock order is bottom-up (leaf first)!
2322 * NOTE: proc_pt_pv can be NULL.
2324 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), NULL);
2325 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2327 if (pmap_enter_debug > 0) {
2329 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2331 (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2338 * xpv is the page table page pv from the shared object
2339 * (for convenience), from above.
2341 * Calculate the pte value for the PT to load into the process PD.
2342 * If we have to change it we must properly dispose of the previous
2345 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2346 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2347 (pmap->pmap_bits[PG_U_IDX] |
2348 pmap->pmap_bits[PG_RW_IDX] |
2349 pmap->pmap_bits[PG_V_IDX] |
2350 pmap->pmap_bits[PG_A_IDX] |
2351 pmap->pmap_bits[PG_M_IDX]);
2354 * Dispose of previous page table page if it was local to the
2355 * process pmap. If the old pt is not empty we cannot dispose of it
2356 * until we clean it out. This case should not arise very often so
2357 * it is not optimized.
2360 pmap_inval_bulk_t bulk;
2362 if (proc_pt_pv->pv_m->wire_count != 1) {
2368 va & ~(vm_offset_t)SEG_MASK,
2369 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2374 * The release call will indirectly clean out *pt
2376 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2377 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2378 pmap_inval_bulk_flush(&bulk);
2381 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2385 * Handle remaining cases.
2388 atomic_swap_long(pt, npte);
2389 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2390 vm_page_wire_quick(proc_pd_pv->pv_m); /* proc pd for sh pt */
2391 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2392 } else if (*pt != npte) {
2393 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2396 opte = pte_load_clear(pt);
2397 KKASSERT(opte && opte != npte);
2401 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2404 * Clean up opte, bump the wire_count for the process
2405 * PD page representing the new entry if it was
2408 * If the entry was not previously empty and we have
2409 * a PT in the proc pmap then opte must match that
2410 * pt. The proc pt must be retired (this is done
2411 * later on in this procedure).
2413 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2416 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2417 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2418 if (vm_page_unwire_quick(m)) {
2419 panic("pmap_allocpte_seg: "
2420 "bad wire count %p",
2426 * The existing process page table was replaced and must be destroyed
2440 * Release any resources held by the given physical map.
2442 * Called when a pmap initialized by pmap_pinit is being released. Should
2443 * only be called if the map contains no valid mappings.
2445 struct pmap_release_info {
2451 static int pmap_release_callback(pv_entry_t pv, void *data);
2454 pmap_release(struct pmap *pmap)
2456 struct pmap_release_info info;
2458 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2459 ("pmap still active! %016jx",
2460 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2463 * There is no longer a pmap_list, if there were we would remove the
2464 * pmap from it here.
2468 * Pull pv's off the RB tree in order from low to high and release
2476 spin_lock(&pmap->pm_spin);
2477 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2478 pmap_release_callback, &info);
2479 spin_unlock(&pmap->pm_spin);
2483 } while (info.retry);
2487 * One resident page (the pml4 page) should remain.
2488 * No wired pages should remain.
2491 if (pmap->pm_stats.resident_count !=
2492 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1) ||
2493 pmap->pm_stats.wired_count != 0) {
2494 kprintf("fatal pmap problem - pmap %p flags %08x "
2495 "rescnt=%jd wirecnt=%jd\n",
2498 pmap->pm_stats.resident_count,
2499 pmap->pm_stats.wired_count);
2500 tsleep(pmap, 0, "DEAD", 0);
2503 KKASSERT(pmap->pm_stats.resident_count ==
2504 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2505 KKASSERT(pmap->pm_stats.wired_count == 0);
2510 * Called from low to high. We must cache the proper parent pv so we
2511 * can adjust its wired count.
2514 pmap_release_callback(pv_entry_t pv, void *data)
2516 struct pmap_release_info *info = data;
2517 pmap_t pmap = info->pmap;
2522 * Acquire a held and locked pv, check for release race
2524 pindex = pv->pv_pindex;
2525 if (info->pvp == pv) {
2526 spin_unlock(&pmap->pm_spin);
2528 } else if (pv_hold_try(pv)) {
2529 spin_unlock(&pmap->pm_spin);
2531 spin_unlock(&pmap->pm_spin);
2534 if (pv->pv_pmap != pmap || pindex != pv->pv_pindex) {
2536 spin_lock(&pmap->pm_spin);
2541 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2543 * I am PTE, parent is PT
2545 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2546 pindex += NUPTE_TOTAL;
2547 } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2549 * I am PT, parent is PD
2551 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2552 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2553 } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2555 * I am PD, parent is PDP
2557 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2559 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2560 } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2562 * I am PDP, parent is PML4 (there's only one)
2565 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2566 NUPD_TOTAL) >> NPML4EPGSHIFT;
2567 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2569 pindex = pmap_pml4_pindex();
2581 if (info->pvp && info->pvp->pv_pindex != pindex) {
2585 if (info->pvp == NULL)
2586 info->pvp = pv_get(pmap, pindex, NULL);
2593 r = pmap_release_pv(pv, info->pvp, NULL);
2594 spin_lock(&pmap->pm_spin);
2600 * Called with held (i.e. also locked) pv. This function will dispose of
2601 * the lock along with the pv.
2603 * If the caller already holds the locked parent page table for pv it
2604 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2605 * pass NULL for pvp.
2608 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2613 * The pmap is currently not spinlocked, pv is held+locked.
2614 * Remove the pv's page from its parent's page table. The
2615 * parent's page table page's wire_count will be decremented.
2617 * This will clean out the pte at any level of the page table.
2618 * If smp != 0 all cpus are affected.
2620 * Do not tear-down recursively, its faster to just let the
2621 * release run its course.
2623 pmap_remove_pv_pte(pv, pvp, bulk, 0);
2626 * Terminal pvs are unhooked from their vm_pages. Because
2627 * terminal pages aren't page table pages they aren't wired
2628 * by us, so we have to be sure not to unwire them either.
2630 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2631 pmap_remove_pv_page(pv);
2636 * We leave the top-level page table page cached, wired, and
2637 * mapped in the pmap until the dtor function (pmap_puninit())
2640 * Since we are leaving the top-level pv intact we need
2641 * to break out of what would otherwise be an infinite loop.
2643 if (pv->pv_pindex == pmap_pml4_pindex()) {
2649 * For page table pages (other than the top-level page),
2650 * remove and free the vm_page. The representitive mapping
2651 * removed above by pmap_remove_pv_pte() did not undo the
2652 * last wire_count so we have to do that as well.
2654 p = pmap_remove_pv_page(pv);
2655 vm_page_busy_wait(p, FALSE, "pmaprl");
2656 if (p->wire_count != 1) {
2657 kprintf("p->wire_count was %016lx %d\n",
2658 pv->pv_pindex, p->wire_count);
2660 KKASSERT(p->wire_count == 1);
2661 KKASSERT(p->flags & PG_UNMANAGED);
2663 vm_page_unwire(p, 0);
2664 KKASSERT(p->wire_count == 0);
2674 * This function will remove the pte associated with a pv from its parent.
2675 * Terminal pv's are supported. All cpus specified by (bulk) are properly
2678 * The wire count will be dropped on the parent page table. The wire
2679 * count on the page being removed (pv->pv_m) from the parent page table
2680 * is NOT touched. Note that terminal pages will not have any additional
2681 * wire counts while page table pages will have at least one representing
2682 * the mapping, plus others representing sub-mappings.
2684 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2685 * pages and user page table and terminal pages.
2687 * The pv must be locked. The pvp, if supplied, must be locked. All
2688 * supplied pv's will remain locked on return.
2690 * XXX must lock parent pv's if they exist to remove pte XXX
2694 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2697 vm_pindex_t ptepindex = pv->pv_pindex;
2698 pmap_t pmap = pv->pv_pmap;
2704 if (ptepindex == pmap_pml4_pindex()) {
2706 * We are the top level PML4E table, there is no parent.
2708 p = pmap->pm_pmlpv->pv_m;
2709 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2711 * Remove a PDP page from the PML4E. This can only occur
2712 * with user page tables. We do not have to lock the
2713 * pml4 PV so just ignore pvp.
2715 vm_pindex_t pml4_pindex;
2716 vm_pindex_t pdp_index;
2719 pdp_index = ptepindex - pmap_pdp_pindex(0);
2721 pml4_pindex = pmap_pml4_pindex();
2722 pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2727 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2728 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2729 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2730 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2731 } else if (ptepindex >= pmap_pd_pindex(0)) {
2733 * Remove a PD page from the PDP
2735 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2736 * of a simple pmap because it stops at
2739 vm_pindex_t pdp_pindex;
2740 vm_pindex_t pd_index;
2743 pd_index = ptepindex - pmap_pd_pindex(0);
2746 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2747 (pd_index >> NPML4EPGSHIFT);
2748 pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2753 pd = pv_pte_lookup(pvp, pd_index &
2754 ((1ul << NPDPEPGSHIFT) - 1));
2755 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2756 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2757 pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2759 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2760 p = pv->pv_m; /* degenerate test later */
2762 } else if (ptepindex >= pmap_pt_pindex(0)) {
2764 * Remove a PT page from the PD
2766 vm_pindex_t pd_pindex;
2767 vm_pindex_t pt_index;
2770 pt_index = ptepindex - pmap_pt_pindex(0);
2773 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2774 (pt_index >> NPDPEPGSHIFT);
2775 pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2780 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2782 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2783 ("*pt unexpectedly invalid %016jx "
2784 "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2785 *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2786 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2788 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2789 kprintf("*pt unexpectedly invalid %016jx "
2790 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2792 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2793 tsleep(pt, 0, "DEAD", 0);
2796 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2799 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2802 * Remove a managed PTE from the PT page. Userland pmaps
2803 * manage PT/PD/PDP page tables pages but the kernel_pmap
2806 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2807 * pv is a pte_pv so we can safely lock pt_pv.
