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 uint8_t *base);
261 extern int std_subyte (uint8_t *base, uint8_t byte);
262 extern int32_t std_fuword32 (const uint32_t *base);
263 extern int64_t std_fuword64 (const uint64_t *base);
264 extern int std_suword64 (uint64_t *base, uint64_t word);
265 extern int std_suword32 (uint32_t *base, int word);
266 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
267 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
269 static void pv_hold(pv_entry_t pv);
270 static int _pv_hold_try(pv_entry_t pv
272 static void pv_drop(pv_entry_t pv);
273 static void _pv_lock(pv_entry_t pv
275 static void pv_unlock(pv_entry_t pv);
276 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
278 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
280 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
281 vm_pindex_t **pmarkp, int *errorp);
282 static void pv_put(pv_entry_t pv);
283 static void pv_free(pv_entry_t pv, pv_entry_t pvp);
284 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
285 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
287 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
288 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
289 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
290 pmap_inval_bulk_t *bulk, int destroy);
291 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
292 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
293 pmap_inval_bulk_t *bulk);
295 struct pmap_scan_info;
296 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
297 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
298 pv_entry_t pt_pv, int sharept,
299 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
300 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
301 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
302 pv_entry_t pt_pv, int sharept,
303 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
305 static void i386_protection_init (void);
306 static void create_pagetables(vm_paddr_t *firstaddr);
307 static void pmap_remove_all (vm_page_t m);
308 static boolean_t pmap_testbit (vm_page_t m, int bit);
310 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
311 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
313 static void pmap_pinit_defaults(struct pmap *pmap);
314 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
315 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
317 static unsigned pdir4mb;
320 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
322 if (pv1->pv_pindex < pv2->pv_pindex)
324 if (pv1->pv_pindex > pv2->pv_pindex)
329 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
330 pv_entry_compare, vm_pindex_t, pv_pindex);
334 pmap_page_stats_adding(vm_page_t m)
336 globaldata_t gd = mycpu;
338 if (TAILQ_EMPTY(&m->md.pv_list)) {
339 ++gd->gd_vmtotal.t_arm;
340 } else if (TAILQ_FIRST(&m->md.pv_list) ==
341 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
342 ++gd->gd_vmtotal.t_armshr;
343 ++gd->gd_vmtotal.t_avmshr;
345 ++gd->gd_vmtotal.t_avmshr;
351 pmap_page_stats_deleting(vm_page_t m)
353 globaldata_t gd = mycpu;
355 if (TAILQ_EMPTY(&m->md.pv_list)) {
356 --gd->gd_vmtotal.t_arm;
357 } else if (TAILQ_FIRST(&m->md.pv_list) ==
358 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
359 --gd->gd_vmtotal.t_armshr;
360 --gd->gd_vmtotal.t_avmshr;
362 --gd->gd_vmtotal.t_avmshr;
367 * Move the kernel virtual free pointer to the next
368 * 2MB. This is used to help improve performance
369 * by using a large (2MB) page for much of the kernel
370 * (.text, .data, .bss)
374 pmap_kmem_choose(vm_offset_t addr)
376 vm_offset_t newaddr = addr;
378 newaddr = roundup2(addr, NBPDR);
385 * Super fast pmap_pte routine best used when scanning the pv lists.
386 * This eliminates many course-grained invltlb calls. Note that many of
387 * the pv list scans are across different pmaps and it is very wasteful
388 * to do an entire invltlb when checking a single mapping.
390 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
394 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
396 return pmap_pte(pmap, va);
400 * Returns the pindex of a page table entry (representing a terminal page).
401 * There are NUPTE_TOTAL page table entries possible (a huge number)
403 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
404 * We want to properly translate negative KVAs.
408 pmap_pte_pindex(vm_offset_t va)
410 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
414 * Returns the pindex of a page table.
418 pmap_pt_pindex(vm_offset_t va)
420 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
424 * Returns the pindex of a page directory.
428 pmap_pd_pindex(vm_offset_t va)
430 return (NUPTE_TOTAL + NUPT_TOTAL +
431 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
436 pmap_pdp_pindex(vm_offset_t va)
438 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
439 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
444 pmap_pml4_pindex(void)
446 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
450 * Return various clipped indexes for a given VA
452 * Returns the index of a pt in a page directory, representing a page
457 pmap_pt_index(vm_offset_t va)
459 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
463 * Returns the index of a pd in a page directory page, representing a page
468 pmap_pd_index(vm_offset_t va)
470 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
474 * Returns the index of a pdp in the pml4 table, representing a page
479 pmap_pdp_index(vm_offset_t va)
481 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
485 * The placemarker hash must be broken up into four zones so lock
486 * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
488 * Placemarkers are used to 'lock' page table indices that do not have
489 * a pv_entry. This allows the pmap to support managed and unmanaged
490 * pages and shared page tables.
492 #define PM_PLACE_BASE (PM_PLACEMARKS >> 2)
496 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
500 if (pindex < pmap_pt_pindex(0)) /* zone 0 - PTE */
502 else if (pindex < pmap_pd_pindex(0)) /* zone 1 - PT */
504 else if (pindex < pmap_pdp_pindex(0)) /* zone 2 - PD */
505 hi = PM_PLACE_BASE << 1;
506 else /* zone 3 - PDP (and PML4E) */
507 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
508 hi += pindex & (PM_PLACE_BASE - 1);
510 return (&pmap->pm_placemarks[hi]);
515 * Generic procedure to index a pte from a pt, pd, or pdp.
517 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
518 * a page table page index but is instead of PV lookup index.
522 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
526 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
527 return(&pte[pindex]);
531 * Return pointer to PDP slot in the PML4
535 pmap_pdp(pmap_t pmap, vm_offset_t va)
537 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
541 * Return pointer to PD slot in the PDP given a pointer to the PDP
545 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
549 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
550 return (&pd[pmap_pd_index(va)]);
554 * Return pointer to PD slot in the PDP.
558 pmap_pd(pmap_t pmap, vm_offset_t va)
562 pdp = pmap_pdp(pmap, va);
563 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
565 return (pmap_pdp_to_pd(*pdp, va));
569 * Return pointer to PT slot in the PD given a pointer to the PD
573 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
577 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
578 return (&pt[pmap_pt_index(va)]);
582 * Return pointer to PT slot in the PD
584 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
585 * so we cannot lookup the PD via the PDP. Instead we
586 * must look it up via the pmap.
590 pmap_pt(pmap_t pmap, vm_offset_t va)
594 vm_pindex_t pd_pindex;
596 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
597 pd_pindex = pmap_pd_pindex(va);
598 spin_lock(&pmap->pm_spin);
599 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
600 spin_unlock(&pmap->pm_spin);
601 if (pv == NULL || pv->pv_m == NULL)
603 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
605 pd = pmap_pd(pmap, va);
606 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
608 return (pmap_pd_to_pt(*pd, va));
613 * Return pointer to PTE slot in the PT given a pointer to the PT
617 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
621 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
622 return (&pte[pmap_pte_index(va)]);
626 * Return pointer to PTE slot in the PT
630 pmap_pte(pmap_t pmap, vm_offset_t va)
634 pt = pmap_pt(pmap, va);
635 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
637 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
638 return ((pt_entry_t *)pt);
639 return (pmap_pt_to_pte(*pt, va));
643 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
644 * the PT layer. This will speed up core pmap operations considerably.
646 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
647 * must be in a known associated state (typically by being locked when
648 * the pmap spinlock isn't held). We allow the race for that case.
650 * NOTE: pm_pvhint is only accessed (read) with the spin-lock held, using
651 * cpu_ccfence() to prevent compiler optimizations from reloading the
656 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
658 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0)) {
660 pv->pv_pmap->pm_pvhint = pv;
666 * Return address of PT slot in PD (KVM only)
668 * Cannot be used for user page tables because it might interfere with
669 * the shared page-table-page optimization (pmap_mmu_optimize).
673 vtopt(vm_offset_t va)
675 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
676 NPML4EPGSHIFT)) - 1);
678 return (PDmap + ((va >> PDRSHIFT) & mask));
682 * KVM - return address of PTE slot in PT
686 vtopte(vm_offset_t va)
688 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
689 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
691 return (PTmap + ((va >> PAGE_SHIFT) & mask));
695 allocpages(vm_paddr_t *firstaddr, long n)
700 bzero((void *)ret, n * PAGE_SIZE);
701 *firstaddr += n * PAGE_SIZE;
707 create_pagetables(vm_paddr_t *firstaddr)
709 long i; /* must be 64 bits */
715 * We are running (mostly) V=P at this point
717 * Calculate NKPT - number of kernel page tables. We have to
718 * accomodoate prealloction of the vm_page_array, dump bitmap,
719 * MSGBUF_SIZE, and other stuff. Be generous.
721 * Maxmem is in pages.
723 * ndmpdp is the number of 1GB pages we wish to map.
725 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
726 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
728 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
731 * Starting at the beginning of kvm (not KERNBASE).
733 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
734 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
735 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
736 ndmpdp) + 511) / 512;
740 * Starting at KERNBASE - map 2G worth of page table pages.
741 * KERNBASE is offset -2G from the end of kvm.
743 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
748 KPTbase = allocpages(firstaddr, nkpt_base);
749 KPTphys = allocpages(firstaddr, nkpt_phys);
750 KPML4phys = allocpages(firstaddr, 1);
751 KPDPphys = allocpages(firstaddr, NKPML4E);
752 KPDphys = allocpages(firstaddr, NKPDPE);
755 * Calculate the page directory base for KERNBASE,
756 * that is where we start populating the page table pages.
757 * Basically this is the end - 2.
759 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
761 DMPDPphys = allocpages(firstaddr, NDMPML4E);
762 if ((amd_feature & AMDID_PAGE1GB) == 0)
763 DMPDphys = allocpages(firstaddr, ndmpdp);
764 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
767 * Fill in the underlying page table pages for the area around
768 * KERNBASE. This remaps low physical memory to KERNBASE.
770 * Read-only from zero to physfree
771 * XXX not fully used, underneath 2M pages
773 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
774 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
775 ((pt_entry_t *)KPTbase)[i] |=
776 pmap_bits_default[PG_RW_IDX] |
777 pmap_bits_default[PG_V_IDX] |
778 pmap_bits_default[PG_G_IDX];
782 * Now map the initial kernel page tables. One block of page
783 * tables is placed at the beginning of kernel virtual memory,
784 * and another block is placed at KERNBASE to map the kernel binary,
785 * data, bss, and initial pre-allocations.
787 for (i = 0; i < nkpt_base; i++) {
788 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
789 ((pd_entry_t *)KPDbase)[i] |=
790 pmap_bits_default[PG_RW_IDX] |
791 pmap_bits_default[PG_V_IDX];
793 for (i = 0; i < nkpt_phys; i++) {
794 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
795 ((pd_entry_t *)KPDphys)[i] |=
796 pmap_bits_default[PG_RW_IDX] |
797 pmap_bits_default[PG_V_IDX];
801 * Map from zero to end of allocations using 2M pages as an
802 * optimization. This will bypass some of the KPTBase pages
803 * above in the KERNBASE area.
805 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
806 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
807 ((pd_entry_t *)KPDbase)[i] |=
808 pmap_bits_default[PG_RW_IDX] |
809 pmap_bits_default[PG_V_IDX] |
810 pmap_bits_default[PG_PS_IDX] |
811 pmap_bits_default[PG_G_IDX];
815 * And connect up the PD to the PDP. The kernel pmap is expected
816 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
818 for (i = 0; i < NKPDPE; i++) {
819 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
820 KPDphys + (i << PAGE_SHIFT);
821 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
822 pmap_bits_default[PG_RW_IDX] |
823 pmap_bits_default[PG_V_IDX] |
824 pmap_bits_default[PG_U_IDX];
828 * Now set up the direct map space using either 2MB or 1GB pages
829 * Preset PG_M and PG_A because demotion expects it.
831 * When filling in entries in the PD pages make sure any excess
832 * entries are set to zero as we allocated enough PD pages
834 if ((amd_feature & AMDID_PAGE1GB) == 0) {
835 for (i = 0; i < NPDEPG * ndmpdp; i++) {
836 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
837 ((pd_entry_t *)DMPDphys)[i] |=
838 pmap_bits_default[PG_RW_IDX] |
839 pmap_bits_default[PG_V_IDX] |
840 pmap_bits_default[PG_PS_IDX] |
841 pmap_bits_default[PG_G_IDX] |
842 pmap_bits_default[PG_M_IDX] |
843 pmap_bits_default[PG_A_IDX];
847 * And the direct map space's PDP
849 for (i = 0; i < ndmpdp; i++) {
850 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
852 ((pdp_entry_t *)DMPDPphys)[i] |=
853 pmap_bits_default[PG_RW_IDX] |
854 pmap_bits_default[PG_V_IDX] |
855 pmap_bits_default[PG_U_IDX];
858 for (i = 0; i < ndmpdp; i++) {
859 ((pdp_entry_t *)DMPDPphys)[i] =
860 (vm_paddr_t)i << PDPSHIFT;
861 ((pdp_entry_t *)DMPDPphys)[i] |=
862 pmap_bits_default[PG_RW_IDX] |
863 pmap_bits_default[PG_V_IDX] |
864 pmap_bits_default[PG_PS_IDX] |
865 pmap_bits_default[PG_G_IDX] |
866 pmap_bits_default[PG_M_IDX] |
867 pmap_bits_default[PG_A_IDX];
871 /* And recursively map PML4 to itself in order to get PTmap */
872 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
873 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
874 pmap_bits_default[PG_RW_IDX] |
875 pmap_bits_default[PG_V_IDX] |
876 pmap_bits_default[PG_U_IDX];
879 * Connect the Direct Map slots up to the PML4
881 for (j = 0; j < NDMPML4E; ++j) {
882 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
883 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
884 pmap_bits_default[PG_RW_IDX] |
885 pmap_bits_default[PG_V_IDX] |
886 pmap_bits_default[PG_U_IDX];
890 * Connect the KVA slot up to the PML4
892 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
893 ((pdp_entry_t *)KPML4phys)[KPML4I] |=
894 pmap_bits_default[PG_RW_IDX] |
895 pmap_bits_default[PG_V_IDX] |
896 pmap_bits_default[PG_U_IDX];
900 * Bootstrap the system enough to run with virtual memory.
902 * On the i386 this is called after mapping has already been enabled
903 * and just syncs the pmap module with what has already been done.
904 * [We can't call it easily with mapping off since the kernel is not
905 * mapped with PA == VA, hence we would have to relocate every address
906 * from the linked base (virtual) address "KERNBASE" to the actual
907 * (physical) address starting relative to 0]
910 pmap_bootstrap(vm_paddr_t *firstaddr)
916 KvaStart = VM_MIN_KERNEL_ADDRESS;
917 KvaEnd = VM_MAX_KERNEL_ADDRESS;
918 KvaSize = KvaEnd - KvaStart;
920 avail_start = *firstaddr;
923 * Create an initial set of page tables to run the kernel in.
