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.
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
48 * Manages physical address maps.
51 #include "opt_msgbuf.h"
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
57 #include <sys/msgbuf.h>
58 #include <sys/vmmeter.h>
60 #include <sys/vmspace.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/globaldata.h>
86 #include <machine/pmap.h>
87 #include <machine/pmap_inval.h>
96 #define PMAP_KEEP_PDIRS
97 #ifndef PMAP_SHPGPERPROC
98 #define PMAP_SHPGPERPROC 1000
101 #if defined(DIAGNOSTIC)
102 #define PMAP_DIAGNOSTIC
107 #if !defined(PMAP_DIAGNOSTIC)
108 #define PMAP_INLINE __inline
114 * Get PDEs and PTEs for user/kernel address space
116 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
117 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
119 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & VPTE_V) != 0)
120 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & VPTE_WIRED) != 0)
121 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & VPTE_M) != 0)
122 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & VPTE_A) != 0)
123 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & VPTE_V) != 0)
126 * Given a map and a machine independent protection code,
127 * convert to a vax protection code.
129 #define pte_prot(m, p) \
130 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
131 static int protection_codes[8];
133 struct pmap kernel_pmap;
135 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
137 static struct vm_object kptobj;
140 static uint64_t KPDphys; /* phys addr of kernel level 2 */
141 uint64_t KPDPphys; /* phys addr of kernel level 3 */
142 uint64_t KPML4phys; /* phys addr of kernel level 4 */
144 extern int vmm_enabled;
145 extern void *vkernel_stack;
148 * Data for the pv entry allocation mechanism
150 static vm_zone_t pvzone;
151 static struct vm_zone pvzone_store;
152 static int pv_entry_count = 0;
153 static int pv_entry_max = 0;
154 static int pv_entry_high_water = 0;
155 static int pmap_pagedaemon_waken = 0;
156 static struct pv_entry *pvinit;
159 * All those kernel PT submaps that BSD is so fond of
161 pt_entry_t *CMAP1 = NULL, *ptmmap;
162 caddr_t CADDR1 = NULL;
163 static pt_entry_t *msgbufmap;
167 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
168 static pv_entry_t get_pv_entry (void);
169 static void i386_protection_init (void);
170 static __inline void pmap_clearbit (vm_page_t m, int bit);
172 static void pmap_remove_all (vm_page_t m);
173 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
174 pt_entry_t oldpte, vm_offset_t sva);
175 static void pmap_remove_page (struct pmap *pmap, vm_offset_t va);
176 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
178 static boolean_t pmap_testbit (vm_page_t m, int bit);
179 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
180 vm_page_t mpte, vm_page_t m, pv_entry_t);
182 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
184 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
185 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
186 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
187 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t);
190 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
192 if (pv1->pv_va < pv2->pv_va)
194 if (pv1->pv_va > pv2->pv_va)
199 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
200 pv_entry_compare, vm_offset_t, pv_va);
202 static __inline vm_pindex_t
203 pmap_pt_pindex(vm_offset_t va)
205 return va >> PDRSHIFT;
208 static __inline vm_pindex_t
209 pmap_pte_index(vm_offset_t va)
211 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
214 static __inline vm_pindex_t
215 pmap_pde_index(vm_offset_t va)
217 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
220 static __inline vm_pindex_t
221 pmap_pdpe_index(vm_offset_t va)
223 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
226 static __inline vm_pindex_t
227 pmap_pml4e_index(vm_offset_t va)
229 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
232 /* Return a pointer to the PML4 slot that corresponds to a VA */
233 static __inline pml4_entry_t *
234 pmap_pml4e(pmap_t pmap, vm_offset_t va)
236 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
239 /* Return a pointer to the PDP slot that corresponds to a VA */
240 static __inline pdp_entry_t *
241 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
245 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & VPTE_FRAME);
246 return (&pdpe[pmap_pdpe_index(va)]);
249 /* Return a pointer to the PDP slot that corresponds to a VA */
250 static __inline pdp_entry_t *
251 pmap_pdpe(pmap_t pmap, vm_offset_t va)
255 pml4e = pmap_pml4e(pmap, va);
256 if ((*pml4e & VPTE_V) == 0)
258 return (pmap_pml4e_to_pdpe(pml4e, va));
261 /* Return a pointer to the PD slot that corresponds to a VA */
262 static __inline pd_entry_t *
263 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
267 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & VPTE_FRAME);
268 return (&pde[pmap_pde_index(va)]);
271 /* Return a pointer to the PD slot that corresponds to a VA */
272 static __inline pd_entry_t *
273 pmap_pde(pmap_t pmap, vm_offset_t va)
277 pdpe = pmap_pdpe(pmap, va);
278 if (pdpe == NULL || (*pdpe & VPTE_V) == 0)
280 return (pmap_pdpe_to_pde(pdpe, va));
283 /* Return a pointer to the PT slot that corresponds to a VA */
284 static __inline pt_entry_t *
285 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
289 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & VPTE_FRAME);
290 return (&pte[pmap_pte_index(va)]);
294 * Hold pt_m for page table scans to prevent it from getting reused out
295 * from under us across blocking conditions in the body of the loop.
299 pmap_hold_pt_page(pd_entry_t *pde, vm_offset_t va)
304 pte = (pt_entry_t)*pde;
306 pt_m = PHYS_TO_VM_PAGE(pte & VPTE_FRAME);
312 /* Return a pointer to the PT slot that corresponds to a VA */
313 static __inline pt_entry_t *
314 pmap_pte(pmap_t pmap, vm_offset_t va)
318 pde = pmap_pde(pmap, va);
319 if (pde == NULL || (*pde & VPTE_V) == 0)
321 if ((*pde & VPTE_PS) != 0) /* compat with i386 pmap_pte() */
322 return ((pt_entry_t *)pde);
323 return (pmap_pde_to_pte(pde, va));
326 static PMAP_INLINE pt_entry_t *
327 vtopte(vm_offset_t va)
330 x = pmap_pte(&kernel_pmap, va);
335 static __inline pd_entry_t *
336 vtopde(vm_offset_t va)
339 x = pmap_pde(&kernel_pmap, va);
345 allocpages(vm_paddr_t *firstaddr, int n)
350 /*bzero((void *)ret, n * PAGE_SIZE); not mapped yet */
351 *firstaddr += n * PAGE_SIZE;
356 create_dmap_vmm(vm_paddr_t *firstaddr)
359 int pml4_stack_index;
366 uint64_t KPDP_DMAP_phys = allocpages(firstaddr, NDMPML4E);
367 uint64_t KPDP_VSTACK_phys = allocpages(firstaddr, 1);
368 uint64_t KPD_VSTACK_phys = allocpages(firstaddr, 1);
370 pml4_entry_t *KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
371 pdp_entry_t *KPDP_DMAP_virt = (pdp_entry_t *)PHYS_TO_DMAP(KPDP_DMAP_phys);
372 pdp_entry_t *KPDP_VSTACK_virt = (pdp_entry_t *)PHYS_TO_DMAP(KPDP_VSTACK_phys);
373 pd_entry_t *KPD_VSTACK_virt = (pd_entry_t *)PHYS_TO_DMAP(KPD_VSTACK_phys);
375 bzero(KPDP_DMAP_virt, NDMPML4E * PAGE_SIZE);
376 bzero(KPDP_VSTACK_virt, 1 * PAGE_SIZE);
377 bzero(KPD_VSTACK_virt, 1 * PAGE_SIZE);
379 do_cpuid(0x80000001, regs);
380 amd_feature = regs[3];
382 /* Build the mappings for the first 512GB */
383 if (amd_feature & AMDID_PAGE1GB) {
384 /* In pages of 1 GB, if supported */
385 for (i = 0; i < NPDPEPG; i++) {
386 KPDP_DMAP_virt[i] = ((uint64_t)i << PDPSHIFT);
387 KPDP_DMAP_virt[i] |= VPTE_RW | VPTE_V | VPTE_PS | VPTE_U;
390 /* In page of 2MB, otherwise */
391 for (i = 0; i < NPDPEPG; i++) {
392 uint64_t KPD_DMAP_phys;
393 pd_entry_t *KPD_DMAP_virt;
395 KPD_DMAP_phys = allocpages(firstaddr, 1);
397 (pd_entry_t *)PHYS_TO_DMAP(KPD_DMAP_phys);
399 bzero(KPD_DMAP_virt, PAGE_SIZE);
401 KPDP_DMAP_virt[i] = KPD_DMAP_phys;
402 KPDP_DMAP_virt[i] |= VPTE_RW | VPTE_V | VPTE_U;
404 /* For each PD, we have to allocate NPTEPG PT */
405 for (j = 0; j < NPTEPG; j++) {
406 KPD_DMAP_virt[j] = (i << PDPSHIFT) |
408 KPD_DMAP_virt[j] |= VPTE_RW | VPTE_V |
414 /* DMAP for the first 512G */
415 KPML4virt[0] = KPDP_DMAP_phys;
416 KPML4virt[0] |= VPTE_RW | VPTE_V | VPTE_U;
418 /* create a 2 MB map of the new stack */
419 pml4_stack_index = (uint64_t)&stack_addr >> PML4SHIFT;
420 KPML4virt[pml4_stack_index] = KPDP_VSTACK_phys;
421 KPML4virt[pml4_stack_index] |= VPTE_RW | VPTE_V | VPTE_U;
423 pdp_stack_index = ((uint64_t)&stack_addr & PML4MASK) >> PDPSHIFT;
424 KPDP_VSTACK_virt[pdp_stack_index] = KPD_VSTACK_phys;
425 KPDP_VSTACK_virt[pdp_stack_index] |= VPTE_RW | VPTE_V | VPTE_U;
427 pd_stack_index = ((uint64_t)&stack_addr & PDPMASK) >> PDRSHIFT;
428 KPD_VSTACK_virt[pd_stack_index] = (uint64_t) vkernel_stack;
429 KPD_VSTACK_virt[pd_stack_index] |= VPTE_RW | VPTE_V | VPTE_U | VPTE_PS;
433 create_pagetables(vm_paddr_t *firstaddr, int64_t ptov_offset)
436 pml4_entry_t *KPML4virt;
437 pdp_entry_t *KPDPvirt;
440 int kpml4i = pmap_pml4e_index(ptov_offset);
441 int kpdpi = pmap_pdpe_index(ptov_offset);
442 int kpdi = pmap_pde_index(ptov_offset);
445 * Calculate NKPT - number of kernel page tables. We have to
446 * accomodoate prealloction of the vm_page_array, dump bitmap,
447 * MSGBUF_SIZE, and other stuff. Be generous.
