2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
9 * Copyright (c) 2011-2012 Matthew Dillon
10 * All rights reserved.
12 * This code is derived from software contributed to Berkeley by
13 * the Systems Programming Group of the University of Utah Computer
14 * Science Department and William Jolitz of UUNET Technologies Inc.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * Manage physical address maps for x86-64 systems.
49 #include "opt_disable_pse.h"
52 #include "opt_msgbuf.h"
54 #include <sys/param.h>
55 #include <sys/kernel.h>
57 #include <sys/msgbuf.h>
58 #include <sys/vmmeter.h>
60 #include <sys/systm.h>
63 #include <vm/vm_param.h>
64 #include <sys/sysctl.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_pageout.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_zone.h>
76 #include <sys/thread2.h>
77 #include <sys/sysref2.h>
78 #include <sys/spinlock2.h>
79 #include <vm/vm_page2.h>
81 #include <machine/cputypes.h>
82 #include <machine/md_var.h>
83 #include <machine/specialreg.h>
84 #include <machine/smp.h>
85 #include <machine_base/apic/apicreg.h>
86 #include <machine/globaldata.h>
87 #include <machine/pmap.h>
88 #include <machine/pmap_inval.h>
89 #include <machine/inttypes.h>
93 #define PMAP_KEEP_PDIRS
94 #ifndef PMAP_SHPGPERPROC
95 #define PMAP_SHPGPERPROC 2000
98 #if defined(DIAGNOSTIC)
99 #define PMAP_DIAGNOSTIC
105 * pmap debugging will report who owns a pv lock when blocking.
109 #define PMAP_DEBUG_DECL ,const char *func, int lineno
110 #define PMAP_DEBUG_ARGS , __func__, __LINE__
111 #define PMAP_DEBUG_COPY , func, lineno
113 #define pv_get(pmap, pindex) _pv_get(pmap, pindex \
115 #define pv_lock(pv) _pv_lock(pv \
117 #define pv_hold_try(pv) _pv_hold_try(pv \
119 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \
124 #define PMAP_DEBUG_DECL
125 #define PMAP_DEBUG_ARGS
126 #define PMAP_DEBUG_COPY
128 #define pv_get(pmap, pindex) _pv_get(pmap, pindex)
129 #define pv_lock(pv) _pv_lock(pv)
130 #define pv_hold_try(pv) _pv_hold_try(pv)
131 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp)
136 * Get PDEs and PTEs for user/kernel address space
138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
140 #define pmap_pde_v(pmap, pte) ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
141 #define pmap_pte_w(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
142 #define pmap_pte_m(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
143 #define pmap_pte_u(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
144 #define pmap_pte_v(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
147 * Given a map and a machine independent protection code,
148 * convert to a vax protection code.
150 #define pte_prot(m, p) \
151 (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
152 static int protection_codes[PROTECTION_CODES_SIZE];
154 struct pmap kernel_pmap;
156 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
158 vm_paddr_t avail_start; /* PA of first available physical page */
159 vm_paddr_t avail_end; /* PA of last available physical page */
160 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
161 vm_offset_t virtual2_end;
162 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
163 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
164 vm_offset_t KvaStart; /* VA start of KVA space */
165 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
166 vm_offset_t KvaSize; /* max size of kernel virtual address space */
167 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
168 //static int pgeflag; /* PG_G or-in */
169 //static int pseflag; /* PG_PS or-in */
173 static vm_paddr_t dmaplimit;
175 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
177 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE]; /* PAT -> PG_ bits */
178 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
180 static uint64_t KPTbase;
181 static uint64_t KPTphys;
182 static uint64_t KPDphys; /* phys addr of kernel level 2 */
183 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
184 uint64_t KPDPphys; /* phys addr of kernel level 3 */
185 uint64_t KPML4phys; /* phys addr of kernel level 4 */
187 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
188 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
191 * Data for the pv entry allocation mechanism
193 static vm_zone_t pvzone;
194 static struct vm_zone pvzone_store;
195 static struct vm_object pvzone_obj;
196 static int pv_entry_max=0, pv_entry_high_water=0;
197 static int pmap_pagedaemon_waken = 0;
198 static struct pv_entry *pvinit;
201 * All those kernel PT submaps that BSD is so fond of
203 pt_entry_t *CMAP1 = NULL, *ptmmap;
204 caddr_t CADDR1 = NULL, ptvmmap = NULL;
205 static pt_entry_t *msgbufmap;
206 struct msgbuf *msgbufp=NULL;
209 * PMAP default PG_* bits. Needed to be able to add
210 * EPT/NPT pagetable pmap_bits for the VMM module
212 uint64_t pmap_bits_default[] = {
213 REGULAR_PMAP, /* TYPE_IDX 0 */
214 X86_PG_V, /* PG_V_IDX 1 */
215 X86_PG_RW, /* PG_RW_IDX 2 */
216 X86_PG_U, /* PG_U_IDX 3 */
217 X86_PG_A, /* PG_A_IDX 4 */
218 X86_PG_M, /* PG_M_IDX 5 */
219 X86_PG_PS, /* PG_PS_IDX3 6 */
220 X86_PG_G, /* PG_G_IDX 7 */
221 X86_PG_AVAIL1, /* PG_AVAIL1_IDX 8 */
222 X86_PG_AVAIL2, /* PG_AVAIL2_IDX 9 */
223 X86_PG_AVAIL3, /* PG_AVAIL3_IDX 10 */
224 X86_PG_NC_PWT | X86_PG_NC_PCD, /* PG_N_IDX 11 */
229 static pt_entry_t *pt_crashdumpmap;
230 static caddr_t crashdumpmap;
232 static int pmap_debug = 0;
233 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
234 &pmap_debug, 0, "Debug pmap's");
236 static int pmap_enter_debug = 0;
237 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
238 &pmap_enter_debug, 0, "Debug pmap_enter's");
240 static int pmap_yield_count = 64;
241 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
242 &pmap_yield_count, 0, "Yield during init_pt/release");
243 static int pmap_mmu_optimize = 0;
244 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
245 &pmap_mmu_optimize, 0, "Share page table pages when possible");
246 int pmap_fast_kernel_cpusync = 0;
247 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
248 &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
249 int pmap_dynamic_delete = -1;
250 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
251 &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
255 /* Standard user access funtions */
256 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
258 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
259 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
260 extern int std_fubyte (const void *base);
261 extern int std_subyte (void *base, int byte);
262 extern long std_fuword (const void *base);
263 extern int std_suword (void *base, long word);
264 extern int std_suword32 (void *base, int word);
266 static void pv_hold(pv_entry_t pv);
267 static int _pv_hold_try(pv_entry_t pv
269 static void pv_drop(pv_entry_t pv);
270 static void _pv_lock(pv_entry_t pv
272 static void pv_unlock(pv_entry_t pv);
273 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
275 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
277 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
278 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
279 static void pv_put(pv_entry_t pv);
280 static void pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway);
281 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
282 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
284 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
285 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
286 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
287 pmap_inval_bulk_t *bulk, int destroy);
288 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
289 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
290 pmap_inval_bulk_t *bulk);
292 struct pmap_scan_info;
293 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
294 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
295 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
296 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
297 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
298 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
300 static void i386_protection_init (void);
301 static void create_pagetables(vm_paddr_t *firstaddr);
302 static void pmap_remove_all (vm_page_t m);
303 static boolean_t pmap_testbit (vm_page_t m, int bit);
305 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
306 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
308 static void pmap_pinit_defaults(struct pmap *pmap);
310 static unsigned pdir4mb;
313 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
315 if (pv1->pv_pindex < pv2->pv_pindex)
317 if (pv1->pv_pindex > pv2->pv_pindex)
322 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
323 pv_entry_compare, vm_pindex_t, pv_pindex);
327 pmap_page_stats_adding(vm_page_t m)
329 globaldata_t gd = mycpu;
331 if (TAILQ_EMPTY(&m->md.pv_list)) {
332 ++gd->gd_vmtotal.t_arm;
333 } else if (TAILQ_FIRST(&m->md.pv_list) ==
334 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
335 ++gd->gd_vmtotal.t_armshr;
336 ++gd->gd_vmtotal.t_avmshr;
338 ++gd->gd_vmtotal.t_avmshr;
344 pmap_page_stats_deleting(vm_page_t m)
346 globaldata_t gd = mycpu;
348 if (TAILQ_EMPTY(&m->md.pv_list)) {
349 --gd->gd_vmtotal.t_arm;
350 } else if (TAILQ_FIRST(&m->md.pv_list) ==
351 TAILQ_LAST(&m->md.pv_list, md_page_pv_list)) {
352 --gd->gd_vmtotal.t_armshr;
353 --gd->gd_vmtotal.t_avmshr;
355 --gd->gd_vmtotal.t_avmshr;
360 * Move the kernel virtual free pointer to the next
361 * 2MB. This is used to help improve performance
362 * by using a large (2MB) page for much of the kernel
363 * (.text, .data, .bss)
367 pmap_kmem_choose(vm_offset_t addr)
369 vm_offset_t newaddr = addr;
371 newaddr = roundup2(addr, NBPDR);
378 * Super fast pmap_pte routine best used when scanning the pv lists.
379 * This eliminates many course-grained invltlb calls. Note that many of
380 * the pv list scans are across different pmaps and it is very wasteful
381 * to do an entire invltlb when checking a single mapping.
383 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
387 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
389 return pmap_pte(pmap, va);
393 * Returns the pindex of a page table entry (representing a terminal page).
394 * There are NUPTE_TOTAL page table entries possible (a huge number)
396 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
397 * We want to properly translate negative KVAs.
401 pmap_pte_pindex(vm_offset_t va)
403 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
407 * Returns the pindex of a page table.
411 pmap_pt_pindex(vm_offset_t va)
413 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
417 * Returns the pindex of a page directory.
421 pmap_pd_pindex(vm_offset_t va)
423 return (NUPTE_TOTAL + NUPT_TOTAL +
424 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
429 pmap_pdp_pindex(vm_offset_t va)
431 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
432 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
437 pmap_pml4_pindex(void)
439 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
443 * Return various clipped indexes for a given VA
445 * Returns the index of a pt in a page directory, representing a page
450 pmap_pt_index(vm_offset_t va)
452 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
456 * Returns the index of a pd in a page directory page, representing a page
461 pmap_pd_index(vm_offset_t va)
463 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
467 * Returns the index of a pdp in the pml4 table, representing a page
472 pmap_pdp_index(vm_offset_t va)
474 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
478 * Generic procedure to index a pte from a pt, pd, or pdp.
480 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
481 * a page table page index but is instead of PV lookup index.
485 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
489 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
490 return(&pte[pindex]);
494 * Return pointer to PDP slot in the PML4
498 pmap_pdp(pmap_t pmap, vm_offset_t va)
500 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
504 * Return pointer to PD slot in the PDP given a pointer to the PDP
508 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
512 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
513 return (&pd[pmap_pd_index(va)]);
517 * Return pointer to PD slot in the PDP.
521 pmap_pd(pmap_t pmap, vm_offset_t va)
525 pdp = pmap_pdp(pmap, va);
526 if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
528 return (pmap_pdp_to_pd(*pdp, va));
532 * Return pointer to PT slot in the PD given a pointer to the PD
536 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
540 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
541 return (&pt[pmap_pt_index(va)]);
545 * Return pointer to PT slot in the PD
547 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
548 * so we cannot lookup the PD via the PDP. Instead we
549 * must look it up via the pmap.
553 pmap_pt(pmap_t pmap, vm_offset_t va)
557 vm_pindex_t pd_pindex;
559 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
560 pd_pindex = pmap_pd_pindex(va);
561 spin_lock(&pmap->pm_spin);
562 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
563 spin_unlock(&pmap->pm_spin);
564 if (pv == NULL || pv->pv_m == NULL)
566 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
568 pd = pmap_pd(pmap, va);
569 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
571 return (pmap_pd_to_pt(*pd, va));
576 * Return pointer to PTE slot in the PT given a pointer to the PT
580 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
584 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
585 return (&pte[pmap_pte_index(va)]);
589 * Return pointer to PTE slot in the PT
593 pmap_pte(pmap_t pmap, vm_offset_t va)
597 pt = pmap_pt(pmap, va);
598 if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
600 if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
601 return ((pt_entry_t *)pt);
602 return (pmap_pt_to_pte(*pt, va));
606 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
607 * the PT layer. This will speed up core pmap operations considerably.
609 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
610 * must be in a known associated state (typically by being locked when
611 * the pmap spinlock isn't held). We allow the race for that case.
615 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
617 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
618 pv->pv_pmap->pm_pvhint = pv;
623 * Return address of PT slot in PD (KVM only)
625 * Cannot be used for user page tables because it might interfere with
626 * the shared page-table-page optimization (pmap_mmu_optimize).
630 vtopt(vm_offset_t va)
632 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
633 NPML4EPGSHIFT)) - 1);
635 return (PDmap + ((va >> PDRSHIFT) & mask));
639 * KVM - return address of PTE slot in PT
643 vtopte(vm_offset_t va)
645 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
646 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
648 return (PTmap + ((va >> PAGE_SHIFT) & mask));
652 allocpages(vm_paddr_t *firstaddr, long n)
657 bzero((void *)ret, n * PAGE_SIZE);
658 *firstaddr += n * PAGE_SIZE;
664 create_pagetables(vm_paddr_t *firstaddr)
666 long i; /* must be 64 bits */
672 * We are running (mostly) V=P at this point
674 * Calculate NKPT - number of kernel page tables. We have to
675 * accomodoate prealloction of the vm_page_array, dump bitmap,
676 * MSGBUF_SIZE, and other stuff. Be generous.
678 * Maxmem is in pages.
680 * ndmpdp is the number of 1GB pages we wish to map.
682 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
683 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
685 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
688 * Starting at the beginning of kvm (not KERNBASE).
690 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
691 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
692 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
693 ndmpdp) + 511) / 512;
697 * Starting at KERNBASE - map 2G worth of page table pages.
698 * KERNBASE is offset -2G from the end of kvm.
700 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
705 KPTbase = allocpages(firstaddr, nkpt_base);
706 KPTphys = allocpages(firstaddr, nkpt_phys);
707 KPML4phys = allocpages(firstaddr, 1);
708 KPDPphys = allocpages(firstaddr, NKPML4E);
709 KPDphys = allocpages(firstaddr, NKPDPE);
712 * Calculate the page directory base for KERNBASE,
713 * that is where we start populating the page table pages.
714 * Basically this is the end - 2.
716 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
718 DMPDPphys = allocpages(firstaddr, NDMPML4E);
719 if ((amd_feature & AMDID_PAGE1GB) == 0)
720 DMPDphys = allocpages(firstaddr, ndmpdp);
721 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
724 * Fill in the underlying page table pages for the area around
725 * KERNBASE. This remaps low physical memory to KERNBASE.
727 * Read-only from zero to physfree
728 * XXX not fully used, underneath 2M pages
730 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
731 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
732 ((pt_entry_t *)KPTbase)[i] |=
733 pmap_bits_default[PG_RW_IDX] |
734 pmap_bits_default[PG_V_IDX] |
735 pmap_bits_default[PG_G_IDX];
739 * Now map the initial kernel page tables. One block of page
740 * tables is placed at the beginning of kernel virtual memory,
741 * and another block is placed at KERNBASE to map the kernel binary,
742 * data, bss, and initial pre-allocations.
