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) 2011-2017 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.
51 #include "opt_msgbuf.h"
53 #include <sys/param.h>
54 #include <sys/kernel.h>
56 #include <sys/msgbuf.h>
57 #include <sys/vmmeter.h>
59 #include <sys/systm.h>
62 #include <vm/vm_param.h>
63 #include <sys/sysctl.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_extern.h>
70 #include <vm/vm_pageout.h>
71 #include <vm/vm_pager.h>
72 #include <vm/vm_zone.h>
75 #include <sys/thread2.h>
76 #include <sys/spinlock2.h>
77 #include <vm/vm_page2.h>
79 #include <machine/cputypes.h>
80 #include <machine/md_var.h>
81 #include <machine/specialreg.h>
82 #include <machine/smp.h>
83 #include <machine_base/apic/apicreg.h>
84 #include <machine/globaldata.h>
85 #include <machine/pmap.h>
86 #include <machine/pmap_inval.h>
87 #include <machine/inttypes.h>
91 #define PMAP_KEEP_PDIRS
92 #ifndef PMAP_SHPGPERPROC
93 #define PMAP_SHPGPERPROC 2000
96 #if defined(DIAGNOSTIC)
97 #define PMAP_DIAGNOSTIC
103 * pmap debugging will report who owns a pv lock when blocking.
107 #define PMAP_DEBUG_DECL ,const char *func, int lineno
108 #define PMAP_DEBUG_ARGS , __func__, __LINE__
109 #define PMAP_DEBUG_COPY , func, lineno
111 #define pv_get(pmap, pindex, pmarkp) _pv_get(pmap, pindex, pmarkp \
113 #define pv_lock(pv) _pv_lock(pv \
115 #define pv_hold_try(pv) _pv_hold_try(pv \
117 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \
120 #define pv_free(pv, pvp) _pv_free(pv, pvp PMAP_DEBUG_ARGS)
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)
132 #define pv_free(pv, pvp) _pv_free(pv, pvp)
137 * Get PDEs and PTEs for user/kernel address space
139 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
141 #define pmap_pde_v(pmap, pte) ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
142 #define pmap_pte_w(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
143 #define pmap_pte_m(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
144 #define pmap_pte_u(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
145 #define pmap_pte_v(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
148 * Given a map and a machine independent protection code,
149 * convert to a vax protection code.
151 #define pte_prot(m, p) \
152 (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
153 static uint64_t protection_codes[PROTECTION_CODES_SIZE];
155 struct pmap kernel_pmap;
157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
159 vm_paddr_t avail_start; /* PA of first available physical page */
160 vm_paddr_t avail_end; /* PA of last available physical page */
161 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
162 vm_offset_t virtual2_end;
163 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
164 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
165 vm_offset_t KvaStart; /* VA start of KVA space */
166 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
167 vm_offset_t KvaSize; /* max size of kernel virtual address space */
168 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
169 //static int pgeflag; /* PG_G or-in */
173 static vm_paddr_t dmaplimit;
174 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
176 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE]; /* PAT -> PG_ bits */
177 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
179 static uint64_t KPTbase;
180 static uint64_t KPTphys;
181 static uint64_t KPDphys; /* phys addr of kernel level 2 */
182 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
183 uint64_t KPDPphys; /* phys addr of kernel level 3 */
184 uint64_t KPML4phys; /* phys addr of kernel level 4 */
186 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
187 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
190 * Data for the pv entry allocation mechanism
192 static vm_zone_t pvzone;
193 static struct vm_zone pvzone_store;
194 static vm_pindex_t pv_entry_max=0, pv_entry_high_water=0;
195 static int pmap_pagedaemon_waken = 0;
196 static struct pv_entry *pvinit;
199 * All those kernel PT submaps that BSD is so fond of
201 pt_entry_t *CMAP1 = NULL, *ptmmap;
202 caddr_t CADDR1 = NULL, ptvmmap = NULL;
203 static pt_entry_t *msgbufmap;
204 struct msgbuf *msgbufp=NULL;
207 * PMAP default PG_* bits. Needed to be able to add
208 * EPT/NPT pagetable pmap_bits for the VMM module
210 uint64_t pmap_bits_default[] = {
211 REGULAR_PMAP, /* TYPE_IDX 0 */
212 X86_PG_V, /* PG_V_IDX 1 */
213 X86_PG_RW, /* PG_RW_IDX 2 */
214 X86_PG_U, /* PG_U_IDX 3 */
215 X86_PG_A, /* PG_A_IDX 4 */
216 X86_PG_M, /* PG_M_IDX 5 */
217 X86_PG_PS, /* PG_PS_IDX3 6 */
218 X86_PG_G, /* PG_G_IDX 7 */
219 X86_PG_AVAIL1, /* PG_AVAIL1_IDX 8 */
220 X86_PG_AVAIL2, /* PG_AVAIL2_IDX 9 */
221 X86_PG_AVAIL3, /* PG_AVAIL3_IDX 10 */
222 X86_PG_NC_PWT | X86_PG_NC_PCD, /* PG_N_IDX 11 */
223 X86_PG_NX, /* PG_NX_IDX 12 */
228 static pt_entry_t *pt_crashdumpmap;
229 static caddr_t crashdumpmap;
231 static int pmap_debug = 0;
232 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
233 &pmap_debug, 0, "Debug pmap's");
235 static int pmap_enter_debug = 0;
236 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
237 &pmap_enter_debug, 0, "Debug pmap_enter's");
239 static int pmap_yield_count = 64;
240 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
241 &pmap_yield_count, 0, "Yield during init_pt/release");
242 static int pmap_mmu_optimize = 0;
243 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
244 &pmap_mmu_optimize, 0, "Share page table pages when possible");
245 int pmap_fast_kernel_cpusync = 0;
246 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
247 &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
248 int pmap_dynamic_delete = 0;
249 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
250 &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
251 int pmap_lock_delay = 100;
252 SYSCTL_INT(_machdep, OID_AUTO, pmap_lock_delay, CTLFLAG_RW,
253 &pmap_lock_delay, 0, "Spin loops");
255 static int pmap_nx_enable = 0;
256 /* needs manual TUNABLE in early probe, see below */
258 /* Standard user access funtions */
259 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
261 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
262 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
263 extern int std_fubyte (const uint8_t *base);
264 extern int std_subyte (uint8_t *base, uint8_t byte);
265 extern int32_t std_fuword32 (const uint32_t *base);
266 extern int64_t std_fuword64 (const uint64_t *base);
267 extern int std_suword64 (uint64_t *base, uint64_t word);
268 extern int std_suword32 (uint32_t *base, int word);
269 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
270 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
272 static void pv_hold(pv_entry_t pv);
273 static int _pv_hold_try(pv_entry_t pv
275 static void pv_drop(pv_entry_t pv);
276 static void _pv_lock(pv_entry_t pv
278 static void pv_unlock(pv_entry_t pv);
279 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
281 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
283 static void _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL);
284 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
285 vm_pindex_t **pmarkp, int *errorp);
286 static void pv_put(pv_entry_t pv);
287 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
288 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
290 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
291 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
292 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
293 pmap_inval_bulk_t *bulk, int destroy);
294 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
295 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
296 pmap_inval_bulk_t *bulk);
298 struct pmap_scan_info;
299 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
300 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
301 pv_entry_t pt_pv, int sharept,
302 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
303 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
304 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
305 pv_entry_t pt_pv, int sharept,
306 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
308 static void x86_64_protection_init (void);
309 static void create_pagetables(vm_paddr_t *firstaddr);
310 static void pmap_remove_all (vm_page_t m);
311 static boolean_t pmap_testbit (vm_page_t m, int bit);
313 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
314 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
316 static void pmap_pinit_defaults(struct pmap *pmap);
317 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
318 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
321 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
323 if (pv1->pv_pindex < pv2->pv_pindex)
325 if (pv1->pv_pindex > pv2->pv_pindex)
330 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
331 pv_entry_compare, vm_pindex_t, pv_pindex);
335 pmap_page_stats_adding(vm_page_t m)
337 globaldata_t gd = mycpu;
339 if (TAILQ_EMPTY(&m->md.pv_list)) {
340 ++gd->gd_vmtotal.t_arm;
341 } else if (TAILQ_FIRST(&m->md.pv_list) ==
342 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
343 ++gd->gd_vmtotal.t_armshr;
344 ++gd->gd_vmtotal.t_avmshr;
346 ++gd->gd_vmtotal.t_avmshr;
352 pmap_page_stats_deleting(vm_page_t m)
354 globaldata_t gd = mycpu;
356 if (TAILQ_EMPTY(&m->md.pv_list)) {
357 --gd->gd_vmtotal.t_arm;
358 } else if (TAILQ_FIRST(&m->md.pv_list) ==
359 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
360 --gd->gd_vmtotal.t_armshr;
361 --gd->gd_vmtotal.t_avmshr;
363 --gd->gd_vmtotal.t_avmshr;
368 * This is an ineligent crowbar to prevent heavily threaded programs
369 * from creating long live-locks in the pmap code when pmap_mmu_optimize
370 * is enabled. Without it a pmap-local page table page can wind up being
371 * constantly created and destroyed (without injury, but also without
372 * progress) as the optimization tries to switch to the object's shared page
376 pmap_softwait(pmap_t pmap)
378 while (pmap->pm_softhold) {
379 tsleep_interlock(&pmap->pm_softhold, 0);
380 if (pmap->pm_softhold)
381 tsleep(&pmap->pm_softhold, PINTERLOCKED, "mmopt", 0);
386 pmap_softhold(pmap_t pmap)
388 while (atomic_swap_int(&pmap->pm_softhold, 1) == 1) {
389 tsleep_interlock(&pmap->pm_softhold, 0);
390 if (atomic_swap_int(&pmap->pm_softhold, 1) == 1)
391 tsleep(&pmap->pm_softhold, PINTERLOCKED, "mmopt", 0);
396 pmap_softdone(pmap_t pmap)
398 atomic_swap_int(&pmap->pm_softhold, 0);
399 wakeup(&pmap->pm_softhold);
403 * Move the kernel virtual free pointer to the next
404 * 2MB. This is used to help improve performance
405 * by using a large (2MB) page for much of the kernel
406 * (.text, .data, .bss)
410 pmap_kmem_choose(vm_offset_t addr)
412 vm_offset_t newaddr = addr;
414 newaddr = roundup2(addr, NBPDR);
419 * Returns the pindex of a page table entry (representing a terminal page).
420 * There are NUPTE_TOTAL page table entries possible (a huge number)
422 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
423 * We want to properly translate negative KVAs.
427 pmap_pte_pindex(vm_offset_t va)
429 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
433 * Returns the pindex of a page table.
437 pmap_pt_pindex(vm_offset_t va)
439 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
443 * Returns the pindex of a page directory.
447 pmap_pd_pindex(vm_offset_t va)
449 return (NUPTE_TOTAL + NUPT_TOTAL +
450 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
455 pmap_pdp_pindex(vm_offset_t va)
457 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
458 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
463 pmap_pml4_pindex(void)
465 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
469 * Return various clipped indexes for a given VA
471 * Returns the index of a pt in a page directory, representing a page
476 pmap_pt_index(vm_offset_t va)
478 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
482 * Returns the index of a pd in a page directory page, representing a page
487 pmap_pd_index(vm_offset_t va)
489 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
493 * Returns the index of a pdp in the pml4 table, representing a page
498 pmap_pdp_index(vm_offset_t va)
500 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
504 * Locate the requested pt_entry
508 pv_entry_lookup(pmap_t pmap, vm_pindex_t pindex)
512 if (pindex < pmap_pt_pindex(0))
513 pv = pmap->pm_pvhint_pte;
514 else if (pindex < pmap_pd_pindex(0))
515 pv = pmap->pm_pvhint_pt;
519 if (pv == NULL || pv->pv_pmap != pmap) {
520 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
522 } else if (pv->pv_pindex != pindex) {
523 pv = pv_entry_rb_tree_RB_LOOKUP_REL(&pmap->pm_pvroot,
532 * Super fast pmap_pte routine best used when scanning the pv lists.
533 * This eliminates many course-grained invltlb calls. Note that many of
534 * the pv list scans are across different pmaps and it is very wasteful
535 * to do an entire invltlb when checking a single mapping.
537 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
541 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
543 return pmap_pte(pmap, va);
547 * The placemarker hash must be broken up into four zones so lock
548 * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
550 * Placemarkers are used to 'lock' page table indices that do not have
551 * a pv_entry. This allows the pmap to support managed and unmanaged
552 * pages and shared page tables.
554 #define PM_PLACE_BASE (PM_PLACEMARKS >> 2)
558 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
562 if (pindex < pmap_pt_pindex(0)) /* zone 0 - PTE */
564 else if (pindex < pmap_pd_pindex(0)) /* zone 1 - PT */
566 else if (pindex < pmap_pdp_pindex(0)) /* zone 2 - PD */
567 hi = PM_PLACE_BASE << 1;
568 else /* zone 3 - PDP (and PML4E) */
569 hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
570 hi += pindex & (PM_PLACE_BASE - 1);
572 return (&pmap->pm_placemarks[hi]);
577 * Generic procedure to index a pte from a pt, pd, or pdp.
579 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
580 * a page table page index but is instead of PV lookup index.
584 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
588 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
589 return(&pte[pindex]);
593 * Return pointer to PDP slot in the PML4
597 pmap_pdp(pmap_t pmap, vm_offset_t va)
599 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
603 * Return pointer to PD slot in the PDP given a pointer to the PDP
607 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
611 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
612 return (&pd[pmap_pd_index(va)]);
616 * Return pointer to PD slot in the PDP.
620 pmap_pd(pmap_t pmap, vm_offset_t va)
624 pdp = pmap_pdp(pmap, va);
625 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
627 return (pmap_pdp_to_pd(*pdp, va));
631 * Return pointer to PT slot in the PD given a pointer to the PD
635 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
639 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
640 return (&pt[pmap_pt_index(va)]);
644 * Return pointer to PT slot in the PD
646 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
647 * so we cannot lookup the PD via the PDP. Instead we
648 * must look it up via the pmap.
652 pmap_pt(pmap_t pmap, vm_offset_t va)
656 vm_pindex_t pd_pindex;
659 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
660 pd_pindex = pmap_pd_pindex(va);
661 spin_lock_shared(&pmap->pm_spin);
662 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
663 if (pv == NULL || pv->pv_m == NULL) {
664 spin_unlock_shared(&pmap->pm_spin);
667 phys = VM_PAGE_TO_PHYS(pv->pv_m);
668 spin_unlock_shared(&pmap->pm_spin);
669 return (pmap_pd_to_pt(phys, va));
671 pd = pmap_pd(pmap, va);
672 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
674 return (pmap_pd_to_pt(*pd, va));
679 * Return pointer to PTE slot in the PT given a pointer to the PT
683 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
687 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
688 return (&pte[pmap_pte_index(va)]);
692 * Return pointer to PTE slot in the PT
696 pmap_pte(pmap_t pmap, vm_offset_t va)
700 pt = pmap_pt(pmap, va);
701 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
703 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
704 return ((pt_entry_t *)pt);
705 return (pmap_pt_to_pte(*pt, va));
709 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
710 * the PT layer. This will speed up core pmap operations considerably.
712 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
713 * must be in a known associated state (typically by being locked when
714 * the pmap spinlock isn't held). We allow the race for that case.
716 * NOTE: pm_pvhint* is only accessed (read) with the spin-lock held, using
717 * cpu_ccfence() to prevent compiler optimizations from reloading the
722 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
724 if (pindex < pmap_pt_pindex(0)) {
726 pv->pv_pmap->pm_pvhint_pte = pv;
727 } else if (pindex < pmap_pd_pindex(0)) {
729 pv->pv_pmap->pm_pvhint_pt = pv;
735 * Return address of PT slot in PD (KVM only)
737 * Cannot be used for user page tables because it might interfere with
738 * the shared page-table-page optimization (pmap_mmu_optimize).
742 vtopt(vm_offset_t va)
744 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
745 NPML4EPGSHIFT)) - 1);
747 return (PDmap + ((va >> PDRSHIFT) & mask));
751 * KVM - return address of PTE slot in PT
755 vtopte(vm_offset_t va)
757 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
758 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
760 return (PTmap + ((va >> PAGE_SHIFT) & mask));
764 * Returns the physical address translation from va for a user address.
765 * (vm_paddr_t)-1 is returned on failure.
768 uservtophys(vm_offset_t va)
770 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
771 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
776 pmap = vmspace_pmap(mycpu->gd_curthread->td_lwp->lwp_vmspace);
778 if (va < VM_MAX_USER_ADDRESS) {
779 pte = kreadmem64(PTmap + ((va >> PAGE_SHIFT) & mask));
780 if (pte & pmap->pmap_bits[PG_V_IDX])
781 pa = (pte & PG_FRAME) | (va & PAGE_MASK);
787 allocpages(vm_paddr_t *firstaddr, long n)
792 bzero((void *)ret, n * PAGE_SIZE);
793 *firstaddr += n * PAGE_SIZE;
799 create_pagetables(vm_paddr_t *firstaddr)
801 long i; /* must be 64 bits */
808 * We are running (mostly) V=P at this point
810 * Calculate how many 1GB PD entries in our PDP pages are needed
811 * for the DMAP. This is only allocated if the system does not
812 * support 1GB pages. Otherwise ndmpdp is simply a count of
813 * the number of 1G terminal entries in our PDP pages are needed.
815 * NOTE: Maxmem is in pages
817 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
818 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
820 KKASSERT(ndmpdp <= NDMPML4E * NPML4EPG);
823 * Starting at KERNBASE - map all 2G worth of page table pages.
824 * KERNBASE is offset -2G from the end of kvm. This will accomodate
825 * all KVM allocations above KERNBASE, including the SYSMAPs below.