449 * Maxmem is in pages.
451 nkpt = (Maxmem * (sizeof(struct vm_page) * 2) + MSGBUF_SIZE) / NBPDR;
455 KPML4phys = allocpages(firstaddr, 1);
456 KPDPphys = allocpages(firstaddr, NKPML4E);
457 KPDphys = allocpages(firstaddr, NKPDPE);
458 KPTphys = allocpages(firstaddr, nkpt);
460 KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
461 KPDPvirt = (pdp_entry_t *)PHYS_TO_DMAP(KPDPphys);
462 KPDvirt = (pd_entry_t *)PHYS_TO_DMAP(KPDphys);
463 KPTvirt = (pt_entry_t *)PHYS_TO_DMAP(KPTphys);
465 bzero(KPML4virt, 1 * PAGE_SIZE);
466 bzero(KPDPvirt, NKPML4E * PAGE_SIZE);
467 bzero(KPDvirt, NKPDPE * PAGE_SIZE);
468 bzero(KPTvirt, nkpt * PAGE_SIZE);
470 /* Now map the page tables at their location within PTmap */
471 for (i = 0; i < nkpt; i++) {
472 KPDvirt[i + kpdi] = KPTphys + (i << PAGE_SHIFT);
473 KPDvirt[i + kpdi] |= VPTE_RW | VPTE_V | VPTE_U;
476 /* And connect up the PD to the PDP */
477 for (i = 0; i < NKPDPE; i++) {
478 KPDPvirt[i + kpdpi] = KPDphys + (i << PAGE_SHIFT);
479 KPDPvirt[i + kpdpi] |= VPTE_RW | VPTE_V | VPTE_U;
482 /* And recursively map PML4 to itself in order to get PTmap */
483 KPML4virt[PML4PML4I] = KPML4phys;
484 KPML4virt[PML4PML4I] |= VPTE_RW | VPTE_V | VPTE_U;
486 /* Connect the KVA slot up to the PML4 */
487 KPML4virt[kpml4i] = KPDPphys;
488 KPML4virt[kpml4i] |= VPTE_RW | VPTE_V | VPTE_U;
492 * Typically used to initialize a fictitious page by vm/device_pager.c
495 pmap_page_init(struct vm_page *m)
498 TAILQ_INIT(&m->md.pv_list);
502 * Bootstrap the system enough to run with virtual memory.
504 * On the i386 this is called after mapping has already been enabled
505 * and just syncs the pmap module with what has already been done.
506 * [We can't call it easily with mapping off since the kernel is not
507 * mapped with PA == VA, hence we would have to relocate every address
508 * from the linked base (virtual) address "KERNBASE" to the actual
509 * (physical) address starting relative to 0]
512 pmap_bootstrap(vm_paddr_t *firstaddr, int64_t ptov_offset)
518 * Create an initial set of page tables to run the kernel in.
520 create_pagetables(firstaddr, ptov_offset);
522 /* Create the DMAP for the VMM */
524 create_dmap_vmm(firstaddr);
527 virtual_start = KvaStart;
528 virtual_end = KvaEnd;
531 * Initialize protection array.
533 i386_protection_init();
536 * The kernel's pmap is statically allocated so we don't have to use
537 * pmap_create, which is unlikely to work correctly at this part of
538 * the boot sequence (XXX and which no longer exists).
540 * The kernel_pmap's pm_pteobj is used only for locking and not
543 kernel_pmap.pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
544 kernel_pmap.pm_count = 1;
545 /* don't allow deactivation */
546 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
547 kernel_pmap.pm_pteobj = NULL; /* see pmap_init */
548 RB_INIT(&kernel_pmap.pm_pvroot);
549 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
552 * Reserve some special page table entries/VA space for temporary
555 #define SYSMAP(c, p, v, n) \
556 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
559 pte = pmap_pte(&kernel_pmap, va);
561 * CMAP1/CMAP2 are used for zeroing and copying pages.
563 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
569 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
573 * ptvmmap is used for reading arbitrary physical pages via
576 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
579 * msgbufp is used to map the system message buffer.
580 * XXX msgbufmap is not used.
582 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
583 atop(round_page(MSGBUF_SIZE)))
588 /* Not ready to do an invltlb yet for VMM*/
595 * Initialize the pmap module.
596 * Called by vm_init, to initialize any structures that the pmap
597 * system needs to map virtual memory.
598 * pmap_init has been enhanced to support in a fairly consistant
599 * way, discontiguous physical memory.
608 * object for kernel page table pages
610 /* JG I think the number can be arbitrary */
611 vm_object_init(&kptobj, 5);
612 kernel_pmap.pm_pteobj = &kptobj;
615 * Allocate memory for random pmap data structures. Includes the
618 for(i = 0; i < vm_page_array_size; i++) {
621 m = &vm_page_array[i];
622 TAILQ_INIT(&m->md.pv_list);
623 m->md.pv_list_count = 0;
627 * init the pv free list
629 initial_pvs = vm_page_array_size;
630 if (initial_pvs < MINPV)
632 pvzone = &pvzone_store;
633 pvinit = (struct pv_entry *)
634 kmem_alloc(&kernel_map,
635 initial_pvs * sizeof (struct pv_entry),
637 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
641 * Now it is safe to enable pv_table recording.
643 pmap_initialized = TRUE;
647 * Initialize the address space (zone) for the pv_entries. Set a
648 * high water mark so that the system can recover from excessive
649 * numbers of pv entries.
654 int shpgperproc = PMAP_SHPGPERPROC;
656 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
657 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
658 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
659 pv_entry_high_water = 9 * (pv_entry_max / 10);
660 zinitna(pvzone, NULL, 0, pv_entry_max, ZONE_INTERRUPT);
664 /***************************************************
665 * Low level helper routines.....
666 ***************************************************/
669 * The modification bit is not tracked for any pages in this range. XXX
670 * such pages in this maps should always use pmap_k*() functions and not
673 * XXX User and kernel address spaces are independant for virtual kernels,
674 * this function only applies to the kernel pmap.
677 pmap_track_modified(pmap_t pmap, vm_offset_t va)
679 if (pmap != &kernel_pmap)
681 if ((va < clean_sva) || (va >= clean_eva))
688 * Extract the physical page address associated with the map/VA pair.
693 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
697 pd_entry_t pde, *pdep;
699 vm_object_hold(pmap->pm_pteobj);
701 pdep = pmap_pde(pmap, va);
705 if ((pde & VPTE_PS) != 0) {
707 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
709 pte = pmap_pde_to_pte(pdep, va);
710 rtval = (*pte & VPTE_FRAME) | (va & PAGE_MASK);
715 *handlep = NULL; /* XXX */
716 vm_object_drop(pmap->pm_pteobj);
722 pmap_extract_done(void *handle)
728 vm_object_drop(pmap->pm_pteobj);
733 * Similar to extract but checks protections, SMP-friendly short-cut for
734 * vm_fault_page[_quick]().
736 * WARNING! THE RETURNED PAGE IS ONLY HELD AND NEITHER IT NOR ITS TARGET
737 * DATA IS SUITABLE FOR WRITING. Writing can interfere with
738 * pageouts flushes, msync, etc. The hold_count is not enough
739 * to avoid races against pageouts and other flush code doesn't
740 * care about hold_count.
743 pmap_fault_page_quick(pmap_t pmap __unused, vm_offset_t vaddr __unused,
744 vm_prot_t prot __unused, int *busyp __unused)
750 * Routine: pmap_kextract
752 * Extract the physical page address associated
753 * kernel virtual address.
756 pmap_kextract(vm_offset_t va)
761 KKASSERT(va >= KvaStart && va < KvaEnd);
764 * The DMAP region is not included in [KvaStart, KvaEnd)
767 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
768 pa = DMAP_TO_PHYS(va);
774 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
777 * Beware of a concurrent promotion that changes the
778 * PDE at this point! For example, vtopte() must not
779 * be used to access the PTE because it would use the
780 * new PDE. It is, however, safe to use the old PDE
781 * because the page table page is preserved by the
784 pa = *pmap_pde_to_pte(&pde, va);
785 pa = (pa & VPTE_FRAME) | (va & PAGE_MASK);
793 /***************************************************
794 * Low level mapping routines.....
795 ***************************************************/
798 * Enter a mapping into kernel_pmap. Mappings created in this fashion
799 * are not managed. Mappings must be immediately accessible on all cpus.
801 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the
802 * real pmap and handle related races before storing the new vpte. The
803 * new semantics for kenter require use to do an UNCONDITIONAL invalidation,
804 * because the entry may have previously been cleared without an invalidation.