925 create_pagetables(firstaddr);
927 virtual2_start = KvaStart;
928 virtual2_end = PTOV_OFFSET;
930 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
931 virtual_start = pmap_kmem_choose(virtual_start);
933 virtual_end = VM_MAX_KERNEL_ADDRESS;
935 /* XXX do %cr0 as well */
936 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
940 * Initialize protection array.
942 i386_protection_init();
945 * The kernel's pmap is statically allocated so we don't have to use
946 * pmap_create, which is unlikely to work correctly at this part of
947 * the boot sequence (XXX and which no longer exists).
949 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
950 kernel_pmap.pm_count = 1;
951 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
952 RB_INIT(&kernel_pmap.pm_pvroot);
953 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
954 for (i = 0; i < PM_PLACEMARKS; ++i)
955 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
958 * Reserve some special page table entries/VA space for temporary
961 #define SYSMAP(c, p, v, n) \
962 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
968 * CMAP1/CMAP2 are used for zeroing and copying pages.
970 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
975 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
978 * ptvmmap is used for reading arbitrary physical pages via
981 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
984 * msgbufp is used to map the system message buffer.
985 * XXX msgbufmap is not used.
987 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
988 atop(round_page(MSGBUF_SIZE)))
991 virtual_start = pmap_kmem_choose(virtual_start);
996 * PG_G is terribly broken on SMP because we IPI invltlb's in some
997 * cases rather then invl1pg. Actually, I don't even know why it
998 * works under UP because self-referential page table mappings
1003 * Initialize the 4MB page size flag
1007 * The 4MB page version of the initial
1008 * kernel page mapping.
1012 #if !defined(DISABLE_PSE)
1013 if (cpu_feature & CPUID_PSE) {
1016 * Note that we have enabled PSE mode
1018 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1019 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
1020 ptditmp &= ~(NBPDR - 1);
1021 ptditmp |= pmap_bits_default[PG_V_IDX] |
1022 pmap_bits_default[PG_RW_IDX] |
1023 pmap_bits_default[PG_PS_IDX] |
1024 pmap_bits_default[PG_U_IDX];
1031 /* Initialize the PAT MSR */
1033 pmap_pinit_defaults(&kernel_pmap);
1035 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1036 &pmap_fast_kernel_cpusync);
1041 * Setup the PAT MSR.
1050 * Default values mapping PATi,PCD,PWT bits at system reset.
1051 * The default values effectively ignore the PATi bit by
1052 * repeating the encodings for 0-3 in 4-7, and map the PCD
1053 * and PWT bit combinations to the expected PAT types.
1055 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
1056 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
1057 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
1058 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
1059 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
1060 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
1061 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
1062 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
1063 pat_pte_index[PAT_WRITE_BACK] = 0;
1064 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
1065 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
1066 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
1067 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1068 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1070 if (cpu_feature & CPUID_PAT) {
1072 * If we support the PAT then set-up entries for
1073 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1076 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1077 PAT_VALUE(4, PAT_WRITE_PROTECTED);
1078 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1079 PAT_VALUE(5, PAT_WRITE_COMBINING);
1080 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1081 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1084 * Then enable the PAT
1089 load_cr4(cr4 & ~CR4_PGE);
1091 /* Disable caches (CD = 1, NW = 0). */
1093 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1095 /* Flushes caches and TLBs. */
1099 /* Update PAT and index table. */
1100 wrmsr(MSR_PAT, pat_msr);
1102 /* Flush caches and TLBs again. */
1106 /* Restore caches and PGE. */
1114 * Set 4mb pdir for mp startup
1119 if (cpu_feature & CPUID_PSE) {
1120 load_cr4(rcr4() | CR4_PSE);
1121 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
1128 * Initialize the pmap module.
1129 * Called by vm_init, to initialize any structures that the pmap
1130 * system needs to map virtual memory.
1131 * pmap_init has been enhanced to support in a fairly consistant
1132 * way, discontiguous physical memory.
1141 * Allocate memory for random pmap data structures. Includes the
1145 for (i = 0; i < vm_page_array_size; i++) {
1148 m = &vm_page_array[i];
1149 TAILQ_INIT(&m->md.pv_list);
1153 * init the pv free list
1155 initial_pvs = vm_page_array_size;
1156 if (initial_pvs < MINPV)
1157 initial_pvs = MINPV;
1158 pvzone = &pvzone_store;
1159 pvinit = (void *)kmem_alloc(&kernel_map,
1160 initial_pvs * sizeof (struct pv_entry),
1162 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1163 pvinit, initial_pvs);
1166 * Now it is safe to enable pv_table recording.
1168 pmap_initialized = TRUE;
1172 * Initialize the address space (zone) for the pv_entries. Set a
1173 * high water mark so that the system can recover from excessive
1174 * numbers of pv entries.
1179 int shpgperproc = PMAP_SHPGPERPROC;
1182 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1183 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1184 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1185 pv_entry_high_water = 9 * (pv_entry_max / 10);
1188 * Subtract out pages already installed in the zone (hack)
1190 entry_max = pv_entry_max - vm_page_array_size;
1194 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1197 * Enable dynamic deletion of empty higher-level page table pages
1198 * by default only if system memory is < 8GB (use 7GB for slop).
1199 * This can save a little memory, but imposes significant
1200 * performance overhead for things like bulk builds, and for programs
1201 * which do a lot of memory mapping and memory unmapping.
1203 if (pmap_dynamic_delete < 0) {
1204 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1205 pmap_dynamic_delete = 1;
1207 pmap_dynamic_delete = 0;
1212 * Typically used to initialize a fictitious page by vm/device_pager.c
1215 pmap_page_init(struct vm_page *m)
1218 TAILQ_INIT(&m->md.pv_list);
1221 /***************************************************
1222 * Low level helper routines.....
1223 ***************************************************/
1226 * this routine defines the region(s) of memory that should
1227 * not be tested for the modified bit.
1231 pmap_track_modified(vm_pindex_t pindex)
1233 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1234 if ((va < clean_sva) || (va >= clean_eva))
1241 * Extract the physical page address associated with the map/VA pair.
1242 * The page must be wired for this to work reliably.
1245 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1252 if (va >= VM_MAX_USER_ADDRESS) {
1254 * Kernel page directories might be direct-mapped and
1255 * there is typically no PV tracking of pte's
1259 pt = pmap_pt(pmap, va);
1260 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1261 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1262 rtval = *pt & PG_PS_FRAME;
1263 rtval |= va & PDRMASK;
1265 ptep = pmap_pt_to_pte(*pt, va);
1266 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1267 rtval = *ptep & PG_FRAME;
1268 rtval |= va & PAGE_MASK;
1276 * User pages currently do not direct-map the page directory
1277 * and some pages might not used managed PVs. But all PT's
1280 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1282 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1283 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1284 rtval = *ptep & PG_FRAME;
1285 rtval |= va & PAGE_MASK;
1288 *handlep = pt_pv; /* locked until done */
1291 } else if (handlep) {
1299 pmap_extract_done(void *handle)
1302 pv_put((pv_entry_t)handle);
1306 * Similar to extract but checks protections, SMP-friendly short-cut for
1307 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1308 * fall-through to the real fault code. Does not work with HVM page
1311 * The returned page, if not NULL, is held (and not busied).
1313 * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1317 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1320 va < VM_MAX_USER_ADDRESS &&
1321 (pmap->pm_flags & PMAP_HVM) == 0) {
1329 req = pmap->pmap_bits[PG_V_IDX] |
1330 pmap->pmap_bits[PG_U_IDX];
1331 if (prot & VM_PROT_WRITE)
1332 req |= pmap->pmap_bits[PG_RW_IDX];
1334 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1337 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1338 if ((*ptep & req) != req) {
1342 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1343 if (pte_pv && error == 0) {
1345 if (prot & VM_PROT_WRITE) {
1346 /* interlocked by presence of pv_entry */
1350 if (prot & VM_PROT_WRITE) {
1351 if (vm_page_busy_try(m, TRUE))
1362 } else if (pte_pv) {
1366 /* error can be 0 or 1 */
1377 * Extract the physical page address associated kernel virtual address.
1380 pmap_kextract(vm_offset_t va)
1382 pd_entry_t pt; /* pt entry in pd */
1385 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1386 pa = DMAP_TO_PHYS(va);
1389 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1390 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1393 * Beware of a concurrent promotion that changes the
1394 * PDE at this point! For example, vtopte() must not
1395 * be used to access the PTE because it would use the
1396 * new PDE. It is, however, safe to use the old PDE
1397 * because the page table page is preserved by the
1400 pa = *pmap_pt_to_pte(pt, va);
1401 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1407 /***************************************************
1408 * Low level mapping routines.....
1409 ***************************************************/
1412 * Routine: pmap_kenter
1414 * Add a wired page to the KVA
1415 * NOTE! note that in order for the mapping to take effect -- you
1416 * should do an invltlb after doing the pmap_kenter().
1419 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1425 kernel_pmap.pmap_bits[PG_RW_IDX] |
1426 kernel_pmap.pmap_bits[PG_V_IDX];
1430 pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1434 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1441 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1442 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1443 * (caller can conditionalize calling smp_invltlb()).
1446 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1452 npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1453 kernel_pmap.pmap_bits[PG_V_IDX];
1462 atomic_swap_long(ptep, npte);
1463 cpu_invlpg((void *)va);
1469 * Enter addresses into the kernel pmap but don't bother
1470 * doing any tlb invalidations. Caller will do a rollup
1471 * invalidation via pmap_rollup_inval().
1474 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1481 kernel_pmap.pmap_bits[PG_RW_IDX] |
1482 kernel_pmap.pmap_bits[PG_V_IDX];
1491 atomic_swap_long(ptep, npte);
1492 cpu_invlpg((void *)va);
1498 * remove a page from the kernel pagetables
1501 pmap_kremove(vm_offset_t va)
1506 pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1510 pmap_kremove_quick(vm_offset_t va)
1515 (void)pte_load_clear(ptep);
1516 cpu_invlpg((void *)va);
1520 * Remove addresses from the kernel pmap but don't bother
1521 * doing any tlb invalidations. Caller will do a rollup
1522 * invalidation via pmap_rollup_inval().
1525 pmap_kremove_noinval(vm_offset_t va)
1530 (void)pte_load_clear(ptep);
1534 * XXX these need to be recoded. They are not used in any critical path.
1537 pmap_kmodify_rw(vm_offset_t va)
1539 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1540 cpu_invlpg((void *)va);
1545 pmap_kmodify_nc(vm_offset_t va)
1547 atomic_set_long(vtopte(va), PG_N);
1548 cpu_invlpg((void *)va);
1553 * Used to map a range of physical addresses into kernel virtual
1554 * address space during the low level boot, typically to map the
1555 * dump bitmap, message buffer, and vm_page_array.
1557 * These mappings are typically made at some pointer after the end of the
1560 * We could return PHYS_TO_DMAP(start) here and not allocate any
1561 * via (*virtp), but then kmem from userland and kernel dumps won't
1562 * have access to the related pointers.
1565 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1568 vm_offset_t va_start;
1570 /*return PHYS_TO_DMAP(start);*/
1575 while (start < end) {
1576 pmap_kenter_quick(va, start);
1584 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1587 * Remove the specified set of pages from the data and instruction caches.
1589 * In contrast to pmap_invalidate_cache_range(), this function does not
1590 * rely on the CPU's self-snoop feature, because it is intended for use
1591 * when moving pages into a different cache domain.
1594 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1596 vm_offset_t daddr, eva;
1599 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1600 (cpu_feature & CPUID_CLFSH) == 0)
1604 for (i = 0; i < count; i++) {
1605 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1606 eva = daddr + PAGE_SIZE;
1607 for (; daddr < eva; daddr += cpu_clflush_line_size)
1615 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1617 KASSERT((sva & PAGE_MASK) == 0,
1618 ("pmap_invalidate_cache_range: sva not page-aligned"));
1619 KASSERT((eva & PAGE_MASK) == 0,
1620 ("pmap_invalidate_cache_range: eva not page-aligned"));
1622 if (cpu_feature & CPUID_SS) {
1623 ; /* If "Self Snoop" is supported, do nothing. */
1625 /* Globally invalidate caches */
1626 cpu_wbinvd_on_all_cpus();
1631 * Invalidate the specified range of virtual memory on all cpus associated
1635 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1637 pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1641 * Add a list of wired pages to the kva. This routine is used for temporary
1642 * kernel mappings such as those found in buffer cache buffer. Page
1643 * modifications and accesses are not tracked or recorded.
1645 * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1646 * semantics as previous mappings may have been zerod without any
1649 * The page *must* be wired.
1652 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1657 end_va = beg_va + count * PAGE_SIZE;
1659 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1664 pte = VM_PAGE_TO_PHYS(*m) |
1665 kernel_pmap.pmap_bits[PG_RW_IDX] |
1666 kernel_pmap.pmap_bits[PG_V_IDX] |
1667 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1669 atomic_swap_long(ptep, pte);
1672 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1676 * This routine jerks page mappings from the kernel -- it is meant only
1677 * for temporary mappings such as those found in buffer cache buffers.
1678 * No recording modified or access status occurs.
1680 * MPSAFE, INTERRUPT SAFE (cluster callback)
1683 pmap_qremove(vm_offset_t beg_va, int count)
1688 end_va = beg_va + count * PAGE_SIZE;
1690 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1694 (void)pte_load_clear(pte);
1695 cpu_invlpg((void *)va);
1697 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1701 * This routine removes temporary kernel mappings, only invalidating them
1702 * on the current cpu. It should only be used under carefully controlled
1706 pmap_qremove_quick(vm_offset_t beg_va, int count)
1711 end_va = beg_va + count * PAGE_SIZE;
1713 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1717 (void)pte_load_clear(pte);
1718 cpu_invlpg((void *)va);
1723 * This routine removes temporary kernel mappings *without* invalidating
1724 * the TLB. It can only be used on permanent kva reservations such as those
1725 * found in buffer cache buffers, under carefully controlled circumstances.
1727 * NOTE: Repopulating these KVAs requires unconditional invalidation.
1728 * (pmap_qenter() does unconditional invalidation).
1731 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1736 end_va = beg_va + count * PAGE_SIZE;
1738 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1742 (void)pte_load_clear(pte);
1747 * Create a new thread and optionally associate it with a (new) process.
1748 * NOTE! the new thread's cpu may not equal the current cpu.
1751 pmap_init_thread(thread_t td)
1753 /* enforce pcb placement & alignment */
1754 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1755 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1756 td->td_savefpu = &td->td_pcb->pcb_save;
1757 td->td_sp = (char *)td->td_pcb; /* no -16 */
1761 * This routine directly affects the fork perf for a process.