827 * We do this by allocating 2*512 PT pages. Each PT page can map
828 * 2MB, for 2GB total.
830 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
833 * Starting at the beginning of kvm (VM_MIN_KERNEL_ADDRESS),
834 * Calculate how many page table pages we need to preallocate
835 * for early vm_map allocations.
837 * A few extra won't hurt, they will get used up in the running
843 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
844 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
845 nkpt_phys += 128; /* a few extra */
848 * The highest value nkpd_phys can be set to is
849 * NKPDPE - (NPDPEPG - KPDPI) (i.e. NKPDPE - 2).
851 * Doing so would cause all PD pages to be pre-populated for
852 * a maximal KVM space (approximately 16*512 pages, or 32MB.
853 * We can save memory by not doing this.
855 nkpd_phys = (nkpt_phys + NPDPEPG - 1) / NPDPEPG;
860 * Normally NKPML4E=1-16 (1-16 kernel PDP page)
861 * Normally NKPDPE= NKPML4E*512-1 (511 min kernel PD pages)
863 * Only allocate enough PD pages
864 * NOTE: We allocate all kernel PD pages up-front, typically
865 * ~511G of KVM, requiring 511 PD pages.
867 KPTbase = allocpages(firstaddr, nkpt_base); /* KERNBASE to end */
868 KPTphys = allocpages(firstaddr, nkpt_phys); /* KVA start */
869 KPML4phys = allocpages(firstaddr, 1); /* recursive PML4 map */
870 KPDPphys = allocpages(firstaddr, NKPML4E); /* kernel PDP pages */
871 KPDphys = allocpages(firstaddr, nkpd_phys); /* kernel PD pages */
874 * Alloc PD pages for the area starting at KERNBASE.
876 KPDbase = allocpages(firstaddr, NPDPEPG - KPDPI);
881 DMPDPphys = allocpages(firstaddr, NDMPML4E);
882 if ((amd_feature & AMDID_PAGE1GB) == 0)
883 DMPDphys = allocpages(firstaddr, ndmpdp);
884 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
887 * Fill in the underlying page table pages for the area around
888 * KERNBASE. This remaps low physical memory to KERNBASE.
890 * Read-only from zero to physfree
891 * XXX not fully used, underneath 2M pages
893 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
894 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
895 ((pt_entry_t *)KPTbase)[i] |=
896 pmap_bits_default[PG_RW_IDX] |
897 pmap_bits_default[PG_V_IDX] |
898 pmap_bits_default[PG_G_IDX];
902 * Now map the initial kernel page tables. One block of page
903 * tables is placed at the beginning of kernel virtual memory,
904 * and another block is placed at KERNBASE to map the kernel binary,
905 * data, bss, and initial pre-allocations.
907 for (i = 0; i < nkpt_base; i++) {
908 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
909 ((pd_entry_t *)KPDbase)[i] |=
910 pmap_bits_default[PG_RW_IDX] |
911 pmap_bits_default[PG_V_IDX];
913 for (i = 0; i < nkpt_phys; i++) {
914 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
915 ((pd_entry_t *)KPDphys)[i] |=
916 pmap_bits_default[PG_RW_IDX] |
917 pmap_bits_default[PG_V_IDX];
921 * Map from zero to end of allocations using 2M pages as an
922 * optimization. This will bypass some of the KPTBase pages
923 * above in the KERNBASE area.
925 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
926 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
927 ((pd_entry_t *)KPDbase)[i] |=
928 pmap_bits_default[PG_RW_IDX] |
929 pmap_bits_default[PG_V_IDX] |
930 pmap_bits_default[PG_PS_IDX] |
931 pmap_bits_default[PG_G_IDX];
935 * Load PD addresses into the PDP pages for primary KVA space to
936 * cover existing page tables. PD's for KERNBASE are handled in
939 * expected to pre-populate all of its PDs. See NKPDPE in vmparam.h.
941 for (i = 0; i < nkpd_phys; i++) {
942 ((pdp_entry_t *)KPDPphys)[NKPML4E * NPDPEPG - NKPDPE + i] =
943 KPDphys + (i << PAGE_SHIFT);
944 ((pdp_entry_t *)KPDPphys)[NKPML4E * NPDPEPG - NKPDPE + i] |=
945 pmap_bits_default[PG_RW_IDX] |
946 pmap_bits_default[PG_V_IDX] |
947 pmap_bits_default[PG_U_IDX];
951 * Load PDs for KERNBASE to the end
953 i = (NKPML4E - 1) * NPDPEPG + KPDPI;
954 for (j = 0; j < NPDPEPG - KPDPI; ++j) {
955 ((pdp_entry_t *)KPDPphys)[i + j] =
956 KPDbase + (j << PAGE_SHIFT);
957 ((pdp_entry_t *)KPDPphys)[i + j] |=
958 pmap_bits_default[PG_RW_IDX] |
959 pmap_bits_default[PG_V_IDX] |
960 pmap_bits_default[PG_U_IDX];
964 * Now set up the direct map space using either 2MB or 1GB pages
965 * Preset PG_M and PG_A because demotion expects it.
967 * When filling in entries in the PD pages make sure any excess
968 * entries are set to zero as we allocated enough PD pages
970 if ((amd_feature & AMDID_PAGE1GB) == 0) {
971 for (i = 0; i < NPDEPG * ndmpdp; i++) {
972 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
973 ((pd_entry_t *)DMPDphys)[i] |=
974 pmap_bits_default[PG_RW_IDX] |
975 pmap_bits_default[PG_V_IDX] |
976 pmap_bits_default[PG_PS_IDX] |
977 pmap_bits_default[PG_G_IDX] |
978 pmap_bits_default[PG_M_IDX] |
979 pmap_bits_default[PG_A_IDX];
983 * And the direct map space's PDP
985 for (i = 0; i < ndmpdp; i++) {
986 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
988 ((pdp_entry_t *)DMPDPphys)[i] |=
989 pmap_bits_default[PG_RW_IDX] |
990 pmap_bits_default[PG_V_IDX] |
991 pmap_bits_default[PG_U_IDX];
994 for (i = 0; i < ndmpdp; i++) {
995 ((pdp_entry_t *)DMPDPphys)[i] =
996 (vm_paddr_t)i << PDPSHIFT;
997 ((pdp_entry_t *)DMPDPphys)[i] |=
998 pmap_bits_default[PG_RW_IDX] |
999 pmap_bits_default[PG_V_IDX] |
1000 pmap_bits_default[PG_PS_IDX] |
1001 pmap_bits_default[PG_G_IDX] |
1002 pmap_bits_default[PG_M_IDX] |
1003 pmap_bits_default[PG_A_IDX];
1007 /* And recursively map PML4 to itself in order to get PTmap */
1008 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
1009 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
1010 pmap_bits_default[PG_RW_IDX] |
1011 pmap_bits_default[PG_V_IDX] |
1012 pmap_bits_default[PG_U_IDX];
1015 * Connect the Direct Map slots up to the PML4
1017 for (j = 0; j < NDMPML4E; ++j) {
1018 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
1019 (DMPDPphys + ((vm_paddr_t)j << PAGE_SHIFT)) |
1020 pmap_bits_default[PG_RW_IDX] |
1021 pmap_bits_default[PG_V_IDX] |
1022 pmap_bits_default[PG_U_IDX];
1026 * Connect the KVA slot up to the PML4
1028 for (j = 0; j < NKPML4E; ++j) {
1029 ((pdp_entry_t *)KPML4phys)[KPML4I + j] =
1030 KPDPphys + ((vm_paddr_t)j << PAGE_SHIFT);
1031 ((pdp_entry_t *)KPML4phys)[KPML4I + j] |=
1032 pmap_bits_default[PG_RW_IDX] |
1033 pmap_bits_default[PG_V_IDX] |
1034 pmap_bits_default[PG_U_IDX];
1041 * Bootstrap the system enough to run with virtual memory.
1043 * On x86_64 this is called after mapping has already been enabled
1044 * and just syncs the pmap module with what has already been done.
1045 * [We can't call it easily with mapping off since the kernel is not
1046 * mapped with PA == VA, hence we would have to relocate every address
1047 * from the linked base (virtual) address "KERNBASE" to the actual
1048 * (physical) address starting relative to 0]
1051 pmap_bootstrap(vm_paddr_t *firstaddr)
1057 KvaStart = VM_MIN_KERNEL_ADDRESS;
1058 KvaEnd = VM_MAX_KERNEL_ADDRESS;
1059 KvaSize = KvaEnd - KvaStart;
1061 avail_start = *firstaddr;
1064 * Create an initial set of page tables to run the kernel in.
1066 create_pagetables(firstaddr);
1068 virtual2_start = KvaStart;
1069 virtual2_end = PTOV_OFFSET;
1071 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
1072 virtual_start = pmap_kmem_choose(virtual_start);
1074 virtual_end = VM_MAX_KERNEL_ADDRESS;
1076 /* XXX do %cr0 as well */
1077 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
1078 load_cr3(KPML4phys);
1081 * Initialize protection array.
1083 x86_64_protection_init();
1086 * The kernel's pmap is statically allocated so we don't have to use
1087 * pmap_create, which is unlikely to work correctly at this part of
1088 * the boot sequence (XXX and which no longer exists).
1090 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
1091 kernel_pmap.pm_count = 1;
1092 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
1093 RB_INIT(&kernel_pmap.pm_pvroot);
1094 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
1095 for (i = 0; i < PM_PLACEMARKS; ++i)
1096 kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
1099 * Reserve some special page table entries/VA space for temporary
1102 #define SYSMAP(c, p, v, n) \
1103 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
1109 * CMAP1/CMAP2 are used for zeroing and copying pages.
1111 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
1116 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
1119 * ptvmmap is used for reading arbitrary physical pages via
1122 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
1125 * msgbufp is used to map the system message buffer.
1126 * XXX msgbufmap is not used.
1128 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
1129 atop(round_page(MSGBUF_SIZE)))
1132 virtual_start = pmap_kmem_choose(virtual_start);
1137 * PG_G is terribly broken on SMP because we IPI invltlb's in some
1138 * cases rather then invl1pg. Actually, I don't even know why it
1139 * works under UP because self-referential page table mappings
1145 /* Initialize the PAT MSR */
1147 pmap_pinit_defaults(&kernel_pmap);
1149 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1150 &pmap_fast_kernel_cpusync);
1155 * Setup the PAT MSR.
1164 * Default values mapping PATi,PCD,PWT bits at system reset.
1165 * The default values effectively ignore the PATi bit by
1166 * repeating the encodings for 0-3 in 4-7, and map the PCD
1167 * and PWT bit combinations to the expected PAT types.
1169 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
1170 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
1171 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
1172 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
1173 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
1174 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
1175 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
1176 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
1177 pat_pte_index[PAT_WRITE_BACK] = 0;
1178 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
1179 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
1180 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
1181 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1182 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1184 if (cpu_feature & CPUID_PAT) {
1186 * If we support the PAT then set-up entries for
1187 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1190 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1191 PAT_VALUE(5, PAT_WRITE_PROTECTED);
1192 pat_msr = (pat_msr & ~PAT_MASK(6)) |
1193 PAT_VALUE(6, PAT_WRITE_COMBINING);
1194 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1195 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PCD;
1198 * Then enable the PAT
1203 load_cr4(cr4 & ~CR4_PGE);
1205 /* Disable caches (CD = 1, NW = 0). */
1207 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1209 /* Flushes caches and TLBs. */
1213 /* Update PAT and index table. */
1214 wrmsr(MSR_PAT, pat_msr);
1216 /* Flush caches and TLBs again. */
1220 /* Restore caches and PGE. */
1228 * Set 4mb pdir for mp startup
1233 if (cpu_feature & CPUID_PSE) {
1234 load_cr4(rcr4() | CR4_PSE);
1235 if (mycpu->gd_cpuid == 0) /* only on BSP */
1241 * Initialize the pmap module.
1242 * Called by vm_init, to initialize any structures that the pmap
1243 * system needs to map virtual memory.
1244 * pmap_init has been enhanced to support in a fairly consistant
1245 * way, discontiguous physical memory.
1250 vm_pindex_t initial_pvs;
1254 * Allocate memory for random pmap data structures. Includes the
1258 for (i = 0; i < vm_page_array_size; i++) {
1261 m = &vm_page_array[i];
1262 TAILQ_INIT(&m->md.pv_list);
1266 * init the pv free list
1268 initial_pvs = vm_page_array_size;
1269 if (initial_pvs < MINPV)
1270 initial_pvs = MINPV;
1271 pvzone = &pvzone_store;
1272 pvinit = (void *)kmem_alloc(&kernel_map,
1273 initial_pvs * sizeof (struct pv_entry),
1275 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1276 pvinit, initial_pvs);
1279 * Now it is safe to enable pv_table recording.
1281 pmap_initialized = TRUE;
1285 * Initialize the address space (zone) for the pv_entries. Set a
1286 * high water mark so that the system can recover from excessive
1287 * numbers of pv entries.
1292 vm_pindex_t shpgperproc = PMAP_SHPGPERPROC;
1293 vm_pindex_t entry_max;
1295 TUNABLE_LONG_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1296 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1297 TUNABLE_LONG_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1298 pv_entry_high_water = 9 * (pv_entry_max / 10);
1301 * Subtract out pages already installed in the zone (hack)
1303 entry_max = pv_entry_max - vm_page_array_size;
1307 zinitna(pvzone, NULL, 0, entry_max, ZONE_INTERRUPT);
1310 * Enable dynamic deletion of empty higher-level page table pages
1311 * by default only if system memory is < 8GB (use 7GB for slop).
1312 * This can save a little memory, but imposes significant
1313 * performance overhead for things like bulk builds, and for programs
1314 * which do a lot of memory mapping and memory unmapping.
1316 if (pmap_dynamic_delete < 0) {
1317 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1318 pmap_dynamic_delete = 1;
1320 pmap_dynamic_delete = 0;
1325 * Typically used to initialize a fictitious page by vm/device_pager.c
1328 pmap_page_init(struct vm_page *m)
1331 TAILQ_INIT(&m->md.pv_list);
1334 /***************************************************
1335 * Low level helper routines.....
1336 ***************************************************/
1339 * this routine defines the region(s) of memory that should
1340 * not be tested for the modified bit.
1344 pmap_track_modified(vm_pindex_t pindex)
1346 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1347 if ((va < clean_sva) || (va >= clean_eva))
1354 * Extract the physical page address associated with the map/VA pair.
1355 * The page must be wired for this to work reliably.
1358 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1365 if (va >= VM_MAX_USER_ADDRESS) {
1367 * Kernel page directories might be direct-mapped and
1368 * there is typically no PV tracking of pte's
1372 pt = pmap_pt(pmap, va);
1373 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1374 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1375 rtval = *pt & PG_PS_FRAME;
1376 rtval |= va & PDRMASK;
1378 ptep = pmap_pt_to_pte(*pt, va);
1379 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1380 rtval = *ptep & PG_FRAME;
1381 rtval |= va & PAGE_MASK;
1389 * User pages currently do not direct-map the page directory
1390 * and some pages might not used managed PVs. But all PT's
1393 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1395 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1396 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1397 rtval = *ptep & PG_FRAME;
1398 rtval |= va & PAGE_MASK;
1401 *handlep = pt_pv; /* locked until done */
1404 } else if (handlep) {
1412 pmap_extract_done(void *handle)
1415 pv_put((pv_entry_t)handle);
1419 * Similar to extract but checks protections, SMP-friendly short-cut for
1420 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1421 * fall-through to the real fault code. Does not work with HVM page
1424 * if busyp is NULL the returned page, if not NULL, is held (and not busied).
1426 * If busyp is not NULL and this function sets *busyp non-zero, the returned
1427 * page is busied (and not held).
1429 * If busyp is not NULL and this function sets *busyp to zero, the returned
1430 * page is held (and not busied).
1432 * If VM_PROT_WRITE is set in prot, and the pte is already writable, the
1433 * returned page will be dirtied. If the pte is not already writable NULL
1434 * is returned. In otherwords, if the bit is set and a vm_page_t is returned,
1435 * any COW will already have happened and that page can be written by the
1438 * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1442 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1445 va < VM_MAX_USER_ADDRESS &&
1446 (pmap->pm_flags & PMAP_HVM) == 0) {
1454 req = pmap->pmap_bits[PG_V_IDX] |
1455 pmap->pmap_bits[PG_U_IDX];
1456 if (prot & VM_PROT_WRITE)
1457 req |= pmap->pmap_bits[PG_RW_IDX];
1459 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1462 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1463 if ((*ptep & req) != req) {
1467 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1468 if (pte_pv && error == 0) {
1470 if (prot & VM_PROT_WRITE) {
1471 /* interlocked by presence of pv_entry */
1475 if (prot & VM_PROT_WRITE) {
1476 if (vm_page_busy_try(m, TRUE))
1487 } else if (pte_pv) {
1491 /* error, since we didn't request a placemarker */
1502 * Extract the physical page address associated kernel virtual address.
1505 pmap_kextract(vm_offset_t va)
1507 pd_entry_t pt; /* pt entry in pd */
1510 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1511 pa = DMAP_TO_PHYS(va);
1514 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1515 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1518 * Beware of a concurrent promotion that changes the
1519 * PDE at this point! For example, vtopte() must not
1520 * be used to access the PTE because it would use the
1521 * new PDE. It is, however, safe to use the old PDE
1522 * because the page table page is preserved by the
1525 pa = *pmap_pt_to_pte(pt, va);
1526 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1532 /***************************************************
1533 * Low level mapping routines.....
1534 ***************************************************/
1537 * Routine: pmap_kenter
1539 * Add a wired page to the KVA
1540 * NOTE! note that in order for the mapping to take effect -- you
1541 * should do an invltlb after doing the pmap_kenter().