807 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
812 KKASSERT(va >= KvaStart && va < KvaEnd);
813 npte = pa | VPTE_RW | VPTE_V | VPTE_U;
817 pmap_inval_pte(ptep, &kernel_pmap, va);
820 pmap_inval_pte(ptep, &kernel_pmap, va);
822 atomic_swap_long(ptep, npte);
826 * Enter an unmanaged KVA mapping for the private use of the current
829 * It is illegal for the mapping to be accessed by other cpus without
830 * proper invalidation.
833 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
839 KKASSERT(va >= KvaStart && va < KvaEnd);
841 npte = (vpte_t)pa | VPTE_RW | VPTE_V | VPTE_U;
845 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
851 pmap_inval_pte(pte, &kernel_pmap, va);
853 atomic_swap_long(ptep, npte);
859 * Invalidation will occur later, ok to be lazy here.
862 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
868 KKASSERT(va >= KvaStart && va < KvaEnd);
870 npte = (vpte_t)pa | VPTE_RW | VPTE_V | VPTE_U;
878 atomic_swap_long(ptep, npte);
884 * Remove an unmanaged mapping created with pmap_kenter*().
887 pmap_kremove(vm_offset_t va)
891 KKASSERT(va >= KvaStart && va < KvaEnd);
894 atomic_swap_long(ptep, 0);
895 pmap_inval_pte(ptep, &kernel_pmap, va);
899 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
900 * only with this cpu.
902 * Unfortunately because we optimize new entries by testing VPTE_V later
903 * on, we actually still have to synchronize with all the cpus. XXX maybe
904 * store a junk value and test against 0 in the other places instead?
907 pmap_kremove_quick(vm_offset_t va)
911 KKASSERT(va >= KvaStart && va < KvaEnd);
914 atomic_swap_long(ptep, 0);
915 pmap_inval_pte(ptep, &kernel_pmap, va); /* NOT _quick */
919 * Invalidation will occur later, ok to be lazy here.
922 pmap_kremove_noinval(vm_offset_t va)
926 KKASSERT(va >= KvaStart && va < KvaEnd);
929 atomic_swap_long(ptep, 0);
933 * Used to map a range of physical addresses into kernel
934 * virtual address space.
936 * For now, VM is already on, we only need to map the
940 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
942 return PHYS_TO_DMAP(start);
946 * Map a set of unmanaged VM pages into KVM.
949 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
954 end_va = beg_va + count * PAGE_SIZE;
955 KKASSERT(beg_va >= KvaStart && end_va <= KvaEnd);
957 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
961 atomic_swap_long(ptep, VM_PAGE_TO_PHYS(*m) |
962 VPTE_RW | VPTE_V | VPTE_U);
965 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
966 /* pmap_inval_pte(pte, &kernel_pmap, va); */
970 * Undo the effects of pmap_qenter*().
973 pmap_qremove(vm_offset_t beg_va, int count)
978 end_va = beg_va + count * PAGE_SIZE;
979 KKASSERT(beg_va >= KvaStart && end_va < KvaEnd);
981 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
985 atomic_swap_long(ptep, 0);
987 pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
991 * Unlike the real pmap code, we can't avoid calling the real-kernel.
994 pmap_qremove_quick(vm_offset_t va, int count)
996 pmap_qremove(va, count);
1000 pmap_qremove_noinval(vm_offset_t va, int count)
1002 pmap_qremove(va, count);
1006 * This routine works like vm_page_lookup() but also blocks as long as the
1007 * page is busy. This routine does not busy the page it returns.
1009 * Unless the caller is managing objects whos pages are in a known state,
1010 * the call should be made with a critical section held so the page's object
1011 * association remains valid on return.
1014 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1018 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1019 m = vm_page_lookup_busy_wait(object, pindex, TRUE, "pplookp");
1025 * Create a new thread and optionally associate it with a (new) process.
1026 * NOTE! the new thread's cpu may not equal the current cpu.
1029 pmap_init_thread(thread_t td)
1031 /* enforce pcb placement */
1032 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1033 td->td_savefpu = &td->td_pcb->pcb_save;
1034 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
1038 * This routine directly affects the fork perf for a process.
1041 pmap_init_proc(struct proc *p)
1046 * Unwire a page table which has been removed from the pmap. We own the
1047 * wire_count, so the page cannot go away. The page representing the page
1048 * table is passed in unbusied and must be busied if we cannot trivially
1051 * XXX NOTE! This code is not usually run because we do not currently
1052 * implement dynamic page table page removal. The page in
1053 * its parent assumes at least 1 wire count, so no call to this
1054 * function ever sees a wire count less than 2.
1057 pmap_unwire_pgtable(pmap_t pmap, vm_offset_t va, vm_page_t m)
1060 * Try to unwire optimally. If non-zero is returned the wire_count
1061 * is 1 and we must busy the page to unwire it.
1063 if (vm_page_unwire_quick(m) == 0)
1066 vm_page_busy_wait(m, TRUE, "pmuwpt");
1067 KASSERT(m->queue == PQ_NONE,
1068 ("_pmap_unwire_pgtable: %p->queue != PQ_NONE", m));
1070 if (m->wire_count == 1) {
1072 * Unmap the page table page.
1074 /* pmap_inval_add(info, pmap, -1); */
1076 if (m->pindex >= (NUPT_TOTAL + NUPD_TOTAL)) {
1079 pml4 = pmap_pml4e(pmap, va);
1081 } else if (m->pindex >= NUPT_TOTAL) {
1084 pdp = pmap_pdpe(pmap, va);
1089 pd = pmap_pde(pmap, va);
1093 KKASSERT(pmap->pm_stats.resident_count > 0);
1094 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1096 if (pmap->pm_ptphint == m)
1097 pmap->pm_ptphint = NULL;
1099 if (m->pindex < NUPT_TOTAL) {
1100 /* We just released a PT, unhold the matching PD */
1103 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) &
1105 pmap_unwire_pgtable(pmap, va, pdpg);
1107 if (m->pindex >= NUPT_TOTAL &&
1108 m->pindex < (NUPT_TOTAL + NUPD_TOTAL)) {
1109 /* We just released a PD, unhold the matching PDP */
1112 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) &
1114 pmap_unwire_pgtable(pmap, va, pdppg);
1118 * This was our last wire, the page had better be unwired
1119 * after we decrement wire_count.
1121 * FUTURE NOTE: shared page directory page could result in
1122 * multiple wire counts.
1124 vm_page_unwire(m, 0);
1125 KKASSERT(m->wire_count == 0);
1126 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1131 /* XXX SMP race to 1 if not holding vmobj */
1132 vm_page_unwire(m, 0);
1139 * After removing a page table entry, this routine is used to
1140 * conditionally free the page, and manage the hold/wire counts.
1142 * If not NULL the caller owns a wire_count on mpte, so it can't disappear.
1143 * If NULL the caller owns a wire_count on what would be the mpte, we must
1147 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1149 vm_pindex_t ptepindex;
1151 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1155 * page table pages in the kernel_pmap are not managed.
1157 if (pmap == &kernel_pmap)
1159 ptepindex = pmap_pt_pindex(va);
1160 if (pmap->pm_ptphint &&
1161 (pmap->pm_ptphint->pindex == ptepindex)) {
1162 mpte = pmap->pm_ptphint;
1164 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1165 pmap->pm_ptphint = mpte;
1166 vm_page_wakeup(mpte);
1169 return pmap_unwire_pgtable(pmap, va, mpte);
1173 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
1174 * just dummy it up so it works well enough for fork().
1176 * In DragonFly, process pmaps may only be used to manipulate user address
1177 * space, never kernel address space.
1180 pmap_pinit0(struct pmap *pmap)
1186 * Initialize a preallocated and zeroed pmap structure,
1187 * such as one in a vmspace structure.
1190 pmap_pinit(struct pmap *pmap)
1195 * No need to allocate page table space yet but we do need a valid
1196 * page directory table.
1198 if (pmap->pm_pml4 == NULL) {
1199 pmap->pm_pml4 = (pml4_entry_t *)
1200 kmem_alloc_pageable(&kernel_map, PAGE_SIZE,
1205 * Allocate an object for the ptes
1207 if (pmap->pm_pteobj == NULL)
1208 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL + 1);
1211 * Allocate the page directory page, unless we already have
1212 * one cached. If we used the cached page the wire_count will
1213 * already be set appropriately.
1215 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1216 ptdpg = vm_page_grab(pmap->pm_pteobj,
1217 NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL,
1218 VM_ALLOC_NORMAL | VM_ALLOC_RETRY |
1220 pmap->pm_pdirm = ptdpg;
1221 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_WRITEABLE);
1222 vm_page_wire(ptdpg);
1223 vm_page_wakeup(ptdpg);
1224 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1227 CPUMASK_ASSZERO(pmap->pm_active);
1228 pmap->pm_ptphint = NULL;
1229 RB_INIT(&pmap->pm_pvroot);
1230 spin_init(&pmap->pm_spin, "pmapinit");
1231 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1232 pmap->pm_stats.resident_count = 1;
1233 pmap->pm_stats.wired_count = 1;
1237 * Clean up a pmap structure so it can be physically freed. This routine
1238 * is called by the vmspace dtor function. A great deal of pmap data is
1239 * left passively mapped to improve vmspace management so we have a bit
1240 * of cleanup work to do here.