2809 * NOTE: FICTITIOUS pages may have multiple physical mappings
2810 * so PHYS_TO_VM_PAGE() will not necessarily work for
2813 vm_pindex_t pt_pindex;
2818 pt_pindex = ptepindex >> NPTEPGSHIFT;
2819 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2821 if (ptepindex >= NUPTE_USER) {
2822 ptep = vtopte(ptepindex << PAGE_SHIFT);
2823 KKASSERT(pvp == NULL);
2824 /* pvp remains NULL */
2827 pt_pindex = NUPTE_TOTAL +
2828 (ptepindex >> NPDPEPGSHIFT);
2829 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
2833 ptep = pv_pte_lookup(pvp, ptepindex &
2834 ((1ul << NPDPEPGSHIFT) - 1));
2836 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2837 if (bulk == NULL) /* XXX */
2838 cpu_invlpg((void *)va); /* XXX */
2841 * Now update the vm_page_t
2843 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] |
2844 pmap->pmap_bits[PG_V_IDX])) !=
2845 (pmap->pmap_bits[PG_MANAGED_IDX] |
2846 pmap->pmap_bits[PG_V_IDX])) {
2847 kprintf("remove_pte badpte %016lx %016lx %d\n",
2849 pv->pv_pindex < pmap_pt_pindex(0));
2852 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2853 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2854 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2857 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2860 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2861 if (pmap_track_modified(ptepindex))
2864 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2865 vm_page_flag_set(p, PG_REFERENCED);
2867 if (pte & pmap->pmap_bits[PG_W_IDX])
2868 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2869 if (pte & pmap->pmap_bits[PG_G_IDX])
2870 cpu_invlpg((void *)va);
2872 KKASSERT(pv->pv_m == p); /* XXX remove me later */
2875 * If requested, scrap the underlying pv->pv_m and the underlying
2876 * pv. If this is a page-table-page we must also free the page.
2878 * pvp must be returned locked.
2882 * page table page (PT, PD, PDP, PML4), caller was responsible
2883 * for testing wired_count.
2887 KKASSERT(pv->pv_m->wire_count == 1);
2888 p = pmap_remove_pv_page(pv);
2892 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2893 vm_page_busy_wait(p, FALSE, "pgpun");
2896 if (ptepindex != pmap_pml4_pindex()) {
2900 ptep = (void *)PHYS_TO_DMAP(p->phys_addr);
2901 for (i = 0; i < 512; ++i) {
2903 kprintf("PGTBL %016jx@%d "
2904 "not empty %016jx\n",
2905 ptepindex, i, ptep[i]);
2910 vm_page_unwire(p, 0);
2911 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2913 } else if (destroy == 2) {
2915 * Normal page, remove from pmap and leave the underlying
2918 pmap_remove_pv_page(pv);
2920 pv = NULL; /* safety */
2924 * If we acquired pvp ourselves then we are responsible for
2925 * recursively deleting it.
2927 if (pvp && gotpvp) {
2929 * Recursively destroy higher-level page tables.
2931 * This is optional. If we do not, they will still
2932 * be destroyed when the process exits.
2934 * NOTE: Do not destroy pv_entry's with extra hold refs,
2935 * a caller may have unlocked it and intends to
2936 * continue to use it.
2938 if (pmap_dynamic_delete &&
2940 pvp->pv_m->wire_count == 1 &&
2941 (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
2942 pvp->pv_pindex != pmap_pml4_pindex()) {
2943 if (pmap_dynamic_delete == 2)
2944 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
2945 if (pmap != &kernel_pmap) {
2946 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2947 pvp = NULL; /* safety */
2949 kprintf("Attempt to remove kernel_pmap pindex "
2950 "%jd\n", pvp->pv_pindex);
2960 * Remove the vm_page association to a pv. The pv must be locked.
2964 pmap_remove_pv_page(pv_entry_t pv)
2969 vm_page_spin_lock(m);
2970 KKASSERT(m && m == pv->pv_m);
2972 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2973 pmap_page_stats_deleting(m);
2976 atomic_add_int(&m->object->agg_pv_list_count, -1);
2978 if (TAILQ_EMPTY(&m->md.pv_list))
2979 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2980 vm_page_spin_unlock(m);
2986 * Grow the number of kernel page table entries, if needed.
2988 * This routine is always called to validate any address space
2989 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2990 * space below KERNBASE.
2992 * kernel_map must be locked exclusively by the caller.
2995 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2998 vm_offset_t ptppaddr;
3000 pd_entry_t *pt, newpt;
3002 int update_kernel_vm_end;
3005 * bootstrap kernel_vm_end on first real VM use
3007 if (kernel_vm_end == 0) {
3008 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3010 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3011 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3012 ~(PAGE_SIZE * NPTEPG - 1);
3014 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3015 kernel_vm_end = kernel_map.max_offset;
3022 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
3023 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
3024 * do not want to force-fill 128G worth of page tables.
3026 if (kstart < KERNBASE) {
3027 if (kstart > kernel_vm_end)
3028 kstart = kernel_vm_end;
3029 KKASSERT(kend <= KERNBASE);
3030 update_kernel_vm_end = 1;
3032 update_kernel_vm_end = 0;
3035 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
3036 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
3038 if (kend - 1 >= kernel_map.max_offset)
3039 kend = kernel_map.max_offset;
3041 while (kstart < kend) {
3042 pt = pmap_pt(&kernel_pmap, kstart);
3044 /* We need a new PD entry */
3045 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3048 VM_ALLOC_INTERRUPT);
3050 panic("pmap_growkernel: no memory to grow "
3053 paddr = VM_PAGE_TO_PHYS(nkpg);
3054 pmap_zero_page(paddr);
3055 newpd = (pdp_entry_t)
3057 kernel_pmap.pmap_bits[PG_V_IDX] |
3058 kernel_pmap.pmap_bits[PG_RW_IDX] |
3059 kernel_pmap.pmap_bits[PG_A_IDX] |
3060 kernel_pmap.pmap_bits[PG_M_IDX]);
3061 *pmap_pd(&kernel_pmap, kstart) = newpd;
3062 continue; /* try again */
3064 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3065 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3066 ~(PAGE_SIZE * NPTEPG - 1);
3067 if (kstart - 1 >= kernel_map.max_offset) {
3068 kstart = kernel_map.max_offset;
3077 * This index is bogus, but out of the way
3079 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3082 VM_ALLOC_INTERRUPT);
3084 panic("pmap_growkernel: no memory to grow kernel");
3087 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3088 pmap_zero_page(ptppaddr);
3089 newpt = (pd_entry_t)(ptppaddr |
3090 kernel_pmap.pmap_bits[PG_V_IDX] |
3091 kernel_pmap.pmap_bits[PG_RW_IDX] |
3092 kernel_pmap.pmap_bits[PG_A_IDX] |
3093 kernel_pmap.pmap_bits[PG_M_IDX]);
3094 atomic_swap_long(pmap_pt(&kernel_pmap, kstart), newpt);
3096 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3097 ~(PAGE_SIZE * NPTEPG - 1);
3099 if (kstart - 1 >= kernel_map.max_offset) {
3100 kstart = kernel_map.max_offset;
3106 * Only update kernel_vm_end for areas below KERNBASE.
3108 if (update_kernel_vm_end && kernel_vm_end < kstart)
3109 kernel_vm_end = kstart;
3113 * Add a reference to the specified pmap.
3116 pmap_reference(pmap_t pmap)
3119 atomic_add_int(&pmap->pm_count, 1);
3122 /***************************************************
3123 * page management routines.
3124 ***************************************************/
3127 * Hold a pv without locking it
3130 pv_hold(pv_entry_t pv)
3132 atomic_add_int(&pv->pv_hold, 1);
3136 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
3137 * was successfully locked, FALSE if it wasn't. The caller must dispose of
3140 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3141 * pv list via its page) must be held by the caller.
3144 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3149 * Critical path shortcut expects pv to already have one ref
3150 * (for the pv->pv_pmap).
3152 if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3155 pv->pv_line = lineno;
3161 count = pv->pv_hold;
3163 if ((count & PV_HOLD_LOCKED) == 0) {
3164 if (atomic_cmpset_int(&pv->pv_hold, count,
3165 (count + 1) | PV_HOLD_LOCKED)) {
3168 pv->pv_line = lineno;
3173 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3181 * Drop a previously held pv_entry which could not be locked, allowing its
3184 * Must not be called with a spinlock held as we might zfree() the pv if it
3185 * is no longer associated with a pmap and this was the last hold count.
