kernel - pc64 - Fix pv_free() race and add assertions
[dragonfly.git] / sys / platform / pc64 / x86_64 / pmap.c
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * Copyright (c) 1994 John S. Dyson
4  * Copyright (c) 1994 David Greenman
5  * Copyright (c) 2003 Peter Wemm
6  * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7  * Copyright (c) 2008, 2009 The DragonFly Project.
8  * Copyright (c) 2008, 2009 Jordan Gordeev.
9  * Copyright (c) 2011-2012 Matthew Dillon
10  * All rights reserved.
11  *
12  * This code is derived from software contributed to Berkeley by
13  * the Systems Programming Group of the University of Utah Computer
14  * Science Department and William Jolitz of UUNET Technologies Inc.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 3. All advertising materials mentioning features or use of this software
25  *    must display the following acknowledgement:
26  *      This product includes software developed by the University of
27  *      California, Berkeley and its contributors.
28  * 4. Neither the name of the University nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  */
44 /*
45  * Manage physical address maps for x86-64 systems.
46  */
47
48 #if JG
49 #include "opt_disable_pse.h"
50 #include "opt_pmap.h"
51 #endif
52 #include "opt_msgbuf.h"
53
54 #include <sys/param.h>
55 #include <sys/kernel.h>
56 #include <sys/proc.h>
57 #include <sys/msgbuf.h>
58 #include <sys/vmmeter.h>
59 #include <sys/mman.h>
60 #include <sys/systm.h>
61
62 #include <vm/vm.h>
63 #include <vm/vm_param.h>
64 #include <sys/sysctl.h>
65 #include <sys/lock.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>
74
75 #include <sys/user.h>
76 #include <sys/thread2.h>
77 #include <sys/sysref2.h>
78 #include <sys/spinlock2.h>
79 #include <vm/vm_page2.h>
80
81 #include <machine/cputypes.h>
82 #include <machine/md_var.h>
83 #include <machine/specialreg.h>
84 #include <machine/smp.h>
85 #include <machine_base/apic/apicreg.h>
86 #include <machine/globaldata.h>
87 #include <machine/pmap.h>
88 #include <machine/pmap_inval.h>
89 #include <machine/inttypes.h>
90
91 #include <ddb/ddb.h>
92
93 #define PMAP_KEEP_PDIRS
94 #ifndef PMAP_SHPGPERPROC
95 #define PMAP_SHPGPERPROC 2000
96 #endif
97
98 #if defined(DIAGNOSTIC)
99 #define PMAP_DIAGNOSTIC
100 #endif
101
102 #define MINPV 2048
103
104 /*
105  * pmap debugging will report who owns a pv lock when blocking.
106  */
107 #ifdef PMAP_DEBUG
108
109 #define PMAP_DEBUG_DECL         ,const char *func, int lineno
110 #define PMAP_DEBUG_ARGS         , __func__, __LINE__
111 #define PMAP_DEBUG_COPY         , func, lineno
112
113 #define pv_get(pmap, pindex)            _pv_get(pmap, pindex            \
114                                                         PMAP_DEBUG_ARGS)
115 #define pv_lock(pv)                     _pv_lock(pv                     \
116                                                         PMAP_DEBUG_ARGS)
117 #define pv_hold_try(pv)                 _pv_hold_try(pv                 \
118                                                         PMAP_DEBUG_ARGS)
119 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp  \
120                                                         PMAP_DEBUG_ARGS)
121
122 #else
123
124 #define PMAP_DEBUG_DECL
125 #define PMAP_DEBUG_ARGS
126 #define PMAP_DEBUG_COPY
127
128 #define pv_get(pmap, pindex)            _pv_get(pmap, pindex)
129 #define pv_lock(pv)                     _pv_lock(pv)
130 #define pv_hold_try(pv)                 _pv_hold_try(pv)
131 #define pv_alloc(pmap, pindex, isnewp)  _pv_alloc(pmap, pindex, isnewp)
132
133 #endif
134
135 /*
136  * Get PDEs and PTEs for user/kernel address space
137  */
138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
139
140 #define pmap_pde_v(pmap, pte)           ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
141 #define pmap_pte_w(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
142 #define pmap_pte_m(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
143 #define pmap_pte_u(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
144 #define pmap_pte_v(pmap, pte)           ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
145
146 /*
147  * Given a map and a machine independent protection code,
148  * convert to a vax protection code.
149  */
150 #define pte_prot(m, p)          \
151         (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
152 static int protection_codes[PROTECTION_CODES_SIZE];
153
154 struct pmap kernel_pmap;
155 static TAILQ_HEAD(,pmap)        pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
156
157 MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
158
159 vm_paddr_t avail_start;         /* PA of first available physical page */
160 vm_paddr_t avail_end;           /* PA of last available physical page */
161 vm_offset_t virtual2_start;     /* cutout free area prior to kernel start */
162 vm_offset_t virtual2_end;
163 vm_offset_t virtual_start;      /* VA of first avail page (after kernel bss) */
164 vm_offset_t virtual_end;        /* VA of last avail page (end of kernel AS) */
165 vm_offset_t KvaStart;           /* VA start of KVA space */
166 vm_offset_t KvaEnd;             /* VA end of KVA space (non-inclusive) */
167 vm_offset_t KvaSize;            /* max size of kernel virtual address space */
168 static boolean_t pmap_initialized = FALSE;      /* Has pmap_init completed? */
169 //static int pgeflag;           /* PG_G or-in */
170 //static int pseflag;           /* PG_PS or-in */
171 uint64_t PatMsr;
172
173 static int ndmpdp;
174 static vm_paddr_t dmaplimit;
175 static int nkpt;
176 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
177
178 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];        /* PAT -> PG_ bits */
179 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/    /* PAT -> PG_ bits */
180
181 static uint64_t KPTbase;
182 static uint64_t KPTphys;
183 static uint64_t KPDphys;        /* phys addr of kernel level 2 */
184 static uint64_t KPDbase;        /* phys addr of kernel level 2 @ KERNBASE */
185 uint64_t KPDPphys;      /* phys addr of kernel level 3 */
186 uint64_t KPML4phys;     /* phys addr of kernel level 4 */
187
188 static uint64_t DMPDphys;       /* phys addr of direct mapped level 2 */
189 static uint64_t DMPDPphys;      /* phys addr of direct mapped level 3 */
190
191 /*
192  * Data for the pv entry allocation mechanism
193  */
194 static vm_zone_t pvzone;
195 static struct vm_zone pvzone_store;
196 static struct vm_object pvzone_obj;
197 static int pv_entry_max=0, pv_entry_high_water=0;
198 static int pmap_pagedaemon_waken = 0;
199 static struct pv_entry *pvinit;
200
201 /*
202  * All those kernel PT submaps that BSD is so fond of
203  */
204 pt_entry_t *CMAP1 = NULL, *ptmmap;
205 caddr_t CADDR1 = NULL, ptvmmap = NULL;
206 static pt_entry_t *msgbufmap;
207 struct msgbuf *msgbufp=NULL;
208
209 /*
210  * PMAP default PG_* bits. Needed to be able to add
211  * EPT/NPT pagetable pmap_bits for the VMM module
212  */
213 uint64_t pmap_bits_default[] = {
214                 REGULAR_PMAP,                                   /* TYPE_IDX             0 */
215                 X86_PG_V,                                       /* PG_V_IDX             1 */
216                 X86_PG_RW,                                      /* PG_RW_IDX            2 */
217                 X86_PG_U,                                       /* PG_U_IDX             3 */
218                 X86_PG_A,                                       /* PG_A_IDX             4 */
219                 X86_PG_M,                                       /* PG_M_IDX             5 */
220                 X86_PG_PS,                                      /* PG_PS_IDX3           6 */
221                 X86_PG_G,                                       /* PG_G_IDX             7 */
222                 X86_PG_AVAIL1,                                  /* PG_AVAIL1_IDX        8 */
223                 X86_PG_AVAIL2,                                  /* PG_AVAIL2_IDX        9 */
224                 X86_PG_AVAIL3,                                  /* PG_AVAIL3_IDX        10 */
225                 X86_PG_NC_PWT | X86_PG_NC_PCD,                  /* PG_N_IDX     11 */
226 };
227 /*
228  * Crashdump maps.
229  */
230 static pt_entry_t *pt_crashdumpmap;
231 static caddr_t crashdumpmap;
232
233 static int pmap_yield_count = 64;
234 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
235     &pmap_yield_count, 0, "Yield during init_pt/release");
236 static int pmap_mmu_optimize = 0;
237 SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
238     &pmap_mmu_optimize, 0, "Share page table pages when possible");
239
240 #define DISABLE_PSE
241
242 /* Standard user access funtions */
243 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
244     size_t *lencopied);
245 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
246 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
247 extern int std_fubyte (const void *base);
248 extern int std_subyte (void *base, int byte);
249 extern long std_fuword (const void *base);
250 extern int std_suword (void *base, long word);
251 extern int std_suword32 (void *base, int word);
252
253 static void pv_hold(pv_entry_t pv);
254 static int _pv_hold_try(pv_entry_t pv
255                                 PMAP_DEBUG_DECL);
256 static void pv_drop(pv_entry_t pv);
257 static void _pv_lock(pv_entry_t pv
258                                 PMAP_DEBUG_DECL);
259 static void pv_unlock(pv_entry_t pv);
260 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
261                                 PMAP_DEBUG_DECL);
262 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
263                                 PMAP_DEBUG_DECL);
264 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
265 static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
266 static void pv_put(pv_entry_t pv);
267 static void pv_free(pv_entry_t pv);
268 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
269 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
270                       pv_entry_t *pvpp);
271 static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
272                       pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
273 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
274                       struct pmap_inval_info *info);
275 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
276 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp);
277
278 struct pmap_scan_info;
279 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
280                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
281                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
282 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
283                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
284                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
285
286 static void i386_protection_init (void);
287 static void create_pagetables(vm_paddr_t *firstaddr);
288 static void pmap_remove_all (vm_page_t m);
289 static boolean_t pmap_testbit (vm_page_t m, int bit);
290
291 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
292 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
293
294 static void pmap_pinit_defaults(struct pmap *pmap);
295
296 static unsigned pdir4mb;
297
298 static int
299 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
300 {
301         if (pv1->pv_pindex < pv2->pv_pindex)
302                 return(-1);
303         if (pv1->pv_pindex > pv2->pv_pindex)
304                 return(1);
305         return(0);
306 }
307
308 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
309              pv_entry_compare, vm_pindex_t, pv_pindex);
310
311 /*
312  * Move the kernel virtual free pointer to the next
313  * 2MB.  This is used to help improve performance
314  * by using a large (2MB) page for much of the kernel
315  * (.text, .data, .bss)
316  */
317 static
318 vm_offset_t
319 pmap_kmem_choose(vm_offset_t addr)
320 {
321         vm_offset_t newaddr = addr;
322
323         newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
324         return newaddr;
325 }
326
327 /*
328  * pmap_pte_quick:
329  *
330  *      Super fast pmap_pte routine best used when scanning the pv lists.
331  *      This eliminates many course-grained invltlb calls.  Note that many of
332  *      the pv list scans are across different pmaps and it is very wasteful
333  *      to do an entire invltlb when checking a single mapping.
334  */
335 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
336
337 static
338 pt_entry_t *
339 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
340 {
341         return pmap_pte(pmap, va);
342 }
343
344 /*
345  * Returns the pindex of a page table entry (representing a terminal page).
346  * There are NUPTE_TOTAL page table entries possible (a huge number)
347  *
348  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
349  * We want to properly translate negative KVAs.
350  */
351 static __inline
352 vm_pindex_t
353 pmap_pte_pindex(vm_offset_t va)
354 {
355         return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
356 }
357
358 /*
359  * Returns the pindex of a page table.
360  */
361 static __inline
362 vm_pindex_t
363 pmap_pt_pindex(vm_offset_t va)
364 {
365         return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
366 }
367
368 /*
369  * Returns the pindex of a page directory.
370  */
371 static __inline
372 vm_pindex_t
373 pmap_pd_pindex(vm_offset_t va)
374 {
375         return (NUPTE_TOTAL + NUPT_TOTAL +
376                 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
377 }
378
379 static __inline
380 vm_pindex_t
381 pmap_pdp_pindex(vm_offset_t va)
382 {
383         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
384                 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
385 }
386
387 static __inline
388 vm_pindex_t
389 pmap_pml4_pindex(void)
390 {
391         return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
392 }
393
394 /*
395  * Return various clipped indexes for a given VA
396  *
397  * Returns the index of a pte in a page table, representing a terminal
398  * page.
399  */
400 static __inline
401 vm_pindex_t
402 pmap_pte_index(vm_offset_t va)
403 {
404         return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
405 }
406
407 /*
408  * Returns the index of a pt in a page directory, representing a page
409  * table.
410  */
411 static __inline
412 vm_pindex_t
413 pmap_pt_index(vm_offset_t va)
414 {
415         return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
416 }
417
418 /*
419  * Returns the index of a pd in a page directory page, representing a page
420  * directory.
421  */
422 static __inline
423 vm_pindex_t
424 pmap_pd_index(vm_offset_t va)
425 {
426         return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
427 }
428
429 /*
430  * Returns the index of a pdp in the pml4 table, representing a page
431  * directory page.
432  */
433 static __inline
434 vm_pindex_t
435 pmap_pdp_index(vm_offset_t va)
436 {
437         return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
438 }
439
440 /*
441  * Generic procedure to index a pte from a pt, pd, or pdp.
442  *
443  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
444  *       a page table page index but is instead of PV lookup index.
445  */
446 static
447 void *
448 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
449 {
450         pt_entry_t *pte;
451
452         pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
453         return(&pte[pindex]);
454 }
455
456 /*
457  * Return pointer to PDP slot in the PML4
458  */
459 static __inline
460 pml4_entry_t *
461 pmap_pdp(pmap_t pmap, vm_offset_t va)
462 {
463         return (&pmap->pm_pml4[pmap_pdp_index(va)]);
464 }
465
466 /*
467  * Return pointer to PD slot in the PDP given a pointer to the PDP
468  */
469 static __inline
470 pdp_entry_t *
471 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
472 {
473         pdp_entry_t *pd;
474
475         pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
476         return (&pd[pmap_pd_index(va)]);
477 }
478
479 /*
480  * Return pointer to PD slot in the PDP.
481  */
482 static __inline
483 pdp_entry_t *
484 pmap_pd(pmap_t pmap, vm_offset_t va)
485 {
486         pml4_entry_t *pdp;
487
488         pdp = pmap_pdp(pmap, va);
489         if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
490                 return NULL;
491         return (pmap_pdp_to_pd(*pdp, va));
492 }
493
494 /*
495  * Return pointer to PT slot in the PD given a pointer to the PD
496  */
497 static __inline
498 pd_entry_t *
499 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
500 {
501         pd_entry_t *pt;
502
503         pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
504         return (&pt[pmap_pt_index(va)]);
505 }
506
507 /*
508  * Return pointer to PT slot in the PD
509  *
510  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
511  *                   so we cannot lookup the PD via the PDP.  Instead we
512  *                   must look it up via the pmap.
513  */
514 static __inline
515 pd_entry_t *
516 pmap_pt(pmap_t pmap, vm_offset_t va)
517 {
518         pdp_entry_t *pd;
519         pv_entry_t pv;
520         vm_pindex_t pd_pindex;
521
522         if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
523                 pd_pindex = pmap_pd_pindex(va);
524                 spin_lock(&pmap->pm_spin);
525                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
526                 spin_unlock(&pmap->pm_spin);
527                 if (pv == NULL || pv->pv_m == NULL)
528                         return NULL;
529                 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
530         } else {
531                 pd = pmap_pd(pmap, va);
532                 if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
533                          return NULL;
534                 return (pmap_pd_to_pt(*pd, va));
535         }
536 }
537
538 /*
539  * Return pointer to PTE slot in the PT given a pointer to the PT
540  */
541 static __inline
542 pt_entry_t *
543 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
544 {
545         pt_entry_t *pte;
546
547         pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
548         return (&pte[pmap_pte_index(va)]);
549 }
550
551 /*
552  * Return pointer to PTE slot in the PT
553  */
554 static __inline
555 pt_entry_t *
556 pmap_pte(pmap_t pmap, vm_offset_t va)
557 {
558         pd_entry_t *pt;
559
560         pt = pmap_pt(pmap, va);
561         if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
562                  return NULL;
563         if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
564                 return ((pt_entry_t *)pt);
565         return (pmap_pt_to_pte(*pt, va));
566 }
567
568 /*
569  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
570  * the PT layer.  This will speed up core pmap operations considerably.
571  *
572  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
573  *       must be in a known associated state (typically by being locked when
574  *       the pmap spinlock isn't held).  We allow the race for that case.
575  */
576 static __inline
577 void
578 pv_cache(pv_entry_t pv, vm_pindex_t pindex)
579 {
580         if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
581                 pv->pv_pmap->pm_pvhint = pv;
582 }
583
584
585 /*
586  * KVM - return address of PT slot in PD
587  */
588 static __inline
589 pd_entry_t *
590 vtopt(vm_offset_t va)
591 {
592         uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
593                                   NPML4EPGSHIFT)) - 1);
594
595         return (PDmap + ((va >> PDRSHIFT) & mask));
596 }
597
598 /*
599  * KVM - return address of PTE slot in PT
600  */
601 static __inline
602 pt_entry_t *
603 vtopte(vm_offset_t va)
604 {
605         uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
606                                   NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
607
608         return (PTmap + ((va >> PAGE_SHIFT) & mask));
609 }
610
611 static uint64_t
612 allocpages(vm_paddr_t *firstaddr, long n)
613 {
614         uint64_t ret;
615
616         ret = *firstaddr;
617         bzero((void *)ret, n * PAGE_SIZE);
618         *firstaddr += n * PAGE_SIZE;
619         return (ret);
620 }
621
622 static
623 void
624 create_pagetables(vm_paddr_t *firstaddr)
625 {
626         long i;         /* must be 64 bits */
627         long nkpt_base;
628         long nkpt_phys;
629         int j;
630
631         /*
632          * We are running (mostly) V=P at this point
633          *
634          * Calculate NKPT - number of kernel page tables.  We have to
635          * accomodoate prealloction of the vm_page_array, dump bitmap,
636          * MSGBUF_SIZE, and other stuff.  Be generous.
637          *
638          * Maxmem is in pages.
639          *
640          * ndmpdp is the number of 1GB pages we wish to map.
641          */
642         ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
643         if (ndmpdp < 4)         /* Minimum 4GB of dirmap */
644                 ndmpdp = 4;
645         KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
646
647         /*
648          * Starting at the beginning of kvm (not KERNBASE).
649          */
650         nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
651         nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
652         nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
653                        ndmpdp) + 511) / 512;
654         nkpt_phys += 128;
655
656         /*
657          * Starting at KERNBASE - map 2G worth of page table pages.
658          * KERNBASE is offset -2G from the end of kvm.
659          */
660         nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
661
662         /*
663          * Allocate pages
664          */
665         KPTbase = allocpages(firstaddr, nkpt_base);
666         KPTphys = allocpages(firstaddr, nkpt_phys);
667         KPML4phys = allocpages(firstaddr, 1);
668         KPDPphys = allocpages(firstaddr, NKPML4E);
669         KPDphys = allocpages(firstaddr, NKPDPE);
670
671         /*
672          * Calculate the page directory base for KERNBASE,
673          * that is where we start populating the page table pages.
674          * Basically this is the end - 2.
675          */
676         KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
677
678         DMPDPphys = allocpages(firstaddr, NDMPML4E);
679         if ((amd_feature & AMDID_PAGE1GB) == 0)
680                 DMPDphys = allocpages(firstaddr, ndmpdp);
681         dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
682
683         /*
684          * Fill in the underlying page table pages for the area around
685          * KERNBASE.  This remaps low physical memory to KERNBASE.
686          *
687          * Read-only from zero to physfree
688          * XXX not fully used, underneath 2M pages
689          */
690         for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
691                 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
692                 ((pt_entry_t *)KPTbase)[i] |=
693                     pmap_bits_default[PG_RW_IDX] |
694                     pmap_bits_default[PG_V_IDX] |
695                     pmap_bits_default[PG_G_IDX];
696         }
697
698         /*
699          * Now map the initial kernel page tables.  One block of page
700          * tables is placed at the beginning of kernel virtual memory,
701          * and another block is placed at KERNBASE to map the kernel binary,
702          * data, bss, and initial pre-allocations.