1544 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1550 kernel_pmap.pmap_bits[PG_RW_IDX] |
1551 kernel_pmap.pmap_bits[PG_V_IDX];
1555 pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1559 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1566 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1567 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1568 * (caller can conditionalize calling smp_invltlb()).
1571 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1577 npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1578 kernel_pmap.pmap_bits[PG_V_IDX];
1587 atomic_swap_long(ptep, npte);
1588 cpu_invlpg((void *)va);
1594 * Enter addresses into the kernel pmap but don't bother
1595 * doing any tlb invalidations. Caller will do a rollup
1596 * invalidation via pmap_rollup_inval().
1599 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1606 kernel_pmap.pmap_bits[PG_RW_IDX] |
1607 kernel_pmap.pmap_bits[PG_V_IDX];
1616 atomic_swap_long(ptep, npte);
1617 cpu_invlpg((void *)va);
1623 * remove a page from the kernel pagetables
1626 pmap_kremove(vm_offset_t va)
1631 pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1635 pmap_kremove_quick(vm_offset_t va)
1640 (void)pte_load_clear(ptep);
1641 cpu_invlpg((void *)va);
1645 * Remove addresses from the kernel pmap but don't bother
1646 * doing any tlb invalidations. Caller will do a rollup
1647 * invalidation via pmap_rollup_inval().
1650 pmap_kremove_noinval(vm_offset_t va)
1655 (void)pte_load_clear(ptep);
1659 * XXX these need to be recoded. They are not used in any critical path.
1662 pmap_kmodify_rw(vm_offset_t va)
1664 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1665 cpu_invlpg((void *)va);
1670 pmap_kmodify_nc(vm_offset_t va)
1672 atomic_set_long(vtopte(va), PG_N);
1673 cpu_invlpg((void *)va);
1678 * Used to map a range of physical addresses into kernel virtual
1679 * address space during the low level boot, typically to map the
1680 * dump bitmap, message buffer, and vm_page_array.
1682 * These mappings are typically made at some pointer after the end of the
1685 * We could return PHYS_TO_DMAP(start) here and not allocate any
1686 * via (*virtp), but then kmem from userland and kernel dumps won't
1687 * have access to the related pointers.
1690 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1693 vm_offset_t va_start;
1695 /*return PHYS_TO_DMAP(start);*/
1700 while (start < end) {
1701 pmap_kenter_quick(va, start);
1709 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1712 * Remove the specified set of pages from the data and instruction caches.
1714 * In contrast to pmap_invalidate_cache_range(), this function does not
1715 * rely on the CPU's self-snoop feature, because it is intended for use
1716 * when moving pages into a different cache domain.
1719 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1721 vm_offset_t daddr, eva;
1724 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1725 (cpu_feature & CPUID_CLFSH) == 0)
1729 for (i = 0; i < count; i++) {
1730 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1731 eva = daddr + PAGE_SIZE;
1732 for (; daddr < eva; daddr += cpu_clflush_line_size)
1740 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1742 KASSERT((sva & PAGE_MASK) == 0,
1743 ("pmap_invalidate_cache_range: sva not page-aligned"));
1744 KASSERT((eva & PAGE_MASK) == 0,
1745 ("pmap_invalidate_cache_range: eva not page-aligned"));
1747 if (cpu_feature & CPUID_SS) {
1748 ; /* If "Self Snoop" is supported, do nothing. */
1750 /* Globally invalidate caches */
1751 cpu_wbinvd_on_all_cpus();
1756 * Invalidate the specified range of virtual memory on all cpus associated
1760 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1762 pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1766 * Add a list of wired pages to the kva. This routine is used for temporary
1767 * kernel mappings such as those found in buffer cache buffer. Page
1768 * modifications and accesses are not tracked or recorded.
1770 * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1771 * semantics as previous mappings may have been zerod without any
1774 * The page *must* be wired.
1776 static __inline void
1777 _pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count, int doinval)
1782 end_va = beg_va + count * PAGE_SIZE;
1784 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1789 pte = VM_PAGE_TO_PHYS(*m) |
1790 kernel_pmap.pmap_bits[PG_RW_IDX] |
1791 kernel_pmap.pmap_bits[PG_V_IDX] |
1792 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1794 atomic_swap_long(ptep, pte);
1798 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1802 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1804 _pmap_qenter(beg_va, m, count, 1);
1808 pmap_qenter_noinval(vm_offset_t beg_va, vm_page_t *m, int count)
1810 _pmap_qenter(beg_va, m, count, 0);
1814 * This routine jerks page mappings from the kernel -- it is meant only
1815 * for temporary mappings such as those found in buffer cache buffers.
1816 * No recording modified or access status occurs.
1818 * MPSAFE, INTERRUPT SAFE (cluster callback)
1821 pmap_qremove(vm_offset_t beg_va, int count)
1826 end_va = beg_va + count * PAGE_SIZE;
1828 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1832 (void)pte_load_clear(pte);
1833 cpu_invlpg((void *)va);
1835 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1839 * This routine removes temporary kernel mappings, only invalidating them
1840 * on the current cpu. It should only be used under carefully controlled
1844 pmap_qremove_quick(vm_offset_t beg_va, int count)
1849 end_va = beg_va + count * PAGE_SIZE;
1851 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1855 (void)pte_load_clear(pte);
1856 cpu_invlpg((void *)va);
1861 * This routine removes temporary kernel mappings *without* invalidating
1862 * the TLB. It can only be used on permanent kva reservations such as those
1863 * found in buffer cache buffers, under carefully controlled circumstances.
1865 * NOTE: Repopulating these KVAs requires unconditional invalidation.
1866 * (pmap_qenter() does unconditional invalidation).
1869 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1874 end_va = beg_va + count * PAGE_SIZE;
1876 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1880 (void)pte_load_clear(pte);
1885 * Create a new thread and optionally associate it with a (new) process.
1886 * NOTE! the new thread's cpu may not equal the current cpu.
1889 pmap_init_thread(thread_t td)
1891 /* enforce pcb placement & alignment */
1892 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1893 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1894 td->td_savefpu = &td->td_pcb->pcb_save;
1895 td->td_sp = (char *)td->td_pcb; /* no -16 */
1899 * This routine directly affects the fork perf for a process.
1902 pmap_init_proc(struct proc *p)
1907 pmap_pinit_defaults(struct pmap *pmap)
1909 bcopy(pmap_bits_default, pmap->pmap_bits,
1910 sizeof(pmap_bits_default));
1911 bcopy(protection_codes, pmap->protection_codes,
1912 sizeof(protection_codes));
1913 bcopy(pat_pte_index, pmap->pmap_cache_bits,
1914 sizeof(pat_pte_index));
1915 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1916 pmap->copyinstr = std_copyinstr;
1917 pmap->copyin = std_copyin;
1918 pmap->copyout = std_copyout;
1919 pmap->fubyte = std_fubyte;
1920 pmap->subyte = std_subyte;
1921 pmap->fuword32 = std_fuword32;
1922 pmap->fuword64 = std_fuword64;
1923 pmap->suword32 = std_suword32;
1924 pmap->suword64 = std_suword64;
1925 pmap->swapu32 = std_swapu32;
1926 pmap->swapu64 = std_swapu64;
1929 * Initialize pmap0/vmspace0.
1931 * On architectures where the kernel pmap is not integrated into the user
1932 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1933 * kernel_pmap should be used to directly access the kernel_pmap.
1936 pmap_pinit0(struct pmap *pmap)
1940 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1942 CPUMASK_ASSZERO(pmap->pm_active);
1943 pmap->pm_pvhint_pt = NULL;
1944 pmap->pm_pvhint_pte = NULL;
1945 RB_INIT(&pmap->pm_pvroot);
1946 spin_init(&pmap->pm_spin, "pmapinit0");
1947 for (i = 0; i < PM_PLACEMARKS; ++i)
1948 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1949 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1950 pmap_pinit_defaults(pmap);
1954 * Initialize a preallocated and zeroed pmap structure,
1955 * such as one in a vmspace structure.
1958 pmap_pinit_simple(struct pmap *pmap)
1963 * Misc initialization
1966 CPUMASK_ASSZERO(pmap->pm_active);
1967 pmap->pm_pvhint_pt = NULL;
1968 pmap->pm_pvhint_pte = NULL;
1969 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1971 pmap_pinit_defaults(pmap);
1974 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1977 if (pmap->pm_pmlpv == NULL) {
1978 RB_INIT(&pmap->pm_pvroot);
1979 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1980 spin_init(&pmap->pm_spin, "pmapinitsimple");
1981 for (i = 0; i < PM_PLACEMARKS; ++i)
1982 pmap->pm_placemarks[i] = PM_NOPLACEMARK;
1987 pmap_pinit(struct pmap *pmap)
1992 if (pmap->pm_pmlpv) {
1993 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1998 pmap_pinit_simple(pmap);
1999 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
2002 * No need to allocate page table space yet but we do need a valid
2003 * page directory table.
2005 if (pmap->pm_pml4 == NULL) {
2007 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
2013 * Allocate the page directory page, which wires it even though
2014 * it isn't being entered into some higher level page table (it
2015 * being the highest level). If one is already cached we don't
2016 * have to do anything.
2018 if ((pv = pmap->pm_pmlpv) == NULL) {
2019 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2020 pmap->pm_pmlpv = pv;
2021 pmap_kenter((vm_offset_t)pmap->pm_pml4,
2022 VM_PAGE_TO_PHYS(pv->pv_m));
2026 * Install DMAP and KMAP.
2028 for (j = 0; j < NDMPML4E; ++j) {
2029 pmap->pm_pml4[DMPML4I + j] =
2030 (DMPDPphys + ((vm_paddr_t)j << PAGE_SHIFT)) |
2031 pmap->pmap_bits[PG_RW_IDX] |
2032 pmap->pmap_bits[PG_V_IDX] |
2033 pmap->pmap_bits[PG_U_IDX];
2035 for (j = 0; j < NKPML4E; ++j) {
2036 pmap->pm_pml4[KPML4I + j] =
2037 (KPDPphys + ((vm_paddr_t)j << PAGE_SHIFT)) |
2038 pmap->pmap_bits[PG_RW_IDX] |
2039 pmap->pmap_bits[PG_V_IDX] |
2040 pmap->pmap_bits[PG_U_IDX];
2044 * install self-referential address mapping entry
2046 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
2047 pmap->pmap_bits[PG_V_IDX] |
2048 pmap->pmap_bits[PG_RW_IDX] |
2049 pmap->pmap_bits[PG_A_IDX] |
2050 pmap->pmap_bits[PG_M_IDX];
2052 KKASSERT(pv->pv_m->flags & PG_MAPPED);
2053 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
2055 KKASSERT(pmap->pm_pml4[255] == 0);
2056 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
2057 KKASSERT(pv->pv_entry.rbe_left == NULL);
2058 KKASSERT(pv->pv_entry.rbe_right == NULL);
2062 * Clean up a pmap structure so it can be physically freed. This routine
2063 * is called by the vmspace dtor function. A great deal of pmap data is
2064 * left passively mapped to improve vmspace management so we have a bit
2065 * of cleanup work to do here.
2068 pmap_puninit(pmap_t pmap)
2073 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
2074 if ((pv = pmap->pm_pmlpv) != NULL) {
2075 if (pv_hold_try(pv) == 0)
2077 KKASSERT(pv == pmap->pm_pmlpv);
2078 p = pmap_remove_pv_page(pv);
2080 pv = NULL; /* safety */
2081 pmap_kremove((vm_offset_t)pmap->pm_pml4);
2082 vm_page_busy_wait(p, FALSE, "pgpun");
2083 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2084 vm_page_unwire(p, 0);
2085 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2088 * XXX eventually clean out PML4 static entries and
2089 * use vm_page_free_zero()
2092 pmap->pm_pmlpv = NULL;
2094 if (pmap->pm_pml4) {
2095 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
2096 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
2097 pmap->pm_pml4 = NULL;
2099 KKASSERT(pmap->pm_stats.resident_count == 0);
2100 KKASSERT(pmap->pm_stats.wired_count == 0);
2104 * This function is now unused (used to add the pmap to the pmap_list)
2107 pmap_pinit2(struct pmap *pmap)
2112 * This routine is called when various levels in the page table need to
2113 * be populated. This routine cannot fail.
2115 * This function returns two locked pv_entry's, one representing the
2116 * requested pv and one representing the requested pv's parent pv. If
2117 * an intermediate page table does not exist it will be created, mapped,
2118 * wired, and the parent page table will be given an additional hold
2119 * count representing the presence of the child pv_entry.
2123 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
2129 vm_pindex_t pt_pindex;
2135 * If the pv already exists and we aren't being asked for the
2136 * parent page table page we can just return it. A locked+held pv
2137 * is returned. The pv will also have a second hold related to the
2138 * pmap association that we don't have to worry about.
2141 pv = pv_alloc(pmap, ptepindex, &isnew);
2142 if (isnew == 0 && pvpp == NULL)
2146 * Special case terminal PVs. These are not page table pages so
2147 * no vm_page is allocated (the caller supplied the vm_page). If
2148 * pvpp is non-NULL we are being asked to also removed the pt_pv
2151 * Note that pt_pv's are only returned for user VAs. We assert that
2152 * a pt_pv is not being requested for kernel VAs. The kernel
2153 * pre-wires all higher-level page tables so don't overload managed
2154 * higher-level page tables on top of it!
2156 if (ptepindex < pmap_pt_pindex(0)) {
2157 if (ptepindex >= NUPTE_USER) {
2158 /* kernel manages this manually for KVM */
2159 KKASSERT(pvpp == NULL);
2161 KKASSERT(pvpp != NULL);
2162 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
2163 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
2165 vm_page_wire_quick(pvp->pv_m);
2172 * The kernel never uses managed PT/PD/PDP pages.
2174 KKASSERT(pmap != &kernel_pmap);
2177 * Non-terminal PVs allocate a VM page to represent the page table,
2178 * so we have to resolve pvp and calculate ptepindex for the pvp
2179 * and then for the page table entry index in the pvp for
2182 if (ptepindex < pmap_pd_pindex(0)) {
2184 * pv is PT, pvp is PD
2186 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2187 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2188 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2193 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2194 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2196 } else if (ptepindex < pmap_pdp_pindex(0)) {
2198 * pv is PD, pvp is PDP
2200 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2203 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2204 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2206 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2207 KKASSERT(pvpp == NULL);
2210 pvp = pmap_allocpte(pmap, ptepindex, NULL);
2216 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2217 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2218 } else if (ptepindex < pmap_pml4_pindex()) {
2220 * pv is PDP, pvp is the root pml4 table
2222 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2227 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2228 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2231 * pv represents the top-level PML4, there is no parent.
2240 * (isnew) is TRUE, pv is not terminal.
2242 * (1) Add a wire count to the parent page table (pvp).
2243 * (2) Allocate a VM page for the page table.
2244 * (3) Enter the VM page into the parent page table.
2246 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2249 vm_page_wire_quick(pvp->pv_m);
2252 m = vm_page_alloc(NULL, pv->pv_pindex,
2253 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2254 VM_ALLOC_INTERRUPT);
2259 vm_page_wire(m); /* wire for mapping in parent */
2260 vm_page_unmanage(m); /* m must be spinunlocked */
2261 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2262 m->valid = VM_PAGE_BITS_ALL;
2264 vm_page_spin_lock(m);
2265 pmap_page_stats_adding(m);
2266 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2268 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2269 vm_page_spin_unlock(m);
2272 * (isnew) is TRUE, pv is not terminal.
2274 * Wire the page into pvp. Bump the resident_count for the pmap.
2275 * There is no pvp for the top level, address the pm_pml4[] array
2278 * If the caller wants the parent we return it, otherwise
2279 * we just put it away.
2281 * No interlock is needed for pte 0 -> non-zero.
2283 * In the situation where *ptep is valid we might have an unmanaged
2284 * page table page shared from another page table which we need to
2285 * unshare before installing our private page table page.
2288 v = VM_PAGE_TO_PHYS(m) |
2289 (pmap->pmap_bits[PG_U_IDX] |
2290 pmap->pmap_bits[PG_RW_IDX] |
2291 pmap->pmap_bits[PG_V_IDX] |
2292 pmap->pmap_bits[PG_A_IDX] |
2293 pmap->pmap_bits[PG_M_IDX]);
2294 ptep = pv_pte_lookup(pvp, ptepindex);
2295 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2299 panic("pmap_allocpte: unexpected pte %p/%d",
2300 pvp, (int)ptepindex);
2302 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, v);
2303 if (vm_page_unwire_quick(
2304 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2305 panic("pmap_allocpte: shared pgtable "
2306 "pg bad wirecount");
2311 pte = atomic_swap_long(ptep, v);
2313 kprintf("install pgtbl mixup 0x%016jx "
2314 "old/new 0x%016jx/0x%016jx\n",
2315 (intmax_t)ptepindex, pte, v);
2322 * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2326 KKASSERT(pvp->pv_m != NULL);
2327 ptep = pv_pte_lookup(pvp, ptepindex);
2328 v = VM_PAGE_TO_PHYS(pv->pv_m) |
2329 (pmap->pmap_bits[PG_U_IDX] |
2330 pmap->pmap_bits[PG_RW_IDX] |
2331 pmap->pmap_bits[PG_V_IDX] |
2332 pmap->pmap_bits[PG_A_IDX] |
2333 pmap->pmap_bits[PG_M_IDX]);
2335 kprintf("mismatched upper level pt %016jx/%016jx\n",
2347 * This version of pmap_allocpte() checks for possible segment optimizations
2348 * that would allow page-table sharing. It can be called for terminal
2349 * page or page table page ptepindex's.