3188 pv_drop(pv_entry_t pv)
3193 count = pv->pv_hold;
3195 KKASSERT((count & PV_HOLD_MASK) > 0);
3196 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3197 (PV_HOLD_LOCKED | 1));
3198 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3199 if ((count & PV_HOLD_MASK) == 1) {
3201 if (pmap_enter_debug > 0) {
3203 kprintf("pv_drop: free pv %p\n", pv);
3206 KKASSERT(count == 1);
3207 KKASSERT(pv->pv_pmap == NULL);
3217 * Find or allocate the requested PV entry, returning a locked, held pv.
3219 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3220 * for the caller and one representing the pmap and vm_page association.
3222 * If (*isnew) is zero, the returned pv will have only one hold count.
3224 * Since both associations can only be adjusted while the pv is locked,
3225 * together they represent just one additional hold.
3229 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3232 pv_entry_t pnew = NULL;
3234 spin_lock(&pmap->pm_spin);
3239 pv = pmap->pm_pvhint;
3242 pv->pv_pmap != pmap ||
3243 pv->pv_pindex != pindex) {
3244 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3251 * We need to stage a new pv entry
3254 spin_unlock(&pmap->pm_spin);
3255 pnew = zalloc(pvzone);
3256 spin_lock(&pmap->pm_spin);
3261 * We need to block if someone is holding a
3264 pmark = pmap_placemarker_hash(pmap, pindex);
3266 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3267 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3268 ssleep(pmark, &pmap->pm_spin, 0, "pvplc", 0);
3273 * Setup the new entry
3275 pnew->pv_pmap = pmap;
3276 pnew->pv_pindex = pindex;
3277 pnew->pv_hold = PV_HOLD_LOCKED | 2;
3279 pnew->pv_func = func;
3280 pnew->pv_line = lineno;
3282 pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3283 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3284 spin_unlock(&pmap->pm_spin);
3287 KKASSERT(pv == NULL);
3293 * We have an entry, clean up any staged pv we had allocated,
3294 * then block until we can lock the entry.
3297 spin_unlock(&pmap->pm_spin);
3298 zfree(pvzone, pnew);
3300 spin_lock(&pmap->pm_spin);
3303 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3304 spin_unlock(&pmap->pm_spin);
3306 spin_unlock(&pmap->pm_spin);
3307 _pv_lock(pv PMAP_DEBUG_COPY);
3311 * Make sure the pv is still in good shape for return,
3314 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3319 spin_lock(&pmap->pm_spin);
3324 * Find the requested PV entry, returning a locked+held pv or NULL
3328 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3332 spin_lock(&pmap->pm_spin);
3337 pv = pmap->pm_pvhint;
3340 pv->pv_pmap != pmap ||
3341 pv->pv_pindex != pindex) {
3342 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3347 * Block if there is a placemarker
3351 pmark = pmap_placemarker_hash(pmap, pindex);
3353 if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3354 ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3355 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3356 ssleep(pmark, &pmap->pm_spin, 0, "pvpld", 0);
3363 spin_unlock(&pmap->pm_spin);
3366 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3367 spin_unlock(&pmap->pm_spin);
3369 spin_unlock(&pmap->pm_spin);
3370 _pv_lock(pv PMAP_DEBUG_COPY);
3372 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3373 pv_cache(pv, pindex);
3377 spin_lock(&pmap->pm_spin);
3382 * Lookup, hold, and attempt to lock (pmap,pindex).
3384 * If the entry does not exist NULL is returned and *errorp is set to 0
3386 * If the entry exists and could be successfully locked it is returned and
3387 * errorp is set to 0.
3389 * If the entry exists but could NOT be successfully locked it is returned
3390 * held and *errorp is set to 1.
3392 * If the entry is placemarked by someone else NULL is returned and *errorp
3397 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3401 spin_lock_shared(&pmap->pm_spin);
3403 pv = pmap->pm_pvhint;
3406 pv->pv_pmap != pmap ||
3407 pv->pv_pindex != pindex) {
3408 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3414 pmark = pmap_placemarker_hash(pmap, pindex);
3416 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3418 } else if (pmarkp &&
3419 atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3423 * Can't set a placemark with a NULL pmarkp, return
3430 spin_unlock_shared(&pmap->pm_spin);
3434 if (pv_hold_try(pv)) {
3435 pv_cache(pv, pindex);
3436 spin_unlock_shared(&pmap->pm_spin);
3438 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3439 return(pv); /* lock succeeded */
3441 spin_unlock_shared(&pmap->pm_spin);
3443 return (pv); /* lock failed */
3447 * Lock a held pv, keeping the hold count
3451 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3456 count = pv->pv_hold;
3458 if ((count & PV_HOLD_LOCKED) == 0) {
3459 if (atomic_cmpset_int(&pv->pv_hold, count,
3460 count | PV_HOLD_LOCKED)) {
3463 pv->pv_line = lineno;
3469 tsleep_interlock(pv, 0);
3470 if (atomic_cmpset_int(&pv->pv_hold, count,
3471 count | PV_HOLD_WAITING)) {
3473 kprintf("pv waiting on %s:%d\n",
3474 pv->pv_func, pv->pv_line);
3476 tsleep(pv, PINTERLOCKED, "pvwait", hz);
3483 * Unlock a held and locked pv, keeping the hold count.
3487 pv_unlock(pv_entry_t pv)
3492 count = pv->pv_hold;
3494 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3495 (PV_HOLD_LOCKED | 1));
3496 if (atomic_cmpset_int(&pv->pv_hold, count,
3498 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3499 if (count & PV_HOLD_WAITING)
3507 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3508 * and the hold count drops to zero we will free it.
3510 * Caller should not hold any spin locks. We are protected from hold races
3511 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3512 * lock held. A pv cannot be located otherwise.
3516 pv_put(pv_entry_t pv)
3519 if (pmap_enter_debug > 0) {
3521 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3526 * Fast - shortcut most common condition
3528 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3539 * Remove the pmap association from a pv, require that pv_m already be removed,
3540 * then unlock and drop the pv. Any pte operations must have already been
3541 * completed. This call may result in a last-drop which will physically free
3544 * Removing the pmap association entails an additional drop.
3546 * pv must be exclusively locked on call and will be disposed of on return.
3550 pv_free(pv_entry_t pv, pv_entry_t pvp)
3554 KKASSERT(pv->pv_m == NULL);
3555 KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3556 if ((pmap = pv->pv_pmap) != NULL) {
3557 spin_lock(&pmap->pm_spin);
3558 KKASSERT(pv->pv_pmap == pmap);
3559 if (pmap->pm_pvhint == pv)
3560 pmap->pm_pvhint = NULL;
3561 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3562 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3565 spin_unlock(&pmap->pm_spin);
3568 * Try to shortcut three atomic ops, otherwise fall through
3569 * and do it normally. Drop two refs and the lock all in
3572 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3574 if (pmap_enter_debug > 0) {
3576 kprintf("pv_free: free pv %p\n", pv);
3581 vm_page_unwire_quick(pvp->pv_m);
3584 pv_drop(pv); /* ref for pv_pmap */
3586 vm_page_unwire_quick(pvp->pv_m);
3592 * This routine is very drastic, but can save the system
3600 static int warningdone=0;
3602 if (pmap_pagedaemon_waken == 0)
3604 pmap_pagedaemon_waken = 0;
3605 if (warningdone < 5) {
3606 kprintf("pmap_collect: collecting pv entries -- "
3607 "suggest increasing PMAP_SHPGPERPROC\n");
3611 for (i = 0; i < vm_page_array_size; i++) {
3612 m = &vm_page_array[i];
3613 if (m->wire_count || m->hold_count)
3615 if (vm_page_busy_try(m, TRUE) == 0) {
3616 if (m->wire_count == 0 && m->hold_count == 0) {
3625 * Scan the pmap for active page table entries and issue a callback.
3626 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3627 * its parent page table.
3629 * pte_pv will be NULL if the page or page table is unmanaged.
3630 * pt_pv will point to the page table page containing the pte for the page.
3632 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3633 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3634 * process pmap's PD and page to the callback function. This can be
3635 * confusing because the pt_pv is really a pd_pv, and the target page
3636 * table page is simply aliased by the pmap and not owned by it.
3638 * It is assumed that the start and end are properly rounded to the page size.
3640 * It is assumed that PD pages and above are managed and thus in the RB tree,
3641 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3643 struct pmap_scan_info {
3647 vm_pindex_t sva_pd_pindex;
3648 vm_pindex_t eva_pd_pindex;
3649 void (*func)(pmap_t, struct pmap_scan_info *,
3650 pv_entry_t, vm_pindex_t *, pv_entry_t,
3652 pt_entry_t *, void *);
3654 pmap_inval_bulk_t bulk_core;
3655 pmap_inval_bulk_t *bulk;
3660 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3661 static int pmap_scan_callback(pv_entry_t pv, void *data);
3664 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3666 struct pmap *pmap = info->pmap;
3667 pv_entry_t pd_pv; /* A page directory PV */
3668 pv_entry_t pt_pv; /* A page table PV */
3669 pv_entry_t pte_pv; /* A page table entry PV */
3670 vm_pindex_t *pte_placemark;
3671 vm_pindex_t *pt_placemark;
3674 struct pv_entry dummy_pv;
3680 info->bulk = &info->bulk_core;
3681 pmap_inval_bulk_init(&info->bulk_core, pmap);
3687 * Hold the token for stability; if the pmap is empty we have nothing
3691 if (pmap->pm_stats.resident_count == 0) {
3699 * Special handling for scanning one page, which is a very common
3700 * operation (it is?).