703          */
704         for (i = 0; i < nkpt_base; i++) {
705                 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
706                 ((pd_entry_t *)KPDbase)[i] |=
707                     pmap_bits_default[PG_RW_IDX] |
708                     pmap_bits_default[PG_V_IDX];
709         }
710         for (i = 0; i < nkpt_phys; i++) {
711                 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
712                 ((pd_entry_t *)KPDphys)[i] |=
713                     pmap_bits_default[PG_RW_IDX] |
714                     pmap_bits_default[PG_V_IDX];
715         }
716
717         /*
718          * Map from zero to end of allocations using 2M pages as an
719          * optimization.  This will bypass some of the KPTBase pages
720          * above in the KERNBASE area.
721          */
722         for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
723                 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
724                 ((pd_entry_t *)KPDbase)[i] |=
725                     pmap_bits_default[PG_RW_IDX] |
726                     pmap_bits_default[PG_V_IDX] |
727                     pmap_bits_default[PG_PS_IDX] |
728                     pmap_bits_default[PG_G_IDX];
729         }
730
731         /*
732          * And connect up the PD to the PDP.  The kernel pmap is expected
733          * to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
734          */
735         for (i = 0; i < NKPDPE; i++) {
736                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
737                                 KPDphys + (i << PAGE_SHIFT);
738                 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
739                     pmap_bits_default[PG_RW_IDX] |
740                     pmap_bits_default[PG_V_IDX] |
741                     pmap_bits_default[PG_U_IDX];
742         }
743
744         /*
745          * Now set up the direct map space using either 2MB or 1GB pages
746          * Preset PG_M and PG_A because demotion expects it.
747          *
748          * When filling in entries in the PD pages make sure any excess
749          * entries are set to zero as we allocated enough PD pages
750          */
751         if ((amd_feature & AMDID_PAGE1GB) == 0) {
752                 for (i = 0; i < NPDEPG * ndmpdp; i++) {
753                         ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
754                         ((pd_entry_t *)DMPDphys)[i] |=
755                             pmap_bits_default[PG_RW_IDX] |
756                             pmap_bits_default[PG_V_IDX] |
757                             pmap_bits_default[PG_PS_IDX] |
758                             pmap_bits_default[PG_G_IDX] |
759                             pmap_bits_default[PG_M_IDX] |
760                             pmap_bits_default[PG_A_IDX];
761                 }
762
763                 /*
764                  * And the direct map space's PDP
765                  */
766                 for (i = 0; i < ndmpdp; i++) {
767                         ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
768                                                         (i << PAGE_SHIFT);
769                         ((pdp_entry_t *)DMPDPphys)[i] |=
770                             pmap_bits_default[PG_RW_IDX] |
771                             pmap_bits_default[PG_V_IDX] |
772                             pmap_bits_default[PG_U_IDX];
773                 }
774         } else {
775                 for (i = 0; i < ndmpdp; i++) {
776                         ((pdp_entry_t *)DMPDPphys)[i] =
777                                                 (vm_paddr_t)i << PDPSHIFT;
778                         ((pdp_entry_t *)DMPDPphys)[i] |=
779                             pmap_bits_default[PG_RW_IDX] |
780                             pmap_bits_default[PG_V_IDX] |
781                             pmap_bits_default[PG_PS_IDX] |
782                             pmap_bits_default[PG_G_IDX] |
783                             pmap_bits_default[PG_M_IDX] |
784                             pmap_bits_default[PG_A_IDX];
785                 }
786         }
787
788         /* And recursively map PML4 to itself in order to get PTmap */
789         ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
790         ((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
791             pmap_bits_default[PG_RW_IDX] |
792             pmap_bits_default[PG_V_IDX] |
793             pmap_bits_default[PG_U_IDX];
794
795         /*
796          * Connect the Direct Map slots up to the PML4
797          */
798         for (j = 0; j < NDMPML4E; ++j) {
799                 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
800                     (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
801                     pmap_bits_default[PG_RW_IDX] |
802                     pmap_bits_default[PG_V_IDX] |
803                     pmap_bits_default[PG_U_IDX];
804         }
805
806         /*
807          * Connect the KVA slot up to the PML4
808          */
809         ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
810         ((pdp_entry_t *)KPML4phys)[KPML4I] |=
811             pmap_bits_default[PG_RW_IDX] |
812             pmap_bits_default[PG_V_IDX] |
813             pmap_bits_default[PG_U_IDX];
814 }
815
816 /*
817  *      Bootstrap the system enough to run with virtual memory.
818  *
819  *      On the i386 this is called after mapping has already been enabled
820  *      and just syncs the pmap module with what has already been done.
821  *      [We can't call it easily with mapping off since the kernel is not
822  *      mapped with PA == VA, hence we would have to relocate every address
823  *      from the linked base (virtual) address "KERNBASE" to the actual
824  *      (physical) address starting relative to 0]
825  */
826 void
827 pmap_bootstrap(vm_paddr_t *firstaddr)
828 {
829         vm_offset_t va;
830         pt_entry_t *pte;
831
832         KvaStart = VM_MIN_KERNEL_ADDRESS;
833         KvaEnd = VM_MAX_KERNEL_ADDRESS;
834         KvaSize = KvaEnd - KvaStart;
835
836         avail_start = *firstaddr;
837
838         /*
839          * Create an initial set of page tables to run the kernel in.
840          */
841         create_pagetables(firstaddr);
842
843         virtual2_start = KvaStart;
844         virtual2_end = PTOV_OFFSET;
845
846         virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
847         virtual_start = pmap_kmem_choose(virtual_start);
848
849         virtual_end = VM_MAX_KERNEL_ADDRESS;
850
851         /* XXX do %cr0 as well */
852         load_cr4(rcr4() | CR4_PGE | CR4_PSE);
853         load_cr3(KPML4phys);
854
855         /*
856          * Initialize protection array.
857          */
858         i386_protection_init();
859
860         /*
861          * The kernel's pmap is statically allocated so we don't have to use
862          * pmap_create, which is unlikely to work correctly at this part of
863          * the boot sequence (XXX and which no longer exists).
864          */
865         kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
866         kernel_pmap.pm_count = 1;
867         kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
868         RB_INIT(&kernel_pmap.pm_pvroot);
869         spin_init(&kernel_pmap.pm_spin);
870         lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
871
872         /*
873          * Reserve some special page table entries/VA space for temporary
874          * mapping of pages.
875          */
876 #define SYSMAP(c, p, v, n)      \
877         v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
878
879         va = virtual_start;
880         pte = vtopte(va);
881
882         /*
883          * CMAP1/CMAP2 are used for zeroing and copying pages.
884          */
885         SYSMAP(caddr_t, CMAP1, CADDR1, 1)
886
887         /*
888          * Crashdump maps.
889          */
890         SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
891
892         /*
893          * ptvmmap is used for reading arbitrary physical pages via
894          * /dev/mem.
895          */
896         SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
897
898         /*
899          * msgbufp is used to map the system message buffer.
900          * XXX msgbufmap is not used.
901          */
902         SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
903                atop(round_page(MSGBUF_SIZE)))
904
905         virtual_start = va;
906
907         *CMAP1 = 0;
908
909         /*
910          * PG_G is terribly broken on SMP because we IPI invltlb's in some
911          * cases rather then invl1pg.  Actually, I don't even know why it
912          * works under UP because self-referential page table mappings
913          */
914 //      pgeflag = 0;
915
916 /*
917  * Initialize the 4MB page size flag
918  */
919 //      pseflag = 0;
920 /*
921  * The 4MB page version of the initial
922  * kernel page mapping.
923  */
924         pdir4mb = 0;
925
926 #if !defined(DISABLE_PSE)
927         if (cpu_feature & CPUID_PSE) {
928                 pt_entry_t ptditmp;
929                 /*
930                  * Note that we have enabled PSE mode
931                  */
932 //              pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
933                 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
934                 ptditmp &= ~(NBPDR - 1);
935                 ptditmp |= pmap_bits_default[PG_V_IDX] |
936                     pmap_bits_default[PG_RW_IDX] |
937                     pmap_bits_default[PG_PS_IDX] |
938                     pmap_bits_default[PG_U_IDX];
939 //                  pgeflag;
940                 pdir4mb = ptditmp;
941         }
942 #endif
943         cpu_invltlb();
944
945         /* Initialize the PAT MSR */
946         pmap_init_pat();
947
948         pmap_pinit_defaults(&kernel_pmap);
949 }
950
951 /*
952  * Setup the PAT MSR.
953  */
954 void
955 pmap_init_pat(void)
956 {
957         uint64_t pat_msr;
958         u_long cr0, cr4;
959
960         /*
961          * Default values mapping PATi,PCD,PWT bits at system reset.
962          * The default values effectively ignore the PATi bit by
963          * repeating the encodings for 0-3 in 4-7, and map the PCD
964          * and PWT bit combinations to the expected PAT types.
965          */
966         pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |        /* 000 */
967                   PAT_VALUE(1, PAT_WRITE_THROUGH) |     /* 001 */
968                   PAT_VALUE(2, PAT_UNCACHED) |          /* 010 */
969                   PAT_VALUE(3, PAT_UNCACHEABLE) |       /* 011 */
970                   PAT_VALUE(4, PAT_WRITE_BACK) |        /* 100 */
971                   PAT_VALUE(5, PAT_WRITE_THROUGH) |     /* 101 */
972                   PAT_VALUE(6, PAT_UNCACHED) |          /* 110 */
973                   PAT_VALUE(7, PAT_UNCACHEABLE);        /* 111 */
974         pat_pte_index[PAT_WRITE_BACK]   = 0;
975         pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
976         pat_pte_index[PAT_UNCACHED]     = X86_PG_NC_PCD;
977         pat_pte_index[PAT_UNCACHEABLE]  = X86_PG_NC_PCD | X86_PG_NC_PWT;
978         pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
979         pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
980
981         if (cpu_feature & CPUID_PAT) {
982                 /*
983                  * If we support the PAT then set-up entries for
984                  * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
985                  * 4 and 5.
986                  */
987                 pat_msr = (pat_msr & ~PAT_MASK(4)) |
988                           PAT_VALUE(4, PAT_WRITE_PROTECTED);
989                 pat_msr = (pat_msr & ~PAT_MASK(5)) |
990                           PAT_VALUE(5, PAT_WRITE_COMBINING);
991                 pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | 0;
992                 pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
993
994                 /*
995                  * Then enable the PAT
996                  */
997
998                 /* Disable PGE. */
999                 cr4 = rcr4();
1000                 load_cr4(cr4 & ~CR4_PGE);
1001
1002                 /* Disable caches (CD = 1, NW = 0). */
1003                 cr0 = rcr0();
1004                 load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1005
1006                 /* Flushes caches and TLBs. */
1007                 wbinvd();
1008                 cpu_invltlb();
1009
1010                 /* Update PAT and index table. */
1011                 wrmsr(MSR_PAT, pat_msr);
1012
1013                 /* Flush caches and TLBs again. */
1014                 wbinvd();
1015                 cpu_invltlb();
1016
1017                 /* Restore caches and PGE. */
1018                 load_cr0(cr0);
1019                 load_cr4(cr4);
1020                 PatMsr = pat_msr;
1021         }
1022 }
1023
1024 /*
1025  * Set 4mb pdir for mp startup
1026  */
1027 void
1028 pmap_set_opt(void)
1029 {
1030         if (cpu_feature & CPUID_PSE) {
1031                 load_cr4(rcr4() | CR4_PSE);
1032                 if (pdir4mb && mycpu->gd_cpuid == 0) {  /* only on BSP */
1033                         cpu_invltlb();
1034                 }
1035         }
1036 }
1037
1038 /*
1039  *      Initialize the pmap module.
1040  *      Called by vm_init, to initialize any structures that the pmap
1041  *      system needs to map virtual memory.
1042  *      pmap_init has been enhanced to support in a fairly consistant
1043  *      way, discontiguous physical memory.
1044  */
1045 void
1046 pmap_init(void)
1047 {
1048         int i;
1049         int initial_pvs;
1050
1051         /*
1052          * Allocate memory for random pmap data structures.  Includes the
1053          * pv_head_table.
1054          */
1055
1056         for (i = 0; i < vm_page_array_size; i++) {
1057                 vm_page_t m;
1058
1059                 m = &vm_page_array[i];
1060                 TAILQ_INIT(&m->md.pv_list);
1061         }
1062
1063         /*
1064          * init the pv free list
1065          */
1066         initial_pvs = vm_page_array_size;
1067         if (initial_pvs < MINPV)
1068                 initial_pvs = MINPV;
1069         pvzone = &pvzone_store;
1070         pvinit = (void *)kmem_alloc(&kernel_map,
1071                                     initial_pvs * sizeof (struct pv_entry));
1072         zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1073                   pvinit, initial_pvs);
1074
1075         /*
1076          * Now it is safe to enable pv_table recording.
1077          */
1078         pmap_initialized = TRUE;
1079 }
1080
1081 /*
1082  * Initialize the address space (zone) for the pv_entries.  Set a
1083  * high water mark so that the system can recover from excessive
1084  * numbers of pv entries.
1085  */
1086 void
1087 pmap_init2(void)
1088 {
1089         int shpgperproc = PMAP_SHPGPERPROC;
1090         int entry_max;
1091
1092         TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
1093         pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
1094         TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
1095         pv_entry_high_water = 9 * (pv_entry_max / 10);
1096
1097         /*
1098          * Subtract out pages already installed in the zone (hack)
1099          */
1100         entry_max = pv_entry_max - vm_page_array_size;
1101         if (entry_max <= 0)
1102                 entry_max = 1;
1103
1104         zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
1105 }
1106
1107 /*
1108  * Typically used to initialize a fictitious page by vm/device_pager.c
1109  */
1110 void
1111 pmap_page_init(struct vm_page *m)
1112 {
1113         vm_page_init(m);
1114         TAILQ_INIT(&m->md.pv_list);
1115 }
1116
1117 /***************************************************
1118  * Low level helper routines.....
1119  ***************************************************/
1120
1121 /*
1122  * this routine defines the region(s) of memory that should
1123  * not be tested for the modified bit.
1124  */
1125 static __inline
1126 int
1127 pmap_track_modified(vm_pindex_t pindex)
1128 {
1129         vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
1130         if ((va < clean_sva) || (va >= clean_eva)) 
1131                 return 1;
1132         else
1133                 return 0;
1134 }
1135
1136 /*
1137  * Extract the physical page address associated with the map/VA pair.
1138  * The page must be wired for this to work reliably.
1139  *
1140  * XXX for the moment we're using pv_find() instead of pv_get(), as
1141  *     callers might be expecting non-blocking operation.
1142  */
1143 vm_paddr_t 
1144 pmap_extract(pmap_t pmap, vm_offset_t va)
1145 {
1146         vm_paddr_t rtval;
1147         pv_entry_t pt_pv;
1148         pt_entry_t *ptep;
1149
1150         rtval = 0;
1151         if (va >= VM_MAX_USER_ADDRESS) {
1152                 /*
1153                  * Kernel page directories might be direct-mapped and
1154                  * there is typically no PV tracking of pte's
1155                  */
1156                 pd_entry_t *pt;
1157
1158                 pt = pmap_pt(pmap, va);
1159                 if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1160                         if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1161                                 rtval = *pt & PG_PS_FRAME;
1162                                 rtval |= va & PDRMASK;
1163                         } else {
1164                                 ptep = pmap_pt_to_pte(*pt, va);
1165                                 if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1166                                         rtval = *ptep & PG_FRAME;
1167                                         rtval |= va & PAGE_MASK;
1168                                 }
1169                         }
1170                 }
1171         } else {
1172                 /*
1173                  * User pages currently do not direct-map the page directory
1174                  * and some pages might not used managed PVs.  But all PT's
1175                  * will have a PV.
1176                  */
1177                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1178                 if (pt_pv) {
1179                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1180                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1181                                 rtval = *ptep & PG_FRAME;
1182                                 rtval |= va & PAGE_MASK;
1183                         }
1184                         pv_drop(pt_pv);
1185                 }
1186         }
1187         return rtval;
1188 }
1189
1190 /*
1191  * Similar to extract but checks protections, SMP-friendly short-cut for
1192  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1193  * fall-through to the real fault code.
1194  *
1195  * The returned page, if not NULL, is held (and not busied).
1196  */
1197 vm_page_t
1198 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1199 {
1200         if (pmap && va < VM_MAX_USER_ADDRESS) {
1201                 pv_entry_t pt_pv;
1202                 pv_entry_t pte_pv;
1203                 pt_entry_t *ptep;
1204                 pt_entry_t req;
1205                 vm_page_t m;
1206                 int error;
1207
1208                 req = pmap->pmap_bits[PG_V_IDX] |
1209                       pmap->pmap_bits[PG_U_IDX];
1210                 if (prot & VM_PROT_WRITE)
1211                         req |= pmap->pmap_bits[PG_RW_IDX];
1212
1213                 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1214                 if (pt_pv == NULL)
1215                         return (NULL);
1216                 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1217                 if ((*ptep & req) != req) {
1218                         pv_drop(pt_pv);
1219                         return (NULL);
1220                 }
1221                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), &error);
1222                 if (pte_pv && error == 0) {
1223                         m = pte_pv->pv_m;
1224                         vm_page_hold(m);
1225                         if (prot & VM_PROT_WRITE)
1226                                 vm_page_dirty(m);
1227                         pv_put(pte_pv);
1228                 } else if (pte_pv) {
1229                         pv_drop(pte_pv);
1230                         m = NULL;
1231                 } else {
1232                         m = NULL;
1233                 }
1234                 pv_drop(pt_pv);
1235                 return(m);
1236         } else {
1237                 return(NULL);
1238         }
1239 }
1240
1241 /*
1242  * Extract the physical page address associated kernel virtual address.
1243  */
1244 vm_paddr_t
1245 pmap_kextract(vm_offset_t va)
1246 {
1247         pd_entry_t pt;          /* pt entry in pd */
1248         vm_paddr_t pa;
1249
1250         if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1251                 pa = DMAP_TO_PHYS(va);
1252         } else {
1253                 pt = *vtopt(va);
1254                 if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1255                         pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1256                 } else {
1257                         /*
1258                          * Beware of a concurrent promotion that changes the
1259                          * PDE at this point!  For example, vtopte() must not
1260                          * be used to access the PTE because it would use the
1261                          * new PDE.  It is, however, safe to use the old PDE
1262                          * because the page table page is preserved by the
1263                          * promotion.
1264                          */
1265                         pa = *pmap_pt_to_pte(pt, va);
1266                         pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1267                 }
1268         }
1269         return pa;
1270 }
1271
1272 /***************************************************
1273  * Low level mapping routines.....
1274  ***************************************************/
1275
1276 /*
1277  * Routine: pmap_kenter
1278  * Function:
1279  *      Add a wired page to the KVA
1280  *      NOTE! note that in order for the mapping to take effect -- you
1281  *      should do an invltlb after doing the pmap_kenter().
1282  */
1283 void 
1284 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1285 {
1286         pt_entry_t *pte;
1287         pt_entry_t npte;
1288         pmap_inval_info info;
1289
1290         pmap_inval_init(&info);                         /* XXX remove */
1291         npte = pa |
1292             kernel_pmap.pmap_bits[PG_RW_IDX] |
1293             kernel_pmap.pmap_bits[PG_V_IDX];
1294 //          pgeflag;
1295         pte = vtopte(va);
1296         pmap_inval_interlock(&info, &kernel_pmap, va);  /* XXX remove */
1297         *pte = npte;
1298         pmap_inval_deinterlock(&info, &kernel_pmap);    /* XXX remove */
1299         pmap_inval_done(&info);                         /* XXX remove */
1300 }
1301
1302 /*
1303  * Routine: pmap_kenter_quick
1304  * Function:
1305  *      Similar to pmap_kenter(), except we only invalidate the
1306  *      mapping on the current CPU.