1764 pmap_init_proc(struct proc *p)
1769 pmap_pinit_defaults(struct pmap *pmap)
1771 bcopy(pmap_bits_default, pmap->pmap_bits,
1772 sizeof(pmap_bits_default));
1773 bcopy(protection_codes, pmap->protection_codes,
1774 sizeof(protection_codes));
1775 bcopy(pat_pte_index, pmap->pmap_cache_bits,
1776 sizeof(pat_pte_index));
1777 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1778 pmap->copyinstr = std_copyinstr;
1779 pmap->copyin = std_copyin;
1780 pmap->copyout = std_copyout;
1781 pmap->fubyte = std_fubyte;
1782 pmap->subyte = std_subyte;
1783 pmap->fuword32 = std_fuword32;
1784 pmap->fuword64 = std_fuword64;
1785 pmap->suword32 = std_suword32;
1786 pmap->suword64 = std_suword64;
1787 pmap->swapu32 = std_swapu32;
1788 pmap->swapu64 = std_swapu64;
1791 * Initialize pmap0/vmspace0.
1793 * On architectures where the kernel pmap is not integrated into the user
1794 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1795 * kernel_pmap should be used to directly access the kernel_pmap.
1798 pmap_pinit0(struct pmap *pmap)
1802 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1804 CPUMASK_ASSZERO(pmap->pm_active);
1805 pmap->pm_pvhint = NULL;
1806 RB_INIT(&pmap->pm_pvroot);
1807 spin_init(&pmap->pm_spin, "pmapinit0");
1808 for (i = 0; i < PM_PLACEMARKS; ++i)
1809 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1810 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1811 pmap_pinit_defaults(pmap);
1815 * Initialize a preallocated and zeroed pmap structure,
1816 * such as one in a vmspace structure.
1819 pmap_pinit_simple(struct pmap *pmap)
1824 * Misc initialization
1827 CPUMASK_ASSZERO(pmap->pm_active);
1828 pmap->pm_pvhint = NULL;
1829 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1831 pmap_pinit_defaults(pmap);
1834 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1837 if (pmap->pm_pmlpv == NULL) {
1838 RB_INIT(&pmap->pm_pvroot);
1839 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1840 spin_init(&pmap->pm_spin, "pmapinitsimple");
1841 for (i = 0; i < PM_PLACEMARKS; ++i)
1842 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1847 pmap_pinit(struct pmap *pmap)
1852 if (pmap->pm_pmlpv) {
1853 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1858 pmap_pinit_simple(pmap);
1859 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1862 * No need to allocate page table space yet but we do need a valid
1863 * page directory table.
1865 if (pmap->pm_pml4 == NULL) {
1867 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1873 * Allocate the page directory page, which wires it even though
1874 * it isn't being entered into some higher level page table (it
1875 * being the highest level). If one is already cached we don't
1876 * have to do anything.
1878 if ((pv = pmap->pm_pmlpv) == NULL) {
1879 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1880 pmap->pm_pmlpv = pv;
1881 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1882 VM_PAGE_TO_PHYS(pv->pv_m));
1886 * Install DMAP and KMAP.
1888 for (j = 0; j < NDMPML4E; ++j) {
1889 pmap->pm_pml4[DMPML4I + j] =
1890 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1891 pmap->pmap_bits[PG_RW_IDX] |
1892 pmap->pmap_bits[PG_V_IDX] |
1893 pmap->pmap_bits[PG_U_IDX];
1895 pmap->pm_pml4[KPML4I] = KPDPphys |
1896 pmap->pmap_bits[PG_RW_IDX] |
1897 pmap->pmap_bits[PG_V_IDX] |
1898 pmap->pmap_bits[PG_U_IDX];
1901 * install self-referential address mapping entry
1903 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1904 pmap->pmap_bits[PG_V_IDX] |
1905 pmap->pmap_bits[PG_RW_IDX] |
1906 pmap->pmap_bits[PG_A_IDX] |
1907 pmap->pmap_bits[PG_M_IDX];
1909 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1910 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1912 KKASSERT(pmap->pm_pml4[255] == 0);
1913 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1914 KKASSERT(pv->pv_entry.rbe_left == NULL);
1915 KKASSERT(pv->pv_entry.rbe_right == NULL);
1919 * Clean up a pmap structure so it can be physically freed. This routine
1920 * is called by the vmspace dtor function. A great deal of pmap data is
1921 * left passively mapped to improve vmspace management so we have a bit
1922 * of cleanup work to do here.
1925 pmap_puninit(pmap_t pmap)
1930 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1931 if ((pv = pmap->pm_pmlpv) != NULL) {
1932 if (pv_hold_try(pv) == 0)
1934 KKASSERT(pv == pmap->pm_pmlpv);
1935 p = pmap_remove_pv_page(pv);
1937 pv = NULL; /* safety */
1938 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1939 vm_page_busy_wait(p, FALSE, "pgpun");
1940 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1941 vm_page_unwire(p, 0);
1942 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1945 * XXX eventually clean out PML4 static entries and
1946 * use vm_page_free_zero()
1949 pmap->pm_pmlpv = NULL;
1951 if (pmap->pm_pml4) {
1952 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1953 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1954 pmap->pm_pml4 = NULL;
1956 KKASSERT(pmap->pm_stats.resident_count == 0);
1957 KKASSERT(pmap->pm_stats.wired_count == 0);
1961 * This function is now unused (used to add the pmap to the pmap_list)
1964 pmap_pinit2(struct pmap *pmap)
1969 * This routine is called when various levels in the page table need to
1970 * be populated. This routine cannot fail.
1972 * This function returns two locked pv_entry's, one representing the
1973 * requested pv and one representing the requested pv's parent pv. If
1974 * an intermediate page table does not exist it will be created, mapped,
1975 * wired, and the parent page table will be given an additional hold
1976 * count representing the presence of the child pv_entry.
1980 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1986 vm_pindex_t pt_pindex;
1992 * If the pv already exists and we aren't being asked for the
1993 * parent page table page we can just return it. A locked+held pv
1994 * is returned. The pv will also have a second hold related to the
1995 * pmap association that we don't have to worry about.
1998 pv = pv_alloc(pmap, ptepindex, &isnew);
1999 if (isnew == 0 && pvpp == NULL)
2003 * Special case terminal PVs. These are not page table pages so
2004 * no vm_page is allocated (the caller supplied the vm_page). If
2005 * pvpp is non-NULL we are being asked to also removed the pt_pv
2008 * Note that pt_pv's are only returned for user VAs. We assert that
2009 * a pt_pv is not being requested for kernel VAs. The kernel
2010 * pre-wires all higher-level page tables so don't overload managed
2011 * higher-level page tables on top of it!
2013 if (ptepindex < pmap_pt_pindex(0)) {
2014 if (ptepindex >= NUPTE_USER) {
2015 /* kernel manages this manually for KVM */
2016 KKASSERT(pvpp == NULL);
2018 KKASSERT(pvpp != NULL);
2019 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
2020 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
2022 vm_page_wire_quick(pvp->pv_m);
2029 * The kernel never uses managed PT/PD/PDP pages.
2031 KKASSERT(pmap != &kernel_pmap);
2034 * Non-terminal PVs allocate a VM page to represent the page table,
2035 * so we have to resolve pvp and calculate ptepindex for the pvp
2036 * and then for the page table entry index in the pvp for
2039 if (ptepindex < pmap_pd_pindex(0)) {
2041 * pv is PT, pvp is PD
2043 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2044 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2045 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2050 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2051 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2053 } else if (ptepindex < pmap_pdp_pindex(0)) {
2055 * pv is PD, pvp is PDP
2057 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2060 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2061 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2063 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2064 KKASSERT(pvpp == NULL);
2067 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2073 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2074 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2075 } else if (ptepindex < pmap_pml4_pindex()) {
2077 * pv is PDP, pvp is the root pml4 table
2079 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2084 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2085 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2088 * pv represents the top-level PML4, there is no parent.
2097 * (isnew) is TRUE, pv is not terminal.
2099 * (1) Add a wire count to the parent page table (pvp).
2100 * (2) Allocate a VM page for the page table.
2101 * (3) Enter the VM page into the parent page table.
2103 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2106 vm_page_wire_quick(pvp->pv_m);
2109 m = vm_page_alloc(NULL, pv->pv_pindex,
2110 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2111 VM_ALLOC_INTERRUPT);
2116 vm_page_wire(m); /* wire for mapping in parent */
2117 vm_page_unmanage(m); /* m must be spinunlocked */
2118 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2119 m->valid = VM_PAGE_BITS_ALL;
2121 vm_page_spin_lock(m);
2122 pmap_page_stats_adding(m);
2123 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2125 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2126 vm_page_spin_unlock(m);
2129 * (isnew) is TRUE, pv is not terminal.
2131 * Wire the page into pvp. Bump the resident_count for the pmap.
2132 * There is no pvp for the top level, address the pm_pml4[] array
2135 * If the caller wants the parent we return it, otherwise
2136 * we just put it away.
2138 * No interlock is needed for pte 0 -> non-zero.
2140 * In the situation where *ptep is valid we might have an unmanaged
2141 * page table page shared from another page table which we need to
2142 * unshare before installing our private page table page.
2145 v = VM_PAGE_TO_PHYS(m) |
2146 (pmap->pmap_bits[PG_U_IDX] |
2147 pmap->pmap_bits[PG_RW_IDX] |
2148 pmap->pmap_bits[PG_V_IDX] |
2149 pmap->pmap_bits[PG_A_IDX] |
2150 pmap->pmap_bits[PG_M_IDX]);
2151 ptep = pv_pte_lookup(pvp, ptepindex);
2152 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2156 panic("pmap_allocpte: unexpected pte %p/%d",
2157 pvp, (int)ptepindex);
2159 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2160 if (vm_page_unwire_quick(
2161 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2162 panic("pmap_allocpte: shared pgtable "
2163 "pg bad wirecount");
2168 pte = atomic_swap_long(ptep, v);
2170 kprintf("install pgtbl mixup 0x%016jx "
2171 "old/new 0x%016jx/0x%016jx\n",
2172 (intmax_t)ptepindex, pte, v);
2179 * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2183 KKASSERT(pvp->pv_m != NULL);
2184 ptep = pv_pte_lookup(pvp, ptepindex);
2185 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2186 (pmap->pmap_bits[PG_U_IDX] |
2187 pmap->pmap_bits[PG_RW_IDX] |
2188 pmap->pmap_bits[PG_V_IDX] |
2189 pmap->pmap_bits[PG_A_IDX] |
2190 pmap->pmap_bits[PG_M_IDX]);
2192 kprintf("mismatched upper level pt %016jx/%016jx\n",
2204 * This version of pmap_allocpte() checks for possible segment optimizations
2205 * that would allow page-table sharing. It can be called for terminal
2206 * page or page table page ptepindex's.
2208 * The function is called with page table page ptepindex's for fictitious
2209 * and unmanaged terminal pages. That is, we don't want to allocate a
2210 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2213 * This function can return a pv and *pvpp associated with the passed in pmap
2214 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2215 * an unmanaged page table page will be entered into the pass in pmap.
2219 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2220 vm_map_entry_t entry, vm_offset_t va)
2226 pv_entry_t pte_pv; /* in original or shared pmap */
2227 pv_entry_t pt_pv; /* in original or shared pmap */
2228 pv_entry_t proc_pd_pv; /* in original pmap */
2229 pv_entry_t proc_pt_pv; /* in original pmap */
2230 pv_entry_t xpv; /* PT in shared pmap */
2231 pd_entry_t *pt; /* PT entry in PD of original pmap */
2232 pd_entry_t opte; /* contents of *pt */
2233 pd_entry_t npte; /* contents of *pt */
2238 * Basic tests, require a non-NULL vm_map_entry, require proper
2239 * alignment and type for the vm_map_entry, require that the
2240 * underlying object already be allocated.
2242 * We allow almost any type of object to use this optimization.
2243 * The object itself does NOT have to be sized to a multiple of the
2244 * segment size, but the memory mapping does.
2246 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2247 * won't work as expected.
2249 if (entry == NULL ||
2250 pmap_mmu_optimize == 0 || /* not enabled */
2251 (pmap->pm_flags & PMAP_HVM) || /* special pmap */
2252 ptepindex >= pmap_pd_pindex(0) || /* not terminal or pt */
2253 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2254 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2255 entry->object.vm_object == NULL || /* needs VM object */
2256 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2257 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2258 (entry->offset & SEG_MASK) || /* must be aligned */
2259 (entry->start & SEG_MASK)) {
2260 return(pmap_allocpte(pmap, ptepindex, pvpp));
2264 * Make sure the full segment can be represented.
2266 b = va & ~(vm_offset_t)SEG_MASK;
2267 if (b < entry->start || b + SEG_SIZE > entry->end)
2268 return(pmap_allocpte(pmap, ptepindex, pvpp));
2271 * If the full segment can be represented dive the VM object's
2272 * shared pmap, allocating as required.
2274 object = entry->object.vm_object;
2276 if (entry->protection & VM_PROT_WRITE)
2277 obpmapp = &object->md.pmap_rw;
2279 obpmapp = &object->md.pmap_ro;
2282 if (pmap_enter_debug > 0) {
2284 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2286 va, entry->protection, object,
2288 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2289 entry, entry->start, entry->end);
2294 * We allocate what appears to be a normal pmap but because portions
2295 * of this pmap are shared with other unrelated pmaps we have to
2296 * set pm_active to point to all cpus.
2298 * XXX Currently using pmap_spin to interlock the update, can't use
2299 * vm_object_hold/drop because the token might already be held
2300 * shared OR exclusive and we don't know.
2302 while ((obpmap = *obpmapp) == NULL) {
2303 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2304 pmap_pinit_simple(obpmap);
2305 pmap_pinit2(obpmap);
2306 spin_lock(&pmap_spin);
2307 if (*obpmapp != NULL) {
2311 spin_unlock(&pmap_spin);
2312 pmap_release(obpmap);
2313 pmap_puninit(obpmap);
2314 kfree(obpmap, M_OBJPMAP);
2315 obpmap = *obpmapp; /* safety */
2317 obpmap->pm_active = smp_active_mask;
2318 obpmap->pm_flags |= PMAP_SEGSHARED;
2320 spin_unlock(&pmap_spin);
2325 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2326 * pte/pt using the shared pmap from the object but also adjust
2327 * the process pmap's page table page as a side effect.
2331 * Resolve the terminal PTE and PT in the shared pmap. This is what
2332 * we will return. This is true if ptepindex represents a terminal
2333 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2337 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2338 if (ptepindex >= pmap_pt_pindex(0))
2344 * Resolve the PD in the process pmap so we can properly share the
2345 * page table page. Lock order is bottom-up (leaf first)!
2347 * NOTE: proc_pt_pv can be NULL.
2349 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), NULL);
2350 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2352 if (pmap_enter_debug > 0) {
2354 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2356 (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2363 * xpv is the page table page pv from the shared object
2364 * (for convenience), from above.