3702 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3704 if (info->sva + PAGE_SIZE == info->eva) {
3705 if (info->sva >= VM_MAX_USER_ADDRESS) {
3707 * Kernel mappings do not track wire counts on
3708 * page table pages and only maintain pd_pv and
3709 * pte_pv levels so pmap_scan() works.
3712 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3714 ptep = vtopte(info->sva);
3717 * User pages which are unmanaged will not have a
3718 * pte_pv. User page table pages which are unmanaged
3719 * (shared from elsewhere) will also not have a pt_pv.
3720 * The func() callback will pass both pte_pv and pt_pv
3721 * as NULL in that case.
3723 * We hold pte_placemark across the operation for
3726 * WARNING! We must hold pt_placemark across the
3727 * *ptep test to prevent misintepreting
3728 * a non-zero *ptep as a shared page
3729 * table page. Hold it across the function
3730 * callback as well for SMP safety.
3732 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3734 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
3736 if (pt_pv == NULL) {
3737 KKASSERT(pte_pv == NULL);
3738 pd_pv = pv_get(pmap,
3739 pmap_pd_pindex(info->sva),
3742 ptep = pv_pte_lookup(pd_pv,
3743 pmap_pt_index(info->sva));
3745 info->func(pmap, info,
3751 pv_placemarker_wakeup(pmap,
3756 pv_placemarker_wakeup(pmap,
3759 pv_placemarker_wakeup(pmap, pte_placemark);
3762 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3766 * NOTE: *ptep can't be ripped out from under us if we hold
3767 * pte_pv (or pte_placemark) locked, but bits can
3773 KKASSERT(pte_pv == NULL);
3774 pv_placemarker_wakeup(pmap, pte_placemark);
3775 } else if (pte_pv) {
3776 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3777 pmap->pmap_bits[PG_V_IDX])) ==
3778 (pmap->pmap_bits[PG_MANAGED_IDX] |
3779 pmap->pmap_bits[PG_V_IDX]),
3780 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
3781 *ptep, oldpte, info->sva, pte_pv));
3782 info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
3783 info->sva, ptep, info->arg);
3785 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3786 pmap->pmap_bits[PG_V_IDX])) ==
3787 pmap->pmap_bits[PG_V_IDX],
3788 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
3789 *ptep, oldpte, info->sva));
3790 info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
3791 info->sva, ptep, info->arg);
3796 pmap_inval_bulk_flush(info->bulk);
3801 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3804 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3805 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3807 if (info->sva >= VM_MAX_USER_ADDRESS) {
3809 * The kernel does not currently maintain any pv_entry's for
3810 * higher-level page tables.
3812 bzero(&dummy_pv, sizeof(dummy_pv));
3813 dummy_pv.pv_pindex = info->sva_pd_pindex;
3814 spin_lock(&pmap->pm_spin);
3815 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3816 pmap_scan_callback(&dummy_pv, info);
3817 ++dummy_pv.pv_pindex;
3819 spin_unlock(&pmap->pm_spin);
3822 * User page tables maintain local PML4, PDP, and PD
3823 * pv_entry's at the very least. PT pv's might be
3824 * unmanaged and thus not exist. PTE pv's might be
3825 * unmanaged and thus not exist.
3827 spin_lock(&pmap->pm_spin);
3828 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
3829 pmap_scan_callback, info);
3830 spin_unlock(&pmap->pm_spin);
3832 pmap_inval_bulk_flush(info->bulk);
3836 * WARNING! pmap->pm_spin held
3839 pmap_scan_cmp(pv_entry_t pv, void *data)
3841 struct pmap_scan_info *info = data;
3842 if (pv->pv_pindex < info->sva_pd_pindex)
3844 if (pv->pv_pindex >= info->eva_pd_pindex)
3850 * pmap_scan() by PDs
3852 * WARNING! pmap->pm_spin held
3855 pmap_scan_callback(pv_entry_t pv, void *data)
3857 struct pmap_scan_info *info = data;
3858 struct pmap *pmap = info->pmap;
3859 pv_entry_t pd_pv; /* A page directory PV */
3860 pv_entry_t pt_pv; /* A page table PV */
3861 vm_pindex_t *pt_placemark;
3866 vm_offset_t va_next;
3867 vm_pindex_t pd_pindex;
3877 * Pull the PD pindex from the pv before releasing the spinlock.
3879 * WARNING: pv is faked for kernel pmap scans.
3881 pd_pindex = pv->pv_pindex;
3882 spin_unlock(&pmap->pm_spin);
3883 pv = NULL; /* invalid after spinlock unlocked */
3886 * Calculate the page range within the PD. SIMPLE pmaps are
3887 * direct-mapped for the entire 2^64 address space. Normal pmaps
3888 * reflect the user and kernel address space which requires
3889 * cannonicalization w/regards to converting pd_pindex's back
3892 sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
3893 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3894 (sva & PML4_SIGNMASK)) {
3895 sva |= PML4_SIGNMASK;
3897 eva = sva + NBPDP; /* can overflow */
3898 if (sva < info->sva)
3900 if (eva < info->sva || eva > info->eva)
3904 * NOTE: kernel mappings do not track page table pages, only
3907 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3908 * However, for the scan to be efficient we try to
3909 * cache items top-down.
3914 for (; sva < eva; sva = va_next) {
3917 if (sva >= VM_MAX_USER_ADDRESS) {
3926 * PD cache, scan shortcut if it doesn't exist.
3928 if (pd_pv == NULL) {
3929 pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
3930 } else if (pd_pv->pv_pmap != pmap ||
3931 pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3933 pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
3935 if (pd_pv == NULL) {
3936 va_next = (sva + NBPDP) & ~PDPMASK;
3945 if (pt_pv && (pt_pv->pv_pmap != pmap ||
3946 pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
3950 if (pt_pv == NULL) {
3951 pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
3952 &pt_placemark, &error);
3954 pv_put(pd_pv); /* lock order */
3961 pv_placemarker_wait(pmap, pt_placemark);
3966 /* may have to re-check later if pt_pv is NULL here */
3970 * If pt_pv is NULL we either have an shared page table
3971 * page and must issue a callback specific to that case,
3972 * or there is no page table page.
3974 * Either way we can skip the page table page.
3976 * WARNING! pt_pv can also be NULL due to a pv creation
3977 * race where we find it to be NULL and then
3978 * later see a pte_pv. But its possible the pt_pv
3979 * got created inbetween the two operations, so
3982 if (pt_pv == NULL) {
3984 * Possible unmanaged (shared from another pmap)
3987 * WARNING! We must hold pt_placemark across the
3988 * *ptep test to prevent misintepreting
3989 * a non-zero *ptep as a shared page
3990 * table page. Hold it across the function
3991 * callback as well for SMP safety.
3993 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3994 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3995 info->func(pmap, info, NULL, pt_placemark,
3997 sva, ptep, info->arg);
3999 pv_placemarker_wakeup(pmap, pt_placemark);
4003 * Done, move to next page table page.
4005 va_next = (sva + NBPDR) & ~PDRMASK;
4012 * From this point in the loop testing pt_pv for non-NULL
4013 * means we are in UVM, else if it is NULL we are in KVM.
4015 * Limit our scan to either the end of the va represented
4016 * by the current page table page, or to the end of the
4017 * range being removed.
4020 va_next = (sva + NBPDR) & ~PDRMASK;
4027 * Scan the page table for pages. Some pages may not be
4028 * managed (might not have a pv_entry).
4030 * There is no page table management for kernel pages so
4031 * pt_pv will be NULL in that case, but otherwise pt_pv
4032 * is non-NULL, locked, and referenced.
4036 * At this point a non-NULL pt_pv means a UVA, and a NULL
4037 * pt_pv means a KVA.
4040 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4044 while (sva < va_next) {
4046 vm_pindex_t *pte_placemark;
4049 * Yield every 64 pages, stop if requested.
4051 if ((++info->count & 63) == 0)
4057 * We can shortcut our scan if *ptep == 0. This is
4058 * an unlocked check.
4067 * Acquire the related pte_pv, if any. If *ptep == 0
4068 * the related pte_pv should not exist, but if *ptep
4069 * is not zero the pte_pv may or may not exist (e.g.
4070 * will not exist for an unmanaged page).
4072 * However a multitude of races are possible here
4073 * so if we cannot lock definite state we clean out
4074 * our cache and break the inner while() loop to
4075 * force a loop up to the top of the for().
4077 * XXX unlock/relock pd_pv, pt_pv, and re-test their
4078 * validity instead of looping up?