1307  */
1308 void
1309 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1310 {
1311         pt_entry_t *pte;
1312         pt_entry_t npte;
1313
1314         npte = pa |
1315             kernel_pmap.pmap_bits[PG_RW_IDX] |
1316             kernel_pmap.pmap_bits[PG_V_IDX];
1317 //          pgeflag;
1318         pte = vtopte(va);
1319         *pte = npte;
1320         cpu_invlpg((void *)va);
1321 }
1322
1323 void
1324 pmap_kenter_sync(vm_offset_t va)
1325 {
1326         pmap_inval_info info;
1327
1328         pmap_inval_init(&info);
1329         pmap_inval_interlock(&info, &kernel_pmap, va);
1330         pmap_inval_deinterlock(&info, &kernel_pmap);
1331         pmap_inval_done(&info);
1332 }
1333
1334 void
1335 pmap_kenter_sync_quick(vm_offset_t va)
1336 {
1337         cpu_invlpg((void *)va);
1338 }
1339
1340 /*
1341  * remove a page from the kernel pagetables
1342  */
1343 void
1344 pmap_kremove(vm_offset_t va)
1345 {
1346         pt_entry_t *pte;
1347         pmap_inval_info info;
1348
1349         pmap_inval_init(&info);
1350         pte = vtopte(va);
1351         pmap_inval_interlock(&info, &kernel_pmap, va);
1352         (void)pte_load_clear(pte);
1353         pmap_inval_deinterlock(&info, &kernel_pmap);
1354         pmap_inval_done(&info);
1355 }
1356
1357 void
1358 pmap_kremove_quick(vm_offset_t va)
1359 {
1360         pt_entry_t *pte;
1361         pte = vtopte(va);
1362         (void)pte_load_clear(pte);
1363         cpu_invlpg((void *)va);
1364 }
1365
1366 /*
1367  * XXX these need to be recoded.  They are not used in any critical path.
1368  */
1369 void
1370 pmap_kmodify_rw(vm_offset_t va)
1371 {
1372         atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1373         cpu_invlpg((void *)va);
1374 }
1375
1376 /* NOT USED
1377 void
1378 pmap_kmodify_nc(vm_offset_t va)
1379 {
1380         atomic_set_long(vtopte(va), PG_N);
1381         cpu_invlpg((void *)va);
1382 }
1383 */
1384
1385 /*
1386  * Used to map a range of physical addresses into kernel virtual
1387  * address space during the low level boot, typically to map the
1388  * dump bitmap, message buffer, and vm_page_array.
1389  *
1390  * These mappings are typically made at some pointer after the end of the
1391  * kernel text+data.
1392  *
1393  * We could return PHYS_TO_DMAP(start) here and not allocate any
1394  * via (*virtp), but then kmem from userland and kernel dumps won't
1395  * have access to the related pointers.
1396  */
1397 vm_offset_t
1398 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1399 {
1400         vm_offset_t va;
1401         vm_offset_t va_start;
1402
1403         /*return PHYS_TO_DMAP(start);*/
1404
1405         va_start = *virtp;
1406         va = va_start;
1407
1408         while (start < end) {
1409                 pmap_kenter_quick(va, start);
1410                 va += PAGE_SIZE;
1411                 start += PAGE_SIZE;
1412         }
1413         *virtp = va;
1414         return va_start;
1415 }
1416
1417 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
1418
1419 /*
1420  * Remove the specified set of pages from the data and instruction caches.
1421  *
1422  * In contrast to pmap_invalidate_cache_range(), this function does not
1423  * rely on the CPU's self-snoop feature, because it is intended for use
1424  * when moving pages into a different cache domain.
1425  */
1426 void
1427 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
1428 {
1429         vm_offset_t daddr, eva;
1430         int i;
1431
1432         if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
1433             (cpu_feature & CPUID_CLFSH) == 0)
1434                 wbinvd();
1435         else {
1436                 cpu_mfence();
1437                 for (i = 0; i < count; i++) {
1438                         daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
1439                         eva = daddr + PAGE_SIZE;
1440                         for (; daddr < eva; daddr += cpu_clflush_line_size)
1441                                 clflush(daddr);
1442                 }
1443                 cpu_mfence();
1444         }
1445 }
1446
1447 void
1448 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
1449 {
1450         KASSERT((sva & PAGE_MASK) == 0,
1451             ("pmap_invalidate_cache_range: sva not page-aligned"));
1452         KASSERT((eva & PAGE_MASK) == 0,
1453             ("pmap_invalidate_cache_range: eva not page-aligned"));
1454
1455         if (cpu_feature & CPUID_SS) {
1456                 ; /* If "Self Snoop" is supported, do nothing. */
1457         } else {
1458                 /* Globally invalidate caches */
1459                 cpu_wbinvd_on_all_cpus();
1460         }
1461 }
1462 void
1463 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1464 {
1465         smp_invlpg_range(pmap->pm_active, sva, eva);
1466 }
1467
1468 /*
1469  * Add a list of wired pages to the kva
1470  * this routine is only used for temporary
1471  * kernel mappings that do not need to have
1472  * page modification or references recorded.
1473  * Note that old mappings are simply written
1474  * over.  The page *must* be wired.
1475  */
1476 void
1477 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1478 {
1479         vm_offset_t end_va;
1480
1481         end_va = va + count * PAGE_SIZE;
1482                 
1483         while (va < end_va) {
1484                 pt_entry_t *pte;
1485
1486                 pte = vtopte(va);
1487                 *pte = VM_PAGE_TO_PHYS(*m) |
1488                     kernel_pmap.pmap_bits[PG_RW_IDX] |
1489                     kernel_pmap.pmap_bits[PG_V_IDX] |
1490                     kernel_pmap.pmap_cache_bits[(*m)->pat_mode];
1491 //              pgeflag;
1492                 cpu_invlpg((void *)va);
1493                 va += PAGE_SIZE;
1494                 m++;
1495         }
1496         smp_invltlb();
1497 }
1498
1499 /*
1500  * This routine jerks page mappings from the
1501  * kernel -- it is meant only for temporary mappings.
1502  *
1503  * MPSAFE, INTERRUPT SAFE (cluster callback)
1504  */
1505 void
1506 pmap_qremove(vm_offset_t va, int count)
1507 {
1508         vm_offset_t end_va;
1509
1510         end_va = va + count * PAGE_SIZE;
1511
1512         while (va < end_va) {
1513                 pt_entry_t *pte;
1514
1515                 pte = vtopte(va);
1516                 (void)pte_load_clear(pte);
1517                 cpu_invlpg((void *)va);
1518                 va += PAGE_SIZE;
1519         }
1520         smp_invltlb();
1521 }
1522
1523 /*
1524  * Create a new thread and optionally associate it with a (new) process.
1525  * NOTE! the new thread's cpu may not equal the current cpu.
1526  */
1527 void
1528 pmap_init_thread(thread_t td)
1529 {
1530         /* enforce pcb placement & alignment */
1531         td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1532         td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1533         td->td_savefpu = &td->td_pcb->pcb_save;
1534         td->td_sp = (char *)td->td_pcb; /* no -16 */
1535 }
1536
1537 /*
1538  * This routine directly affects the fork perf for a process.
1539  */
1540 void
1541 pmap_init_proc(struct proc *p)
1542 {
1543 }
1544
1545 static void
1546 pmap_pinit_defaults(struct pmap *pmap)
1547 {
1548         bcopy(pmap_bits_default, pmap->pmap_bits,
1549               sizeof(pmap_bits_default));
1550         bcopy(protection_codes, pmap->protection_codes,
1551               sizeof(protection_codes));
1552         bcopy(pat_pte_index, pmap->pmap_cache_bits,
1553               sizeof(pat_pte_index));
1554         pmap->pmap_cache_mask = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
1555         pmap->copyinstr = std_copyinstr;
1556         pmap->copyin = std_copyin;
1557         pmap->copyout = std_copyout;
1558         pmap->fubyte = std_fubyte;
1559         pmap->subyte = std_subyte;
1560         pmap->fuword = std_fuword;
1561         pmap->suword = std_suword;
1562         pmap->suword32 = std_suword32;
1563 }
1564 /*
1565  * Initialize pmap0/vmspace0.  This pmap is not added to pmap_list because
1566  * it, and IdlePTD, represents the template used to update all other pmaps.
1567  *
1568  * On architectures where the kernel pmap is not integrated into the user
1569  * process pmap, this pmap represents the process pmap, not the kernel pmap.
1570  * kernel_pmap should be used to directly access the kernel_pmap.
1571  */
1572 void
1573 pmap_pinit0(struct pmap *pmap)
1574 {
1575         pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1576         pmap->pm_count = 1;
1577         pmap->pm_active = 0;
1578         pmap->pm_pvhint = NULL;
1579         RB_INIT(&pmap->pm_pvroot);
1580         spin_init(&pmap->pm_spin);
1581         lwkt_token_init(&pmap->pm_token, "pmap_tok");
1582         bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1583         pmap_pinit_defaults(pmap);
1584 }
1585
1586 /*
1587  * Initialize a preallocated and zeroed pmap structure,
1588  * such as one in a vmspace structure.
1589  */
1590 static void
1591 pmap_pinit_simple(struct pmap *pmap)
1592 {
1593         /*
1594          * Misc initialization
1595          */
1596         pmap->pm_count = 1;
1597         pmap->pm_active = 0;
1598         pmap->pm_pvhint = NULL;
1599         pmap->pm_flags = PMAP_FLAG_SIMPLE;
1600
1601         pmap_pinit_defaults(pmap);
1602
1603         /*
1604          * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1605          * for this).
1606          */
1607         if (pmap->pm_pmlpv == NULL) {
1608                 RB_INIT(&pmap->pm_pvroot);
1609                 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1610                 spin_init(&pmap->pm_spin);
1611                 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1612         }
1613 }
1614
1615 void
1616 pmap_pinit(struct pmap *pmap)
1617 {
1618         pv_entry_t pv;
1619         int j;
1620
1621         if (pmap->pm_pmlpv) {
1622                 if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
1623                         pmap_puninit(pmap);
1624                 }
1625         }
1626
1627         pmap_pinit_simple(pmap);
1628         pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
1629
1630         /*
1631          * No need to allocate page table space yet but we do need a valid
1632          * page directory table.
1633          */
1634         if (pmap->pm_pml4 == NULL) {
1635                 pmap->pm_pml4 =
1636                     (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1637         }
1638
1639         /*
1640          * Allocate the page directory page, which wires it even though
1641          * it isn't being entered into some higher level page table (it
1642          * being the highest level).  If one is already cached we don't
1643          * have to do anything.
1644          */
1645         if ((pv = pmap->pm_pmlpv) == NULL) {
1646                 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1647                 pmap->pm_pmlpv = pv;
1648                 pmap_kenter((vm_offset_t)pmap->pm_pml4,
1649                             VM_PAGE_TO_PHYS(pv->pv_m));
1650                 pv_put(pv);
1651
1652                 /*
1653                  * Install DMAP and KMAP.
1654                  */
1655                 for (j = 0; j < NDMPML4E; ++j) {
1656                         pmap->pm_pml4[DMPML4I + j] =
1657                             (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1658                             pmap->pmap_bits[PG_RW_IDX] |
1659                             pmap->pmap_bits[PG_V_IDX] |
1660                             pmap->pmap_bits[PG_U_IDX];
1661                 }
1662                 pmap->pm_pml4[KPML4I] = KPDPphys |
1663                     pmap->pmap_bits[PG_RW_IDX] |
1664                     pmap->pmap_bits[PG_V_IDX] |
1665                     pmap->pmap_bits[PG_U_IDX];
1666
1667                 /*
1668                  * install self-referential address mapping entry
1669                  */
1670                 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1671                     pmap->pmap_bits[PG_V_IDX] |
1672                     pmap->pmap_bits[PG_RW_IDX] |
1673                     pmap->pmap_bits[PG_A_IDX] |
1674                     pmap->pmap_bits[PG_M_IDX];
1675         } else {
1676                 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1677                 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
1678         }
1679         KKASSERT(pmap->pm_pml4[255] == 0);
1680         KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1681         KKASSERT(pv->pv_entry.rbe_left == NULL);
1682         KKASSERT(pv->pv_entry.rbe_right == NULL);
1683 }
1684
1685 /*
1686  * Clean up a pmap structure so it can be physically freed.  This routine
1687  * is called by the vmspace dtor function.  A great deal of pmap data is
1688  * left passively mapped to improve vmspace management so we have a bit
1689  * of cleanup work to do here.
1690  */
1691 void
1692 pmap_puninit(pmap_t pmap)
1693 {
1694         pv_entry_t pv;
1695         vm_page_t p;
1696
1697         KKASSERT(pmap->pm_active == 0);
1698         if ((pv = pmap->pm_pmlpv) != NULL) {
1699                 if (pv_hold_try(pv) == 0)
1700                         pv_lock(pv);
1701                 KKASSERT(pv == pmap->pm_pmlpv);
1702                 p = pmap_remove_pv_page(pv);
1703                 pv_free(pv);
1704                 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1705                 vm_page_busy_wait(p, FALSE, "pgpun");
1706                 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
1707                 vm_page_unwire(p, 0);
1708                 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1709
1710                 /*
1711                  * XXX eventually clean out PML4 static entries and
1712                  * use vm_page_free_zero()
1713                  */
1714                 vm_page_free(p);
1715                 pmap->pm_pmlpv = NULL;
1716         }
1717         if (pmap->pm_pml4) {
1718                 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1719                 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1720                 pmap->pm_pml4 = NULL;
1721         }
1722         KKASSERT(pmap->pm_stats.resident_count == 0);
1723         KKASSERT(pmap->pm_stats.wired_count == 0);
1724 }
1725
1726 /*
1727  * Wire in kernel global address entries.  To avoid a race condition
1728  * between pmap initialization and pmap_growkernel, this procedure
1729  * adds the pmap to the master list (which growkernel scans to update),
1730  * then copies the template.
1731  */
1732 void
1733 pmap_pinit2(struct pmap *pmap)
1734 {
1735         spin_lock(&pmap_spin);
1736         TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1737         spin_unlock(&pmap_spin);
1738 }
1739
1740 /*
1741  * This routine is called when various levels in the page table need to
1742  * be populated.  This routine cannot fail.
1743  *
1744  * This function returns two locked pv_entry's, one representing the
1745  * requested pv and one representing the requested pv's parent pv.  If
1746  * the pv did not previously exist it will be mapped into its parent
1747  * and wired, otherwise no additional wire count will be added.
1748  */
1749 static
1750 pv_entry_t
1751 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
1752 {
1753         pt_entry_t *ptep;
1754         pv_entry_t pv;
1755         pv_entry_t pvp;
1756         vm_pindex_t pt_pindex;
1757         vm_page_t m;
1758         int isnew;
1759         int ispt;
1760
1761         /*
1762          * If the pv already exists and we aren't being asked for the
1763          * parent page table page we can just return it.  A locked+held pv
1764          * is returned.  The pv will also have a second hold related to the
1765          * pmap association that we don't have to worry about.
1766          */
1767         ispt = 0;
1768         pv = pv_alloc(pmap, ptepindex, &isnew);
1769         if (isnew == 0 && pvpp == NULL)
1770                 return(pv);
1771
1772         /*
1773          * Special case terminal PVs.  These are not page table pages so
1774          * no vm_page is allocated (the caller supplied the vm_page).  If
1775          * pvpp is non-NULL we are being asked to also removed the pt_pv
1776          * for this pv.
1777          *
1778          * Note that pt_pv's are only returned for user VAs. We assert that
1779          * a pt_pv is not being requested for kernel VAs.
1780          */
1781         if (ptepindex < pmap_pt_pindex(0)) {
1782                 if (ptepindex >= NUPTE_USER)
1783                         KKASSERT(pvpp == NULL);
1784                 else
1785                         KKASSERT(pvpp != NULL);
1786                 if (pvpp) {
1787                         pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1788                         pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1789                         if (isnew)
1790                                 vm_page_wire_quick(pvp->pv_m);
1791                         *pvpp = pvp;
1792                 } else {
1793                         pvp = NULL;
1794                 }
1795                 return(pv);
1796         }
1797
1798         /*
1799          * Non-terminal PVs allocate a VM page to represent the page table,
1800          * so we have to resolve pvp and calculate ptepindex for the pvp
1801          * and then for the page table entry index in the pvp for
1802          * fall-through.
1803          */
1804         if (ptepindex < pmap_pd_pindex(0)) {
1805                 /*
1806                  * pv is PT, pvp is PD
1807                  */
1808                 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1809                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1810                 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1811                 if (!isnew)
1812                         goto notnew;
1813
1814                 /*
1815                  * PT index in PD
1816                  */
1817                 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1818                 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
1819                 ispt = 1;
1820         } else if (ptepindex < pmap_pdp_pindex(0)) {
1821                 /*
1822                  * pv is PD, pvp is PDP
1823                  *
1824                  * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1825                  *                   the PD.
1826                  */
1827                 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1828                 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
1829
1830                 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1831                         KKASSERT(pvpp == NULL);
1832                         pvp = NULL;
1833                 } else {
1834                         pvp = pmap_allocpte(pmap, ptepindex, NULL);
1835                 }
1836                 if (!isnew)
1837                         goto notnew;
1838
1839                 /*
1840                  * PD index in PDP
1841                  */
1842                 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1843                 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1844         } else if (ptepindex < pmap_pml4_pindex()) {
1845                 /*
1846                  * pv is PDP, pvp is the root pml4 table
1847                  */
1848                 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1849                 if (!isnew)
1850                         goto notnew;
1851
1852                 /*
1853                  * PDP index in PML4
1854                  */
1855                 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1856                 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1857         } else {
1858                 /*
1859                  * pv represents the top-level PML4, there is no parent.
1860                  */
1861                 pvp = NULL;
1862                 if (!isnew)
1863                         goto notnew;
1864         }
1865
1866         /*
1867          * This code is only reached if isnew is TRUE and this is not a
1868          * terminal PV.  We need to allocate a vm_page for the page table
1869          * at this level and enter it into the parent page table.
1870          *
1871          * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1872          */
1873         for (;;) {
1874                 m = vm_page_alloc(NULL, pv->pv_pindex,
1875                                   VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1876                                   VM_ALLOC_INTERRUPT);
1877                 if (m)
1878                         break;
1879                 vm_wait(0);
1880         }
1881         vm_page_spin_lock(m);
1882         TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1883         pv->pv_m = m;
1884         vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1885         vm_page_spin_unlock(m);
1886         vm_page_unmanage(m);    /* m must be spinunlocked */
1887
1888         if ((m->flags & PG_ZERO) == 0) {
1889                 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1890         }
1891 #ifdef PMAP_DEBUG
1892         else {
1893                 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1894         }
1895 #endif
1896         m->valid = VM_PAGE_BITS_ALL;
1897         vm_page_flag_clear(m, PG_ZERO);
1898         vm_page_wire(m);        /* wire for mapping in parent */
1899
1900         /*
1901          * Wire the page into pvp, bump the wire-count for pvp's page table
1902          * page.  Bump the resident_count for the pmap.  There is no pvp
1903          * for the top level, address the pm_pml4[] array directly.
1904          *
1905          * If the caller wants the parent we return it, otherwise
1906          * we just put it away.
1907          *
1908          * No interlock is needed for pte 0 -> non-zero.
1909          *
1910          * In the situation where *ptep is valid we might have an unmanaged
1911          * page table page shared from another page table which we need to
1912          * unshare before installing our private page table page.
1913          */
1914         if (pvp) {
1915                 ptep = pv_pte_lookup(pvp, ptepindex);
1916                 if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1917                         pt_entry_t pte;
1918                         pmap_inval_info info;
1919
1920                         if (ispt == 0) {
1921                                 panic("pmap_allocpte: unexpected pte %p/%d",
1922                                       pvp, (int)ptepindex);
1923                         }
1924                         pmap_inval_init(&info);
1925                         pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1926                         pte = pte_load_clear(ptep);
1927                         pmap_inval_deinterlock(&info, pmap);
1928                         pmap_inval_done(&info);
1929                         if (vm_page_unwire_quick(
1930                                         PHYS_TO_VM_PAGE(pte & PG_FRAME))) {
1931                                 panic("pmap_allocpte: shared pgtable "
1932                                       "pg bad wirecount");
1933                         }
1934                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
1935                 } else {
1936                         vm_page_wire_quick(pvp->pv_m);
1937                 }
1938                 *ptep = VM_PAGE_TO_PHYS(m) |
1939                     (pmap->pmap_bits[PG_U_IDX] |
1940                     pmap->pmap_bits[PG_RW_IDX] |
1941                     pmap->pmap_bits[PG_V_IDX] |
1942                     pmap->pmap_bits[PG_A_IDX] |
1943                     pmap->pmap_bits[PG_M_IDX]);
1944         }
1945         vm_page_wakeup(m);
1946 notnew:
1947         if (pvpp)
1948                 *pvpp = pvp;
1949         else if (pvp)
1950                 pv_put(pvp);
1951         return (pv);
1952 }
1953
1954 /*
1955  * This version of pmap_allocpte() checks for possible segment optimizations
1956  * that would allow page-table sharing.  It can be called for terminal
1957  * page or page table page ptepindex's.