2351 * The function is called with page table page ptepindex's for fictitious
2352 * and unmanaged terminal pages. That is, we don't want to allocate a
2353 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2356 * This function can return a pv and *pvpp associated with the passed in pmap
2357 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2358 * an unmanaged page table page will be entered into the pass in pmap.
2362 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2363 vm_map_entry_t entry, vm_offset_t va)
2368 vm_pindex_t *pt_placemark;
2370 pv_entry_t pte_pv; /* in original or shared pmap */
2371 pv_entry_t pt_pv; /* in original or shared pmap */
2372 pv_entry_t proc_pd_pv; /* in original pmap */
2373 pv_entry_t proc_pt_pv; /* in original pmap */
2374 pv_entry_t xpv; /* PT in shared pmap */
2375 pd_entry_t *pt; /* PT entry in PD of original pmap */
2376 pd_entry_t opte; /* contents of *pt */
2377 pd_entry_t npte; /* contents of *pt */
2382 * Basic tests, require a non-NULL vm_map_entry, require proper
2383 * alignment and type for the vm_map_entry, require that the
2384 * underlying object already be allocated.
2386 * We allow almost any type of object to use this optimization.
2387 * The object itself does NOT have to be sized to a multiple of the
2388 * segment size, but the memory mapping does.
2390 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2391 * won't work as expected.
2393 if (entry == NULL ||
2394 pmap_mmu_optimize == 0 || /* not enabled */
2395 (pmap->pm_flags & PMAP_HVM) || /* special pmap */
2396 ptepindex >= pmap_pd_pindex(0) || /* not terminal or pt */
2397 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2398 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2399 entry->object.vm_object == NULL || /* needs VM object */
2400 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2401 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2402 (entry->offset & SEG_MASK) || /* must be aligned */
2403 (entry->start & SEG_MASK)) {
2404 return(pmap_allocpte(pmap, ptepindex, pvpp));
2408 * Make sure the full segment can be represented.
2410 b = va & ~(vm_offset_t)SEG_MASK;
2411 if (b < entry->start || b + SEG_SIZE > entry->end)
2412 return(pmap_allocpte(pmap, ptepindex, pvpp));
2415 * If the full segment can be represented dive the VM object's
2416 * shared pmap, allocating as required.
2418 object = entry->object.vm_object;
2420 if (entry->protection & VM_PROT_WRITE)
2421 obpmapp = &object->md.pmap_rw;
2423 obpmapp = &object->md.pmap_ro;
2426 if (pmap_enter_debug > 0) {
2428 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2430 va, entry->protection, object,
2432 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2433 entry, entry->start, entry->end);
2438 * We allocate what appears to be a normal pmap but because portions
2439 * of this pmap are shared with other unrelated pmaps we have to
2440 * set pm_active to point to all cpus.
2442 * XXX Currently using pmap_spin to interlock the update, can't use
2443 * vm_object_hold/drop because the token might already be held
2444 * shared OR exclusive and we don't know.
2446 while ((obpmap = *obpmapp) == NULL) {
2447 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2448 pmap_pinit_simple(obpmap);
2449 pmap_pinit2(obpmap);
2450 spin_lock(&pmap_spin);
2451 if (*obpmapp != NULL) {
2455 spin_unlock(&pmap_spin);
2456 pmap_release(obpmap);
2457 pmap_puninit(obpmap);
2458 kfree(obpmap, M_OBJPMAP);
2459 obpmap = *obpmapp; /* safety */
2461 obpmap->pm_active = smp_active_mask;
2462 obpmap->pm_flags |= PMAP_SEGSHARED;
2464 spin_unlock(&pmap_spin);
2469 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2470 * pte/pt using the shared pmap from the object but also adjust
2471 * the process pmap's page table page as a side effect.
2475 * Resolve the terminal PTE and PT in the shared pmap. This is what
2476 * we will return. This is true if ptepindex represents a terminal
2477 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2481 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2484 if (ptepindex >= pmap_pt_pindex(0))
2490 * Resolve the PD in the process pmap so we can properly share the
2491 * page table page. Lock order is bottom-up (leaf first)!
2493 * NOTE: proc_pt_pv can be NULL.
2495 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b), &pt_placemark);
2496 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2498 if (pmap_enter_debug > 0) {
2500 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2502 (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2509 * xpv is the page table page pv from the shared object
2510 * (for convenience), from above.
2512 * Calculate the pte value for the PT to load into the process PD.
2513 * If we have to change it we must properly dispose of the previous
2516 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2517 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2518 (pmap->pmap_bits[PG_U_IDX] |
2519 pmap->pmap_bits[PG_RW_IDX] |
2520 pmap->pmap_bits[PG_V_IDX] |
2521 pmap->pmap_bits[PG_A_IDX] |
2522 pmap->pmap_bits[PG_M_IDX]);
2525 * Dispose of previous page table page if it was local to the
2526 * process pmap. If the old pt is not empty we cannot dispose of it
2527 * until we clean it out. This case should not arise very often so
2528 * it is not optimized.
2530 * Leave pt_pv and pte_pv (in our object pmap) locked and intact
2534 pmap_inval_bulk_t bulk;
2536 if (proc_pt_pv->pv_m->wire_count != 1) {
2538 * The page table has a bunch of stuff in it
2539 * which we have to scrap.
2541 if (softhold == 0) {
2543 pmap_softhold(pmap);
2548 va & ~(vm_offset_t)SEG_MASK,
2549 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2552 * The page table is empty and can be destroyed.
2553 * However, doing so leaves the pt slot unlocked,
2554 * so we have to loop-up to handle any races until
2555 * we get a NULL proc_pt_pv and a proper pt_placemark.
2557 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2558 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2559 pmap_inval_bulk_flush(&bulk);
2566 * Handle remaining cases. We are holding pt_placemark to lock
2567 * the page table page in the primary pmap while we manipulate
2571 atomic_swap_long(pt, npte);
2572 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2573 vm_page_wire_quick(proc_pd_pv->pv_m); /* proc pd for sh pt */
2574 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2575 } else if (*pt != npte) {
2576 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2579 opte = pte_load_clear(pt);
2580 KKASSERT(opte && opte != npte);
2584 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2587 * Clean up opte, bump the wire_count for the process
2588 * PD page representing the new entry if it was
2591 * If the entry was not previously empty and we have
2592 * a PT in the proc pmap then opte must match that
2593 * pt. The proc pt must be retired (this is done
2594 * later on in this procedure).
2596 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2599 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2600 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2601 if (vm_page_unwire_quick(m)) {
2602 panic("pmap_allocpte_seg: "
2603 "bad wire count %p",
2609 pmap_softdone(pmap);
2612 * Remove our earmark on the page table page.
2614 pv_placemarker_wakeup(pmap, pt_placemark);
2617 * The existing process page table was replaced and must be destroyed
2630 * Release any resources held by the given physical map.
2632 * Called when a pmap initialized by pmap_pinit is being released. Should
2633 * only be called if the map contains no valid mappings.
2635 struct pmap_release_info {
2641 static int pmap_release_callback(pv_entry_t pv, void *data);
2644 pmap_release(struct pmap *pmap)
2646 struct pmap_release_info info;
2648 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2649 ("pmap still active! %016jx",
2650 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2653 * There is no longer a pmap_list, if there were we would remove the
2654 * pmap from it here.
2658 * Pull pv's off the RB tree in order from low to high and release
2666 spin_lock(&pmap->pm_spin);
2667 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2668 pmap_release_callback, &info);
2669 spin_unlock(&pmap->pm_spin);
2673 } while (info.retry);
2677 * One resident page (the pml4 page) should remain.
2678 * No wired pages should remain.
2681 if (pmap->pm_stats.resident_count !=
2682 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1) ||
2683 pmap->pm_stats.wired_count != 0) {
2684 kprintf("fatal pmap problem - pmap %p flags %08x "
2685 "rescnt=%jd wirecnt=%jd\n",
2688 pmap->pm_stats.resident_count,
2689 pmap->pm_stats.wired_count);
2690 tsleep(pmap, 0, "DEAD", 0);
2693 KKASSERT(pmap->pm_stats.resident_count ==
2694 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2695 KKASSERT(pmap->pm_stats.wired_count == 0);
2700 * Called from low to high. We must cache the proper parent pv so we
2701 * can adjust its wired count.
2704 pmap_release_callback(pv_entry_t pv, void *data)
2706 struct pmap_release_info *info = data;
2707 pmap_t pmap = info->pmap;
2712 * Acquire a held and locked pv, check for release race
2714 pindex = pv->pv_pindex;
2715 if (info->pvp == pv) {
2716 spin_unlock(&pmap->pm_spin);
2718 } else if (pv_hold_try(pv)) {
2719 spin_unlock(&pmap->pm_spin);
2721 spin_unlock(&pmap->pm_spin);
2725 spin_lock(&pmap->pm_spin);
2729 KKASSERT(pv->pv_pmap == pmap && pindex == pv->pv_pindex);
2731 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2733 * I am PTE, parent is PT
2735 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2736 pindex += NUPTE_TOTAL;
2737 } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2739 * I am PT, parent is PD
2741 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2742 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2743 } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2745 * I am PD, parent is PDP
2747 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2749 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2750 } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2752 * I am PDP, parent is PML4 (there's only one)
2755 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2756 NUPD_TOTAL) >> NPML4EPGSHIFT;
2757 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2759 pindex = pmap_pml4_pindex();
2771 if (info->pvp && info->pvp->pv_pindex != pindex) {
2775 if (info->pvp == NULL)
2776 info->pvp = pv_get(pmap, pindex, NULL);
2783 r = pmap_release_pv(pv, info->pvp, NULL);
2784 spin_lock(&pmap->pm_spin);
2790 * Called with held (i.e. also locked) pv. This function will dispose of
2791 * the lock along with the pv.
2793 * If the caller already holds the locked parent page table for pv it
2794 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2795 * pass NULL for pvp.
2798 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2803 * The pmap is currently not spinlocked, pv is held+locked.
2804 * Remove the pv's page from its parent's page table. The
2805 * parent's page table page's wire_count will be decremented.
2807 * This will clean out the pte at any level of the page table.
2808 * If smp != 0 all cpus are affected.
2810 * Do not tear-down recursively, its faster to just let the
2811 * release run its course.
2813 pmap_remove_pv_pte(pv, pvp, bulk, 0);
2816 * Terminal pvs are unhooked from their vm_pages. Because
2817 * terminal pages aren't page table pages they aren't wired
2818 * by us, so we have to be sure not to unwire them either.
2820 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2821 pmap_remove_pv_page(pv);
2826 * We leave the top-level page table page cached, wired, and
2827 * mapped in the pmap until the dtor function (pmap_puninit())
2830 * Since we are leaving the top-level pv intact we need
2831 * to break out of what would otherwise be an infinite loop.
2833 if (pv->pv_pindex == pmap_pml4_pindex()) {
2839 * For page table pages (other than the top-level page),
2840 * remove and free the vm_page. The representitive mapping
2841 * removed above by pmap_remove_pv_pte() did not undo the
2842 * last wire_count so we have to do that as well.
2844 p = pmap_remove_pv_page(pv);
2845 vm_page_busy_wait(p, FALSE, "pmaprl");
2846 if (p->wire_count != 1) {
2847 kprintf("p->wire_count was %016lx %d\n",
2848 pv->pv_pindex, p->wire_count);
2850 KKASSERT(p->wire_count == 1);
2851 KKASSERT(p->flags & PG_UNMANAGED);
2853 vm_page_unwire(p, 0);
2854 KKASSERT(p->wire_count == 0);
2864 * This function will remove the pte associated with a pv from its parent.
2865 * Terminal pv's are supported. All cpus specified by (bulk) are properly
2868 * The wire count will be dropped on the parent page table. The wire
2869 * count on the page being removed (pv->pv_m) from the parent page table
2870 * is NOT touched. Note that terminal pages will not have any additional
2871 * wire counts while page table pages will have at least one representing
2872 * the mapping, plus others representing sub-mappings.
2874 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2875 * pages and user page table and terminal pages.
2877 * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2878 * be freshly allocated and not imply that the pte is managed. In this
2879 * case pv->pv_m should be NULL.
2881 * The pv must be locked. The pvp, if supplied, must be locked. All
2882 * supplied pv's will remain locked on return.
2884 * XXX must lock parent pv's if they exist to remove pte XXX
2888 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2891 vm_pindex_t ptepindex = pv->pv_pindex;
2892 pmap_t pmap = pv->pv_pmap;
2898 if (ptepindex == pmap_pml4_pindex()) {
2900 * We are the top level PML4E table, there is no parent.
2902 p = pmap->pm_pmlpv->pv_m;
2903 KKASSERT(pv->pv_m == p); /* debugging */
2904 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2906 * Remove a PDP page from the PML4E. This can only occur
2907 * with user page tables. We do not have to lock the
2908 * pml4 PV so just ignore pvp.
2910 vm_pindex_t pml4_pindex;
2911 vm_pindex_t pdp_index;
2914 pdp_index = ptepindex - pmap_pdp_pindex(0);
2916 pml4_pindex = pmap_pml4_pindex();
2917 pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2922 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2923 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2924 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2925 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2926 KKASSERT(pv->pv_m == p); /* debugging */
2927 } else if (ptepindex >= pmap_pd_pindex(0)) {
2929 * Remove a PD page from the PDP
2931 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2932 * of a simple pmap because it stops at
2935 vm_pindex_t pdp_pindex;
2936 vm_pindex_t pd_index;
2939 pd_index = ptepindex - pmap_pd_pindex(0);
2942 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2943 (pd_index >> NPML4EPGSHIFT);
2944 pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
2949 pd = pv_pte_lookup(pvp, pd_index &
2950 ((1ul << NPDPEPGSHIFT) - 1));
2951 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2952 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2953 pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2955 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2956 p = pv->pv_m; /* degenerate test later */
2958 KKASSERT(pv->pv_m == p); /* debugging */
2959 } else if (ptepindex >= pmap_pt_pindex(0)) {
2961 * Remove a PT page from the PD
2963 vm_pindex_t pd_pindex;
2964 vm_pindex_t pt_index;
2967 pt_index = ptepindex - pmap_pt_pindex(0);
2970 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2971 (pt_index >> NPDPEPGSHIFT);
2972 pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
2977 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2979 KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
2980 ("*pt unexpectedly invalid %016jx "
2981 "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2982 *pt, gotpvp, ptepindex, pt_index, pv, pvp));
2983 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2985 if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
2986 kprintf("*pt unexpectedly invalid %016jx "
2987 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2989 *pt, gotpvp, ptepindex, pt_index, pv, pvp);
2990 tsleep(pt, 0, "DEAD", 0);
2993 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2996 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2997 KKASSERT(pv->pv_m == p); /* debugging */
3000 * Remove a PTE from the PT page. The PV might exist even if
3001 * the PTE is not managed, in whichcase pv->pv_m should be
3004 * NOTE: Userland pmaps manage the parent PT/PD/PDP page
3005 * table pages but the kernel_pmap does not.
3007 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
3008 * pv is a pte_pv so we can safely lock pt_pv.
3010 * NOTE: FICTITIOUS pages may have multiple physical mappings
3011 * so PHYS_TO_VM_PAGE() will not necessarily work for
3014 vm_pindex_t pt_pindex;
3019 pt_pindex = ptepindex >> NPTEPGSHIFT;
3020 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
3022 if (ptepindex >= NUPTE_USER) {
3023 ptep = vtopte(ptepindex << PAGE_SHIFT);
3024 KKASSERT(pvp == NULL);
3025 /* pvp remains NULL */
3028 pt_pindex = NUPTE_TOTAL +
3029 (ptepindex >> NPDPEPGSHIFT);
3030 pvp = pv_get(pv->pv_pmap, pt_pindex, NULL);
3034 ptep = pv_pte_lookup(pvp, ptepindex &
3035 ((1ul << NPDPEPGSHIFT) - 1));
3037 pte = pmap_inval_bulk(bulk, va, ptep, 0);
3038 if (bulk == NULL) /* XXX */
3039 cpu_invlpg((void *)va); /* XXX */
3042 * Now update the vm_page_t
3044 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3045 (pte & pmap->pmap_bits[PG_V_IDX])) {
3047 * Valid managed page, adjust (p).
3049 if (pte & pmap->pmap_bits[PG_DEVICE_IDX]) {
3052 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
3053 KKASSERT(pv->pv_m == p);
3055 if (pte & pmap->pmap_bits[PG_M_IDX]) {
3056 if (pmap_track_modified(ptepindex))
3059 if (pte & pmap->pmap_bits[PG_A_IDX]) {
3060 vm_page_flag_set(p, PG_REFERENCED);
3064 * Unmanaged page, do not try to adjust the vm_page_t.
3065 * pv could be freshly allocated for a pmap_enter(),
3066 * replacing an unmanaged page with a managed one.
3068 * pv->pv_m might reflect the new page and not the
3071 * We could extract p from the physical address and
3072 * adjust it but we explicitly do not for unmanaged
3077 if (pte & pmap->pmap_bits[PG_W_IDX])
3078 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3079 if (pte & pmap->pmap_bits[PG_G_IDX])
3080 cpu_invlpg((void *)va);
3084 * If requested, scrap the underlying pv->pv_m and the underlying
3085 * pv. If this is a page-table-page we must also free the page.
3087 * pvp must be returned locked.
3091 * page table page (PT, PD, PDP, PML4), caller was responsible
3092 * for testing wired_count.
3094 KKASSERT(pv->pv_m->wire_count == 1);
3095 p = pmap_remove_pv_page(pv);
3099 vm_page_busy_wait(p, FALSE, "pgpun");
3100 vm_page_unwire(p, 0);
3101 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
3103 } else if (destroy == 2) {
3105 * Normal page, remove from pmap and leave the underlying
3108 pmap_remove_pv_page(pv);
3110 pv = NULL; /* safety */
3114 * If we acquired pvp ourselves then we are responsible for
3115 * recursively deleting it.