2366 * Calculate the pte value for the PT to load into the process PD.
2367 * If we have to change it we must properly dispose of the previous
2370 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2371 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2372 (pmap->pmap_bits[PG_U_IDX] |
2373 pmap->pmap_bits[PG_RW_IDX] |
2374 pmap->pmap_bits[PG_V_IDX] |
2375 pmap->pmap_bits[PG_A_IDX] |
2376 pmap->pmap_bits[PG_M_IDX]);
2379 * Dispose of previous page table page if it was local to the
2380 * process pmap. If the old pt is not empty we cannot dispose of it
2381 * until we clean it out. This case should not arise very often so
2382 * it is not optimized.
2385 pmap_inval_bulk_t bulk;
2387 if (proc_pt_pv->pv_m->wire_count != 1) {
2393 va & ~(vm_offset_t)SEG_MASK,
2394 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2399 * The release call will indirectly clean out *pt
2401 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2402 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2403 pmap_inval_bulk_flush(&bulk);
2406 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2410 * Handle remaining cases.
2413 atomic_swap_long(pt, npte);
2414 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2415 vm_page_wire_quick(proc_pd_pv->pv_m); /* proc pd for sh pt */
2416 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2417 } else if (*pt != npte) {
2418 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2421 opte = pte_load_clear(pt);
2422 KKASSERT(opte && opte != npte);
2426 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2429 * Clean up opte, bump the wire_count for the process
2430 * PD page representing the new entry if it was
2433 * If the entry was not previously empty and we have
2434 * a PT in the proc pmap then opte must match that
2435 * pt. The proc pt must be retired (this is done
2436 * later on in this procedure).
2438 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2441 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2442 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2443 if (vm_page_unwire_quick(m)) {
2444 panic("pmap_allocpte_seg: "
2445 "bad wire count %p",
2451 * The existing process page table was replaced and must be destroyed
2465 * Release any resources held by the given physical map.
2467 * Called when a pmap initialized by pmap_pinit is being released. Should
2468 * only be called if the map contains no valid mappings.
2470 struct pmap_release_info {
2476 static int pmap_release_callback(pv_entry_t pv, void *data);
2479 pmap_release(struct pmap *pmap)
2481 struct pmap_release_info info;
2483 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2484 ("pmap still active! %016jx",
2485 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2488 * There is no longer a pmap_list, if there were we would remove the
2489 * pmap from it here.
2493 * Pull pv's off the RB tree in order from low to high and release
2501 spin_lock(&pmap->pm_spin);
2502 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2503 pmap_release_callback, &info);
2504 spin_unlock(&pmap->pm_spin);
2508 } while (info.retry);
2512 * One resident page (the pml4 page) should remain.
2513 * No wired pages should remain.
2516 if (pmap->pm_stats.resident_count !=
2517 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1) ||
2518 pmap->pm_stats.wired_count != 0) {
2519 kprintf("fatal pmap problem - pmap %p flags %08x "
2520 "rescnt=%jd wirecnt=%jd\n",
2523 pmap->pm_stats.resident_count,
2524 pmap->pm_stats.wired_count);
2525 tsleep(pmap, 0, "DEAD", 0);
2528 KKASSERT(pmap->pm_stats.resident_count ==
2529 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2530 KKASSERT(pmap->pm_stats.wired_count == 0);
2535 * Called from low to high. We must cache the proper parent pv so we
2536 * can adjust its wired count.
2539 pmap_release_callback(pv_entry_t pv, void *data)
2541 struct pmap_release_info *info = data;
2542 pmap_t pmap = info->pmap;
2547 * Acquire a held and locked pv, check for release race
2549 pindex = pv->pv_pindex;
2550 if (info->pvp == pv) {
2551 spin_unlock(&pmap->pm_spin);
2553 } else if (pv_hold_try(pv)) {
2554 spin_unlock(&pmap->pm_spin);
2556 spin_unlock(&pmap->pm_spin);
2559 if (pv->pv_pmap != pmap || pindex != pv->pv_pindex) {
2561 spin_lock(&pmap->pm_spin);
2566 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2568 * I am PTE, parent is PT
2570 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2571 pindex += NUPTE_TOTAL;
2572 } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2574 * I am PT, parent is PD
2576 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2577 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2578 } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2580 * I am PD, parent is PDP
2582 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2584 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2585 } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2587 * I am PDP, parent is PML4 (there's only one)
2590 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2591 NUPD_TOTAL) >> NPML4EPGSHIFT;
2592 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2594 pindex = pmap_pml4_pindex();
2606 if (info->pvp && info->pvp->pv_pindex != pindex) {
2610 if (info->pvp == NULL)
2611 info->pvp = pv_get(pmap, pindex, NULL);
2618 r = pmap_release_pv(pv, info->pvp, NULL);
2619 spin_lock(&pmap->pm_spin);
2625 * Called with held (i.e. also locked) pv. This function will dispose of
2626 * the lock along with the pv.
2628 * If the caller already holds the locked parent page table for pv it
2629 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2630 * pass NULL for pvp.
2633 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2638 * The pmap is currently not spinlocked, pv is held+locked.
2639 * Remove the pv's page from its parent's page table. The
2640 * parent's page table page's wire_count will be decremented.
2642 * This will clean out the pte at any level of the page table.
2643 * If smp != 0 all cpus are affected.
2645 * Do not tear-down recursively, its faster to just let the
2646 * release run its course.
2648 pmap_remove_pv_pte(pv, pvp, bulk, 0);
2651 * Terminal pvs are unhooked from their vm_pages. Because
2652 * terminal pages aren't page table pages they aren't wired
2653 * by us, so we have to be sure not to unwire them either.
2655 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2656 pmap_remove_pv_page(pv);
2661 * We leave the top-level page table page cached, wired, and
2662 * mapped in the pmap until the dtor function (pmap_puninit())
2665 * Since we are leaving the top-level pv intact we need
2666 * to break out of what would otherwise be an infinite loop.
2668 if (pv->pv_pindex == pmap_pml4_pindex()) {
2674 * For page table pages (other than the top-level page),
2675 * remove and free the vm_page. The representitive mapping
2676 * removed above by pmap_remove_pv_pte() did not undo the
2677 * last wire_count so we have to do that as well.
2679 p = pmap_remove_pv_page(pv);
2680 vm_page_busy_wait(p, FALSE, "pmaprl");
2681 if (p->wire_count != 1) {
2682 kprintf("p->wire_count was %016lx %d\n",
2683 pv->pv_pindex, p->wire_count);
2685 KKASSERT(p->wire_count == 1);
2686 KKASSERT(p->flags & PG_UNMANAGED);
2688 vm_page_unwire(p, 0);
2689 KKASSERT(p->wire_count == 0);
2699 * This function will remove the pte associated with a pv from its parent.
2700 * Terminal pv's are supported. All cpus specified by (bulk) are properly
2703 * The wire count will be dropped on the parent page table. The wire
2704 * count on the page being removed (pv->pv_m) from the parent page table
2705 * is NOT touched. Note that terminal pages will not have any additional
2706 * wire counts while page table pages will have at least one representing
2707 * the mapping, plus others representing sub-mappings.
2709 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2710 * pages and user page table and terminal pages.
2712 * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2713 * be freshly allocated and not imply that the pte is managed. In this
2714 * case pv->pv_m should be NULL.
2716 * The pv must be locked. The pvp, if supplied, must be locked. All
2717 * supplied pv's will remain locked on return.
2719 * XXX must lock parent pv's if they exist to remove pte XXX
2723 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2726 vm_pindex_t ptepindex = pv->pv_pindex;
2727 pmap_t pmap = pv->pv_pmap;
2733 if (ptepindex == pmap_pml4_pindex()) {
2735 * We are the top level PML4E table, there is no parent.
2737 p = pmap->pm_pmlpv->pv_m;
2738 KKASSERT(pv->pv_m == p); /* debugging */
2739 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2741 * Remove a PDP page from the PML4E. This can only occur
2742 * with user page tables. We do not have to lock the
2743 * pml4 PV so just ignore pvp.
2745 vm_pindex_t pml4_pindex;
2746 vm_pindex_t pdp_index;
2749 pdp_index = ptepindex - pmap_pdp_pindex(0);
2751 pml4_pindex = pmap_pml4_pindex();
2752 pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2757 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2758 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2759 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2760 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2761 KKASSERT(pv->pv_m == p); /* debugging */
2762 } else if (ptepindex >= pmap_pd_pindex(0)) {
2764 * Remove a PD page from the PDP
2766 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2767 * of a simple pmap because it stops at
2770 vm_pindex_t pdp_pindex;
2771 vm_pindex_t pd_index;
2774 pd_index = ptepindex - pmap_pd_pindex(0);
2777 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2778 (pd_index >> NPML4EPGSHIFT);
2779 pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2784 pd = pv_pte_lookup(pvp, pd_index &
2785 ((1ul << NPDPEPGSHIFT) - 1));
2786 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2787 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2788 pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2790 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2791 p = pv->pv_m; /* degenerate test later */
2793 KKASSERT(pv->pv_m == p); /* debugging */
2794 } else if (ptepindex >= pmap_pt_pindex(0)) {
2796 * Remove a PT page from the PD
2798 vm_pindex_t pd_pindex;
2799 vm_pindex_t pt_index;
2802 pt_index = ptepindex - pmap_pt_pindex(0);
2805 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2806 (pt_index >> NPDPEPGSHIFT);
2807 pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2812 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2814 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2815 ("*pt unexpectedly invalid %016jx "
2816 "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2817 *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2818 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2820 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2821 kprintf("*pt unexpectedly invalid %016jx "
2822 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2824 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2825 tsleep(pt, 0, "DEAD", 0);
2828 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2831 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2832 KKASSERT(pv->pv_m == p); /* debugging */
2835 * Remove a PTE from the PT page. The PV might exist even if
2836 * the PTE is not managed, in whichcase pv->pv_m should be
2839 * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2840 * table pages but the kernel_pmap does not.
2842 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2843 * pv is a pte_pv so we can safely lock pt_pv.
2845 * NOTE: FICTITIOUS pages may have multiple physical mappings
2846 * so PHYS_TO_VM_PAGE() will not necessarily work for
2849 vm_pindex_t pt_pindex;
2854 pt_pindex = ptepindex >> NPTEPGSHIFT;
2855 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2857 if (ptepindex >= NUPTE_USER) {
2858 ptep = vtopte(ptepindex << PAGE_SHIFT);
2859 KKASSERT(pvp == NULL);
2860 /* pvp remains NULL */
2863 pt_pindex = NUPTE_TOTAL +
2864 (ptepindex >> NPDPEPGSHIFT);
2865 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
2869 ptep = pv_pte_lookup(pvp, ptepindex &
2870 ((1ul << NPDPEPGSHIFT) - 1));
2872 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2873 if (bulk == NULL) /* XXX */
2874 cpu_invlpg((void *)va); /* XXX */
2877 * Now update the vm_page_t
2879 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
2880 (pte & pmap->pmap_bits[PG_V_IDX])) {
2882 * Valid managed page, adjust (p).
2884 if (pte & pmap->pmap_bits[PG_DEVICE_IDX]) {
2887 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2888 KKASSERT(pv->pv_m == p);
2890 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2891 if (pmap_track_modified(ptepindex))
2894 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2895 vm_page_flag_set(p, PG_REFERENCED);
2899 * Unmanaged page, do not try to adjust the vm_page_t.
2900 * pv could be freshly allocated for a pmap_enter(),
2901 * replacing an unmanaged page with a managed one.
2903 * pv->pv_m might reflect the new page and not the
2906 * We could extract p from the physical address and
2907 * adjust it but we explicitly do not for unmanaged
2912 if (pte & pmap->pmap_bits[PG_W_IDX])
2913 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2914 if (pte & pmap->pmap_bits[PG_G_IDX])
2915 cpu_invlpg((void *)va);
2919 * If requested, scrap the underlying pv->pv_m and the underlying
2920 * pv. If this is a page-table-page we must also free the page.
2922 * pvp must be returned locked.
2926 * page table page (PT, PD, PDP, PML4), caller was responsible
2927 * for testing wired_count.
2929 KKASSERT(pv->pv_m->wire_count == 1);
2930 p = pmap_remove_pv_page(pv);
2934 vm_page_busy_wait(p, FALSE, "pgpun");
2935 vm_page_unwire(p, 0);
2936 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2938 } else if (destroy == 2) {
2940 * Normal page, remove from pmap and leave the underlying
2943 pmap_remove_pv_page(pv);
2945 pv = NULL; /* safety */
2949 * If we acquired pvp ourselves then we are responsible for
2950 * recursively deleting it.
2952 if (pvp && gotpvp) {
2954 * Recursively destroy higher-level page tables.
2956 * This is optional. If we do not, they will still
2957 * be destroyed when the process exits.
2959 * NOTE: Do not destroy pv_entry's with extra hold refs,
2960 * a caller may have unlocked it and intends to
2961 * continue to use it.
2963 if (pmap_dynamic_delete &&
2965 pvp->pv_m->wire_count == 1 &&
2966 (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
2967 pvp->pv_pindex != pmap_pml4_pindex()) {
2968 if (pmap_dynamic_delete == 2)
2969 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
2970 if (pmap != &kernel_pmap) {
2971 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2972 pvp = NULL; /* safety */
2974 kprintf("Attempt to remove kernel_pmap pindex "
2975 "%jd\n", pvp->pv_pindex);
2985 * Remove the vm_page association to a pv. The pv must be locked.
2989 pmap_remove_pv_page(pv_entry_t pv)
2994 vm_page_spin_lock(m);
2995 KKASSERT(m && m == pv->pv_m);
2997 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2998 pmap_page_stats_deleting(m);
2999 if (TAILQ_EMPTY(&m->md.pv_list))
3000 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3001 vm_page_spin_unlock(m);
3007 * Grow the number of kernel page table entries, if needed.
3009 * This routine is always called to validate any address space
3010 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
3011 * space below KERNBASE.
3013 * kernel_map must be locked exclusively by the caller.
3016 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3019 vm_offset_t ptppaddr;
3021 pd_entry_t *pt, newpt;
3023 int update_kernel_vm_end;
3026 * bootstrap kernel_vm_end on first real VM use
3028 if (kernel_vm_end == 0) {
3029 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3031 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3032 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3033 ~(PAGE_SIZE * NPTEPG - 1);
3035 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3036 kernel_vm_end = kernel_map.max_offset;
3043 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
3044 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
3045 * do not want to force-fill 128G worth of page tables.