744 for (i = 0; i < nkpt_base; i++) {
745 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
746 ((pd_entry_t *)KPDbase)[i] |=
747 pmap_bits_default[PG_RW_IDX] |
748 pmap_bits_default[PG_V_IDX];
750 for (i = 0; i < nkpt_phys; i++) {
751 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
752 ((pd_entry_t *)KPDphys)[i] |=
753 pmap_bits_default[PG_RW_IDX] |
754 pmap_bits_default[PG_V_IDX];
758 * Map from zero to end of allocations using 2M pages as an
759 * optimization. This will bypass some of the KPTBase pages
760 * above in the KERNBASE area.
762 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
763 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
764 ((pd_entry_t *)KPDbase)[i] |=
765 pmap_bits_default[PG_RW_IDX] |
766 pmap_bits_default[PG_V_IDX] |
767 pmap_bits_default[PG_PS_IDX] |
768 pmap_bits_default[PG_G_IDX];
772 * And connect up the PD to the PDP. The kernel pmap is expected
773 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
775 for (i = 0; i < NKPDPE; i++) {
776 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
777 KPDphys + (i << PAGE_SHIFT);
778 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
779 pmap_bits_default[PG_RW_IDX] |
780 pmap_bits_default[PG_V_IDX] |
781 pmap_bits_default[PG_U_IDX];
785 * Now set up the direct map space using either 2MB or 1GB pages
786 * Preset PG_M and PG_A because demotion expects it.
788 * When filling in entries in the PD pages make sure any excess
789 * entries are set to zero as we allocated enough PD pages
791 if ((amd_feature & AMDID_PAGE1GB) == 0) {
792 for (i = 0; i < NPDEPG * ndmpdp; i++) {
793 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
794 ((pd_entry_t *)DMPDphys)[i] |=
795 pmap_bits_default[PG_RW_IDX] |
796 pmap_bits_default[PG_V_IDX] |
797 pmap_bits_default[PG_PS_IDX] |
798 pmap_bits_default[PG_G_IDX] |
799 pmap_bits_default[PG_M_IDX] |
800 pmap_bits_default[PG_A_IDX];
804 * And the direct map space's PDP
806 for (i = 0; i < ndmpdp; i++) {
807 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
809 ((pdp_entry_t *)DMPDPphys)[i] |=
810 pmap_bits_default[PG_RW_IDX] |
811 pmap_bits_default[PG_V_IDX] |
812 pmap_bits_default[PG_U_IDX];
815 for (i = 0; i < ndmpdp; i++) {
816 ((pdp_entry_t *)DMPDPphys)[i] =
817 (vm_paddr_t)i << PDPSHIFT;
818 ((pdp_entry_t *)DMPDPphys)[i] |=
819 pmap_bits_default[PG_RW_IDX] |
820 pmap_bits_default[PG_V_IDX] |
821 pmap_bits_default[PG_PS_IDX] |
822 pmap_bits_default[PG_G_IDX] |
823 pmap_bits_default[PG_M_IDX] |
824 pmap_bits_default[PG_A_IDX];
828 /* And recursively map PML4 to itself in order to get PTmap */
829 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
830 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
831 pmap_bits_default[PG_RW_IDX] |
832 pmap_bits_default[PG_V_IDX] |
833 pmap_bits_default[PG_U_IDX];
836 * Connect the Direct Map slots up to the PML4
838 for (j = 0; j < NDMPML4E; ++j) {
839 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
840 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
841 pmap_bits_default[PG_RW_IDX] |
842 pmap_bits_default[PG_V_IDX] |
843 pmap_bits_default[PG_U_IDX];
847 * Connect the KVA slot up to the PML4
849 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
850 ((pdp_entry_t *)KPML4phys)[KPML4I] |=
851 pmap_bits_default[PG_RW_IDX] |
852 pmap_bits_default[PG_V_IDX] |
853 pmap_bits_default[PG_U_IDX];
857 * Bootstrap the system enough to run with virtual memory.
859 * On the i386 this is called after mapping has already been enabled
860 * and just syncs the pmap module with what has already been done.
861 * [We can't call it easily with mapping off since the kernel is not
862 * mapped with PA == VA, hence we would have to relocate every address
863 * from the linked base (virtual) address "KERNBASE" to the actual
864 * (physical) address starting relative to 0]
867 pmap_bootstrap(vm_paddr_t *firstaddr)
872 KvaStart = VM_MIN_KERNEL_ADDRESS;
873 KvaEnd = VM_MAX_KERNEL_ADDRESS;
874 KvaSize = KvaEnd - KvaStart;
876 avail_start = *firstaddr;
879 * Create an initial set of page tables to run the kernel in.
881 create_pagetables(firstaddr);
883 virtual2_start = KvaStart;
884 virtual2_end = PTOV_OFFSET;
886 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
887 virtual_start = pmap_kmem_choose(virtual_start);
889 virtual_end = VM_MAX_KERNEL_ADDRESS;
891 /* XXX do %cr0 as well */
892 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
896 * Initialize protection array.
898 i386_protection_init();
901 * The kernel's pmap is statically allocated so we don't have to use
902 * pmap_create, which is unlikely to work correctly at this part of
903 * the boot sequence (XXX and which no longer exists).
905 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
906 kernel_pmap.pm_count = 1;
907 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
908 RB_INIT(&kernel_pmap.pm_pvroot);
909 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
910 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
913 * Reserve some special page table entries/VA space for temporary
916 #define SYSMAP(c, p, v, n) \
917 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
923 * CMAP1/CMAP2 are used for zeroing and copying pages.
925 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
930 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
933 * ptvmmap is used for reading arbitrary physical pages via
936 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
939 * msgbufp is used to map the system message buffer.
940 * XXX msgbufmap is not used.
942 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
943 atop(round_page(MSGBUF_SIZE)))
946 virtual_start = pmap_kmem_choose(virtual_start);
951 * PG_G is terribly broken on SMP because we IPI invltlb's in some
952 * cases rather then invl1pg. Actually, I don't even know why it
953 * works under UP because self-referential page table mappings
958 * Initialize the 4MB page size flag
962 * The 4MB page version of the initial
963 * kernel page mapping.
967 #if !defined(DISABLE_PSE)
968 if (cpu_feature & CPUID_PSE) {
971 * Note that we have enabled PSE mode
973 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
974 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
975 ptditmp &= ~(NBPDR - 1);
976 ptditmp |= pmap_bits_default[PG_V_IDX] |
977 pmap_bits_default[PG_RW_IDX] |
978 pmap_bits_default[PG_PS_IDX] |
979 pmap_bits_default[PG_U_IDX];
986 /* Initialize the PAT MSR */
988 pmap_pinit_defaults(&kernel_pmap);
990 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
991 &pmap_fast_kernel_cpusync);
1005 * Default values mapping PATi,PCD,PWT bits at system reset.
1006 * The default values effectively ignore the PATi bit by
1007 * repeating the encodings for 0-3 in 4-7, and map the PCD
1008 * and PWT bit combinations to the expected PAT types.
1010 pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | /* 000 */
1011 PAT_VALUE(1, PAT_WRITE_THROUGH) | /* 001 */
1012 PAT_VALUE(2, PAT_UNCACHED) | /* 010 */
1013 PAT_VALUE(3, PAT_UNCACHEABLE) | /* 011 */
1014 PAT_VALUE(4, PAT_WRITE_BACK) | /* 100 */
1015 PAT_VALUE(5, PAT_WRITE_THROUGH) | /* 101 */
1016 PAT_VALUE(6, PAT_UNCACHED) | /* 110 */
1017 PAT_VALUE(7, PAT_UNCACHEABLE); /* 111 */
1018 pat_pte_index[PAT_WRITE_BACK] = 0;
1019 pat_pte_index[PAT_WRITE_THROUGH]= 0 | X86_PG_NC_PWT;
1020 pat_pte_index[PAT_UNCACHED] = X86_PG_NC_PCD;
1021 pat_pte_index[PAT_UNCACHEABLE] = X86_PG_NC_PCD | X86_PG_NC_PWT;
1022 pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1023 pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1025 if (cpu_feature & CPUID_PAT) {
1027 * If we support the PAT then set-up entries for
1028 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1031 pat_msr = (pat_msr & ~PAT_MASK(4)) |
1032 PAT_VALUE(4, PAT_WRITE_PROTECTED);
1033 pat_msr = (pat_msr & ~PAT_MASK(5)) |
1034 PAT_VALUE(5, PAT_WRITE_COMBINING);
1035 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
1036 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1039 * Then enable the PAT
1044 load_cr4(cr4 & ~CR4_PGE);
1046 /* Disable caches (CD = 1, NW = 0). */
1048 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1050 /* Flushes caches and TLBs. */
1054 /* Update PAT and index table. */
1055 wrmsr(MSR_PAT, pat_msr);
1057 /* Flush caches and TLBs again. */
1061 /* Restore caches and PGE. */
1069 * Set 4mb pdir for mp startup
1074 if (cpu_feature & CPUID_PSE) {
1075 load_cr4(rcr4() | CR4_PSE);
1076 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
1083 * Initialize the pmap module.
1084 * Called by vm_init, to initialize any structures that the pmap
1085 * system needs to map virtual memory.
1086 * pmap_init has been enhanced to support in a fairly consistant
1087 * way, discontiguous physical memory.
1096 * Allocate memory for random pmap data structures. Includes the
1100 for (i = 0; i < vm_page_array_size; i++) {
1103 m = &vm_page_array[i];
1104 TAILQ_INIT(&m->md.pv_list);
1108 * init the pv free list
1110 initial_pvs = vm_page_array_size;
1111 if (initial_pvs < MINPV)
1112 initial_pvs = MINPV;
1113 pvzone = &pvzone_store;
1114 pvinit = (void *)kmem_alloc(&kernel_map,
1115 initial_pvs * sizeof (struct pv_entry),
1117 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1118 pvinit, initial_pvs);
1121 * Now it is safe to enable pv_table recording.
1123 pmap_initialized = TRUE;
1127 * Initialize the address space (zone) for the pv_entries. Set a
1128 * high water mark so that the system can recover from excessive
1129 * numbers of pv entries.
1134 int shpgperproc = PMAP_SHPGPERPROC;
1137 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1138 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1139 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1140 pv_entry_high_water = 9 * (pv_entry_max / 10);
1143 * Subtract out pages already installed in the zone (hack)
1145 entry_max = pv_entry_max - vm_page_array_size;
1149 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT);
1152 * Enable dynamic deletion of empty higher-level page table pages
1153 * by default only if system memory is < 8GB (use 7GB for slop).
1154 * This can save a little memory, but imposes significant
1155 * performance overhead for things like bulk builds, and for programs
1156 * which do a lot of memory mapping and memory unmapping.
1158 if (pmap_dynamic_delete < 0) {
1159 if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1160 pmap_dynamic_delete = 1;
1162 pmap_dynamic_delete = 0;
1167 * Typically used to initialize a fictitious page by vm/device_pager.c
1170 pmap_page_init(struct vm_page *m)
1173 TAILQ_INIT(&m->md.pv_list);
1176 /***************************************************
1177 * Low level helper routines.....
1178 ***************************************************/
1181 * this routine defines the region(s) of memory that should
1182 * not be tested for the modified bit.
1186 pmap_track_modified(vm_pindex_t pindex)
1188 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1189 if ((va < clean_sva) || (va >= clean_eva))
1196 * Extract the physical page address associated with the map/VA pair.
1197 * The page must be wired for this to work reliably.
1199 * XXX for the moment we're using pv_find() instead of pv_get(), as
1200 * callers might be expecting non-blocking operation.
1203 pmap_extract(pmap_t pmap, vm_offset_t va)
1210 if (va >= VM_MAX_USER_ADDRESS) {
1212 * Kernel page directories might be direct-mapped and
1213 * there is typically no PV tracking of pte's
1217 pt = pmap_pt(pmap, va);
1218 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1219 if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1220 rtval = *pt & PG_PS_FRAME;
1221 rtval |= va & PDRMASK;
1223 ptep = pmap_pt_to_pte(*pt, va);
1224 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1225 rtval = *ptep & PG_FRAME;
1226 rtval |= va & PAGE_MASK;
1232 * User pages currently do not direct-map the page directory
1233 * and some pages might not used managed PVs. But all PT's
1236 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1238 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1239 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1240 rtval = *ptep & PG_FRAME;
1241 rtval |= va & PAGE_MASK;
1250 * Similar to extract but checks protections, SMP-friendly short-cut for
1251 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1252 * fall-through to the real fault code.
1254 * The returned page, if not NULL, is held (and not busied).
1257 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1259 if (pmap && va < VM_MAX_USER_ADDRESS) {
1267 req = pmap->pmap_bits[PG_V_IDX] |
1268 pmap->pmap_bits[PG_U_IDX];
1269 if (prot & VM_PROT_WRITE)
1270 req |= pmap->pmap_bits[PG_RW_IDX];
1272 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1275 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1276 if ((*ptep & req) != req) {
1280 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1281 if (pte_pv && error == 0) {
1284 if (prot & VM_PROT_WRITE)
1287 } else if (pte_pv) {
1301 * Extract the physical page address associated kernel virtual address.
1304 pmap_kextract(vm_offset_t va)
1306 pd_entry_t pt; /* pt entry in pd */
1309 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1310 pa = DMAP_TO_PHYS(va);
1313 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1314 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1317 * Beware of a concurrent promotion that changes the
1318 * PDE at this point! For example, vtopte() must not
1319 * be used to access the PTE because it would use the
1320 * new PDE. It is, however, safe to use the old PDE
1321 * because the page table page is preserved by the
1324 pa = *pmap_pt_to_pte(pt, va);
1325 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1331 /***************************************************
1332 * Low level mapping routines.....
1333 ***************************************************/
1336 * Routine: pmap_kenter
1338 * Add a wired page to the KVA
1339 * NOTE! note that in order for the mapping to take effect -- you
1340 * should do an invltlb after doing the pmap_kenter().
1343 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1349 kernel_pmap.pmap_bits[PG_RW_IDX] |
1350 kernel_pmap.pmap_bits[PG_V_IDX];
1354 pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1358 pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1365 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1366 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1367 * (caller can conditionalize calling smp_invltlb()).
1370 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1377 kernel_pmap.pmap_bits[PG_RW_IDX] |
1378 kernel_pmap.pmap_bits[PG_V_IDX];
1388 cpu_invlpg((void *)va);
1394 * Enter addresses into the kernel pmap but don't bother
1395 * doing any tlb invalidations. Caller will do a rollup
1396 * invalidation via pmap_rollup_inval().
1399 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1406 kernel_pmap.pmap_bits[PG_RW_IDX] |
1407 kernel_pmap.pmap_bits[PG_V_IDX];
1417 cpu_invlpg((void *)va);
1423 * remove a page from the kernel pagetables
1426 pmap_kremove(vm_offset_t va)
1431 pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1435 pmap_kremove_quick(vm_offset_t va)
1440 (void)pte_load_clear(ptep);
1441 cpu_invlpg((void *)va);
1445 * Remove addresses from the kernel pmap but don't bother
1446 * doing any tlb invalidations. Caller will do a rollup
1447 * invalidation via pmap_rollup_inval().