1958  *
1959  * The function is called with page table page ptepindex's for fictitious
1960  * and unmanaged terminal pages.  That is, we don't want to allocate a
1961  * terminal pv, we just want the pt_pv.  pvpp is usually passed as NULL
1962  * for this case.
1963  *
1964  * This function can return a pv and *pvpp associated with the passed in pmap
1965  * OR a pv and *pvpp associated with the shared pmap.  In the latter case
1966  * an unmanaged page table page will be entered into the pass in pmap.
1967  */
1968 static
1969 pv_entry_t
1970 pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
1971                   vm_map_entry_t entry, vm_offset_t va)
1972 {
1973         struct pmap_inval_info info;
1974         vm_object_t object;
1975         pmap_t obpmap;
1976         pmap_t *obpmapp;
1977         vm_offset_t b;
1978         pv_entry_t pte_pv;      /* in original or shared pmap */
1979         pv_entry_t pt_pv;       /* in original or shared pmap */
1980         pv_entry_t proc_pd_pv;  /* in original pmap */
1981         pv_entry_t proc_pt_pv;  /* in original pmap */
1982         pv_entry_t xpv;         /* PT in shared pmap */
1983         pd_entry_t *pt;         /* PT entry in PD of original pmap */
1984         pd_entry_t opte;        /* contents of *pt */
1985         pd_entry_t npte;        /* contents of *pt */
1986         vm_page_t m;
1987
1988 retry:
1989         /*
1990          * Basic tests, require a non-NULL vm_map_entry, require proper
1991          * alignment and type for the vm_map_entry, require that the
1992          * underlying object already be allocated.
1993          *
1994          * We currently allow any type of object to use this optimization.
1995          * The object itself does NOT have to be sized to a multiple of the
1996          * segment size, but the memory mapping does.
1997          *
1998          * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
1999          *     won't work as expected.
2000          */
2001         if (entry == NULL ||
2002             pmap_mmu_optimize == 0 ||                   /* not enabled */
2003             ptepindex >= pmap_pd_pindex(0) ||           /* not terminal */
2004             entry->inheritance != VM_INHERIT_SHARE ||   /* not shared */
2005             entry->maptype != VM_MAPTYPE_NORMAL ||      /* weird map type */
2006             entry->object.vm_object == NULL ||          /* needs VM object */
2007             entry->object.vm_object->type == OBJT_DEVICE ||     /* ick */
2008             entry->object.vm_object->type == OBJT_MGTDEVICE ||  /* ick */
2009             (entry->offset & SEG_MASK) ||               /* must be aligned */
2010             (entry->start & SEG_MASK)) {
2011                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2012         }
2013
2014         /*
2015          * Make sure the full segment can be represented.
2016          */
2017         b = va & ~(vm_offset_t)SEG_MASK;
2018         if (b < entry->start && b + SEG_SIZE > entry->end)
2019                 return(pmap_allocpte(pmap, ptepindex, pvpp));
2020
2021         /*
2022          * If the full segment can be represented dive the VM object's
2023          * shared pmap, allocating as required.
2024          */
2025         object = entry->object.vm_object;
2026
2027         if (entry->protection & VM_PROT_WRITE)
2028                 obpmapp = &object->md.pmap_rw;
2029         else
2030                 obpmapp = &object->md.pmap_ro;
2031
2032         /*
2033          * We allocate what appears to be a normal pmap but because portions
2034          * of this pmap are shared with other unrelated pmaps we have to
2035          * set pm_active to point to all cpus.
2036          *
2037          * XXX Currently using pmap_spin to interlock the update, can't use
2038          *     vm_object_hold/drop because the token might already be held
2039          *     shared OR exclusive and we don't know.
2040          */
2041         while ((obpmap = *obpmapp) == NULL) {
2042                 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
2043                 pmap_pinit_simple(obpmap);
2044                 pmap_pinit2(obpmap);
2045                 spin_lock(&pmap_spin);
2046                 if (*obpmapp != NULL) {
2047                         /*
2048                          * Handle race
2049                          */
2050                         spin_unlock(&pmap_spin);
2051                         pmap_release(obpmap);
2052                         pmap_puninit(obpmap);
2053                         kfree(obpmap, M_OBJPMAP);
2054                 } else {
2055                         obpmap->pm_active = smp_active_mask;
2056                         *obpmapp = obpmap;
2057                         spin_unlock(&pmap_spin);
2058                 }
2059         }
2060
2061         /*
2062          * Layering is: PTE, PT, PD, PDP, PML4.  We have to return the
2063          * pte/pt using the shared pmap from the object but also adjust
2064          * the process pmap's page table page as a side effect.
2065          */
2066
2067         /*
2068          * Resolve the terminal PTE and PT in the shared pmap.  This is what
2069          * we will return.  This is true if ptepindex represents a terminal
2070          * page, otherwise pte_pv is actually the PT and pt_pv is actually
2071          * the PD.
2072          */
2073         pt_pv = NULL;
2074         pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
2075         if (ptepindex >= pmap_pt_pindex(0))
2076                 xpv = pte_pv;
2077         else
2078                 xpv = pt_pv;
2079
2080         /*
2081          * Resolve the PD in the process pmap so we can properly share the
2082          * page table page.  Lock order is bottom-up (leaf first)!
2083          *
2084          * NOTE: proc_pt_pv can be NULL.
2085          */
2086         proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
2087         proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
2088
2089         /*
2090          * xpv is the page table page pv from the shared object
2091          * (for convenience).
2092          *
2093          * Calculate the pte value for the PT to load into the process PD.
2094          * If we have to change it we must properly dispose of the previous
2095          * entry.
2096          */
2097         pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2098         npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
2099             (pmap->pmap_bits[PG_U_IDX] |
2100             pmap->pmap_bits[PG_RW_IDX] |
2101             pmap->pmap_bits[PG_V_IDX] |
2102             pmap->pmap_bits[PG_A_IDX] |
2103             pmap->pmap_bits[PG_M_IDX]);
2104
2105         /*
2106          * Dispose of previous page table page if it was local to the
2107          * process pmap.  If the old pt is not empty we cannot dispose of it
2108          * until we clean it out.  This case should not arise very often so
2109          * it is not optimized.
2110          */
2111         if (proc_pt_pv) {
2112                 if (proc_pt_pv->pv_m->wire_count != 1) {
2113                         pv_put(proc_pd_pv);
2114                         pv_put(proc_pt_pv);
2115                         pv_put(pt_pv);
2116                         pv_put(pte_pv);
2117                         pmap_remove(pmap,
2118                                     va & ~(vm_offset_t)SEG_MASK,
2119                                     (va + SEG_SIZE) & ~(vm_offset_t)SEG_MASK);
2120                         goto retry;
2121                 }
2122                 pmap_release_pv(proc_pt_pv, proc_pd_pv);
2123                 proc_pt_pv = NULL;
2124                 /* relookup */
2125                 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
2126         }
2127
2128         /*
2129          * Handle remaining cases.
2130          */
2131         if (*pt == 0) {
2132                 *pt = npte;
2133                 vm_page_wire_quick(xpv->pv_m);
2134                 vm_page_wire_quick(proc_pd_pv->pv_m);
2135                 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2136         } else if (*pt != npte) {
2137                 pmap_inval_init(&info);
2138                 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
2139
2140                 opte = pte_load_clear(pt);
2141                 KKASSERT(opte && opte != npte);
2142
2143                 *pt = npte;
2144                 vm_page_wire_quick(xpv->pv_m);  /* pgtable pg that is npte */
2145
2146                 /*
2147                  * Clean up opte, bump the wire_count for the process
2148                  * PD page representing the new entry if it was
2149                  * previously empty.
2150                  *
2151                  * If the entry was not previously empty and we have
2152                  * a PT in the proc pmap then opte must match that
2153                  * pt.  The proc pt must be retired (this is done
2154                  * later on in this procedure).
2155                  *
2156                  * NOTE: replacing valid pte, wire_count on proc_pd_pv
2157                  * stays the same.
2158                  */
2159                 KKASSERT(opte & pmap->pmap_bits[PG_V_IDX]);
2160                 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
2161                 if (vm_page_unwire_quick(m)) {
2162                         panic("pmap_allocpte_seg: "
2163                               "bad wire count %p",
2164                               m);
2165                 }
2166
2167                 pmap_inval_deinterlock(&info, pmap);
2168                 pmap_inval_done(&info);
2169         }
2170
2171         /*
2172          * The existing process page table was replaced and must be destroyed
2173          * here.
2174          */
2175         if (proc_pd_pv)
2176                 pv_put(proc_pd_pv);
2177         if (pvpp)
2178                 *pvpp = pt_pv;
2179         else
2180                 pv_put(pt_pv);
2181
2182         return (pte_pv);
2183 }
2184
2185 /*
2186  * Release any resources held by the given physical map.
2187  *
2188  * Called when a pmap initialized by pmap_pinit is being released.  Should
2189  * only be called if the map contains no valid mappings.
2190  *
2191  * Caller must hold pmap->pm_token
2192  */
2193 struct pmap_release_info {
2194         pmap_t  pmap;
2195         int     retry;
2196 };
2197
2198 static int pmap_release_callback(pv_entry_t pv, void *data);
2199
2200 void
2201 pmap_release(struct pmap *pmap)
2202 {
2203         struct pmap_release_info info;
2204
2205         KASSERT(pmap->pm_active == 0,
2206                 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
2207
2208         spin_lock(&pmap_spin);
2209         TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
2210         spin_unlock(&pmap_spin);
2211
2212         /*
2213          * Pull pv's off the RB tree in order from low to high and release
2214          * each page.
2215          */
2216         info.pmap = pmap;
2217         do {
2218                 info.retry = 0;
2219                 spin_lock(&pmap->pm_spin);
2220                 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2221                         pmap_release_callback, &info);
2222                 spin_unlock(&pmap->pm_spin);
2223         } while (info.retry);
2224
2225
2226         /*
2227          * One resident page (the pml4 page) should remain.
2228          * No wired pages should remain.
2229          */
2230         KKASSERT(pmap->pm_stats.resident_count ==
2231                  ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
2232
2233         KKASSERT(pmap->pm_stats.wired_count == 0);
2234 }
2235
2236 static int
2237 pmap_release_callback(pv_entry_t pv, void *data)
2238 {
2239         struct pmap_release_info *info = data;
2240         pmap_t pmap = info->pmap;
2241         int r;
2242
2243         if (pv_hold_try(pv)) {
2244                 spin_unlock(&pmap->pm_spin);
2245         } else {
2246                 spin_unlock(&pmap->pm_spin);
2247                 pv_lock(pv);
2248         }
2249         if (pv->pv_pmap != pmap) {
2250                 pv_put(pv);
2251                 spin_lock(&pmap->pm_spin);
2252                 info->retry = 1;
2253                 return(-1);
2254         }
2255         r = pmap_release_pv(pv, NULL);
2256         spin_lock(&pmap->pm_spin);
2257         return(r);
2258 }
2259
2260 /*
2261  * Called with held (i.e. also locked) pv.  This function will dispose of
2262  * the lock along with the pv.
2263  *
2264  * If the caller already holds the locked parent page table for pv it
2265  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2266  * pass NULL for pvp.
2267  */
2268 static int
2269 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp)
2270 {
2271         vm_page_t p;
2272
2273         /*
2274          * The pmap is currently not spinlocked, pv is held+locked.
2275          * Remove the pv's page from its parent's page table.  The
2276          * parent's page table page's wire_count will be decremented.
2277          */
2278         pmap_remove_pv_pte(pv, pvp, NULL);
2279
2280         /*
2281          * Terminal pvs are unhooked from their vm_pages.  Because
2282          * terminal pages aren't page table pages they aren't wired
2283          * by us, so we have to be sure not to unwire them either.
2284          */
2285         if (pv->pv_pindex < pmap_pt_pindex(0)) {
2286                 pmap_remove_pv_page(pv);
2287                 goto skip;
2288         }
2289
2290         /*
2291          * We leave the top-level page table page cached, wired, and
2292          * mapped in the pmap until the dtor function (pmap_puninit())
2293          * gets called.
2294          *
2295          * Since we are leaving the top-level pv intact we need
2296          * to break out of what would otherwise be an infinite loop.
2297          */
2298         if (pv->pv_pindex == pmap_pml4_pindex()) {
2299                 pv_put(pv);
2300                 return(-1);
2301         }
2302
2303         /*
2304          * For page table pages (other than the top-level page),
2305          * remove and free the vm_page.  The representitive mapping
2306          * removed above by pmap_remove_pv_pte() did not undo the
2307          * last wire_count so we have to do that as well.
2308          */
2309         p = pmap_remove_pv_page(pv);
2310         vm_page_busy_wait(p, FALSE, "pmaprl");
2311         if (p->wire_count != 1) {
2312                 kprintf("p->wire_count was %016lx %d\n",
2313                         pv->pv_pindex, p->wire_count);
2314         }
2315         KKASSERT(p->wire_count == 1);
2316         KKASSERT(p->flags & PG_UNMANAGED);
2317
2318         vm_page_unwire(p, 0);
2319         KKASSERT(p->wire_count == 0);
2320
2321         /*
2322          * Theoretically this page, if not the pml4 page, should contain
2323          * all-zeros.  But its just too dangerous to mark it PG_ZERO.  Free
2324          * normally.
2325          */
2326         vm_page_free(p);
2327 skip:
2328         pv_free(pv);
2329         return 0;
2330 }
2331
2332 /*
2333  * This function will remove the pte associated with a pv from its parent.
2334  * Terminal pv's are supported.  The removal will be interlocked if info
2335  * is non-NULL.  The caller must dispose of pv instead of just unlocking
2336  * it.
2337  *
2338  * The wire count will be dropped on the parent page table.  The wire
2339  * count on the page being removed (pv->pv_m) from the parent page table
2340  * is NOT touched.  Note that terminal pages will not have any additional
2341  * wire counts while page table pages will have at least one representing
2342  * the mapping, plus others representing sub-mappings.
2343  *
2344  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2345  *       pages and user page table and terminal pages.
2346  *
2347  * The pv must be locked.
2348  *
2349  * XXX must lock parent pv's if they exist to remove pte XXX
2350  */
2351 static
2352 void
2353 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2354 {
2355         vm_pindex_t ptepindex = pv->pv_pindex;
2356         pmap_t pmap = pv->pv_pmap;
2357         vm_page_t p;
2358         int gotpvp = 0;
2359
2360         KKASSERT(pmap);
2361
2362         if (ptepindex == pmap_pml4_pindex()) {
2363                 /*
2364                  * We are the top level pml4 table, there is no parent.
2365                  */
2366                 p = pmap->pm_pmlpv->pv_m;
2367         } else if (ptepindex >= pmap_pdp_pindex(0)) {
2368                 /*
2369                  * Remove a PDP page from the pml4e.  This can only occur
2370                  * with user page tables.  We do not have to lock the
2371                  * pml4 PV so just ignore pvp.
2372                  */
2373                 vm_pindex_t pml4_pindex;
2374                 vm_pindex_t pdp_index;
2375                 pml4_entry_t *pdp;
2376
2377                 pdp_index = ptepindex - pmap_pdp_pindex(0);
2378                 if (pvp == NULL) {
2379                         pml4_pindex = pmap_pml4_pindex();
2380                         pvp = pv_get(pv->pv_pmap, pml4_pindex);
2381                         KKASSERT(pvp);
2382                         gotpvp = 1;
2383                 }
2384                 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2385                 KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2386                 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2387                 *pdp = 0;
2388                 KKASSERT(info == NULL);
2389         } else if (ptepindex >= pmap_pd_pindex(0)) {
2390                 /*
2391                  * Remove a PD page from the pdp
2392                  *
2393                  * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2394                  *                   of a simple pmap because it stops at
2395                  *                   the PD page.
2396                  */
2397                 vm_pindex_t pdp_pindex;
2398                 vm_pindex_t pd_index;
2399                 pdp_entry_t *pd;
2400
2401                 pd_index = ptepindex - pmap_pd_pindex(0);
2402
2403                 if (pvp == NULL) {
2404                         pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2405                                      (pd_index >> NPML4EPGSHIFT);
2406                         pvp = pv_get(pv->pv_pmap, pdp_pindex);
2407                         if (pvp)
2408                                 gotpvp = 1;
2409                 }
2410                 if (pvp) {
2411                         pd = pv_pte_lookup(pvp, pd_index &
2412                                                 ((1ul << NPDPEPGSHIFT) - 1));
2413                         KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
2414                         p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2415                         *pd = 0;
2416                 } else {
2417                         KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2418                         p = pv->pv_m;           /* degenerate test later */
2419                 }
2420                 KKASSERT(info == NULL);
2421         } else if (ptepindex >= pmap_pt_pindex(0)) {
2422                 /*
2423                  *  Remove a PT page from the pd
2424                  */
2425                 vm_pindex_t pd_pindex;
2426                 vm_pindex_t pt_index;
2427                 pd_entry_t *pt;
2428
2429                 pt_index = ptepindex - pmap_pt_pindex(0);
2430
2431                 if (pvp == NULL) {
2432                         pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2433                                     (pt_index >> NPDPEPGSHIFT);
2434                         pvp = pv_get(pv->pv_pmap, pd_pindex);
2435                         KKASSERT(pvp);
2436                         gotpvp = 1;
2437                 }
2438                 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2439                 KKASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0);
2440                 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2441                 *pt = 0;
2442                 KKASSERT(info == NULL);
2443         } else {
2444                 /*
2445                  * Remove a PTE from the PT page
2446                  *
2447                  * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2448                  *       pv is a pte_pv so we can safely lock pt_pv.
2449                  *
2450                  * NOTE: FICTITIOUS pages may have multiple physical mappings
2451                  *       so PHYS_TO_VM_PAGE() will not necessarily work for
2452                  *       terminal ptes.
2453                  */
2454                 vm_pindex_t pt_pindex;
2455                 pt_entry_t *ptep;
2456                 pt_entry_t pte;
2457                 vm_offset_t va;
2458
2459                 pt_pindex = ptepindex >> NPTEPGSHIFT;
2460                 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2461
2462                 if (ptepindex >= NUPTE_USER) {
2463                         ptep = vtopte(ptepindex << PAGE_SHIFT);
2464                         KKASSERT(pvp == NULL);
2465                 } else {
2466                         if (pvp == NULL) {
2467                                 pt_pindex = NUPTE_TOTAL +
2468                                             (ptepindex >> NPDPEPGSHIFT);
2469                                 pvp = pv_get(pv->pv_pmap, pt_pindex);
2470                                 KKASSERT(pvp);
2471                                 gotpvp = 1;
2472                         }
2473                         ptep = pv_pte_lookup(pvp, ptepindex &
2474                                                   ((1ul << NPDPEPGSHIFT) - 1));
2475                 }
2476
2477                 if (info)
2478                         pmap_inval_interlock(info, pmap, va);
2479                 pte = pte_load_clear(ptep);
2480                 if (info)
2481                         pmap_inval_deinterlock(info, pmap);
2482                 else
2483                         cpu_invlpg((void *)va);
2484
2485                 /*
2486                  * Now update the vm_page_t
2487                  */
2488                 if ((pte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) !=
2489                     (pmap->pmap_bits[PG_MANAGED_IDX]|pmap->pmap_bits[PG_V_IDX])) {
2490                         kprintf("remove_pte badpte %016lx %016lx %d\n",
2491                                 pte, pv->pv_pindex,
2492                                 pv->pv_pindex < pmap_pt_pindex(0));
2493                 }
2494                 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2495                 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2496                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
2497                         p = pv->pv_m;
2498                 else
2499                         p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2500                 /* p = pv->pv_m; */
2501
2502                 if (pte & pmap->pmap_bits[PG_M_IDX]) {
2503                         if (pmap_track_modified(ptepindex))
2504                                 vm_page_dirty(p);
2505                 }
2506                 if (pte & pmap->pmap_bits[PG_A_IDX]) {
2507                         vm_page_flag_set(p, PG_REFERENCED);
2508                 }
2509                 if (pte & pmap->pmap_bits[PG_W_IDX])
2510                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
2511                 if (pte & pmap->pmap_bits[PG_G_IDX])
2512                         cpu_invlpg((void *)va);
2513         }
2514
2515         /*
2516          * Unwire the parent page table page.  The wire_count cannot go below
2517          * 1 here because the parent page table page is itself still mapped.