3047 if (kstart < KERNBASE) {
3048 if (kstart > kernel_vm_end)
3049 kstart = kernel_vm_end;
3050 KKASSERT(kend <= KERNBASE);
3051 update_kernel_vm_end = 1;
3053 update_kernel_vm_end = 0;
3056 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
3057 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
3059 if (kend - 1 >= kernel_map.max_offset)
3060 kend = kernel_map.max_offset;
3062 while (kstart < kend) {
3063 pt = pmap_pt(&kernel_pmap, kstart);
3065 /* We need a new PD entry */
3066 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3069 VM_ALLOC_INTERRUPT);
3071 panic("pmap_growkernel: no memory to grow "
3074 paddr = VM_PAGE_TO_PHYS(nkpg);
3075 pmap_zero_page(paddr);
3076 newpd = (pdp_entry_t)
3078 kernel_pmap.pmap_bits[PG_V_IDX] |
3079 kernel_pmap.pmap_bits[PG_RW_IDX] |
3080 kernel_pmap.pmap_bits[PG_A_IDX] |
3081 kernel_pmap.pmap_bits[PG_M_IDX]);
3082 *pmap_pd(&kernel_pmap, kstart) = newpd;
3083 continue; /* try again */
3085 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3086 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3087 ~(PAGE_SIZE * NPTEPG - 1);
3088 if (kstart - 1 >= kernel_map.max_offset) {
3089 kstart = kernel_map.max_offset;
3098 * This index is bogus, but out of the way
3100 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3103 VM_ALLOC_INTERRUPT);
3105 panic("pmap_growkernel: no memory to grow kernel");
3108 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3109 pmap_zero_page(ptppaddr);
3110 newpt = (pd_entry_t)(ptppaddr |
3111 kernel_pmap.pmap_bits[PG_V_IDX] |
3112 kernel_pmap.pmap_bits[PG_RW_IDX] |
3113 kernel_pmap.pmap_bits[PG_A_IDX] |
3114 kernel_pmap.pmap_bits[PG_M_IDX]);
3115 atomic_swap_long(pmap_pt(&kernel_pmap, kstart), newpt);
3117 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3118 ~(PAGE_SIZE * NPTEPG - 1);
3120 if (kstart - 1 >= kernel_map.max_offset) {
3121 kstart = kernel_map.max_offset;
3127 * Only update kernel_vm_end for areas below KERNBASE.
3129 if (update_kernel_vm_end && kernel_vm_end < kstart)
3130 kernel_vm_end = kstart;
3134 * Add a reference to the specified pmap.
3137 pmap_reference(pmap_t pmap)
3140 atomic_add_int(&pmap->pm_count, 1);
3143 /***************************************************
3144 * page management routines.
3145 ***************************************************/
3148 * Hold a pv without locking it
3151 pv_hold(pv_entry_t pv)
3153 atomic_add_int(&pv->pv_hold, 1);
3157 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
3158 * was successfully locked, FALSE if it wasn't. The caller must dispose of
3161 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3162 * pv list via its page) must be held by the caller.
3165 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3170 * Critical path shortcut expects pv to already have one ref
3171 * (for the pv->pv_pmap).
3173 if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3176 pv->pv_line = lineno;
3182 count = pv->pv_hold;
3184 if ((count & PV_HOLD_LOCKED) == 0) {
3185 if (atomic_cmpset_int(&pv->pv_hold, count,
3186 (count + 1) | PV_HOLD_LOCKED)) {
3189 pv->pv_line = lineno;
3194 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3202 * Drop a previously held pv_entry which could not be locked, allowing its
3205 * Must not be called with a spinlock held as we might zfree() the pv if it
3206 * is no longer associated with a pmap and this was the last hold count.
3209 pv_drop(pv_entry_t pv)
3214 count = pv->pv_hold;
3216 KKASSERT((count & PV_HOLD_MASK) > 0);
3217 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3218 (PV_HOLD_LOCKED | 1));
3219 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3220 if ((count & PV_HOLD_MASK) == 1) {
3222 if (pmap_enter_debug > 0) {
3224 kprintf("pv_drop: free pv %p\n", pv);
3227 KKASSERT(count == 1);
3228 KKASSERT(pv->pv_pmap == NULL);
3238 * Find or allocate the requested PV entry, returning a locked, held pv.
3240 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3241 * for the caller and one representing the pmap and vm_page association.
3243 * If (*isnew) is zero, the returned pv will have only one hold count.
3245 * Since both associations can only be adjusted while the pv is locked,
3246 * together they represent just one additional hold.
3250 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3253 pv_entry_t pnew = NULL;
3255 spin_lock(&pmap->pm_spin);
3260 pv = pmap->pm_pvhint;
3263 pv->pv_pmap != pmap ||
3264 pv->pv_pindex != pindex) {
3265 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3272 * We need to stage a new pv entry
3275 spin_unlock(&pmap->pm_spin);
3276 pnew = zalloc(pvzone);
3277 spin_lock(&pmap->pm_spin);
3282 * We need to block if someone is holding a
3285 pmark = pmap_placemarker_hash(pmap, pindex);
3287 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3288 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3289 ssleep(pmark, &pmap->pm_spin, 0, "pvplc", 0);
3294 * Setup the new entry
3296 pnew->pv_pmap = pmap;
3297 pnew->pv_pindex = pindex;
3298 pnew->pv_hold = PV_HOLD_LOCKED | 2;
3300 pnew->pv_func = func;
3301 pnew->pv_line = lineno;
3303 pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3304 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3305 spin_unlock(&pmap->pm_spin);
3308 KKASSERT(pv == NULL);
3314 * We have an entry, clean up any staged pv we had allocated,
3315 * then block until we can lock the entry.
3318 spin_unlock(&pmap->pm_spin);
3319 zfree(pvzone, pnew);
3321 spin_lock(&pmap->pm_spin);
3324 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3325 spin_unlock(&pmap->pm_spin);
3327 spin_unlock(&pmap->pm_spin);
3328 _pv_lock(pv PMAP_DEBUG_COPY);
3332 * Make sure the pv is still in good shape for return,
3335 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3340 spin_lock(&pmap->pm_spin);
3345 * Find the requested PV entry, returning a locked+held pv or NULL
3349 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3353 spin_lock(&pmap->pm_spin);
3358 pv = pmap->pm_pvhint;
3361 pv->pv_pmap != pmap ||
3362 pv->pv_pindex != pindex) {
3363 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3366 while (pv == NULL) {
3368 * Block if there is a placemarker. If we are to
3369 * return it, we must also aquire the spot.
3373 pmark = pmap_placemarker_hash(pmap, pindex);
3375 if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3376 ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3377 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3378 ssleep(pmark, &pmap->pm_spin, 0, "pvpld", 0);
3382 if (atomic_swap_long(pmark, pindex) !=
3384 panic("_pv_get: pmark race");
3388 spin_unlock(&pmap->pm_spin);
3391 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3392 spin_unlock(&pmap->pm_spin);
3394 spin_unlock(&pmap->pm_spin);
3395 _pv_lock(pv PMAP_DEBUG_COPY);
3397 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3398 pv_cache(pv, pindex);
3402 spin_lock(&pmap->pm_spin);
3407 * Lookup, hold, and attempt to lock (pmap,pindex).
3409 * If the entry does not exist NULL is returned and *errorp is set to 0
3411 * If the entry exists and could be successfully locked it is returned and
3412 * errorp is set to 0.
3414 * If the entry exists but could NOT be successfully locked it is returned
3415 * held and *errorp is set to 1.
3417 * If the entry is placemarked by someone else NULL is returned and *errorp
3422 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3426 spin_lock_shared(&pmap->pm_spin);
3428 pv = pmap->pm_pvhint;
3431 pv->pv_pmap != pmap ||
3432 pv->pv_pindex != pindex) {
3433 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3439 pmark = pmap_placemarker_hash(pmap, pindex);
3441 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3443 } else if (pmarkp &&
3444 atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3448 * Can't set a placemark with a NULL pmarkp, or if
3449 * pmarkp is non-NULL but we failed to set our
3456 spin_unlock_shared(&pmap->pm_spin);
3460 if (pv_hold_try(pv)) {
3461 pv_cache(pv, pindex);
3462 spin_unlock_shared(&pmap->pm_spin);
3464 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3465 return(pv); /* lock succeeded */
3467 spin_unlock_shared(&pmap->pm_spin);
3469 return (pv); /* lock failed */
3473 * Lock a held pv, keeping the hold count
3477 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3482 count = pv->pv_hold;
3484 if ((count & PV_HOLD_LOCKED) == 0) {
3485 if (atomic_cmpset_int(&pv->pv_hold, count,
3486 count | PV_HOLD_LOCKED)) {
3489 pv->pv_line = lineno;
3495 tsleep_interlock(pv, 0);
3496 if (atomic_cmpset_int(&pv->pv_hold, count,
3497 count | PV_HOLD_WAITING)) {
3499 kprintf("pv waiting on %s:%d\n",
3500 pv->pv_func, pv->pv_line);
3502 tsleep(pv, PINTERLOCKED, "pvwait", hz);
3509 * Unlock a held and locked pv, keeping the hold count.
3513 pv_unlock(pv_entry_t pv)
3518 count = pv->pv_hold;
3520 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3521 (PV_HOLD_LOCKED | 1));
3522 if (atomic_cmpset_int(&pv->pv_hold, count,
3524 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3525 if (count & PV_HOLD_WAITING)
3533 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3534 * and the hold count drops to zero we will free it.
3536 * Caller should not hold any spin locks. We are protected from hold races
3537 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3538 * lock held. A pv cannot be located otherwise.
3542 pv_put(pv_entry_t pv)
3545 if (pmap_enter_debug > 0) {
3547 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3552 * Fast - shortcut most common condition
3554 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3565 * Remove the pmap association from a pv, require that pv_m already be removed,
3566 * then unlock and drop the pv. Any pte operations must have already been
3567 * completed. This call may result in a last-drop which will physically free
3570 * Removing the pmap association entails an additional drop.
3572 * pv must be exclusively locked on call and will be disposed of on return.
3576 pv_free(pv_entry_t pv, pv_entry_t pvp)
3580 KKASSERT(pv->pv_m == NULL);
3581 KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3582 if ((pmap = pv->pv_pmap) != NULL) {
3583 spin_lock(&pmap->pm_spin);
3584 KKASSERT(pv->pv_pmap == pmap);
3585 if (pmap->pm_pvhint == pv)
3586 pmap->pm_pvhint = NULL;
3587 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3588 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3591 spin_unlock(&pmap->pm_spin);
3594 * Try to shortcut three atomic ops, otherwise fall through
3595 * and do it normally. Drop two refs and the lock all in
3598 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3600 if (pmap_enter_debug > 0) {
3602 kprintf("pv_free: free pv %p\n", pv);
3607 vm_page_unwire_quick(pvp->pv_m);
3610 pv_drop(pv); /* ref for pv_pmap */
3612 vm_page_unwire_quick(pvp->pv_m);
3618 * This routine is very drastic, but can save the system
3626 static int warningdone=0;
3628 if (pmap_pagedaemon_waken == 0)
3630 pmap_pagedaemon_waken = 0;
3631 if (warningdone < 5) {
3632 kprintf("pmap_collect: collecting pv entries -- "
3633 "suggest increasing PMAP_SHPGPERPROC\n");
3637 for (i = 0; i < vm_page_array_size; i++) {
3638 m = &vm_page_array[i];
3639 if (m->wire_count || m->hold_count)
3641 if (vm_page_busy_try(m, TRUE) == 0) {
3642 if (m->wire_count == 0 && m->hold_count == 0) {
3651 * Scan the pmap for active page table entries and issue a callback.
3652 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3653 * its parent page table.
3655 * pte_pv will be NULL if the page or page table is unmanaged.
3656 * pt_pv will point to the page table page containing the pte for the page.
3658 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3659 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3660 * process pmap's PD and page to the callback function. This can be
3661 * confusing because the pt_pv is really a pd_pv, and the target page
3662 * table page is simply aliased by the pmap and not owned by it.
3664 * It is assumed that the start and end are properly rounded to the page size.
3666 * It is assumed that PD pages and above are managed and thus in the RB tree,
3667 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3669 struct pmap_scan_info {
3673 vm_pindex_t sva_pd_pindex;
3674 vm_pindex_t eva_pd_pindex;
3675 void (*func)(pmap_t, struct pmap_scan_info *,
3676 pv_entry_t, vm_pindex_t *, pv_entry_t,
3678 pt_entry_t *, void *);
3680 pmap_inval_bulk_t bulk_core;
3681 pmap_inval_bulk_t *bulk;
3686 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3687 static int pmap_scan_callback(pv_entry_t pv, void *data);
3690 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3692 struct pmap *pmap = info->pmap;
3693 pv_entry_t pd_pv; /* A page directory PV */
3694 pv_entry_t pt_pv; /* A page table PV */
3695 pv_entry_t pte_pv; /* A page table entry PV */
3696 vm_pindex_t *pte_placemark;
3697 vm_pindex_t *pt_placemark;
3700 struct pv_entry dummy_pv;
3706 info->bulk = &info->bulk_core;
3707 pmap_inval_bulk_init(&info->bulk_core, pmap);
3713 * Hold the token for stability; if the pmap is empty we have nothing
3717 if (pmap->pm_stats.resident_count == 0) {
3725 * Special handling for scanning one page, which is a very common
3726 * operation (it is?).
3728 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3730 if (info->sva + PAGE_SIZE == info->eva) {
3731 if (info->sva >= VM_MAX_USER_ADDRESS) {
3733 * Kernel mappings do not track wire counts on
3734 * page table pages and only maintain pd_pv and
3735 * pte_pv levels so pmap_scan() works.
3738 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3740 ptep = vtopte(info->sva);
3743 * User pages which are unmanaged will not have a
3744 * pte_pv. User page table pages which are unmanaged
3745 * (shared from elsewhere) will also not have a pt_pv.
3746 * The func() callback will pass both pte_pv and pt_pv
3747 * as NULL in that case.
3749 * We hold pte_placemark across the operation for
3752 * WARNING! We must hold pt_placemark across the
3753 * *ptep test to prevent misintepreting
3754 * a non-zero *ptep as a shared page
3755 * table page. Hold it across the function
3756 * callback as well for SMP safety.
3758 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
3760 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
3762 if (pt_pv == NULL) {
3763 KKASSERT(pte_pv == NULL);
3764 pd_pv = pv_get(pmap,
3765 pmap_pd_pindex(info->sva),
3768 ptep = pv_pte_lookup(pd_pv,
3769 pmap_pt_index(info->sva));
3771 info->func(pmap, info,
3777 pv_placemarker_wakeup(pmap,
3782 pv_placemarker_wakeup(pmap,
3785 pv_placemarker_wakeup(pmap, pte_placemark);
3788 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3792 * NOTE: *ptep can't be ripped out from under us if we hold
3793 * pte_pv (or pte_placemark) locked, but bits can
3799 KKASSERT(pte_pv == NULL);
3800 pv_placemarker_wakeup(pmap, pte_placemark);
3801 } else if (pte_pv) {
3802 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3803 pmap->pmap_bits[PG_V_IDX])) ==
3804 (pmap->pmap_bits[PG_MANAGED_IDX] |
3805 pmap->pmap_bits[PG_V_IDX]),
3806 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
3807 *ptep, oldpte, info->sva, pte_pv));
3808 info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
3809 info->sva, ptep, info->arg);
3811 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3812 pmap->pmap_bits[PG_V_IDX])) ==
3813 pmap->pmap_bits[PG_V_IDX],
3814 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
3815 *ptep, oldpte, info->sva));
3816 info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
3817 info->sva, ptep, info->arg);
3822 pmap_inval_bulk_flush(info->bulk);
3827 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3830 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3831 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3833 if (info->sva >= VM_MAX_USER_ADDRESS) {
3835 * The kernel does not currently maintain any pv_entry's for
3836 * higher-level page tables.