1450 pmap_kremove_noinval(vm_offset_t va)
1455 (void)pte_load_clear(ptep);
1459 * XXX these need to be recoded. They are not used in any critical path.
1462 pmap_kmodify_rw(vm_offset_t va)
1464 atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1465 cpu_invlpg((void *)va);
1470 pmap_kmodify_nc(vm_offset_t va)
1472 atomic_set_long(vtopte(va), PG_N);
1473 cpu_invlpg((void *)va);
1478 * Used to map a range of physical addresses into kernel virtual
1479 * address space during the low level boot, typically to map the
1480 * dump bitmap, message buffer, and vm_page_array.
1482 * These mappings are typically made at some pointer after the end of the
1485 * We could return PHYS_TO_DMAP(start) here and not allocate any
1486 * via (*virtp), but then kmem from userland and kernel dumps won't
1487 * have access to the related pointers.
1490 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1493 vm_offset_t va_start;
1495 /*return PHYS_TO_DMAP(start);*/
1500 while (start < end) {
1501 pmap_kenter_quick(va, start);
1509 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1512 * Remove the specified set of pages from the data and instruction caches.
1514 * In contrast to pmap_invalidate_cache_range(), this function does not
1515 * rely on the CPU's self-snoop feature, because it is intended for use
1516 * when moving pages into a different cache domain.
1519 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1521 vm_offset_t daddr, eva;
1524 if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1525 (cpu_feature & CPUID_CLFSH) == 0)
1529 for (i = 0; i < count; i++) {
1530 daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1531 eva = daddr + PAGE_SIZE;
1532 for (; daddr < eva; daddr += cpu_clflush_line_size)
1540 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1542 KASSERT((sva & PAGE_MASK) == 0,
1543 ("pmap_invalidate_cache_range: sva not page-aligned"));
1544 KASSERT((eva & PAGE_MASK) == 0,
1545 ("pmap_invalidate_cache_range: eva not page-aligned"));
1547 if (cpu_feature & CPUID_SS) {
1548 ; /* If "Self Snoop" is supported, do nothing. */
1550 /* Globally invalidate caches */
1551 cpu_wbinvd_on_all_cpus();
1556 * Invalidate the specified range of virtual memory on all cpus associated
1560 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1562 pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
1566 * Add a list of wired pages to the kva. This routine is used for temporary
1567 * kernel mappings such as those found in buffer cache buffer. Page
1568 * modifications and accesses are not tracked or recorded.
1570 * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1571 * semantics as previous mappings may have been zerod without any
1574 * The page *must* be wired.
1577 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
1582 end_va = beg_va + count * PAGE_SIZE;
1584 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1588 *pte = VM_PAGE_TO_PHYS(*m) |
1589 kernel_pmap.pmap_bits[PG_RW_IDX] |
1590 kernel_pmap.pmap_bits[PG_V_IDX] |
1591 kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1595 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1599 * This routine jerks page mappings from the kernel -- it is meant only
1600 * for temporary mappings such as those found in buffer cache buffers.
1601 * No recording modified or access status occurs.
1603 * MPSAFE, INTERRUPT SAFE (cluster callback)
1606 pmap_qremove(vm_offset_t beg_va, int count)
1611 end_va = beg_va + count * PAGE_SIZE;
1613 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1617 (void)pte_load_clear(pte);
1618 cpu_invlpg((void *)va);
1620 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
1624 * This routine removes temporary kernel mappings, only invalidating them
1625 * on the current cpu. It should only be used under carefully controlled
1629 pmap_qremove_quick(vm_offset_t beg_va, int count)
1634 end_va = beg_va + count * PAGE_SIZE;
1636 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1640 (void)pte_load_clear(pte);
1641 cpu_invlpg((void *)va);
1646 * This routine removes temporary kernel mappings *without* invalidating
1647 * the TLB. It can only be used on permanent kva reservations such as those
1648 * found in buffer cache buffers, under carefully controlled circumstances.
1650 * NOTE: Repopulating these KVAs requires unconditional invalidation.
1651 * (pmap_qenter() does unconditional invalidation).
1654 pmap_qremove_noinval(vm_offset_t beg_va, int count)
1659 end_va = beg_va + count * PAGE_SIZE;
1661 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
1665 (void)pte_load_clear(pte);
1670 * Create a new thread and optionally associate it with a (new) process.
1671 * NOTE! the new thread's cpu may not equal the current cpu.
1674 pmap_init_thread(thread_t td)
1676 /* enforce pcb placement & alignment */
1677 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1678 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1679 td->td_savefpu = &td->td_pcb->pcb_save;
1680 td->td_sp = (char *)td->td_pcb; /* no -16 */
1684 * This routine directly affects the fork perf for a process.
1687 pmap_init_proc(struct proc *p)
1692 pmap_pinit_defaults(struct pmap *pmap)
1694 bcopy(pmap_bits_default, pmap->pmap_bits,
1695 sizeof(pmap_bits_default));
1696 bcopy(protection_codes, pmap->protection_codes,
1697 sizeof(protection_codes));
1698 bcopy(pat_pte_index, pmap->pmap_cache_bits,
1699 sizeof(pat_pte_index));
1700 pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1701 pmap->copyinstr = std_copyinstr;
1702 pmap->copyin = std_copyin;
1703 pmap->copyout = std_copyout;
1704 pmap->fubyte = std_fubyte;
1705 pmap->subyte = std_subyte;
1706 pmap->fuword = std_fuword;
1707 pmap->suword = std_suword;
1708 pmap->suword32 = std_suword32;
1711 * Initialize pmap0/vmspace0.
1713 * On architectures where the kernel pmap is not integrated into the user
1714 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1715 * kernel_pmap should be used to directly access the kernel_pmap.
1718 pmap_pinit0(struct pmap *pmap)
1720 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1722 CPUMASK_ASSZERO(pmap->pm_active);
1723 pmap->pm_pvhint = NULL;
1724 RB_INIT(&pmap->pm_pvroot);
1725 spin_init(&pmap->pm_spin, "pmapinit0");
1726 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1727 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1728 pmap_pinit_defaults(pmap);
1732 * Initialize a preallocated and zeroed pmap structure,
1733 * such as one in a vmspace structure.
1736 pmap_pinit_simple(struct pmap *pmap)
1739 * Misc initialization
1742 CPUMASK_ASSZERO(pmap->pm_active);
1743 pmap->pm_pvhint = NULL;
1744 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1746 pmap_pinit_defaults(pmap);
1749 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1752 if (pmap->pm_pmlpv == NULL) {
1753 RB_INIT(&pmap->pm_pvroot);
1754 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1755 spin_init(&pmap->pm_spin, "pmapinitsimple");
1756 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1761 pmap_pinit(struct pmap *pmap)
1766 if (pmap->pm_pmlpv) {
1767 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1772 pmap_pinit_simple(pmap);
1773 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1776 * No need to allocate page table space yet but we do need a valid
1777 * page directory table.
1779 if (pmap->pm_pml4 == NULL) {
1781 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
1787 * Allocate the page directory page, which wires it even though
1788 * it isn't being entered into some higher level page table (it
1789 * being the highest level). If one is already cached we don't
1790 * have to do anything.
1792 if ((pv = pmap->pm_pmlpv) == NULL) {
1793 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1794 pmap->pm_pmlpv = pv;
1795 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1796 VM_PAGE_TO_PHYS(pv->pv_m));
1800 * Install DMAP and KMAP.
1802 for (j = 0; j < NDMPML4E; ++j) {
1803 pmap->pm_pml4[DMPML4I + j] =
1804 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1805 pmap->pmap_bits[PG_RW_IDX] |
1806 pmap->pmap_bits[PG_V_IDX] |
1807 pmap->pmap_bits[PG_U_IDX];
1809 pmap->pm_pml4[KPML4I] = KPDPphys |
1810 pmap->pmap_bits[PG_RW_IDX] |
1811 pmap->pmap_bits[PG_V_IDX] |
1812 pmap->pmap_bits[PG_U_IDX];
1815 * install self-referential address mapping entry
1817 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1818 pmap->pmap_bits[PG_V_IDX] |
1819 pmap->pmap_bits[PG_RW_IDX] |
1820 pmap->pmap_bits[PG_A_IDX] |
1821 pmap->pmap_bits[PG_M_IDX];
1823 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1824 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1826 KKASSERT(pmap->pm_pml4[255] == 0);
1827 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1828 KKASSERT(pv->pv_entry.rbe_left == NULL);
1829 KKASSERT(pv->pv_entry.rbe_right == NULL);
1833 * Clean up a pmap structure so it can be physically freed. This routine
1834 * is called by the vmspace dtor function. A great deal of pmap data is
1835 * left passively mapped to improve vmspace management so we have a bit
1836 * of cleanup work to do here.
1839 pmap_puninit(pmap_t pmap)
1844 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1845 if ((pv = pmap->pm_pmlpv) != NULL) {
1846 if (pv_hold_try(pv) == 0)
1848 KKASSERT(pv == pmap->pm_pmlpv);
1849 p = pmap_remove_pv_page(pv);
1850 pv_free(pv, NULL, 1);
1851 pv = NULL; /* safety */
1852 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1853 vm_page_busy_wait(p, FALSE, "pgpun");
1854 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1855 vm_page_unwire(p, 0);
1856 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1859 * XXX eventually clean out PML4 static entries and
1860 * use vm_page_free_zero()
1863 pmap->pm_pmlpv = NULL;
1865 if (pmap->pm_pml4) {
1866 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1867 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1868 pmap->pm_pml4 = NULL;
1870 KKASSERT(pmap->pm_stats.resident_count == 0);
1871 KKASSERT(pmap->pm_stats.wired_count == 0);
1875 * This function is now unused (used to add the pmap to the pmap_list)
1878 pmap_pinit2(struct pmap *pmap)
1883 * This routine is called when various levels in the page table need to
1884 * be populated. This routine cannot fail.
1886 * This function returns two locked pv_entry's, one representing the
1887 * requested pv and one representing the requested pv's parent pv. If
1888 * an intermediate page table does not exist it will be created, mapped,
1889 * wired, and the parent page table will be given an additional hold
1890 * count representing the presence of the child pv_entry.
1894 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1899 vm_pindex_t pt_pindex;
1905 * If the pv already exists and we aren't being asked for the
1906 * parent page table page we can just return it. A locked+held pv
1907 * is returned. The pv will also have a second hold related to the
1908 * pmap association that we don't have to worry about.
1911 pv = pv_alloc(pmap, ptepindex, &isnew);
1912 if (isnew == 0 && pvpp == NULL)
1916 * Special case terminal PVs. These are not page table pages so
1917 * no vm_page is allocated (the caller supplied the vm_page). If
1918 * pvpp is non-NULL we are being asked to also removed the pt_pv
1921 * Note that pt_pv's are only returned for user VAs. We assert that
1922 * a pt_pv is not being requested for kernel VAs. The kernel
1923 * pre-wires all higher-level page tables so don't overload managed
1924 * higher-level page tables on top of it!
1926 if (ptepindex < pmap_pt_pindex(0)) {
1927 if (ptepindex >= NUPTE_USER) {
1928 /* kernel manages this manually for KVM */
1929 KKASSERT(pvpp == NULL);
1931 KKASSERT(pvpp != NULL);
1932 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1933 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1935 vm_page_wire_quick(pvp->pv_m);
1948 * The kernel never uses managed PT/PD/PDP pages.
1950 KKASSERT(pmap != &kernel_pmap);
1953 * Non-terminal PVs allocate a VM page to represent the page table,
1954 * so we have to resolve pvp and calculate ptepindex for the pvp
1955 * and then for the page table entry index in the pvp for
1958 if (ptepindex < pmap_pd_pindex(0)) {
1960 * pv is PT, pvp is PD
1962 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1963 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1964 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1971 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1972 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1974 } else if (ptepindex < pmap_pdp_pindex(0)) {
1976 * pv is PD, pvp is PDP
1978 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1981 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1982 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1984 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1985 KKASSERT(pvpp == NULL);
1988 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1996 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1997 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1998 } else if (ptepindex < pmap_pml4_pindex()) {
2000 * pv is PDP, pvp is the root pml4 table
2002 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2009 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2010 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2013 * pv represents the top-level PML4, there is no parent.
2021 * (isnew) is TRUE, pv is not terminal.
2023 * (1) Add a wire count to the parent page table (pvp).
2024 * (2) Allocate a VM page for the page table.
2025 * (3) Enter the VM page into the parent page table.
2027 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2030 vm_page_wire_quick(pvp->pv_m);
2033 m = vm_page_alloc(NULL, pv->pv_pindex,
2034 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2035 VM_ALLOC_INTERRUPT);
2040 vm_page_spin_lock(m);
2041 pmap_page_stats_adding(m);
2042 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2044 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
2045 vm_page_spin_unlock(m);
2046 vm_page_unmanage(m); /* m must be spinunlocked */
2048 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2049 m->valid = VM_PAGE_BITS_ALL;
2050 vm_page_wire(m); /* wire for mapping in parent */
2053 * Wire the page into pvp. Bump the resident_count for the pmap.
2054 * There is no pvp for the top level, address the pm_pml4[] array
2057 * If the caller wants the parent we return it, otherwise
2058 * we just put it away.
2060 * No interlock is needed for pte 0 -> non-zero.
2062 * In the situation where *ptep is valid we might have an unmanaged
2063 * page table page shared from another page table which we need to
2064 * unshare before installing our private page table page.
2067 ptep = pv_pte_lookup(pvp, ptepindex);
2068 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2072 panic("pmap_allocpte: unexpected pte %p/%d",
2073 pvp, (int)ptepindex);
2075 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
2076 if (vm_page_unwire_quick(
2077 PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
2078 panic("pmap_allocpte: shared pgtable "
2079 "pg bad wirecount");
2081 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2083 *ptep = VM_PAGE_TO_PHYS(m) |
2084 (pmap->pmap_bits[PG_U_IDX] |
2085 pmap->pmap_bits[PG_RW_IDX] |
2086 pmap->pmap_bits[PG_V_IDX] |
2087 pmap->pmap_bits[PG_A_IDX] |
2088 pmap->pmap_bits[PG_M_IDX]);
2100 * This version of pmap_allocpte() checks for possible segment optimizations
2101 * that would allow page-table sharing. It can be called for terminal
2102 * page or page table page ptepindex's.
2104 * The function is called with page table page ptepindex's for fictitious
2105 * and unmanaged terminal pages. That is, we don't want to allocate a
2106 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2109 * This function can return a pv and *pvpp associated with the passed in pmap
2110 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2111 * an unmanaged page table page will be entered into the pass in pmap.