2518          *
2519          * XXX remove the assertions later.
2520          */
2521         KKASSERT(pv->pv_m == p);
2522         if (pvp && vm_page_unwire_quick(pvp->pv_m))
2523                 panic("pmap_remove_pv_pte: Insufficient wire_count");
2524
2525         if (gotpvp)
2526                 pv_put(pvp);
2527 }
2528
2529 /*
2530  * Remove the vm_page association to a pv.  The pv must be locked.
2531  */
2532 static
2533 vm_page_t
2534 pmap_remove_pv_page(pv_entry_t pv)
2535 {
2536         vm_page_t m;
2537
2538         m = pv->pv_m;
2539         KKASSERT(m);
2540         vm_page_spin_lock(m);
2541         pv->pv_m = NULL;
2542         TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2543         /*
2544         if (m->object)
2545                 atomic_add_int(&m->object->agg_pv_list_count, -1);
2546         */
2547         if (TAILQ_EMPTY(&m->md.pv_list))
2548                 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2549         vm_page_spin_unlock(m);
2550         return(m);
2551 }
2552
2553 /*
2554  * Grow the number of kernel page table entries, if needed.
2555  *
2556  * This routine is always called to validate any address space
2557  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
2558  * space below KERNBASE.
2559  */
2560 void
2561 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
2562 {
2563         vm_paddr_t paddr;
2564         vm_offset_t ptppaddr;
2565         vm_page_t nkpg;
2566         pd_entry_t *pt, newpt;
2567         pdp_entry_t newpd;
2568         int update_kernel_vm_end;
2569
2570         /*
2571          * bootstrap kernel_vm_end on first real VM use
2572          */
2573         if (kernel_vm_end == 0) {
2574                 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
2575                 nkpt = 0;
2576                 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2577                         kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2578                                         ~(PAGE_SIZE * NPTEPG - 1);
2579                         nkpt++;
2580                         if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2581                                 kernel_vm_end = kernel_map.max_offset;
2582                                 break;                       
2583                         }
2584                 }
2585         }
2586
2587         /*
2588          * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
2589          * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
2590          * do not want to force-fill 128G worth of page tables.
2591          */
2592         if (kstart < KERNBASE) {
2593                 if (kstart > kernel_vm_end)
2594                         kstart = kernel_vm_end;
2595                 KKASSERT(kend <= KERNBASE);
2596                 update_kernel_vm_end = 1;
2597         } else {
2598                 update_kernel_vm_end = 0;
2599         }
2600
2601         kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2602         kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2603
2604         if (kend - 1 >= kernel_map.max_offset)
2605                 kend = kernel_map.max_offset;
2606
2607         while (kstart < kend) {
2608                 pt = pmap_pt(&kernel_pmap, kstart);
2609                 if (pt == NULL) {
2610                         /* We need a new PDP entry */
2611                         nkpg = vm_page_alloc(NULL, nkpt,
2612                                              VM_ALLOC_NORMAL |
2613                                              VM_ALLOC_SYSTEM |
2614                                              VM_ALLOC_INTERRUPT);
2615                         if (nkpg == NULL) {
2616                                 panic("pmap_growkernel: no memory to grow "
2617                                       "kernel");
2618                         }
2619                         paddr = VM_PAGE_TO_PHYS(nkpg);
2620                         if ((nkpg->flags & PG_ZERO) == 0)
2621                                 pmap_zero_page(paddr);
2622                         vm_page_flag_clear(nkpg, PG_ZERO);
2623                         newpd = (pdp_entry_t)
2624                             (paddr |
2625                             kernel_pmap.pmap_bits[PG_V_IDX] |
2626                             kernel_pmap.pmap_bits[PG_RW_IDX] |
2627                             kernel_pmap.pmap_bits[PG_A_IDX] |
2628                             kernel_pmap.pmap_bits[PG_M_IDX]);
2629                         *pmap_pd(&kernel_pmap, kstart) = newpd;
2630                         nkpt++;
2631                         continue; /* try again */
2632                 }
2633                 if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
2634                         kstart = (kstart + PAGE_SIZE * NPTEPG) &
2635                                  ~(PAGE_SIZE * NPTEPG - 1);
2636                         if (kstart - 1 >= kernel_map.max_offset) {
2637                                 kstart = kernel_map.max_offset;
2638                                 break;                       
2639                         }
2640                         continue;
2641                 }
2642
2643                 /*
2644                  * This index is bogus, but out of the way
2645                  */
2646                 nkpg = vm_page_alloc(NULL, nkpt,
2647                                      VM_ALLOC_NORMAL |
2648                                      VM_ALLOC_SYSTEM |
2649                                      VM_ALLOC_INTERRUPT);
2650                 if (nkpg == NULL)
2651                         panic("pmap_growkernel: no memory to grow kernel");
2652
2653                 vm_page_wire(nkpg);
2654                 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2655                 pmap_zero_page(ptppaddr);
2656                 vm_page_flag_clear(nkpg, PG_ZERO);
2657                 newpt = (pd_entry_t) (ptppaddr |
2658                     kernel_pmap.pmap_bits[PG_V_IDX] |
2659                     kernel_pmap.pmap_bits[PG_RW_IDX] |
2660                     kernel_pmap.pmap_bits[PG_A_IDX] |
2661                     kernel_pmap.pmap_bits[PG_M_IDX]);
2662                 *pmap_pt(&kernel_pmap, kstart) = newpt;
2663                 nkpt++;
2664
2665                 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2666                           ~(PAGE_SIZE * NPTEPG - 1);
2667
2668                 if (kstart - 1 >= kernel_map.max_offset) {
2669                         kstart = kernel_map.max_offset;
2670                         break;                       
2671                 }
2672         }
2673
2674         /*
2675          * Only update kernel_vm_end for areas below KERNBASE.
2676          */
2677         if (update_kernel_vm_end && kernel_vm_end < kstart)
2678                 kernel_vm_end = kstart;
2679 }
2680
2681 /*
2682  *      Add a reference to the specified pmap.
2683  */
2684 void
2685 pmap_reference(pmap_t pmap)
2686 {
2687         if (pmap != NULL) {
2688                 lwkt_gettoken(&pmap->pm_token);
2689                 ++pmap->pm_count;
2690                 lwkt_reltoken(&pmap->pm_token);
2691         }
2692 }
2693
2694 /***************************************************
2695  * page management routines.
2696  ***************************************************/
2697
2698 /*
2699  * Hold a pv without locking it
2700  */
2701 static void
2702 pv_hold(pv_entry_t pv)
2703 {
2704         u_int count;
2705
2706         if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2707                 return;
2708
2709         for (;;) {
2710                 count = pv->pv_hold;
2711                 cpu_ccfence();
2712                 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2713                         return;
2714                 /* retry */
2715         }
2716 }
2717
2718 /*
2719  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
2720  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
2721  * the pv properly.
2722  *
2723  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2724  * pv list via its page) must be held by the caller.
2725  */
2726 static int
2727 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
2728 {
2729         u_int count;
2730
2731         /*
2732          * Critical path shortcut expects pv to already have one ref
2733          * (for the pv->pv_pmap).
2734          */
2735         if (atomic_cmpset_int(&pv->pv_hold, 1, PV_HOLD_LOCKED | 2)) {
2736 #ifdef PMAP_DEBUG
2737                 pv->pv_func = func;
2738                 pv->pv_line = lineno;
2739 #endif
2740                 return TRUE;
2741         }
2742
2743         for (;;) {
2744                 count = pv->pv_hold;
2745                 cpu_ccfence();
2746                 if ((count & PV_HOLD_LOCKED) == 0) {
2747                         if (atomic_cmpset_int(&pv->pv_hold, count,
2748                                               (count + 1) | PV_HOLD_LOCKED)) {
2749 #ifdef PMAP_DEBUG
2750                                 pv->pv_func = func;
2751                                 pv->pv_line = lineno;
2752 #endif
2753                                 return TRUE;
2754                         }
2755                 } else {
2756                         if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2757                                 return FALSE;
2758                 }
2759                 /* retry */
2760         }
2761 }
2762
2763 /*
2764  * Drop a previously held pv_entry which could not be locked, allowing its
2765  * destruction.
2766  *
2767  * Must not be called with a spinlock held as we might zfree() the pv if it
2768  * is no longer associated with a pmap and this was the last hold count.
2769  */
2770 static void
2771 pv_drop(pv_entry_t pv)
2772 {
2773         u_int count;
2774
2775         for (;;) {
2776                 count = pv->pv_hold;
2777                 cpu_ccfence();
2778                 KKASSERT((count & PV_HOLD_MASK) > 0);
2779                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2780                          (PV_HOLD_LOCKED | 1));
2781                 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2782                         if ((count & PV_HOLD_MASK) == 1) {
2783                                 KKASSERT(count == 1);
2784                                 KKASSERT(pv->pv_pmap == NULL);
2785                                 zfree(pvzone, pv);
2786                         }
2787                         return;
2788                 }
2789                 /* retry */
2790         }
2791 }
2792
2793 /*
2794  * Find or allocate the requested PV entry, returning a locked, held pv.
2795  *
2796  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
2797  * for the caller and one representing the pmap and vm_page association.
2798  *
2799  * If (*isnew) is zero, the returned pv will have only one hold count.
2800  *
2801  * Since both associations can only be adjusted while the pv is locked,
2802  * together they represent just one additional hold.
2803  */
2804 static
2805 pv_entry_t
2806 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
2807 {
2808         pv_entry_t pv;
2809         pv_entry_t pnew = NULL;
2810
2811         spin_lock(&pmap->pm_spin);
2812         for (;;) {
2813                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2814                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2815                                                         pindex);
2816                 }
2817                 if (pv == NULL) {
2818                         if (pnew == NULL) {
2819                                 spin_unlock(&pmap->pm_spin);
2820                                 pnew = zalloc(pvzone);
2821                                 spin_lock(&pmap->pm_spin);
2822                                 continue;
2823                         }
2824                         pnew->pv_pmap = pmap;
2825                         pnew->pv_pindex = pindex;
2826                         pnew->pv_hold = PV_HOLD_LOCKED | 2;
2827 #ifdef PMAP_DEBUG
2828                         pnew->pv_func = func;
2829                         pnew->pv_line = lineno;
2830 #endif
2831                         pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2832                         ++pmap->pm_generation;
2833                         atomic_add_long(&pmap->pm_stats.resident_count, 1);
2834                         spin_unlock(&pmap->pm_spin);
2835                         *isnew = 1;
2836                         return(pnew);
2837                 }
2838                 if (pnew) {
2839                         spin_unlock(&pmap->pm_spin);
2840                         zfree(pvzone, pnew);
2841                         pnew = NULL;
2842                         spin_lock(&pmap->pm_spin);
2843                         continue;
2844                 }
2845                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2846                         spin_unlock(&pmap->pm_spin);
2847                 } else {
2848                         spin_unlock(&pmap->pm_spin);
2849                         _pv_lock(pv PMAP_DEBUG_COPY);
2850                 }
2851                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2852                         *isnew = 0;
2853                         return(pv);
2854                 }
2855                 pv_put(pv);
2856                 spin_lock(&pmap->pm_spin);
2857         }
2858 }
2859
2860 /*
2861  * Find the requested PV entry, returning a locked+held pv or NULL
2862  */
2863 static
2864 pv_entry_t
2865 _pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2866 {
2867         pv_entry_t pv;
2868
2869         spin_lock(&pmap->pm_spin);
2870         for (;;) {
2871                 /*
2872                  * Shortcut cache
2873                  */
2874                 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2875                         pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2876                                                         pindex);
2877                 }
2878                 if (pv == NULL) {
2879                         spin_unlock(&pmap->pm_spin);
2880                         return NULL;
2881                 }
2882                 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2883                         spin_unlock(&pmap->pm_spin);
2884                 } else {
2885                         spin_unlock(&pmap->pm_spin);
2886                         _pv_lock(pv PMAP_DEBUG_COPY);
2887                 }
2888                 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2889                         pv_cache(pv, pindex);
2890                         return(pv);
2891                 }
2892                 pv_put(pv);
2893                 spin_lock(&pmap->pm_spin);
2894         }
2895 }
2896
2897 /*
2898  * Lookup, hold, and attempt to lock (pmap,pindex).
2899  *
2900  * If the entry does not exist NULL is returned and *errorp is set to 0
2901  *
2902  * If the entry exists and could be successfully locked it is returned and
2903  * errorp is set to 0.
2904  *
2905  * If the entry exists but could NOT be successfully locked it is returned
2906  * held and *errorp is set to 1.
2907  */
2908 static
2909 pv_entry_t
2910 pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
2911 {
2912         pv_entry_t pv;
2913
2914         spin_lock_shared(&pmap->pm_spin);
2915         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2916                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2917         if (pv == NULL) {
2918                 spin_unlock_shared(&pmap->pm_spin);
2919                 *errorp = 0;
2920                 return NULL;
2921         }
2922         if (pv_hold_try(pv)) {
2923                 pv_cache(pv, pindex);
2924                 spin_unlock_shared(&pmap->pm_spin);
2925                 *errorp = 0;
2926                 KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
2927                 return(pv);     /* lock succeeded */
2928         }
2929         spin_unlock_shared(&pmap->pm_spin);
2930         *errorp = 1;
2931         return (pv);            /* lock failed */
2932 }
2933
2934 /*
2935  * Find the requested PV entry, returning a held pv or NULL
2936  */
2937 static
2938 pv_entry_t
2939 pv_find(pmap_t pmap, vm_pindex_t pindex)
2940 {
2941         pv_entry_t pv;
2942
2943         spin_lock_shared(&pmap->pm_spin);
2944
2945         if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2946                 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2947         if (pv == NULL) {
2948                 spin_unlock_shared(&pmap->pm_spin);
2949                 return NULL;
2950         }
2951         pv_hold(pv);
2952         pv_cache(pv, pindex);
2953         spin_unlock_shared(&pmap->pm_spin);
2954         return(pv);
2955 }
2956
2957 /*
2958  * Lock a held pv, keeping the hold count
2959  */
2960 static
2961 void
2962 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2963 {
2964         u_int count;
2965
2966         for (;;) {
2967                 count = pv->pv_hold;
2968                 cpu_ccfence();
2969                 if ((count & PV_HOLD_LOCKED) == 0) {
2970                         if (atomic_cmpset_int(&pv->pv_hold, count,
2971                                               count | PV_HOLD_LOCKED)) {
2972 #ifdef PMAP_DEBUG
2973                                 pv->pv_func = func;
2974                                 pv->pv_line = lineno;
2975 #endif
2976                                 return;
2977                         }
2978                         continue;
2979                 }
2980                 tsleep_interlock(pv, 0);
2981                 if (atomic_cmpset_int(&pv->pv_hold, count,
2982                                       count | PV_HOLD_WAITING)) {
2983 #ifdef PMAP_DEBUG
2984                         kprintf("pv waiting on %s:%d\n",
2985                                         pv->pv_func, pv->pv_line);
2986 #endif
2987                         tsleep(pv, PINTERLOCKED, "pvwait", hz);
2988                 }
2989                 /* retry */
2990         }
2991 }
2992
2993 /*
2994  * Unlock a held and locked pv, keeping the hold count.
2995  */
2996 static
2997 void
2998 pv_unlock(pv_entry_t pv)
2999 {
3000         u_int count;
3001
3002         for (;;) {
3003                 count = pv->pv_hold;
3004                 cpu_ccfence();
3005                 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3006                          (PV_HOLD_LOCKED | 1));
3007                 if (atomic_cmpset_int(&pv->pv_hold, count,
3008                                       count &
3009                                       ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3010                         if (count & PV_HOLD_WAITING)
3011                                 wakeup(pv);
3012                         break;
3013                 }
3014         }
3015 }
3016
3017 /*
3018  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3019  * and the hold count drops to zero we will free it.
3020  *
3021  * Caller should not hold any spin locks.  We are protected from hold races
3022  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3023  * lock held.  A pv cannot be located otherwise.
3024  */
3025 static
3026 void
3027 pv_put(pv_entry_t pv)
3028 {
3029         /*
3030          * Fast - shortcut most common condition
3031          */
3032         if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3033                 return;
3034
3035         /*
3036          * Slow
3037          */
3038         pv_unlock(pv);
3039         pv_drop(pv);
3040 }
3041
3042 /*
3043  * Remove the pmap association from a pv, require that pv_m already be removed,
3044  * then unlock and drop the pv.  Any pte operations must have already been
3045  * completed.  This call may result in a last-drop which will physically free
3046  * the pv.
3047  *
3048  * Removing the pmap association entails an additional drop.
3049  *
3050  * pv must be exclusively locked on call and will be disposed of on return.
3051  */
3052 static
3053 void
3054 pv_free(pv_entry_t pv)
3055 {
3056         pmap_t pmap;
3057
3058         KKASSERT(pv->pv_m == NULL);
3059         KKASSERT((pv->pv_hold & PV_HOLD_MASK) >= 2);
3060         if ((pmap = pv->pv_pmap) != NULL) {
3061                 spin_lock(&pmap->pm_spin);
3062                 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3063                 ++pmap->pm_generation;
3064                 if (pmap->pm_pvhint == pv)
3065                         pmap->pm_pvhint = NULL;
3066                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3067                 pv->pv_pmap = NULL;
3068                 pv->pv_pindex = 0;
3069                 spin_unlock(&pmap->pm_spin);
3070
3071                 /*
3072                  * Try to shortcut three atomic ops, otherwise fall through
3073                  * and do it normally.  Drop two refs and the lock all in
3074                  * one go.
3075                  */
3076                 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3077                         zfree(pvzone, pv);
3078                         return;
3079                 }
3080                 pv_drop(pv);    /* ref for pv_pmap */
3081         }
3082         pv_put(pv);
3083 }
3084
3085 /*
3086  * This routine is very drastic, but can save the system
3087  * in a pinch.
3088  */
3089 void
3090 pmap_collect(void)
3091 {
3092         int i;
3093         vm_page_t m;
3094         static int warningdone=0;
3095
3096         if (pmap_pagedaemon_waken == 0)
3097                 return;
3098         pmap_pagedaemon_waken = 0;
3099         if (warningdone < 5) {
3100                 kprintf("pmap_collect: collecting pv entries -- "
3101                         "suggest increasing PMAP_SHPGPERPROC\n");
3102                 warningdone++;
3103         }
3104
3105         for (i = 0; i < vm_page_array_size; i++) {
3106                 m = &vm_page_array[i];
3107                 if (m->wire_count || m->hold_count)
3108                         continue;
3109                 if (vm_page_busy_try(m, TRUE) == 0) {
3110                         if (m->wire_count == 0 && m->hold_count == 0) {
3111                                 pmap_remove_all(m);
3112                         }
3113                         vm_page_wakeup(m);
3114                 }
3115         }
3116 }
3117
3118 /*
3119  * Scan the pmap for active page table entries and issue a callback.
3120  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3121  * its parent page table.
3122  *
3123  * pte_pv will be NULL if the page or page table is unmanaged.
3124  * pt_pv will point to the page table page containing the pte for the page.
3125  *
3126  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3127  *       we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3128  *       process pmap's PD and page to the callback function.  This can be
3129  *       confusing because the pt_pv is really a pd_pv, and the target page
3130  *       table page is simply aliased by the pmap and not owned by it.
3131  *
3132  * It is assumed that the start and end are properly rounded to the page size.
3133  *
3134  * It is assumed that PD pages and above are managed and thus in the RB tree,
3135  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3136  */
3137 struct pmap_scan_info {
3138         struct pmap *pmap;
3139         vm_offset_t sva;
3140         vm_offset_t eva;
3141         vm_pindex_t sva_pd_pindex;
3142         vm_pindex_t eva_pd_pindex;
3143         void (*func)(pmap_t, struct pmap_scan_info *,
3144                      pv_entry_t, pv_entry_t, int, vm_offset_t,
3145                      pt_entry_t *, void *);
3146         void *arg;
3147         int doinval;
3148         struct pmap_inval_info inval;
3149 };
3150
3151 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3152 static int pmap_scan_callback(pv_entry_t pv, void *data);
3153
3154 static void
3155 pmap_scan(struct pmap_scan_info *info)
3156 {
3157         struct pmap *pmap = info->pmap;
3158         pv_entry_t pd_pv;       /* A page directory PV */
3159         pv_entry_t pt_pv;       /* A page table PV */
3160         pv_entry_t pte_pv;      /* A page table entry PV */
3161         pt_entry_t *ptep;
3162         pt_entry_t oldpte;
3163         struct pv_entry dummy_pv;
3164         int generation;
3165
3166         if (pmap == NULL)
3167                 return;
3168
3169         /*
3170          * Hold the token for stability; if the pmap is empty we have nothing
3171          * to do.
