2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
8 * Copyright (c) 2005 Alan L. Cox <alc@cs.rice.edu>
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
46 * Copyright (c) 2003 Networks Associates Technology, Inc.
47 * All rights reserved.
49 * This software was developed for the FreeBSD Project by Jake Burkholder,
50 * Safeport Network Services, and Network Associates Laboratories, the
51 * Security Research Division of Network Associates, Inc. under
52 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA
53 * CHATS research program.
55 * Redistribution and use in source and binary forms, with or without
56 * modification, are permitted provided that the following conditions
58 * 1. Redistributions of source code must retain the above copyright
59 * notice, this list of conditions and the following disclaimer.
60 * 2. Redistributions in binary form must reproduce the above copyright
61 * notice, this list of conditions and the following disclaimer in the
62 * documentation and/or other materials provided with the distribution.
64 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
65 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
66 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
67 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
68 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
69 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
70 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
71 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
72 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
73 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
77 #include <sys/cdefs.h>
78 __FBSDID("$FreeBSD$");
81 * Manages physical address maps.
83 * In addition to hardware address maps, this
84 * module is called upon to provide software-use-only
85 * maps which may or may not be stored in the same
86 * form as hardware maps. These pseudo-maps are
87 * used to store intermediate results from copy
88 * operations to and from address spaces.
90 * Since the information managed by this module is
91 * also stored by the logical address mapping module,
92 * this module may throw away valid virtual-to-physical
93 * mappings at almost any time. However, invalidations
94 * of virtual-to-physical mappings must be done as
97 * In order to cope with hardware architectures which
98 * make virtual-to-physical map invalidates expensive,
99 * this module may delay invalidate or reduced protection
100 * operations until such time as they are actually
101 * necessary. This module is given full information as
102 * to which processors are currently using which maps,
103 * and to when physical maps must be made correct.
106 #define PMAP_DIAGNOSTIC
109 #include "opt_pmap.h"
110 #include "opt_msgbuf.h"
112 #include "opt_xbox.h"
114 #include <sys/param.h>
115 #include <sys/systm.h>
116 #include <sys/kernel.h>
118 #include <sys/lock.h>
119 #include <sys/malloc.h>
120 #include <sys/mman.h>
121 #include <sys/msgbuf.h>
122 #include <sys/mutex.h>
123 #include <sys/proc.h>
125 #include <sys/vmmeter.h>
126 #include <sys/sched.h>
127 #include <sys/sysctl.h>
133 #include <vm/vm_param.h>
134 #include <vm/vm_kern.h>
135 #include <vm/vm_page.h>
136 #include <vm/vm_map.h>
137 #include <vm/vm_object.h>
138 #include <vm/vm_extern.h>
139 #include <vm/vm_pageout.h>
140 #include <vm/vm_pager.h>
143 #include <machine/cpu.h>
144 #include <machine/cputypes.h>
145 #include <machine/md_var.h>
146 #include <machine/pcb.h>
147 #include <machine/specialreg.h>
149 #include <machine/smp.h>
153 #include <machine/xbox.h>
156 #include <xen/interface/xen.h>
157 #include <machine/xen/hypervisor.h>
158 #include <machine/xen/hypercall.h>
159 #include <machine/xen/xenvar.h>
160 #include <machine/xen/xenfunc.h>
162 #if !defined(CPU_DISABLE_SSE) && defined(I686_CPU)
163 #define CPU_ENABLE_SSE
166 #ifndef PMAP_SHPGPERPROC
167 #define PMAP_SHPGPERPROC 200
170 #if defined(DIAGNOSTIC)
171 #define PMAP_DIAGNOSTIC
174 #if !defined(PMAP_DIAGNOSTIC)
175 #define PMAP_INLINE __gnu89_inline
182 #define PV_STAT(x) do { x ; } while (0)
184 #define PV_STAT(x) do { } while (0)
187 #define pa_index(pa) ((pa) >> PDRSHIFT)
188 #define pa_to_pvh(pa) (&pv_table[pa_index(pa)])
191 * Get PDEs and PTEs for user/kernel address space
193 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
194 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
196 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
197 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
198 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
199 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
200 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
202 #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v)))
204 struct pmap kernel_pmap_store;
205 LIST_HEAD(pmaplist, pmap);
206 static struct pmaplist allpmaps;
207 static struct mtx allpmaps_lock;
209 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
210 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
211 int pgeflag = 0; /* PG_G or-in */
212 int pseflag = 0; /* PG_PS or-in */
215 vm_offset_t kernel_vm_end;
216 extern u_int32_t KERNend;
221 static uma_zone_t pdptzone;
226 * Data for the pv entry allocation mechanism
228 static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0;
229 static struct md_page *pv_table;
230 static int shpgperproc = PMAP_SHPGPERPROC;
232 struct pv_chunk *pv_chunkbase; /* KVA block for pv_chunks */
233 int pv_maxchunks; /* How many chunks we have KVA for */
234 vm_offset_t pv_vafree; /* freelist stored in the PTE */
237 * All those kernel PT submaps that BSD is so fond of
246 static struct sysmaps sysmaps_pcpu[MAXCPU];
247 pt_entry_t *CMAP1 = 0;
248 static pt_entry_t *CMAP3;
249 caddr_t CADDR1 = 0, ptvmmap = 0;
250 static caddr_t CADDR3;
251 struct msgbuf *msgbufp = 0;
256 static caddr_t crashdumpmap;
258 static pt_entry_t *PMAP1 = 0, *PMAP2;
259 static pt_entry_t *PADDR1 = 0, *PADDR2;
262 static int PMAP1changedcpu;
263 SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD,
265 "Number of times pmap_pte_quick changed CPU with same PMAP1");
267 static int PMAP1changed;
268 SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD,
270 "Number of times pmap_pte_quick changed PMAP1");
271 static int PMAP1unchanged;
272 SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD,
274 "Number of times pmap_pte_quick didn't change PMAP1");
275 static struct mtx PMAP2mutex;
277 SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters");
278 static int pg_ps_enabled;
279 SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RD, &pg_ps_enabled, 0,
280 "Are large page mappings enabled?");
282 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_max, CTLFLAG_RD, &pv_entry_max, 0,
283 "Max number of PV entries");
284 SYSCTL_INT(_vm_pmap, OID_AUTO, shpgperproc, CTLFLAG_RD, &shpgperproc, 0,
285 "Page share factor per proc");
287 static void free_pv_entry(pmap_t pmap, pv_entry_t pv);
288 static pv_entry_t get_pv_entry(pmap_t locked_pmap, int try);
290 static vm_page_t pmap_enter_quick_locked(multicall_entry_t **mcl, int *count, pmap_t pmap, vm_offset_t va,
291 vm_page_t m, vm_prot_t prot, vm_page_t mpte);
292 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva,
294 static void pmap_remove_page(struct pmap *pmap, vm_offset_t va,
296 static void pmap_remove_entry(struct pmap *pmap, vm_page_t m,
298 static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m);
299 static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va,
302 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags);
304 static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags);
305 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, vm_page_t *free);
306 static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va);
307 static void pmap_pte_release(pt_entry_t *pte);
308 static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t *);
309 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
310 static boolean_t pmap_is_prefaultable_locked(pmap_t pmap, vm_offset_t addr);
311 static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode);
314 #if defined(PAE) && !defined(XEN)
315 static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait);
318 CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t));
319 CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t));
322 * If you get an error here, then you set KVA_PAGES wrong! See the
323 * description of KVA_PAGES in sys/i386/include/pmap.h. It must be
324 * multiple of 4 for a normal kernel, or a multiple of 8 for a PAE.
326 CTASSERT(KERNBASE % (1 << 24) == 0);
333 #if defined(I686_CPU)
334 if (cpu_class == CPUCLASS_686) {
335 #if defined(CPU_ENABLE_SSE)
336 if (cpu_feature & CPUID_SSE2)
343 bzero(page, PAGE_SIZE);
347 pd_set(struct pmap *pmap, int ptepindex, vm_paddr_t val, int type)
349 vm_paddr_t pdir_ma = vtomach(&pmap->pm_pdir[ptepindex]);
354 xen_queue_pt_update(shadow_pdir_ma,
355 xpmap_ptom(val & ~(PG_RW)));
357 xen_queue_pt_update(pdir_ma,
360 case SH_PD_SET_VA_MA:
362 xen_queue_pt_update(shadow_pdir_ma,
365 xen_queue_pt_update(pdir_ma, val);
367 case SH_PD_SET_VA_CLEAR:
369 xen_queue_pt_update(shadow_pdir_ma, 0);
371 xen_queue_pt_update(pdir_ma, 0);
377 * Move the kernel virtual free pointer to the next
378 * 4MB. This is used to help improve performance
379 * by using a large (4MB) page for much of the kernel
380 * (.text, .data, .bss)
383 pmap_kmem_choose(vm_offset_t addr)
385 vm_offset_t newaddr = addr;
388 if (cpu_feature & CPUID_PSE)
389 newaddr = (addr + PDRMASK) & ~PDRMASK;
395 * Bootstrap the system enough to run with virtual memory.
397 * On the i386 this is called after mapping has already been enabled
398 * and just syncs the pmap module with what has already been done.
399 * [We can't call it easily with mapping off since the kernel is not
400 * mapped with PA == VA, hence we would have to relocate every address
401 * from the linked base (virtual) address "KERNBASE" to the actual
402 * (physical) address starting relative to 0]
405 pmap_bootstrap(vm_paddr_t firstaddr)
408 pt_entry_t *pte, *unused;
409 struct sysmaps *sysmaps;
413 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too
414 * large. It should instead be correctly calculated in locore.s and
415 * not based on 'first' (which is a physical address, not a virtual
416 * address, for the start of unused physical memory). The kernel
417 * page tables are NOT double mapped and thus should not be included
418 * in this calculation.
420 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
421 virtual_avail = pmap_kmem_choose(virtual_avail);
423 virtual_end = VM_MAX_KERNEL_ADDRESS;
426 * Initialize the kernel pmap (which is statically allocated).
428 PMAP_LOCK_INIT(kernel_pmap);
429 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD);
431 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT);
433 kernel_pmap->pm_active = -1; /* don't allow deactivation */
434 TAILQ_INIT(&kernel_pmap->pm_pvchunk);
435 LIST_INIT(&allpmaps);
436 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN);
437 mtx_lock_spin(&allpmaps_lock);
438 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list);
439 mtx_unlock_spin(&allpmaps_lock);
443 * Reserve some special page table entries/VA space for temporary
446 #define SYSMAP(c, p, v, n) \
447 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
453 * CMAP1/CMAP2 are used for zeroing and copying pages.
454 * CMAP3 is used for the idle process page zeroing.
456 for (i = 0; i < MAXCPU; i++) {
457 sysmaps = &sysmaps_pcpu[i];
458 mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF);
459 SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1)
460 SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1)
462 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
463 SYSMAP(caddr_t, CMAP3, CADDR3, 1)
464 PT_SET_MA(CADDR3, 0);
469 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS)
472 * ptvmmap is used for reading arbitrary physical pages via /dev/mem.
474 SYSMAP(caddr_t, unused, ptvmmap, 1)
477 * msgbufp is used to map the system message buffer.
479 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE)))
482 * ptemap is used for pmap_pte_quick
484 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1);
485 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1);
487 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF);
490 PT_SET_MA(CADDR1, 0);
493 * Leave in place an identity mapping (virt == phys) for the low 1 MB
494 * physical memory region that is used by the ACPI wakeup code. This
495 * mapping must not have PG_G set.
499 * leave here deliberately to show that this is not supported
502 /* FIXME: This is gross, but needed for the XBOX. Since we are in such
503 * an early stadium, we cannot yet neatly map video memory ... :-(
504 * Better fixes are very welcome! */
505 if (!arch_i386_is_xbox)
507 for (i = 1; i < NKPT; i++)
510 /* Initialize the PAT MSR if present. */
513 /* Turn on PG_G on kernel page(s) */
526 /* Bail if this CPU doesn't implement PAT. */
527 if (!(cpu_feature & CPUID_PAT))
532 * Leave the indices 0-3 at the default of WB, WT, UC, and UC-.
533 * Program 4 and 5 as WP and WC.