2115 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
2116 vm_map_entry_t entry, vm_offset_t va)
2122 pv_entry_t pte_pv; /* in original or shared pmap */
2123 pv_entry_t pt_pv; /* in original or shared pmap */
2124 pv_entry_t proc_pd_pv; /* in original pmap */
2125 pv_entry_t proc_pt_pv; /* in original pmap */
2126 pv_entry_t xpv; /* PT in shared pmap */
2127 pd_entry_t *pt; /* PT entry in PD of original pmap */
2128 pd_entry_t opte; /* contents of *pt */
2129 pd_entry_t npte; /* contents of *pt */
2134 * Basic tests, require a non-NULL vm_map_entry, require proper
2135 * alignment and type for the vm_map_entry, require that the
2136 * underlying object already be allocated.
2138 * We allow almost any type of object to use this optimization.
2139 * The object itself does NOT have to be sized to a multiple of the
2140 * segment size, but the memory mapping does.
2142 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2143 * won't work as expected.
2145 if (entry == NULL ||
2146 pmap_mmu_optimize == 0 || /* not enabled */
2147 (pmap->pm_flags & PMAP_HVM) || /* special pmap */
2148 ptepindex >= pmap_pd_pindex(0) || /* not terminal or pt */
2149 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
2150 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
2151 entry->object.vm_object == NULL || /* needs VM object */
2152 entry->object.vm_object->type == OBJT_DEVICE || /* ick */
2153 entry->object.vm_object->type == OBJT_MGTDEVICE || /* ick */
2154 (entry->offset & SEG_MASK) || /* must be aligned */
2155 (entry->start & SEG_MASK)) {
2156 return(pmap_allocpte(pmap, ptepindex, pvpp));
2160 * Make sure the full segment can be represented.
2162 b = va & ~(vm_offset_t)SEG_MASK;
2163 if (b < entry->start || b + SEG_SIZE > entry->end)
2164 return(pmap_allocpte(pmap, ptepindex, pvpp));
2167 * If the full segment can be represented dive the VM object's
2168 * shared pmap, allocating as required.
2170 object = entry->object.vm_object;
2172 if (entry->protection & VM_PROT_WRITE)
2173 obpmapp = &object->md.pmap_rw;
2175 obpmapp = &object->md.pmap_ro;
2178 if (pmap_enter_debug > 0) {
2180 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2182 va, entry->protection, object,
2184 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2185 entry, entry->start, entry->end);
2190 * We allocate what appears to be a normal pmap but because portions
2191 * of this pmap are shared with other unrelated pmaps we have to
2192 * set pm_active to point to all cpus.
2194 * XXX Currently using pmap_spin to interlock the update, can't use
2195 * vm_object_hold/drop because the token might already be held
2196 * shared OR exclusive and we don't know.
2198 while ((obpmap = *obpmapp) == NULL) {
2199 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2200 pmap_pinit_simple(obpmap);
2201 pmap_pinit2(obpmap);
2202 spin_lock(&pmap_spin);
2203 if (*obpmapp != NULL) {
2207 spin_unlock(&pmap_spin);
2208 pmap_release(obpmap);
2209 pmap_puninit(obpmap);
2210 kfree(obpmap, M_OBJPMAP);
2211 obpmap = *obpmapp; /* safety */
2213 obpmap->pm_active = smp_active_mask;
2214 obpmap->pm_flags |= PMAP_SEGSHARED;
2216 spin_unlock(&pmap_spin);
2221 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2222 * pte/pt using the shared pmap from the object but also adjust
2223 * the process pmap's page table page as a side effect.
2227 * Resolve the terminal PTE and PT in the shared pmap. This is what
2228 * we will return. This is true if ptepindex represents a terminal
2229 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2233 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2234 if (ptepindex >= pmap_pt_pindex(0))
2240 * Resolve the PD in the process pmap so we can properly share the
2241 * page table page. Lock order is bottom-up (leaf first)!
2243 * NOTE: proc_pt_pv can be NULL.
2245 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2246 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2248 if (pmap_enter_debug > 0) {
2250 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2252 (proc_pt_pv ? proc_pt_pv->pv_m->wire_count : -1),
2259 * xpv is the page table page pv from the shared object
2260 * (for convenience), from above.
2262 * Calculate the pte value for the PT to load into the process PD.
2263 * If we have to change it we must properly dispose of the previous
2266 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2267 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2268 (pmap->pmap_bits[PG_U_IDX] |
2269 pmap->pmap_bits[PG_RW_IDX] |
2270 pmap->pmap_bits[PG_V_IDX] |
2271 pmap->pmap_bits[PG_A_IDX] |
2272 pmap->pmap_bits[PG_M_IDX]);
2275 * Dispose of previous page table page if it was local to the
2276 * process pmap. If the old pt is not empty we cannot dispose of it
2277 * until we clean it out. This case should not arise very often so
2278 * it is not optimized.
2281 pmap_inval_bulk_t bulk;
2283 if (proc_pt_pv->pv_m->wire_count != 1) {
2289 va & ~(vm_offset_t)SEG_MASK,
2290 (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2295 * The release call will indirectly clean out *pt
2297 pmap_inval_bulk_init(&bulk, proc_pt_pv->pv_pmap);
2298 pmap_release_pv(proc_pt_pv, proc_pd_pv, &bulk);
2299 pmap_inval_bulk_flush(&bulk);
2302 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2306 * Handle remaining cases.
2310 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2311 vm_page_wire_quick(proc_pd_pv->pv_m); /* proc pd for sh pt */
2312 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2313 } else if (*pt != npte) {
2314 opte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, pt, npte);
2317 opte = pte_load_clear(pt);
2318 KKASSERT(opte && opte != npte);
2322 vm_page_wire_quick(xpv->pv_m); /* shared pt -> proc */
2325 * Clean up opte, bump the wire_count for the process
2326 * PD page representing the new entry if it was
2329 * If the entry was not previously empty and we have
2330 * a PT in the proc pmap then opte must match that
2331 * pt. The proc pt must be retired (this is done
2332 * later on in this procedure).
2334 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2337 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2338 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2339 if (vm_page_unwire_quick(m)) {
2340 panic("pmap_allocpte_seg: "
2341 "bad wire count %p",
2347 * The existing process page table was replaced and must be destroyed
2361 * Release any resources held by the given physical map.
2363 * Called when a pmap initialized by pmap_pinit is being released. Should
2364 * only be called if the map contains no valid mappings.
2366 * Caller must hold pmap->pm_token
2368 struct pmap_release_info {
2374 static int pmap_release_callback(pv_entry_t pv, void *data);
2377 pmap_release(struct pmap *pmap)
2379 struct pmap_release_info info;
2381 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2382 ("pmap still active! %016jx",
2383 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2386 * There is no longer a pmap_list, if there were we would remove the
2387 * pmap from it here.
2391 * Pull pv's off the RB tree in order from low to high and release
2399 spin_lock(&pmap->pm_spin);
2400 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2401 pmap_release_callback, &info);
2402 spin_unlock(&pmap->pm_spin);
2406 } while (info.retry);
2410 * One resident page (the pml4 page) should remain.
2411 * No wired pages should remain.
2413 KKASSERT(pmap->pm_stats.resident_count ==
2414 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2416 KKASSERT(pmap->pm_stats.wired_count == 0);
2420 * Called from low to high. We must cache the proper parent pv so we
2421 * can adjust its wired count.
2424 pmap_release_callback(pv_entry_t pv, void *data)
2426 struct pmap_release_info *info = data;
2427 pmap_t pmap = info->pmap;
2431 if (info->pvp == pv) {
2432 spin_unlock(&pmap->pm_spin);
2434 } else if (pv_hold_try(pv)) {
2435 spin_unlock(&pmap->pm_spin);
2437 spin_unlock(&pmap->pm_spin);
2440 if (pv->pv_pmap != pmap) {
2442 spin_lock(&pmap->pm_spin);
2447 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2451 pindex = pv->pv_pindex >> NPTEPGSHIFT;
2452 pindex += NUPTE_TOTAL;
2453 } else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2457 pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2458 pindex += NUPTE_TOTAL + NUPT_TOTAL;
2459 } else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2463 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2465 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2466 } else if (pv->pv_pindex < pmap_pml4_pindex()) {
2468 * parent is PML4 (there's only one)
2471 pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL -
2472 NUPD_TOTAL) >> NPML4EPGSHIFT;
2473 pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL;
2475 pindex = pmap_pml4_pindex();
2487 if (info->pvp && info->pvp->pv_pindex != pindex) {
2491 if (info->pvp == NULL)
2492 info->pvp = pv_get(pmap, pindex);
2499 r = pmap_release_pv(pv, info->pvp, NULL);
2500 spin_lock(&pmap->pm_spin);
2505 * Called with held (i.e. also locked) pv. This function will dispose of
2506 * the lock along with the pv.
2508 * If the caller already holds the locked parent page table for pv it
2509 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2510 * pass NULL for pvp.
2513 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2518 * The pmap is currently not spinlocked, pv is held+locked.
2519 * Remove the pv's page from its parent's page table. The
2520 * parent's page table page's wire_count will be decremented.
2522 * This will clean out the pte at any level of the page table.
2523 * If smp != 0 all cpus are affected.
2525 * Do not tear-down recursively, its faster to just let the
2526 * release run its course.
2528 pmap_remove_pv_pte(pv, pvp, bulk, 0);
2531 * Terminal pvs are unhooked from their vm_pages. Because
2532 * terminal pages aren't page table pages they aren't wired
2533 * by us, so we have to be sure not to unwire them either.
2535 if (pv->pv_pindex < pmap_pt_pindex(0)) {
2536 pmap_remove_pv_page(pv);
2541 * We leave the top-level page table page cached, wired, and
2542 * mapped in the pmap until the dtor function (pmap_puninit())
2545 * Since we are leaving the top-level pv intact we need
2546 * to break out of what would otherwise be an infinite loop.
2548 if (pv->pv_pindex == pmap_pml4_pindex()) {
2554 * For page table pages (other than the top-level page),
2555 * remove and free the vm_page. The representitive mapping
2556 * removed above by pmap_remove_pv_pte() did not undo the
2557 * last wire_count so we have to do that as well.
2559 p = pmap_remove_pv_page(pv);
2560 vm_page_busy_wait(p, FALSE, "pmaprl");
2561 if (p->wire_count != 1) {
2562 kprintf("p->wire_count was %016lx %d\n",
2563 pv->pv_pindex, p->wire_count);
2565 KKASSERT(p->wire_count == 1);
2566 KKASSERT(p->flags & PG_UNMANAGED);
2568 vm_page_unwire(p, 0);
2569 KKASSERT(p->wire_count == 0);
2573 pv_free(pv, pvp, 1);
2579 * This function will remove the pte associated with a pv from its parent.
2580 * Terminal pv's are supported. All cpus specified by (bulk) are properly
2583 * The wire count will be dropped on the parent page table. The wire
2584 * count on the page being removed (pv->pv_m) from the parent page table
2585 * is NOT touched. Note that terminal pages will not have any additional
2586 * wire counts while page table pages will have at least one representing
2587 * the mapping, plus others representing sub-mappings.
2589 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2590 * pages and user page table and terminal pages.
2592 * The pv must be locked. The pvp, if supplied, must be locked. All
2593 * supplied pv's will remain locked on return.
2595 * XXX must lock parent pv's if they exist to remove pte XXX
2599 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2602 vm_pindex_t ptepindex = pv->pv_pindex;
2603 pmap_t pmap = pv->pv_pmap;
2609 if (ptepindex == pmap_pml4_pindex()) {
2611 * We are the top level pml4 table, there is no parent.
2613 p = pmap->pm_pmlpv->pv_m;
2614 } else if (ptepindex >= pmap_pdp_pindex(0)) {
2616 * Remove a PDP page from the pml4e. This can only occur
2617 * with user page tables. We do not have to lock the
2618 * pml4 PV so just ignore pvp.
2620 vm_pindex_t pml4_pindex;
2621 vm_pindex_t pdp_index;
2624 pdp_index = ptepindex - pmap_pdp_pindex(0);
2626 pml4_pindex = pmap_pml4_pindex();
2627 pvp = pv_get(pv->pv_pmap, pml4_pindex);
2631 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2632 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2633 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2634 pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2635 } else if (ptepindex >= pmap_pd_pindex(0)) {
2637 * Remove a PD page from the pdp
2639 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2640 * of a simple pmap because it stops at
2643 vm_pindex_t pdp_pindex;
2644 vm_pindex_t pd_index;
2647 pd_index = ptepindex - pmap_pd_pindex(0);
2650 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2651 (pd_index >> NPML4EPGSHIFT);
2652 pvp = pv_get(pv->pv_pmap, pdp_pindex);
2656 pd = pv_pte_lookup(pvp, pd_index &
2657 ((1ul << NPDPEPGSHIFT) - 1));
2658 KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2659 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2660 pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
2662 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2663 p = pv->pv_m; /* degenerate test later */
2665 } else if (ptepindex >= pmap_pt_pindex(0)) {
2667 * Remove a PT page from the pd
2669 vm_pindex_t pd_pindex;
2670 vm_pindex_t pt_index;
2673 pt_index = ptepindex - pmap_pt_pindex(0);
2676 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2677 (pt_index >> NPDPEPGSHIFT);
2678 pvp = pv_get(pv->pv_pmap, pd_pindex);
2682 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2683 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2684 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2685 pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
2688 * Remove a managed PTE from the PT page. Userland pmaps
2689 * manage PT/PD/PDP page tables pages but the kernel_pmap
2692 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2693 * pv is a pte_pv so we can safely lock pt_pv.
2695 * NOTE: FICTITIOUS pages may have multiple physical mappings
2696 * so PHYS_TO_VM_PAGE() will not necessarily work for
2699 vm_pindex_t pt_pindex;
2704 pt_pindex = ptepindex >> NPTEPGSHIFT;
2705 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2707 if (ptepindex >= NUPTE_USER) {
2708 ptep = vtopte(ptepindex << PAGE_SHIFT);
2709 KKASSERT(pvp == NULL);
2710 /* pvp remains NULL */
2713 pt_pindex = NUPTE_TOTAL +
2714 (ptepindex >> NPDPEPGSHIFT);
2715 pvp = pv_get(pv->pv_pmap, pt_pindex);
2719 ptep = pv_pte_lookup(pvp, ptepindex &
2720 ((1ul << NPDPEPGSHIFT) - 1));
2722 pte = pmap_inval_bulk(bulk, va, ptep, 0);
2723 if (bulk == NULL) /* XXX */
2724 cpu_invlpg((void *)va); /* XXX */
2727 * Now update the vm_page_t
2729 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] |
2730 pmap->pmap_bits[PG_V_IDX])) !=
2731 (pmap->pmap_bits[PG_MANAGED_IDX] |
2732 pmap->pmap_bits[PG_V_IDX])) {
2733 kprintf("remove_pte badpte %016lx %016lx %d\n",
2735 pv->pv_pindex < pmap_pt_pindex(0));
2738 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2739 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2740 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2743 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2746 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2747 if (pmap_track_modified(ptepindex))
2750 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2751 vm_page_flag_set(p, PG_REFERENCED);
2753 if (pte & pmap->pmap_bits[PG_W_IDX])
2754 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2755 if (pte & pmap->pmap_bits[PG_G_IDX])
2756 cpu_invlpg((void *)va);
2758 KKASSERT(pv->pv_m == p); /* XXX remove me later */
2761 * If requested, scrap the underlying pv->pv_m and the underlying
2762 * pv. If this is a page-table-page we must also free the page.