3172          */
3173         lwkt_gettoken(&pmap->pm_token);
3174 #if 0
3175         if (pmap->pm_stats.resident_count == 0) {
3176                 lwkt_reltoken(&pmap->pm_token);
3177                 return;
3178         }
3179 #endif
3180
3181         pmap_inval_init(&info->inval);
3182
3183 again:
3184         /*
3185          * Special handling for scanning one page, which is a very common
3186          * operation (it is?).
3187          *
3188          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3189          */
3190         if (info->sva + PAGE_SIZE == info->eva) {
3191                 generation = pmap->pm_generation;
3192                 if (info->sva >= VM_MAX_USER_ADDRESS) {
3193                         /*
3194                          * Kernel mappings do not track wire counts on
3195                          * page table pages and only maintain pd_pv and
3196                          * pte_pv levels so pmap_scan() works.
3197                          */
3198                         pt_pv = NULL;
3199                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3200                         ptep = vtopte(info->sva);
3201                 } else {
3202                         /*
3203                          * User pages which are unmanaged will not have a
3204                          * pte_pv.  User page table pages which are unmanaged
3205                          * (shared from elsewhere) will also not have a pt_pv.
3206                          * The func() callback will pass both pte_pv and pt_pv
3207                          * as NULL in that case.
3208                          */
3209                         pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
3210                         pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
3211                         if (pt_pv == NULL) {
3212                                 KKASSERT(pte_pv == NULL);
3213                                 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
3214                                 if (pd_pv) {
3215                                         ptep = pv_pte_lookup(pd_pv,
3216                                                     pmap_pt_index(info->sva));
3217                                         if (*ptep) {
3218                                                 info->func(pmap, info,
3219                                                      NULL, pd_pv, 1,
3220                                                      info->sva, ptep,
3221                                                      info->arg);
3222                                         }
3223                                         pv_put(pd_pv);
3224                                 }
3225                                 goto fast_skip;
3226                         }
3227                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
3228                 }
3229
3230                 /*
3231                  * NOTE: *ptep can't be ripped out from under us if we hold
3232                  *       pte_pv locked, but bits can change.  However, there is
3233                  *       a race where another thread may be inserting pte_pv
3234                  *       and setting *ptep just after our pte_pv lookup fails.
3235                  *
3236                  *       In this situation we can end up with a NULL pte_pv
3237                  *       but find that we have a managed *ptep.  We explicitly
3238                  *       check for this race.
3239                  */
3240                 oldpte = *ptep;
3241                 cpu_ccfence();
3242                 if (oldpte == 0) {
3243                         /*
3244                          * Unlike the pv_find() case below we actually
3245                          * acquired a locked pv in this case so any
3246                          * race should have been resolved.  It is expected
3247                          * to not exist.
3248                          */
3249                         KKASSERT(pte_pv == NULL);
3250                 } else if (pte_pv) {
3251                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3252                                            pmap->pmap_bits[PG_V_IDX])) ==
3253                                 (pmap->pmap_bits[PG_MANAGED_IDX] |
3254                                  pmap->pmap_bits[PG_V_IDX]),
3255                             ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3256                              "generation %d/%d",
3257                             *ptep, oldpte, info->sva, pte_pv,
3258                             generation, pmap->pm_generation));
3259                         info->func(pmap, info, pte_pv, pt_pv, 0,
3260                                    info->sva, ptep, info->arg);
3261                 } else {
3262                         /*
3263                          * Check for insertion race
3264                          */
3265                         if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3266                             pt_pv) {
3267                                 pte_pv = pv_find(pmap,
3268                                                  pmap_pte_pindex(info->sva));
3269                                 if (pte_pv) {
3270                                         pv_drop(pte_pv);
3271                                         pv_put(pt_pv);
3272                                         kprintf("pmap_scan: RACE1 "
3273                                                 "%016jx, %016lx\n",
3274                                                 info->sva, oldpte);
3275                                         goto again;
3276                                 }
3277                         }
3278
3279                         /*
3280                          * Didn't race
3281                          */
3282                         KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] |
3283                                            pmap->pmap_bits[PG_V_IDX])) ==
3284                             pmap->pmap_bits[PG_V_IDX],
3285                             ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3286                              "generation %d/%d",
3287                             *ptep, oldpte, info->sva,
3288                             generation, pmap->pm_generation));
3289                         info->func(pmap, info, NULL, pt_pv, 0,
3290                             info->sva, ptep, info->arg);
3291                 }
3292                 if (pt_pv)
3293                         pv_put(pt_pv);
3294 fast_skip:
3295                 pmap_inval_done(&info->inval);
3296                 lwkt_reltoken(&pmap->pm_token);
3297                 return;
3298         }
3299
3300         /*
3301          * Nominal scan case, RB_SCAN() for PD pages and iterate from
3302          * there.
3303          */
3304         info->sva_pd_pindex = pmap_pd_pindex(info->sva);
3305         info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
3306
3307         if (info->sva >= VM_MAX_USER_ADDRESS) {
3308                 /*
3309                  * The kernel does not currently maintain any pv_entry's for
3310                  * higher-level page tables.
3311                  */
3312                 bzero(&dummy_pv, sizeof(dummy_pv));
3313                 dummy_pv.pv_pindex = info->sva_pd_pindex;
3314                 spin_lock(&pmap->pm_spin);
3315                 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
3316                         pmap_scan_callback(&dummy_pv, info);
3317                         ++dummy_pv.pv_pindex;
3318                 }
3319                 spin_unlock(&pmap->pm_spin);
3320         } else {
3321                 /*
3322                  * User page tables maintain local PML4, PDP, and PD
3323                  * pv_entry's at the very least.  PT pv's might be
3324                  * unmanaged and thus not exist.  PTE pv's might be
3325                  * unmanaged and thus not exist.
3326                  */
3327                 spin_lock(&pmap->pm_spin);
3328                 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
3329                         pmap_scan_cmp, pmap_scan_callback, info);
3330                 spin_unlock(&pmap->pm_spin);
3331         }
3332         pmap_inval_done(&info->inval);
3333         lwkt_reltoken(&pmap->pm_token);
3334 }
3335
3336 /*
3337  * WARNING! pmap->pm_spin held
3338  */
3339 static int
3340 pmap_scan_cmp(pv_entry_t pv, void *data)
3341 {
3342         struct pmap_scan_info *info = data;
3343         if (pv->pv_pindex < info->sva_pd_pindex)
3344                 return(-1);
3345         if (pv->pv_pindex >= info->eva_pd_pindex)
3346                 return(1);
3347         return(0);
3348 }
3349
3350 /*
3351  * WARNING! pmap->pm_spin held
3352  */
3353 static int
3354 pmap_scan_callback(pv_entry_t pv, void *data)
3355 {
3356         struct pmap_scan_info *info = data;
3357         struct pmap *pmap = info->pmap;
3358         pv_entry_t pd_pv;       /* A page directory PV */
3359         pv_entry_t pt_pv;       /* A page table PV */
3360         pv_entry_t pte_pv;      /* A page table entry PV */
3361         pt_entry_t *ptep;
3362         pt_entry_t oldpte;
3363         vm_offset_t sva;
3364         vm_offset_t eva;
3365         vm_offset_t va_next;
3366         vm_pindex_t pd_pindex;
3367         int error;
3368         int generation;
3369
3370         /*
3371          * Pull the PD pindex from the pv before releasing the spinlock.
3372          *
3373          * WARNING: pv is faked for kernel pmap scans.
3374          */
3375         pd_pindex = pv->pv_pindex;
3376         spin_unlock(&pmap->pm_spin);
3377         pv = NULL;      /* invalid after spinlock unlocked */
3378
3379         /*
3380          * Calculate the page range within the PD.  SIMPLE pmaps are
3381          * direct-mapped for the entire 2^64 address space.  Normal pmaps
3382          * reflect the user and kernel address space which requires
3383          * cannonicalization w/regards to converting pd_pindex's back
3384          * into addresses.
3385          */
3386         sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
3387         if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
3388             (sva & PML4_SIGNMASK)) {
3389                 sva |= PML4_SIGNMASK;
3390         }
3391         eva = sva + NBPDP;      /* can overflow */
3392         if (sva < info->sva)
3393                 sva = info->sva;
3394         if (eva < info->sva || eva > info->eva)
3395                 eva = info->eva;
3396
3397         /*
3398          * NOTE: kernel mappings do not track page table pages, only
3399          *       terminal pages.
3400          *
3401          * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3402          *       However, for the scan to be efficient we try to
3403          *       cache items top-down.
3404          */
3405         pd_pv = NULL;
3406         pt_pv = NULL;
3407
3408         for (; sva < eva; sva = va_next) {
3409                 if (sva >= VM_MAX_USER_ADDRESS) {
3410                         if (pt_pv) {
3411                                 pv_put(pt_pv);
3412                                 pt_pv = NULL;
3413                         }
3414                         goto kernel_skip;
3415                 }
3416
3417                 /*
3418                  * PD cache (degenerate case if we skip).  It is possible
3419                  * for the PD to not exist due to races.  This is ok.
3420                  */
3421                 if (pd_pv == NULL) {
3422                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3423                 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3424                         pv_put(pd_pv);
3425                         pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3426                 }
3427                 if (pd_pv == NULL) {
3428                         va_next = (sva + NBPDP) & ~PDPMASK;
3429                         if (va_next < sva)
3430                                 va_next = eva;
3431                         continue;
3432                 }
3433
3434                 /*
3435                  * PT cache
3436                  */
3437                 if (pt_pv == NULL) {
3438                         if (pd_pv) {
3439                                 pv_put(pd_pv);
3440                                 pd_pv = NULL;
3441                         }
3442                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3443                 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
3444                         if (pd_pv) {
3445                                 pv_put(pd_pv);
3446                                 pd_pv = NULL;
3447                         }
3448                         pv_put(pt_pv);
3449                         pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3450                 }
3451
3452                 /*
3453                  * If pt_pv is NULL we either have an shared page table
3454                  * page and must issue a callback specific to that case,
3455                  * or there is no page table page.
3456                  *
3457                  * Either way we can skip the page table page.
3458                  */
3459                 if (pt_pv == NULL) {
3460                         /*
3461                          * Possible unmanaged (shared from another pmap)
3462                          * page table page.
3463                          */
3464                         if (pd_pv == NULL)
3465                                 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3466                         KKASSERT(pd_pv != NULL);
3467                         ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3468                         if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
3469                                 info->func(pmap, info, NULL, pd_pv, 1,
3470                                            sva, ptep, info->arg);
3471                         }
3472
3473                         /*
3474                          * Done, move to next page table page.
3475                          */
3476                         va_next = (sva + NBPDR) & ~PDRMASK;
3477                         if (va_next < sva)
3478                                 va_next = eva;
3479                         continue;
3480                 }
3481
3482                 /*
3483                  * From this point in the loop testing pt_pv for non-NULL
3484                  * means we are in UVM, else if it is NULL we are in KVM.
3485                  *
3486                  * Limit our scan to either the end of the va represented
3487                  * by the current page table page, or to the end of the
3488                  * range being removed.
3489                  */
3490 kernel_skip:
3491                 va_next = (sva + NBPDR) & ~PDRMASK;
3492                 if (va_next < sva)
3493                         va_next = eva;
3494                 if (va_next > eva)
3495                         va_next = eva;
3496
3497                 /*
3498                  * Scan the page table for pages.  Some pages may not be
3499                  * managed (might not have a pv_entry).
3500                  *
3501                  * There is no page table management for kernel pages so
3502                  * pt_pv will be NULL in that case, but otherwise pt_pv
3503                  * is non-NULL, locked, and referenced.
3504                  */
3505
3506                 /*
3507                  * At this point a non-NULL pt_pv means a UVA, and a NULL
3508                  * pt_pv means a KVA.
3509                  */
3510                 if (pt_pv)
3511                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3512                 else
3513                         ptep = vtopte(sva);
3514
3515                 while (sva < va_next) {
3516                         /*
3517                          * Acquire the related pte_pv, if any.  If *ptep == 0
3518                          * the related pte_pv should not exist, but if *ptep
3519                          * is not zero the pte_pv may or may not exist (e.g.
3520                          * will not exist for an unmanaged page).
3521                          *
3522                          * However a multitude of races are possible here.
3523                          *
3524                          * In addition, the (pt_pv, pte_pv) lock order is
3525                          * backwards, so we have to be careful in aquiring
3526                          * a properly locked pte_pv.
3527                          */
3528                         generation = pmap->pm_generation;
3529                         if (pt_pv) {
3530                                 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3531                                                     &error);
3532                                 if (error) {
3533                                         if (pd_pv) {
3534                                                 pv_put(pd_pv);
3535                                                 pd_pv = NULL;
3536                                         }
3537                                         pv_put(pt_pv);   /* must be non-NULL */
3538                                         pt_pv = NULL;
3539                                         pv_lock(pte_pv); /* safe to block now */
3540                                         pv_put(pte_pv);
3541                                         pte_pv = NULL;
3542                                         pt_pv = pv_get(pmap,
3543                                                        pmap_pt_pindex(sva));
3544                                         /*
3545                                          * pt_pv reloaded, need new ptep
3546                                          */
3547                                         KKASSERT(pt_pv != NULL);
3548                                         ptep = pv_pte_lookup(pt_pv,
3549                                                         pmap_pte_index(sva));
3550                                         continue;
3551                                 }
3552                         } else {
3553                                 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3554                         }
3555
3556                         /*
3557                          * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3558                          */
3559                         oldpte = *ptep;
3560                         if (oldpte == 0) {
3561                                 if (pte_pv) {
3562                                         kprintf("Unexpected non-NULL pte_pv "
3563                                                 "%p pt_pv %p "
3564                                                 "*ptep = %016lx/%016lx\n",
3565                                                 pte_pv, pt_pv, *ptep, oldpte);
3566                                         panic("Unexpected non-NULL pte_pv");
3567                                 }
3568                                 sva += PAGE_SIZE;
3569                                 ++ptep;
3570                                 continue;
3571                         }
3572
3573                         /*
3574                          * Ready for the callback.  The locked pte_pv (if any)
3575                          * is consumed by the callback.  pte_pv will exist if
3576                          *  the page is managed, and will not exist if it
3577                          * isn't.
3578                          */
3579                         if (pte_pv) {
3580                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3581                                     (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX]),
3582                                     ("badC *ptep %016lx/%016lx sva %016lx "
3583                                     "pte_pv %p pm_generation %d/%d",
3584                                     *ptep, oldpte, sva, pte_pv,
3585                                     generation, pmap->pm_generation));
3586                                 info->func(pmap, info, pte_pv, pt_pv, 0,
3587                                     sva, ptep, info->arg);
3588                         } else {
3589                                 /*
3590                                  * Check for insertion race.  Since there is no
3591                                  * pte_pv to guard us it is possible for us
3592                                  * to race another thread doing an insertion.
3593                                  * Our lookup misses the pte_pv but our *ptep
3594                                  * check sees the inserted pte.
3595                                  *
3596                                  * XXX panic case seems to occur within a
3597                                  * vm_fork() of /bin/sh, which frankly
3598                                  * shouldn't happen since no other threads
3599                                  * should be inserting to our pmap in that
3600                                  * situation.  Removing, possibly.  Inserting,
3601                                  * shouldn't happen.
3602                                  */
3603                                 if ((oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) &&
3604                                     pt_pv) {
3605                                         pte_pv = pv_find(pmap,
3606                                                          pmap_pte_pindex(sva));
3607                                         if (pte_pv) {
3608                                                 pv_drop(pte_pv);
3609                                                 kprintf("pmap_scan: RACE2 "
3610                                                         "%016jx, %016lx\n",
3611                                                         sva, oldpte);
3612                                                 continue;
3613                                         }
3614                                 }
3615
3616                                 /*
3617                                  * Didn't race
3618                                  */
3619                                 KASSERT((oldpte & (pmap->pmap_bits[PG_MANAGED_IDX] | pmap->pmap_bits[PG_V_IDX])) ==
3620                                     pmap->pmap_bits[PG_V_IDX],
3621                                     ("badD *ptep %016lx/%016lx sva %016lx "
3622                                     "pte_pv NULL pm_generation %d/%d",
3623                                      *ptep, oldpte, sva,
3624                                      generation, pmap->pm_generation));
3625                                 info->func(pmap, info, NULL, pt_pv, 0,
3626                                     sva, ptep, info->arg);
3627                         }
3628                         pte_pv = NULL;
3629                         sva += PAGE_SIZE;
3630                         ++ptep;
3631                 }
3632                 lwkt_yield();
3633         }
3634         if (pd_pv) {
3635                 pv_put(pd_pv);
3636                 pd_pv = NULL;
3637         }
3638         if (pt_pv) {
3639                 pv_put(pt_pv);
3640                 pt_pv = NULL;
3641         }
3642         lwkt_yield();
3643
3644         /*
3645          * Relock before returning.
3646          */
3647         spin_lock(&pmap->pm_spin);
3648         return (0);
3649 }
3650
3651 void
3652 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3653 {
3654         struct pmap_scan_info info;
3655
3656         info.pmap = pmap;
3657         info.sva = sva;
3658         info.eva = eva;
3659         info.func = pmap_remove_callback;
3660         info.arg = NULL;
3661         info.doinval = 1;       /* normal remove requires pmap inval */
3662         pmap_scan(&info);
3663 }
3664
3665 static void
3666 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3667 {
3668         struct pmap_scan_info info;
3669
3670         info.pmap = pmap;
3671         info.sva = sva;
3672         info.eva = eva;
3673         info.func = pmap_remove_callback;
3674         info.arg = NULL;
3675         info.doinval = 0;       /* normal remove requires pmap inval */
3676         pmap_scan(&info);
3677 }
3678
3679 static void
3680 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
3681                      pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3682                      vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3683 {
3684         pt_entry_t pte;
3685
3686         if (pte_pv) {
3687                 /*
3688                  * This will also drop pt_pv's wire_count. Note that
3689                  * terminal pages are not wired based on mmu presence.
3690                  */
3691                 if (info->doinval)
3692                         pmap_remove_pv_pte(pte_pv, pt_pv, &info->inval);
3693                 else
3694                         pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
3695                 pmap_remove_pv_page(pte_pv);
3696                 pv_free(pte_pv);
3697         } else if (sharept == 0) {
3698                 /*
3699                  * Unmanaged page
3700                  *
3701                  * pt_pv's wire_count is still bumped by unmanaged pages
3702                  * so we must decrement it manually.
3703                  */
3704                 if (info->doinval)
3705                         pmap_inval_interlock(&info->inval, pmap, va);
3706                 pte = pte_load_clear(ptep);
3707                 if (info->doinval)
3708                         pmap_inval_deinterlock(&info->inval, pmap);
3709                 if (pte & pmap->pmap_bits[PG_W_IDX])
3710                         atomic_add_long(&pmap->pm_stats.wired_count, -1);
3711                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3712                 if (vm_page_unwire_quick(pt_pv->pv_m))
3713                         panic("pmap_remove: insufficient wirecount");
3714         } else {
3715                 /*
3716                  * Unmanaged page table, pt_pv is actually the pd_pv
3717                  * for our pmap (not the share object pmap).
3718                  *
3719                  * We have to unwire the target page table page and we
3720                  * have to unwire our page directory page.
3721                  */
3722                 if (info->doinval)
3723                         pmap_inval_interlock(&info->inval, pmap, va);
3724                 pte = pte_load_clear(ptep);
3725                 if (info->doinval)
3726                         pmap_inval_deinterlock(&info->inval, pmap);
3727                 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3728                 KKASSERT((pte & pmap->pmap_bits[PG_DEVICE_IDX]) == 0);
3729                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3730                         panic("pmap_remove: shared pgtable1 bad wirecount");
3731                 if (vm_page_unwire_quick(pt_pv->pv_m))
3732                         panic("pmap_remove: shared pgtable2 bad wirecount");
3733         }
3734 }
3735
3736 /*
3737  * Removes this physical page from all physical maps in which it resides.