534 * Leave 6 and 7 as UC and UC-.
536 pat_msr = rdmsr(MSR_PAT);
537 pat_msr &= ~(PAT_MASK(4) | PAT_MASK(5));
538 pat_msr |= PAT_VALUE(4, PAT_WRITE_PROTECTED) |
539 PAT_VALUE(5, PAT_WRITE_COMBINING);
542 * Due to some Intel errata, we can only safely use the lower 4
543 * PAT entries. Thus, just replace PAT Index 2 with WC instead
546 * Intel Pentium III Processor Specification Update
547 * Errata E.27 (Upper Four PAT Entries Not Usable With Mode B
550 * Intel Pentium IV Processor Specification Update
551 * Errata N46 (PAT Index MSB May Be Calculated Incorrectly)
553 pat_msr = rdmsr(MSR_PAT);
554 pat_msr &= ~PAT_MASK(2);
555 pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING);
557 wrmsr(MSR_PAT, pat_msr);
561 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on.
568 vm_offset_t va, endva;
575 endva = KERNBASE + KERNend;
578 va = KERNBASE + KERNLOAD;
580 pdir = kernel_pmap->pm_pdir[KPTDI+i];
582 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir;
583 invltlb(); /* Play it safe, invltlb() every time */
588 va = (vm_offset_t)btext;
593 invltlb(); /* Play it safe, invltlb() every time */
600 * Initialize a vm_page's machine-dependent fields.
603 pmap_page_init(vm_page_t m)
606 TAILQ_INIT(&m->md.pv_list);
609 #if defined(PAE) && !defined(XEN)
611 static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt");
614 pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
616 *flags = UMA_SLAB_PRIV;
617 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL,
623 * ABuse the pte nodes for unmapped kva to thread a kva freelist through.
625 * - Must deal with pages in order to ensure that none of the PG_* bits
626 * are ever set, PG_V in particular.
627 * - Assumes we can write to ptes without pte_store() atomic ops, even
628 * on PAE systems. This should be ok.
629 * - Assumes nothing will ever test these addresses for 0 to indicate
630 * no mapping instead of correctly checking PG_V.
631 * - Assumes a vm_offset_t will fit in a pte (true for i386).
632 * Because PG_V is never set, there can be no mappings to invalidate.
634 static int ptelist_count = 0;
636 pmap_ptelist_alloc(vm_offset_t *head)
639 vm_offset_t *phead = (vm_offset_t *)*head;
641 if (ptelist_count == 0) {
642 printf("out of memory!!!!!!\n");
643 return (0); /* Out of memory */
646 va = phead[ptelist_count];
651 pmap_ptelist_free(vm_offset_t *head, vm_offset_t va)
653 vm_offset_t *phead = (vm_offset_t *)*head;
655 phead[ptelist_count++] = va;
659 pmap_ptelist_init(vm_offset_t *head, void *base, int npages)
665 nstackpages = (npages + PAGE_SIZE/sizeof(vm_offset_t) - 1)/ (PAGE_SIZE/sizeof(vm_offset_t));
666 for (i = 0; i < nstackpages; i++) {
667 va = (vm_offset_t)base + i * PAGE_SIZE;
668 m = vm_page_alloc(NULL, i,
669 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
671 pmap_qenter(va, &m, 1);
674 *head = (vm_offset_t)base;
675 for (i = npages - 1; i >= nstackpages; i--) {
676 va = (vm_offset_t)base + i * PAGE_SIZE;
677 pmap_ptelist_free(head, va);
683 * Initialize the pmap module.
684 * Called by vm_init, to initialize any structures that the pmap
685 * system needs to map virtual memory.
695 * Initialize the vm page array entries for the kernel pmap's
698 for (i = 0; i < nkpt; i++) {
699 mpte = PHYS_TO_VM_PAGE(xpmap_mtop(PTD[i + KPTDI] & PG_FRAME));
700 KASSERT(mpte >= vm_page_array &&
701 mpte < &vm_page_array[vm_page_array_size],
702 ("pmap_init: page table page is out of range"));
703 mpte->pindex = i + KPTDI;
704 mpte->phys_addr = xpmap_mtop(PTD[i + KPTDI] & PG_FRAME);
708 * Initialize the address space (zone) for the pv entries. Set a
709 * high water mark so that the system can recover from excessive
710 * numbers of pv entries.
712 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
713 pv_entry_max = shpgperproc * maxproc + cnt.v_page_count;
714 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
715 pv_entry_max = roundup(pv_entry_max, _NPCPV);
716 pv_entry_high_water = 9 * (pv_entry_max / 10);
719 * Are large page mappings enabled?
721 TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled);
724 * Calculate the size of the pv head table for superpages.
726 for (i = 0; phys_avail[i + 1]; i += 2);
727 pv_npg = round_4mpage(phys_avail[(i - 2) + 1]) / NBPDR;
730 * Allocate memory for the pv head table for superpages.
732 s = (vm_size_t)(pv_npg * sizeof(struct md_page));
734 pv_table = (struct md_page *)kmem_alloc(kernel_map, s);
735 for (i = 0; i < pv_npg; i++)
736 TAILQ_INIT(&pv_table[i].pv_list);
738 pv_maxchunks = MAX(pv_entry_max / _NPCPV, maxproc);
739 pv_chunkbase = (struct pv_chunk *)kmem_alloc_nofault(kernel_map,
740 PAGE_SIZE * pv_maxchunks);
741 if (pv_chunkbase == NULL)
742 panic("pmap_init: not enough kvm for pv chunks");
743 pmap_ptelist_init(&pv_vafree, pv_chunkbase, pv_maxchunks);
744 #if defined(PAE) && !defined(XEN)
745 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL,
746 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1,
747 UMA_ZONE_VM | UMA_ZONE_NOFREE);
748 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf);
753 /***************************************************
754 * Low level helper routines.....
755 ***************************************************/
758 * Determine the appropriate bits to set in a PTE or PDE for a specified
762 pmap_cache_bits(int mode, boolean_t is_pde)
764 int pat_flag, pat_index, cache_bits;
766 /* The PAT bit is different for PTE's and PDE's. */
767 pat_flag = is_pde ? PG_PDE_PAT : PG_PTE_PAT;
769 /* If we don't support PAT, map extended modes to older ones. */
770 if (!(cpu_feature & CPUID_PAT)) {
772 case PAT_UNCACHEABLE:
773 case PAT_WRITE_THROUGH:
777 case PAT_WRITE_COMBINING:
778 case PAT_WRITE_PROTECTED:
779 mode = PAT_UNCACHEABLE;
784 /* Map the caching mode to a PAT index. */
787 case PAT_UNCACHEABLE:
790 case PAT_WRITE_THROUGH:
799 case PAT_WRITE_COMBINING:
802 case PAT_WRITE_PROTECTED:
807 case PAT_UNCACHEABLE:
808 case PAT_WRITE_PROTECTED:
811 case PAT_WRITE_THROUGH:
817 case PAT_WRITE_COMBINING:
822 panic("Unknown caching mode %d\n", mode);
825 /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */
828 cache_bits |= pat_flag;
830 cache_bits |= PG_NC_PCD;
832 cache_bits |= PG_NC_PWT;
837 * For SMP, these functions have to use the IPI mechanism for coherence.
839 * N.B.: Before calling any of the following TLB invalidation functions,
840 * the calling processor must ensure that all stores updating a non-
841 * kernel page table are globally performed. Otherwise, another
842 * processor could cache an old, pre-update entry without being
843 * invalidated. This can happen one of two ways: (1) The pmap becomes
844 * active on another processor after its pm_active field is checked by
845 * one of the following functions but before a store updating the page
846 * table is globally performed. (2) The pmap becomes active on another
847 * processor before its pm_active field is checked but due to
848 * speculative loads one of the following functions stills reads the
849 * pmap as inactive on the other processor.
851 * The kernel page table is exempt because its pm_active field is
852 * immutable. The kernel page table is always active on every
856 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
861 CTR2(KTR_PMAP, "pmap_invalidate_page: pmap=%p va=0x%x",
865 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
869 cpumask = PCPU_GET(cpumask);
870 other_cpus = PCPU_GET(other_cpus);
871 if (pmap->pm_active & cpumask)
873 if (pmap->pm_active & other_cpus)
874 smp_masked_invlpg(pmap->pm_active & other_cpus, va);
881 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
887 CTR3(KTR_PMAP, "pmap_invalidate_page: pmap=%p eva=0x%x sva=0x%x",
891 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
892 for (addr = sva; addr < eva; addr += PAGE_SIZE)
894 smp_invlpg_range(sva, eva);
896 cpumask = PCPU_GET(cpumask);
897 other_cpus = PCPU_GET(other_cpus);
898 if (pmap->pm_active & cpumask)
899 for (addr = sva; addr < eva; addr += PAGE_SIZE)
901 if (pmap->pm_active & other_cpus)
902 smp_masked_invlpg_range(pmap->pm_active & other_cpus,
910 pmap_invalidate_all(pmap_t pmap)
915 CTR1(KTR_PMAP, "pmap_invalidate_page: pmap=%p", pmap);
918 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
922 cpumask = PCPU_GET(cpumask);
923 other_cpus = PCPU_GET(other_cpus);
924 if (pmap->pm_active & cpumask)
926 if (pmap->pm_active & other_cpus)
927 smp_masked_invltlb(pmap->pm_active & other_cpus);
933 pmap_invalidate_cache(void)
943 * Normal, non-SMP, 486+ invalidation functions.
944 * We inline these within pmap.c for speed.
947 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
949 CTR2(KTR_PMAP, "pmap_invalidate_page: pmap=%p va=0x%x",
952 if (pmap == kernel_pmap || pmap->pm_active)
958 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
962 if (eva - sva > PAGE_SIZE)
963 CTR3(KTR_PMAP, "pmap_invalidate_range: pmap=%p sva=0x%x eva=0x%x",
966 if (pmap == kernel_pmap || pmap->pm_active)
967 for (addr = sva; addr < eva; addr += PAGE_SIZE)
973 pmap_invalidate_all(pmap_t pmap)
976 CTR1(KTR_PMAP, "pmap_invalidate_all: pmap=%p", pmap);
978 if (pmap == kernel_pmap || pmap->pm_active)
983 pmap_invalidate_cache(void)
991 * Are we current address space or kernel? N.B. We return FALSE when
992 * a pmap's page table is in use because a kernel thread is borrowing
993 * it. The borrowed page table can change spontaneously, making any
994 * dependence on its continued use subject to a race condition.
997 pmap_is_current(pmap_t pmap)
1000 return (pmap == kernel_pmap ||
1001 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) &&
1002 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)));
1006 * If the given pmap is not the current or kernel pmap, the returned pte must
1007 * be released by passing it to pmap_pte_release().
1010 pmap_pte(pmap_t pmap, vm_offset_t va)
1015 pde = pmap_pde(pmap, va);
1019 /* are we current address space or kernel? */
1020 if (pmap_is_current(pmap))
1021 return (vtopte(va));
1022 mtx_lock(&PMAP2mutex);
1023 newpf = *pde & PG_FRAME;
1024 if ((*PMAP2 & PG_FRAME) != newpf) {
1025 PT_SET_MA(PADDR2, newpf | PG_V | PG_A | PG_M);
1026 CTR3(KTR_PMAP, "pmap_pte: pmap=%p va=0x%x newpte=0x%08x",
1027 pmap, va, (*PMAP2 & 0xffffffff));
1030 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1)));
1036 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte
1039 static __inline void
1040 pmap_pte_release(pt_entry_t *pte)
1043 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) {
1044 CTR1(KTR_PMAP, "pmap_pte_release: pte=0x%jx",
1046 PT_SET_VA(PMAP2, 0, TRUE);
1047 mtx_unlock(&PMAP2mutex);
1051 static __inline void
1052 invlcaddr(void *caddr)
1055 invlpg((u_int)caddr);
1060 * Super fast pmap_pte routine best used when scanning
1061 * the pv lists. This eliminates many coarse-grained
1062 * invltlb calls. Note that many of the pv list
1063 * scans are across different pmaps. It is very wasteful
1064 * to do an entire invltlb for checking a single mapping.
1066 * If the given pmap is not the current pmap, vm_page_queue_mtx
1067 * must be held and curthread pinned to a CPU.