2764 * pvp must be returned locked.
2768 * page table page (PT, PD, PDP, PML4), caller was responsible
2769 * for testing wired_count.
2773 KKASSERT(pv->pv_m->wire_count == 1);
2774 p = pmap_remove_pv_page(pv);
2775 pv_free(pv, pvp, 1);
2778 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
2779 vm_page_busy_wait(p, FALSE, "pgpun");
2780 vm_page_unwire(p, 0);
2781 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2783 } else if (destroy == 2) {
2785 * Normal page, remove from pmap and leave the underlying
2788 pmap_remove_pv_page(pv);
2789 pv_free(pv, pvp, 1);
2790 pv = NULL; /* safety */
2794 * If we acquired pvp ourselves then we are responsible for
2795 * recursively deleting it.
2797 if (pvp && gotpvp) {
2799 * Recursively destroy higher-level page tables.
2801 * This is optional. If we do not, they will still
2802 * be destroyed when the process exits.
2804 * NOTE: Do not destroy pv_entry's with extra hold refs,
2805 * a caller may have unlocked it and intends to
2806 * continue to use it.
2808 if (pmap_dynamic_delete &&
2810 pvp->pv_m->wire_count == 1 &&
2811 (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
2812 pvp->pv_pindex != pmap_pml4_pindex()) {
2813 if (pmap_dynamic_delete == 2)
2814 kprintf("A %jd %08x\n", pvp->pv_pindex, pvp->pv_hold);
2815 if (pmap != &kernel_pmap) {
2816 pmap_remove_pv_pte(pvp, NULL, bulk, 1);
2817 pvp = NULL; /* safety */
2819 kprintf("Attempt to remove kernel_pmap pindex "
2820 "%jd\n", pvp->pv_pindex);
2830 * Remove the vm_page association to a pv. The pv must be locked.
2834 pmap_remove_pv_page(pv_entry_t pv)
2840 vm_page_spin_lock(m);
2842 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2843 pmap_page_stats_deleting(m);
2846 atomic_add_int(&m->object->agg_pv_list_count, -1);
2848 if (TAILQ_EMPTY(&m->md.pv_list))
2849 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2850 vm_page_spin_unlock(m);
2856 * Grow the number of kernel page table entries, if needed.
2858 * This routine is always called to validate any address space
2859 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2860 * space below KERNBASE.
2862 * kernel_map must be locked exclusively by the caller.
2865 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2868 vm_offset_t ptppaddr;
2870 pd_entry_t *pt, newpt;
2872 int update_kernel_vm_end;
2875 * bootstrap kernel_vm_end on first real VM use
2877 if (kernel_vm_end == 0) {
2878 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2880 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2881 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2882 ~(PAGE_SIZE * NPTEPG - 1);
2884 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2885 kernel_vm_end = kernel_map.max_offset;
2892 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2893 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2894 * do not want to force-fill 128G worth of page tables.
2896 if (kstart < KERNBASE) {
2897 if (kstart > kernel_vm_end)
2898 kstart = kernel_vm_end;
2899 KKASSERT(kend <= KERNBASE);
2900 update_kernel_vm_end = 1;
2902 update_kernel_vm_end = 0;
2905 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2906 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2908 if (kend - 1 >= kernel_map.max_offset)
2909 kend = kernel_map.max_offset;
2911 while (kstart < kend) {
2912 pt = pmap_pt(&kernel_pmap, kstart);
2914 /* We need a new PD entry */
2915 nkpg = vm_page_alloc(NULL, nkpt,
2918 VM_ALLOC_INTERRUPT);
2920 panic("pmap_growkernel: no memory to grow "
2923 paddr = VM_PAGE_TO_PHYS(nkpg);
2924 pmap_zero_page(paddr);
2925 newpd = (pdp_entry_t)
2927 kernel_pmap.pmap_bits[PG_V_IDX] |
2928 kernel_pmap.pmap_bits[PG_RW_IDX] |
2929 kernel_pmap.pmap_bits[PG_A_IDX] |
2930 kernel_pmap.pmap_bits[PG_M_IDX]);
2931 *pmap_pd(&kernel_pmap, kstart) = newpd;
2933 continue; /* try again */
2935 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2936 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2937 ~(PAGE_SIZE * NPTEPG - 1);
2938 if (kstart - 1 >= kernel_map.max_offset) {
2939 kstart = kernel_map.max_offset;
2948 * This index is bogus, but out of the way
2950 nkpg = vm_page_alloc(NULL, nkpt,
2953 VM_ALLOC_INTERRUPT);
2955 panic("pmap_growkernel: no memory to grow kernel");
2958 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2959 pmap_zero_page(ptppaddr);
2960 newpt = (pd_entry_t) (ptppaddr |
2961 kernel_pmap.pmap_bits[PG_V_IDX] |
2962 kernel_pmap.pmap_bits[PG_RW_IDX] |
2963 kernel_pmap.pmap_bits[PG_A_IDX] |
2964 kernel_pmap.pmap_bits[PG_M_IDX]);
2965 *pmap_pt(&kernel_pmap, kstart) = newpt;
2968 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2969 ~(PAGE_SIZE * NPTEPG - 1);
2971 if (kstart - 1 >= kernel_map.max_offset) {
2972 kstart = kernel_map.max_offset;
2978 * Only update kernel_vm_end for areas below KERNBASE.
2980 if (update_kernel_vm_end && kernel_vm_end < kstart)
2981 kernel_vm_end = kstart;
2985 * Add a reference to the specified pmap.
2988 pmap_reference(pmap_t pmap)
2991 lwkt_gettoken(&pmap->pm_token);
2993 lwkt_reltoken(&pmap->pm_token);
2997 /***************************************************
2998 * page management routines.
2999 ***************************************************/
3002 * Hold a pv without locking it
3005 pv_hold(pv_entry_t pv)
3007 atomic_add_int(&pv->pv_hold, 1);
3011 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
3012 * was successfully locked, FALSE if it wasn't. The caller must dispose of
3015 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3016 * pv list via its page) must be held by the caller.
3019 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3024 * Critical path shortcut expects pv to already have one ref
3025 * (for the pv->pv_pmap).
3027 if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
3030 pv->pv_line = lineno;
3036 count = pv->pv_hold;
3038 if ((count & PV_HOLD_LOCKED) == 0) {
3039 if (atomic_cmpset_int(&pv->pv_hold, count,
3040 (count + 1) | PV_HOLD_LOCKED)) {
3043 pv->pv_line = lineno;
3048 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
3056 * Drop a previously held pv_entry which could not be locked, allowing its
3059 * Must not be called with a spinlock held as we might zfree() the pv if it
3060 * is no longer associated with a pmap and this was the last hold count.
3063 pv_drop(pv_entry_t pv)
3068 count = pv->pv_hold;
3070 KKASSERT((count & PV_HOLD_MASK) > 0);
3071 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3072 (PV_HOLD_LOCKED | 1));
3073 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3074 if ((count & PV_HOLD_MASK) == 1) {
3076 if (pmap_enter_debug > 0) {
3078 kprintf("pv_drop: free pv %p\n", pv);
3081 KKASSERT(count == 1);
3082 KKASSERT(pv->pv_pmap == NULL);
3092 * Find or allocate the requested PV entry, returning a locked, held pv.
3094 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3095 * for the caller and one representing the pmap and vm_page association.
3097 * If (*isnew) is zero, the returned pv will have only one hold count.
3099 * Since both associations can only be adjusted while the pv is locked,
3100 * together they represent just one additional hold.
3104 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3107 pv_entry_t pnew = NULL;
3109 spin_lock(&pmap->pm_spin);
3111 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3112 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3117 spin_unlock(&pmap->pm_spin);
3118 pnew = zalloc(pvzone);
3119 spin_lock(&pmap->pm_spin);
3122 pnew->pv_pmap = pmap;
3123 pnew->pv_pindex = pindex;
3124 pnew->pv_hold = PV_HOLD_LOCKED | 2;
3126 pnew->pv_func = func;
3127 pnew->pv_line = lineno;
3129 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3130 ++pmap->pm_generation;
3131 atomic_add_long(&pmap->pm_stats.resident_count, 1);
3132 spin_unlock(&pmap->pm_spin);
3137 spin_unlock(&pmap->pm_spin);
3138 zfree(pvzone, pnew);
3140 spin_lock(&pmap->pm_spin);
3143 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3144 spin_unlock(&pmap->pm_spin);
3146 spin_unlock(&pmap->pm_spin);
3147 _pv_lock(pv PMAP_DEBUG_COPY);
3149 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3154 spin_lock(&pmap->pm_spin);
3159 * Find the requested PV entry, returning a locked+held pv or NULL
3163 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
3167 spin_lock(&pmap->pm_spin);
3172 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
3173 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
3177 spin_unlock(&pmap->pm_spin);
3180 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3181 spin_unlock(&pmap->pm_spin);
3183 spin_unlock(&pmap->pm_spin);
3184 _pv_lock(pv PMAP_DEBUG_COPY);
3186 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
3187 pv_cache(pv, pindex);
3191 spin_lock(&pmap->pm_spin);
3196 * Lookup, hold, and attempt to lock (pmap,pindex).
3198 * If the entry does not exist NULL is returned and *errorp is set to 0
3200 * If the entry exists and could be successfully locked it is returned and
3201 * errorp is set to 0.
3203 * If the entry exists but could NOT be successfully locked it is returned
3204 * held and *errorp is set to 1.
3208 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
3212 spin_lock_shared(&pmap->pm_spin);
3213 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3214 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3216 spin_unlock_shared(&pmap->pm_spin);
3220 if (pv_hold_try(pv)) {
3221 pv_cache(pv, pindex);
3222 spin_unlock_shared(&pmap->pm_spin);
3224 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3225 return(pv); /* lock succeeded */
3227 spin_unlock_shared(&pmap->pm_spin);
3229 return (pv); /* lock failed */
3233 * Find the requested PV entry, returning a held pv or NULL
3237 pv_find(pmap_t pmap, vm_pindex_t pindex)
3241 spin_lock_shared(&pmap->pm_spin);
3243 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
3244 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
3246 spin_unlock_shared(&pmap->pm_spin);
3250 pv_cache(pv, pindex);
3251 spin_unlock_shared(&pmap->pm_spin);
3256 * Lock a held pv, keeping the hold count
3260 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3265 count = pv->pv_hold;
3267 if ((count & PV_HOLD_LOCKED) == 0) {
3268 if (atomic_cmpset_int(&pv->pv_hold, count,
3269 count | PV_HOLD_LOCKED)) {
3272 pv->pv_line = lineno;
3278 tsleep_interlock(pv, 0);
3279 if (atomic_cmpset_int(&pv->pv_hold, count,
3280 count | PV_HOLD_WAITING)) {
3282 kprintf("pv waiting on %s:%d\n",
3283 pv->pv_func, pv->pv_line);
3285 tsleep(pv, PINTERLOCKED, "pvwait", hz);
3292 * Unlock a held and locked pv, keeping the hold count.
3296 pv_unlock(pv_entry_t pv)
3301 count = pv->pv_hold;
3303 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3304 (PV_HOLD_LOCKED | 1));
3305 if (atomic_cmpset_int(&pv->pv_hold, count,
3307 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3308 if (count & PV_HOLD_WAITING)
3316 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3317 * and the hold count drops to zero we will free it.
3319 * Caller should not hold any spin locks. We are protected from hold races
3320 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3321 * lock held. A pv cannot be located otherwise.
3325 pv_put(pv_entry_t pv)
3328 if (pmap_enter_debug > 0) {
3330 kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3335 * Fast - shortcut most common condition
3337 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3348 * Remove the pmap association from a pv, require that pv_m already be removed,
3349 * then unlock and drop the pv. Any pte operations must have already been
3350 * completed. This call may result in a last-drop which will physically free
3353 * Removing the pmap association entails an additional drop.
3355 * pv must be exclusively locked on call and will be disposed of on return.
3359 pv_free(pv_entry_t pv, pv_entry_t pvp, int putaway)
3363 KKASSERT(pv->pv_m == NULL);
3364 KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3365 if ((pmap = pv->pv_pmap) != NULL) {
3366 spin_lock(&pmap->pm_spin);
3367 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3368 ++pmap->pm_generation;
3369 if (pmap->pm_pvhint == pv)
3370 pmap->pm_pvhint = NULL;
3371 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3374 spin_unlock(&pmap->pm_spin);
3377 * Try to shortcut three atomic ops, otherwise fall through
3378 * and do it normally. Drop two refs and the lock all in
3382 atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3384 if (pmap_enter_debug > 0) {
3386 kprintf("pv_free: free pv %p\n", pv);
3391 vm_page_unwire_quick(pvp->pv_m);
3394 pv_drop(pv); /* ref for pv_pmap */
3399 vm_page_unwire_quick(pvp->pv_m);
3403 * This routine is very drastic, but can save the system
3411 static int warningdone=0;
3413 if (pmap_pagedaemon_waken == 0)
3415 pmap_pagedaemon_waken = 0;
3416 if (warningdone < 5) {
3417 kprintf("pmap_collect: collecting pv entries -- "
3418 "suggest increasing PMAP_SHPGPERPROC\n");
3422 for (i = 0; i < vm_page_array_size; i++) {
3423 m = &vm_page_array[i];
3424 if (m->wire_count || m->hold_count)
3426 if (vm_page_busy_try(m, TRUE) == 0) {
3427 if (m->wire_count == 0 && m->hold_count == 0) {
3436 * Scan the pmap for active page table entries and issue a callback.
3437 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3438 * its parent page table.
3440 * pte_pv will be NULL if the page or page table is unmanaged.
3441 * pt_pv will point to the page table page containing the pte for the page.
3443 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3444 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3445 * process pmap's PD and page to the callback function. This can be
3446 * confusing because the pt_pv is really a pd_pv, and the target page
3447 * table page is simply aliased by the pmap and not owned by it.
3449 * It is assumed that the start and end are properly rounded to the page size.
3451 * It is assumed that PD pages and above are managed and thus in the RB tree,
3452 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3454 struct pmap_scan_info {
3458 vm_pindex_t sva_pd_pindex;
3459 vm_pindex_t eva_pd_pindex;
3460 void (*func)(pmap_t, struct pmap_scan_info *,
3461 pv_entry_t, pv_entry_t, int, vm_offset_t,
3462 pt_entry_t *, void *);
3464 pmap_inval_bulk_t bulk_core;
3465 pmap_inval_bulk_t *bulk;
3470 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3471 static int pmap_scan_callback(pv_entry_t pv, void *data);
3474 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3476 struct pmap *pmap = info->pmap;
3477 pv_entry_t pd_pv; /* A page directory PV */
3478 pv_entry_t pt_pv; /* A page table PV */
3479 pv_entry_t pte_pv; /* A page table entry PV */
3482 struct pv_entry dummy_pv;
3489 info->bulk = &info->bulk_core;
3490 pmap_inval_bulk_init(&info->bulk_core, pmap);
3496 * Hold the token for stability; if the pmap is empty we have nothing
3499 lwkt_gettoken(&pmap->pm_token);
3501 if (pmap->pm_stats.resident_count == 0) {
3502 lwkt_reltoken(&pmap->pm_token);
3511 * Special handling for scanning one page, which is a very common
3512 * operation (it is?).