4080 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4081 &pte_placemark, &error);
4083 pv_put(pd_pv); /* lock order */
4086 pv_put(pt_pv); /* lock order */
4089 if (pte_pv) { /* block */
4094 pv_placemarker_wait(pmap,
4097 va_next = sva; /* retry */
4102 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
4107 kprintf("Unexpected non-NULL pte_pv "
4109 "*ptep = %016lx/%016lx\n",
4110 pte_pv, pt_pv, *ptep, oldpte);
4111 panic("Unexpected non-NULL pte_pv");
4113 pv_placemarker_wakeup(pmap, pte_placemark);
4120 * We can't hold pd_pv across the callback (because
4121 * we don't pass it to the callback and the callback
4125 vm_page_wire_quick(pd_pv->pv_m);
4130 * Ready for the callback. The locked pte_pv (if any)
4131 * is consumed by the callback. pte_pv will exist if
4132 * the page is managed, and will not exist if it
4136 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
4137 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
4138 ("badC *ptep %016lx/%016lx sva %016lx "
4140 *ptep, oldpte, sva, pte_pv));
4142 * We must unlock pd_pv across the callback
4143 * to avoid deadlocks on any recursive
4144 * disposal. Re-check that it still exists
4147 * Call target disposes of pte_pv and may
4148 * destroy but will not dispose of pt_pv.
4150 info->func(pmap, info, pte_pv, NULL,
4152 sva, ptep, info->arg);
4155 * We must unlock pd_pv across the callback
4156 * to avoid deadlocks on any recursive
4157 * disposal. Re-check that it still exists
4160 * Call target disposes of pte_pv and may
4161 * destroy but will not dispose of pt_pv.
4163 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
4164 pmap->pmap_bits[PG_V_IDX],
4165 ("badD *ptep %016lx/%016lx sva %016lx "
4167 *ptep, oldpte, sva));
4169 info->func(pmap, info, NULL, pte_placemark,
4171 sva, ptep, info->arg);
4175 vm_page_unwire_quick(pd_pv->pv_m);
4176 if (pd_pv->pv_pmap == NULL) {
4177 va_next = sva; /* retry */
4194 if ((++info->count & 7) == 0)
4198 * Relock before returning.
4200 spin_lock(&pmap->pm_spin);
4205 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4207 struct pmap_scan_info info;
4212 info.func = pmap_remove_callback;
4214 pmap_scan(&info, 1);
4218 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4220 struct pmap_scan_info info;
4225 info.func = pmap_remove_callback;
4227 pmap_scan(&info, 0);
4231 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4232 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4233 pv_entry_t pt_pv, int sharept,
4234 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4240 * This will also drop pt_pv's wire_count. Note that
4241 * terminal pages are not wired based on mmu presence.
4243 * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4245 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4246 pte_pv = NULL; /* safety */
4249 * Recursively destroy higher-level page tables.
4251 * This is optional. If we do not, they will still
4252 * be destroyed when the process exits.
4254 * NOTE: Do not destroy pv_entry's with extra hold refs,
4255 * a caller may have unlocked it and intends to
4256 * continue to use it.
4258 if (pmap_dynamic_delete &&
4261 pt_pv->pv_m->wire_count == 1 &&
4262 (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4263 pt_pv->pv_pindex != pmap_pml4_pindex()) {
4264 if (pmap_dynamic_delete == 2)
4265 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4266 pv_hold(pt_pv); /* extra hold */
4267 pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4268 pv_lock(pt_pv); /* prior extra hold + relock */
4270 } else if (sharept == 0) {
4272 * Unmanaged page table (pt, pd, or pdp. Not pte).
4274 * pt_pv's wire_count is still bumped by unmanaged pages
4275 * so we must decrement it manually.
4277 * We have to unwire the target page table page.
4279 * It is unclear how we can invalidate a segment so we
4280 * invalidate -1 which invlidates the tlb.
4282 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4283 if (pte & pmap->pmap_bits[PG_W_IDX])
4284 atomic_add_long(&pmap->pm_stats.wired_count, -1);
4285 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4286 if (vm_page_unwire_quick(pt_pv->pv_m))
4287 panic("pmap_remove: insufficient wirecount");
4288 pv_placemarker_wakeup(pmap, pte_placemark);
4291 * Unmanaged page table (pt, pd, or pdp. Not pte) for
4292 * a shared page table.
4294 * pt_pv is actually the pd_pv for our pmap (not the shared
4297 * We have to unwire the target page table page and we
4298 * have to unwire our page directory page.
4300 * It is unclear how we can invalidate a segment so we
4301 * invalidate -1 which invlidates the tlb.
4303 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4304 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4305 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4306 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4307 panic("pmap_remove: shared pgtable1 bad wirecount");
4308 if (vm_page_unwire_quick(pt_pv->pv_m))
4309 panic("pmap_remove: shared pgtable2 bad wirecount");
4310 pv_placemarker_wakeup(pmap, pte_placemark);
4315 * Removes this physical page from all physical maps in which it resides.
4316 * Reflects back modify bits to the pager.
4318 * This routine may not be called from an interrupt.
4322 pmap_remove_all(vm_page_t m)
4325 pmap_inval_bulk_t bulk;
4327 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4330 vm_page_spin_lock(m);
4331 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4332 KKASSERT(pv->pv_m == m);
4333 if (pv_hold_try(pv)) {
4334 vm_page_spin_unlock(m);
4336 vm_page_spin_unlock(m);
4339 if (pv->pv_pmap == NULL || pv->pv_m != m) {
4341 vm_page_spin_lock(m);
4346 * Holding no spinlocks, pv is locked.
4348 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4349 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4350 pv = NULL; /* safety */
4351 pmap_inval_bulk_flush(&bulk);
4352 vm_page_spin_lock(m);
4354 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4355 vm_page_spin_unlock(m);
4359 * Removes the page from a particular pmap
4362 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4365 pmap_inval_bulk_t bulk;
4367 if (!pmap_initialized)
4371 vm_page_spin_lock(m);
4372 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4373 if (pv->pv_pmap != pmap)
4375 KKASSERT(pv->pv_m == m);
4376 if (pv_hold_try(pv)) {
4377 vm_page_spin_unlock(m);
4379 vm_page_spin_unlock(m);
4382 if (pv->pv_pmap != pmap || pv->pv_m != m) {
4388 * Holding no spinlocks, pv is locked.
4390 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4391 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4392 pv = NULL; /* safety */
4393 pmap_inval_bulk_flush(&bulk);
4396 vm_page_spin_unlock(m);
4400 * Set the physical protection on the specified range of this map
4401 * as requested. This function is typically only used for debug watchpoints
4404 * This function may not be called from an interrupt if the map is
4405 * not the kernel_pmap.
4407 * NOTE! For shared page table pages we just unmap the page.
4410 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4412 struct pmap_scan_info info;
4413 /* JG review for NX */
4417 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4418 pmap_remove(pmap, sva, eva);
4421 if (prot & VM_PROT_WRITE)
4426 info.func = pmap_protect_callback;
4428 pmap_scan(&info, 1);
4433 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4434 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4435 pv_entry_t pt_pv, int sharept,
4436 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4448 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4449 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4450 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4451 KKASSERT(m == pte_pv->pv_m);
4452 vm_page_flag_set(m, PG_REFERENCED);
4454 cbits &= ~pmap->pmap_bits[PG_A_IDX];
4456 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4457 if (pmap_track_modified(pte_pv->pv_pindex)) {
4458 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4460 m = PHYS_TO_VM_PAGE(pbits &
4465 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4468 } else if (sharept) {
4470 * Unmanaged page table, pt_pv is actually the pd_pv
4471 * for our pmap (not the object's shared pmap).
4473 * When asked to protect something in a shared page table
4474 * page we just unmap the page table page. We have to
4475 * invalidate the tlb in this situation.
4477 * XXX Warning, shared page tables will not be used for
4478 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4479 * so PHYS_TO_VM_PAGE() should be safe here.
4481 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4482 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4483 panic("pmap_protect: pgtable1 pg bad wirecount");
4484 if (vm_page_unwire_quick(pt_pv->pv_m))
4485 panic("pmap_protect: pgtable2 pg bad wirecount");
4488 /* else unmanaged page, adjust bits, no wire changes */
4491 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4493 if (pmap_enter_debug > 0) {
4495 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4496 "pt_pv=%p cbits=%08lx\n",
4502 if (pbits != cbits) {
4503 if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4504 ptep, pbits, cbits)) {
4512 pv_placemarker_wakeup(pmap, pte_placemark);
4516 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4517 * mapping at that address. Set protection and wiring as requested.
4519 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4520 * possible. If it is we enter the page into the appropriate shared pmap
4521 * hanging off the related VM object instead of the passed pmap, then we
4522 * share the page table page from the VM object's pmap into the current pmap.
4524 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4528 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4529 boolean_t wired, vm_map_entry_t entry)
4531 pv_entry_t pt_pv; /* page table */
4532 pv_entry_t pte_pv; /* page table entry */
4533 vm_pindex_t *pte_placemark;
4536 pt_entry_t origpte, newpte;
4541 va = trunc_page(va);
4542 #ifdef PMAP_DIAGNOSTIC
4544 panic("pmap_enter: toobig");
4545 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4546 panic("pmap_enter: invalid to pmap_enter page table "
4547 "pages (va: 0x%lx)", va);
4549 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4550 kprintf("Warning: pmap_enter called on UVA with "
4553 db_print_backtrace();
4556 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4557 kprintf("Warning: pmap_enter called on KVA without"
4560 db_print_backtrace();
4565 * Get locked PV entries for our new page table entry (pte_pv)
4566 * and for its parent page table (pt_pv). We need the parent
4567 * so we can resolve the location of the ptep.