1245 pmap_puninit(pmap_t pmap)
1249 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1250 if ((p = pmap->pm_pdirm) != NULL) {
1251 KKASSERT(pmap->pm_pml4 != NULL);
1252 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1253 vm_page_busy_wait(p, TRUE, "pgpun");
1254 vm_page_unwire(p, 0);
1255 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1257 pmap->pm_pdirm = NULL;
1258 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1259 KKASSERT(pmap->pm_stats.wired_count == 0);
1261 if (pmap->pm_pml4) {
1262 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1263 pmap->pm_pml4 = NULL;
1265 if (pmap->pm_pteobj) {
1266 vm_object_deallocate(pmap->pm_pteobj);
1267 pmap->pm_pteobj = NULL;
1272 * This function is now unused (used to add the pmap to the pmap_list)
1275 pmap_pinit2(struct pmap *pmap)
1280 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1281 * 0 on failure (if the procedure had to sleep).
1283 * When asked to remove the page directory page itself, we actually just
1284 * leave it cached so we do not have to incur the SMP inval overhead of
1285 * removing the kernel mapping. pmap_puninit() will take care of it.
1288 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1291 * This code optimizes the case of freeing non-busy
1292 * page-table pages. Those pages are zero now, and
1293 * might as well be placed directly into the zero queue.
1295 if (vm_page_busy_try(p, TRUE)) {
1296 vm_page_sleep_busy(p, TRUE, "pmaprl");
1301 * Remove the page table page from the processes address space.
1303 if (p->pindex == NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL) {
1305 * We are the pml4 table itself.
1307 /* XXX anything to do here? */
1308 } else if (p->pindex >= (NUPT_TOTAL + NUPD_TOTAL)) {
1310 * We are a PDP page.
1311 * We look for the PML4 entry that points to us.
1317 m4 = vm_page_lookup(pmap->pm_pteobj,
1318 NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
1319 KKASSERT(m4 != NULL);
1320 pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1321 idx = (p->pindex - (NUPT_TOTAL + NUPD_TOTAL)) % NPML4EPG;
1322 KKASSERT(pml4[idx] != 0);
1324 kprintf("pmap_release: Unmapped PML4\n");
1326 vm_page_unwire_quick(m4);
1327 } else if (p->pindex >= NUPT_TOTAL) {
1330 * We look for the PDP entry that points to us.
1336 m3 = vm_page_lookup(pmap->pm_pteobj,
1337 NUPT_TOTAL + NUPD_TOTAL +
1338 (p->pindex - NUPT_TOTAL) / NPDPEPG);
1339 KKASSERT(m3 != NULL);
1340 pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1341 idx = (p->pindex - NUPT_TOTAL) % NPDPEPG;
1342 KKASSERT(pdp[idx] != 0);
1344 kprintf("pmap_release: Unmapped PDP %d\n", idx);
1346 vm_page_unwire_quick(m3);
1348 /* We are a PT page.
1349 * We look for the PD entry that points to us.
1355 m2 = vm_page_lookup(pmap->pm_pteobj,
1356 NUPT_TOTAL + p->pindex / NPDEPG);
1357 KKASSERT(m2 != NULL);
1358 pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1359 idx = p->pindex % NPDEPG;
1361 kprintf("pmap_release: Unmapped PD %d\n", idx);
1363 vm_page_unwire_quick(m2);
1365 KKASSERT(pmap->pm_stats.resident_count > 0);
1366 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1368 if (p->wire_count > 1) {
1369 panic("pmap_release: freeing held pt page "
1370 "pmap=%p pg=%p dmap=%p pi=%ld {%ld,%ld,%ld}",
1371 pmap, p, (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(p)),
1372 p->pindex, NUPT_TOTAL, NUPD_TOTAL, NUPDP_TOTAL);
1375 if (pmap->pm_ptphint == p)
1376 pmap->pm_ptphint = NULL;
1379 * We leave the top-level page table page cached, wired, and mapped in
1380 * the pmap until the dtor function (pmap_puninit()) gets called.
1381 * However, still clean it up.
1383 if (p->pindex == NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL) {
1384 bzero(pmap->pm_pml4, PAGE_SIZE);
1387 vm_page_unwire(p, 0);
1388 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1390 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1396 * Locate the requested PT, PD, or PDP page table page.
1398 * Returns a busied page, caller must vm_page_wakeup() when done.
1401 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1410 * Find or fabricate a new pagetable page. A non-zero wire_count
1411 * indicates that the page has already been mapped into its parent.
1413 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1414 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1415 if (m->wire_count != 0)
1419 * Map the page table page into its parent, giving it 1 wire count.
1422 vm_page_unmanage(m);
1423 atomic_add_long(&pmap->pm_stats.resident_count, 1);
1424 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1426 data = VM_PAGE_TO_PHYS(m) |
1427 VPTE_RW | VPTE_V | VPTE_U | VPTE_A | VPTE_M | VPTE_WIRED;
1428 atomic_add_long(&pmap->pm_stats.wired_count, 1);
1430 if (ptepindex >= (NUPT_TOTAL + NUPD_TOTAL)) {
1432 * Map PDP into the PML4
1434 pindex = ptepindex - (NUPT_TOTAL + NUPD_TOTAL);
1435 pindex &= (NUPDP_TOTAL - 1);
1436 ptep = (pt_entry_t *)pmap->pm_pml4;
1438 } else if (ptepindex >= NUPT_TOTAL) {
1440 * Map PD into its PDP
1442 pindex = (ptepindex - NUPT_TOTAL) >> NPDPEPGSHIFT;
1443 pindex += NUPT_TOTAL + NUPD_TOTAL;
1444 pm = _pmap_allocpte(pmap, pindex);
1445 pindex = (ptepindex - NUPT_TOTAL) & (NPDPEPG - 1);
1446 ptep = (void *)PHYS_TO_DMAP(pm->phys_addr);
1449 * Map PT into its PD
1451 pindex = ptepindex >> NPDPEPGSHIFT;
1452 pindex += NUPT_TOTAL;
1453 pm = _pmap_allocpte(pmap, pindex);
1454 pindex = ptepindex & (NPTEPG - 1);
1455 ptep = (void *)PHYS_TO_DMAP(pm->phys_addr);
1459 * Install the pte in (pm). (m) prevents races.
1462 data = atomic_swap_long(ptep, data);
1464 vm_page_wire_quick(pm);
1467 pmap->pm_ptphint = pm;
1473 * Determine the page table page required to access the VA in the pmap
1474 * and allocate it if necessary. Return a held vm_page_t for the page.
1476 * Only used with user pmaps.
1479 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1481 vm_pindex_t ptepindex;
1484 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1487 * Calculate pagetable page index, and return the PT page to
1490 ptepindex = pmap_pt_pindex(va);
1491 m = _pmap_allocpte(pmap, ptepindex);
1496 /***************************************************
1497 * Pmap allocation/deallocation routines.
1498 ***************************************************/
1501 * Release any resources held by the given physical map.
1502 * Called when a pmap initialized by pmap_pinit is being released.
1503 * Should only be called if the map contains no valid mappings.
1505 static int pmap_release_callback(struct vm_page *p, void *data);
1508 pmap_release(struct pmap *pmap)
1510 vm_object_t object = pmap->pm_pteobj;
1511 struct rb_vm_page_scan_info info;
1513 KKASSERT(pmap != &kernel_pmap);
1515 #if defined(DIAGNOSTIC)
1516 if (object->ref_count != 1)
1517 panic("pmap_release: pteobj reference count != 1");
1521 info.object = object;
1523 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
1524 ("pmap %p still active! %016jx",
1526 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
1528 vm_object_hold(object);
1532 info.limit = object->generation;
1534 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1535 pmap_release_callback, &info);
1536 if (info.error == 0 && info.mpte) {
1537 if (pmap_release_free_page(pmap, info.mpte))
1540 } while (info.error);
1542 pmap->pm_ptphint = NULL;
1544 KASSERT((pmap->pm_stats.wired_count == (pmap->pm_pdirm != NULL)),
1545 ("pmap_release: dangling count %p %ld",
1546 pmap, pmap->pm_stats.wired_count));
1548 vm_object_drop(object);
1552 pmap_release_callback(struct vm_page *p, void *data)
1554 struct rb_vm_page_scan_info *info = data;
1556 if (p->pindex == NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL) {
1560 if (pmap_release_free_page(info->pmap, p)) {
1564 if (info->object->generation != info->limit) {
1572 * Grow the number of kernel page table entries, if needed.
1574 * kernel_map must be locked exclusively by the caller.
1577 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1581 vm_offset_t ptppaddr;
1583 pd_entry_t *pde, newpdir;
1588 vm_object_hold(&kptobj);
1589 if (kernel_vm_end == 0) {
1590 kernel_vm_end = KvaStart;
1592 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & VPTE_V) != 0) {
1593 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1595 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1596 kernel_vm_end = kernel_map.max_offset;
1601 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1602 if (addr - 1 >= kernel_map.max_offset)
1603 addr = kernel_map.max_offset;
1604 while (kernel_vm_end < addr) {
1605 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1607 /* We need a new PDP entry */
1608 nkpg = vm_page_alloc(&kptobj, nkpt,
1611 VM_ALLOC_INTERRUPT);
1613 panic("pmap_growkernel: no memory to "
1616 paddr = VM_PAGE_TO_PHYS(nkpg);
1617 pmap_zero_page(paddr);
1618 newpdp = (pdp_entry_t)(paddr |
1619 VPTE_V | VPTE_RW | VPTE_U |
1620 VPTE_A | VPTE_M | VPTE_WIRED);
1621 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
1622 atomic_add_long(&kernel_pmap.pm_stats.wired_count, 1);
1624 continue; /* try again */
1626 if ((*pde & VPTE_V) != 0) {
1627 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1628 ~(PAGE_SIZE * NPTEPG - 1);
1629 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1630 kernel_vm_end = kernel_map.max_offset;
1637 * This index is bogus, but out of the way
1639 nkpg = vm_page_alloc(&kptobj, nkpt,
1642 VM_ALLOC_INTERRUPT);
1644 panic("pmap_growkernel: no memory to grow kernel");
1647 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1648 pmap_zero_page(ptppaddr);
1649 newpdir = (pd_entry_t)(ptppaddr |
1650 VPTE_V | VPTE_RW | VPTE_U |
1651 VPTE_A | VPTE_M | VPTE_WIRED);
1652 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1653 atomic_add_long(&kernel_pmap.pm_stats.wired_count, 1);
1656 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1657 ~(PAGE_SIZE * NPTEPG - 1);
1658 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1659 kernel_vm_end = kernel_map.max_offset;
1663 vm_object_drop(&kptobj);
1667 * Add a reference to the specified pmap.