1070 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
1075 pde = pmap_pde(pmap, va);
1079 /* are we current address space or kernel? */
1080 if (pmap_is_current(pmap))
1081 return (vtopte(va));
1082 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1083 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
1084 newpf = *pde & PG_FRAME;
1085 if ((*PMAP1 & PG_FRAME) != newpf) {
1086 PT_SET_MA(PADDR1, newpf | PG_V | PG_A | PG_M);
1087 CTR3(KTR_PMAP, "pmap_pte_quick: pmap=%p va=0x%x newpte=0x%08x",
1088 pmap, va, (u_long)*PMAP1);
1091 PMAP1cpu = PCPU_GET(cpuid);
1096 if (PMAP1cpu != PCPU_GET(cpuid)) {
1097 PMAP1cpu = PCPU_GET(cpuid);
1103 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1)));
1109 * Routine: pmap_extract
1111 * Extract the physical page address associated
1112 * with the given map/virtual_address pair.
1115 pmap_extract(pmap_t pmap, vm_offset_t va)
1124 pde = pmap->pm_pdir[va >> PDRSHIFT];
1126 if ((pde & PG_PS) != 0) {
1127 rtval = xpmap_mtop(pde & PG_PS_FRAME) | (va & PDRMASK);
1131 pte = pmap_pte(pmap, va);
1132 pteval = *pte ? xpmap_mtop(*pte) : 0;
1133 rtval = (pteval & PG_FRAME) | (va & PAGE_MASK);
1134 pmap_pte_release(pte);
1141 * Routine: pmap_extract_ma
1143 * Like pmap_extract, but returns machine address
1146 pmap_extract_ma(pmap_t pmap, vm_offset_t va)
1154 pde = pmap->pm_pdir[va >> PDRSHIFT];
1156 if ((pde & PG_PS) != 0) {
1157 rtval = (pde & ~PDRMASK) | (va & PDRMASK);
1161 pte = pmap_pte(pmap, va);
1162 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
1163 pmap_pte_release(pte);
1170 * Routine: pmap_extract_and_hold
1172 * Atomically extract and hold the physical page
1173 * with the given pmap and virtual address pair
1174 * if that mapping permits the given protection.
1177 pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1184 vm_page_lock_queues();
1186 pde = PT_GET(pmap_pde(pmap, va));
1189 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) {
1190 m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) |
1196 pte = PT_GET(pmap_pte_quick(pmap, va));
1198 PT_SET_MA(PADDR1, 0);
1200 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) {
1201 m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
1207 vm_page_unlock_queues();
1212 /***************************************************
1213 * Low level mapping routines.....
1214 ***************************************************/
1217 * Add a wired page to the kva.
1218 * Note: not SMP coherent.
1221 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1223 PT_SET_MA(va, xpmap_ptom(pa)| PG_RW | PG_V | pgeflag);
1227 pmap_kenter_ma(vm_offset_t va, vm_paddr_t ma)
1232 pte_store_ma(pte, ma | PG_RW | PG_V | pgeflag);
1236 static __inline void
1237 pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode)
1239 PT_SET_MA(va, pa | PG_RW | PG_V | pgeflag | pmap_cache_bits(mode, 0));
1243 * Remove a page from the kernel pagetables.
1244 * Note: not SMP coherent.
1247 pmap_kremove(vm_offset_t va)
1252 PT_CLEAR_VA(pte, FALSE);
1256 * Used to map a range of physical addresses into kernel
1257 * virtual address space.
1259 * The value passed in '*virt' is a suggested virtual address for
1260 * the mapping. Architectures which can support a direct-mapped
1261 * physical to virtual region can return the appropriate address
1262 * within that region, leaving '*virt' unchanged. Other
1263 * architectures should map the pages starting at '*virt' and
1264 * update '*virt' with the first usable address after the mapped
1268 pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot)
1270 vm_offset_t va, sva;
1273 CTR4(KTR_PMAP, "pmap_map: va=0x%x start=0x%jx end=0x%jx prot=0x%x",
1274 va, start, end, prot);
1275 while (start < end) {
1276 pmap_kenter(va, start);
1280 pmap_invalidate_range(kernel_pmap, sva, va);
1287 * Add a list of wired pages to the kva
1288 * this routine is only used for temporary
1289 * kernel mappings that do not need to have
1290 * page modification or references recorded.
1291 * Note that old mappings are simply written
1292 * over. The page *must* be wired.
1293 * Note: SMP coherent. Uses a ranged shootdown IPI.
1296 pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count)
1298 pt_entry_t *endpte, *pte;
1300 vm_offset_t va = sva;
1302 multicall_entry_t mcl[16];
1303 multicall_entry_t *mclp = mcl;
1306 CTR2(KTR_PMAP, "pmap_qenter:sva=0x%x count=%d", va, count);
1308 endpte = pte + count;
1309 while (pte < endpte) {
1310 pa = xpmap_ptom(VM_PAGE_TO_PHYS(*ma)) | pgeflag | PG_RW | PG_V | PG_M | PG_A;
1312 mclp->op = __HYPERVISOR_update_va_mapping;
1314 mclp->args[1] = (uint32_t)(pa & 0xffffffff);
1315 mclp->args[2] = (uint32_t)(pa >> 32);
1316 mclp->args[3] = (*pte & PG_V) ? UVMF_INVLPG|UVMF_ALL : 0;
1323 if (mclcount == 16) {
1324 error = HYPERVISOR_multicall(mcl, mclcount);
1327 KASSERT(error == 0, ("bad multicall %d", error));
1331 error = HYPERVISOR_multicall(mcl, mclcount);
1332 KASSERT(error == 0, ("bad multicall %d", error));
1336 for (pte = vtopte(sva), mclcount = 0; mclcount < count; mclcount++, pte++)
1337 KASSERT(*pte, ("pte not set for va=0x%x", sva + mclcount*PAGE_SIZE));
1343 * This routine tears out page mappings from the
1344 * kernel -- it is meant only for temporary mappings.
1345 * Note: SMP coherent. Uses a ranged shootdown IPI.
1348 pmap_qremove(vm_offset_t sva, int count)
1352 CTR2(KTR_PMAP, "pmap_qremove: sva=0x%x count=%d", sva, count);
1354 vm_page_lock_queues();
1356 while (count-- > 0) {
1360 pmap_invalidate_range(kernel_pmap, sva, va);
1362 vm_page_unlock_queues();
1365 /***************************************************
1366 * Page table page management routines.....
1367 ***************************************************/
1368 static __inline void
1369 pmap_free_zero_pages(vm_page_t free)
1373 while (free != NULL) {
1376 vm_page_free_zero(m);
1381 * This routine unholds page table pages, and if the hold count
1382 * drops to zero, then it decrements the wire count.
1385 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, vm_page_t *free)
1389 if (m->wire_count == 0)
1390 return _pmap_unwire_pte_hold(pmap, m, free);
1396 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, vm_page_t *free)
1402 * unmap the page table page
1404 xen_pt_unpin(pmap->pm_pdir[m->pindex]);
1406 * page *might* contain residual mapping :-/
1408 PD_CLEAR_VA(pmap, m->pindex, TRUE);
1410 --pmap->pm_stats.resident_count;
1413 * This is a release store so that the ordinary store unmapping
1414 * the page table page is globally performed before TLB shoot-
1417 atomic_subtract_rel_int(&cnt.v_wire_count, 1);
1420 * Do an invltlb to make the invalidated mapping
1421 * take effect immediately.
1423 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex);
1424 pmap_invalidate_page(pmap, pteva);
1427 * Put page on a list so that it is released after
1428 * *ALL* TLB shootdown is done
1437 * After removing a page table entry, this routine is used to
1438 * conditionally free the page, and manage the hold/wire counts.
1441 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t *free)
1446 if (va >= VM_MAXUSER_ADDRESS)
1448 ptepde = PT_GET(pmap_pde(pmap, va));
1449 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME);
1450 return pmap_unwire_pte_hold(pmap, mpte, free);
1454 pmap_pinit0(pmap_t pmap)
1457 PMAP_LOCK_INIT(pmap);
1458 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD);
1460 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT);
1462 pmap->pm_active = 0;
1463 PCPU_SET(curpmap, pmap);
1464 TAILQ_INIT(&pmap->pm_pvchunk);
1465 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1466 mtx_lock_spin(&allpmaps_lock);
1467 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1468 mtx_unlock_spin(&allpmaps_lock);
1472 * Initialize a preallocated and zeroed pmap structure,
1473 * such as one in a vmspace structure.
1476 pmap_pinit(pmap_t pmap)
1478 vm_page_t m, ptdpg[NPGPTD + 1];
1479 int npgptd = NPGPTD + 1;
1483 PMAP_LOCK_INIT(pmap);
1486 * No need to allocate page table space yet but we do need a valid
1487 * page directory table.
1489 if (pmap->pm_pdir == NULL) {
1490 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map,
1492 if (pmap->pm_pdir == NULL) {
1493 PMAP_LOCK_DESTROY(pmap);
1496 #if defined(XEN) && defined(PAE)
1497 pmap->pm_pdpt = (pd_entry_t *)kmem_alloc_nofault(kernel_map, 1);
1500 #if defined(PAE) && !defined(XEN)
1501 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO);
1502 KASSERT(((vm_offset_t)pmap->pm_pdpt &
1503 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0,
1504 ("pmap_pinit: pdpt misaligned"));
1505 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30),
1506 ("pmap_pinit: pdpt above 4g"));
1511 * allocate the page directory page(s)
1513 for (i = 0; i < npgptd;) {
1514 m = vm_page_alloc(NULL, color++,
1515 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
1523 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD);
1524 for (i = 0; i < NPGPTD; i++) {
1525 if ((ptdpg[i]->flags & PG_ZERO) == 0)
1526 pagezero(&pmap->pm_pdir[i*NPTEPG]);
1529 mtx_lock_spin(&allpmaps_lock);
1530 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1531 mtx_unlock_spin(&allpmaps_lock);
1532 /* Wire in kernel global address entries. */
1534 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t));
1537 pmap_qenter((vm_offset_t)pmap->pm_pdpt, &ptdpg[NPGPTD], 1);
1538 if ((ptdpg[NPGPTD]->flags & PG_ZERO) == 0)
1539 bzero(pmap->pm_pdpt, PAGE_SIZE);
1541 for (i = 0; i < NPGPTD; i++) {
1544 ma = xpmap_ptom(VM_PAGE_TO_PHYS(ptdpg[i]));
1545 pmap->pm_pdpt[i] = ma | PG_V;
1550 for (i = 0; i < NPGPTD; i++) {
1554 ma = xpmap_ptom(VM_PAGE_TO_PHYS(ptdpg[i]));
1555 pd = pmap->pm_pdir + (i * NPDEPG);
1556 PT_SET_MA(pd, *vtopte((vm_offset_t)pd) & ~(PG_M|PG_A|PG_U|PG_RW));
1563 PT_SET_MA(pmap->pm_pdpt, *vtopte((vm_offset_t)pmap->pm_pdpt) & ~PG_RW);
1565 vm_page_lock_queues();
1567 xen_pgdpt_pin(xpmap_ptom(VM_PAGE_TO_PHYS(ptdpg[NPGPTD])));
1568 for (i = 0; i < NPGPTD; i++) {
1569 vm_paddr_t ma = xpmap_ptom(VM_PAGE_TO_PHYS(ptdpg[i]));
1570 PT_SET_VA_MA(&pmap->pm_pdir[PTDPTDI + i], ma | PG_V | PG_A, FALSE);
1573 vm_page_unlock_queues();
1575 pmap->pm_active = 0;
1576 TAILQ_INIT(&pmap->pm_pvchunk);
1577 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1583 * this routine is called if the page table page is not
1587 _pmap_allocpte(pmap_t pmap, unsigned int ptepindex, int flags)
1592 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT ||
1593 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK,
1594 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK"));
1597 * Allocate a page table page.
1599 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ |
1600 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
1601 if (flags & M_WAITOK) {
1603 vm_page_unlock_queues();
1605 vm_page_lock_queues();
1610 * Indicate the need to retry. While waiting, the page table
1611 * page may have been allocated.