4569 * Only hardware MMU actions can modify the ptep out from
4572 * if (m) is fictitious or unmanaged we do not create a managing
4573 * pte_pv for it. Any pre-existing page's management state must
4574 * match (avoiding code complexity).
4576 * If the pmap is still being initialized we assume existing
4579 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4581 if (pmap_initialized == FALSE) {
4584 pte_placemark = NULL;
4587 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4588 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4589 KKASSERT(pte_pv == NULL);
4590 if (va >= VM_MAX_USER_ADDRESS) {
4594 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4596 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4600 KASSERT(origpte == 0 ||
4601 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4602 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4604 if (va >= VM_MAX_USER_ADDRESS) {
4606 * Kernel map, pv_entry-tracked.
4609 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4615 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4617 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4619 pte_placemark = NULL;
4622 KASSERT(origpte == 0 ||
4623 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4624 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4627 pa = VM_PAGE_TO_PHYS(m);
4628 opa = origpte & PG_FRAME;
4630 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4631 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4633 newpte |= pmap->pmap_bits[PG_W_IDX];
4634 if (va < VM_MAX_USER_ADDRESS)
4635 newpte |= pmap->pmap_bits[PG_U_IDX];
4637 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4638 // if (pmap == &kernel_pmap)
4639 // newpte |= pgeflag;
4640 newpte |= pmap->pmap_cache_bits[m->pat_mode];
4641 if (m->flags & PG_FICTITIOUS)
4642 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4645 * It is possible for multiple faults to occur in threaded
4646 * environments, the existing pte might be correct.
4648 if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4649 pmap->pmap_bits[PG_A_IDX])) == 0)
4653 * Ok, either the address changed or the protection or wiring
4656 * Clear the current entry, interlocking the removal. For managed
4657 * pte's this will also flush the modified state to the vm_page.
4658 * Atomic ops are mandatory in order to ensure that PG_M events are
4659 * not lost during any transition.
4661 * WARNING: The caller has busied the new page but not the original
4662 * vm_page which we are trying to replace. Because we hold
4663 * the pte_pv lock, but have not busied the page, PG bits
4664 * can be cleared out from under us.
4669 * NOTE: pt_pv won't exist for a kernel page
4670 * (managed or otherwise).
4672 * NOTE: We are reusing the pte_pv so we do not
4673 * destroy it in pmap_remove_pv_pte().
4675 if (prot & VM_PROT_NOSYNC) {
4676 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4678 pmap_inval_bulk_t bulk;
4680 pmap_inval_bulk_init(&bulk, pmap);
4681 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4682 pmap_inval_bulk_flush(&bulk);
4685 pmap_remove_pv_page(pte_pv);
4686 } else if (prot & VM_PROT_NOSYNC) {
4688 * Unmanaged page, NOSYNC (no mmu sync) requested.
4690 * Leave wire count on PT page intact.
4692 (void)pte_load_clear(ptep);
4693 cpu_invlpg((void *)va);
4694 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4697 * Unmanaged page, normal enter.
4699 * Leave wire count on PT page intact.
4701 pmap_inval_smp(pmap, va, 1, ptep, 0);
4702 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4704 KKASSERT(*ptep == 0);
4708 if (pmap_enter_debug > 0) {
4710 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4711 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4713 origpte, newpte, ptep,
4714 pte_pv, pt_pv, opa, prot);
4720 * Enter on the PV list if part of our managed memory.
4721 * Wiring of the PT page is already handled.
4723 KKASSERT(pte_pv->pv_m == NULL);
4724 vm_page_spin_lock(m);
4726 pmap_page_stats_adding(m);
4727 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4728 vm_page_flag_set(m, PG_MAPPED);
4729 vm_page_spin_unlock(m);
4730 } else if (pt_pv && opa == 0) {
4732 * We have to adjust the wire count on the PT page ourselves
4733 * for unmanaged entries. If opa was non-zero we retained
4734 * the existing wire count from the removal.
4736 vm_page_wire_quick(pt_pv->pv_m);
4740 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4742 * User VMAs do not because those will be zero->non-zero, so no
4743 * stale entries to worry about at this point.
4745 * For KVM there appear to still be issues. Theoretically we
4746 * should be able to scrap the interlocks entirely but we
4749 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4750 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4752 *(volatile pt_entry_t *)ptep = newpte;
4754 cpu_invlpg((void *)va);
4759 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4762 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4765 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4766 vm_page_flag_set(m, PG_WRITEABLE);
4769 * Unmanaged pages need manual resident_count tracking.
4771 if (pte_pv == NULL && pt_pv) {
4772 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4779 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4780 (m->flags & PG_MAPPED));
4783 * Cleanup the pv entry, allowing other accessors.
4787 else if (pte_placemark)
4788 pv_placemarker_wakeup(pmap, pte_placemark);
4795 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4796 * This code also assumes that the pmap has no pre-existing entry for this
4799 * This code currently may only be used on user pmaps, not kernel_pmap.
4802 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4804 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4808 * Make a temporary mapping for a physical address. This is only intended
4809 * to be used for panic dumps.
4811 * The caller is responsible for calling smp_invltlb().
4814 pmap_kenter_temporary(vm_paddr_t pa, long i)
4816 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4817 return ((void *)crashdumpmap);
4820 #define MAX_INIT_PT (96)
4823 * This routine preloads the ptes for a given object into the specified pmap.
4824 * This eliminates the blast of soft faults on process startup and
4825 * immediately after an mmap.
4827 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4830 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4831 vm_object_t object, vm_pindex_t pindex,
4832 vm_size_t size, int limit)
4834 struct rb_vm_page_scan_info info;
4839 * We can't preinit if read access isn't set or there is no pmap
4842 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4846 * We can't preinit if the pmap is not the current pmap
4848 lp = curthread->td_lwp;
4849 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4853 * Misc additional checks
4855 psize = x86_64_btop(size);
4857 if ((object->type != OBJT_VNODE) ||
4858 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4859 (object->resident_page_count > MAX_INIT_PT))) {
4863 if (pindex + psize > object->size) {
4864 if (object->size < pindex)
4866 psize = object->size - pindex;
4873 * If everything is segment-aligned do not pre-init here. Instead
4874 * allow the normal vm_fault path to pass a segment hint to
4875 * pmap_enter() which will then use an object-referenced shared
4878 if ((addr & SEG_MASK) == 0 &&
4879 (ctob(psize) & SEG_MASK) == 0 &&
4880 (ctob(pindex) & SEG_MASK) == 0) {
4885 * Use a red-black scan to traverse the requested range and load
4886 * any valid pages found into the pmap.
4888 * We cannot safely scan the object's memq without holding the
4891 info.start_pindex = pindex;
4892 info.end_pindex = pindex + psize - 1;
4898 vm_object_hold_shared(object);
4899 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4900 pmap_object_init_pt_callback, &info);
4901 vm_object_drop(object);
4906 pmap_object_init_pt_callback(vm_page_t p, void *data)
4908 struct rb_vm_page_scan_info *info = data;
4909 vm_pindex_t rel_index;
4912 * don't allow an madvise to blow away our really
4913 * free pages allocating pv entries.
4915 if ((info->limit & MAP_PREFAULT_MADVISE) &&
4916 vmstats.v_free_count < vmstats.v_free_reserved) {
4921 * Ignore list markers and ignore pages we cannot instantly
4922 * busy (while holding the object token).
4924 if (p->flags & PG_MARKER)
4926 if (vm_page_busy_try(p, TRUE))
4928 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4929 (p->flags & PG_FICTITIOUS) == 0) {
4930 if ((p->queue - p->pc) == PQ_CACHE)
4931 vm_page_deactivate(p);
4932 rel_index = p->pindex - info->start_pindex;
4933 pmap_enter_quick(info->pmap,
4934 info->addr + x86_64_ptob(rel_index), p);
4942 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4945 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4948 * XXX This is safe only because page table pages are not freed.
4951 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4955 /*spin_lock(&pmap->pm_spin);*/
4956 if ((pte = pmap_pte(pmap, addr)) != NULL) {
4957 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4958 /*spin_unlock(&pmap->pm_spin);*/
4962 /*spin_unlock(&pmap->pm_spin);*/
4967 * Change the wiring attribute for a pmap/va pair. The mapping must already
4968 * exist in the pmap. The mapping may or may not be managed. The wiring in
4969 * the page is not changed, the page is returned so the caller can adjust
4970 * its wiring (the page is not locked in any way).
4972 * Wiring is not a hardware characteristic so there is no need to invalidate
4973 * TLB. However, in an SMP environment we must use a locked bus cycle to
4974 * update the pte (if we are not using the pmap_inval_*() API that is)...
4975 * it's ok to do this for simple wiring changes.