3514 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3516 if (info->sva + PAGE_SIZE == info->eva) {
3517 generation = pmap->pm_generation;
3518 if (info->sva >= VM_MAX_USER_ADDRESS) {
3520 * Kernel mappings do not track wire counts on
3521 * page table pages and only maintain pd_pv and
3522 * pte_pv levels so pmap_scan() works.
3525 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3526 ptep = vtopte(info->sva);
3529 * User pages which are unmanaged will not have a
3530 * pte_pv. User page table pages which are unmanaged
3531 * (shared from elsewhere) will also not have a pt_pv.
3532 * The func() callback will pass both pte_pv and pt_pv
3533 * as NULL in that case.
3535 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3536 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3537 if (pt_pv == NULL) {
3538 KKASSERT(pte_pv == NULL);
3539 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3541 ptep = pv_pte_lookup(pd_pv,
3542 pmap_pt_index(info->sva));
3544 info->func(pmap, info,
3553 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3557 * NOTE: *ptep can't be ripped out from under us if we hold
3558 * pte_pv locked, but bits can change. However, there is
3559 * a race where another thread may be inserting pte_pv
3560 * and setting *ptep just after our pte_pv lookup fails.
3562 * In this situation we can end up with a NULL pte_pv
3563 * but find that we have a managed *ptep. We explicitly
3564 * check for this race.
3570 * Unlike the pv_find() case below we actually
3571 * acquired a locked pv in this case so any
3572 * race should have been resolved. It is expected
3575 KKASSERT(pte_pv == NULL);
3576 } else if (pte_pv) {
3577 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3578 pmap->pmap_bits[PG_V_IDX])) ==
3579 (pmap->pmap_bits[PG_MANAGED_IDX] |
3580 pmap->pmap_bits[PG_V_IDX]),
3581 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3583 *ptep, oldpte, info->sva, pte_pv,
3584 generation, pmap->pm_generation));
3585 info->func(pmap, info, pte_pv, pt_pv, 0,
3586 info->sva, ptep, info->arg);
3589 * Check for insertion race
3591 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3593 pte_pv = pv_find(pmap,
3594 pmap_pte_pindex(info->sva));
3598 kprintf("pmap_scan: RACE1 "
3608 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3609 pmap->pmap_bits[PG_V_IDX])) ==
3610 pmap->pmap_bits[PG_V_IDX],
3611 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3613 *ptep, oldpte, info->sva,
3614 generation, pmap->pm_generation));
3615 info->func(pmap, info, NULL, pt_pv, 0,
3616 info->sva, ptep, info->arg);
3621 pmap_inval_bulk_flush(info->bulk);
3622 lwkt_reltoken(&pmap->pm_token);
3627 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3630 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3631 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3633 if (info->sva >= VM_MAX_USER_ADDRESS) {
3635 * The kernel does not currently maintain any pv_entry's for
3636 * higher-level page tables.
3638 bzero(&dummy_pv, sizeof(dummy_pv));
3639 dummy_pv.pv_pindex = info->sva_pd_pindex;
3640 spin_lock(&pmap->pm_spin);
3641 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3642 pmap_scan_callback(&dummy_pv, info);
3643 ++dummy_pv.pv_pindex;
3645 spin_unlock(&pmap->pm_spin);
3648 * User page tables maintain local PML4, PDP, and PD
3649 * pv_entry's at the very least. PT pv's might be
3650 * unmanaged and thus not exist. PTE pv's might be
3651 * unmanaged and thus not exist.
3653 spin_lock(&pmap->pm_spin);
3654 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3655 pmap_scan_cmp, pmap_scan_callback, info);
3656 spin_unlock(&pmap->pm_spin);
3658 pmap_inval_bulk_flush(info->bulk);
3659 lwkt_reltoken(&pmap->pm_token);
3663 * WARNING! pmap->pm_spin held
3666 pmap_scan_cmp(pv_entry_t pv, void *data)
3668 struct pmap_scan_info *info = data;
3669 if (pv->pv_pindex < info->sva_pd_pindex)
3671 if (pv->pv_pindex >= info->eva_pd_pindex)
3677 * WARNING! pmap->pm_spin held
3680 pmap_scan_callback(pv_entry_t pv, void *data)
3682 struct pmap_scan_info *info = data;
3683 struct pmap *pmap = info->pmap;
3684 pv_entry_t pd_pv; /* A page directory PV */
3685 pv_entry_t pt_pv; /* A page table PV */
3686 pv_entry_t pte_pv; /* A page table entry PV */
3691 vm_offset_t va_next;
3692 vm_pindex_t pd_pindex;
3703 * Pull the PD pindex from the pv before releasing the spinlock.
3705 * WARNING: pv is faked for kernel pmap scans.
3707 pd_pindex = pv->pv_pindex;
3708 spin_unlock(&pmap->pm_spin);
3709 pv = NULL; /* invalid after spinlock unlocked */
3712 * Calculate the page range within the PD. SIMPLE pmaps are
3713 * direct-mapped for the entire 2^64 address space. Normal pmaps
3714 * reflect the user and kernel address space which requires
3715 * cannonicalization w/regards to converting pd_pindex's back
3718 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3719 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3720 (sva & PML4_SIGNMASK)) {
3721 sva |= PML4_SIGNMASK;
3723 eva = sva + NBPDP; /* can overflow */
3724 if (sva < info->sva)
3726 if (eva < info->sva || eva > info->eva)
3730 * NOTE: kernel mappings do not track page table pages, only
3733 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3734 * However, for the scan to be efficient we try to
3735 * cache items top-down.
3740 for (; sva < eva; sva = va_next) {
3743 if (sva >= VM_MAX_USER_ADDRESS) {
3752 * PD cache (degenerate case if we skip). It is possible
3753 * for the PD to not exist due to races. This is ok.
3755 if (pd_pv == NULL) {
3756 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3757 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3759 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3761 if (pd_pv == NULL) {
3762 va_next = (sva + NBPDP) & ~PDPMASK;
3771 if (pt_pv == NULL) {
3772 vm_page_wire_quick(pd_pv->pv_m);
3774 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3776 vm_page_unwire_quick(pd_pv->pv_m);
3777 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3778 vm_page_wire_quick(pd_pv->pv_m);
3781 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3783 vm_page_unwire_quick(pd_pv->pv_m);
3787 * If pt_pv is NULL we either have an shared page table
3788 * page and must issue a callback specific to that case,
3789 * or there is no page table page.
3791 * Either way we can skip the page table page.
3793 if (pt_pv == NULL) {
3795 * Possible unmanaged (shared from another pmap)
3798 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3799 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3800 info->func(pmap, info, NULL, pd_pv, 1,
3801 sva, ptep, info->arg);
3805 * Done, move to next page table page.
3807 va_next = (sva + NBPDR) & ~PDRMASK;
3814 * From this point in the loop testing pt_pv for non-NULL
3815 * means we are in UVM, else if it is NULL we are in KVM.
3817 * Limit our scan to either the end of the va represented
3818 * by the current page table page, or to the end of the
3819 * range being removed.
3822 va_next = (sva + NBPDR) & ~PDRMASK;
3829 * Scan the page table for pages. Some pages may not be
3830 * managed (might not have a pv_entry).
3832 * There is no page table management for kernel pages so
3833 * pt_pv will be NULL in that case, but otherwise pt_pv
3834 * is non-NULL, locked, and referenced.
3838 * At this point a non-NULL pt_pv means a UVA, and a NULL
3839 * pt_pv means a KVA.
3842 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3846 while (sva < va_next) {
3848 * Yield every 64 pages, stop if requested.
3850 if ((++info->count & 63) == 0)
3856 * Check if pt_pv has been lost (probably due to
3857 * a remove of the underlying pages).
3859 if (pt_pv && pt_pv->pv_pmap == NULL)
3863 * Acquire the related pte_pv, if any. If *ptep == 0
3864 * the related pte_pv should not exist, but if *ptep
3865 * is not zero the pte_pv may or may not exist (e.g.
3866 * will not exist for an unmanaged page).
3868 * However a multitude of races are possible here.
3870 * In addition, the (pt_pv, pte_pv) lock order is
3871 * backwards, so we have to be careful in aquiring
3872 * a properly locked pte_pv.
3874 generation = pmap->pm_generation;
3876 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3880 vm_page_wire_quick(pd_pv->pv_m);
3883 vm_page_wire_quick(pt_pv->pv_m);
3884 pv_unlock(pt_pv);/* must be non-NULL */
3885 pv_lock(pte_pv); /* safe to block now */
3889 vm_page_unwire_quick(pt_pv->pv_m);
3892 * pt_pv reloaded, need new ptep
3894 KKASSERT(pt_pv != NULL);
3895 ptep = pv_pte_lookup(pt_pv,
3896 pmap_pte_index(sva));
3899 vm_page_unwire_quick(pd_pv->pv_m);
3904 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3908 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3913 kprintf("Unexpected non-NULL pte_pv "
3915 "*ptep = %016lx/%016lx\n",
3916 pte_pv, pt_pv, *ptep, oldpte);
3917 panic("Unexpected non-NULL pte_pv");
3925 * Ready for the callback. The locked pte_pv (if any)
3926 * is consumed by the callback. pte_pv will exist if
3927 * the page is managed, and will not exist if it
3931 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3932 (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3933 ("badC *ptep %016lx/%016lx sva %016lx "
3934 "pte_pv %p pm_generation %d/%d",
3935 *ptep, oldpte, sva, pte_pv,
3936 generation, pmap->pm_generation));
3938 * We must unlock pd_pv across the callback
3939 * to avoid deadlocks on any recursive
3940 * disposal. Re-check that it still exists
3945 info->func(pmap, info, pte_pv, pt_pv, 0,
3946 sva, ptep, info->arg);
3949 if (pd_pv->pv_pmap == NULL) {
3956 * Check for insertion race. Since there is no
3957 * pte_pv to guard us it is possible for us
3958 * to race another thread doing an insertion.
3959 * Our lookup misses the pte_pv but our *ptep
3960 * check sees the inserted pte.
3962 * XXX panic case seems to occur within a
3963 * vm_fork() of /bin/sh, which frankly
3964 * shouldn't happen since no other threads
3965 * should be inserting to our pmap in that
3966 * situation. Removing, possibly. Inserting,
3969 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3971 pte_pv = pv_find(pmap,
3972 pmap_pte_pindex(sva));
3975 kprintf("pmap_scan: RACE2 "
3985 * We must unlock pd_pv across the callback
3986 * to avoid deadlocks on any recursive
3987 * disposal. Re-check that it still exists
3990 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3991 pmap->pmap_bits[PG_V_IDX],
3992 ("badD *ptep %016lx/%016lx sva %016lx "
3993 "pte_pv NULL pm_generation %d/%d",
3995 generation, pmap->pm_generation));
3998 info->func(pmap, info, NULL, pt_pv, 0,
3999 sva, ptep, info->arg);
4002 if (pd_pv->pv_pmap == NULL) {
4021 if ((++info->count & 7) == 0)
4025 * Relock before returning.
4027 spin_lock(&pmap->pm_spin);
4032 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4034 struct pmap_scan_info info;
4039 info.func = pmap_remove_callback;
4041 pmap_scan(&info, 1);
4045 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4047 struct pmap_scan_info info;
4052 info.func = pmap_remove_callback;
4054 pmap_scan(&info, 0);
4058 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4059 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4060 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4066 * This will also drop pt_pv's wire_count. Note that
4067 * terminal pages are not wired based on mmu presence.
4069 * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4071 pmap_remove_pv_pte(pte_pv, pt_pv, info->bulk, 2);
4072 pte_pv = NULL; /* safety */
4075 * Recursively destroy higher-level page tables.
4077 * This is optional. If we do not, they will still
4078 * be destroyed when the process exits.
4080 * NOTE: Do not destroy pv_entry's with extra hold refs,
4081 * a caller may have unlocked it and intends to
4082 * continue to use it.
4084 if (pmap_dynamic_delete &&
4087 pt_pv->pv_m->wire_count == 1 &&
4088 (pt_pv->pv_hold & PV_HOLD_MASK) == 2 &&
4089 pt_pv->pv_pindex != pmap_pml4_pindex()) {
4090 if (pmap_dynamic_delete == 2)
4091 kprintf("B %jd %08x\n", pt_pv->pv_pindex, pt_pv->pv_hold);
4093 pmap_remove_pv_pte(pt_pv, NULL, info->bulk, 1);
4096 } else if (sharept == 0) {
4098 * Unmanaged page table (pt, pd, or pdp. Not pte).
4100 * pt_pv's wire_count is still bumped by unmanaged pages
4101 * so we must decrement it manually.
4103 * We have to unwire the target page table page.
4105 * It is unclear how we can invalidate a segment so we
4106 * invalidate -1 which invlidates the tlb.
4108 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4109 if (pte & pmap->pmap_bits[PG_W_IDX])
4110 atomic_add_long(&pmap->pm_stats.wired_count, -1);
4111 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4112 if (vm_page_unwire_quick(pt_pv->pv_m))
4113 panic("pmap_remove: insufficient wirecount");
4116 * Unmanaged page table (pt, pd, or pdp. Not pte) for
4117 * a shared page table.
4119 * pt_pv is actually the pd_pv for our pmap (not the shared
4122 * We have to unwire the target page table page and we
4123 * have to unwire our page directory page.
4125 * It is unclear how we can invalidate a segment so we
4126 * invalidate -1 which invlidates the tlb.
4128 pte = pmap_inval_bulk(info->bulk, (vm_offset_t)-1, ptep, 0);
4129 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4130 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
4131 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4132 panic("pmap_remove: shared pgtable1 bad wirecount");
4133 if (vm_page_unwire_quick(pt_pv->pv_m))
4134 panic("pmap_remove: shared pgtable2 bad wirecount");
4139 * Removes this physical page from all physical maps in which it resides.
4140 * Reflects back modify bits to the pager.
4142 * This routine may not be called from an interrupt.
4146 pmap_remove_all(vm_page_t m)
4149 pmap_inval_bulk_t bulk;
4151 if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
4154 vm_page_spin_lock(m);
4155 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
4156 KKASSERT(pv->pv_m == m);
4157 if (pv_hold_try(pv)) {
4158 vm_page_spin_unlock(m);
4160 vm_page_spin_unlock(m);
4163 if (pv->pv_m != m) {
4165 vm_page_spin_lock(m);
4170 * Holding no spinlocks, pv is locked.
4172 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4173 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4174 pv = NULL; /* safety */
4175 pmap_inval_bulk_flush(&bulk);
4177 pmap_remove_pv_page(pv);
4180 vm_page_spin_lock(m);
4182 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
4183 vm_page_spin_unlock(m);
4187 * Removes the page from a particular pmap
4190 pmap_remove_specific(pmap_t pmap, vm_page_t m)
4193 pmap_inval_bulk_t bulk;
4195 if (!pmap_initialized)
4199 vm_page_spin_lock(m);
4200 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4201 if (pv->pv_pmap != pmap)
4203 KKASSERT(pv->pv_m == m);
4204 if (pv_hold_try(pv)) {
4205 vm_page_spin_unlock(m);
4207 vm_page_spin_unlock(m);
4210 if (pv->pv_m != m) {
4216 * Holding no spinlocks, pv is locked.