1615 if ((m->flags & PG_ZERO) == 0)
1619 * Map the pagetable page into the process address space, if
1620 * it isn't already there.
1622 pmap->pm_stats.resident_count++;
1624 ptema = xpmap_ptom(VM_PAGE_TO_PHYS(m));
1626 PT_SET_VA_MA(&pmap->pm_pdir[ptepindex],
1627 (ptema | PG_U | PG_RW | PG_V | PG_A | PG_M), TRUE);
1629 KASSERT(pmap->pm_pdir[ptepindex],
1630 ("_pmap_allocpte: ptepindex=%d did not get mapped", ptepindex));
1635 pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags)
1641 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT ||
1642 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK,
1643 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK"));
1646 * Calculate pagetable page index
1648 ptepindex = va >> PDRSHIFT;
1651 * Get the page directory entry
1653 ptema = pmap->pm_pdir[ptepindex];
1656 * This supports switching from a 4MB page to a
1659 if (ptema & PG_PS) {
1663 pmap->pm_pdir[ptepindex] = 0;
1665 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1666 pmap_invalidate_all(kernel_pmap);
1670 * If the page table page is mapped, we just increment the
1671 * hold count, and activate it.
1674 m = PHYS_TO_VM_PAGE(xpmap_mtop(ptema) & PG_FRAME);
1678 * Here if the pte page isn't mapped, or if it has
1681 CTR3(KTR_PMAP, "pmap_allocpte: pmap=%p va=0x%08x flags=0x%x",
1683 m = _pmap_allocpte(pmap, ptepindex, flags);
1684 if (m == NULL && (flags & M_WAITOK))
1687 KASSERT(pmap->pm_pdir[ptepindex], ("ptepindex=%d did not get mapped", ptepindex));
1693 /***************************************************
1694 * Pmap allocation/deallocation routines.
1695 ***************************************************/
1699 * Deal with a SMP shootdown of other users of the pmap that we are
1700 * trying to dispose of. This can be a bit hairy.
1702 static u_int *lazymask;
1703 static u_int lazyptd;
1704 static volatile u_int lazywait;
1706 void pmap_lazyfix_action(void);
1709 pmap_lazyfix_action(void)
1711 u_int mymask = PCPU_GET(cpumask);
1714 (*ipi_lazypmap_counts[PCPU_GET(cpuid)])++;
1716 if (rcr3() == lazyptd)
1717 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1718 atomic_clear_int(lazymask, mymask);
1719 atomic_store_rel_int(&lazywait, 1);
1723 pmap_lazyfix_self(u_int mymask)
1726 if (rcr3() == lazyptd)
1727 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1728 atomic_clear_int(lazymask, mymask);
1733 pmap_lazyfix(pmap_t pmap)
1739 while ((mask = pmap->pm_active) != 0) {
1741 mask = mask & -mask; /* Find least significant set bit */
1742 mtx_lock_spin(&smp_ipi_mtx);
1744 lazyptd = vtophys(pmap->pm_pdpt);
1746 lazyptd = vtophys(pmap->pm_pdir);
1748 mymask = PCPU_GET(cpumask);
1749 if (mask == mymask) {
1750 lazymask = &pmap->pm_active;
1751 pmap_lazyfix_self(mymask);
1753 atomic_store_rel_int((u_int *)&lazymask,
1754 (u_int)&pmap->pm_active);
1755 atomic_store_rel_int(&lazywait, 0);
1756 ipi_selected(mask, IPI_LAZYPMAP);
1757 while (lazywait == 0) {
1763 mtx_unlock_spin(&smp_ipi_mtx);
1765 printf("pmap_lazyfix: spun for 50000000\n");
1772 * Cleaning up on uniprocessor is easy. For various reasons, we're
1773 * unlikely to have to even execute this code, including the fact
1774 * that the cleanup is deferred until the parent does a wait(2), which
1775 * means that another userland process has run.
1778 pmap_lazyfix(pmap_t pmap)
1782 cr3 = vtophys(pmap->pm_pdir);
1783 if (cr3 == rcr3()) {
1784 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1785 pmap->pm_active &= ~(PCPU_GET(cpumask));
1791 * Release any resources held by the given physical map.
1792 * Called when a pmap initialized by pmap_pinit is being released.
1793 * Should only be called if the map contains no valid mappings.
1796 pmap_release(pmap_t pmap)
1798 vm_page_t m, ptdpg[2*NPGPTD+1];
1803 int npgptd = NPGPTD + 1;
1805 int npgptd = NPGPTD;
1808 int npgptd = NPGPTD;
1810 KASSERT(pmap->pm_stats.resident_count == 0,
1811 ("pmap_release: pmap resident count %ld != 0",
1812 pmap->pm_stats.resident_count));
1816 mtx_lock_spin(&allpmaps_lock);
1817 LIST_REMOVE(pmap, pm_list);
1818 mtx_unlock_spin(&allpmaps_lock);
1820 for (i = 0; i < NPGPTD; i++)
1821 ptdpg[i] = PHYS_TO_VM_PAGE(vtophys(pmap->pm_pdir + (i*NPDEPG)) & PG_FRAME);
1822 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD);
1823 #if defined(PAE) && defined(XEN)
1824 ptdpg[NPGPTD] = PHYS_TO_VM_PAGE(vtophys(pmap->pm_pdpt));
1827 for (i = 0; i < npgptd; i++) {
1829 ma = xpmap_ptom(VM_PAGE_TO_PHYS(m));
1830 /* unpinning L1 and L2 treated the same */
1833 KASSERT(xpmap_ptom(VM_PAGE_TO_PHYS(m)) == (pmap->pm_pdpt[i] & PG_FRAME),
1834 ("pmap_release: got wrong ptd page"));
1837 atomic_subtract_int(&cnt.v_wire_count, 1);
1840 PMAP_LOCK_DESTROY(pmap);
1844 kvm_size(SYSCTL_HANDLER_ARGS)
1846 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1848 return sysctl_handle_long(oidp, &ksize, 0, req);
1850 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1851 0, 0, kvm_size, "IU", "Size of KVM");
1854 kvm_free(SYSCTL_HANDLER_ARGS)
1856 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1858 return sysctl_handle_long(oidp, &kfree, 0, req);
1860 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1861 0, 0, kvm_free, "IU", "Amount of KVM free");
1864 * grow the number of kernel page table entries, if needed
1867 pmap_growkernel(vm_offset_t addr)
1870 vm_paddr_t ptppaddr;
1874 mtx_assert(&kernel_map->system_mtx, MA_OWNED);
1875 if (kernel_vm_end == 0) {
1876 kernel_vm_end = KERNBASE;
1878 while (pdir_pde(PTD, kernel_vm_end)) {
1879 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1881 if (kernel_vm_end - 1 >= kernel_map->max_offset) {
1882 kernel_vm_end = kernel_map->max_offset;
1887 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1888 if (addr - 1 >= kernel_map->max_offset)
1889 addr = kernel_map->max_offset;
1890 while (kernel_vm_end < addr) {
1891 if (pdir_pde(PTD, kernel_vm_end)) {
1892 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1893 if (kernel_vm_end - 1 >= kernel_map->max_offset) {
1894 kernel_vm_end = kernel_map->max_offset;
1901 * This index is bogus, but out of the way
1903 nkpg = vm_page_alloc(NULL, nkpt,
1904 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
1906 panic("pmap_growkernel: no memory to grow kernel");
1910 pmap_zero_page(nkpg);
1911 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1912 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1913 vm_page_lock_queues();
1914 PD_SET_VA(kernel_pmap, (kernel_vm_end >> PDRSHIFT), newpdir, TRUE);
1915 mtx_lock_spin(&allpmaps_lock);
1916 LIST_FOREACH(pmap, &allpmaps, pm_list)
1917 PD_SET_VA(pmap, (kernel_vm_end >> PDRSHIFT), newpdir, TRUE);
1919 mtx_unlock_spin(&allpmaps_lock);
1920 vm_page_unlock_queues();
1922 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1923 if (kernel_vm_end - 1 >= kernel_map->max_offset) {
1924 kernel_vm_end = kernel_map->max_offset;
1931 /***************************************************
1932 * page management routines.
1933 ***************************************************/
1935 CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE);
1936 CTASSERT(_NPCM == 11);
1938 static __inline struct pv_chunk *
1939 pv_to_chunk(pv_entry_t pv)
1942 return (struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK);
1945 #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap)
1947 #define PC_FREE0_9 0xfffffffful /* Free values for index 0 through 9 */
1948 #define PC_FREE10 0x0000fffful /* Free values for index 10 */
1950 static uint32_t pc_freemask[11] = {
1951 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9,
1952 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9,
1953 PC_FREE0_9, PC_FREE0_9, PC_FREE0_9,
1954 PC_FREE0_9, PC_FREE10
1957 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0,
1958 "Current number of pv entries");
1961 static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail;
1963 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0,
1964 "Current number of pv entry chunks");
1965 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0,
1966 "Current number of pv entry chunks allocated");
1967 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0,
1968 "Current number of pv entry chunks frees");
1969 SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0,
1970 "Number of times tried to get a chunk page but failed.");
1972 static long pv_entry_frees, pv_entry_allocs;
1973 static int pv_entry_spare;
1975 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0,
1976 "Current number of pv entry frees");
1977 SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0,
1978 "Current number of pv entry allocs");
1979 SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0,
1980 "Current number of spare pv entries");
1982 static int pmap_collect_inactive, pmap_collect_active;
1984 SYSCTL_INT(_vm_pmap, OID_AUTO, pmap_collect_inactive, CTLFLAG_RD, &pmap_collect_inactive, 0,
1985 "Current number times pmap_collect called on inactive queue");
1986 SYSCTL_INT(_vm_pmap, OID_AUTO, pmap_collect_active, CTLFLAG_RD, &pmap_collect_active, 0,
1987 "Current number times pmap_collect called on active queue");
1991 * We are in a serious low memory condition. Resort to
1992 * drastic measures to free some pages so we can allocate
1993 * another pv entry chunk. This is normally called to
1994 * unmap inactive pages, and if necessary, active pages.
1997 pmap_collect(pmap_t locked_pmap, struct vpgqueues *vpq)
2000 pt_entry_t *pte, tpte;
2001 pv_entry_t next_pv, pv;
2006 TAILQ_FOREACH(m, &vpq->pl, pageq) {
2007 if (m->hold_count || m->busy)
2009 TAILQ_FOREACH_SAFE(pv, &m->md.pv_list, pv_list, next_pv) {
2012 /* Avoid deadlock and lock recursion. */
2013 if (pmap > locked_pmap)
2015 else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap))
2017 pmap->pm_stats.resident_count--;
2018 pte = pmap_pte_quick(pmap, va);
2019 tpte = pte_load_clear(pte);
2020 KASSERT((tpte & PG_W) == 0,
2021 ("pmap_collect: wired pte %#jx", (uintmax_t)tpte));
2023 vm_page_flag_set(m, PG_REFERENCED);
2025 KASSERT((tpte & PG_RW),
2026 ("pmap_collect: modified page not writable: va: %#x, pte: %#jx",
2027 va, (uintmax_t)tpte));
2031 pmap_unuse_pt(pmap, va, &free);
2032 pmap_invalidate_page(pmap, va);
2033 pmap_free_zero_pages(free);
2034 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2035 if (TAILQ_EMPTY(&m->md.pv_list))
2036 vm_page_flag_clear(m, PG_WRITEABLE);
2037 free_pv_entry(pmap, pv);
2038 if (pmap != locked_pmap)
2047 * free the pv_entry back to the free list
2050 free_pv_entry(pmap_t pmap, pv_entry_t pv)
2053 struct pv_chunk *pc;
2054 int idx, field, bit;
2056 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2057 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2058 PV_STAT(pv_entry_frees++);
2059 PV_STAT(pv_entry_spare++);
2061 pc = pv_to_chunk(pv);
2062 idx = pv - &pc->pc_pventry[0];
2065 pc->pc_map[field] |= 1ul << bit;
2066 /* move to head of list */
2067 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2068 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
2069 for (idx = 0; idx < _NPCM; idx++)
2070 if (pc->pc_map[idx] != pc_freemask[idx])
2072 PV_STAT(pv_entry_spare -= _NPCPV);
2073 PV_STAT(pc_chunk_count--);
2074 PV_STAT(pc_chunk_frees++);
2075 /* entire chunk is free, return it */
2076 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2077 m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc));
2078 pmap_qremove((vm_offset_t)pc, 1);
2079 vm_page_unwire(m, 0);
2081 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc);
2085 * get a new pv_entry, allocating a block from the system
2089 get_pv_entry(pmap_t pmap, int try)
2091 static const struct timeval printinterval = { 60, 0 };
2092 static struct timeval lastprint;
2093 static vm_pindex_t colour;
2094 struct vpgqueues *pq;
2097 struct pv_chunk *pc;
2100 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2101 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2102 PV_STAT(pv_entry_allocs++);
2104 if (pv_entry_count > pv_entry_high_water)
2105 if (ratecheck(&lastprint, &printinterval))
2106 printf("Approaching the limit on PV entries, consider "
2107 "increasing either the vm.pmap.shpgperproc or the "
2108 "vm.pmap.pv_entry_max tunable.\n");
2111 pc = TAILQ_FIRST(&pmap->pm_pvchunk);
2113 for (field = 0; field < _NPCM; field++) {
2114 if (pc->pc_map[field]) {
2115 bit = bsfl(pc->pc_map[field]);
2119 if (field < _NPCM) {
2120 pv = &pc->pc_pventry[field * 32 + bit];
2121 pc->pc_map[field] &= ~(1ul << bit);
2122 /* If this was the last item, move it to tail */
2123 for (field = 0; field < _NPCM; field++)
2124 if (pc->pc_map[field] != 0) {
2125 PV_STAT(pv_entry_spare--);
2126 return (pv); /* not full, return */
2128 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2129 TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
2130 PV_STAT(pv_entry_spare--);
2135 * Access to the ptelist "pv_vafree" is synchronized by the page
2136 * queues lock. If "pv_vafree" is currently non-empty, it will
2137 * remain non-empty until pmap_ptelist_alloc() completes.