3117 if (pvp && gotpvp) {
3119 * Recursively destroy higher-level page tables.
3121 * This is optional. If we do not, they will still
3122 * be destroyed when the process exits.
3124 * NOTE: Do not destroy pv_entry's with extra hold refs,
3125 * a caller may have unlocked it and intends to
3126 * continue to use it.
3128 if (pmap_dynamic_delete &&
3130 pvp->pv_m->wire_count == 1 &&
3131 (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
3132 pvp->pv_pindex != pmap_pml4_pindex()) {
3133 if (pmap_dynamic_delete == 2)
3134 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
3135 if (pmap != &kernel_pmap) {
3136 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
3137 pvp = NULL; /* safety */
3139 kprintf("Attempt to remove kernel_pmap pindex "
3140 "%jd\n", pvp->pv_pindex);
3150 * Remove the vm_page association to a pv. The pv must be locked.
3154 pmap_remove_pv_page(pv_entry_t pv)
3159 vm_page_spin_lock(m);
3160 KKASSERT(m && m == pv->pv_m);
3162 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3163 pmap_page_stats_deleting(m);
3164 if (TAILQ_EMPTY(&m->md.pv_list))
3165 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3166 vm_page_spin_unlock(m);
3172 * Grow the number of kernel page table entries, if needed.
3174 * This routine is always called to validate any address space
3175 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
3176 * space below KERNBASE.
3178 * kernel_map must be locked exclusively by the caller.
3181 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3184 vm_offset_t ptppaddr;
3186 pd_entry_t *pt, newpt;
3187 pdp_entry_t *pd, newpd;
3188 int update_kernel_vm_end;
3191 * bootstrap kernel_vm_end on first real VM use
3193 if (kernel_vm_end == 0) {
3194 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3197 pt = pmap_pt(&kernel_pmap, kernel_vm_end);
3200 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) == 0)
3202 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3203 ~(vm_offset_t)(PAGE_SIZE * NPTEPG - 1);
3204 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
3205 kernel_vm_end = kernel_map.max_offset;
3212 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
3213 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
3214 * do not want to force-fill 128G worth of page tables.
3216 if (kstart < KERNBASE) {
3217 if (kstart > kernel_vm_end)
3218 kstart = kernel_vm_end;
3219 KKASSERT(kend <= KERNBASE);
3220 update_kernel_vm_end = 1;
3222 update_kernel_vm_end = 0;
3225 kstart = rounddown2(kstart, (vm_offset_t)(PAGE_SIZE * NPTEPG));
3226 kend = roundup2(kend, (vm_offset_t)(PAGE_SIZE * NPTEPG));
3228 if (kend - 1 >= kernel_map.max_offset)
3229 kend = kernel_map.max_offset;
3231 while (kstart < kend) {
3232 pt = pmap_pt(&kernel_pmap, kstart);
3235 * We need a new PD entry
3237 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3240 VM_ALLOC_INTERRUPT);
3242 panic("pmap_growkernel: no memory to grow "
3245 paddr = VM_PAGE_TO_PHYS(nkpg);
3246 pmap_zero_page(paddr);
3247 pd = pmap_pd(&kernel_pmap, kstart);
3249 newpd = (pdp_entry_t)
3251 kernel_pmap.pmap_bits[PG_V_IDX] |
3252 kernel_pmap.pmap_bits[PG_RW_IDX] |
3253 kernel_pmap.pmap_bits[PG_A_IDX] |
3254 kernel_pmap.pmap_bits[PG_M_IDX]);
3255 atomic_swap_long(pd, newpd);
3258 kprintf("NEWPD pd=%p pde=%016jx phys=%016jx\n",
3262 continue; /* try again */
3265 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3266 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3267 ~(vm_offset_t)(PAGE_SIZE * NPTEPG - 1);
3268 if (kstart - 1 >= kernel_map.max_offset) {
3269 kstart = kernel_map.max_offset;
3278 * This index is bogus, but out of the way
3280 nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3283 VM_ALLOC_INTERRUPT);
3285 panic("pmap_growkernel: no memory to grow kernel");
3288 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3289 pmap_zero_page(ptppaddr);
3290 newpt = (pd_entry_t)(ptppaddr |
3291 kernel_pmap.pmap_bits[PG_V_IDX] |
3292 kernel_pmap.pmap_bits[PG_RW_IDX] |
3293 kernel_pmap.pmap_bits[PG_A_IDX] |
3294 kernel_pmap.pmap_bits[PG_M_IDX]);
3295 atomic_swap_long(pt, newpt);
3297 kstart = (kstart + PAGE_SIZE * NPTEPG) &
3298 ~(vm_offset_t)(PAGE_SIZE * NPTEPG - 1);
3300 if (kstart - 1 >= kernel_map.max_offset) {
3301 kstart = kernel_map.max_offset;
3307 * Only update kernel_vm_end for areas below KERNBASE.
3309 if (update_kernel_vm_end && kernel_vm_end < kstart)
3310 kernel_vm_end = kstart;
3314 * Add a reference to the specified pmap.
3317 pmap_reference(pmap_t pmap)
3320 atomic_add_int(&pmap->pm_count, 1);
3323 /***************************************************
3324 * page management routines.
3325 ***************************************************/
3328 * Hold a pv without locking it
3331 pv_hold(pv_entry_t pv)
3333 atomic_add_int(&pv->pv_hold, 1);
3337 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
3338 * was successfully locked, FALSE if it wasn't. The caller must dispose of
3341 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3342 * pv list via its page) must be held by the caller in order to stabilize
3346 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3351 * Critical path shortcut expects pv to already have one ref
3352 * (for the pv->pv_pmap).
3354 if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3357 pv->pv_line = lineno;
3363 count = pv->pv_hold;
3365 if ((count & PV_HOLD_LOCKED) == 0) {
3366 if (atomic_cmpset_int(&pv->pv_hold, count,
3367 (count + 1) | PV_HOLD_LOCKED)) {
3370 pv->pv_line = lineno;
3375 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3383 * Drop a previously held pv_entry which could not be locked, allowing its
3386 * Must not be called with a spinlock held as we might zfree() the pv if it
3387 * is no longer associated with a pmap and this was the last hold count.
3390 pv_drop(pv_entry_t pv)
3395 count = pv->pv_hold;
3397 KKASSERT((count & PV_HOLD_MASK) > 0);
3398 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3399 (PV_HOLD_LOCKED | 1));
3400 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3401 if ((count & PV_HOLD_MASK) == 1) {
3403 if (pmap_enter_debug > 0) {
3405 kprintf("pv_drop: free pv %p\n", pv);
3408 KKASSERT(count == 1);
3409 KKASSERT(pv->pv_pmap == NULL);
3419 * Find or allocate the requested PV entry, returning a locked, held pv.
3421 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3422 * for the caller and one representing the pmap and vm_page association.
3424 * If (*isnew) is zero, the returned pv will have only one hold count.
3426 * Since both associations can only be adjusted while the pv is locked,
3427 * together they represent just one additional hold.
3431 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3433 struct mdglobaldata *md = mdcpu;
3441 pnew = atomic_swap_ptr((void *)&md->gd_newpv, NULL);
3444 pnew = md->gd_newpv; /* might race NULL */
3445 md->gd_newpv = NULL;
3450 pnew = zalloc(pvzone);
3452 spin_lock_shared(&pmap->pm_spin);
3457 pv = pv_entry_lookup(pmap, pindex);
3462 * Requires exclusive pmap spinlock
3464 if (pmap_excl == 0) {
3466 if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3467 spin_unlock_shared(&pmap->pm_spin);
3468 spin_lock(&pmap->pm_spin);
3474 * We need to block if someone is holding our
3475 * placemarker. As long as we determine the
3476 * placemarker has not been aquired we do not
3477 * need to get it as acquision also requires
3478 * the pmap spin lock.
3480 * However, we can race the wakeup.
3482 pmark = pmap_placemarker_hash(pmap, pindex);
3484 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3485 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3486 tsleep_interlock(pmark, 0);
3487 if (((*pmark ^ pindex) &
3488 ~PM_PLACEMARK_WAKEUP) == 0) {
3489 spin_unlock(&pmap->pm_spin);
3490 tsleep(pmark, PINTERLOCKED, "pvplc", 0);
3491 spin_lock(&pmap->pm_spin);
3497 * Setup the new entry
3499 pnew->pv_pmap = pmap;
3500 pnew->pv_pindex = pindex;
3501 pnew->pv_hold = PV_HOLD_LOCKED | 2;
3503 pnew->pv_func = func;
3504 pnew->pv_line = lineno;
3505 if (pnew->pv_line_lastfree > 0) {
3506 pnew->pv_line_lastfree =
3507 -pnew->pv_line_lastfree;
3510 pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3511 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3512 spin_unlock(&pmap->pm_spin);
3515 KKASSERT(pv == NULL);
3520 * We already have an entry, cleanup the staged pnew if
3521 * we can get the lock, otherwise block and retry.
3523 if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY))) {
3525 spin_unlock(&pmap->pm_spin);
3527 spin_unlock_shared(&pmap->pm_spin);
3529 pnew = atomic_swap_ptr((void *)&md->gd_newpv, pnew);
3531 zfree(pvzone, pnew);
3534 if (md->gd_newpv == NULL)
3535 md->gd_newpv = pnew;
3537 zfree(pvzone, pnew);
3540 KKASSERT(pv->pv_pmap == pmap &&
3541 pv->pv_pindex == pindex);
3546 spin_unlock(&pmap->pm_spin);
3547 _pv_lock(pv PMAP_DEBUG_COPY);
3549 spin_lock(&pmap->pm_spin);
3551 spin_unlock_shared(&pmap->pm_spin);
3552 _pv_lock(pv PMAP_DEBUG_COPY);
3554 spin_lock_shared(&pmap->pm_spin);
3561 * Find the requested PV entry, returning a locked+held pv or NULL
3565 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3570 spin_lock_shared(&pmap->pm_spin);
3575 pv = pv_entry_lookup(pmap, pindex);
3578 * Block if there is ANY placemarker. If we are to
3579 * return it, we must also aquire the spot, so we
3580 * have to block even if the placemarker is held on
3581 * a different address.
3583 * OPTIMIZATION: If pmarkp is passed as NULL the
3584 * caller is just probing (or looking for a real
3585 * pv_entry), and in this case we only need to check
3586 * to see if the placemarker matches pindex.
3591 * Requires exclusive pmap spinlock
3593 if (pmap_excl == 0) {
3595 if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3596 spin_unlock_shared(&pmap->pm_spin);
3597 spin_lock(&pmap->pm_spin);
3602 pmark = pmap_placemarker_hash(pmap, pindex);
3604 if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3605 ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3606 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3607 tsleep_interlock(pmark, 0);
3608 if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3609 ((*pmark ^ pindex) &
3610 ~PM_PLACEMARK_WAKEUP) == 0) {
3611 spin_unlock(&pmap->pm_spin);
3612 tsleep(pmark, PINTERLOCKED, "pvpld", 0);
3613 spin_lock(&pmap->pm_spin);
3618 if (atomic_swap_long(pmark, pindex) !=
3620 panic("_pv_get: pmark race");
3624 spin_unlock(&pmap->pm_spin);
3627 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3628 pv_cache(pv, pindex);
3630 spin_unlock(&pmap->pm_spin);
3632 spin_unlock_shared(&pmap->pm_spin);
3633 KKASSERT(pv->pv_pmap == pmap &&
3634 pv->pv_pindex == pindex);
3638 spin_unlock(&pmap->pm_spin);
3639 _pv_lock(pv PMAP_DEBUG_COPY);
3641 spin_lock(&pmap->pm_spin);
3643 spin_unlock_shared(&pmap->pm_spin);
3644 _pv_lock(pv PMAP_DEBUG_COPY);
3646 spin_lock_shared(&pmap->pm_spin);
3652 * Lookup, hold, and attempt to lock (pmap,pindex).
3654 * If the entry does not exist NULL is returned and *errorp is set to 0
3656 * If the entry exists and could be successfully locked it is returned and
3657 * errorp is set to 0.
3659 * If the entry exists but could NOT be successfully locked it is returned
3660 * held and *errorp is set to 1.
3662 * If the entry is placemarked by someone else NULL is returned and *errorp
3667 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3671 spin_lock_shared(&pmap->pm_spin);
3673 pv = pv_entry_lookup(pmap, pindex);
3677 pmark = pmap_placemarker_hash(pmap, pindex);
3679 if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3681 } else if (pmarkp &&
3682 atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3686 * Can't set a placemark with a NULL pmarkp, or if
3687 * pmarkp is non-NULL but we failed to set our
3694 spin_unlock_shared(&pmap->pm_spin);
3700 * XXX This has problems if the lock is shared, why?
3702 if (pv_hold_try(pv)) {
3703 pv_cache(pv, pindex); /* overwrite ok (shared lock) */
3704 spin_unlock_shared(&pmap->pm_spin);
3706 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3707 return(pv); /* lock succeeded */
3709 spin_unlock_shared(&pmap->pm_spin);
3712 return (pv); /* lock failed */
3716 * Lock a held pv, keeping the hold count
3720 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3725 count = pv->pv_hold;
3727 if ((count & PV_HOLD_LOCKED) == 0) {
3728 if (atomic_cmpset_int(&pv->pv_hold, count,
3729 count | PV_HOLD_LOCKED)) {
3732 pv->pv_line = lineno;
3738 tsleep_interlock(pv, 0);
3739 if (atomic_cmpset_int(&pv->pv_hold, count,
3740 count | PV_HOLD_WAITING)) {
3742 if (pmap_enter_debug > 0) {
3744 kprintf("pv waiting on %s:%d\n",
3745 pv->pv_func, pv->pv_line);
3748 tsleep(pv, PINTERLOCKED, "pvwait", hz);
3755 * Unlock a held and locked pv, keeping the hold count.
3759 pv_unlock(pv_entry_t pv)
3764 count = pv->pv_hold;
3766 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3767 (PV_HOLD_LOCKED | 1));
3768 if (atomic_cmpset_int(&pv->pv_hold, count,
3770 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3771 if (count & PV_HOLD_WAITING)
3779 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3780 * and the hold count drops to zero we will free it.
3782 * Caller should not hold any spin locks. We are protected from hold races
3783 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3784 * lock held. A pv cannot be located otherwise.
3788 pv_put(pv_entry_t pv)
3791 if (pmap_enter_debug > 0) {
3793 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3798 * Normal put-aways must have a pv_m associated with the pv,
3799 * but allow the case where the pv has been destructed due
3800 * to pmap_dynamic_delete.
3802 KKASSERT(pv->pv_pmap == NULL || pv->pv_m != NULL);
3805 * Fast - shortcut most common condition
3807 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3818 * Remove the pmap association from a pv, require that pv_m already be removed,
3819 * then unlock and drop the pv. Any pte operations must have already been
3820 * completed. This call may result in a last-drop which will physically free
3823 * Removing the pmap association entails an additional drop.
3825 * pv must be exclusively locked on call and will be disposed of on return.
3829 _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL)
3834 pv->pv_func_lastfree = func;
3835 pv->pv_line_lastfree = lineno;
3837 KKASSERT(pv->pv_m == NULL);
3838 KKASSERT((pv->pv_hold & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
3839 (PV_HOLD_LOCKED|1));
3840 if ((pmap = pv->pv_pmap) != NULL) {
3841 spin_lock(&pmap->pm_spin);
3842 KKASSERT(pv->pv_pmap == pmap);
3843 if (pmap->pm_pvhint_pt == pv)
3844 pmap->pm_pvhint_pt = NULL;
3845 if (pmap->pm_pvhint_pte == pv)
3846 pmap->pm_pvhint_pte = NULL;
3847 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3848 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3851 spin_unlock(&pmap->pm_spin);
3854 * Try to shortcut three atomic ops, otherwise fall through
3855 * and do it normally. Drop two refs and the lock all in
3859 vm_page_unwire_quick(pvp->pv_m);
3860 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3862 if (pmap_enter_debug > 0) {
3864 kprintf("pv_free: free pv %p\n", pv);
3870 pv_drop(pv); /* ref for pv_pmap */
3877 * This routine is very drastic, but can save the system
3885 static int warningdone=0;
3887 if (pmap_pagedaemon_waken == 0)
3889 pmap_pagedaemon_waken = 0;
3890 if (warningdone < 5) {
3891 kprintf("pmap_collect: collecting pv entries -- "
3892 "suggest increasing PMAP_SHPGPERPROC\n");
3896 for (i = 0; i < vm_page_array_size; i++) {
3897 m = &vm_page_array[i];
3898 if (m->wire_count || m->hold_count)
3900 if (vm_page_busy_try(m, TRUE) == 0) {
3901 if (m->wire_count == 0 && m->hold_count == 0) {
3910 * Scan the pmap for active page table entries and issue a callback.
3911 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3912 * its parent page table.
3914 * pte_pv will be NULL if the page or page table is unmanaged.
3915 * pt_pv will point to the page table page containing the pte for the page.
3917 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3918 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3919 * process pmap's PD and page to the callback function. This can be
3920 * confusing because the pt_pv is really a pd_pv, and the target page
3921 * table page is simply aliased by the pmap and not owned by it.
3923 * It is assumed that the start and end are properly rounded to the page size.