4218 pmap_inval_bulk_init(&bulk, pv->pv_pmap);
4219 pmap_remove_pv_pte(pv, NULL, &bulk, 2);
4220 pv = NULL; /* safety */
4221 pmap_inval_bulk_flush(&bulk);
4223 pmap_remove_pv_page(pv);
4228 vm_page_spin_unlock(m);
4232 * Set the physical protection on the specified range of this map
4233 * as requested. This function is typically only used for debug watchpoints
4236 * This function may not be called from an interrupt if the map is
4237 * not the kernel_pmap.
4239 * NOTE! For shared page table pages we just unmap the page.
4242 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4244 struct pmap_scan_info info;
4245 /* JG review for NX */
4249 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
4250 pmap_remove(pmap, sva, eva);
4253 if (prot & VM_PROT_WRITE)
4258 info.func = pmap_protect_callback;
4260 pmap_scan(&info, 1);
4265 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4266 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
4267 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4279 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
4280 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4281 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4282 KKASSERT(m == pte_pv->pv_m);
4283 vm_page_flag_set(m, PG_REFERENCED);
4285 cbits &= ~pmap->pmap_bits[PG_A_IDX];
4287 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
4288 if (pmap_track_modified(pte_pv->pv_pindex)) {
4289 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
4291 m = PHYS_TO_VM_PAGE(pbits &
4296 cbits &= ~pmap->pmap_bits[PG_M_IDX];
4299 } else if (sharept) {
4301 * Unmanaged page table, pt_pv is actually the pd_pv
4302 * for our pmap (not the object's shared pmap).
4304 * When asked to protect something in a shared page table
4305 * page we just unmap the page table page. We have to
4306 * invalidate the tlb in this situation.
4308 * XXX Warning, shared page tables will not be used for
4309 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4310 * so PHYS_TO_VM_PAGE() should be safe here.
4312 pte = pmap_inval_smp(pmap, (vm_offset_t)-1, 1, ptep, 0);
4313 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
4314 panic("pmap_protect: pgtable1 pg bad wirecount");
4315 if (vm_page_unwire_quick(pt_pv->pv_m))
4316 panic("pmap_protect: pgtable2 pg bad wirecount");
4319 /* else unmanaged page, adjust bits, no wire changes */
4322 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4324 if (pmap_enter_debug > 0) {
4326 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4327 "pt_pv=%p cbits=%08lx\n",
4333 if (pbits != cbits) {
4334 if (!pmap_inval_smp_cmpset(pmap, (vm_offset_t)-1,
4335 ptep, pbits, cbits)) {
4345 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4346 * mapping at that address. Set protection and wiring as requested.
4348 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4349 * possible. If it is we enter the page into the appropriate shared pmap
4350 * hanging off the related VM object instead of the passed pmap, then we
4351 * share the page table page from the VM object's pmap into the current pmap.
4353 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4357 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4358 boolean_t wired, vm_map_entry_t entry)
4360 pv_entry_t pt_pv; /* page table */
4361 pv_entry_t pte_pv; /* page table entry */
4364 pt_entry_t origpte, newpte;
4369 va = trunc_page(va);
4370 #ifdef PMAP_DIAGNOSTIC
4372 panic("pmap_enter: toobig");
4373 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4374 panic("pmap_enter: invalid to pmap_enter page table "
4375 "pages (va: 0x%lx)", va);
4377 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4378 kprintf("Warning: pmap_enter called on UVA with "
4381 db_print_backtrace();
4384 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4385 kprintf("Warning: pmap_enter called on KVA without"
4388 db_print_backtrace();
4393 * Get locked PV entries for our new page table entry (pte_pv)
4394 * and for its parent page table (pt_pv). We need the parent
4395 * so we can resolve the location of the ptep.
4397 * Only hardware MMU actions can modify the ptep out from
4400 * if (m) is fictitious or unmanaged we do not create a managing
4401 * pte_pv for it. Any pre-existing page's management state must
4402 * match (avoiding code complexity).
4404 * If the pmap is still being initialized we assume existing
4407 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4409 if (pmap_initialized == FALSE) {
4414 } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
4416 if (va >= VM_MAX_USER_ADDRESS) {
4420 pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
4422 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4426 KASSERT(origpte == 0 ||
4427 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0,
4428 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4430 if (va >= VM_MAX_USER_ADDRESS) {
4432 * Kernel map, pv_entry-tracked.
4435 pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
4441 pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
4443 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4447 KASSERT(origpte == 0 ||
4448 (origpte & pmap->pmap_bits[PG_MANAGED_IDX]),
4449 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte, va));
4452 pa = VM_PAGE_TO_PHYS(m);
4453 opa = origpte & PG_FRAME;
4455 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4456 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4458 newpte |= pmap->pmap_bits[PG_W_IDX];
4459 if (va < VM_MAX_USER_ADDRESS)
4460 newpte |= pmap->pmap_bits[PG_U_IDX];
4462 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4463 // if (pmap == &kernel_pmap)
4464 // newpte |= pgeflag;
4465 newpte |= pmap->pmap_cache_bits[m->pat_mode];
4466 if (m->flags & PG_FICTITIOUS)
4467 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4470 * It is possible for multiple faults to occur in threaded
4471 * environments, the existing pte might be correct.
4473 if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4474 pmap->pmap_bits[PG_A_IDX])) == 0)
4478 * Ok, either the address changed or the protection or wiring
4481 * Clear the current entry, interlocking the removal. For managed
4482 * pte's this will also flush the modified state to the vm_page.
4483 * Atomic ops are mandatory in order to ensure that PG_M events are
4484 * not lost during any transition.
4486 * WARNING: The caller has busied the new page but not the original
4487 * vm_page which we are trying to replace. Because we hold
4488 * the pte_pv lock, but have not busied the page, PG bits
4489 * can be cleared out from under us.
4494 * pt_pv won't exist for a kernel page (managed or
4497 if (prot & VM_PROT_NOSYNC) {
4498 pmap_remove_pv_pte(pte_pv, pt_pv, NULL, 0);
4500 pmap_inval_bulk_t bulk;
4502 pmap_inval_bulk_init(&bulk, pmap);
4503 pmap_remove_pv_pte(pte_pv, pt_pv, &bulk, 0);
4504 pmap_inval_bulk_flush(&bulk);
4507 pmap_remove_pv_page(pte_pv);
4508 } else if (prot & VM_PROT_NOSYNC) {
4510 * Unmanaged page, NOSYNC (no mmu sync) requested.
4512 * Leave wire count on PT page intact.
4514 (void)pte_load_clear(ptep);
4515 cpu_invlpg((void *)va);
4516 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4519 * Unmanaged page, normal enter.
4521 * Leave wire count on PT page intact.
4523 pmap_inval_smp(pmap, va, 1, ptep, 0);
4524 atomic_add_long(&pmap->pm_stats.resident_count, -1);
4526 KKASSERT(*ptep == 0);
4530 if (pmap_enter_debug > 0) {
4532 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4533 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4535 origpte, newpte, ptep,
4536 pte_pv, pt_pv, opa, prot);
4542 * Enter on the PV list if part of our managed memory.
4543 * Wiring of the PT page is already handled.
4545 KKASSERT(pte_pv->pv_m == NULL);
4546 vm_page_spin_lock(m);
4548 pmap_page_stats_adding(m);
4549 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4550 vm_page_flag_set(m, PG_MAPPED);
4551 vm_page_spin_unlock(m);
4552 } else if (pt_pv && opa == 0) {
4554 * We have to adjust the wire count on the PT page ourselves
4555 * for unmanaged entries. If opa was non-zero we retained
4556 * the existing wire count from the removal.
4558 vm_page_wire_quick(pt_pv->pv_m);
4562 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4564 * User VMAs do not because those will be zero->non-zero, so no
4565 * stale entries to worry about at this point.
4567 * For KVM there appear to still be issues. Theoretically we
4568 * should be able to scrap the interlocks entirely but we
4571 if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL) {
4572 pmap_inval_smp(pmap, va, 1, ptep, newpte);
4574 *(volatile pt_entry_t *)ptep = newpte;
4576 cpu_invlpg((void *)va);
4581 atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4584 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4587 if (newpte & pmap->pmap_bits[PG_RW_IDX])
4588 vm_page_flag_set(m, PG_WRITEABLE);
4591 * Unmanaged pages need manual resident_count tracking.
4593 if (pte_pv == NULL && pt_pv)
4594 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4600 KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4601 (m->flags & PG_MAPPED));
4604 * Cleanup the pv entry, allowing other accessors.
4613 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4614 * This code also assumes that the pmap has no pre-existing entry for this
4617 * This code currently may only be used on user pmaps, not kernel_pmap.
4620 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4622 pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4626 * Make a temporary mapping for a physical address. This is only intended
4627 * to be used for panic dumps.
4629 * The caller is responsible for calling smp_invltlb().
4632 pmap_kenter_temporary(vm_paddr_t pa, long i)
4634 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4635 return ((void *)crashdumpmap);
4638 #define MAX_INIT_PT (96)
4641 * This routine preloads the ptes for a given object into the specified pmap.
4642 * This eliminates the blast of soft faults on process startup and
4643 * immediately after an mmap.
4645 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4648 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4649 vm_object_t object, vm_pindex_t pindex,
4650 vm_size_t size, int limit)
4652 struct rb_vm_page_scan_info info;
4657 * We can't preinit if read access isn't set or there is no pmap
4660 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4664 * We can't preinit if the pmap is not the current pmap
4666 lp = curthread->td_lwp;
4667 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4671 * Misc additional checks
4673 psize = x86_64_btop(size);
4675 if ((object->type != OBJT_VNODE) ||
4676 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4677 (object->resident_page_count > MAX_INIT_PT))) {
4681 if (pindex + psize > object->size) {
4682 if (object->size < pindex)
4684 psize = object->size - pindex;
4691 * If everything is segment-aligned do not pre-init here. Instead
4692 * allow the normal vm_fault path to pass a segment hint to
4693 * pmap_enter() which will then use an object-referenced shared
4696 if ((addr & SEG_MASK) == 0 &&
4697 (ctob(psize) & SEG_MASK) == 0 &&
4698 (ctob(pindex) & SEG_MASK) == 0) {
4703 * Use a red-black scan to traverse the requested range and load
4704 * any valid pages found into the pmap.
4706 * We cannot safely scan the object's memq without holding the
4709 info.start_pindex = pindex;
4710 info.end_pindex = pindex + psize - 1;
4716 vm_object_hold_shared(object);
4717 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4718 pmap_object_init_pt_callback, &info);
4719 vm_object_drop(object);
4724 pmap_object_init_pt_callback(vm_page_t p, void *data)
4726 struct rb_vm_page_scan_info *info = data;
4727 vm_pindex_t rel_index;
4730 * don't allow an madvise to blow away our really
4731 * free pages allocating pv entries.
4733 if ((info->limit & MAP_PREFAULT_MADVISE) &&
4734 vmstats.v_free_count < vmstats.v_free_reserved) {
4739 * Ignore list markers and ignore pages we cannot instantly
4740 * busy (while holding the object token).
4742 if (p->flags & PG_MARKER)
4744 if (vm_page_busy_try(p, TRUE))
4746 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4747 (p->flags & PG_FICTITIOUS) == 0) {
4748 if ((p->queue - p->pc) == PQ_CACHE)
4749 vm_page_deactivate(p);
4750 rel_index = p->pindex - info->start_pindex;
4751 pmap_enter_quick(info->pmap,
4752 info->addr + x86_64_ptob(rel_index), p);
4760 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4763 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4766 * XXX This is safe only because page table pages are not freed.
4769 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4773 /*spin_lock(&pmap->pm_spin);*/
4774 if ((pte = pmap_pte(pmap, addr)) != NULL) {
4775 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4776 /*spin_unlock(&pmap->pm_spin);*/
4780 /*spin_unlock(&pmap->pm_spin);*/
4785 * Change the wiring attribute for a pmap/va pair. The mapping must already
4786 * exist in the pmap. The mapping may or may not be managed.
4788 * Wiring is not a hardware characteristic so there is no need to invalidate
4789 * TLB. However, in an SMP environment we must use a locked bus cycle to
4790 * update the pte (if we are not using the pmap_inval_*() API that is)...
4791 * it's ok to do this for simple wiring changes.
4794 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4795 vm_map_entry_t entry)
4803 lwkt_gettoken(&pmap->pm_token);
4804 if (pmap == &kernel_pmap) {
4806 * The kernel may have managed pages, but not managed
4809 ptep = pmap_pte_quick(pmap, va);
4811 if (wired && !pmap_pte_w(pmap, ptep))
4812 atomic_add_long(&pmap->pm_stats.wired_count, 1);
4813 else if (!wired && pmap_pte_w(pmap, ptep))
4814 atomic_add_long(&pmap->pm_stats.wired_count, -1);
4817 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4819 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4822 * Userland, the pmap of the possibly shared segment might
4825 pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL,
4827 ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4829 if (wired && !pmap_pte_w(pmap, ptep))
4830 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4831 else if (!wired && pmap_pte_w(pmap, ptep))
4832 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4835 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4837 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4840 lwkt_reltoken(&pmap->pm_token);
4846 * Copy the range specified by src_addr/len from the source map to
4847 * the range dst_addr/len in the destination map.
4849 * This routine is only advisory and need not do anything.
4852 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4853 vm_size_t len, vm_offset_t src_addr)
4860 * Zero the specified physical page.
4862 * This function may be called from an interrupt and no locking is
4866 pmap_zero_page(vm_paddr_t phys)
4868 vm_offset_t va = PHYS_TO_DMAP(phys);
4870 pagezero((void *)va);
4876 * Zero part of a physical page by mapping it into memory and clearing
4877 * its contents with bzero.
4879 * off and size may not cover an area beyond a single hardware page.
4882 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4884 vm_offset_t virt = PHYS_TO_DMAP(phys);
4886 bzero((char *)virt + off, size);
4892 * Copy the physical page from the source PA to the target PA.
4893 * This function may be called from an interrupt. No locking
4897 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4899 vm_offset_t src_virt, dst_virt;
4901 src_virt = PHYS_TO_DMAP(src);
4902 dst_virt = PHYS_TO_DMAP(dst);
4903 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4907 * pmap_copy_page_frag:
4909 * Copy the physical page from the source PA to the target PA.
4910 * This function may be called from an interrupt. No locking
4914 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4916 vm_offset_t src_virt, dst_virt;
4918 src_virt = PHYS_TO_DMAP(src);
4919 dst_virt = PHYS_TO_DMAP(dst);
4921 bcopy((char *)src_virt + (src & PAGE_MASK),
4922 (char *)dst_virt + (dst & PAGE_MASK),
4927 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4928 * this page. This count may be changed upwards or downwards in the future;
4929 * it is only necessary that true be returned for a small subset of pmaps
4930 * for proper page aging.