3738  * Reflects back modify bits to the pager.
3739  *
3740  * This routine may not be called from an interrupt.
3741  */
3742 static
3743 void
3744 pmap_remove_all(vm_page_t m)
3745 {
3746         struct pmap_inval_info info;
3747         pv_entry_t pv;
3748
3749         if (!pmap_initialized /* || (m->flags & PG_FICTITIOUS)*/)
3750                 return;
3751
3752         pmap_inval_init(&info);
3753         vm_page_spin_lock(m);
3754         while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3755                 KKASSERT(pv->pv_m == m);
3756                 if (pv_hold_try(pv)) {
3757                         vm_page_spin_unlock(m);
3758                 } else {
3759                         vm_page_spin_unlock(m);
3760                         pv_lock(pv);
3761                 }
3762                 if (pv->pv_m != m) {
3763                         pv_put(pv);
3764                         vm_page_spin_lock(m);
3765                         continue;
3766                 }
3767                 /*
3768                  * Holding no spinlocks, pv is locked.
3769                  */
3770                 pmap_remove_pv_pte(pv, NULL, &info);
3771                 pmap_remove_pv_page(pv);
3772                 pv_free(pv);
3773                 vm_page_spin_lock(m);
3774         }
3775         KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
3776         vm_page_spin_unlock(m);
3777         pmap_inval_done(&info);
3778 }
3779
3780 /*
3781  * Set the physical protection on the specified range of this map
3782  * as requested.  This function is typically only used for debug watchpoints
3783  * and COW pages.
3784  *
3785  * This function may not be called from an interrupt if the map is
3786  * not the kernel_pmap.
3787  *
3788  * NOTE!  For shared page table pages we just unmap the page.
3789  */
3790 void
3791 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3792 {
3793         struct pmap_scan_info info;
3794         /* JG review for NX */
3795
3796         if (pmap == NULL)
3797                 return;
3798         if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3799                 pmap_remove(pmap, sva, eva);
3800                 return;
3801         }
3802         if (prot & VM_PROT_WRITE)
3803                 return;
3804         info.pmap = pmap;
3805         info.sva = sva;
3806         info.eva = eva;
3807         info.func = pmap_protect_callback;
3808         info.arg = &prot;
3809         info.doinval = 1;
3810         pmap_scan(&info);
3811 }
3812
3813 static
3814 void
3815 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
3816                       pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3817                       vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
3818 {
3819         pt_entry_t pbits;
3820         pt_entry_t cbits;
3821         pt_entry_t pte;
3822         vm_page_t m;
3823
3824         /*
3825          * XXX non-optimal.
3826          */
3827         pmap_inval_interlock(&info->inval, pmap, va);
3828 again:
3829         pbits = *ptep;
3830         cbits = pbits;
3831         if (pte_pv) {
3832                 m = NULL;
3833                 if (pbits & pmap->pmap_bits[PG_A_IDX]) {
3834                         if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3835                                 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3836                                 KKASSERT(m == pte_pv->pv_m);
3837                                 vm_page_flag_set(m, PG_REFERENCED);
3838                         }
3839                         cbits &= ~pmap->pmap_bits[PG_A_IDX];
3840                 }
3841                 if (pbits & pmap->pmap_bits[PG_M_IDX]) {
3842                         if (pmap_track_modified(pte_pv->pv_pindex)) {
3843                                 if ((pbits & pmap->pmap_bits[PG_DEVICE_IDX]) == 0) {
3844                                         if (m == NULL) {
3845                                                 m = PHYS_TO_VM_PAGE(pbits &
3846                                                                     PG_FRAME);
3847                                         }
3848                                         vm_page_dirty(m);
3849                                 }
3850                                 cbits &= ~pmap->pmap_bits[PG_M_IDX];
3851                         }
3852                 }
3853         } else if (sharept) {
3854                 /*
3855                  * Unmanaged page table, pt_pv is actually the pd_pv
3856                  * for our pmap (not the share object pmap).
3857                  *
3858                  * When asked to protect something in a shared page table
3859                  * page we just unmap the page table page.  We have to
3860                  * invalidate the tlb in this situation.
3861                  *
3862                  * XXX Warning, shared page tables will not be used for
3863                  * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
3864                  * so PHYS_TO_VM_PAGE() should be safe here.
3865                  */
3866                 pte = pte_load_clear(ptep);
3867                 pmap_inval_invltlb(&info->inval);
3868                 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3869                         panic("pmap_protect: pgtable1 pg bad wirecount");
3870                 if (vm_page_unwire_quick(pt_pv->pv_m))
3871                         panic("pmap_protect: pgtable2 pg bad wirecount");
3872                 ptep = NULL;
3873         }
3874         /* else unmanaged page, adjust bits, no wire changes */
3875
3876         if (ptep) {
3877                 cbits &= ~pmap->pmap_bits[PG_RW_IDX];
3878                 if (pbits != cbits && !atomic_cmpset_long(ptep, pbits, cbits)) {
3879                         goto again;
3880                 }
3881         }
3882         pmap_inval_deinterlock(&info->inval, pmap);
3883         if (pte_pv)
3884                 pv_put(pte_pv);
3885 }
3886
3887 /*
3888  * Insert the vm_page (m) at the virtual address (va), replacing any prior
3889  * mapping at that address.  Set protection and wiring as requested.
3890  *
3891  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
3892  * possible.  If it is we enter the page into the appropriate shared pmap
3893  * hanging off the related VM object instead of the passed pmap, then we
3894  * share the page table page from the VM object's pmap into the current pmap.
3895  *
3896  * NOTE: This routine MUST insert the page into the pmap now, it cannot
3897  *       lazy-evaluate.
3898  */
3899 void
3900 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3901            boolean_t wired, vm_map_entry_t entry)
3902 {
3903         pmap_inval_info info;
3904         pv_entry_t pt_pv;       /* page table */
3905         pv_entry_t pte_pv;      /* page table entry */
3906         pt_entry_t *ptep;
3907         vm_paddr_t opa;
3908         pt_entry_t origpte, newpte;
3909         vm_paddr_t pa;
3910
3911         if (pmap == NULL)
3912                 return;
3913         va = trunc_page(va);
3914 #ifdef PMAP_DIAGNOSTIC
3915         if (va >= KvaEnd)
3916                 panic("pmap_enter: toobig");
3917         if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
3918                 panic("pmap_enter: invalid to pmap_enter page table "
3919                       "pages (va: 0x%lx)", va);
3920 #endif
3921         if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
3922                 kprintf("Warning: pmap_enter called on UVA with "
3923                         "kernel_pmap\n");
3924 #ifdef DDB
3925                 db_print_backtrace();
3926 #endif
3927         }
3928         if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
3929                 kprintf("Warning: pmap_enter called on KVA without"
3930                         "kernel_pmap\n");
3931 #ifdef DDB
3932                 db_print_backtrace();
3933 #endif
3934         }
3935
3936         /*
3937          * Get locked PV entries for our new page table entry (pte_pv)
3938          * and for its parent page table (pt_pv).  We need the parent
3939          * so we can resolve the location of the ptep.
3940          *
3941          * Only hardware MMU actions can modify the ptep out from
3942          * under us.
3943          *
3944          * if (m) is fictitious or unmanaged we do not create a managing
3945          * pte_pv for it.  Any pre-existing page's management state must
3946          * match (avoiding code complexity).
3947          *
3948          * If the pmap is still being initialized we assume existing
3949          * page tables.
3950          *
3951          * Kernel mapppings do not track page table pages (i.e. pt_pv).
3952          */
3953         if (pmap_initialized == FALSE) {
3954                 pte_pv = NULL;
3955                 pt_pv = NULL;
3956                 ptep = vtopte(va);
3957                 origpte = *ptep;
3958         } else if (m->flags & (/*PG_FICTITIOUS |*/ PG_UNMANAGED)) { /* XXX */
3959                 pte_pv = NULL;
3960                 if (va >= VM_MAX_USER_ADDRESS) {
3961                         pt_pv = NULL;
3962                         ptep = vtopte(va);
3963                 } else {
3964                         pt_pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va),
3965                                                   NULL, entry, va);
3966                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3967                 }
3968                 origpte = *ptep;
3969                 cpu_ccfence();
3970                 KKASSERT(origpte == 0 ||
3971                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0);
3972         } else {
3973                 if (va >= VM_MAX_USER_ADDRESS) {
3974                         /*
3975                          * Kernel map, pv_entry-tracked.
3976                          */
3977                         pt_pv = NULL;
3978                         pte_pv = pmap_allocpte(pmap, pmap_pte_pindex(va), NULL);
3979                         ptep = vtopte(va);
3980                 } else {
3981                         /*
3982                          * User map
3983                          */
3984                         pte_pv = pmap_allocpte_seg(pmap, pmap_pte_pindex(va),
3985                                                    &pt_pv, entry, va);
3986                         ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
3987                 }
3988                 origpte = *ptep;
3989                 cpu_ccfence();
3990                 KKASSERT(origpte == 0 ||
3991                          (origpte & pmap->pmap_bits[PG_MANAGED_IDX]));
3992         }
3993
3994         pa = VM_PAGE_TO_PHYS(m);
3995         opa = origpte & PG_FRAME;
3996
3997         newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
3998                  pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
3999         if (wired)
4000                 newpte |= pmap->pmap_bits[PG_W_IDX];
4001         if (va < VM_MAX_USER_ADDRESS)
4002                 newpte |= pmap->pmap_bits[PG_U_IDX];
4003         if (pte_pv)
4004                 newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4005 //      if (pmap == &kernel_pmap)
4006 //              newpte |= pgeflag;
4007         newpte |= pmap->pmap_cache_bits[m->pat_mode];
4008         if (m->flags & PG_FICTITIOUS)
4009                 newpte |= pmap->pmap_bits[PG_DEVICE_IDX];
4010
4011         /*
4012          * It is possible for multiple faults to occur in threaded
4013          * environments, the existing pte might be correct.
4014          */
4015         if (((origpte ^ newpte) & ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4016             pmap->pmap_bits[PG_A_IDX])) == 0)
4017                 goto done;
4018
4019         if ((prot & VM_PROT_NOSYNC) == 0)
4020                 pmap_inval_init(&info);
4021
4022         /*
4023          * Ok, either the address changed or the protection or wiring
4024          * changed.
4025          *
4026          * Clear the current entry, interlocking the removal.  For managed
4027          * pte's this will also flush the modified state to the vm_page.
4028          * Atomic ops are mandatory in order to ensure that PG_M events are
4029          * not lost during any transition.
4030          *
4031          * WARNING: The caller has busied the new page but not the original
4032          *          vm_page which we are trying to replace.  Because we hold
4033          *          the pte_pv lock, but have not busied the page, PG bits
4034          *          can be cleared out from under us.
4035          */
4036         if (opa) {
4037                 if (pte_pv) {
4038                         /*
4039                          * pmap_remove_pv_pte() unwires pt_pv and assumes
4040                          * we will free pte_pv, but since we are reusing
4041                          * pte_pv we want to retain the wire count.
4042                          *
4043                          * pt_pv won't exist for a kernel page (managed or
4044                          * otherwise).
4045                          */
4046                         if (pt_pv)
4047                                 vm_page_wire_quick(pt_pv->pv_m);
4048                         if (prot & VM_PROT_NOSYNC)
4049                                 pmap_remove_pv_pte(pte_pv, pt_pv, NULL);
4050                         else
4051                                 pmap_remove_pv_pte(pte_pv, pt_pv, &info);
4052                         if (pte_pv->pv_m)
4053                                 pmap_remove_pv_page(pte_pv);
4054                 } else if (prot & VM_PROT_NOSYNC) {
4055                         /*
4056                          * Unmanaged page, NOSYNC (no mmu sync) requested.
4057                          *
4058                          * Leave wire count on PT page intact.
4059                          */
4060                         (void)pte_load_clear(ptep);
4061                         cpu_invlpg((void *)va);
4062                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4063                 } else {
4064                         /*
4065                          * Unmanaged page, normal enter.
4066                          *
4067                          * Leave wire count on PT page intact.
4068                          */
4069                         pmap_inval_interlock(&info, pmap, va);
4070                         (void)pte_load_clear(ptep);
4071                         pmap_inval_deinterlock(&info, pmap);
4072                         atomic_add_long(&pmap->pm_stats.resident_count, -1);
4073                 }
4074                 KKASSERT(*ptep == 0);
4075         }
4076
4077         if (pte_pv) {
4078                 /*
4079                  * Enter on the PV list if part of our managed memory.
4080                  * Wiring of the PT page is already handled.
4081                  */
4082                 KKASSERT(pte_pv->pv_m == NULL);
4083                 vm_page_spin_lock(m);
4084                 pte_pv->pv_m = m;
4085                 TAILQ_INSERT_TAIL(&m->md.pv_list, pte_pv, pv_list);
4086                 /*
4087                 if (m->object)
4088                         atomic_add_int(&m->object->agg_pv_list_count, 1);
4089                 */
4090                 vm_page_flag_set(m, PG_MAPPED);
4091                 vm_page_spin_unlock(m);
4092         } else if (pt_pv && opa == 0) {
4093                 /*
4094                  * We have to adjust the wire count on the PT page ourselves
4095                  * for unmanaged entries.  If opa was non-zero we retained
4096                  * the existing wire count from the removal.
4097                  */
4098                 vm_page_wire_quick(pt_pv->pv_m);
4099         }
4100
4101         /*
4102          * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4103          *
4104          * User VMAs do not because those will be zero->non-zero, so no
4105          * stale entries to worry about at this point.
4106          *
4107          * For KVM there appear to still be issues.  Theoretically we
4108          * should be able to scrap the interlocks entirely but we
4109          * get crashes.
4110          */
4111         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
4112                 pmap_inval_interlock(&info, pmap, va);
4113
4114         /*
4115          * Set the pte
4116          */
4117         *(volatile pt_entry_t *)ptep = newpte;
4118
4119         if ((prot & VM_PROT_NOSYNC) == 0 && pt_pv == NULL)
4120                 pmap_inval_deinterlock(&info, pmap);
4121         else if (pt_pv == NULL)
4122                 cpu_invlpg((void *)va);
4123
4124         if (wired) {
4125                 if (pte_pv) {
4126                         atomic_add_long(&pte_pv->pv_pmap->pm_stats.wired_count,
4127                                         1);
4128                 } else {
4129                         atomic_add_long(&pmap->pm_stats.wired_count, 1);
4130                 }
4131         }
4132         if (newpte & pmap->pmap_bits[PG_RW_IDX])
4133                 vm_page_flag_set(m, PG_WRITEABLE);
4134
4135         /*
4136          * Unmanaged pages need manual resident_count tracking.
4137          */
4138         if (pte_pv == NULL && pt_pv)
4139                 atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
4140
4141         /*
4142          * Cleanup
4143          */
4144         if ((prot & VM_PROT_NOSYNC) == 0 || pte_pv == NULL)
4145                 pmap_inval_done(&info);
4146 done:
4147         KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
4148                  (m->flags & PG_MAPPED));
4149
4150         /*
4151          * Cleanup the pv entry, allowing other accessors.
4152          */
4153         if (pte_pv)
4154                 pv_put(pte_pv);
4155         if (pt_pv)
4156                 pv_put(pt_pv);
4157 }
4158
4159 /*
4160  * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4161  * This code also assumes that the pmap has no pre-existing entry for this
4162  * VA.
4163  *
4164  * This code currently may only be used on user pmaps, not kernel_pmap.
4165  */
4166 void
4167 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
4168 {
4169         pmap_enter(pmap, va, m, VM_PROT_READ, FALSE, NULL);
4170 }
4171
4172 /*
4173  * Make a temporary mapping for a physical address.  This is only intended
4174  * to be used for panic dumps.
4175  *
4176  * The caller is responsible for calling smp_invltlb().
4177  */
4178 void *
4179 pmap_kenter_temporary(vm_paddr_t pa, long i)
4180 {
4181         pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
4182         return ((void *)crashdumpmap);
4183 }
4184
4185 #define MAX_INIT_PT (96)
4186
4187 /*
4188  * This routine preloads the ptes for a given object into the specified pmap.
4189  * This eliminates the blast of soft faults on process startup and
4190  * immediately after an mmap.
4191  */
4192 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
4193
4194 void
4195 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
4196                     vm_object_t object, vm_pindex_t pindex,
4197                     vm_size_t size, int limit)
4198 {
4199         struct rb_vm_page_scan_info info;
4200         struct lwp *lp;
4201         vm_size_t psize;
4202
4203         /*
4204          * We can't preinit if read access isn't set or there is no pmap
4205          * or object.
4206          */
4207         if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
4208                 return;
4209
4210         /*
4211          * We can't preinit if the pmap is not the current pmap
4212          */
4213         lp = curthread->td_lwp;
4214         if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
4215                 return;
4216
4217         /*
4218          * Misc additional checks
4219          */
4220         psize = x86_64_btop(size);
4221
4222         if ((object->type != OBJT_VNODE) ||
4223                 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
4224                         (object->resident_page_count > MAX_INIT_PT))) {
4225                 return;
4226         }
4227
4228         if (pindex + psize > object->size) {
4229                 if (object->size < pindex)
4230                         return;           
4231                 psize = object->size - pindex;
4232         }
4233
4234         if (psize == 0)
4235                 return;
4236
4237         /*
4238          * If everything is segment-aligned do not pre-init here.  Instead
4239          * allow the normal vm_fault path to pass a segment hint to
4240          * pmap_enter() which will then use an object-referenced shared
4241          * page table page.
4242          */
4243         if ((addr & SEG_MASK) == 0 &&
4244             (ctob(psize) & SEG_MASK) == 0 &&
4245             (ctob(pindex) & SEG_MASK) == 0) {
4246                 return;
4247         }
4248
4249         /*
4250          * Use a red-black scan to traverse the requested range and load
4251          * any valid pages found into the pmap.
4252          *
4253          * We cannot safely scan the object's memq without holding the
4254          * object token.
4255          */
4256         info.start_pindex = pindex;
4257         info.end_pindex = pindex + psize - 1;
4258         info.limit = limit;
4259         info.mpte = NULL;
4260         info.addr = addr;
4261         info.pmap = pmap;
4262
4263         vm_object_hold_shared(object);
4264         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
4265                                 pmap_object_init_pt_callback, &info);
4266         vm_object_drop(object);
4267 }
4268
4269 static
4270 int
4271 pmap_object_init_pt_callback(vm_page_t p, void *data)
4272 {
4273         struct rb_vm_page_scan_info *info = data;
4274         vm_pindex_t rel_index;
4275
4276         /*
4277          * don't allow an madvise to blow away our really
4278          * free pages allocating pv entries.
4279          */
4280         if ((info->limit & MAP_PREFAULT_MADVISE) &&
4281                 vmstats.v_free_count < vmstats.v_free_reserved) {
4282                     return(-1);
4283         }
4284
4285         /*
4286          * Ignore list markers and ignore pages we cannot instantly
4287          * busy (while holding the object token).
4288          */
4289         if (p->flags & PG_MARKER)
4290                 return 0;
4291         if (vm_page_busy_try(p, TRUE))
4292                 return 0;
4293         if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
4294             (p->flags & PG_FICTITIOUS) == 0) {
4295                 if ((p->queue - p->pc) == PQ_CACHE)
4296                         vm_page_deactivate(p);
4297                 rel_index = p->pindex - info->start_pindex;
4298                 pmap_enter_quick(info->pmap,
4299                                  info->addr + x86_64_ptob(rel_index), p);
4300         }
4301         vm_page_wakeup(p);
4302         lwkt_yield();
4303         return(0);
4304 }
4305
4306 /*
4307  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4308  * address.
4309  *
4310  * Returns FALSE if it would be non-trivial or if a pte is already loaded
4311  * into the slot.
4312  *
4313  * XXX This is safe only because page table pages are not freed.
4314  */
4315 int
4316 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
4317 {
4318         pt_entry_t *pte;
4319
4320         /*spin_lock(&pmap->pm_spin);*/
4321         if ((pte = pmap_pte(pmap, addr)) != NULL) {
4322                 if (*pte & pmap->pmap_bits[PG_V_IDX]) {
4323                         /*spin_unlock(&pmap->pm_spin);*/
4324                         return FALSE;
4325                 }
4326         }
4327         /*spin_unlock(&pmap->pm_spin);*/
4328         return TRUE;
4329 }
4330
4331 /*
4332  * Change the wiring attribute for a pmap/va pair.  The mapping must already
4333  * exist in the pmap.  The mapping may or may not be managed.