1672 pmap_reference(pmap_t pmap)
1675 atomic_add_int(&pmap->pm_count, 1);
1678 /************************************************************************
1679 * VMSPACE MANAGEMENT *
1680 ************************************************************************
1682 * The VMSPACE management we do in our virtual kernel must be reflected
1683 * in the real kernel. This is accomplished by making vmspace system
1684 * calls to the real kernel.
1687 cpu_vmspace_alloc(struct vmspace *vm)
1694 * If VMM enable, don't do nothing, we
1695 * are able to use real page tables
1700 #define USER_SIZE (VM_MAX_USER_ADDRESS - VM_MIN_USER_ADDRESS)
1702 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
1703 panic("vmspace_create() failed");
1705 rp = vmspace_mmap(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1706 PROT_READ|PROT_WRITE,
1707 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
1709 if (rp == MAP_FAILED)
1710 panic("vmspace_mmap: failed");
1711 vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1713 vpte = VM_PAGE_TO_PHYS(vmspace_pmap(vm)->pm_pdirm) |
1714 VPTE_RW | VPTE_V | VPTE_U;
1715 r = vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1718 panic("vmspace_mcontrol: failed");
1722 cpu_vmspace_free(struct vmspace *vm)
1725 * If VMM enable, don't do nothing, we
1726 * are able to use real page tables
1731 if (vmspace_destroy(&vm->vm_pmap) < 0)
1732 panic("vmspace_destroy() failed");
1735 /***************************************************
1736 * page management routines.
1737 ***************************************************/
1740 * free the pv_entry back to the free list. This function may be
1741 * called from an interrupt.
1743 static __inline void
1744 free_pv_entry(pv_entry_t pv)
1746 atomic_add_int(&pv_entry_count, -1);
1747 KKASSERT(pv_entry_count >= 0);
1752 * get a new pv_entry, allocating a block from the system
1753 * when needed. This function may be called from an interrupt.
1758 atomic_add_int(&pv_entry_count, 1);
1759 if (pv_entry_high_water &&
1760 (pv_entry_count > pv_entry_high_water) &&
1761 atomic_swap_int(&pmap_pagedaemon_waken, 1) == 0) {
1762 wakeup(&vm_pages_needed);
1764 return zalloc(pvzone);
1768 * This routine is very drastic, but can save the system
1778 static int warningdone=0;
1780 if (pmap_pagedaemon_waken == 0)
1782 pmap_pagedaemon_waken = 0;
1784 if (warningdone < 5) {
1785 kprintf("pmap_collect: collecting pv entries -- "
1786 "suggest increasing PMAP_SHPGPERPROC\n");
1790 for (i = 0; i < vm_page_array_size; i++) {
1791 m = &vm_page_array[i];
1792 if (m->wire_count || m->hold_count)
1794 if (vm_page_busy_try(m, TRUE) == 0) {
1795 if (m->wire_count == 0 && m->hold_count == 0) {
1805 * If it is the first entry on the list, it is actually
1806 * in the header and we must copy the following entry up
1807 * to the header. Otherwise we must search the list for
1808 * the entry. In either case we free the now unused entry.
1810 * pmap->pm_pteobj must be held and (m) must be spin-locked by the caller.
1813 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1818 vm_page_spin_lock(m);
1819 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, va);
1822 * Note that pv_ptem is NULL if the page table page itself is not
1823 * managed, even if the page being removed IS managed.
1827 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1828 if (TAILQ_EMPTY(&m->md.pv_list))
1829 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1830 m->md.pv_list_count--;
1831 KKASSERT(m->md.pv_list_count >= 0);
1832 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
1833 atomic_add_int(&pmap->pm_generation, 1);
1834 vm_page_spin_unlock(m);
1835 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1838 vm_page_spin_unlock(m);
1839 kprintf("pmap_remove_entry: could not find "
1840 "pmap=%p m=%p va=%016jx\n",
1847 * Create a pv entry for page at pa for (pmap, va). If the page table page
1848 * holding the VA is managed, mpte will be non-NULL.
1850 * pmap->pm_pteobj must be held and (m) must be spin-locked by the caller.
1853 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m,
1860 m->md.pv_list_count++;
1861 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1862 pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pv);
1863 vm_page_flag_set(m, PG_MAPPED);
1864 KKASSERT(pv == NULL);
1868 * pmap_remove_pte: do the things to unmap a page in a process
1870 * Caller holds pmap->pm_pteobj and holds the associated page table
1871 * page busy to prevent races.
1874 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, pt_entry_t oldpte,
1881 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1883 if (oldpte & VPTE_WIRED)
1884 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1885 KKASSERT(pmap->pm_stats.wired_count >= 0);
1889 * Machines that don't support invlpg, also don't support
1890 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1894 cpu_invlpg((void *)va);
1896 KKASSERT(pmap->pm_stats.resident_count > 0);
1897 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1898 if (oldpte & VPTE_MANAGED) {
1899 m = PHYS_TO_VM_PAGE(oldpte);
1902 * NOTE: pmap_remove_entry() will spin-lock the page
1904 if (oldpte & VPTE_M) {
1905 #if defined(PMAP_DIAGNOSTIC)
1906 if (pmap_nw_modified(oldpte)) {
1907 kprintf("pmap_remove: modified page not "
1908 "writable: va: 0x%lx, pte: 0x%lx\n",
1912 if (pmap_track_modified(pmap, va))
1915 if (oldpte & VPTE_A)
1916 vm_page_flag_set(m, PG_REFERENCED);
1917 error = pmap_remove_entry(pmap, m, va);
1919 error = pmap_unuse_pt(pmap, va, NULL);
1927 * Remove a single page from a process address space.
1929 * This function may not be called from an interrupt if the pmap is
1932 * Caller holds pmap->pm_pteobj
1935 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
1939 pte = pmap_pte(pmap, va);
1942 if ((*pte & VPTE_V) == 0)
1944 pmap_remove_pte(pmap, pte, 0, va);
1948 * Remove the given range of addresses from the specified map.
1950 * It is assumed that the start and end are properly rounded to
1953 * This function may not be called from an interrupt if the pmap is
1959 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1961 vm_offset_t va_next;
1962 pml4_entry_t *pml4e;
1964 pd_entry_t ptpaddr, *pde;
1971 vm_object_hold(pmap->pm_pteobj);
1972 KKASSERT(pmap->pm_stats.resident_count >= 0);
1973 if (pmap->pm_stats.resident_count == 0) {
1974 vm_object_drop(pmap->pm_pteobj);
1979 * special handling of removing one page. a very
1980 * common operation and easy to short circuit some
1983 if (sva + PAGE_SIZE == eva) {
1984 pde = pmap_pde(pmap, sva);
1985 if (pde && (*pde & VPTE_PS) == 0) {
1986 pmap_remove_page(pmap, sva);
1987 vm_object_drop(pmap->pm_pteobj);
1992 for (; sva < eva; sva = va_next) {
1993 pml4e = pmap_pml4e(pmap, sva);
1994 if ((*pml4e & VPTE_V) == 0) {
1995 va_next = (sva + NBPML4) & ~PML4MASK;
2001 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2002 if ((*pdpe & VPTE_V) == 0) {
2003 va_next = (sva + NBPDP) & ~PDPMASK;
2010 * Calculate index for next page table.
2012 va_next = (sva + NBPDR) & ~PDRMASK;
2016 pde = pmap_pdpe_to_pde(pdpe, sva);
2020 * Weed out invalid mappings.
2026 * Check for large page.
2028 if ((ptpaddr & VPTE_PS) != 0) {
2029 /* JG FreeBSD has more complex treatment here */
2030 KKASSERT(*pde != 0);
2031 pmap_inval_pde(pde, pmap, sva);
2032 atomic_add_long(&pmap->pm_stats.resident_count,
2033 -NBPDR / PAGE_SIZE);
2038 * Limit our scan to either the end of the va represented
2039 * by the current page table page, or to the end of the
2040 * range being removed.
2046 * NOTE: pmap_remove_pte() can block.
2048 pt_m = pmap_hold_pt_page(pde, sva);
2049 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2052 if (pmap_remove_pte(pmap, pte, 0, sva))
2056 vm_page_unhold(pt_m);
2058 vm_object_drop(pmap->pm_pteobj);
2062 * Removes this physical page from all physical maps in which it resides.
2063 * Reflects back modify bits to the pager.
2065 * This routine may not be called from an interrupt.