3925 * It is assumed that PD pages and above are managed and thus in the RB tree,
3926 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3928 struct pmap_scan_info {
3932 vm_pindex_t sva_pd_pindex;
3933 vm_pindex_t eva_pd_pindex;
3934 void (*func)(pmap_t, struct pmap_scan_info *,
3935 pv_entry_t, vm_pindex_t *, pv_entry_t,
3937 pt_entry_t *, void *);
3939 pmap_inval_bulk_t bulk_core;
3940 pmap_inval_bulk_t *bulk;
3945 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3946 static int pmap_scan_callback(pv_entry_t pv, void *data);
3949 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3951 struct pmap *pmap = info->pmap;
3952 pv_entry_t pd_pv; /* A page directory PV */
3953 pv_entry_t pt_pv; /* A page table PV */
3954 pv_entry_t pte_pv; /* A page table entry PV */
3955 vm_pindex_t *pte_placemark;
3956 vm_pindex_t *pt_placemark;
3959 struct pv_entry dummy_pv;
3964 if (info->sva == info->eva)
3967 info->bulk = &info->bulk_core;
3968 pmap_inval_bulk_init(&info->bulk_core, pmap);
3974 * Hold the token for stability; if the pmap is empty we have nothing
3978 if (pmap->pm_stats.resident_count == 0) {
3986 * Special handling for scanning one page, which is a very common
3987 * operation (it is?).
3989 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3991 if (info->sva + PAGE_SIZE == info->eva) {
3992 if (info->sva >= VM_MAX_USER_ADDRESS) {
3994 * Kernel mappings do not track wire counts on
3995 * page table pages and only maintain pd_pv and
3996 * pte_pv levels so pmap_scan() works.
3999 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
4001 ptep = vtopte(info->sva);
4004 * User pages which are unmanaged will not have a
4005 * pte_pv. User page table pages which are unmanaged
4006 * (shared from elsewhere) will also not have a pt_pv.
4007 * The func() callback will pass both pte_pv and pt_pv
4008 * as NULL in that case.
4010 * We hold pte_placemark across the operation for
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 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
4021 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
4023 if (pt_pv == NULL) {
4024 KKASSERT(pte_pv == NULL);
4025 pd_pv = pv_get(pmap,
4026 pmap_pd_pindex(info->sva),
4029 ptep = pv_pte_lookup(pd_pv,
4030 pmap_pt_index(info->sva));
4032 info->func(pmap, info,
4038 pv_placemarker_wakeup(pmap,
4043 pv_placemarker_wakeup(pmap,
4046 pv_placemarker_wakeup(pmap, pte_placemark);
4049 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
4053 * NOTE: *ptep can't be ripped out from under us if we hold
4054 * pte_pv (or pte_placemark) locked, but bits can
4060 KKASSERT(pte_pv == NULL);
4061 pv_placemarker_wakeup(pmap, pte_placemark);
4062 } else if (pte_pv) {
4063 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
4064 pmap->pmap_bits[PG_V_IDX])) ==
4065 (pmap->pmap_bits[PG_MANAGED_IDX] |
4066 pmap->pmap_bits[PG_V_IDX]),
4067 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
4068 *ptep, oldpte, info->sva, pte_pv));
4069 info->func(pmap, info, pte_pv, NULL, pt_pv, 0,
4070 info->sva, ptep, info->arg);
4072 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
4073 pmap->pmap_bits[PG_V_IDX])) ==
4074 pmap->pmap_bits[PG_V_IDX],
4075 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
4076 *ptep, oldpte, info->sva));
4077 info->func(pmap, info, NULL, pte_placemark, pt_pv, 0,
4078 info->sva, ptep, info->arg);
4083 pmap_inval_bulk_flush(info->bulk);
4088 * Nominal scan case, RB_SCAN() for PD pages and iterate from
4091 * WARNING! eva can overflow our standard ((N + mask) >> bits)
4092 * bounds, resulting in a pd_pindex of 0. To solve the
4093 * problem we use an inclusive range.
4095 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
4096 info->eva_pd_pindex = pmap_pd_pindex(info->eva - PAGE_SIZE);
4098 if (info->sva >= VM_MAX_USER_ADDRESS) {
4100 * The kernel does not currently maintain any pv_entry's for
4101 * higher-level page tables.
4103 bzero(&dummy_pv, sizeof(dummy_pv));
4104 dummy_pv.pv_pindex = info->sva_pd_pindex;
4105 spin_lock(&pmap->pm_spin);
4106 while (dummy_pv.pv_pindex <= info->eva_pd_pindex) {
4107 pmap_scan_callback(&dummy_pv, info);
4108 ++dummy_pv.pv_pindex;
4109 if (dummy_pv.pv_pindex < info->sva_pd_pindex) /*wrap*/
4112 spin_unlock(&pmap->pm_spin);
4115 * User page tables maintain local PML4, PDP, and PD
4116 * pv_entry's at the very least. PT pv's might be
4117 * unmanaged and thus not exist. PTE pv's might be
4118 * unmanaged and thus not exist.
4120 spin_lock(&pmap->pm_spin);
4121 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
4122 pmap_scan_callback, info);
4123 spin_unlock(&pmap->pm_spin);
4125 pmap_inval_bulk_flush(info->bulk);
4129 * WARNING! pmap->pm_spin held
4131 * WARNING! eva can overflow our standard ((N + mask) >> bits)
4132 * bounds, resulting in a pd_pindex of 0. To solve the
4133 * problem we use an inclusive range.
4136 pmap_scan_cmp(pv_entry_t pv, void *data)
4138 struct pmap_scan_info *info = data;
4139 if (pv->pv_pindex < info->sva_pd_pindex)
4141 if (pv->pv_pindex > info->eva_pd_pindex)
4147 * pmap_scan() by PDs
4149 * WARNING! pmap->pm_spin held
4152 pmap_scan_callback(pv_entry_t pv, void *data)
4154 struct pmap_scan_info *info = data;
4155 struct pmap *pmap = info->pmap;
4156 pv_entry_t pd_pv; /* A page directory PV */
4157 pv_entry_t pt_pv; /* A page table PV */
4158 vm_pindex_t *pt_placemark;
4163 vm_offset_t va_next;
4164 vm_pindex_t pd_pindex;
4174 * Pull the PD pindex from the pv before releasing the spinlock.
4176 * WARNING: pv is faked for kernel pmap scans.
4178 pd_pindex = pv->pv_pindex;
4179 spin_unlock(&pmap->pm_spin);
4180 pv = NULL; /* invalid after spinlock unlocked */
4183 * Calculate the page range within the PD. SIMPLE pmaps are
4184 * direct-mapped for the entire 2^64 address space. Normal pmaps
4185 * reflect the user and kernel address space which requires
4186 * cannonicalization w/regards to converting pd_pindex's back
4189 sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
4190 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
4191 (sva & PML4_SIGNMASK)) {
4192 sva |= PML4_SIGNMASK;
4194 eva = sva + NBPDP; /* can overflow */
4195 if (sva < info->sva)
4197 if (eva < info->sva || eva > info->eva)
4201 * NOTE: kernel mappings do not track page table pages, only
4204 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
4205 * However, for the scan to be efficient we try to
4206 * cache items top-down.
4211 for (; sva < eva; sva = va_next) {
4214 if (sva >= VM_MAX_USER_ADDRESS) {
4223 * PD cache, scan shortcut if it doesn't exist.
4225 if (pd_pv == NULL) {
4226 pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4227 } else if (pd_pv->pv_pmap != pmap ||
4228 pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
4230 pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4232 if (pd_pv == NULL) {
4233 va_next = (sva + NBPDP) & ~PDPMASK;
4242 * NOTE: The cached pt_pv can be removed from the pmap when
4243 * pmap_dynamic_delete is enabled.
4245 if (pt_pv && (pt_pv->pv_pmap != pmap ||
4246 pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
4250 if (pt_pv == NULL) {
4251 pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
4252 &pt_placemark, &error);
4254 pv_put(pd_pv); /* lock order */
4261 pv_placemarker_wait(pmap, pt_placemark);
4266 /* may have to re-check later if pt_pv is NULL here */
4270 * If pt_pv is NULL we either have an shared page table
4271 * page and must issue a callback specific to that case,
4272 * or there is no page table page.
4274 * Either way we can skip the page table page.
4276 * WARNING! pt_pv can also be NULL due to a pv creation
4277 * race where we find it to be NULL and then
4278 * later see a pte_pv. But its possible the pt_pv
4279 * got created inbetween the two operations, so
4282 if (pt_pv == NULL) {
4284 * Possible unmanaged (shared from another pmap)
4287 * WARNING! We must hold pt_placemark across the
4288 * *ptep test to prevent misintepreting
4289 * a non-zero *ptep as a shared page
4290 * table page. Hold it across the function
4291 * callback as well for SMP safety.
4293 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4294 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4295 info->func(pmap, info, NULL, pt_placemark,
4297 sva, ptep, info->arg);
4299 pv_placemarker_wakeup(pmap, pt_placemark);
4303 * Done, move to next page table page.
4305 va_next = (sva + NBPDR) & ~PDRMASK;
4312 * From this point in the loop testing pt_pv for non-NULL
4313 * means we are in UVM, else if it is NULL we are in KVM.
4315 * Limit our scan to either the end of the va represented
4316 * by the current page table page, or to the end of the
4317 * range being removed.
4320 va_next = (sva + NBPDR) & ~PDRMASK;
4327 * Scan the page table for pages. Some pages may not be
4328 * managed (might not have a pv_entry).
4330 * There is no page table management for kernel pages so
4331 * pt_pv will be NULL in that case, but otherwise pt_pv
4332 * is non-NULL, locked, and referenced.
4336 * At this point a non-NULL pt_pv means a UVA, and a NULL
4337 * pt_pv means a KVA.
4340 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4344 while (sva < va_next) {
4346 vm_pindex_t *pte_placemark;
4349 * Yield every 64 pages, stop if requested.
4351 if ((++info->count & 63) == 0)
4357 * We can shortcut our scan if *ptep == 0. This is
4358 * an unlocked check.
4368 * Acquire the related pte_pv, if any. If *ptep == 0
4369 * the related pte_pv should not exist, but if *ptep
4370 * is not zero the pte_pv may or may not exist (e.g.
4371 * will not exist for an unmanaged page).
4373 * However a multitude of races are possible here
4374 * so if we cannot lock definite state we clean out
4375 * our cache and break the inner while() loop to
4376 * force a loop up to the top of the for().
4378 * XXX unlock/relock pd_pv, pt_pv, and re-test their
4379 * validity instead of looping up?
4381 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4382 &pte_placemark, &error);
4384 pv_put(pd_pv); /* lock order */
4387 pv_put(pt_pv); /* lock order */
4390 if (pte_pv) { /* block */
4395 pv_placemarker_wait(pmap,
4398 va_next = sva; /* retry */
4403 * Reload *ptep after successfully locking the
4404 * pindex. If *ptep == 0 we had better NOT have a
4411 kprintf("Unexpected non-NULL pte_pv "
4413 "*ptep = %016lx/%016lx\n",
4414 pte_pv, pt_pv, *ptep, oldpte);
4415 panic("Unexpected non-NULL pte_pv");
4417 pv_placemarker_wakeup(pmap, pte_placemark);
4425 * We can't hold pd_pv across the callback (because
4426 * we don't pass it to the callback and the callback
4430 vm_page_wire_quick(pd_pv->pv_m);
4435 * Ready for the callback. The locked pte_pv (if any)
4436 * is consumed by the callback. pte_pv will exist if
4437 * the page is managed, and will not exist if it
4440 if (oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4445 (oldpte & pmap->pmap_bits[PG_V_IDX]),
4446 ("badC *ptep %016lx/%016lx sva %016lx "
4448 *ptep, oldpte, sva, pte_pv));
4450 * We must unlock pd_pv across the callback
4451 * to avoid deadlocks on any recursive
4452 * disposal. Re-check that it still exists
4455 * Call target disposes of pte_pv and may
4456 * destroy but will not dispose of pt_pv.
4458 info->func(pmap, info, pte_pv, NULL,
4460 sva, ptep, info->arg);
4465 * We must unlock pd_pv across the callback
4466 * to avoid deadlocks on any recursive
4467 * disposal. Re-check that it still exists
4470 * Call target disposes of pte_pv or
4471 * pte_placemark and may destroy but will
4472 * not dispose of pt_pv.
4474 KASSERT(pte_pv == NULL &&
4475 (oldpte & pmap->pmap_bits[PG_V_IDX]),
4476 ("badD *ptep %016lx/%016lx sva %016lx "
4477 "pte_pv %p pte_pv->pv_m %p ",
4479 pte_pv, (pte_pv ? pte_pv->pv_m : NULL)));
4483 info->func(pmap, info,
4486 sva, ptep, info->arg);
4488 info->func(pmap, info,
4489 NULL, pte_placemark,
4491 sva, ptep, info->arg);
4496 vm_page_unwire_quick(pd_pv->pv_m);
4497 if (pd_pv->pv_pmap == NULL) {
4498 va_next = sva; /* retry */
4504 * NOTE: The cached pt_pv can be removed from the
4505 * pmap when pmap_dynamic_delete is enabled,
4506 * which will cause ptep to become stale.
4508 * This also means that no pages remain under
4509 * the PT, so we can just break out of the inner
4510 * loop and let the outer loop clean everything
4513 if (pt_pv && pt_pv->pv_pmap != pmap)
4528 if ((++info->count & 7) == 0)
4532 * Relock before returning.
4534 spin_lock(&pmap->pm_spin);
4539 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4541 struct pmap_scan_info info;
4546 info.func = pmap_remove_callback;
4548 pmap_scan(&info, 1);
4551 if (eva - sva < 1024*1024) {
4553 cpu_invlpg((void *)sva);
4561 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4563 struct pmap_scan_info info;
4568 info.func = pmap_remove_callback;
4570 pmap_scan(&info, 0);
4574 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4575 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4576 pv_entry_t pt_pv, int sharept,
4577 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4585 * This will also drop pt_pv's wire_count. Note that
4586 * terminal pages are not wired based on mmu presence.
4588 * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4590 KKASSERT(pte_pv->pv_m != NULL);
4591 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4592 pte_pv = NULL; /* safety */
4595 * Recursively destroy higher-level page tables.
4597 * This is optional. If we do not, they will still
4598 * be destroyed when the process exits.
4600 * NOTE: Do not destroy pv_entry's with extra hold refs,
4601 * a caller may have unlocked it and intends to
4602 * continue to use it.
4604 if (pmap_dynamic_delete &&
4607 pt_pv->pv_m->wire_count == 1 &&
4608 (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4609 pt_pv->pv_pindex != pmap_pml4_pindex()) {
4610 if (pmap_dynamic_delete == 2)
4611 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4612 pv_hold(pt_pv); /* extra hold */
4613 pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4614 pv_lock(pt_pv); /* prior extra hold + relock */
4616 } else if (sharept == 0) {
4618 * Unmanaged pte (pte_placemark is non-NULL)
4620 * pt_pv's wire_count is still bumped by unmanaged pages
4621 * so we must decrement it manually.
4623 * We have to unwire the target page table page.
4625 pte = pmap_inval_bulk(info->bulk, va, ptep, 0);
4626 if (pte & pmap->pmap_bits[PG_W_IDX])
4627 atomic_add_long(&pmap->pm_stats.wired_count, -1);
4628 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4629 if (vm_page_unwire_quick(pt_pv->pv_m))
4630 panic("pmap_remove: insufficient wirecount");
4631 pv_placemarker_wakeup(pmap, pte_placemark);
4634 * Unmanaged page table (pt, pd, or pdp. Not pte) for
4635 * a shared page table.
4637 * pt_pv is actually the pd_pv for our pmap (not the shared
4640 * We have to unwire the target page table page and we
4641 * have to unwire our page directory page.
4643 * It is unclear how we can invalidate a segment so we
4644 * invalidate -1 which invlidates the tlb.
4646 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4647 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4648 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4649 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4650 panic("pmap_remove: shared pgtable1 bad wirecount");
4651 if (vm_page_unwire_quick(pt_pv->pv_m))
4652 panic("pmap_remove: shared pgtable2 bad wirecount");
4653 pv_placemarker_wakeup(pmap, pte_placemark);
4658 * Removes this physical page from all physical maps in which it resides.
4659 * Reflects back modify bits to the pager.
4661 * This routine may not be called from an interrupt.
4665 pmap_remove_all(vm_page_t m)
4668 pmap_inval_bulk_t bulk;
4670 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4673 vm_page_spin_lock(m);
4674 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4675 KKASSERT(pv->pv_m == m);
4676 if (pv_hold_try(pv)) {
4677 vm_page_spin_unlock(m);
4679 vm_page_spin_unlock(m);
4682 vm_page_spin_lock(m);
4685 KKASSERT(pv->pv_pmap && pv->pv_m == m);
4688 * Holding no spinlocks, pv is locked. Once we scrap
4689 * pv we can no longer use it as a list iterator (but
4690 * we are doing a TAILQ_FIRST() so we are ok).
4692 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4693 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4694 pv = NULL; /* safety */
4695 pmap_inval_bulk_flush(&bulk);
4696 vm_page_spin_lock(m);
4698 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4699 vm_page_spin_unlock(m);
4703 * Removes the page from a particular pmap
4706 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4709 pmap_inval_bulk_t bulk;
4711 if (!pmap_initialized)
4715 vm_page_spin_lock(m);
4716 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4717 if (pv->pv_pmap != pmap)
4719 KKASSERT(pv->pv_m == m);
4720 if (pv_hold_try(pv)) {
4721 vm_page_spin_unlock(m);
4723 vm_page_spin_unlock(m);
4728 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
4731 * Holding no spinlocks, pv is locked. Once gone it can't
4732 * be used as an iterator. In fact, because we couldn't
4733 * necessarily lock it atomically it may have moved within
4734 * the list and ALSO cannot be used as an iterator.