3838 bzero(&dummy_pv, sizeof(dummy_pv));
3839 dummy_pv.pv_pindex = info->sva_pd_pindex;
3840 spin_lock(&pmap->pm_spin);
3841 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3842 pmap_scan_callback(&dummy_pv, info);
3843 ++dummy_pv.pv_pindex;
3845 spin_unlock(&pmap->pm_spin);
3848 * User page tables maintain local PML4, PDP, and PD
3849 * pv_entry's at the very least. PT pv's might be
3850 * unmanaged and thus not exist. PTE pv's might be
3851 * unmanaged and thus not exist.
3853 spin_lock(&pmap->pm_spin);
3854 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
3855 pmap_scan_callback, info);
3856 spin_unlock(&pmap->pm_spin);
3858 pmap_inval_bulk_flush(info->bulk);
3862 * WARNING! pmap->pm_spin held
3865 pmap_scan_cmp(pv_entry_t pv, void *data)
3867 struct pmap_scan_info *info = data;
3868 if (pv->pv_pindex < info->sva_pd_pindex)
3870 if (pv->pv_pindex >= info->eva_pd_pindex)
3876 * pmap_scan() by PDs
3878 * WARNING! pmap->pm_spin held
3881 pmap_scan_callback(pv_entry_t pv, void *data)
3883 struct pmap_scan_info *info = data;
3884 struct pmap *pmap = info->pmap;
3885 pv_entry_t pd_pv; /* A page directory PV */
3886 pv_entry_t pt_pv; /* A page table PV */
3887 vm_pindex_t *pt_placemark;
3892 vm_offset_t va_next;
3893 vm_pindex_t pd_pindex;
3903 * Pull the PD pindex from the pv before releasing the spinlock.
3905 * WARNING: pv is faked for kernel pmap scans.
3907 pd_pindex = pv->pv_pindex;
3908 spin_unlock(&pmap->pm_spin);
3909 pv = NULL; /* invalid after spinlock unlocked */
3912 * Calculate the page range within the PD. SIMPLE pmaps are
3913 * direct-mapped for the entire 2^64 address space. Normal pmaps
3914 * reflect the user and kernel address space which requires
3915 * cannonicalization w/regards to converting pd_pindex's back
3918 sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
3919 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3920 (sva & PML4_SIGNMASK)) {
3921 sva |= PML4_SIGNMASK;
3923 eva = sva + NBPDP; /* can overflow */
3924 if (sva < info->sva)
3926 if (eva < info->sva || eva > info->eva)
3930 * NOTE: kernel mappings do not track page table pages, only
3933 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3934 * However, for the scan to be efficient we try to
3935 * cache items top-down.
3940 for (; sva < eva; sva = va_next) {
3943 if (sva >= VM_MAX_USER_ADDRESS) {
3952 * PD cache, scan shortcut if it doesn't exist.
3954 if (pd_pv == NULL) {
3955 pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
3956 } else if (pd_pv->pv_pmap != pmap ||
3957 pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3959 pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
3961 if (pd_pv == NULL) {
3962 va_next = (sva + NBPDP) & ~PDPMASK;
3971 if (pt_pv && (pt_pv->pv_pmap != pmap ||
3972 pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
3976 if (pt_pv == NULL) {
3977 pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
3978 &pt_placemark, &error);
3980 pv_put(pd_pv); /* lock order */
3987 pv_placemarker_wait(pmap, pt_placemark);
3992 /* may have to re-check later if pt_pv is NULL here */
3996 * If pt_pv is NULL we either have an shared page table
3997 * page and must issue a callback specific to that case,
3998 * or there is no page table page.
4000 * Either way we can skip the page table page.
4002 * WARNING! pt_pv can also be NULL due to a pv creation
4003 * race where we find it to be NULL and then
4004 * later see a pte_pv. But its possible the pt_pv
4005 * got created inbetween the two operations, so
4008 if (pt_pv == NULL) {
4010 * Possible unmanaged (shared from another pmap)
4013 * WARNING! We must hold pt_placemark across the
4014 * *ptep test to prevent misintepreting
4015 * a non-zero *ptep as a shared page
4016 * table page. Hold it across the function
4017 * callback as well for SMP safety.
4019 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4020 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4021 info->func(pmap, info, NULL, pt_placemark,
4023 sva, ptep, info->arg);
4025 pv_placemarker_wakeup(pmap, pt_placemark);
4029 * Done, move to next page table page.
4031 va_next = (sva + NBPDR) & ~PDRMASK;
4038 * From this point in the loop testing pt_pv for non-NULL
4039 * means we are in UVM, else if it is NULL we are in KVM.
4041 * Limit our scan to either the end of the va represented
4042 * by the current page table page, or to the end of the
4043 * range being removed.
4046 va_next = (sva + NBPDR) & ~PDRMASK;
4053 * Scan the page table for pages. Some pages may not be
4054 * managed (might not have a pv_entry).
4056 * There is no page table management for kernel pages so
4057 * pt_pv will be NULL in that case, but otherwise pt_pv
4058 * is non-NULL, locked, and referenced.
4062 * At this point a non-NULL pt_pv means a UVA, and a NULL
4063 * pt_pv means a KVA.
4066 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4070 while (sva < va_next) {
4072 vm_pindex_t *pte_placemark;
4075 * Yield every 64 pages, stop if requested.
4077 if ((++info->count & 63) == 0)
4083 * We can shortcut our scan if *ptep == 0. This is
4084 * an unlocked check.
4094 * Acquire the related pte_pv, if any. If *ptep == 0
4095 * the related pte_pv should not exist, but if *ptep
4096 * is not zero the pte_pv may or may not exist (e.g.
4097 * will not exist for an unmanaged page).
4099 * However a multitude of races are possible here
4100 * so if we cannot lock definite state we clean out
4101 * our cache and break the inner while() loop to
4102 * force a loop up to the top of the for().
4104 * XXX unlock/relock pd_pv, pt_pv, and re-test their
4105 * validity instead of looping up?
4107 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4108 &pte_placemark, &error);
4110 pv_put(pd_pv); /* lock order */
4113 pv_put(pt_pv); /* lock order */
4116 if (pte_pv) { /* block */
4121 pv_placemarker_wait(pmap,
4124 va_next = sva; /* retry */
4129 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
4134 kprintf("Unexpected non-NULL pte_pv "
4136 "*ptep = %016lx/%016lx\n",
4137 pte_pv, pt_pv, *ptep, oldpte);
4138 panic("Unexpected non-NULL pte_pv");
4140 pv_placemarker_wakeup(pmap, pte_placemark);
4147 * We can't hold pd_pv across the callback (because
4148 * we don't pass it to the callback and the callback
4152 vm_page_wire_quick(pd_pv->pv_m);
4157 * Ready for the callback. The locked pte_pv (if any)
4158 * is consumed by the callback. pte_pv will exist if
4159 * the page is managed, and will not exist if it
4163 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
4164 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
4165 ("badC *ptep %016lx/%016lx sva %016lx "
4167 *ptep, oldpte, sva, pte_pv));
4169 * We must unlock pd_pv across the callback
4170 * to avoid deadlocks on any recursive
4171 * disposal. Re-check that it still exists
4174 * Call target disposes of pte_pv and may
4175 * destroy but will not dispose of pt_pv.
4177 info->func(pmap, info, pte_pv, NULL,
4179 sva, ptep, info->arg);
4182 * We must unlock pd_pv across the callback
4183 * to avoid deadlocks on any recursive
4184 * disposal. Re-check that it still exists
4187 * Call target disposes of pte_pv and may
4188 * destroy but will not dispose of pt_pv.
4190 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
4191 pmap->pmap_bits[PG_V_IDX],
4192 ("badD *ptep %016lx/%016lx sva %016lx "
4194 *ptep, oldpte, sva));
4196 info->func(pmap, info, NULL, pte_placemark,
4198 sva, ptep, info->arg);
4202 vm_page_unwire_quick(pd_pv->pv_m);
4203 if (pd_pv->pv_pmap == NULL) {
4204 va_next = sva; /* retry */
4221 if ((++info->count & 7) == 0)
4225 * Relock before returning.
4227 spin_lock(&pmap->pm_spin);
4232 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4234 struct pmap_scan_info info;
4239 info.func = pmap_remove_callback;
4241 pmap_scan(&info, 1);
4245 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4247 struct pmap_scan_info info;
4252 info.func = pmap_remove_callback;
4254 pmap_scan(&info, 0);
4258 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4259 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4260 pv_entry_t pt_pv, int sharept,
4261 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4267 * This will also drop pt_pv's wire_count. Note that
4268 * terminal pages are not wired based on mmu presence.
4270 * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4272 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4273 pte_pv = NULL; /* safety */
4276 * Recursively destroy higher-level page tables.
4278 * This is optional. If we do not, they will still
4279 * be destroyed when the process exits.
4281 * NOTE: Do not destroy pv_entry's with extra hold refs,
4282 * a caller may have unlocked it and intends to
4283 * continue to use it.
4285 if (pmap_dynamic_delete &&
4288 pt_pv->pv_m->wire_count == 1 &&
4289 (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4290 pt_pv->pv_pindex != pmap_pml4_pindex()) {
4291 if (pmap_dynamic_delete == 2)
4292 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4293 pv_hold(pt_pv); /* extra hold */
4294 pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4295 pv_lock(pt_pv); /* prior extra hold + relock */
4297 } else if (sharept == 0) {
4299 * Unmanaged page table (pt, pd, or pdp. Not pte).
4301 * pt_pv's wire_count is still bumped by unmanaged pages
4302 * so we must decrement it manually.
4304 * We have to unwire the target page table page.
4306 * It is unclear how we can invalidate a segment so we
4307 * invalidate -1 which invlidates the tlb.
4309 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4310 if (pte & pmap->pmap_bits[PG_W_IDX])
4311 atomic_add_long(&pmap->pm_stats.wired_count, -1);
4312 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4313 if (vm_page_unwire_quick(pt_pv->pv_m))
4314 panic("pmap_remove: insufficient wirecount");
4315 pv_placemarker_wakeup(pmap, pte_placemark);
4318 * Unmanaged page table (pt, pd, or pdp. Not pte) for
4319 * a shared page table.
4321 * pt_pv is actually the pd_pv for our pmap (not the shared
4324 * We have to unwire the target page table page and we
4325 * have to unwire our page directory page.
4327 * It is unclear how we can invalidate a segment so we
4328 * invalidate -1 which invlidates the tlb.
4330 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4331 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4332 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4333 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4334 panic("pmap_remove: shared pgtable1 bad wirecount");
4335 if (vm_page_unwire_quick(pt_pv->pv_m))
4336 panic("pmap_remove: shared pgtable2 bad wirecount");
4337 pv_placemarker_wakeup(pmap, pte_placemark);
4342 * Removes this physical page from all physical maps in which it resides.
4343 * Reflects back modify bits to the pager.
4345 * This routine may not be called from an interrupt.
4349 pmap_remove_all(vm_page_t m)
4352 pmap_inval_bulk_t bulk;
4354 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4357 vm_page_spin_lock(m);
4358 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4359 KKASSERT(pv->pv_m == m);
4360 if (pv_hold_try(pv)) {
4361 vm_page_spin_unlock(m);
4363 vm_page_spin_unlock(m);
4366 if (pv->pv_pmap == NULL || pv->pv_m != m) {
4368 vm_page_spin_lock(m);
4373 * Holding no spinlocks, pv is locked. Once we scrap
4374 * pv we can no longer use it as a list iterator (but
4375 * we are doing a TAILQ_FIRST() so we are ok).
4377 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4378 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4379 pv = NULL; /* safety */
4380 pmap_inval_bulk_flush(&bulk);
4381 vm_page_spin_lock(m);
4383 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4384 vm_page_spin_unlock(m);
4388 * Removes the page from a particular pmap
4391 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4394 pmap_inval_bulk_t bulk;
4396 if (!pmap_initialized)
4400 vm_page_spin_lock(m);
4401 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4402 if (pv->pv_pmap != pmap)
4404 KKASSERT(pv->pv_m == m);
4405 if (pv_hold_try(pv)) {
4406 vm_page_spin_unlock(m);
4408 vm_page_spin_unlock(m);
4411 if (pv->pv_pmap != pmap || pv->pv_m != m) {
4417 * Holding no spinlocks, pv is locked. Once gone it can't
4418 * be used as an iterator. In fact, because we couldn't
4419 * necessarily lock it atomically it may have moved within
4420 * the list and ALSO cannot be used as an iterator.
4422 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4423 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4424 pv = NULL; /* safety */
4425 pmap_inval_bulk_flush(&bulk);
4428 vm_page_spin_unlock(m);
4432 * Set the physical protection on the specified range of this map
4433 * as requested. This function is typically only used for debug watchpoints
4436 * This function may not be called from an interrupt if the map is
4437 * not the kernel_pmap.
4439 * NOTE! For shared page table pages we just unmap the page.
4442 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4444 struct pmap_scan_info info;
4445 /* JG review for NX */
4449 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4450 pmap_remove(pmap, sva, eva);
4453 if (prot & VM_PROT_WRITE)
4458 info.func = pmap_protect_callback;
4460 pmap_scan(&info, 1);
4465 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4466 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4467 pv_entry_t pt_pv, int sharept,
4468 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4480 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4481 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4482 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4483 KKASSERT(m == pte_pv->pv_m);
4484 vm_page_flag_set(m, PG_REFERENCED);
4486 cbits &= ~pmap->pmap_bits[PG_A_IDX];
4488 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4489 if (pmap_track_modified(pte_pv->pv_pindex)) {
4490 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4492 m = PHYS_TO_VM_PAGE(pbits &
4497 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4500 } else if (sharept) {
4502 * Unmanaged page table, pt_pv is actually the pd_pv
4503 * for our pmap (not the object's shared pmap).
4505 * When asked to protect something in a shared page table
4506 * page we just unmap the page table page. We have to
4507 * invalidate the tlb in this situation.
4509 * XXX Warning, shared page tables will not be used for
4510 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4511 * so PHYS_TO_VM_PAGE() should be safe here.