4933 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4938 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4941 vm_page_spin_lock(m);
4942 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4943 if (pv->pv_pmap == pmap) {
4944 vm_page_spin_unlock(m);
4951 vm_page_spin_unlock(m);
4956 * Remove all pages from specified address space this aids process exit
4957 * speeds. Also, this code may be special cased for the current process
4961 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4963 pmap_remove_noinval(pmap, sva, eva);
4968 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4969 * routines are inline, and a lot of things compile-time evaluate.
4973 pmap_testbit(vm_page_t m, int bit)
4979 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4982 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4984 vm_page_spin_lock(m);
4985 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4986 vm_page_spin_unlock(m);
4990 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4992 #if defined(PMAP_DIAGNOSTIC)
4993 if (pv->pv_pmap == NULL) {
4994 kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
5002 * If the bit being tested is the modified bit, then
5003 * mark clean_map and ptes as never
5006 * WARNING! Because we do not lock the pv, *pte can be in a
5007 * state of flux. Despite this the value of *pte
5008 * will still be related to the vm_page in some way
5009 * because the pv cannot be destroyed as long as we
5010 * hold the vm_page spin lock.
5012 if (bit == PG_A_IDX || bit == PG_M_IDX) {
5013 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5014 if (!pmap_track_modified(pv->pv_pindex))
5018 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5019 if (*pte & pmap->pmap_bits[bit]) {
5020 vm_page_spin_unlock(m);
5024 vm_page_spin_unlock(m);
5029 * This routine is used to modify bits in ptes. Only one bit should be
5030 * specified. PG_RW requires special handling.
5032 * Caller must NOT hold any spin locks
5036 pmap_clearbit(vm_page_t m, int bit_index)
5043 if (bit_index == PG_RW_IDX)
5044 vm_page_flag_clear(m, PG_WRITEABLE);
5045 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
5052 * Loop over all current mappings setting/clearing as appropos If
5053 * setting RO do we need to clear the VAC?
5055 * NOTE: When clearing PG_M we could also (not implemented) drop
5056 * through to the PG_RW code and clear PG_RW too, forcing
5057 * a fault on write to redetect PG_M for virtual kernels, but
5058 * it isn't necessary since virtual kernels invalidate the
5059 * pte when they clear the VPTE_M bit in their virtual page
5062 * NOTE: Does not re-dirty the page when clearing only PG_M.
5064 * NOTE: Because we do not lock the pv, *pte can be in a state of
5065 * flux. Despite this the value of *pte is still somewhat
5066 * related while we hold the vm_page spin lock.
5068 * *pte can be zero due to this race. Since we are clearing
5069 * bits we basically do no harm when this race ccurs.
5071 if (bit_index != PG_RW_IDX) {
5072 vm_page_spin_lock(m);
5073 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5074 #if defined(PMAP_DIAGNOSTIC)
5075 if (pv->pv_pmap == NULL) {
5076 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5082 pte = pmap_pte_quick(pv->pv_pmap,
5083 pv->pv_pindex << PAGE_SHIFT);
5085 if (pbits & pmap->pmap_bits[bit_index])
5086 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
5088 vm_page_spin_unlock(m);
5093 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
5097 vm_page_spin_lock(m);
5098 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5100 * don't write protect pager mappings
5102 if (!pmap_track_modified(pv->pv_pindex))
5105 #if defined(PMAP_DIAGNOSTIC)
5106 if (pv->pv_pmap == NULL) {
5107 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
5114 * Skip pages which do not have PG_RW set.
5116 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5117 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
5123 if (pv_hold_try(pv)) {
5124 vm_page_spin_unlock(m);
5126 vm_page_spin_unlock(m);
5127 pv_lock(pv); /* held, now do a blocking lock */
5129 if (pv->pv_pmap != pmap || pv->pv_m != m) {
5130 pv_put(pv); /* and release */
5131 goto restart; /* anything could have happened */
5133 KKASSERT(pv->pv_pmap == pmap);
5139 nbits = pbits & ~(pmap->pmap_bits[PG_RW_IDX] |
5140 pmap->pmap_bits[PG_M_IDX]);
5141 if (pmap_inval_smp_cmpset(pmap,
5142 ((vm_offset_t)pv->pv_pindex << PAGE_SHIFT),
5143 pte, pbits, nbits)) {
5148 vm_page_spin_lock(m);
5151 * If PG_M was found to be set while we were clearing PG_RW
5152 * we also clear PG_M (done above) and mark the page dirty.
5153 * Callers expect this behavior.
5155 if (pbits & pmap->pmap_bits[PG_M_IDX])
5159 vm_page_spin_unlock(m);
5163 * Lower the permission for all mappings to a given page.
5165 * Page must be busied by caller. Because page is busied by caller this
5166 * should not be able to race a pmap_enter().
5169 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5171 /* JG NX support? */
5172 if ((prot & VM_PROT_WRITE) == 0) {
5173 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5175 * NOTE: pmap_clearbit(.. PG_RW) also clears
5176 * the PG_WRITEABLE flag in (m).
5178 pmap_clearbit(m, PG_RW_IDX);
5186 pmap_phys_address(vm_pindex_t ppn)
5188 return (x86_64_ptob(ppn));
5192 * Return a count of reference bits for a page, clearing those bits.
5193 * It is not necessary for every reference bit to be cleared, but it
5194 * is necessary that 0 only be returned when there are truly no
5195 * reference bits set.
5197 * XXX: The exact number of bits to check and clear is a matter that
5198 * should be tested and standardized at some point in the future for
5199 * optimal aging of shared pages.
5201 * This routine may not block.
5204 pmap_ts_referenced(vm_page_t m)
5211 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5214 vm_page_spin_lock(m);
5215 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
5216 if (!pmap_track_modified(pv->pv_pindex))
5219 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
5220 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
5221 atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
5227 vm_page_spin_unlock(m);
5234 * Return whether or not the specified physical page was modified
5235 * in any physical maps.
5238 pmap_is_modified(vm_page_t m)
5242 res = pmap_testbit(m, PG_M_IDX);
5247 * Clear the modify bits on the specified physical page.
5250 pmap_clear_modify(vm_page_t m)
5252 pmap_clearbit(m, PG_M_IDX);
5256 * pmap_clear_reference:
5258 * Clear the reference bit on the specified physical page.
5261 pmap_clear_reference(vm_page_t m)
5263 pmap_clearbit(m, PG_A_IDX);
5267 * Miscellaneous support routines follow
5272 i386_protection_init(void)
5276 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
5277 kp = protection_codes;
5278 for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5280 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5282 * Read access is also 0. There isn't any execute bit,
5283 * so just make it readable.
5285 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5286 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5287 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5290 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5291 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5292 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5293 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5294 *kp++ = pmap_bits_default[PG_RW_IDX];
5301 * Map a set of physical memory pages into the kernel virtual
5302 * address space. Return a pointer to where it is mapped. This
5303 * routine is intended to be used for mapping device memory,
5306 * NOTE: We can't use pgeflag unless we invalidate the pages one at
5309 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5310 * work whether the cpu supports PAT or not. The remaining PAT
5311 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5315 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5317 return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5321 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5323 return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5327 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5329 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5333 * Map a set of physical memory pages into the kernel virtual
5334 * address space. Return a pointer to where it is mapped. This
5335 * routine is intended to be used for mapping device memory,
5339 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5341 vm_offset_t va, tmpva, offset;
5345 offset = pa & PAGE_MASK;
5346 size = roundup(offset + size, PAGE_SIZE);
5348 va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5350 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5352 pa = pa & ~PAGE_MASK;
5353 for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5354 pte = vtopte(tmpva);
5356 kernel_pmap.pmap_bits[PG_RW_IDX] |
5357 kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5358 kernel_pmap.pmap_cache_bits[mode];
5359 tmpsize -= PAGE_SIZE;
5363 pmap_invalidate_range(&kernel_pmap, va, va + size);
5364 pmap_invalidate_cache_range(va, va + size);
5366 return ((void *)(va + offset));
5370 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5372 vm_offset_t base, offset;
5374 base = va & ~PAGE_MASK;
5375 offset = va & PAGE_MASK;
5376 size = roundup(offset + size, PAGE_SIZE);
5377 pmap_qremove(va, size >> PAGE_SHIFT);
5378 kmem_free(&kernel_map, base, size);
5382 * Sets the memory attribute for the specified page.
5385 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5391 * If "m" is a normal page, update its direct mapping. This update
5392 * can be relied upon to perform any cache operations that are
5393 * required for data coherence.
5395 if ((m->flags & PG_FICTITIOUS) == 0)
5396 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5400 * Change the PAT attribute on an existing kernel memory map. Caller
5401 * must ensure that the virtual memory in question is not accessed
5402 * during the adjustment.
5405 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5412 panic("pmap_change_attr: va is NULL");
5413 base = trunc_page(va);
5417 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
5418 kernel_pmap.pmap_cache_bits[mode];
5423 changed = 1; /* XXX: not optimal */
5426 * Flush CPU caches if required to make sure any data isn't cached that
5427 * shouldn't be, etc.
5430 pmap_invalidate_range(&kernel_pmap, base, va);
5431 pmap_invalidate_cache_range(base, va);
5436 * perform the pmap work for mincore
5439 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5441 pt_entry_t *ptep, pte;
5445 lwkt_gettoken(&pmap->pm_token);
5446 ptep = pmap_pte(pmap, addr);
5448 if (ptep && (pte = *ptep) != 0) {
5451 val = MINCORE_INCORE;
5452 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5455 pa = pte & PG_FRAME;
5457 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5460 m = PHYS_TO_VM_PAGE(pa);
5465 if (pte & pmap->pmap_bits[PG_M_IDX])
5466 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5468 * Modified by someone
5470 else if (m && (m->dirty || pmap_is_modified(m)))
5471 val |= MINCORE_MODIFIED_OTHER;
5475 if (pte & pmap->pmap_bits[PG_A_IDX])
5476 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5479 * Referenced by someone
5481 else if (m && ((m->flags & PG_REFERENCED) ||
5482 pmap_ts_referenced(m))) {
5483 val |= MINCORE_REFERENCED_OTHER;
5484 vm_page_flag_set(m, PG_REFERENCED);
5488 lwkt_reltoken(&pmap->pm_token);
5494 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5495 * vmspace will be ref'd and the old one will be deref'd.
5497 * The vmspace for all lwps associated with the process will be adjusted
5498 * and cr3 will be reloaded if any lwp is the current lwp.
5500 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5503 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5505 struct vmspace *oldvm;
5508 oldvm = p->p_vmspace;
5509 if (oldvm != newvm) {
5512 p->p_vmspace = newvm;
5513 KKASSERT(p->p_nthreads == 1);
5514 lp = RB_ROOT(&p->p_lwp_tree);
5515 pmap_setlwpvm(lp, newvm);
5522 * Set the vmspace for a LWP. The vmspace is almost universally set the
5523 * same as the process vmspace, but virtual kernels need to swap out contexts
5524 * on a per-lwp basis.
5526 * Caller does not necessarily hold any vmspace tokens. Caller must control
5527 * the lwp (typically be in the context of the lwp). We use a critical
5528 * section to protect against statclock and hardclock (statistics collection).
5531 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5533 struct vmspace *oldvm;
5536 oldvm = lp->lwp_vmspace;
5538 if (oldvm != newvm) {
5540 lp->lwp_vmspace = newvm;
5541 if (curthread->td_lwp == lp) {
5542 pmap = vmspace_pmap(newvm);
5543 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
5544 if (pmap->pm_active_lock & CPULOCK_EXCL)
5545 pmap_interlock_wait(newvm);
5546 #if defined(SWTCH_OPTIM_STATS)
5549 if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5550 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5551 } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5552 curthread->td_pcb->pcb_cr3 = KPML4phys;
5554 panic("pmap_setlwpvm: unknown pmap type\n");
5556 load_cr3(curthread->td_pcb->pcb_cr3);
5557 pmap = vmspace_pmap(oldvm);
5558 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
5566 * Called when switching to a locked pmap, used to interlock against pmaps
5567 * undergoing modifications to prevent us from activating the MMU for the
5568 * target pmap until all such modifications have completed. We have to do
5569 * this because the thread making the modifications has already set up its
5570 * SMP synchronization mask.
5572 * This function cannot sleep!
5577 pmap_interlock_wait(struct vmspace *vm)
5579 struct pmap *pmap = &vm->vm_pmap;
5581 if (pmap->pm_active_lock & CPULOCK_EXCL) {
5583 KKASSERT(curthread->td_critcount >= 2);
5584 DEBUG_PUSH_INFO("pmap_interlock_wait");
5585 while (pmap->pm_active_lock & CPULOCK_EXCL) {
5587 lwkt_process_ipiq();
5595 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5598 if ((obj == NULL) || (size < NBPDR) ||
5599 ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5603 addr = roundup2(addr, NBPDR);
5608 * Used by kmalloc/kfree, page already exists at va
5611 pmap_kvtom(vm_offset_t va)
5613 pt_entry_t *ptep = vtopte(va);
5615 KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5616 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5620 * Initialize machine-specific shared page directory support. This
5621 * is executed when a VM object is created.
5624 pmap_object_init(vm_object_t object)
5626 object->md.pmap_rw = NULL;
5627 object->md.pmap_ro = NULL;
5631 * Clean up machine-specific shared page directory support. This
5632 * is executed when a VM object is destroyed.
5635 pmap_object_free(vm_object_t object)
5639 if ((pmap = object->md.pmap_rw) != NULL) {
5640 object->md.pmap_rw = NULL;
5641 pmap_remove_noinval(pmap,
5642 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5643 CPUMASK_ASSZERO(pmap->pm_active);
5646 kfree(pmap, M_OBJPMAP);
5648 if ((pmap = object->md.pmap_ro) != NULL) {
5649 object->md.pmap_ro = NULL;
5650 pmap_remove_noinval(pmap,
5651 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5652 CPUMASK_ASSZERO(pmap->pm_active);
5655 kfree(pmap, M_OBJPMAP);
5660 * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
5661 * VM page and issue a pginfo->callback.
5663 * We are expected to dispose of any non-NULL pte_pv.
5667 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
5668 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
5669 vm_offset_t va, pt_entry_t *ptep, void *arg)
5671 struct pmap_pgscan_info *pginfo = arg;
5676 * Try to busy the page while we hold the pte_pv locked.
5678 m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
5679 if (vm_page_busy_try(m, TRUE) == 0) {
5680 if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
5682 * The callback is issued with the pte_pv
5683 * unlocked and put away, and the pt_pv
5689 if (pginfo->callback(pginfo, va, m) < 0)
5698 ++pginfo->busycount;
5701 } else if (sharept) {
5702 /* shared page table */
5704 /* else unmanaged page */
5709 pmap_pgscan(struct pmap_pgscan_info *pginfo)
5711 struct pmap_scan_info info;
5713 pginfo->offset = pginfo->beg_addr;
5714 info.pmap = pginfo->pmap;
5715 info.sva = pginfo->beg_addr;
5716 info.eva = pginfo->end_addr;
5717 info.func = pmap_pgscan_callback;
5719 pmap_scan(&info, 0);
5721 pginfo->offset = pginfo->end_addr;