2070 pmap_remove_all(vm_page_t m)
2072 pt_entry_t *pte, tpte;
2077 #if defined(PMAP_DIAGNOSTIC)
2079 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
2082 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
2083 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
2088 vm_page_spin_lock(m);
2089 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2091 pmobj = pmap->pm_pteobj;
2094 * Handle reversed lock ordering
2096 if (vm_object_hold_try(pmobj) == 0) {
2097 refcount_acquire(&pmobj->hold_count);
2098 vm_page_spin_unlock(m);
2099 vm_object_lock(pmobj);
2100 vm_page_spin_lock(m);
2101 if (pv != TAILQ_FIRST(&m->md.pv_list) ||
2102 pmap != pv->pv_pmap ||
2103 pmobj != pmap->pm_pteobj) {
2104 vm_page_spin_unlock(m);
2105 vm_object_drop(pmobj);
2110 KKASSERT(pmap->pm_stats.resident_count > 0);
2111 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2113 pte = pmap_pte(pmap, pv->pv_va);
2114 KKASSERT(pte != NULL);
2116 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2117 if (tpte & VPTE_WIRED)
2118 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2119 KKASSERT(pmap->pm_stats.wired_count >= 0);
2122 vm_page_flag_set(m, PG_REFERENCED);
2125 * Update the vm_page_t clean and reference bits.
2127 if (tpte & VPTE_M) {
2128 #if defined(PMAP_DIAGNOSTIC)
2129 if (pmap_nw_modified(tpte)) {
2131 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2135 if (pmap_track_modified(pmap, pv->pv_va))
2138 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2139 if (TAILQ_EMPTY(&m->md.pv_list))
2140 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2141 m->md.pv_list_count--;
2142 KKASSERT(m->md.pv_list_count >= 0);
2143 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2144 atomic_add_int(&pmap->pm_generation, 1);
2145 vm_page_spin_unlock(m);
2146 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2149 vm_object_drop(pmobj);
2150 vm_page_spin_lock(m);
2152 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2153 vm_page_spin_unlock(m);
2157 * Removes the page from a particular pmap
2160 pmap_remove_specific(pmap_t pmap, vm_page_t m)
2162 pt_entry_t *pte, tpte;
2165 vm_object_hold(pmap->pm_pteobj);
2167 vm_page_spin_lock(m);
2168 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2169 if (pv->pv_pmap != pmap)
2172 KKASSERT(pmap->pm_stats.resident_count > 0);
2173 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2175 pte = pmap_pte(pmap, pv->pv_va);
2176 KKASSERT(pte != NULL);
2178 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2179 if (tpte & VPTE_WIRED)
2180 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2181 KKASSERT(pmap->pm_stats.wired_count >= 0);
2184 vm_page_flag_set(m, PG_REFERENCED);
2187 * Update the vm_page_t clean and reference bits.
2189 if (tpte & VPTE_M) {
2190 if (pmap_track_modified(pmap, pv->pv_va))
2193 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2194 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2195 atomic_add_int(&pmap->pm_generation, 1);
2196 m->md.pv_list_count--;
2197 KKASSERT(m->md.pv_list_count >= 0);
2198 if (TAILQ_EMPTY(&m->md.pv_list))
2199 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2200 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2201 vm_page_spin_unlock(m);
2205 vm_page_spin_unlock(m);
2206 vm_object_drop(pmap->pm_pteobj);
2210 * Set the physical protection on the specified range of this map
2213 * This function may not be called from an interrupt if the map is
2214 * not the kernel_pmap.
2219 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2221 vm_offset_t va_next;
2222 pml4_entry_t *pml4e;
2224 pd_entry_t ptpaddr, *pde;
2231 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2232 pmap_remove(pmap, sva, eva);
2236 if (prot & VM_PROT_WRITE)
2239 vm_object_hold(pmap->pm_pteobj);
2241 for (; sva < eva; sva = va_next) {
2242 pml4e = pmap_pml4e(pmap, sva);
2243 if ((*pml4e & VPTE_V) == 0) {
2244 va_next = (sva + NBPML4) & ~PML4MASK;
2250 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2251 if ((*pdpe & VPTE_V) == 0) {
2252 va_next = (sva + NBPDP) & ~PDPMASK;
2258 va_next = (sva + NBPDR) & ~PDRMASK;
2262 pde = pmap_pdpe_to_pde(pdpe, sva);
2267 * Check for large page.
2269 if ((ptpaddr & VPTE_PS) != 0) {
2271 pmap_clean_pde(pde, pmap, sva);
2272 atomic_add_long(&pmap->pm_stats.resident_count,
2273 -NBPDR / PAGE_SIZE);
2279 * Weed out invalid mappings. Note: we assume that the page
2280 * directory table is always allocated, and in kernel virtual.
2288 pt_m = pmap_hold_pt_page(pde, sva);
2289 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2292 * Clean managed pages and also check the accessed
2293 * bit. Just remove write perms for unmanaged
2294 * pages. Be careful of races, turning off write
2295 * access will force a fault rather then setting
2296 * the modified bit at an unexpected time.
2298 pmap_clean_pte(pte, pmap, sva, NULL);
2300 vm_page_unhold(pt_m);
2302 vm_object_drop(pmap->pm_pteobj);
2306 * Enter a managed page into a pmap. If the page is not wired related pmap
2307 * data can be destroyed at any time for later demand-operation.
2309 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
2310 * specified protection, and wire the mapping if requested.
2312 * NOTE: This routine may not lazy-evaluate or lose information. The
2313 * page must actually be inserted into the given map NOW.
2315 * NOTE: When entering a page at a KVA address, the pmap must be the
2321 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2322 boolean_t wired, vm_map_entry_t entry __unused)
2327 pt_entry_t origpte, newpte;
2334 va = trunc_page(va);
2336 vm_object_hold(pmap->pm_pteobj);
2339 * Get the page table page. The kernel_pmap's page table pages
2340 * are preallocated and have no associated vm_page_t.
2342 * If not NULL, mpte will be busied and we must vm_page_wakeup()
2343 * to cleanup. There will already be at least one wire count from
2344 * it being mapped into its parent.
2346 if (pmap == &kernel_pmap) {
2350 mpte = pmap_allocpte(pmap, va);
2351 pte = (void *)PHYS_TO_DMAP(mpte->phys_addr);
2352 pte += pmap_pte_index(va);
2356 * Deal with races against the kernel's real MMU by cleaning the
2357 * page, even if we are re-entering the same page.
2359 pa = VM_PAGE_TO_PHYS(m);
2360 origpte = pmap_inval_loadandclear(pte, pmap, va);
2361 /*origpte = pmap_clean_pte(pte, pmap, va, NULL);*/
2362 opa = origpte & VPTE_FRAME;
2364 if (origpte & VPTE_PS)
2365 panic("pmap_enter: attempted pmap_enter on 2MB page");
2367 if ((origpte & (VPTE_MANAGED|VPTE_M)) == (VPTE_MANAGED|VPTE_M)) {
2368 if (pmap_track_modified(pmap, va)) {
2369 vm_page_t om = PHYS_TO_VM_PAGE(opa);
2375 * Mapping has not changed, must be protection or wiring change.
2377 if (origpte && (opa == pa)) {
2379 * Wiring change, just update stats. We don't worry about
2380 * wiring PT pages as they remain resident as long as there
2381 * are valid mappings in them. Hence, if a user page is wired,
2382 * the PT page will be also.
2384 if (wired && ((origpte & VPTE_WIRED) == 0))
2385 atomic_add_long(&pmap->pm_stats.wired_count, 1);
2386 else if (!wired && (origpte & VPTE_WIRED))
2387 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2389 if (origpte & VPTE_MANAGED) {
2391 KKASSERT(m->flags & PG_MAPPED);
2392 KKASSERT(!(m->flags & (PG_FICTITIOUS|PG_UNMANAGED)));
2394 KKASSERT((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)));
2396 vm_page_spin_lock(m);
2401 * Bump the wire_count for the page table page.
2404 vm_page_wire_quick(mpte);
2407 * Mapping has changed, invalidate old range and fall through to
2408 * handle validating new mapping. Don't inherit anything from
2413 err = pmap_remove_pte(pmap, NULL, origpte, va);
2416 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2420 * Enter on the PV list if part of our managed memory. Note that we
2421 * raise IPL while manipulating pv_table since pmap_enter can be
2422 * called at interrupt time.
2424 if (pmap_initialized) {
2425 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2427 * WARNING! We are using m's spin-lock as a
2428 * man's pte lock to interlock against
2429 * pmap_page_protect() operations.
2431 * This is a bad hack (obviously).
2433 pv = get_pv_entry();
2434 vm_page_spin_lock(m);
2435 pmap_insert_entry(pmap, va, mpte, m, pv);
2437 /* vm_page_spin_unlock(m); */
2439 vm_page_spin_lock(m);
2442 vm_page_spin_lock(m);
2446 * Increment counters
2448 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2450 atomic_add_long(&pmap->pm_stats.wired_count, 1);
2454 * Now validate mapping with desired protection/wiring.
2456 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | VPTE_V | VPTE_U);
2460 newpte |= VPTE_WIRED;
2461 // if (pmap != &kernel_pmap)
2463 if (newpte & VPTE_RW)
2464 vm_page_flag_set(m, PG_WRITEABLE);
2465 KKASSERT((newpte & VPTE_MANAGED) == 0 || (m->flags & PG_MAPPED));
2467 origpte = atomic_swap_long(pte, newpte);
2468 if (origpte & VPTE_M) {
2469 kprintf("pmap [M] race @ %016jx\n", va);
2470 atomic_set_long(pte, VPTE_M);
2472 vm_page_spin_unlock(m);
2475 vm_page_wakeup(mpte);
2476 vm_object_drop(pmap->pm_pteobj);
2480 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2482 * Currently this routine may only be used on user pmaps, not kernel_pmap.
2487 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2489 pmap_enter(pmap, va, m, VM_PROT_READ, 0, NULL);
2493 * Make a temporary mapping for a physical address. This is only intended
2494 * to be used for panic dumps.
2496 * The caller is responsible for calling smp_invltlb().