4513 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4514 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4515 panic("pmap_protect: pgtable1 pg bad wirecount");
4516 if (vm_page_unwire_quick(pt_pv->pv_m))
4517 panic("pmap_protect: pgtable2 pg bad wirecount");
4520 /* else unmanaged page, adjust bits, no wire changes */
4523 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4525 if (pmap_enter_debug > 0) {
4527 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4528 "pt_pv=%p cbits=%08lx\n",
4534 if (pbits != cbits) {
4535 if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4536 ptep, pbits, cbits)) {
4544 pv_placemarker_wakeup(pmap, pte_placemark);
4548 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4549 * mapping at that address. Set protection and wiring as requested.
4551 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4552 * possible. If it is we enter the page into the appropriate shared pmap
4553 * hanging off the related VM object instead of the passed pmap, then we
4554 * share the page table page from the VM object's pmap into the current pmap.
4556 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4560 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4561 boolean_t wired, vm_map_entry_t entry)
4563 pv_entry_t pt_pv; /* page table */
4564 pv_entry_t pte_pv; /* page table entry */
4565 vm_pindex_t *pte_placemark;
4568 pt_entry_t origpte, newpte;
4573 va = trunc_page(va);
4574 #ifdef PMAP_DIAGNOSTIC
4576 panic("pmap_enter: toobig");
4577 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4578 panic("pmap_enter: invalid to pmap_enter page table "
4579 "pages (va: 0x%lx)", va);
4581 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4582 kprintf("Warning: pmap_enter called on UVA with "
4585 db_print_backtrace();
4588 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4589 kprintf("Warning: pmap_enter called on KVA without"
4592 db_print_backtrace();
4597 * Get locked PV entries for our new page table entry (pte_pv)
4598 * and for its parent page table (pt_pv). We need the parent
4599 * so we can resolve the location of the ptep.
4601 * Only hardware MMU actions can modify the ptep out from
4604 * if (m) is fictitious or unmanaged we do not create a managing
4605 * pte_pv for it. Any pre-existing page's management state must
4606 * match (avoiding code complexity).
4608 * If the pmap is still being initialized we assume existing
4611 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4613 if (pmap_initialized == FALSE) {
4616 pte_placemark = NULL;
4619 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4620 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4621 KKASSERT(pte_pv == NULL);
4622 if (va >= VM_MAX_USER_ADDRESS) {
4626 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4628 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4632 KASSERT(origpte == 0 ||
4633 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4634 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4636 if (va >= VM_MAX_USER_ADDRESS) {
4638 * Kernel map, pv_entry-tracked.
4641 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4647 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4649 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4651 pte_placemark = NULL; /* safety */
4654 KASSERT(origpte == 0 ||
4655 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4656 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4659 pa = VM_PAGE_TO_PHYS(m);
4660 opa = origpte & PG_FRAME;
4663 * Calculate the new PTE. Note that pte_pv alone does not mean
4664 * the new pte_pv is managed, it could exist because the old pte
4665 * was managed even if the new one is not.
4667 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4668 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4670 newpte |= pmap->pmap_bits[PG_W_IDX];
4671 if (va < VM_MAX_USER_ADDRESS)
4672 newpte |= pmap->pmap_bits[PG_U_IDX];
4673 if (pte_pv && (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) == 0)
4674 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4675 // if (pmap == &kernel_pmap)
4676 // newpte |= pgeflag;
4677 newpte |= pmap->pmap_cache_bits[m->pat_mode];
4678 if (m->flags & PG_FICTITIOUS)
4679 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4682 * It is possible for multiple faults to occur in threaded
4683 * environments, the existing pte might be correct.
4685 if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4686 pmap->pmap_bits[PG_A_IDX])) == 0)
4690 * Ok, either the address changed or the protection or wiring
4693 * Clear the current entry, interlocking the removal. For managed
4694 * pte's this will also flush the modified state to the vm_page.
4695 * Atomic ops are mandatory in order to ensure that PG_M events are
4696 * not lost during any transition.
4698 * WARNING: The caller has busied the new page but not the original
4699 * vm_page which we are trying to replace. Because we hold
4700 * the pte_pv lock, but have not busied the page, PG bits
4701 * can be cleared out from under us.
4706 * NOTE: pt_pv won't exist for a kernel page
4707 * (managed or otherwise).
4709 * NOTE: We are reusing the pte_pv so we do not
4710 * destroy it in pmap_remove_pv_pte(). Also,
4711 * pte_pv might be freshly allocated with opa
4712 * being unmanaged and the new page being
4713 * managed, so pte_pv->pv_m may be NULL.
4715 if (prot & VM_PROT_NOSYNC) {
4716 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4718 pmap_inval_bulk_t bulk;
4720 pmap_inval_bulk_init(&bulk, pmap);
4721 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4722 pmap_inval_bulk_flush(&bulk);
4724 if (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4725 KKASSERT(pte_pv->pv_m);
4726 pmap_remove_pv_page(pte_pv);
4728 KKASSERT(pte_pv->pv_m == NULL);
4730 } else if (prot & VM_PROT_NOSYNC) {
4732 * Unmanaged page, NOSYNC (no mmu sync) requested.
4734 * Leave wire count on PT page intact.
4736 (void)pte_load_clear(ptep);
4737 cpu_invlpg((void *)va);
4738 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4741 * Unmanaged page, normal enter.
4743 * Leave wire count on PT page intact.
4745 pmap_inval_smp(pmap, va, 1, ptep, 0);
4746 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4748 KKASSERT(*ptep == 0);
4752 if (pmap_enter_debug > 0) {
4754 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4755 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4757 origpte, newpte, ptep,
4758 pte_pv, pt_pv, opa, prot);
4764 * Enter on the PV list if part of our managed memory.
4765 * Wiring of the PT page is already handled.
4767 KKASSERT(pte_pv->pv_m == NULL);
4768 vm_page_spin_lock(m);
4770 pmap_page_stats_adding(m);
4771 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4772 vm_page_flag_set(m, PG_MAPPED);
4773 vm_page_spin_unlock(m);
4774 } else if (pt_pv && opa == 0) {
4776 * We have to adjust the wire count on the PT page ourselves
4777 * for unmanaged entries. If opa was non-zero we retained
4778 * the existing wire count from the removal.
4780 vm_page_wire_quick(pt_pv->pv_m);
4784 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4786 * User VMAs do not because those will be zero->non-zero, so no
4787 * stale entries to worry about at this point.
4789 * For KVM there appear to still be issues. Theoretically we
4790 * should be able to scrap the interlocks entirely but we
4793 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4794 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4796 atomic_swap_long(ptep, newpte);
4798 cpu_invlpg((void *)va);
4803 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4806 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4809 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4810 vm_page_flag_set(m, PG_WRITEABLE);
4813 * Unmanaged pages need manual resident_count tracking.
4815 if (pte_pv == NULL && pt_pv) {
4816 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4823 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4824 (m->flags & PG_MAPPED));
4827 * Cleanup the pv entry, allowing other accessors.
4831 else if (pte_placemark)
4832 pv_placemarker_wakeup(pmap, pte_placemark);
4839 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4840 * This code also assumes that the pmap has no pre-existing entry for this
4843 * This code currently may only be used on user pmaps, not kernel_pmap.
4846 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4848 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4852 * Make a temporary mapping for a physical address. This is only intended
4853 * to be used for panic dumps.
4855 * The caller is responsible for calling smp_invltlb().
4858 pmap_kenter_temporary(vm_paddr_t pa, long i)
4860 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4861 return ((void *)crashdumpmap);
4864 #define MAX_INIT_PT (96)
4867 * This routine preloads the ptes for a given object into the specified pmap.
4868 * This eliminates the blast of soft faults on process startup and
4869 * immediately after an mmap.
4871 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4874 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4875 vm_object_t object, vm_pindex_t pindex,
4876 vm_size_t size, int limit)
4878 struct rb_vm_page_scan_info info;
4883 * We can't preinit if read access isn't set or there is no pmap
4886 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4890 * We can't preinit if the pmap is not the current pmap
4892 lp = curthread->td_lwp;
4893 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4897 * Misc additional checks
4899 psize = x86_64_btop(size);
4901 if ((object->type != OBJT_VNODE) ||
4902 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4903 (object->resident_page_count > MAX_INIT_PT))) {
4907 if (pindex + psize > object->size) {
4908 if (object->size < pindex)
4910 psize = object->size - pindex;
4917 * If everything is segment-aligned do not pre-init here. Instead
4918 * allow the normal vm_fault path to pass a segment hint to
4919 * pmap_enter() which will then use an object-referenced shared
4922 if ((addr & SEG_MASK) == 0 &&
4923 (ctob(psize) & SEG_MASK) == 0 &&
4924 (ctob(pindex) & SEG_MASK) == 0) {
4929 * Use a red-black scan to traverse the requested range and load
4930 * any valid pages found into the pmap.
4932 * We cannot safely scan the object's memq without holding the
4935 info.start_pindex = pindex;
4936 info.end_pindex = pindex + psize - 1;
4942 vm_object_hold_shared(object);
4943 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4944 pmap_object_init_pt_callback, &info);
4945 vm_object_drop(object);
4950 pmap_object_init_pt_callback(vm_page_t p, void *data)
4952 struct rb_vm_page_scan_info *info = data;
4953 vm_pindex_t rel_index;
4956 * don't allow an madvise to blow away our really
4957 * free pages allocating pv entries.
4959 if ((info->limit & MAP_PREFAULT_MADVISE) &&
4960 vmstats.v_free_count < vmstats.v_free_reserved) {
4965 * Ignore list markers and ignore pages we cannot instantly
4966 * busy (while holding the object token).
4968 if (p->flags & PG_MARKER)
4970 if (vm_page_busy_try(p, TRUE))
4972 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4973 (p->flags & PG_FICTITIOUS) == 0) {
4974 if ((p->queue - p->pc) == PQ_CACHE)
4975 vm_page_deactivate(p);
4976 rel_index = p->pindex - info->start_pindex;
4977 pmap_enter_quick(info->pmap,
4978 info->addr + x86_64_ptob(rel_index), p);
4986 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4989 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4992 * XXX This is safe only because page table pages are not freed.
4995 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4999 /*spin_lock(&pmap->pm_spin);*/
5000 if ((pte = pmap_pte(pmap, addr)) != NULL) {
5001 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5002 /*spin_unlock(&pmap->pm_spin);*/
5006 /*spin_unlock(&pmap->pm_spin);*/
5011 * Change the wiring attribute for a pmap/va pair. The mapping must already
5012 * exist in the pmap. The mapping may or may not be managed. The wiring in
5013 * the page is not changed, the page is returned so the caller can adjust
5014 * its wiring (the page is not locked in any way).
5016 * Wiring is not a hardware characteristic so there is no need to invalidate
5017 * TLB. However, in an SMP environment we must use a locked bus cycle to
5018 * update the pte (if we are not using the pmap_inval_*() API that is)...
5019 * it's ok to do this for simple wiring changes.
5022 pmap_unwire(pmap_t pmap, vm_offset_t va)
5033 * Assume elements in the kernel pmap are stable
5035 if (pmap == &kernel_pmap) {
5036 if (pmap_pt(pmap, va) == 0)
5038 ptep = pmap_pte_quick(pmap, va);
5040 if (pmap_pte_w(pmap, ptep))
5041 atomic_add_long(&pmap->pm_stats.wired_count,-1);
5042 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5043 pa = *ptep & PG_FRAME;
5044 m = PHYS_TO_VM_PAGE(pa);
5047 * We can only [un]wire pmap-local pages (we cannot wire
5050 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5054 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5055 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5060 if (pmap_pte_w(pmap, ptep)) {
5061 atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5064 /* XXX else return NULL so caller doesn't unwire m ? */
5066 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5068 pa = *ptep & PG_FRAME;
5069 m = PHYS_TO_VM_PAGE(pa); /* held by wired count */
5076 * Copy the range specified by src_addr/len from the source map to
5077 * the range dst_addr/len in the destination map.
5079 * This routine is only advisory and need not do anything.
5082 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5083 vm_size_t len, vm_offset_t src_addr)
5090 * Zero the specified physical page.
5092 * This function may be called from an interrupt and no locking is
5096 pmap_zero_page(vm_paddr_t phys)
5098 vm_offset_t va = PHYS_TO_DMAP(phys);
5100 pagezero((void *)va);
5106 * Zero part of a physical page by mapping it into memory and clearing
5107 * its contents with bzero.
5109 * off and size may not cover an area beyond a single hardware page.
5112 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5114 vm_offset_t virt = PHYS_TO_DMAP(phys);
5116 bzero((char *)virt + off, size);
5122 * Copy the physical page from the source PA to the target PA.
5123 * This function may be called from an interrupt. No locking
5127 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5129 vm_offset_t src_virt, dst_virt;
5131 src_virt = PHYS_TO_DMAP(src);
5132 dst_virt = PHYS_TO_DMAP(dst);
5133 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5137 * pmap_copy_page_frag:
5139 * Copy the physical page from the source PA to the target PA.
5140 * This function may be called from an interrupt. No locking
5144 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5146 vm_offset_t src_virt, dst_virt;
5148 src_virt = PHYS_TO_DMAP(src);
5149 dst_virt = PHYS_TO_DMAP(dst);
5151 bcopy((char *)src_virt + (src & PAGE_MASK),
5152 (char *)dst_virt + (dst & PAGE_MASK),
5157 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5158 * this page. This count may be changed upwards or downwards in the future;
5159 * it is only necessary that true be returned for a small subset of pmaps
5160 * for proper page aging.
5163 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5168 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5171 vm_page_spin_lock(m);
5172 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5173 if (pv->pv_pmap == pmap) {
5174 vm_page_spin_unlock(m);
5181 vm_page_spin_unlock(m);
5186 * Remove all pages from specified address space this aids process exit
5187 * speeds. Also, this code may be special cased for the current process
5191 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5193 pmap_remove_noinval(pmap, sva, eva);
5198 * pmap_testbit tests bits in pte's note that the testbit/clearbit
5199 * routines are inline, and a lot of things compile-time evaluate.
5203 pmap_testbit(vm_page_t m, int bit)
5209 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5212 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5214 vm_page_spin_lock(m);
5215 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5216 vm_page_spin_unlock(m);
5220 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5222 #if defined(PMAP_DIAGNOSTIC)
5223 if (pv->pv_pmap == NULL) {
5224 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5232 * If the bit being tested is the modified bit, then
5233 * mark clean_map and ptes as never
5236 * WARNING! Because we do not lock the pv, *pte can be in a
5237 * state of flux. Despite this the value of *pte
5238 * will still be related to the vm_page in some way
5239 * because the pv cannot be destroyed as long as we
5240 * hold the vm_page spin lock.