2139 if (pv_vafree == 0 || (m = vm_page_alloc(NULL, colour, (pq ==
2140 &vm_page_queues[PQ_ACTIVE] ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL) |
2141 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) {
2144 PV_STAT(pc_chunk_tryfail++);
2148 * Reclaim pv entries: At first, destroy mappings to
2149 * inactive pages. After that, if a pv chunk entry
2150 * is still needed, destroy mappings to active pages.
2153 PV_STAT(pmap_collect_inactive++);
2154 pq = &vm_page_queues[PQ_INACTIVE];
2155 } else if (pq == &vm_page_queues[PQ_INACTIVE]) {
2156 PV_STAT(pmap_collect_active++);
2157 pq = &vm_page_queues[PQ_ACTIVE];
2159 panic("get_pv_entry: increase vm.pmap.shpgperproc");
2160 pmap_collect(pmap, pq);
2163 PV_STAT(pc_chunk_count++);
2164 PV_STAT(pc_chunk_allocs++);
2166 pc = (struct pv_chunk *)pmap_ptelist_alloc(&pv_vafree);
2167 pmap_qenter((vm_offset_t)pc, &m, 1);
2168 if ((m->flags & PG_ZERO) == 0)
2171 pc->pc_map[0] = pc_freemask[0] & ~1ul; /* preallocated bit 0 */
2172 for (field = 1; field < _NPCM; field++)
2173 pc->pc_map[field] = pc_freemask[field];
2174 pv = &pc->pc_pventry[0];
2175 TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
2176 PV_STAT(pv_entry_spare += _NPCPV - 1);
2181 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
2185 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2186 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2187 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2188 if (pmap == PV_PMAP(pv) && va == pv->pv_va)
2191 KASSERT(pv != NULL, ("pmap_remove_entry: pv not found"));
2192 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2193 if (TAILQ_EMPTY(&m->md.pv_list))
2194 vm_page_flag_clear(m, PG_WRITEABLE);
2195 free_pv_entry(pmap, pv);
2199 * Create a pv entry for page at pa for
2203 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
2207 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2208 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2209 pv = get_pv_entry(pmap, FALSE);
2211 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2215 * Conditionally create a pv entry.
2218 pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
2222 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2223 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2224 if (pv_entry_count < pv_entry_high_water &&
2225 (pv = get_pv_entry(pmap, TRUE)) != NULL) {
2227 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2234 * pmap_remove_pte: do the things to unmap a page in a process
2237 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va, vm_page_t *free)
2242 CTR3(KTR_PMAP, "pmap_remove_pte: pmap=%p *ptq=0x%x va=0x%x",
2243 pmap, (u_long)*ptq, va);
2245 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2246 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2248 PT_SET_VA_MA(ptq, 0, TRUE);
2250 pmap->pm_stats.wired_count -= 1;
2252 * Machines that don't support invlpg, also don't support
2256 pmap_invalidate_page(kernel_pmap, va);
2257 pmap->pm_stats.resident_count -= 1;
2259 * XXX This is not strictly correctly, but somewhere along the line
2260 * we are losing the managed bit on some pages. It is unclear to me
2261 * why, but I think the most likely explanation is that xen's writable
2262 * page table implementation doesn't respect the unused bits.
2264 if ((oldpte & PG_MANAGED) || ((oldpte & PG_V) && (va < VM_MAXUSER_ADDRESS))
2266 m = PHYS_TO_VM_PAGE(xpmap_mtop(oldpte) & PG_FRAME);
2268 if (!(oldpte & PG_MANAGED))
2269 printf("va=0x%x is unmanaged :-( pte=0x%llx\n", va, oldpte);
2271 if (oldpte & PG_M) {
2272 KASSERT((oldpte & PG_RW),
2273 ("pmap_remove_pte: modified page not writable: va: %#x, pte: %#jx",
2274 va, (uintmax_t)oldpte));
2278 vm_page_flag_set(m, PG_REFERENCED);
2279 pmap_remove_entry(pmap, m, va);
2280 } else if ((va < VM_MAXUSER_ADDRESS) && (oldpte & PG_V))
2281 printf("va=0x%x is unmanaged :-( pte=0x%llx\n", va, oldpte);
2283 return (pmap_unuse_pt(pmap, va, free));
2287 * Remove a single page from a process address space
2290 pmap_remove_page(pmap_t pmap, vm_offset_t va, vm_page_t *free)
2294 CTR2(KTR_PMAP, "pmap_remove_page: pmap=%p va=0x%x",
2297 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2298 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
2299 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2300 if ((pte = pmap_pte_quick(pmap, va)) == NULL || (*pte & PG_V) == 0)
2302 pmap_remove_pte(pmap, pte, va, free);
2303 pmap_invalidate_page(pmap, va);
2305 PT_SET_MA(PADDR1, 0);
2310 * Remove the given range of addresses from the specified map.
2312 * It is assumed that the start and end are properly
2313 * rounded to the page size.
2316 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2321 vm_page_t free = NULL;
2324 CTR3(KTR_PMAP, "pmap_remove: pmap=%p sva=0x%x eva=0x%x",
2328 * Perform an unsynchronized read. This is, however, safe.
2330 if (pmap->pm_stats.resident_count == 0)
2335 vm_page_lock_queues();
2340 * special handling of removing one page. a very
2341 * common operation and easy to short circuit some
2344 if ((sva + PAGE_SIZE == eva) &&
2345 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
2346 pmap_remove_page(pmap, sva, &free);
2350 for (; sva < eva; sva = pdnxt) {
2354 * Calculate index for next page table.
2356 pdnxt = (sva + NBPDR) & ~PDRMASK;
2357 if (pmap->pm_stats.resident_count == 0)
2360 pdirindex = sva >> PDRSHIFT;
2361 ptpaddr = pmap->pm_pdir[pdirindex];
2364 * Weed out invalid mappings. Note: we assume that the page
2365 * directory table is always allocated, and in kernel virtual.
2371 * Check for large page.
2373 if ((ptpaddr & PG_PS) != 0) {
2374 PD_CLEAR_VA(pmap, pdirindex, TRUE);
2375 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2381 * Limit our scan to either the end of the va represented
2382 * by the current page table page, or to the end of the
2383 * range being removed.
2388 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
2390 if ((*pte & PG_V) == 0)
2394 * The TLB entry for a PG_G mapping is invalidated
2395 * by pmap_remove_pte().
2397 if ((*pte & PG_G) == 0)
2399 if (pmap_remove_pte(pmap, pte, sva, &free))
2405 PT_SET_VA_MA(PMAP1, 0, TRUE);
2408 pmap_invalidate_all(pmap);
2410 vm_page_unlock_queues();
2412 pmap_free_zero_pages(free);
2416 * Routine: pmap_remove_all
2418 * Removes this physical page from
2419 * all physical maps in which it resides.
2420 * Reflects back modify bits to the pager.
2423 * Original versions of this routine were very
2424 * inefficient because they iteratively called
2425 * pmap_remove (slow...)
2429 pmap_remove_all(vm_page_t m)
2433 pt_entry_t *pte, tpte;
2436 #if defined(PMAP_DIAGNOSTIC)
2438 * XXX This makes pmap_remove_all() illegal for non-managed pages!
2440 if (m->flags & PG_FICTITIOUS) {
2441 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%jx",
2442 VM_PAGE_TO_PHYS(m) & 0xffffffff);
2445 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2447 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2450 pmap->pm_stats.resident_count--;
2451 pte = pmap_pte_quick(pmap, pv->pv_va);
2454 PT_SET_VA_MA(pte, 0, TRUE);
2456 pmap->pm_stats.wired_count--;
2458 vm_page_flag_set(m, PG_REFERENCED);
2461 * Update the vm_page_t clean and reference bits.
2464 KASSERT((tpte & PG_RW),
2465 ("pmap_remove_all: modified page not writable: va: %#x, pte: %#jx",
2466 pv->pv_va, (uintmax_t)tpte));
2470 pmap_unuse_pt(pmap, pv->pv_va, &free);
2471 pmap_invalidate_page(pmap, pv->pv_va);
2472 pmap_free_zero_pages(free);
2473 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2474 free_pv_entry(pmap, pv);
2477 vm_page_flag_clear(m, PG_WRITEABLE);
2480 PT_SET_MA(PADDR1, 0);
2485 * Set the physical protection on the
2486 * specified range of this map as requested.
2489 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2496 CTR4(KTR_PMAP, "pmap_protect: pmap=%p sva=0x%x eva=0x%x prot=0x%x",
2497 pmap, sva, eva, prot);
2499 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2500 pmap_remove(pmap, sva, eva);
2505 if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) ==
2506 (VM_PROT_WRITE|VM_PROT_EXECUTE))
2509 if (prot & VM_PROT_WRITE)
2515 vm_page_lock_queues();
2518 for (; sva < eva; sva = pdnxt) {
2519 pt_entry_t obits, pbits;
2522 pdnxt = (sva + NBPDR) & ~PDRMASK;
2524 pdirindex = sva >> PDRSHIFT;
2525 ptpaddr = pmap->pm_pdir[pdirindex];
2528 * Weed out invalid mappings. Note: we assume that the page
2529 * directory table is always allocated, and in kernel virtual.
2535 * Check for large page.
2537 if ((ptpaddr & PG_PS) != 0) {
2538 if ((prot & VM_PROT_WRITE) == 0)
2539 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2541 if ((prot & VM_PROT_EXECUTE) == 0)
2542 pmap->pm_pdir[pdirindex] |= pg_nx;
2551 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
2557 * Regardless of whether a pte is 32 or 64 bits in
2558 * size, PG_RW, PG_A, and PG_M are among the least
2559 * significant 32 bits.
2561 obits = pbits = *pte;
2562 if ((pbits & PG_V) == 0)
2564 if (pbits & PG_MANAGED) {
2567 m = PHYS_TO_VM_PAGE(xpmap_mtop(pbits) & PG_FRAME);
2568 vm_page_flag_set(m, PG_REFERENCED);
2571 if ((pbits & PG_M) != 0) {
2573 m = PHYS_TO_VM_PAGE(xpmap_mtop(pbits) & PG_FRAME);
2578 if ((prot & VM_PROT_WRITE) == 0)
2579 pbits &= ~(PG_RW | PG_M);
2581 if ((prot & VM_PROT_EXECUTE) == 0)
2585 if (pbits != obits) {
2588 PT_SET_VA_MA(pte, pbits, TRUE);
2593 if (!atomic_cmpset_64(pte, obits, pbits))
2596 if (!atomic_cmpset_int((u_int *)pte, obits,
2602 pmap_invalidate_page(pmap, sva);
2610 PT_SET_VA_MA(PMAP1, 0, TRUE);
2612 pmap_invalidate_all(pmap);
2614 vm_page_unlock_queues();
2619 * Insert the given physical page (p) at
2620 * the specified virtual address (v) in the
2621 * target physical map with the protection requested.