2499 pmap_kenter_temporary(vm_paddr_t pa, long i)
2501 pmap_kenter_quick(crashdumpmap + (i * PAGE_SIZE), pa);
2502 return ((void *)crashdumpmap);
2505 #define MAX_INIT_PT (96)
2508 * This routine preloads the ptes for a given object into the specified pmap.
2509 * This eliminates the blast of soft faults on process startup and
2510 * immediately after an mmap.
2514 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2517 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2518 vm_object_t object, vm_pindex_t pindex,
2519 vm_size_t size, int limit)
2521 struct rb_vm_page_scan_info info;
2526 * We can't preinit if read access isn't set or there is no pmap
2529 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2533 * We can't preinit if the pmap is not the current pmap
2535 lp = curthread->td_lwp;
2536 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2540 * Misc additional checks
2542 psize = x86_64_btop(size);
2544 if ((object->type != OBJT_VNODE) ||
2545 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2546 (object->resident_page_count > MAX_INIT_PT))) {
2550 if (psize + pindex > object->size) {
2551 if (object->size < pindex)
2553 psize = object->size - pindex;
2560 * Use a red-black scan to traverse the requested range and load
2561 * any valid pages found into the pmap.
2563 * We cannot safely scan the object's memq unless we are in a
2564 * critical section since interrupts can remove pages from objects.
2566 info.start_pindex = pindex;
2567 info.end_pindex = pindex + psize - 1;
2573 vm_object_hold_shared(object);
2574 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2575 pmap_object_init_pt_callback, &info);
2576 vm_object_drop(object);
2581 pmap_object_init_pt_callback(vm_page_t p, void *data)
2583 struct rb_vm_page_scan_info *info = data;
2584 vm_pindex_t rel_index;
2586 * don't allow an madvise to blow away our really
2587 * free pages allocating pv entries.
2589 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2590 vmstats.v_free_count < vmstats.v_free_reserved) {
2595 * Ignore list markers and ignore pages we cannot instantly
2596 * busy (while holding the object token).
2598 if (p->flags & PG_MARKER)
2600 if (vm_page_busy_try(p, TRUE))
2602 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2603 (p->flags & PG_FICTITIOUS) == 0) {
2604 if ((p->queue - p->pc) == PQ_CACHE)
2605 vm_page_deactivate(p);
2606 rel_index = p->pindex - info->start_pindex;
2607 pmap_enter_quick(info->pmap,
2608 info->addr + x86_64_ptob(rel_index), p);
2615 * Return TRUE if the pmap is in shape to trivially
2616 * pre-fault the specified address.
2618 * Returns FALSE if it would be non-trivial or if a
2619 * pte is already loaded into the slot.
2624 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2630 vm_object_hold(pmap->pm_pteobj);
2631 pde = pmap_pde(pmap, addr);
2632 if (pde == NULL || *pde == 0) {
2635 pte = pmap_pde_to_pte(pde, addr);
2636 ret = (*pte) ? 0 : 1;
2638 vm_object_drop(pmap->pm_pteobj);
2644 * Change the wiring attribute for a map/virtual-address pair.
2646 * The mapping must already exist in the pmap.
2647 * No other requirements.
2650 pmap_unwire(pmap_t pmap, vm_offset_t va)
2659 vm_object_hold(pmap->pm_pteobj);
2660 pte = pmap_pte(pmap, va);
2662 if (pte == NULL || (*pte & VPTE_V) == 0) {
2663 vm_object_drop(pmap->pm_pteobj);
2668 * Wiring is not a hardware characteristic so there is no need to
2669 * invalidate TLB. However, in an SMP environment we must use
2670 * a locked bus cycle to update the pte (if we are not using
2671 * the pmap_inval_*() API that is)... it's ok to do this for simple
2674 if (pmap_pte_w(pte))
2675 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2676 /* XXX else return NULL so caller doesn't unwire m ? */
2677 atomic_clear_long(pte, VPTE_WIRED);
2679 pa = *pte & VPTE_FRAME;
2680 m = PHYS_TO_VM_PAGE(pa); /* held by wired count */
2682 vm_object_drop(pmap->pm_pteobj);
2688 * Copy the range specified by src_addr/len
2689 * from the source map to the range dst_addr/len
2690 * in the destination map.
2692 * This routine is only advisory and need not do anything.
2695 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2696 vm_size_t len, vm_offset_t src_addr)
2699 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2700 * valid through blocking calls, and that's just not going to
2711 * Zero the specified physical page.
2713 * This function may be called from an interrupt and no locking is
2717 pmap_zero_page(vm_paddr_t phys)
2719 vm_offset_t va = PHYS_TO_DMAP(phys);
2721 bzero((void *)va, PAGE_SIZE);
2727 * Zero part of a physical page by mapping it into memory and clearing
2728 * its contents with bzero.
2730 * off and size may not cover an area beyond a single hardware page.
2733 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2735 vm_offset_t virt = PHYS_TO_DMAP(phys);
2737 bzero((char *)virt + off, size);
2743 * Copy the physical page from the source PA to the target PA.
2744 * This function may be called from an interrupt. No locking
2748 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2750 vm_offset_t src_virt, dst_virt;
2752 src_virt = PHYS_TO_DMAP(src);
2753 dst_virt = PHYS_TO_DMAP(dst);
2754 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
2758 * pmap_copy_page_frag:
2760 * Copy the physical page from the source PA to the target PA.
2761 * This function may be called from an interrupt. No locking
2765 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2767 vm_offset_t src_virt, dst_virt;
2769 src_virt = PHYS_TO_DMAP(src);
2770 dst_virt = PHYS_TO_DMAP(dst);
2771 bcopy((char *)src_virt + (src & PAGE_MASK),
2772 (char *)dst_virt + (dst & PAGE_MASK),
2777 * Returns true if the pmap's pv is one of the first 16 pvs linked to
2778 * from this page. This count may be changed upwards or downwards
2779 * in the future; it is only necessary that true be returned for a small
2780 * subset of pmaps for proper page aging.
2782 * No other requirements.
2785 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2790 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2793 vm_page_spin_lock(m);
2794 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2795 if (pv->pv_pmap == pmap) {
2796 vm_page_spin_unlock(m);
2803 vm_page_spin_unlock(m);
2809 * Remove all pages from specified address space this aids process
2810 * exit speeds. Also, this code is special cased for current
2811 * process only, but can have the more generic (and slightly slower)
2812 * mode enabled. This is much faster than pmap_remove in the case
2813 * of running down an entire address space.
2815 * No other requirements.
2818 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2820 pmap_remove(pmap, sva, eva);
2822 pt_entry_t *pte, tpte;
2825 int save_generation;
2827 if (pmap->pm_pteobj)
2828 vm_object_hold(pmap->pm_pteobj);
2830 pmap_invalidate_range(pmap, sva, eva);
2832 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2833 if (pv->pv_va >= eva || pv->pv_va < sva) {
2834 npv = TAILQ_NEXT(pv, pv_plist);
2838 KKASSERT(pmap == pv->pv_pmap);
2840 pte = pmap_pte(pmap, pv->pv_va);
2843 * We cannot remove wired pages from a process' mapping
2846 if (*pte & VPTE_WIRED) {
2847 npv = TAILQ_NEXT(pv, pv_plist);
2850 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2852 m = PHYS_TO_VM_PAGE(tpte & VPTE_FRAME);
2853 vm_page_spin_lock(m);
2855 KASSERT(m < &vm_page_array[vm_page_array_size],
2856 ("pmap_remove_pages: bad tpte %lx", tpte));
2858 KKASSERT(pmap->pm_stats.resident_count > 0);
2859 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2862 * Update the vm_page_t clean and reference bits.
2864 if (tpte & VPTE_M) {
2868 npv = TAILQ_NEXT(pv, pv_plist);
2869 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2870 atomic_add_int(&pmap->pm_generation, 1);
2871 save_generation = pmap->pm_generation;
2872 m->md.pv_list_count--;
2873 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2874 if (TAILQ_EMPTY(&m->md.pv_list))
2875 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2876 vm_page_spin_unlock(m);
2878 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2882 * Restart the scan if we blocked during the unuse or free
2883 * calls and other removals were made.
2885 if (save_generation != pmap->pm_generation) {
2886 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2887 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2890 if (pmap->pm_pteobj)
2891 vm_object_drop(pmap->pm_pteobj);
2892 pmap_remove(pmap, sva, eva);
2897 * pmap_testbit tests bits in active mappings of a VM page.
2900 pmap_testbit(vm_page_t m, int bit)
2905 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2908 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2911 vm_page_spin_lock(m);
2912 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2914 * if the bit being tested is the modified bit, then
2915 * mark clean_map and ptes as never
2918 if (bit & (VPTE_A|VPTE_M)) {
2919 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2923 #if defined(PMAP_DIAGNOSTIC)
2924 if (pv->pv_pmap == NULL) {
2925 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
2929 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2931 vm_page_spin_unlock(m);
2935 vm_page_spin_unlock(m);
2940 * This routine is used to clear bits in ptes. Certain bits require special
2941 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2943 * This routine is only called with certain VPTE_* bit combinations.
2945 static __inline void
2946 pmap_clearbit(vm_page_t m, int bit)
2954 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
2956 vm_page_flag_clear(m, PG_WRITEABLE);
2961 * Loop over all current mappings setting/clearing as appropos If
2962 * setting RO do we need to clear the VAC?
2965 vm_page_spin_lock(m);
2966 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2968 * Need the pmap object lock(?)