4978 pmap_unwire(pmap_t pmap, vm_offset_t va)
4989 * Assume elements in the kernel pmap are stable
4991 if (pmap == &kernel_pmap) {
4992 if (pmap_pt(pmap, va) == 0)
4994 ptep = pmap_pte_quick(pmap, va);
4996 if (pmap_pte_w(pmap, ptep))
4997 atomic_add_long(&pmap->pm_stats.wired_count,-1);
4998 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4999 pa = *ptep & PG_FRAME;
5000 m = PHYS_TO_VM_PAGE(pa);
5003 * We can only [un]wire pmap-local pages (we cannot wire
5006 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5010 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5011 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5016 if (pmap_pte_w(pmap, ptep)) {
5017 atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5020 /* XXX else return NULL so caller doesn't unwire m ? */
5022 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5024 pa = *ptep & PG_FRAME;
5025 m = PHYS_TO_VM_PAGE(pa); /* held by wired count */
5032 * Copy the range specified by src_addr/len from the source map to
5033 * the range dst_addr/len in the destination map.
5035 * This routine is only advisory and need not do anything.
5038 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5039 vm_size_t len, vm_offset_t src_addr)
5046 * Zero the specified physical page.
5048 * This function may be called from an interrupt and no locking is
5052 pmap_zero_page(vm_paddr_t phys)
5054 vm_offset_t va = PHYS_TO_DMAP(phys);
5056 pagezero((void *)va);
5062 * Zero part of a physical page by mapping it into memory and clearing
5063 * its contents with bzero.
5065 * off and size may not cover an area beyond a single hardware page.
5068 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5070 vm_offset_t virt = PHYS_TO_DMAP(phys);
5072 bzero((char *)virt + off, size);
5078 * Copy the physical page from the source PA to the target PA.
5079 * This function may be called from an interrupt. No locking
5083 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5085 vm_offset_t src_virt, dst_virt;
5087 src_virt = PHYS_TO_DMAP(src);
5088 dst_virt = PHYS_TO_DMAP(dst);
5089 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5093 * pmap_copy_page_frag:
5095 * Copy the physical page from the source PA to the target PA.
5096 * This function may be called from an interrupt. No locking
5100 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5102 vm_offset_t src_virt, dst_virt;
5104 src_virt = PHYS_TO_DMAP(src);
5105 dst_virt = PHYS_TO_DMAP(dst);
5107 bcopy((char *)src_virt + (src & PAGE_MASK),
5108 (char *)dst_virt + (dst & PAGE_MASK),
5113 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5114 * this page. This count may be changed upwards or downwards in the future;
5115 * it is only necessary that true be returned for a small subset of pmaps
5116 * for proper page aging.
5119 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5124 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5127 vm_page_spin_lock(m);
5128 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5129 if (pv->pv_pmap == pmap) {
5130 vm_page_spin_unlock(m);
5137 vm_page_spin_unlock(m);
5142 * Remove all pages from specified address space this aids process exit
5143 * speeds. Also, this code may be special cased for the current process
5147 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5149 pmap_remove_noinval(pmap, sva, eva);
5154 * pmap_testbit tests bits in pte's note that the testbit/clearbit
5155 * routines are inline, and a lot of things compile-time evaluate.
5159 pmap_testbit(vm_page_t m, int bit)
5165 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5168 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5170 vm_page_spin_lock(m);
5171 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5172 vm_page_spin_unlock(m);
5176 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5178 #if defined(PMAP_DIAGNOSTIC)
5179 if (pv->pv_pmap == NULL) {
5180 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5188 * If the bit being tested is the modified bit, then
5189 * mark clean_map and ptes as never
5192 * WARNING! Because we do not lock the pv, *pte can be in a
5193 * state of flux. Despite this the value of *pte
5194 * will still be related to the vm_page in some way
5195 * because the pv cannot be destroyed as long as we
5196 * hold the vm_page spin lock.
5198 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5199 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5200 if (!pmap_track_modified(pv->pv_pindex))
5204 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5205 if (*pte & pmap->pmap_bits[bit]) {
5206 vm_page_spin_unlock(m);
5210 vm_page_spin_unlock(m);
5215 * This routine is used to modify bits in ptes. Only one bit should be
5216 * specified. PG_RW requires special handling.
5218 * Caller must NOT hold any spin locks
5222 pmap_clearbit(vm_page_t m, int bit_index)
5229 if (bit_index == PG_RW_IDX)
5230 vm_page_flag_clear(m, PG_WRITEABLE);
5231 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5238 * Loop over all current mappings setting/clearing as appropos If
5239 * setting RO do we need to clear the VAC?
5241 * NOTE: When clearing PG_M we could also (not implemented) drop
5242 * through to the PG_RW code and clear PG_RW too, forcing
5243 * a fault on write to redetect PG_M for virtual kernels, but
5244 * it isn't necessary since virtual kernels invalidate the
5245 * pte when they clear the VPTE_M bit in their virtual page
5248 * NOTE: Does not re-dirty the page when clearing only PG_M.
5250 * NOTE: Because we do not lock the pv, *pte can be in a state of
5251 * flux. Despite this the value of *pte is still somewhat
5252 * related while we hold the vm_page spin lock.
5254 * *pte can be zero due to this race. Since we are clearing
5255 * bits we basically do no harm when this race ccurs.
5257 if (bit_index != PG_RW_IDX) {
5258 vm_page_spin_lock(m);
5259 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5260 #if defined(PMAP_DIAGNOSTIC)
5261 if (pv->pv_pmap == NULL) {
5262 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5268 pte = pmap_pte_quick(pv->pv_pmap,
5269 pv->pv_pindex << PAGE_SHIFT);
5271 if (pbits & pmap->pmap_bits[bit_index])
5272 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5274 vm_page_spin_unlock(m);
5279 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
5283 vm_page_spin_lock(m);
5284 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5286 * don't write protect pager mappings
5288 if (!pmap_track_modified(pv->pv_pindex))
5291 #if defined(PMAP_DIAGNOSTIC)
5292 if (pv->pv_pmap == NULL) {
5293 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5300 * Skip pages which do not have PG_RW set.
5302 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5303 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5309 if (pv_hold_try(pv)) {
5310 vm_page_spin_unlock(m);
5312 vm_page_spin_unlock(m);
5313 pv_lock(pv); /* held, now do a blocking lock */
5315 if (pv->pv_pmap != pmap || pv->pv_m != m) {
5316 pv_put(pv); /* and release */
5317 goto restart; /* anything could have happened */
5319 KKASSERT(pv->pv_pmap == pmap);
5325 nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5326 pmap->pmap_bits[PG_M_IDX]);
5327 if (pmap_inval_smp_cmpset(pmap,
5328 ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5329 pte, pbits, nbits)) {
5334 vm_page_spin_lock(m);
5337 * If PG_M was found to be set while we were clearing PG_RW
5338 * we also clear PG_M (done above) and mark the page dirty.
5339 * Callers expect this behavior.
5341 if (pbits & pmap->pmap_bits[PG_M_IDX])
5345 vm_page_spin_unlock(m);
5349 * Lower the permission for all mappings to a given page.
5351 * Page must be busied by caller. Because page is busied by caller this
5352 * should not be able to race a pmap_enter().
5355 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5357 /* JG NX support? */
5358 if ((prot & VM_PROT_WRITE) == 0) {
5359 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5361 * NOTE: pmap_clearbit(.. PG_RW) also clears
5362 * the PG_WRITEABLE flag in (m).
5364 pmap_clearbit(m, PG_RW_IDX);
5372 pmap_phys_address(vm_pindex_t ppn)
5374 return (x86_64_ptob(ppn));
5378 * Return a count of reference bits for a page, clearing those bits.
5379 * It is not necessary for every reference bit to be cleared, but it
5380 * is necessary that 0 only be returned when there are truly no
5381 * reference bits set.
5383 * XXX: The exact number of bits to check and clear is a matter that
5384 * should be tested and standardized at some point in the future for
5385 * optimal aging of shared pages.
5387 * This routine may not block.
5390 pmap_ts_referenced(vm_page_t m)
5397 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5400 vm_page_spin_lock(m);
5401 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5402 if (!pmap_track_modified(pv->pv_pindex))
5405 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5406 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5407 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5413 vm_page_spin_unlock(m);
5420 * Return whether or not the specified physical page was modified
5421 * in any physical maps.
5424 pmap_is_modified(vm_page_t m)
5428 res = pmap_testbit(m, PG_M_IDX);
5433 * Clear the modify bits on the specified physical page.
5436 pmap_clear_modify(vm_page_t m)
5438 pmap_clearbit(m, PG_M_IDX);
5442 * pmap_clear_reference:
5444 * Clear the reference bit on the specified physical page.
5447 pmap_clear_reference(vm_page_t m)
5449 pmap_clearbit(m, PG_A_IDX);
5453 * Miscellaneous support routines follow
5458 i386_protection_init(void)
5462 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
5463 kp = protection_codes;
5464 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5466 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5468 * Read access is also 0. There isn't any execute bit,
5469 * so just make it readable.