2623 * If specified, the page will be wired down, meaning
2624 * that the related pte can not be reclaimed.
2626 * NB: This is the only routine which MAY NOT lazy-evaluate
2627 * or lose information. That is, this routine must actually
2628 * insert this page into the given map NOW.
2631 pmap_enter(pmap_t pmap, vm_offset_t va, vm_prot_t access, vm_page_t m,
2632 vm_prot_t prot, boolean_t wired)
2638 pt_entry_t origpte, newpte;
2642 CTR6(KTR_PMAP, "pmap_enter: pmap=%08p va=0x%08x access=0x%x ma=0x%08x prot=0x%x wired=%d",
2643 pmap, va, access, xpmap_ptom(VM_PAGE_TO_PHYS(m)), prot, wired);
2644 va = trunc_page(va);
2645 #ifdef PMAP_DIAGNOSTIC
2646 if (va > VM_MAX_KERNEL_ADDRESS)
2647 panic("pmap_enter: toobig");
2648 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2649 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
2654 vm_page_lock_queues();
2659 * In the case that a page table page is not
2660 * resident, we are creating it here.
2662 if (va < VM_MAXUSER_ADDRESS) {
2663 mpte = pmap_allocpte(pmap, va, M_WAITOK);
2665 #if 0 && defined(PMAP_DIAGNOSTIC)
2667 pd_entry_t *pdeaddr = pmap_pde(pmap, va);
2669 if ((origpte & PG_V) == 0) {
2670 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n",
2671 pmap->pm_pdir[PTDPTDI], origpte, va);
2676 pde = pmap_pde(pmap, va);
2677 if ((*pde & PG_PS) != 0)
2678 panic("pmap_enter: attempted pmap_enter on 4MB page");
2679 pte = pmap_pte_quick(pmap, va);
2682 * Page Directory table entry not valid, we need a new PT page
2685 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n",
2686 (uintmax_t)pmap->pm_pdir[va >> PDRSHIFT], va);
2689 pa = VM_PAGE_TO_PHYS(m);
2694 KASSERT((*pte & PG_V) || (*pte == 0), ("address set but not valid pte=%p *pte=0x%016jx",
2699 origpte = xpmap_mtop(origpte);
2700 opa = origpte & PG_FRAME;
2703 * Mapping has not changed, must be protection or wiring change.
2705 if (origpte && (opa == pa)) {
2707 * Wiring change, just update stats. We don't worry about
2708 * wiring PT pages as they remain resident as long as there
2709 * are valid mappings in them. Hence, if a user page is wired,
2710 * the PT page will be also.
2712 if (wired && ((origpte & PG_W) == 0))
2713 pmap->pm_stats.wired_count++;
2714 else if (!wired && (origpte & PG_W))
2715 pmap->pm_stats.wired_count--;
2718 * Remove extra pte reference
2724 * We might be turning off write access to the page,
2725 * so we go ahead and sense modify status.
2727 if (origpte & PG_MANAGED) {
2734 * Mapping has changed, invalidate old range and fall through to
2735 * handle validating new mapping.
2739 pmap->pm_stats.wired_count--;
2740 if (origpte & PG_MANAGED) {
2741 om = PHYS_TO_VM_PAGE(opa);
2742 pmap_remove_entry(pmap, om, va);
2743 } else if (va < VM_MAXUSER_ADDRESS)
2744 printf("va=0x%x is unmanaged :-( \n", va);
2748 KASSERT(mpte->wire_count > 0,
2749 ("pmap_enter: missing reference to page table page,"
2753 pmap->pm_stats.resident_count++;
2756 * Enter on the PV list if part of our managed memory.
2758 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) {
2759 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva,
2760 ("pmap_enter: managed mapping within the clean submap"));
2761 pmap_insert_entry(pmap, va, m);
2766 * Increment counters
2769 pmap->pm_stats.wired_count++;
2773 * Now validate mapping with desired protection/wiring.
2775 newpte = (pt_entry_t)(pa | PG_V);
2776 if ((prot & VM_PROT_WRITE) != 0) {
2778 vm_page_flag_set(m, PG_WRITEABLE);
2781 if ((prot & VM_PROT_EXECUTE) == 0)
2786 if (va < VM_MAXUSER_ADDRESS)
2788 if (pmap == kernel_pmap)
2793 * if the mapping or permission bits are different, we need
2794 * to update the pte.
2796 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2800 PT_SET_VA(pte, newpte | PG_A, FALSE);
2801 if (origpte & PG_A) {
2802 if (origpte & PG_MANAGED)
2803 vm_page_flag_set(om, PG_REFERENCED);
2804 if (opa != VM_PAGE_TO_PHYS(m))
2807 if ((origpte & PG_NX) == 0 &&
2808 (newpte & PG_NX) != 0)
2812 if (origpte & PG_M) {
2813 KASSERT((origpte & PG_RW),
2814 ("pmap_enter: modified page not writable: va: %#x, pte: %#jx",
2815 va, (uintmax_t)origpte));
2816 if ((origpte & PG_MANAGED) != 0)
2818 if ((prot & VM_PROT_WRITE) == 0)
2822 pmap_invalidate_page(pmap, va);
2824 PT_SET_VA(pte, newpte | PG_A, FALSE);
2831 PT_SET_VA_MA(PMAP1, 0, TRUE);
2833 vm_page_unlock_queues();
2838 * Maps a sequence of resident pages belonging to the same object.
2839 * The sequence begins with the given page m_start. This page is
2840 * mapped at the given virtual address start. Each subsequent page is
2841 * mapped at a virtual address that is offset from start by the same
2842 * amount as the page is offset from m_start within the object. The
2843 * last page in the sequence is the page with the largest offset from
2844 * m_start that can be mapped at a virtual address less than the given
2845 * virtual address end. Not every virtual page between start and end
2846 * is mapped; only those for which a resident page exists with the
2847 * corresponding offset from m_start are mapped.
2850 pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end,
2851 vm_page_t m_start, vm_prot_t prot)
2854 vm_pindex_t diff, psize;
2855 multicall_entry_t mcl[16];
2856 multicall_entry_t *mclp = mcl;
2857 int error, count = 0;
2859 VM_OBJECT_LOCK_ASSERT(m_start->object, MA_OWNED);
2860 psize = atop(end - start);
2865 while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
2866 mpte = pmap_enter_quick_locked(&mclp, &count, pmap, start + ptoa(diff), m,
2868 m = TAILQ_NEXT(m, listq);
2870 error = HYPERVISOR_multicall(mcl, count);
2871 KASSERT(error == 0, ("bad multicall %d", error));
2877 error = HYPERVISOR_multicall(mcl, count);
2878 KASSERT(error == 0, ("bad multicall %d", error));
2885 * this code makes some *MAJOR* assumptions:
2886 * 1. Current pmap & pmap exists.
2889 * 4. No page table pages.
2890 * but is *MUCH* faster than pmap_enter...
2894 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot)
2896 multicall_entry_t mcl, *mclp;
2900 CTR4(KTR_PMAP, "pmap_enter_quick: pmap=%p va=0x%x m=%p prot=0x%x",
2904 (void) pmap_enter_quick_locked(&mclp, &count, pmap, va, m, prot, NULL);
2906 HYPERVISOR_multicall(&mcl, count);
2912 pmap_enter_quick_range(pmap_t pmap, vm_offset_t *addrs, vm_page_t *pages, vm_prot_t *prots, int count)
2914 int i, error, index = 0;
2915 multicall_entry_t mcl[16];
2916 multicall_entry_t *mclp = mcl;
2919 for (i = 0; i < count; i++, addrs++, pages++, prots++) {
2920 if (!pmap_is_prefaultable_locked(pmap, *addrs))
2923 (void) pmap_enter_quick_locked(&mclp, &index, pmap, *addrs, *pages, *prots, NULL);
2925 error = HYPERVISOR_multicall(mcl, index);
2928 KASSERT(error == 0, ("bad multicall %d", error));
2932 error = HYPERVISOR_multicall(mcl, index);
2933 KASSERT(error == 0, ("bad multicall %d", error));
2941 pmap_enter_quick_locked(multicall_entry_t **mclpp, int *count, pmap_t pmap, vm_offset_t va, vm_page_t m,
2942 vm_prot_t prot, vm_page_t mpte)
2947 multicall_entry_t *mcl = *mclpp;
2949 KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva ||
2950 (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0,
2951 ("pmap_enter_quick_locked: managed mapping within the clean submap"));
2952 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2953 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2956 * In the case that a page table page is not
2957 * resident, we are creating it here.
2959 if (va < VM_MAXUSER_ADDRESS) {
2964 * Calculate pagetable page index
2966 ptepindex = va >> PDRSHIFT;
2967 if (mpte && (mpte->pindex == ptepindex)) {
2971 * Get the page directory entry
2973 ptema = pmap->pm_pdir[ptepindex];
2976 * If the page table page is mapped, we just increment
2977 * the hold count, and activate it.
2981 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2982 mpte = PHYS_TO_VM_PAGE(xpmap_mtop(ptema) & PG_FRAME);
2985 mpte = _pmap_allocpte(pmap, ptepindex,
2996 * This call to vtopte makes the assumption that we are
2997 * entering the page into the current pmap. In order to support
2998 * quick entry into any pmap, one would likely use pmap_pte_quick.
2999 * But that isn't as quick as vtopte.
3001 KASSERT(pmap_is_current(pmap), ("entering pages in non-current pmap"));
3012 * Enter on the PV list if part of our managed memory.
3014 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0 &&
3015 !pmap_try_insert_pv_entry(pmap, va, m)) {
3018 if (pmap_unwire_pte_hold(pmap, mpte, &free)) {
3019 pmap_invalidate_page(pmap, va);
3020 pmap_free_zero_pages(free);
3029 * Increment counters
3031 pmap->pm_stats.resident_count++;
3033 pa = VM_PAGE_TO_PHYS(m);
3035 if ((prot & VM_PROT_EXECUTE) == 0)
3041 * Now validate mapping with RO protection
3043 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
3044 pte_store(pte, pa | PG_V | PG_U);
3046 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED);
3049 * Now validate mapping with RO protection
3051 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
3052 pa = xpmap_ptom(pa | PG_V | PG_U);
3054 pa = xpmap_ptom(pa | PG_V | PG_U | PG_MANAGED);
3056 mcl->op = __HYPERVISOR_update_va_mapping;
3058 mcl->args[1] = (uint32_t)(pa & 0xffffffff);
3059 mcl->args[2] = (uint32_t)(pa >> 32);
3062 *count = *count + 1;
3068 * Make a temporary mapping for a physical address. This is only intended
3069 * to be used for panic dumps.
3072 pmap_kenter_temporary(vm_paddr_t pa, int i)
3076 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
3077 pmap_kenter(va, pa);
3079 return ((void *)crashdumpmap);
3083 * This code maps large physical mmap regions into the
3084 * processor address space. Note that some shortcuts
3085 * are taken, but the code works.