4736 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4737 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4738 pv = NULL; /* safety */
4739 pmap_inval_bulk_flush(&bulk);
4742 vm_page_spin_unlock(m);
4746 * Set the physical protection on the specified range of this map
4747 * as requested. This function is typically only used for debug watchpoints
4750 * This function may not be called from an interrupt if the map is
4751 * not the kernel_pmap.
4753 * NOTE! For shared page table pages we just unmap the page.
4756 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4758 struct pmap_scan_info info;
4759 /* JG review for NX */
4763 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == VM_PROT_NONE) {
4764 pmap_remove(pmap, sva, eva);
4767 if (prot & VM_PROT_WRITE)
4772 info.func = pmap_protect_callback;
4774 pmap_scan(&info, 1);
4779 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4780 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
4781 pv_entry_t pt_pv, int sharept,
4782 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4793 KKASSERT(pte_pv->pv_m != NULL);
4795 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4796 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4797 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4798 KKASSERT(m == pte_pv->pv_m);
4799 vm_page_flag_set(m, PG_REFERENCED);
4801 cbits &= ~pmap->pmap_bits[PG_A_IDX];
4803 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4804 if (pmap_track_modified(pte_pv->pv_pindex)) {
4805 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4807 m = PHYS_TO_VM_PAGE(pbits &
4812 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4815 } else if (sharept) {
4817 * Unmanaged page table, pt_pv is actually the pd_pv
4818 * for our pmap (not the object's shared pmap).
4820 * When asked to protect something in a shared page table
4821 * page we just unmap the page table page. We have to
4822 * invalidate the tlb in this situation.
4824 * XXX Warning, shared page tables will not be used for
4825 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4826 * so PHYS_TO_VM_PAGE() should be safe here.
4828 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4829 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4830 panic("pmap_protect: pgtable1 pg bad wirecount");
4831 if (vm_page_unwire_quick(pt_pv->pv_m))
4832 panic("pmap_protect: pgtable2 pg bad wirecount");
4835 /* else unmanaged page, adjust bits, no wire changes */
4838 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4840 if (pmap_enter_debug > 0) {
4842 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4843 "pt_pv=%p cbits=%08lx\n",
4849 if (pbits != cbits) {
4852 xva = (sharept) ? (vm_offset_t)-1 : va;
4853 if (!pmap_inval_smp_cmpset(pmap, xva,
4854 ptep, pbits, cbits)) {
4862 pv_placemarker_wakeup(pmap, pte_placemark);
4866 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4867 * mapping at that address. Set protection and wiring as requested.
4869 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4870 * possible. If it is we enter the page into the appropriate shared pmap
4871 * hanging off the related VM object instead of the passed pmap, then we
4872 * share the page table page from the VM object's pmap into the current pmap.
4874 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4877 * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4881 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4882 boolean_t wired, vm_map_entry_t entry)
4884 pv_entry_t pt_pv; /* page table */
4885 pv_entry_t pte_pv; /* page table entry */
4886 vm_pindex_t *pte_placemark;
4889 pt_entry_t origpte, newpte;
4894 va = trunc_page(va);
4895 #ifdef PMAP_DIAGNOSTIC
4897 panic("pmap_enter: toobig");
4898 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4899 panic("pmap_enter: invalid to pmap_enter page table "
4900 "pages (va: 0x%lx)", va);
4902 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4903 kprintf("Warning: pmap_enter called on UVA with "
4906 db_print_backtrace();
4909 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4910 kprintf("Warning: pmap_enter called on KVA without"
4913 db_print_backtrace();
4918 * Get locked PV entries for our new page table entry (pte_pv or
4919 * pte_placemark) and for its parent page table (pt_pv). We need
4920 * the parent so we can resolve the location of the ptep.
4922 * Only hardware MMU actions can modify the ptep out from
4925 * if (m) is fictitious or unmanaged we do not create a managing
4926 * pte_pv for it. Any pre-existing page's management state must
4927 * match (avoiding code complexity).
4929 * If the pmap is still being initialized we assume existing
4932 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4934 * WARNING! If replacing a managed mapping with an unmanaged mapping
4935 * pte_pv will wind up being non-NULL and must be handled
4938 if (pmap_initialized == FALSE) {
4941 pte_placemark = NULL;
4944 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4945 pmap_softwait(pmap);
4946 pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4947 KKASSERT(pte_pv == NULL);
4948 if (va >= VM_MAX_USER_ADDRESS) {
4952 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4954 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4958 KASSERT(origpte == 0 ||
4959 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4960 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4962 pmap_softwait(pmap);
4963 if (va >= VM_MAX_USER_ADDRESS) {
4965 * Kernel map, pv_entry-tracked.
4968 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4974 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4976 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4978 pte_placemark = NULL; /* safety */
4981 KASSERT(origpte == 0 ||
4982 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4983 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4986 pa = VM_PAGE_TO_PHYS(m);
4987 opa = origpte & PG_FRAME;
4990 * Calculate the new PTE. Note that pte_pv alone does not mean
4991 * the new pte_pv is managed, it could exist because the old pte
4992 * was managed even if the new one is not.
4994 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4995 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4997 newpte |= pmap->pmap_bits[PG_W_IDX];
4998 if (va < VM_MAX_USER_ADDRESS)
4999 newpte |= pmap->pmap_bits[PG_U_IDX];
5000 if (pte_pv && (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) == 0)
5001 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
5002 // if (pmap == &kernel_pmap)
5003 // newpte |= pgeflag;
5004 newpte |= pmap->pmap_cache_bits[m->pat_mode];
5005 if (m->flags & PG_FICTITIOUS)
5006 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
5009 * It is possible for multiple faults to occur in threaded
5010 * environments, the existing pte might be correct.
5012 if (((origpte ^ newpte) &
5013 ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
5014 pmap->pmap_bits[PG_A_IDX])) == 0) {
5019 * Ok, either the address changed or the protection or wiring
5022 * Clear the current entry, interlocking the removal. For managed
5023 * pte's this will also flush the modified state to the vm_page.
5024 * Atomic ops are mandatory in order to ensure that PG_M events are
5025 * not lost during any transition.
5027 * WARNING: The caller has busied the new page but not the original
5028 * vm_page which we are trying to replace. Because we hold
5029 * the pte_pv lock, but have not busied the page, PG bits
5030 * can be cleared out from under us.
5033 if (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
5035 * Old page was managed. Expect pte_pv to exist.
5036 * (it might also exist if the old page was unmanaged).
5038 * NOTE: pt_pv won't exist for a kernel page
5039 * (managed or otherwise).
5041 * NOTE: We may be reusing the pte_pv so we do not
5042 * destroy it in pmap_remove_pv_pte().
5044 KKASSERT(pte_pv && pte_pv->pv_m);
5045 if (prot & VM_PROT_NOSYNC) {
5046 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
5048 pmap_inval_bulk_t bulk;
5050 pmap_inval_bulk_init(&bulk, pmap);
5051 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
5052 pmap_inval_bulk_flush(&bulk);
5054 pmap_remove_pv_page(pte_pv);
5055 /* will either set pte_pv->pv_m or pv_free() later */
5058 * Old page was not managed. If we have a pte_pv
5059 * it better not have a pv_m assigned to it. If the
5060 * new page is managed the pte_pv will be destroyed
5061 * near the end (we need its interlock).
5063 * NOTE: We leave the wire count on the PT page
5064 * intact for the followup enter, but adjust
5065 * the wired-pages count on the pmap.
5067 KKASSERT(pte_pv == NULL);
5068 if (prot & VM_PROT_NOSYNC) {
5070 * NOSYNC (no mmu sync) requested.
5072 (void)pte_load_clear(ptep);
5073 cpu_invlpg((void *)va);
5078 pmap_inval_smp(pmap, va, 1, ptep, 0);
5082 * We must adjust pm_stats manually for unmanaged
5086 atomic_add_long(&pmap->pm_stats.
5087 resident_count, -1);
5089 if (origpte & pmap->pmap_bits[PG_W_IDX]) {
5090 atomic_add_long(&pmap->pm_stats.
5094 KKASSERT(*ptep == 0);
5098 if (pmap_enter_debug > 0) {
5100 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
5101 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
5103 origpte, newpte, ptep,
5104 pte_pv, pt_pv, opa, prot);
5108 if ((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5110 * Entering an unmanaged page. We must wire the pt_pv unless
5111 * we retained the wiring from an unmanaged page we had
5112 * removed (if we retained it via pte_pv that will go away
5115 if (pt_pv && (opa == 0 ||
5116 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]))) {
5117 vm_page_wire_quick(pt_pv->pv_m);
5120 atomic_add_long(&pmap->pm_stats.wired_count, 1);
5123 * Unmanaged pages need manual resident_count tracking.
5126 atomic_add_long(&pt_pv->pv_pmap->pm_stats.
5129 if (newpte & pmap->pmap_bits[PG_RW_IDX])
5130 vm_page_flag_set(m, PG_WRITEABLE);
5133 * Entering a managed page. Our pte_pv takes care of the
5134 * PT wiring, so if we had removed an unmanaged page before
5137 * We have to take care of the pmap wired count ourselves.
5139 * Enter on the PV list if part of our managed memory.
5141 KKASSERT(pte_pv && (pte_pv->pv_m == NULL || pte_pv->pv_m == m));
5142 vm_page_spin_lock(m);
5144 pmap_page_stats_adding(m);
5145 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
5146 vm_page_flag_set(m, PG_MAPPED);
5147 if (newpte & pmap->pmap_bits[PG_RW_IDX])
5148 vm_page_flag_set(m, PG_WRITEABLE);
5149 vm_page_spin_unlock(m);
5152 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5153 vm_page_unwire_quick(pt_pv->pv_m);
5157 * Adjust pmap wired pages count for new entry.
5160 atomic_add_long(&pte_pv->pv_pmap->pm_stats.
5166 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
5168 * User VMAs do not because those will be zero->non-zero, so no
5169 * stale entries to worry about at this point.
5171 * For KVM there appear to still be issues. Theoretically we
5172 * should be able to scrap the interlocks entirely but we
5175 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
5176 pmap_inval_smp(pmap, va, 1, ptep, newpte);
5178 origpte = atomic_swap_long(ptep, newpte);
5179 if (origpte & pmap->pmap_bits[PG_M_IDX]) {
5180 kprintf("pmap [M] race @ %016jx\n", va);
5181 atomic_set_long(ptep, pmap->pmap_bits[PG_M_IDX]);
5184 cpu_invlpg((void *)va);
5191 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
5192 (m->flags & PG_MAPPED));
5195 * Cleanup the pv entry, allowing other accessors. If the new page
5196 * is not managed but we have a pte_pv (which was locking our
5197 * operation), we can free it now. pte_pv->pv_m should be NULL.
5199 if (pte_pv && (newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0) {
5200 pv_free(pte_pv, pt_pv);
5201 } else if (pte_pv) {
5203 } else if (pte_placemark) {
5204 pv_placemarker_wakeup(pmap, pte_placemark);
5211 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
5212 * This code also assumes that the pmap has no pre-existing entry for this
5215 * This code currently may only be used on user pmaps, not kernel_pmap.
5218 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
5220 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
5224 * Make a temporary mapping for a physical address. This is only intended
5225 * to be used for panic dumps.
5227 * The caller is responsible for calling smp_invltlb().
5230 pmap_kenter_temporary(vm_paddr_t pa, long i)
5232 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
5233 return ((void *)crashdumpmap);
5236 #define MAX_INIT_PT (96)
5239 * This routine preloads the ptes for a given object into the specified pmap.
5240 * This eliminates the blast of soft faults on process startup and
5241 * immediately after an mmap.
5243 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
5246 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
5247 vm_object_t object, vm_pindex_t pindex,
5248 vm_size_t size, int limit)
5250 struct rb_vm_page_scan_info info;
5255 * We can't preinit if read access isn't set or there is no pmap
5258 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
5262 * We can't preinit if the pmap is not the current pmap
5264 lp = curthread->td_lwp;
5265 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
5269 * Misc additional checks
5271 psize = x86_64_btop(size);
5273 if ((object->type != OBJT_VNODE) ||
5274 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
5275 (object->resident_page_count > MAX_INIT_PT))) {
5279 if (pindex + psize > object->size) {
5280 if (object->size < pindex)
5282 psize = object->size - pindex;
5289 * If everything is segment-aligned do not pre-init here. Instead
5290 * allow the normal vm_fault path to pass a segment hint to
5291 * pmap_enter() which will then use an object-referenced shared
5294 if ((addr & SEG_MASK) == 0 &&
5295 (ctob(psize) & SEG_MASK) == 0 &&
5296 (ctob(pindex) & SEG_MASK) == 0) {
5301 * Use a red-black scan to traverse the requested range and load
5302 * any valid pages found into the pmap.
5304 * We cannot safely scan the object's memq without holding the
5307 info.start_pindex = pindex;
5308 info.end_pindex = pindex + psize - 1;
5313 info.object = object;
5316 * By using the NOLK scan, the callback function must be sure
5317 * to return -1 if the VM page falls out of the object.
5319 vm_object_hold_shared(object);
5320 vm_page_rb_tree_RB_SCAN_NOLK(&object->rb_memq, rb_vm_page_scancmp,
5321 pmap_object_init_pt_callback, &info);
5322 vm_object_drop(object);
5327 pmap_object_init_pt_callback(vm_page_t p, void *data)
5329 struct rb_vm_page_scan_info *info = data;
5330 vm_pindex_t rel_index;
5334 * don't allow an madvise to blow away our really
5335 * free pages allocating pv entries.
5337 if ((info->limit & MAP_PREFAULT_MADVISE) &&
5338 vmstats.v_free_count < vmstats.v_free_reserved) {
5343 * Ignore list markers and ignore pages we cannot instantly
5344 * busy (while holding the object token).
5346 if (p->flags & PG_MARKER)
5351 if (vm_page_busy_try(p, TRUE))
5354 if (vm_page_sbusy_try(p))
5357 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
5358 (p->flags & PG_FICTITIOUS) == 0) {
5359 if ((p->queue - p->pc) == PQ_CACHE) {
5360 if (hard_busy == 0) {
5361 vm_page_sbusy_drop(p);
5365 vm_page_deactivate(p);
5367 rel_index = p->pindex - info->start_pindex;
5368 pmap_enter_quick(info->pmap,
5369 info->addr + x86_64_ptob(rel_index), p);
5374 vm_page_sbusy_drop(p);
5377 * We are using an unlocked scan (that is, the scan expects its
5378 * current element to remain in the tree on return). So we have
5379 * to check here and abort the scan if it isn't.
5381 if (p->object != info->object)
5388 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5391 * Returns FALSE if it would be non-trivial or if a pte is already loaded
5394 * XXX This is safe only because page table pages are not freed.
5397 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
5401 /*spin_lock(&pmap->pm_spin);*/
5402 if ((pte = pmap_pte(pmap, addr)) != NULL) {
5403 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5404 /*spin_unlock(&pmap->pm_spin);*/
5408 /*spin_unlock(&pmap->pm_spin);*/
5413 * Change the wiring attribute for a pmap/va pair. The mapping must already
5414 * exist in the pmap. The mapping may or may not be managed. The wiring in
5415 * the page is not changed, the page is returned so the caller can adjust
5416 * its wiring (the page is not locked in any way).
5418 * Wiring is not a hardware characteristic so there is no need to invalidate
5419 * TLB. However, in an SMP environment we must use a locked bus cycle to
5420 * update the pte (if we are not using the pmap_inval_*() API that is)...
5421 * it's ok to do this for simple wiring changes.
5424 pmap_unwire(pmap_t pmap, vm_offset_t va)
5435 * Assume elements in the kernel pmap are stable
5437 if (pmap == &kernel_pmap) {
5438 if (pmap_pt(pmap, va) == 0)
5440 ptep = pmap_pte_quick(pmap, va);
5441 if (pmap_pte_v(pmap, ptep)) {
5442 if (pmap_pte_w(pmap, ptep))
5443 atomic_add_long(&pmap->pm_stats.wired_count,-1);
5444 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5445 pa = *ptep & PG_FRAME;
5446 m = PHYS_TO_VM_PAGE(pa);
5452 * We can only [un]wire pmap-local pages (we cannot wire
5455 pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5459 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5460 if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5465 if (pmap_pte_w(pmap, ptep)) {
5466 atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5469 /* XXX else return NULL so caller doesn't unwire m ? */
5471 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5473 pa = *ptep & PG_FRAME;
5474 m = PHYS_TO_VM_PAGE(pa); /* held by wired count */
5481 * Copy the range specified by src_addr/len from the source map to
5482 * the range dst_addr/len in the destination map.
5484 * This routine is only advisory and need not do anything.
5487 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5488 vm_size_t len, vm_offset_t src_addr)
5495 * Zero the specified physical page.
5497 * This function may be called from an interrupt and no locking is
5501 pmap_zero_page(vm_paddr_t phys)
5503 vm_offset_t va = PHYS_TO_DMAP(phys);
5505 pagezero((void *)va);
5511 * Zero part of a physical page by mapping it into memory and clearing
5512 * its contents with bzero.
5514 * off and size may not cover an area beyond a single hardware page.
5517 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5519 vm_offset_t virt = PHYS_TO_DMAP(phys);
5521 bzero((char *)virt + off, size);
5527 * Copy the physical page from the source PA to the target PA.
5528 * This function may be called from an interrupt. No locking
5532 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5534 vm_offset_t src_virt, dst_virt;
5536 src_virt = PHYS_TO_DMAP(src);
5537 dst_virt = PHYS_TO_DMAP(dst);
5538 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5542 * pmap_copy_page_frag:
5544 * Copy the physical page from the source PA to the target PA.