2971 pmobj = pmap->pm_pteobj;
2973 if (vm_object_hold_try(pmobj) == 0) {
2974 refcount_acquire(&pmobj->hold_count);
2975 vm_page_spin_unlock(m);
2976 vm_object_lock(pmobj);
2977 vm_object_drop(pmobj);
2982 * don't write protect pager mappings
2984 if (bit == VPTE_RW) {
2985 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va)) {
2986 vm_object_drop(pmobj);
2991 #if defined(PMAP_DIAGNOSTIC)
2992 if (pv->pv_pmap == NULL) {
2993 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
2994 vm_object_drop(pmobj);
3000 * Careful here. We can use a locked bus instruction to
3001 * clear VPTE_A or VPTE_M safely but we need to synchronize
3002 * with the target cpus when we mess with VPTE_RW.
3004 * On virtual kernels we must force a new fault-on-write
3005 * in the real kernel if we clear the Modify bit ourselves,
3006 * otherwise the real kernel will not get a new fault and
3007 * will never set our Modify bit again.
3009 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3011 if (bit == VPTE_RW) {
3013 * We must also clear VPTE_M when clearing
3014 * VPTE_RW and synchronize its state to
3017 pbits = pmap_clean_pte(pte, pv->pv_pmap,
3019 } else if (bit == VPTE_M) {
3021 * We must invalidate the real-kernel pte
3022 * when clearing VPTE_M bit to force the
3023 * real-kernel to take a new fault to re-set
3026 atomic_clear_long(pte, VPTE_M);
3027 if (*pte & VPTE_RW) {
3028 pmap_invalidate_range(pv->pv_pmap,
3030 pv->pv_va + PAGE_SIZE);
3032 } else if ((bit & (VPTE_RW|VPTE_M)) ==
3035 * We've been asked to clear W & M, I guess
3036 * the caller doesn't want us to update
3037 * the dirty status of the VM page.
3039 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va, m);
3040 panic("shouldn't be called");
3043 * We've been asked to clear bits that do
3044 * not interact with hardware.
3046 atomic_clear_long(pte, bit);
3049 vm_object_drop(pmobj);
3052 vm_page_flag_clear(m, PG_WRITEABLE);
3053 vm_page_spin_unlock(m);
3057 * Lower the permission for all mappings to a given page.
3059 * No other requirements.
3062 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3064 /* JG NX support? */
3065 if ((prot & VM_PROT_WRITE) == 0) {
3066 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3067 pmap_clearbit(m, VPTE_RW);
3075 pmap_phys_address(vm_pindex_t ppn)
3077 return (x86_64_ptob(ppn));
3081 * Return a count of reference bits for a page, clearing those bits.
3082 * It is not necessary for every reference bit to be cleared, but it
3083 * is necessary that 0 only be returned when there are truly no
3084 * reference bits set.
3086 * XXX: The exact number of bits to check and clear is a matter that
3087 * should be tested and standardized at some point in the future for
3088 * optimal aging of shared pages.
3090 * No other requirements.
3093 pmap_ts_referenced(vm_page_t m)
3095 pv_entry_t pv, pvf, pvn;
3099 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3102 vm_page_spin_lock(m);
3103 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3106 pvn = TAILQ_NEXT(pv, pv_list);
3107 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3108 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3110 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
3113 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3115 if (pte && (*pte & VPTE_A)) {
3116 atomic_clear_long(pte, VPTE_A);
3122 } while ((pv = pvn) != NULL && pv != pvf);
3124 vm_page_spin_unlock(m);
3130 * Return whether or not the specified physical page was modified
3131 * in any physical maps.
3133 * No other requirements.
3136 pmap_is_modified(vm_page_t m)
3140 res = pmap_testbit(m, VPTE_M);
3146 * Clear the modify bits on the specified physical page. For the vkernel
3147 * we really need to clean the page, which clears VPTE_RW and VPTE_M, in
3148 * order to ensure that we take a fault on the next write to the page.
3149 * Otherwise the page may become dirty without us knowing it.
3151 * No other requirements.
3154 pmap_clear_modify(vm_page_t m)
3156 pmap_clearbit(m, VPTE_RW);
3160 * Clear the reference bit on the specified physical page.
3162 * No other requirements.
3165 pmap_clear_reference(vm_page_t m)
3167 pmap_clearbit(m, VPTE_A);
3171 * Miscellaneous support routines follow
3175 i386_protection_init(void)
3179 kp = protection_codes;
3180 for (prot = 0; prot < 8; prot++) {
3181 if (prot & VM_PROT_READ)
3182 *kp |= 0; /* if it's VALID is readeable */
3183 if (prot & VM_PROT_WRITE)
3185 if (prot & VM_PROT_EXECUTE)
3186 *kp |= 0; /* if it's VALID is executable */
3192 * Sets the memory attribute for the specified page.
3195 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
3197 /* This is a vkernel, do nothing */
3201 * Change the PAT attribute on an existing kernel memory map. Caller
3202 * must ensure that the virtual memory in question is not accessed
3203 * during the adjustment.
3206 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
3208 /* This is a vkernel, do nothing */
3212 * Perform the pmap work for mincore
3214 * No other requirements.
3217 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3219 pt_entry_t *ptep, pte;
3223 vm_object_hold(pmap->pm_pteobj);
3224 ptep = pmap_pte(pmap, addr);
3226 if (ptep && (pte = *ptep) != 0) {
3229 val = MINCORE_INCORE;
3230 if ((pte & VPTE_MANAGED) == 0)
3233 pa = pte & VPTE_FRAME;
3235 m = PHYS_TO_VM_PAGE(pa);
3241 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3243 * Modified by someone
3245 else if (m->dirty || pmap_is_modified(m))
3246 val |= MINCORE_MODIFIED_OTHER;
3251 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3254 * Referenced by someone
3256 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3257 val |= MINCORE_REFERENCED_OTHER;
3258 vm_page_flag_set(m, PG_REFERENCED);
3262 vm_object_drop(pmap->pm_pteobj);
3268 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3269 * vmspace will be ref'd and the old one will be deref'd.
3271 * Caller must hold vmspace->vm_map.token for oldvm and newvm
3274 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3276 struct vmspace *oldvm;
3279 oldvm = p->p_vmspace;
3280 if (oldvm != newvm) {
3283 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
3284 p->p_vmspace = newvm;
3285 KKASSERT(p->p_nthreads == 1);
3286 lp = RB_ROOT(&p->p_lwp_tree);
3287 pmap_setlwpvm(lp, newvm);
3294 * Set the vmspace for a LWP. The vmspace is almost universally set the
3295 * same as the process vmspace, but virtual kernels need to swap out contexts
3296 * on a per-lwp basis.
3299 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3301 struct vmspace *oldvm;
3304 oldvm = lp->lwp_vmspace;
3305 if (oldvm != newvm) {
3307 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
3308 lp->lwp_vmspace = newvm;
3309 if (curthread->td_lwp == lp) {
3310 pmap = vmspace_pmap(newvm);
3311 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
3312 if (pmap->pm_active_lock & CPULOCK_EXCL)
3313 pmap_interlock_wait(newvm);
3314 #if defined(SWTCH_OPTIM_STATS)
3317 pmap = vmspace_pmap(oldvm);
3318 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
3326 * The swtch code tried to switch in a heavy weight process whos pmap
3327 * is locked by another cpu. We have to wait for the lock to clear before
3328 * the pmap can be used.
3331 pmap_interlock_wait (struct vmspace *vm)
3333 pmap_t pmap = vmspace_pmap(vm);
3335 if (pmap->pm_active_lock & CPULOCK_EXCL) {
3337 while (pmap->pm_active_lock & CPULOCK_EXCL) {
3346 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3349 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3353 addr = roundup2(addr, NBPDR);
3358 * Used by kmalloc/kfree, page already exists at va
3361 pmap_kvtom(vm_offset_t va)
3365 KKASSERT(va >= KvaStart && va < KvaEnd);
3367 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
3371 pmap_object_init(vm_object_t object)
3377 pmap_object_free(vm_object_t object)
3383 pmap_pgscan(struct pmap_pgscan_info *pginfo)
3385 pmap_t pmap = pginfo->pmap;
3386 vm_offset_t sva = pginfo->beg_addr;
3387 vm_offset_t eva = pginfo->end_addr;
3388 vm_offset_t va_next;
3389 pml4_entry_t *pml4e;
3391 pd_entry_t ptpaddr, *pde;
3396 vm_object_hold(pmap->pm_pteobj);
3398 for (; sva < eva; sva = va_next) {
3402 pml4e = pmap_pml4e(pmap, sva);
3403 if ((*pml4e & VPTE_V) == 0) {
3404 va_next = (sva + NBPML4) & ~PML4MASK;
3410 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
3411 if ((*pdpe & VPTE_V) == 0) {
3412 va_next = (sva + NBPDP) & ~PDPMASK;
3418 va_next = (sva + NBPDR) & ~PDRMASK;
3422 pde = pmap_pdpe_to_pde(pdpe, sva);
3427 * Check for large page (ignore).
3429 if ((ptpaddr & VPTE_PS) != 0) {
3431 pmap_clean_pde(pde, pmap, sva);
3432 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
3439 * Weed out invalid mappings. Note: we assume that the page
3440 * directory table is always allocated, and in kernel virtual.
3448 pt_m = pmap_hold_pt_page(pde, sva);
3449 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
3455 if ((*pte & VPTE_MANAGED) == 0)
3458 m = PHYS_TO_VM_PAGE(*pte & VPTE_FRAME);
3459 if (vm_page_busy_try(m, TRUE) == 0) {
3460 if (pginfo->callback(pginfo, sva, m) < 0)
3464 vm_page_unhold(pt_m);
3466 vm_object_drop(pmap->pm_pteobj);