3088 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr,
3089 vm_object_t object, vm_pindex_t pindex,
3094 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
3095 KASSERT(object->type == OBJT_DEVICE,
3096 ("pmap_object_init_pt: non-device object"));
3098 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) {
3101 unsigned int ptepindex;
3106 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)])
3110 p = vm_page_lookup(object, pindex);
3112 if (vm_page_sleep_if_busy(p, FALSE, "init4p"))
3115 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
3120 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) {
3121 vm_page_lock_queues();
3123 vm_page_unlock_queues();
3127 p = vm_page_lookup(object, pindex);
3128 vm_page_lock_queues();
3130 vm_page_unlock_queues();
3133 ptepa = VM_PAGE_TO_PHYS(p);
3134 if (ptepa & (NBPDR - 1))
3137 p->valid = VM_PAGE_BITS_ALL;
3140 pmap->pm_stats.resident_count += size >> PAGE_SHIFT;
3141 npdes = size >> PDRSHIFT;
3143 for(i = 0; i < npdes; i++) {
3144 PD_SET_VA(pmap, ptepindex,
3145 ptepa | PG_U | PG_M | PG_RW | PG_V | PG_PS, FALSE);
3149 pmap_invalidate_all(pmap);
3157 * Routine: pmap_change_wiring
3158 * Function: Change the wiring attribute for a map/virtual-address
3160 * In/out conditions:
3161 * The mapping must already exist in the pmap.
3164 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
3168 vm_page_lock_queues();
3170 pte = pmap_pte(pmap, va);
3172 if (wired && !pmap_pte_w(pte)) {
3173 PT_SET_VA_MA((pte), *(pte) | PG_W, TRUE);
3174 pmap->pm_stats.wired_count++;
3175 } else if (!wired && pmap_pte_w(pte)) {
3176 PT_SET_VA_MA((pte), *(pte) & ~PG_W, TRUE);
3177 pmap->pm_stats.wired_count--;
3181 * Wiring is not a hardware characteristic so there is no need to
3184 pmap_pte_release(pte);
3186 vm_page_unlock_queues();
3192 * Copy the range specified by src_addr/len
3193 * from the source map to the range dst_addr/len
3194 * in the destination map.
3196 * This routine is only advisory and need not do anything.
3200 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
3201 vm_offset_t src_addr)
3205 vm_offset_t end_addr = src_addr + len;
3208 if (dst_addr != src_addr)
3211 if (!pmap_is_current(src_pmap)) {
3213 "pmap_copy, skipping: pdir[PTDPTDI]=0x%jx PTDpde[0]=0x%jx",
3214 (src_pmap->pm_pdir[PTDPTDI] & PG_FRAME), (PTDpde[0] & PG_FRAME));
3218 CTR5(KTR_PMAP, "pmap_copy: dst_pmap=%p src_pmap=%p dst_addr=0x%x len=%d src_addr=0x%x",
3219 dst_pmap, src_pmap, dst_addr, len, src_addr);
3221 vm_page_lock_queues();
3222 if (dst_pmap < src_pmap) {
3223 PMAP_LOCK(dst_pmap);
3224 PMAP_LOCK(src_pmap);
3226 PMAP_LOCK(src_pmap);
3227 PMAP_LOCK(dst_pmap);
3230 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3231 pt_entry_t *src_pte, *dst_pte;
3232 vm_page_t dstmpte, srcmpte;
3233 pd_entry_t srcptepaddr;
3236 if (addr >= UPT_MIN_ADDRESS)
3237 panic("pmap_copy: invalid to pmap_copy page tables");
3239 pdnxt = (addr + NBPDR) & ~PDRMASK;
3240 ptepindex = addr >> PDRSHIFT;
3242 srcptepaddr = PT_GET(&src_pmap->pm_pdir[ptepindex]);
3243 if (srcptepaddr == 0)
3246 if (srcptepaddr & PG_PS) {
3247 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3248 PD_SET_VA(dst_pmap, ptepindex, srcptepaddr & ~PG_W, TRUE);
3249 dst_pmap->pm_stats.resident_count +=
3255 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr & PG_FRAME);
3256 if (srcmpte->wire_count == 0)
3257 panic("pmap_copy: source page table page is unused");
3259 if (pdnxt > end_addr)
3262 src_pte = vtopte(addr);
3263 while (addr < pdnxt) {
3267 * we only virtual copy managed pages
3269 if ((ptetemp & PG_MANAGED) != 0) {
3270 dstmpte = pmap_allocpte(dst_pmap, addr,
3272 if (dstmpte == NULL)
3274 dst_pte = pmap_pte_quick(dst_pmap, addr);
3275 if (*dst_pte == 0 &&
3276 pmap_try_insert_pv_entry(dst_pmap, addr,
3277 PHYS_TO_VM_PAGE(xpmap_mtop(ptetemp) & PG_FRAME))) {
3279 * Clear the wired, modified, and
3280 * accessed (referenced) bits
3283 KASSERT(ptetemp != 0, ("src_pte not set"));
3284 PT_SET_VA_MA(dst_pte, ptetemp & ~(PG_W | PG_M | PG_A), TRUE /* XXX debug */);
3285 KASSERT(*dst_pte == (ptetemp & ~(PG_W | PG_M | PG_A)),
3286 ("no pmap copy expected: 0x%jx saw: 0x%jx",
3287 ptetemp & ~(PG_W | PG_M | PG_A), *dst_pte));
3288 dst_pmap->pm_stats.resident_count++;
3291 if (pmap_unwire_pte_hold(dst_pmap,
3293 pmap_invalidate_page(dst_pmap,
3295 pmap_free_zero_pages(free);
3298 if (dstmpte->wire_count >= srcmpte->wire_count)
3307 vm_page_unlock_queues();
3308 PMAP_UNLOCK(src_pmap);
3309 PMAP_UNLOCK(dst_pmap);
3313 * pmap_zero_page zeros the specified hardware page by mapping
3314 * the page into KVM and using bzero to clear its contents.
3317 pmap_zero_page(vm_page_t m)
3319 struct sysmaps *sysmaps;
3321 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
3322 mtx_lock(&sysmaps->lock);
3323 if (*sysmaps->CMAP2)
3324 panic("pmap_zero_page: CMAP2 busy");
3326 PT_SET_MA(sysmaps->CADDR2, PG_V | PG_RW | xpmap_ptom(VM_PAGE_TO_PHYS(m)) | PG_A | PG_M);
3327 pagezero(sysmaps->CADDR2);
3328 PT_SET_MA(sysmaps->CADDR2, 0);
3330 mtx_unlock(&sysmaps->lock);
3334 * pmap_zero_page_area zeros the specified hardware page by mapping
3335 * the page into KVM and using bzero to clear its contents.
3337 * off and size may not cover an area beyond a single hardware page.
3340 pmap_zero_page_area(vm_page_t m, int off, int size)
3342 struct sysmaps *sysmaps;
3344 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
3345 mtx_lock(&sysmaps->lock);
3346 if (*sysmaps->CMAP2)
3347 panic("pmap_zero_page: CMAP2 busy");
3349 PT_SET_MA(sysmaps->CADDR2, PG_V | PG_RW | xpmap_ptom(VM_PAGE_TO_PHYS(m)) | PG_A | PG_M);
3351 if (off == 0 && size == PAGE_SIZE)
3352 pagezero(sysmaps->CADDR2);
3354 bzero((char *)sysmaps->CADDR2 + off, size);
3355 PT_SET_MA(sysmaps->CADDR2, 0);
3357 mtx_unlock(&sysmaps->lock);
3361 * pmap_zero_page_idle zeros the specified hardware page by mapping
3362 * the page into KVM and using bzero to clear its contents. This
3363 * is intended to be called from the vm_pagezero process only and
3367 pmap_zero_page_idle(vm_page_t m)
3371 panic("pmap_zero_page: CMAP3 busy");
3373 PT_SET_MA(CADDR3, PG_V | PG_RW | xpmap_ptom(VM_PAGE_TO_PHYS(m)) | PG_A | PG_M);
3375 PT_SET_MA(CADDR3, 0);
3380 * pmap_copy_page copies the specified (machine independent)
3381 * page by mapping the page into virtual memory and using
3382 * bcopy to copy the page, one machine dependent page at a
3386 pmap_copy_page(vm_page_t src, vm_page_t dst)
3388 struct sysmaps *sysmaps;
3390 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
3391 mtx_lock(&sysmaps->lock);
3392 if (*sysmaps->CMAP1)
3393 panic("pmap_copy_page: CMAP1 busy");
3394 if (*sysmaps->CMAP2)
3395 panic("pmap_copy_page: CMAP2 busy");
3397 PT_SET_MA(sysmaps->CADDR1, PG_V | xpmap_ptom(VM_PAGE_TO_PHYS(src)) | PG_A);
3398 PT_SET_MA(sysmaps->CADDR2, PG_V | PG_RW | xpmap_ptom(VM_PAGE_TO_PHYS(dst)) | PG_A | PG_M);
3399 bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE);
3400 PT_SET_MA(sysmaps->CADDR1, 0);
3401 PT_SET_MA(sysmaps->CADDR2, 0);
3403 mtx_unlock(&sysmaps->lock);
3407 * Returns true if the pmap's pv is one of the first
3408 * 16 pvs linked to from this page. This count may
3409 * be changed upwards or downwards in the future; it
3410 * is only necessary that true be returned for a small
3411 * subset of pmaps for proper page aging.
3414 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3419 if (m->flags & PG_FICTITIOUS)
3422 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3423 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3424 if (PV_PMAP(pv) == pmap) {
3435 * pmap_page_wired_mappings:
3437 * Return the number of managed mappings to the given physical page
3441 pmap_page_wired_mappings(vm_page_t m)
3449 if ((m->flags & PG_FICTITIOUS) != 0)
3451 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3453 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3456 pte = pmap_pte_quick(pmap, pv->pv_va);
3457 if ((*pte & PG_W) != 0)
3466 * Returns TRUE if the given page is mapped individually or as part of
3467 * a 4mpage. Otherwise, returns FALSE.
3470 pmap_page_is_mapped(vm_page_t m)
3472 struct md_page *pvh;
3474 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
3476 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3477 if (TAILQ_EMPTY(&m->md.pv_list)) {
3478 pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
3479 return (!TAILQ_EMPTY(&pvh->pv_list));
3485 * Remove all pages from specified address space
3486 * this aids process exit speeds. Also, this code
3487 * is special cased for current process only, but
3488 * can have the more generic (and slightly slower)
3489 * mode enabled. This is much faster than pmap_remove
3490 * in the case of running down an entire address space.
3493 pmap_remove_pages(pmap_t pmap)
3495 pt_entry_t *pte, tpte;
3496 vm_page_t m, free = NULL;
3498 struct pv_chunk *pc, *npc;
3501 uint32_t inuse, bitmask;
3504 CTR1(KTR_PMAP, "pmap_remove_pages: pmap=%p", pmap);
3506 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) {
3507 printf("warning: pmap_remove_pages called with non-current pmap\n");
3510 vm_page_lock_queues();
3511 KASSERT(pmap_is_current(pmap), ("removing pages from non-current pmap"));
3514 TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) {
3516 for (field = 0; field < _NPCM; field++) {
3517 inuse = (~(pc->pc_map[field])) & pc_freemask[field];
3518 while (inuse != 0) {
3520 bitmask = 1UL << bit;
3521 idx = field * 32 + bit;
3522 pv = &pc->pc_pventry[idx];
3525 pte = vtopte(pv->pv_va);
3526 tpte = *pte ? xpmap_mtop(*pte) : 0;
3530 "TPTE at %p IS ZERO @ VA %08x\n",
3536 * We cannot remove wired pages from a process' mapping at this time
3543 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3544 KASSERT(m->phys_addr == (tpte & PG_FRAME),
3545 ("vm_page_t %p phys_addr mismatch %016jx %016jx",
3546 m, (uintmax_t)m->phys_addr,
3549 KASSERT(m < &vm_page_array[vm_page_array_size],
3550 ("pmap_remove_pages: bad tpte %#jx",
3554 PT_CLEAR_VA(pte, FALSE);
3557 * Update the vm_page_t clean/reference bits.
3562 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3563 if (TAILQ_EMPTY(&m->md.pv_list))
3564 vm_page_flag_clear(m, PG_WRITEABLE);
3566 pmap_unuse_pt(pmap, pv->pv_va, &free);
3569 PV_STAT(pv_entry_frees++);
3570 PV_STAT(pv_entry_spare++);
3572 pc->pc_map[field] |= bitmask;
3573 pmap->pm_stats.resident_count--;
3578 PV_STAT(pv_entry_spare -= _NPCPV);
3579 PV_STAT(pc_chunk_count--);
3580 PV_STAT(pc_chunk_frees++);
3581 TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
3582 m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc));
3583 pmap_qremove((vm_offset_t)pc, 1);
3584 vm_page_unwire(m, 0);
3586 pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc);
3591 PT_SET_MA(PADDR1, 0);
3594 pmap_invalidate_all(pmap);
3595 vm_page_unlock_queues();
3597 pmap_free_zero_pages(free);
3603 * Return whether or not the specified physical page was modified
3604 * in any physical maps.