5545 * This function may be called from an interrupt. No locking
5549 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5551 vm_offset_t src_virt, dst_virt;
5553 src_virt = PHYS_TO_DMAP(src);
5554 dst_virt = PHYS_TO_DMAP(dst);
5556 bcopy((char *)src_virt + (src & PAGE_MASK),
5557 (char *)dst_virt + (dst & PAGE_MASK),
5562 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5563 * this page. This count may be changed upwards or downwards in the future;
5564 * it is only necessary that true be returned for a small subset of pmaps
5565 * for proper page aging.
5568 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
5573 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5576 vm_page_spin_lock(m);
5577 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5578 if (pv->pv_pmap == pmap) {
5579 vm_page_spin_unlock(m);
5586 vm_page_spin_unlock(m);
5591 * Remove all pages from specified address space this aids process exit
5592 * speeds. Also, this code may be special cased for the current process
5596 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5598 pmap_remove_noinval(pmap, sva, eva);
5603 * pmap_testbit tests bits in pte's note that the testbit/clearbit
5604 * routines are inline, and a lot of things compile-time evaluate.
5609 pmap_testbit(vm_page_t m, int bit)
5615 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5618 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
5620 vm_page_spin_lock(m);
5621 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
5622 vm_page_spin_unlock(m);
5626 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5627 #if defined(PMAP_DIAGNOSTIC)
5628 if (pv->pv_pmap == NULL) {
5629 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5637 * If the bit being tested is the modified bit, then
5638 * mark clean_map and ptes as never
5641 * WARNING! Because we do not lock the pv, *pte can be in a
5642 * state of flux. Despite this the value of *pte
5643 * will still be related to the vm_page in some way
5644 * because the pv cannot be destroyed as long as we
5645 * hold the vm_page spin lock.
5647 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5648 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5649 if (!pmap_track_modified(pv->pv_pindex))
5653 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5654 if (*pte & pmap->pmap_bits[bit]) {
5655 vm_page_spin_unlock(m);
5659 vm_page_spin_unlock(m);
5664 * This routine is used to modify bits in ptes. Only one bit should be
5665 * specified. PG_RW requires special handling.
5667 * Caller must NOT hold any spin locks
5671 pmap_clearbit(vm_page_t m, int bit_index)
5678 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5679 if (bit_index == PG_RW_IDX)
5680 vm_page_flag_clear(m, PG_WRITEABLE);
5687 * Loop over all current mappings setting/clearing as appropos If
5688 * setting RO do we need to clear the VAC?
5690 * NOTE: When clearing PG_M we could also (not implemented) drop
5691 * through to the PG_RW code and clear PG_RW too, forcing
5692 * a fault on write to redetect PG_M for virtual kernels, but
5693 * it isn't necessary since virtual kernels invalidate the
5694 * pte when they clear the VPTE_M bit in their virtual page
5697 * NOTE: Does not re-dirty the page when clearing only PG_M.
5699 * NOTE: Because we do not lock the pv, *pte can be in a state of
5700 * flux. Despite this the value of *pte is still somewhat
5701 * related while we hold the vm_page spin lock.
5703 * *pte can be zero due to this race. Since we are clearing
5704 * bits we basically do no harm when this race occurs.
5706 if (bit_index != PG_RW_IDX) {
5707 vm_page_spin_lock(m);
5708 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5709 #if defined(PMAP_DIAGNOSTIC)
5710 if (pv->pv_pmap == NULL) {
5711 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5717 pte = pmap_pte_quick(pv->pv_pmap,
5718 pv->pv_pindex << PAGE_SHIFT);
5720 if (pbits & pmap->pmap_bits[bit_index])
5721 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5723 vm_page_spin_unlock(m);
5728 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
5732 vm_page_spin_lock(m);
5733 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5735 * don't write protect pager mappings
5737 if (!pmap_track_modified(pv->pv_pindex))
5740 #if defined(PMAP_DIAGNOSTIC)
5741 if (pv->pv_pmap == NULL) {
5742 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5750 * Skip pages which do not have PG_RW set.
5752 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5753 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5757 * We must lock the PV to be able to safely test the pte.
5759 if (pv_hold_try(pv)) {
5760 vm_page_spin_unlock(m);
5762 vm_page_spin_unlock(m);
5763 pv_lock(pv); /* held, now do a blocking lock */
5769 * Reload pte after acquiring pv.
5771 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5773 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0) {
5779 KKASSERT(pv->pv_pmap == pmap && pv->pv_m == m);
5785 nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5786 pmap->pmap_bits[PG_M_IDX]);
5787 if (pmap_inval_smp_cmpset(pmap,
5788 ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5789 pte, pbits, nbits)) {
5796 * If PG_M was found to be set while we were clearing PG_RW
5797 * we also clear PG_M (done above) and mark the page dirty.
5798 * Callers expect this behavior.
5800 * we lost pv so it cannot be used as an iterator. In fact,
5801 * because we couldn't necessarily lock it atomically it may
5802 * have moved within the list and ALSO cannot be used as an
5805 vm_page_spin_lock(m);
5806 if (pbits & pmap->pmap_bits[PG_M_IDX])
5808 vm_page_spin_unlock(m);
5812 if (bit_index == PG_RW_IDX)
5813 vm_page_flag_clear(m, PG_WRITEABLE);
5814 vm_page_spin_unlock(m);
5818 * Lower the permission for all mappings to a given page.
5820 * Page must be busied by caller. Because page is busied by caller this
5821 * should not be able to race a pmap_enter().
5824 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5826 /* JG NX support? */
5827 if ((prot & VM_PROT_WRITE) == 0) {
5828 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5830 * NOTE: pmap_clearbit(.. PG_RW) also clears
5831 * the PG_WRITEABLE flag in (m).
5833 pmap_clearbit(m, PG_RW_IDX);
5841 pmap_phys_address(vm_pindex_t ppn)
5843 return (x86_64_ptob(ppn));
5847 * Return a count of reference bits for a page, clearing those bits.
5848 * It is not necessary for every reference bit to be cleared, but it
5849 * is necessary that 0 only be returned when there are truly no
5850 * reference bits set.
5852 * XXX: The exact number of bits to check and clear is a matter that
5853 * should be tested and standardized at some point in the future for
5854 * optimal aging of shared pages.
5856 * This routine may not block.
5859 pmap_ts_referenced(vm_page_t m)
5866 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5869 vm_page_spin_lock(m);
5870 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5871 if (!pmap_track_modified(pv->pv_pindex))
5874 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5875 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5876 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5882 vm_page_spin_unlock(m);
5889 * Return whether or not the specified physical page was modified
5890 * in any physical maps.
5893 pmap_is_modified(vm_page_t m)
5897 res = pmap_testbit(m, PG_M_IDX);
5902 * Clear the modify bits on the specified physical page.
5905 pmap_clear_modify(vm_page_t m)
5907 pmap_clearbit(m, PG_M_IDX);
5911 * pmap_clear_reference:
5913 * Clear the reference bit on the specified physical page.
5916 pmap_clear_reference(vm_page_t m)
5918 pmap_clearbit(m, PG_A_IDX);
5922 * Miscellaneous support routines follow
5927 x86_64_protection_init(void)
5933 * NX supported? (boot time loader.conf override only)
5935 TUNABLE_INT_FETCH("machdep.pmap_nx_enable", &pmap_nx_enable);
5936 if (pmap_nx_enable == 0 || (amd_feature & AMDID_NX) == 0)
5937 pmap_bits_default[PG_NX_IDX] = 0;
5940 * 0 is basically read-only access, but also set the NX (no-execute)
5941 * bit when VM_PROT_EXECUTE is not specified.
5943 kp = protection_codes;
5944 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5946 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5948 * This case handled elsewhere
5952 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5956 *kp++ = pmap_bits_default[PG_NX_IDX];
5958 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5959 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5961 * Execute requires read access
5965 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5966 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5968 * Write without execute is RW|NX
5970 *kp++ = pmap_bits_default[PG_RW_IDX] |
5971 pmap_bits_default[PG_NX_IDX];
5973 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5974 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5976 * Write with execute is RW
5978 *kp++ = pmap_bits_default[PG_RW_IDX];
5985 * Map a set of physical memory pages into the kernel virtual
5986 * address space. Return a pointer to where it is mapped. This
5987 * routine is intended to be used for mapping device memory,
5990 * NOTE: We can't use pgeflag unless we invalidate the pages one at
5993 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5994 * work whether the cpu supports PAT or not. The remaining PAT
5995 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5999 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
6001 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
6005 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
6007 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
6011 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
6013 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
6017 * Map a set of physical memory pages into the kernel virtual
6018 * address space. Return a pointer to where it is mapped. This
6019 * routine is intended to be used for mapping device memory,
6023 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
6025 vm_offset_t va, tmpva, offset;
6029 offset = pa & PAGE_MASK;
6030 size = roundup(offset + size, PAGE_SIZE);
6032 va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
6034 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
6036 pa = pa & ~PAGE_MASK;
6037 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
6038 pte = vtopte(tmpva);
6040 kernel_pmap.pmap_bits[PG_RW_IDX] |
6041 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
6042 kernel_pmap.pmap_cache_bits[mode];
6043 tmpsize -= PAGE_SIZE;
6047 pmap_invalidate_range(&kernel_pmap, va, va + size);
6048 pmap_invalidate_cache_range(va, va + size);
6050 return ((void *)(va + offset));
6054 pmap_unmapdev(vm_offset_t va, vm_size_t size)
6056 vm_offset_t base, offset;
6058 base = va & ~PAGE_MASK;
6059 offset = va & PAGE_MASK;
6060 size = roundup(offset + size, PAGE_SIZE);
6061 pmap_qremove(va, size >> PAGE_SHIFT);
6062 kmem_free(&kernel_map, base, size);
6066 * Sets the memory attribute for the specified page.
6069 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
6075 * If "m" is a normal page, update its direct mapping. This update
6076 * can be relied upon to perform any cache operations that are
6077 * required for data coherence.
6079 if ((m->flags & PG_FICTITIOUS) == 0)
6080 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
6084 * Change the PAT attribute on an existing kernel memory map. Caller
6085 * must ensure that the virtual memory in question is not accessed
6086 * during the adjustment.
6089 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
6096 panic("pmap_change_attr: va is NULL");
6097 base = trunc_page(va);
6101 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
6102 kernel_pmap.pmap_cache_bits[mode];
6107 changed = 1; /* XXX: not optimal */
6110 * Flush CPU caches if required to make sure any data isn't cached that
6111 * shouldn't be, etc.
6114 pmap_invalidate_range(&kernel_pmap, base, va);
6115 pmap_invalidate_cache_range(base, va);
6120 * perform the pmap work for mincore
6123 pmap_mincore(pmap_t pmap, vm_offset_t addr)
6125 pt_entry_t *ptep, pte;
6129 ptep = pmap_pte(pmap, addr);
6131 if (ptep && (pte = *ptep) != 0) {
6134 val = MINCORE_INCORE;
6135 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
6138 pa = pte & PG_FRAME;
6140 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
6143 m = PHYS_TO_VM_PAGE(pa);
6148 if (pte & pmap->pmap_bits[PG_M_IDX])
6149 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
6151 * Modified by someone
6153 else if (m && (m->dirty || pmap_is_modified(m)))
6154 val |= MINCORE_MODIFIED_OTHER;
6158 if (pte & pmap->pmap_bits[PG_A_IDX])
6159 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
6162 * Referenced by someone
6164 else if (m && ((m->flags & PG_REFERENCED) ||
6165 pmap_ts_referenced(m))) {
6166 val |= MINCORE_REFERENCED_OTHER;
6167 vm_page_flag_set(m, PG_REFERENCED);
6176 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
6177 * vmspace will be ref'd and the old one will be deref'd.
6179 * The vmspace for all lwps associated with the process will be adjusted
6180 * and cr3 will be reloaded if any lwp is the current lwp.
6182 * The process must hold the vmspace->vm_map.token for oldvm and newvm
6185 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
6187 struct vmspace *oldvm;
6190 oldvm = p->p_vmspace;
6191 if (oldvm != newvm) {
6194 p->p_vmspace = newvm;
6195 KKASSERT(p->p_nthreads == 1);
6196 lp = RB_ROOT(&p->p_lwp_tree);
6197 pmap_setlwpvm(lp, newvm);
6204 * Set the vmspace for a LWP. The vmspace is almost universally set the
6205 * same as the process vmspace, but virtual kernels need to swap out contexts
6206 * on a per-lwp basis.
6208 * Caller does not necessarily hold any vmspace tokens. Caller must control
6209 * the lwp (typically be in the context of the lwp). We use a critical
6210 * section to protect against statclock and hardclock (statistics collection).
6213 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
6215 struct vmspace *oldvm;
6218 oldvm = lp->lwp_vmspace;
6220 if (oldvm != newvm) {
6222 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
6223 lp->lwp_vmspace = newvm;
6224 if (curthread->td_lwp == lp) {
6225 pmap = vmspace_pmap(newvm);
6226 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
6227 if (pmap->pm_active_lock & CPULOCK_EXCL)
6228 pmap_interlock_wait(newvm);
6229 #if defined(SWTCH_OPTIM_STATS)
6232 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
6233 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
6234 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
6235 curthread->td_pcb->pcb_cr3 = KPML4phys;
6237 panic("pmap_setlwpvm: unknown pmap type\n");
6239 load_cr3(curthread->td_pcb->pcb_cr3);
6240 pmap = vmspace_pmap(oldvm);
6241 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
6249 * Called when switching to a locked pmap, used to interlock against pmaps
6250 * undergoing modifications to prevent us from activating the MMU for the
6251 * target pmap until all such modifications have completed. We have to do
6252 * this because the thread making the modifications has already set up its
6253 * SMP synchronization mask.
6255 * This function cannot sleep!
6260 pmap_interlock_wait(struct vmspace *vm)
6262 struct pmap *pmap = &vm->vm_pmap;
6264 if (pmap->pm_active_lock & CPULOCK_EXCL) {
6266 KKASSERT(curthread->td_critcount >= 2);
6267 DEBUG_PUSH_INFO("pmap_interlock_wait");
6268 while (pmap->pm_active_lock & CPULOCK_EXCL) {
6270 lwkt_process_ipiq();
6278 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
6281 if ((obj == NULL) || (size < NBPDR) ||
6282 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
6286 addr = roundup2(addr, NBPDR);
6291 * Used by kmalloc/kfree, page already exists at va
6294 pmap_kvtom(vm_offset_t va)
6296 pt_entry_t *ptep = vtopte(va);
6298 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
6299 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
6303 * Initialize machine-specific shared page directory support. This
6304 * is executed when a VM object is created.
6307 pmap_object_init(vm_object_t object)
6309 object->md.pmap_rw = NULL;
6310 object->md.pmap_ro = NULL;
6314 * Clean up machine-specific shared page directory support. This
6315 * is executed when a VM object is destroyed.
6318 pmap_object_free(vm_object_t object)
6322 if ((pmap = object->md.pmap_rw) != NULL) {
6323 object->md.pmap_rw = NULL;
6324 pmap_remove_noinval(pmap,
6325 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6326 CPUMASK_ASSZERO(pmap->pm_active);
6329 kfree(pmap, M_OBJPMAP);
6331 if ((pmap = object->md.pmap_ro) != NULL) {
6332 object->md.pmap_ro = NULL;
6333 pmap_remove_noinval(pmap,
6334 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
6335 CPUMASK_ASSZERO(pmap->pm_active);
6338 kfree(pmap, M_OBJPMAP);
6343 * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6344 * VM page and issue a pginfo->callback.
6346 * We are expected to dispose of any non-NULL pte_pv.
6350 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
6351 pv_entry_t pte_pv, vm_pindex_t *pte_placemark,
6352 pv_entry_t pt_pv, int sharept,
6353 vm_offset_t va, pt_entry_t *ptep, void *arg)
6355 struct pmap_pgscan_info *pginfo = arg;
6360 * Try to busy the page while we hold the pte_pv locked.
6362 KKASSERT(pte_pv->pv_m);
6363 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
6364 if (vm_page_busy_try(m, TRUE) == 0) {
6365 if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
6367 * The callback is issued with the pte_pv
6368 * unlocked and put away, and the pt_pv
6373 vm_page_wire_quick(pt_pv->pv_m);
6376 if (pginfo->callback(pginfo, va, m) < 0)
6380 vm_page_unwire_quick(pt_pv->pv_m);
6387 ++pginfo->busycount;
6392 * Shared page table or unmanaged page (sharept or !sharept)
6394 pv_placemarker_wakeup(pmap, pte_placemark);
6399 pmap_pgscan(struct pmap_pgscan_info *pginfo)
6401 struct pmap_scan_info info;
6403 pginfo->offset = pginfo->beg_addr;
6404 info.pmap = pginfo->pmap;
6405 info.sva = pginfo->beg_addr;
6406 info.eva = pginfo->end_addr;
6407 info.func = pmap_pgscan_callback;
6409 pmap_scan(&info, 0);
6411 pginfo->offset = pginfo->end_addr;
6415 * Wait for a placemarker that we do not own to clear. The placemarker
6416 * in question is not necessarily set to the pindex we want, we may have
6417 * to wait on the element because we want to reserve it ourselves.
6419 * NOTE: PM_PLACEMARK_WAKEUP sets a bit which is already set in
6420 * PM_NOPLACEMARK, so it does not interfere with placemarks
6421 * which have already been woken up.
6425 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
6427 if (*pmark != PM_NOPLACEMARK) {
6428 atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
6429 tsleep_interlock(pmark, 0);
6430 if (*pmark != PM_NOPLACEMARK)
6431 tsleep(pmark, PINTERLOCKED, "pvplw", 0);
6436 * Wakeup a placemarker that we own. Replace the entry with
6437 * PM_NOPLACEMARK and issue a wakeup() if necessary.
6441 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
6445 pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
6446 KKASSERT(pindex != PM_NOPLACEMARK);
6447 if (pindex & PM_PLACEMARK_WAKEUP)