5242 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5243 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5244 if (!pmap_track_modified(pv->pv_pindex))
5248 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5249 if (*pte & pmap->pmap_bits[bit]) {
5250 vm_page_spin_unlock(m);
5254 vm_page_spin_unlock(m);
5259 * This routine is used to modify bits in ptes. Only one bit should be
5260 * specified. PG_RW requires special handling.
5262 * Caller must NOT hold any spin locks
5266 pmap_clearbit(vm_page_t m, int bit_index)
5273 if (bit_index == PG_RW_IDX)
5274 vm_page_flag_clear(m, PG_WRITEABLE);
5275 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5282 * Loop over all current mappings setting/clearing as appropos If
5283 * setting RO do we need to clear the VAC?
5285 * NOTE: When clearing PG_M we could also (not implemented) drop
5286 * through to the PG_RW code and clear PG_RW too, forcing
5287 * a fault on write to redetect PG_M for virtual kernels, but
5288 * it isn't necessary since virtual kernels invalidate the
5289 * pte when they clear the VPTE_M bit in their virtual page
5292 * NOTE: Does not re-dirty the page when clearing only PG_M.
5294 * NOTE: Because we do not lock the pv, *pte can be in a state of
5295 * flux. Despite this the value of *pte is still somewhat
5296 * related while we hold the vm_page spin lock.
5298 * *pte can be zero due to this race. Since we are clearing
5299 * bits we basically do no harm when this race ccurs.
5301 if (bit_index != PG_RW_IDX) {
5302 vm_page_spin_lock(m);
5303 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5304 #if defined(PMAP_DIAGNOSTIC)
5305 if (pv->pv_pmap == NULL) {
5306 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5312 pte = pmap_pte_quick(pv->pv_pmap,
5313 pv->pv_pindex << PAGE_SHIFT);
5315 if (pbits & pmap->pmap_bits[bit_index])
5316 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5318 vm_page_spin_unlock(m);
5323 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
5327 vm_page_spin_lock(m);
5329 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5331 * don't write protect pager mappings
5333 if (!pmap_track_modified(pv->pv_pindex))
5336 #if defined(PMAP_DIAGNOSTIC)
5337 if (pv->pv_pmap == NULL) {
5338 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5346 * Skip pages which do not have PG_RW set.
5348 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5349 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5355 if (pv_hold_try(pv)) {
5356 vm_page_spin_unlock(m);
5358 vm_page_spin_unlock(m);
5359 pv_lock(pv); /* held, now do a blocking lock */
5361 if (pv->pv_pmap != pmap || pv->pv_m != m) {
5362 pv_put(pv); /* and release */
5363 goto restart; /* anything could have happened */
5365 KKASSERT(pv->pv_pmap == pmap);
5371 nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5372 pmap->pmap_bits[PG_M_IDX]);
5373 if (pmap_inval_smp_cmpset(pmap,
5374 ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5375 pte, pbits, nbits)) {
5380 vm_page_spin_lock(m);
5383 * If PG_M was found to be set while we were clearing PG_RW
5384 * we also clear PG_M (done above) and mark the page dirty.
5385 * Callers expect this behavior.
5387 * we lost pv so it cannot be used as an iterator. In fact,
5388 * because we couldn't necessarily lock it atomically it may
5389 * have moved within the list and ALSO cannot be used as an
5392 if (pbits & pmap->pmap_bits[PG_M_IDX])
5395 goto restart_locked;
5397 vm_page_spin_unlock(m);
5401 * Lower the permission for all mappings to a given page.
5403 * Page must be busied by caller. Because page is busied by caller this
5404 * should not be able to race a pmap_enter().
5407 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5409 /* JG NX support? */
5410 if ((prot & VM_PROT_WRITE) == 0) {
5411 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5413 * NOTE: pmap_clearbit(.. PG_RW) also clears
5414 * the PG_WRITEABLE flag in (m).
5416 pmap_clearbit(m, PG_RW_IDX);
5424 pmap_phys_address(vm_pindex_t ppn)
5426 return (x86_64_ptob(ppn));
5430 * Return a count of reference bits for a page, clearing those bits.
5431 * It is not necessary for every reference bit to be cleared, but it
5432 * is necessary that 0 only be returned when there are truly no
5433 * reference bits set.
5435 * XXX: The exact number of bits to check and clear is a matter that
5436 * should be tested and standardized at some point in the future for
5437 * optimal aging of shared pages.
5439 * This routine may not block.
5442 pmap_ts_referenced(vm_page_t m)
5449 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5452 vm_page_spin_lock(m);
5453 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5454 if (!pmap_track_modified(pv->pv_pindex))
5457 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5458 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5459 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5465 vm_page_spin_unlock(m);
5472 * Return whether or not the specified physical page was modified
5473 * in any physical maps.
5476 pmap_is_modified(vm_page_t m)
5480 res = pmap_testbit(m, PG_M_IDX);
5485 * Clear the modify bits on the specified physical page.
5488 pmap_clear_modify(vm_page_t m)
5490 pmap_clearbit(m, PG_M_IDX);
5494 * pmap_clear_reference:
5496 * Clear the reference bit on the specified physical page.
5499 pmap_clear_reference(vm_page_t m)
5501 pmap_clearbit(m, PG_A_IDX);
5505 * Miscellaneous support routines follow
5510 i386_protection_init(void)
5514 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
5515 kp = protection_codes;
5516 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5518 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5520 * Read access is also 0. There isn't any execute bit,
5521 * so just make it readable.
5523 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5524 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5525 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5528 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5529 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5530 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5531 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5532 *kp++ = pmap_bits_default[PG_RW_IDX];
5539 * Map a set of physical memory pages into the kernel virtual
5540 * address space. Return a pointer to where it is mapped. This
5541 * routine is intended to be used for mapping device memory,
5544 * NOTE: We can't use pgeflag unless we invalidate the pages one at
5547 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5548 * work whether the cpu supports PAT or not. The remaining PAT
5549 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5553 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5555 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5559 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5561 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5565 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5567 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5571 * Map a set of physical memory pages into the kernel virtual
5572 * address space. Return a pointer to where it is mapped. This
5573 * routine is intended to be used for mapping device memory,
5577 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5579 vm_offset_t va, tmpva, offset;
5583 offset = pa & PAGE_MASK;
5584 size = roundup(offset + size, PAGE_SIZE);
5586 va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5588 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5590 pa = pa & ~PAGE_MASK;
5591 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5592 pte = vtopte(tmpva);
5594 kernel_pmap.pmap_bits[PG_RW_IDX] |
5595 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5596 kernel_pmap.pmap_cache_bits[mode];
5597 tmpsize -= PAGE_SIZE;
5601 pmap_invalidate_range(&kernel_pmap, va, va + size);
5602 pmap_invalidate_cache_range(va, va + size);
5604 return ((void *)(va + offset));
5608 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5610 vm_offset_t base, offset;
5612 base = va & ~PAGE_MASK;
5613 offset = va & PAGE_MASK;
5614 size = roundup(offset + size, PAGE_SIZE);
5615 pmap_qremove(va, size >> PAGE_SHIFT);
5616 kmem_free(&kernel_map, base, size);
5620 * Sets the memory attribute for the specified page.
5623 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5629 * If "m" is a normal page, update its direct mapping. This update
5630 * can be relied upon to perform any cache operations that are
5631 * required for data coherence.
5633 if ((m->flags & PG_FICTITIOUS) == 0)
5634 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5638 * Change the PAT attribute on an existing kernel memory map. Caller
5639 * must ensure that the virtual memory in question is not accessed
5640 * during the adjustment.
5643 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5650 panic("pmap_change_attr: va is NULL");
5651 base = trunc_page(va);
5655 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5656 kernel_pmap.pmap_cache_bits[mode];
5661 changed = 1; /* XXX: not optimal */
5664 * Flush CPU caches if required to make sure any data isn't cached that
5665 * shouldn't be, etc.
5668 pmap_invalidate_range(&kernel_pmap, base, va);
5669 pmap_invalidate_cache_range(base, va);
5674 * perform the pmap work for mincore
5677 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5679 pt_entry_t *ptep, pte;
5683 ptep = pmap_pte(pmap, addr);
5685 if (ptep && (pte = *ptep) != 0) {
5688 val = MINCORE_INCORE;
5689 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5692 pa = pte & PG_FRAME;
5694 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5697 m = PHYS_TO_VM_PAGE(pa);
5702 if (pte & pmap->pmap_bits[PG_M_IDX])
5703 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5705 * Modified by someone
5707 else if (m && (m->dirty || pmap_is_modified(m)))
5708 val |= MINCORE_MODIFIED_OTHER;
5712 if (pte & pmap->pmap_bits[PG_A_IDX])
5713 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5716 * Referenced by someone
5718 else if (m && ((m->flags & PG_REFERENCED) ||
5719 pmap_ts_referenced(m))) {
5720 val |= MINCORE_REFERENCED_OTHER;
5721 vm_page_flag_set(m, PG_REFERENCED);
5730 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5731 * vmspace will be ref'd and the old one will be deref'd.
5733 * The vmspace for all lwps associated with the process will be adjusted
5734 * and cr3 will be reloaded if any lwp is the current lwp.
5736 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5739 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5741 struct vmspace *oldvm;
5744 oldvm = p->p_vmspace;
5745 if (oldvm != newvm) {
5748 p->p_vmspace = newvm;
5749 KKASSERT(p->p_nthreads == 1);
5750 lp = RB_ROOT(&p->p_lwp_tree);
5751 pmap_setlwpvm(lp, newvm);
5758 * Set the vmspace for a LWP. The vmspace is almost universally set the
5759 * same as the process vmspace, but virtual kernels need to swap out contexts
5760 * on a per-lwp basis.
5762 * Caller does not necessarily hold any vmspace tokens. Caller must control
5763 * the lwp (typically be in the context of the lwp). We use a critical
5764 * section to protect against statclock and hardclock (statistics collection).
5767 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5769 struct vmspace *oldvm;
5772 oldvm = lp->lwp_vmspace;
5774 if (oldvm != newvm) {
5776 lp->lwp_vmspace = newvm;
5777 if (curthread->td_lwp == lp) {
5778 pmap = vmspace_pmap(newvm);
5779 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5780 if (pmap->pm_active_lock & CPULOCK_EXCL)
5781 pmap_interlock_wait(newvm);
5782 #if defined(SWTCH_OPTIM_STATS)
5785 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5786 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5787 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5788 curthread->td_pcb->pcb_cr3 = KPML4phys;
5790 panic("pmap_setlwpvm: unknown pmap type\n");
5792 load_cr3(curthread->td_pcb->pcb_cr3);
5793 pmap = vmspace_pmap(oldvm);
5794 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5802 * Called when switching to a locked pmap, used to interlock against pmaps
5803 * undergoing modifications to prevent us from activating the MMU for the
5804 * target pmap until all such modifications have completed. We have to do
5805 * this because the thread making the modifications has already set up its
5806 * SMP synchronization mask.
5808 * This function cannot sleep!
5813 pmap_interlock_wait(struct vmspace *vm)
5815 struct pmap *pmap = &vm->vm_pmap;
5817 if (pmap->pm_active_lock & CPULOCK_EXCL) {
5819 KKASSERT(curthread->td_critcount >= 2);
5820 DEBUG_PUSH_INFO("pmap_interlock_wait");
5821 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5823 lwkt_process_ipiq();
5831 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5834 if ((obj == NULL) || (size < NBPDR) ||
5835 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5839 addr = roundup2(addr, NBPDR);
5844 * Used by kmalloc/kfree, page already exists at va
5847 pmap_kvtom(vm_offset_t va)
5849 pt_entry_t *ptep = vtopte(va);
5851 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5852 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5856 * Initialize machine-specific shared page directory support. This
5857 * is executed when a VM object is created.
5860 pmap_object_init(vm_object_t object)
5862 object->md.pmap_rw = NULL;
5863 object->md.pmap_ro = NULL;
5867 * Clean up machine-specific shared page directory support. This
5868 * is executed when a VM object is destroyed.
5871 pmap_object_free(vm_object_t object)
5875 if ((pmap = object->md.pmap_rw) != NULL) {
5876 object->md.pmap_rw = NULL;
5877 pmap_remove_noinval(pmap,
5878 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5879 CPUMASK_ASSZERO(pmap->pm_active);
5882 kfree(pmap, M_OBJPMAP);
5884 if ((pmap = object->md.pmap_ro) != NULL) {
5885 object->md.pmap_ro = NULL;
5886 pmap_remove_noinval(pmap,
5887 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5888 CPUMASK_ASSZERO(pmap->pm_active);
5891 kfree(pmap, M_OBJPMAP);
5896 * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5897 * VM page and issue a pginfo->callback.
5899 * We are expected to dispose of any non-NULL pte_pv.
5903 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5904 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
5905 pv_entry_t pt_pv, int sharept,
5906 vm_offset_t va, pt_entry_t *ptep, void *arg)
5908 struct pmap_pgscan_info *pginfo = arg;
5913 * Try to busy the page while we hold the pte_pv locked.
5915 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
5916 if (vm_page_busy_try(m, TRUE) == 0) {
5917 if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
5919 * The callback is issued with the pte_pv
5920 * unlocked and put away, and the pt_pv
5925 vm_page_wire_quick(pt_pv->pv_m);
5928 if (pginfo->callback(pginfo, va, m) < 0)
5932 vm_page_unwire_quick(pt_pv->pv_m);
5939 ++pginfo->busycount;
5942 } else if (sharept) {
5943 /* shared page table */
5944 pv_placemarker_wakeup(pmap, pte_placemark);
5946 /* else unmanaged page */
5947 pv_placemarker_wakeup(pmap, pte_placemark);
5952 pmap_pgscan(struct pmap_pgscan_info *pginfo)
5954 struct pmap_scan_info info;
5956 pginfo->offset = pginfo->beg_addr;
5957 info.pmap = pginfo->pmap;
5958 info.sva = pginfo->beg_addr;
5959 info.eva = pginfo->end_addr;
5960 info.func = pmap_pgscan_callback;
5962 pmap_scan(&info, 0);
5964 pginfo->offset = pginfo->end_addr;
5968 * Wait for a placemarker that we do not own to clear. The placemarker
5969 * in question is not necessary set to the pindex we want, we may have
5970 * to wait on the element because we want to reserve it ourselves.
5974 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
5976 spin_lock(&pmap->pm_spin);
5977 if (*pmark != PM_NOPLACEMARK) {
5978 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
5979 ssleep(pmark, &pmap->pm_spin, 0, "pvplw", 0);
5981 spin_unlock(&pmap->pm_spin);
5985 * Wakeup a placemarker that we own. Replace the entry with
5986 * PM_NOPLACEMARK and issue a wakeup() if necessary.
5990 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
5994 spin_lock(&pmap->pm_spin);
5995 pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
5996 spin_unlock(&pmap->pm_spin);
5997 KKASSERT(pindex != PM_NOPLACEMARK);
5998 if (pindex & PM_PLACEMARK_WAKEUP)