5471 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5472 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5473 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5476 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5477 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5478 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5479 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5480 *kp++ = pmap_bits_default[PG_RW_IDX];
5487 * Map a set of physical memory pages into the kernel virtual
5488 * address space. Return a pointer to where it is mapped. This
5489 * routine is intended to be used for mapping device memory,
5492 * NOTE: We can't use pgeflag unless we invalidate the pages one at
5495 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5496 * work whether the cpu supports PAT or not. The remaining PAT
5497 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5501 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5503 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5507 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5509 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5513 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5515 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5519 * Map a set of physical memory pages into the kernel virtual
5520 * address space. Return a pointer to where it is mapped. This
5521 * routine is intended to be used for mapping device memory,
5525 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5527 vm_offset_t va, tmpva, offset;
5531 offset = pa & PAGE_MASK;
5532 size = roundup(offset + size, PAGE_SIZE);
5534 va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5536 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5538 pa = pa & ~PAGE_MASK;
5539 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5540 pte = vtopte(tmpva);
5542 kernel_pmap.pmap_bits[PG_RW_IDX] |
5543 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5544 kernel_pmap.pmap_cache_bits[mode];
5545 tmpsize -= PAGE_SIZE;
5549 pmap_invalidate_range(&kernel_pmap, va, va + size);
5550 pmap_invalidate_cache_range(va, va + size);
5552 return ((void *)(va + offset));
5556 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5558 vm_offset_t base, offset;
5560 base = va & ~PAGE_MASK;
5561 offset = va & PAGE_MASK;
5562 size = roundup(offset + size, PAGE_SIZE);
5563 pmap_qremove(va, size >> PAGE_SHIFT);
5564 kmem_free(&kernel_map, base, size);
5568 * Sets the memory attribute for the specified page.
5571 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5577 * If "m" is a normal page, update its direct mapping. This update
5578 * can be relied upon to perform any cache operations that are
5579 * required for data coherence.
5581 if ((m->flags & PG_FICTITIOUS) == 0)
5582 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5586 * Change the PAT attribute on an existing kernel memory map. Caller
5587 * must ensure that the virtual memory in question is not accessed
5588 * during the adjustment.
5591 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5598 panic("pmap_change_attr: va is NULL");
5599 base = trunc_page(va);
5603 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5604 kernel_pmap.pmap_cache_bits[mode];
5609 changed = 1; /* XXX: not optimal */
5612 * Flush CPU caches if required to make sure any data isn't cached that
5613 * shouldn't be, etc.
5616 pmap_invalidate_range(&kernel_pmap, base, va);
5617 pmap_invalidate_cache_range(base, va);
5622 * perform the pmap work for mincore
5625 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5627 pt_entry_t *ptep, pte;
5631 ptep = pmap_pte(pmap, addr);
5633 if (ptep && (pte = *ptep) != 0) {
5636 val = MINCORE_INCORE;
5637 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5640 pa = pte & PG_FRAME;
5642 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5645 m = PHYS_TO_VM_PAGE(pa);
5650 if (pte & pmap->pmap_bits[PG_M_IDX])
5651 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5653 * Modified by someone
5655 else if (m && (m->dirty || pmap_is_modified(m)))
5656 val |= MINCORE_MODIFIED_OTHER;
5660 if (pte & pmap->pmap_bits[PG_A_IDX])
5661 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5664 * Referenced by someone
5666 else if (m && ((m->flags & PG_REFERENCED) ||
5667 pmap_ts_referenced(m))) {
5668 val |= MINCORE_REFERENCED_OTHER;
5669 vm_page_flag_set(m, PG_REFERENCED);
5678 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5679 * vmspace will be ref'd and the old one will be deref'd.
5681 * The vmspace for all lwps associated with the process will be adjusted
5682 * and cr3 will be reloaded if any lwp is the current lwp.
5684 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5687 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5689 struct vmspace *oldvm;
5692 oldvm = p->p_vmspace;
5693 if (oldvm != newvm) {
5696 p->p_vmspace = newvm;
5697 KKASSERT(p->p_nthreads == 1);
5698 lp = RB_ROOT(&p->p_lwp_tree);
5699 pmap_setlwpvm(lp, newvm);
5706 * Set the vmspace for a LWP. The vmspace is almost universally set the
5707 * same as the process vmspace, but virtual kernels need to swap out contexts
5708 * on a per-lwp basis.
5710 * Caller does not necessarily hold any vmspace tokens. Caller must control
5711 * the lwp (typically be in the context of the lwp). We use a critical
5712 * section to protect against statclock and hardclock (statistics collection).
5715 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5717 struct vmspace *oldvm;
5720 oldvm = lp->lwp_vmspace;
5722 if (oldvm != newvm) {
5724 lp->lwp_vmspace = newvm;
5725 if (curthread->td_lwp == lp) {
5726 pmap = vmspace_pmap(newvm);
5727 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5728 if (pmap->pm_active_lock & CPULOCK_EXCL)
5729 pmap_interlock_wait(newvm);
5730 #if defined(SWTCH_OPTIM_STATS)
5733 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5734 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5735 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5736 curthread->td_pcb->pcb_cr3 = KPML4phys;
5738 panic("pmap_setlwpvm: unknown pmap type\n");
5740 load_cr3(curthread->td_pcb->pcb_cr3);
5741 pmap = vmspace_pmap(oldvm);
5742 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5750 * Called when switching to a locked pmap, used to interlock against pmaps
5751 * undergoing modifications to prevent us from activating the MMU for the
5752 * target pmap until all such modifications have completed. We have to do
5753 * this because the thread making the modifications has already set up its
5754 * SMP synchronization mask.
5756 * This function cannot sleep!
5761 pmap_interlock_wait(struct vmspace *vm)
5763 struct pmap *pmap = &vm->vm_pmap;
5765 if (pmap->pm_active_lock & CPULOCK_EXCL) {
5767 KKASSERT(curthread->td_critcount >= 2);
5768 DEBUG_PUSH_INFO("pmap_interlock_wait");
5769 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5771 lwkt_process_ipiq();
5779 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5782 if ((obj == NULL) || (size < NBPDR) ||
5783 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5787 addr = roundup2(addr, NBPDR);
5792 * Used by kmalloc/kfree, page already exists at va
5795 pmap_kvtom(vm_offset_t va)
5797 pt_entry_t *ptep = vtopte(va);
5799 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5800 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5804 * Initialize machine-specific shared page directory support. This
5805 * is executed when a VM object is created.
5808 pmap_object_init(vm_object_t object)
5810 object->md.pmap_rw = NULL;
5811 object->md.pmap_ro = NULL;
5815 * Clean up machine-specific shared page directory support. This
5816 * is executed when a VM object is destroyed.
5819 pmap_object_free(vm_object_t object)
5823 if ((pmap = object->md.pmap_rw) != NULL) {
5824 object->md.pmap_rw = NULL;
5825 pmap_remove_noinval(pmap,
5826 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5827 CPUMASK_ASSZERO(pmap->pm_active);
5830 kfree(pmap, M_OBJPMAP);
5832 if ((pmap = object->md.pmap_ro) != NULL) {
5833 object->md.pmap_ro = NULL;
5834 pmap_remove_noinval(pmap,
5835 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5836 CPUMASK_ASSZERO(pmap->pm_active);
5839 kfree(pmap, M_OBJPMAP);
5844 * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5845 * VM page and issue a pginfo->callback.
5847 * We are expected to dispose of any non-NULL pte_pv.
5851 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5852 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
5853 pv_entry_t pt_pv, int sharept,
5854 vm_offset_t va, pt_entry_t *ptep, void *arg)
5856 struct pmap_pgscan_info *pginfo = arg;
5861 * Try to busy the page while we hold the pte_pv locked.
5863 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
5864 if (vm_page_busy_try(m, TRUE) == 0) {
5865 if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
5867 * The callback is issued with the pte_pv
5868 * unlocked and put away, and the pt_pv
5873 vm_page_wire_quick(pt_pv->pv_m);
5876 if (pginfo->callback(pginfo, va, m) < 0)
5880 vm_page_unwire_quick(pt_pv->pv_m);
5887 ++pginfo->busycount;
5890 } else if (sharept) {
5891 /* shared page table */
5892 pv_placemarker_wakeup(pmap, pte_placemark);
5894 /* else unmanaged page */
5895 pv_placemarker_wakeup(pmap, pte_placemark);
5900 pmap_pgscan(struct pmap_pgscan_info *pginfo)
5902 struct pmap_scan_info info;
5904 pginfo->offset = pginfo->beg_addr;
5905 info.pmap = pginfo->pmap;
5906 info.sva = pginfo->beg_addr;
5907 info.eva = pginfo->end_addr;
5908 info.func = pmap_pgscan_callback;
5910 pmap_scan(&info, 0);
5912 pginfo->offset = pginfo->end_addr;
5916 * Wait for a placemarker that we do not own to clear. The placemarker
5917 * in question is not necessary set to the pindex we want, we may have
5918 * to wait on the element because we want to reserve it ourselves.
5922 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
5924 spin_lock(&pmap->pm_spin);
5925 if (*pmark != PM_NOPLACEMARK) {
5926 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
5927 ssleep(pmark, &pmap->pm_spin, 0, "pvplw", 0);
5929 spin_unlock(&pmap->pm_spin);
5933 * Wakeup a placemarker that we own. Replace the entry with
5934 * PM_NOPLACEMARK and issue a wakeup() if necessary.
5938 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
5942 spin_lock(&pmap->pm_spin);
5943 pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
5944 spin_unlock(&pmap->pm_spin);
5945 KKASSERT(pindex != PM_NOPLACEMARK);
5946 if (pindex & PM_PLACEMARK_WAKEUP)