3607 pmap_is_modified(vm_page_t m)
3615 if (m->flags & PG_FICTITIOUS)
3619 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3620 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3623 pte = pmap_pte_quick(pmap, pv->pv_va);
3624 rv = (*pte & PG_M) != 0;
3630 PT_SET_MA(PADDR1, 0);
3636 * pmap_is_prefaultable:
3638 * Return whether or not the specified virtual address is elgible
3642 pmap_is_prefaultable_locked(pmap_t pmap, vm_offset_t addr)
3645 boolean_t rv = FALSE;
3649 if (pmap_is_current(pmap) && *pmap_pde(pmap, addr)) {
3657 pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr)
3662 rv = pmap_is_prefaultable_locked(pmap, addr);
3668 pmap_map_readonly(pmap_t pmap, vm_offset_t va, int len)
3670 int i, npages = round_page(len) >> PAGE_SHIFT;
3671 for (i = 0; i < npages; i++) {
3673 pte = pmap_pte(pmap, (vm_offset_t)(va + i*PAGE_SIZE));
3674 pte_store(pte, xpmap_mtop(*pte & ~(PG_RW|PG_M)));
3675 PMAP_MARK_PRIV(xpmap_mtop(*pte));
3676 pmap_pte_release(pte);
3681 pmap_map_readwrite(pmap_t pmap, vm_offset_t va, int len)
3683 int i, npages = round_page(len) >> PAGE_SHIFT;
3684 for (i = 0; i < npages; i++) {
3686 pte = pmap_pte(pmap, (vm_offset_t)(va + i*PAGE_SIZE));
3687 PMAP_MARK_UNPRIV(xpmap_mtop(*pte));
3688 pte_store(pte, xpmap_mtop(*pte) | (PG_RW|PG_M));
3689 pmap_pte_release(pte);
3694 * Clear the write and modified bits in each of the given page's mappings.
3697 pmap_remove_write(vm_page_t m)
3701 pt_entry_t oldpte, *pte;
3703 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3704 if ((m->flags & PG_FICTITIOUS) != 0 ||
3705 (m->flags & PG_WRITEABLE) == 0)
3708 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3711 pte = pmap_pte_quick(pmap, pv->pv_va);
3714 if ((oldpte & PG_RW) != 0) {
3716 * Regardless of whether a pte is 32 or 64 bits
3717 * in size, PG_RW and PG_M are among the least
3718 * significant 32 bits.
3720 if (!atomic_cmpset_int((u_int *)pte, oldpte,
3721 oldpte & ~(PG_RW | PG_M)))
3723 if ((oldpte & PG_M) != 0)
3725 pmap_invalidate_page(pmap, pv->pv_va);
3729 vm_page_flag_clear(m, PG_WRITEABLE);
3732 PT_SET_MA(PADDR1, 0);
3737 * pmap_ts_referenced:
3739 * Return a count of reference bits for a page, clearing those bits.
3740 * It is not necessary for every reference bit to be cleared, but it
3741 * is necessary that 0 only be returned when there are truly no
3742 * reference bits set.
3744 * XXX: The exact number of bits to check and clear is a matter that
3745 * should be tested and standardized at some point in the future for
3746 * optimal aging of shared pages.
3749 pmap_ts_referenced(vm_page_t m)
3751 pv_entry_t pv, pvf, pvn;
3756 if (m->flags & PG_FICTITIOUS)
3759 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3760 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3763 pvn = TAILQ_NEXT(pv, pv_list);
3764 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3765 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3768 pte = pmap_pte_quick(pmap, pv->pv_va);
3769 if ((*pte & PG_A) != 0) {
3770 PT_SET_VA_MA(pte, *pte & ~PG_A, FALSE);
3771 pmap_invalidate_page(pmap, pv->pv_va);
3777 } while ((pv = pvn) != NULL && pv != pvf);
3781 PT_SET_MA(PADDR1, 0);
3788 * Clear the modify bits on the specified physical page.
3791 pmap_clear_modify(vm_page_t m)
3797 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3798 if ((m->flags & PG_FICTITIOUS) != 0)
3801 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3804 pte = pmap_pte_quick(pmap, pv->pv_va);
3805 if ((*pte & PG_M) != 0) {
3807 * Regardless of whether a pte is 32 or 64 bits
3808 * in size, PG_M is among the least significant
3811 PT_SET_VA_MA(pte, *pte & ~PG_M, FALSE);
3812 pmap_invalidate_page(pmap, pv->pv_va);
3820 * pmap_clear_reference:
3822 * Clear the reference bit on the specified physical page.
3825 pmap_clear_reference(vm_page_t m)
3831 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3832 if ((m->flags & PG_FICTITIOUS) != 0)
3835 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3838 pte = pmap_pte_quick(pmap, pv->pv_va);
3839 if ((*pte & PG_A) != 0) {
3841 * Regardless of whether a pte is 32 or 64 bits
3842 * in size, PG_A is among the least significant
3845 PT_SET_VA_MA(pte, *pte & ~PG_A, FALSE);
3846 pmap_invalidate_page(pmap, pv->pv_va);
3854 * Miscellaneous support routines follow
3858 * Map a set of physical memory pages into the kernel virtual
3859 * address space. Return a pointer to where it is mapped. This
3860 * routine is intended to be used for mapping device memory,
3864 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
3866 vm_offset_t va, tmpva, offset;
3868 offset = pa & PAGE_MASK;
3869 size = roundup(offset + size, PAGE_SIZE);
3872 if (pa < KERNLOAD && pa + size <= KERNLOAD)
3875 va = kmem_alloc_nofault(kernel_map, size);
3877 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3879 for (tmpva = va; size > 0; ) {
3880 pmap_kenter_attr(tmpva, pa, mode);
3885 pmap_invalidate_range(kernel_pmap, va, tmpva);
3886 pmap_invalidate_cache();
3887 return ((void *)(va + offset));
3891 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3894 return (pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
3898 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
3901 return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
3905 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3907 vm_offset_t base, offset, tmpva;
3909 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD)
3911 base = trunc_page(va);
3912 offset = va & PAGE_MASK;
3913 size = roundup(offset + size, PAGE_SIZE);
3915 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE)
3916 pmap_kremove(tmpva);
3917 pmap_invalidate_range(kernel_pmap, va, tmpva);
3919 kmem_free(kernel_map, base, size);
3923 pmap_change_attr(va, size, mode)
3928 vm_offset_t base, offset, tmpva;
3933 base = trunc_page(va);
3934 offset = va & PAGE_MASK;
3935 size = roundup(offset + size, PAGE_SIZE);
3937 /* Only supported on kernel virtual addresses. */
3938 if (base <= VM_MAXUSER_ADDRESS)
3941 /* 4MB pages and pages that aren't mapped aren't supported. */
3942 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) {
3943 pde = pmap_pde(kernel_pmap, tmpva);
3946 if ((*pde & PG_V) == 0)
3949 if ((*pte & PG_V) == 0)
3954 * Ok, all the pages exist and are 4k, so run through them updating
3957 for (tmpva = base; size > 0; ) {
3958 pte = vtopte(tmpva);
3961 * The cache mode bits are all in the low 32-bits of the
3962 * PTE, so we can just spin on updating the low 32-bits.
3965 opte = *(u_int *)pte;
3966 npte = opte & ~(PG_PTE_PAT | PG_NC_PCD | PG_NC_PWT);
3967 npte |= pmap_cache_bits(mode, 0);
3968 PT_SET_VA_MA(pte, npte, TRUE);
3969 } while (npte != opte && (*pte != npte));
3975 * Flush CPU caches to make sure any data isn't cached that shouldn't
3978 pmap_invalidate_range(kernel_pmap, base, tmpva);
3979 pmap_invalidate_cache();
3984 * perform the pmap work for mincore
3987 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3989 pt_entry_t *ptep, pte;
3994 ptep = pmap_pte(pmap, addr);
3995 pte = (ptep != NULL) ? PT_GET(ptep) : 0;
3996 pmap_pte_release(ptep);
4002 val = MINCORE_INCORE;
4003 if ((pte & PG_MANAGED) == 0)
4006 pa = pte & PG_FRAME;
4008 m = PHYS_TO_VM_PAGE(pa);
4014 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
4017 * Modified by someone else
4019 vm_page_lock_queues();
4020 if (m->dirty || pmap_is_modified(m))
4021 val |= MINCORE_MODIFIED_OTHER;
4022 vm_page_unlock_queues();
4028 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
4031 * Referenced by someone else
4033 vm_page_lock_queues();
4034 if ((m->flags & PG_REFERENCED) ||
4035 pmap_ts_referenced(m)) {
4036 val |= MINCORE_REFERENCED_OTHER;
4037 vm_page_flag_set(m, PG_REFERENCED);
4039 vm_page_unlock_queues();
4046 pmap_activate(struct thread *td)
4048 pmap_t pmap, oldpmap;
4052 pmap = vmspace_pmap(td->td_proc->p_vmspace);
4053 oldpmap = PCPU_GET(curpmap);
4055 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask));
4056 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask));
4058 oldpmap->pm_active &= ~1;
4059 pmap->pm_active |= 1;
4062 cr3 = vtophys(pmap->pm_pdpt);
4064 cr3 = vtophys(pmap->pm_pdir);
4067 * pmap_activate is for the current thread on the current cpu
4069 td->td_pcb->pcb_cr3 = cr3;
4073 PCPU_SET(curpmap, pmap);
4078 * Increase the starting virtual address of the given mapping if a
4079 * different alignment might result in more superpage mappings.
4082 pmap_align_superpage(vm_object_t object, vm_ooffset_t offset,
4083 vm_offset_t *addr, vm_size_t size)
4085 vm_offset_t superpage_offset;
4089 if (object != NULL && (object->flags & OBJ_COLORED) != 0)
4090 offset += ptoa(object->pg_color);
4091 superpage_offset = offset & PDRMASK;
4092 if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR ||
4093 (*addr & PDRMASK) == superpage_offset)
4095 if ((*addr & PDRMASK) < superpage_offset)
4096 *addr = (*addr & ~PDRMASK) + superpage_offset;
4098 *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset;
4101 #if defined(PMAP_DEBUG)
4102 pmap_pid_dump(int pid)
4109 sx_slock(&allproc_lock);
4110 FOREACH_PROC_IN_SYSTEM(p) {
4111 if (p->p_pid != pid)
4117 pmap = vmspace_pmap(p->p_vmspace);
4118 for (i = 0; i < NPDEPTD; i++) {
4121 vm_offset_t base = i << PDRSHIFT;
4123 pde = &pmap->pm_pdir[i];
4124 if (pde && pmap_pde_v(pde)) {
4125 for (j = 0; j < NPTEPG; j++) {
4126 vm_offset_t va = base + (j << PAGE_SHIFT);
4127 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) {
4132 sx_sunlock(&allproc_lock);
4135 pte = pmap_pte(pmap, va);
4136 if (pte && pmap_pte_v(pte)) {
4140 m = PHYS_TO_VM_PAGE(pa & PG_FRAME);
4141 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x",
4142 va, pa, m->hold_count, m->wire_count, m->flags);
4157 sx_sunlock(&allproc_lock);
4164 static void pads(pmap_t pm);
4165 void pmap_pvdump(vm_paddr_t pa);
4167 /* print address space of pmap*/
4175 if (pm == kernel_pmap)
4177 for (i = 0; i < NPDEPTD; i++)
4179 for (j = 0; j < NPTEPG; j++) {
4180 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
4181 if (pm == kernel_pmap && va < KERNBASE)
4183 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
4185 ptep = pmap_pte(pm, va);
4186 if (pmap_pte_v(ptep))
4187 printf("%x:%x ", va, *ptep);
4193 pmap_pvdump(vm_paddr_t pa)
4199 printf("pa %x", pa);
4200 m = PHYS_TO_VM_PAGE(pa);
4201 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4203 printf(" -> pmap %p, va %x", (void *)pmap, pv->pv_va);