4334  */
4335 void
4336 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
4337                    vm_map_entry_t entry)
4338 {
4339         pt_entry_t *ptep;
4340         pv_entry_t pv;
4341
4342         if (pmap == NULL)
4343                 return;
4344         lwkt_gettoken(&pmap->pm_token);
4345         pv = pmap_allocpte_seg(pmap, pmap_pt_pindex(va), NULL, entry, va);
4346         ptep = pv_pte_lookup(pv, pmap_pte_index(va));
4347
4348         if (wired && !pmap_pte_w(pmap, ptep))
4349                 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, 1);
4350         else if (!wired && pmap_pte_w(pmap, ptep))
4351                 atomic_add_long(&pv->pv_pmap->pm_stats.wired_count, -1);
4352
4353         /*
4354          * Wiring is not a hardware characteristic so there is no need to
4355          * invalidate TLB.  However, in an SMP environment we must use
4356          * a locked bus cycle to update the pte (if we are not using 
4357          * the pmap_inval_*() API that is)... it's ok to do this for simple
4358          * wiring changes.
4359          */
4360         if (wired)
4361                 atomic_set_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4362         else
4363                 atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
4364         pv_put(pv);
4365         lwkt_reltoken(&pmap->pm_token);
4366 }
4367
4368
4369
4370 /*
4371  * Copy the range specified by src_addr/len from the source map to
4372  * the range dst_addr/len in the destination map.
4373  *
4374  * This routine is only advisory and need not do anything.
4375  */
4376 void
4377 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, 
4378           vm_size_t len, vm_offset_t src_addr)
4379 {
4380 }       
4381
4382 /*
4383  * pmap_zero_page:
4384  *
4385  *      Zero the specified physical page.
4386  *
4387  *      This function may be called from an interrupt and no locking is
4388  *      required.
4389  */
4390 void
4391 pmap_zero_page(vm_paddr_t phys)
4392 {
4393         vm_offset_t va = PHYS_TO_DMAP(phys);
4394
4395         pagezero((void *)va);
4396 }
4397
4398 /*
4399  * pmap_page_assertzero:
4400  *
4401  *      Assert that a page is empty, panic if it isn't.
4402  */
4403 void
4404 pmap_page_assertzero(vm_paddr_t phys)
4405 {
4406         vm_offset_t va = PHYS_TO_DMAP(phys);
4407         size_t i;
4408
4409         for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
4410                 if (*(long *)((char *)va + i) != 0) {
4411                         panic("pmap_page_assertzero() @ %p not zero!",
4412                               (void *)(intptr_t)va);
4413                 }
4414         }
4415 }
4416
4417 /*
4418  * pmap_zero_page:
4419  *
4420  *      Zero part of a physical page by mapping it into memory and clearing
4421  *      its contents with bzero.
4422  *
4423  *      off and size may not cover an area beyond a single hardware page.
4424  */
4425 void
4426 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
4427 {
4428         vm_offset_t virt = PHYS_TO_DMAP(phys);
4429
4430         bzero((char *)virt + off, size);
4431 }
4432
4433 /*
4434  * pmap_copy_page:
4435  *
4436  *      Copy the physical page from the source PA to the target PA.
4437  *      This function may be called from an interrupt.  No locking
4438  *      is required.
4439  */
4440 void
4441 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
4442 {
4443         vm_offset_t src_virt, dst_virt;
4444
4445         src_virt = PHYS_TO_DMAP(src);
4446         dst_virt = PHYS_TO_DMAP(dst);
4447         bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
4448 }
4449
4450 /*
4451  * pmap_copy_page_frag:
4452  *
4453  *      Copy the physical page from the source PA to the target PA.
4454  *      This function may be called from an interrupt.  No locking
4455  *      is required.
4456  */
4457 void
4458 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
4459 {
4460         vm_offset_t src_virt, dst_virt;
4461
4462         src_virt = PHYS_TO_DMAP(src);
4463         dst_virt = PHYS_TO_DMAP(dst);
4464
4465         bcopy((char *)src_virt + (src & PAGE_MASK),
4466               (char *)dst_virt + (dst & PAGE_MASK),
4467               bytes);
4468 }
4469
4470 /*
4471  * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4472  * this page.  This count may be changed upwards or downwards in the future;
4473  * it is only necessary that true be returned for a small subset of pmaps
4474  * for proper page aging.
4475  */
4476 boolean_t
4477 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
4478 {
4479         pv_entry_t pv;
4480         int loops = 0;
4481
4482         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4483                 return FALSE;
4484
4485         vm_page_spin_lock(m);
4486         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4487                 if (pv->pv_pmap == pmap) {
4488                         vm_page_spin_unlock(m);
4489                         return TRUE;
4490                 }
4491                 loops++;
4492                 if (loops >= 16)
4493                         break;
4494         }
4495         vm_page_spin_unlock(m);
4496         return (FALSE);
4497 }
4498
4499 /*
4500  * Remove all pages from specified address space this aids process exit
4501  * speeds.  Also, this code may be special cased for the current process
4502  * only.
4503  */
4504 void
4505 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4506 {
4507         pmap_remove_noinval(pmap, sva, eva);
4508         cpu_invltlb();
4509 }
4510
4511 /*
4512  * pmap_testbit tests bits in pte's note that the testbit/clearbit
4513  * routines are inline, and a lot of things compile-time evaluate.
4514  */
4515 static
4516 boolean_t
4517 pmap_testbit(vm_page_t m, int bit)
4518 {
4519         pv_entry_t pv;
4520         pt_entry_t *pte;
4521         pmap_t pmap;
4522
4523         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4524                 return FALSE;
4525
4526         if (TAILQ_FIRST(&m->md.pv_list) == NULL)
4527                 return FALSE;
4528         vm_page_spin_lock(m);
4529         if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
4530                 vm_page_spin_unlock(m);
4531                 return FALSE;
4532         }
4533
4534         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4535
4536 #if defined(PMAP_DIAGNOSTIC)
4537                 if (pv->pv_pmap == NULL) {
4538                         kprintf("Null pmap (tb) at pindex: %"PRIu64"\n",
4539                             pv->pv_pindex);
4540                         continue;
4541                 }
4542 #endif
4543                 pmap = pv->pv_pmap;
4544
4545                 /*
4546                  * If the bit being tested is the modified bit, then
4547                  * mark clean_map and ptes as never
4548                  * modified.
4549                  *
4550                  * WARNING!  Because we do not lock the pv, *pte can be in a
4551                  *           state of flux.  Despite this the value of *pte
4552                  *           will still be related to the vm_page in some way
4553                  *           because the pv cannot be destroyed as long as we
4554                  *           hold the vm_page spin lock.
4555                  */
4556                 if (bit == PG_A_IDX || bit == PG_M_IDX) {
4557                                 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4558                         if (!pmap_track_modified(pv->pv_pindex))
4559                                 continue;
4560                 }
4561
4562                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4563                 if (*pte & pmap->pmap_bits[bit]) {
4564                         vm_page_spin_unlock(m);
4565                         return TRUE;
4566                 }
4567         }
4568         vm_page_spin_unlock(m);
4569         return (FALSE);
4570 }
4571
4572 /*
4573  * This routine is used to modify bits in ptes.  Only one bit should be
4574  * specified.  PG_RW requires special handling.
4575  *
4576  * Caller must NOT hold any spin locks
4577  */
4578 static __inline
4579 void
4580 pmap_clearbit(vm_page_t m, int bit_index)
4581 {
4582         struct pmap_inval_info info;
4583         pv_entry_t pv;
4584         pt_entry_t *pte;
4585         pt_entry_t pbits;
4586         pmap_t pmap;
4587
4588         if (bit_index == PG_RW_IDX)
4589                 vm_page_flag_clear(m, PG_WRITEABLE);
4590         if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
4591                 return;
4592         }
4593
4594         /*
4595          * PG_M or PG_A case
4596          *
4597          * Loop over all current mappings setting/clearing as appropos If
4598          * setting RO do we need to clear the VAC?
4599          *
4600          * NOTE: When clearing PG_M we could also (not implemented) drop
4601          *       through to the PG_RW code and clear PG_RW too, forcing
4602          *       a fault on write to redetect PG_M for virtual kernels, but
4603          *       it isn't necessary since virtual kernels invalidate the
4604          *       pte when they clear the VPTE_M bit in their virtual page
4605          *       tables.
4606          *
4607          * NOTE: Does not re-dirty the page when clearing only PG_M.
4608          *
4609          * NOTE: Because we do not lock the pv, *pte can be in a state of
4610          *       flux.  Despite this the value of *pte is still somewhat
4611          *       related while we hold the vm_page spin lock.
4612          *
4613          *       *pte can be zero due to this race.  Since we are clearing
4614          *       bits we basically do no harm when this race  ccurs.
4615          */
4616         if (bit_index != PG_RW_IDX) {
4617                 vm_page_spin_lock(m);
4618                 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4619 #if defined(PMAP_DIAGNOSTIC)
4620                         if (pv->pv_pmap == NULL) {
4621                                 kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4622                                     pv->pv_pindex);
4623                                 continue;
4624                         }
4625 #endif
4626                         pmap = pv->pv_pmap;
4627                         pte = pmap_pte_quick(pv->pv_pmap,
4628                                              pv->pv_pindex << PAGE_SHIFT);
4629                         pbits = *pte;
4630                         if (pbits & pmap->pmap_bits[bit_index])
4631                                 atomic_clear_long(pte, pmap->pmap_bits[bit_index]);
4632                 }
4633                 vm_page_spin_unlock(m);
4634                 return;
4635         }
4636
4637         /*
4638          * Clear PG_RW.  Also clears PG_M and marks the page dirty if PG_M
4639          * was set.
4640          */
4641         pmap_inval_init(&info);
4642
4643 restart:
4644         vm_page_spin_lock(m);
4645         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4646                 /*
4647                  * don't write protect pager mappings
4648                  */
4649                 if (!pmap_track_modified(pv->pv_pindex))
4650                         continue;
4651
4652 #if defined(PMAP_DIAGNOSTIC)
4653                 if (pv->pv_pmap == NULL) {
4654                         kprintf("Null pmap (cb) at pindex: %"PRIu64"\n",
4655                             pv->pv_pindex);
4656                         continue;
4657                 }
4658 #endif
4659                 pmap = pv->pv_pmap;
4660                 /*
4661                  * Skip pages which do not have PG_RW set.
4662                  */
4663                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4664                 if ((*pte & pmap->pmap_bits[PG_RW_IDX]) == 0)
4665                         continue;
4666
4667                 /*
4668                  * Lock the PV
4669                  */
4670                 if (pv_hold_try(pv)) {
4671                         vm_page_spin_unlock(m);
4672                 } else {
4673                         vm_page_spin_unlock(m);
4674                         pv_lock(pv);    /* held, now do a blocking lock */
4675                 }
4676                 if (pv->pv_pmap != pmap || pv->pv_m != m) {
4677                         pv_put(pv);     /* and release */
4678                         goto restart;   /* anything could have happened */
4679                 }
4680                 pmap_inval_interlock(&info, pmap,
4681                                      (vm_offset_t)pv->pv_pindex << PAGE_SHIFT);
4682                 KKASSERT(pv->pv_pmap == pmap);
4683                 for (;;) {
4684                         pbits = *pte;
4685                         cpu_ccfence();
4686                         if (atomic_cmpset_long(pte, pbits, pbits &
4687                             ~(pmap->pmap_bits[PG_RW_IDX] |
4688                             pmap->pmap_bits[PG_M_IDX]))) {
4689                                 break;
4690                         }
4691                 }
4692                 pmap_inval_deinterlock(&info, pmap);
4693                 vm_page_spin_lock(m);
4694
4695                 /*
4696                  * If PG_M was found to be set while we were clearing PG_RW
4697                  * we also clear PG_M (done above) and mark the page dirty.
4698                  * Callers expect this behavior.
4699                  */
4700                 if (pbits & pmap->pmap_bits[PG_M_IDX])
4701                         vm_page_dirty(m);
4702                 pv_put(pv);
4703         }
4704         vm_page_spin_unlock(m);
4705         pmap_inval_done(&info);
4706 }
4707
4708 /*
4709  * Lower the permission for all mappings to a given page.
4710  *
4711  * Page must be busied by caller.  Because page is busied by caller this
4712  * should not be able to race a pmap_enter().
4713  */
4714 void
4715 pmap_page_protect(vm_page_t m, vm_prot_t prot)
4716 {
4717         /* JG NX support? */
4718         if ((prot & VM_PROT_WRITE) == 0) {
4719                 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
4720                         /*
4721                          * NOTE: pmap_clearbit(.. PG_RW) also clears
4722                          *       the PG_WRITEABLE flag in (m).
4723                          */
4724                         pmap_clearbit(m, PG_RW_IDX);
4725                 } else {
4726                         pmap_remove_all(m);
4727                 }
4728         }
4729 }
4730
4731 vm_paddr_t
4732 pmap_phys_address(vm_pindex_t ppn)
4733 {
4734         return (x86_64_ptob(ppn));
4735 }
4736
4737 /*
4738  * Return a count of reference bits for a page, clearing those bits.
4739  * It is not necessary for every reference bit to be cleared, but it
4740  * is necessary that 0 only be returned when there are truly no
4741  * reference bits set.
4742  *
4743  * XXX: The exact number of bits to check and clear is a matter that
4744  * should be tested and standardized at some point in the future for
4745  * optimal aging of shared pages.
4746  *
4747  * This routine may not block.
4748  */
4749 int
4750 pmap_ts_referenced(vm_page_t m)
4751 {
4752         pv_entry_t pv;
4753         pt_entry_t *pte;
4754         pmap_t pmap;
4755         int rtval = 0;
4756
4757         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
4758                 return (rtval);
4759
4760         vm_page_spin_lock(m);
4761         TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
4762                 if (!pmap_track_modified(pv->pv_pindex))
4763                         continue;
4764                 pmap = pv->pv_pmap;
4765                 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_pindex << PAGE_SHIFT);
4766                 if (pte && (*pte & pmap->pmap_bits[PG_A_IDX])) {
4767                         atomic_clear_long(pte, pmap->pmap_bits[PG_A_IDX]);
4768                         rtval++;
4769                         if (rtval > 4)
4770                                 break;
4771                 }
4772         }
4773         vm_page_spin_unlock(m);
4774         return (rtval);
4775 }
4776
4777 /*
4778  *      pmap_is_modified:
4779  *
4780  *      Return whether or not the specified physical page was modified
4781  *      in any physical maps.
4782  */
4783 boolean_t
4784 pmap_is_modified(vm_page_t m)
4785 {
4786         boolean_t res;
4787
4788         res = pmap_testbit(m, PG_M_IDX);
4789         return (res);
4790 }
4791
4792 /*
4793  *      Clear the modify bits on the specified physical page.
4794  */
4795 void
4796 pmap_clear_modify(vm_page_t m)
4797 {
4798         pmap_clearbit(m, PG_M_IDX);
4799 }
4800
4801 /*
4802  *      pmap_clear_reference:
4803  *
4804  *      Clear the reference bit on the specified physical page.
4805  */
4806 void
4807 pmap_clear_reference(vm_page_t m)
4808 {
4809         pmap_clearbit(m, PG_A_IDX);
4810 }
4811
4812 /*
4813  * Miscellaneous support routines follow
4814  */
4815
4816 static
4817 void
4818 i386_protection_init(void)
4819 {
4820         int *kp, prot;
4821
4822         /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit  */
4823         kp = protection_codes;
4824         for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
4825                 switch (prot) {
4826                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
4827                         /*
4828                          * Read access is also 0. There isn't any execute bit,
4829                          * so just make it readable.
4830                          */
4831                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
4832                 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
4833                 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
4834                         *kp++ = 0;
4835                         break;
4836                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
4837                 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
4838                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
4839                 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
4840                         *kp++ = pmap_bits_default[PG_RW_IDX];
4841                         break;
4842                 }
4843         }
4844 }
4845
4846 /*
4847  * Map a set of physical memory pages into the kernel virtual
4848  * address space. Return a pointer to where it is mapped. This
4849  * routine is intended to be used for mapping device memory,
4850  * NOT real memory.
4851  *
4852  * NOTE: We can't use pgeflag unless we invalidate the pages one at
4853  *       a time.
4854  *
4855  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
4856  *       work whether the cpu supports PAT or not.  The remaining PAT
4857  *       attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
4858  *       supports PAT.
4859  */
4860 void *
4861 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
4862 {
4863         return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4864 }
4865
4866 void *
4867 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
4868 {
4869         return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
4870 }
4871
4872 void *
4873 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
4874 {
4875         return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
4876 }
4877
4878 /*
4879  * Map a set of physical memory pages into the kernel virtual
4880  * address space. Return a pointer to where it is mapped. This
4881  * routine is intended to be used for mapping device memory,
4882  * NOT real memory.
4883  */
4884 void *
4885 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
4886 {
4887         vm_offset_t va, tmpva, offset;
4888         pt_entry_t *pte;
4889         vm_size_t tmpsize;
4890
4891         offset = pa & PAGE_MASK;
4892         size = roundup(offset + size, PAGE_SIZE);
4893
4894         va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
4895         if (va == 0)
4896                 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
4897
4898         pa = pa & ~PAGE_MASK;
4899         for (tmpva = va, tmpsize = size; tmpsize > 0;) {
4900                 pte = vtopte(tmpva);
4901                 *pte = pa |
4902                     kernel_pmap.pmap_bits[PG_RW_IDX] |
4903                     kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
4904                     kernel_pmap.pmap_cache_bits[mode];
4905                 tmpsize -= PAGE_SIZE;
4906                 tmpva += PAGE_SIZE;
4907                 pa += PAGE_SIZE;
4908         }
4909         pmap_invalidate_range(&kernel_pmap, va, va + size);
4910         pmap_invalidate_cache_range(va, va + size);
4911
4912         return ((void *)(va + offset));
4913 }
4914
4915 void
4916 pmap_unmapdev(vm_offset_t va, vm_size_t size)
4917 {
4918         vm_offset_t base, offset;
4919
4920         base = va & ~PAGE_MASK;
4921         offset = va & PAGE_MASK;
4922         size = roundup(offset + size, PAGE_SIZE);
4923         pmap_qremove(va, size >> PAGE_SHIFT);
4924         kmem_free(&kernel_map, base, size);
4925 }
4926
4927 /*
4928  * Sets the memory attribute for the specified page.
4929  */
4930 void
4931 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
4932 {
4933
4934     m->pat_mode = ma;
4935
4936     /*
4937      * If "m" is a normal page, update its direct mapping.  This update
4938      * can be relied upon to perform any cache operations that are
4939      * required for data coherence.
4940      */
4941     if ((m->flags & PG_FICTITIOUS) == 0)
4942         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE,
4943         m->pat_mode);
4944 }
4945
4946 /*
4947  * Change the PAT attribute on an existing kernel memory map.  Caller
4948  * must ensure that the virtual memory in question is not accessed
4949  * during the adjustment.
4950  */
4951 void
4952 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
4953 {
4954         pt_entry_t *pte;
4955         vm_offset_t base;
4956         int changed = 0;
4957
4958         if (va == 0)
4959                 panic("pmap_change_attr: va is NULL");
4960         base = trunc_page(va);
4961
4962         while (count) {
4963                 pte = vtopte(va);
4964                 *pte = (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask)) |
4965                        kernel_pmap.pmap_cache_bits[mode];
4966                 --count;
4967                 va += PAGE_SIZE;
4968         }
4969
4970         changed = 1;    /* XXX: not optimal */
4971
4972         /*
4973          * Flush CPU caches if required to make sure any data isn't cached that
4974          * shouldn't be, etc.
4975          */
4976         if (changed) {
4977                 pmap_invalidate_range(&kernel_pmap, base, va);
4978                 pmap_invalidate_cache_range(base, va);
4979         }
4980 }
4981
4982 /*
4983  * perform the pmap work for mincore
4984  */
4985 int
4986 pmap_mincore(pmap_t pmap, vm_offset_t addr)
4987 {
4988         pt_entry_t *ptep, pte;
4989         vm_page_t m;
4990         int val = 0;
4991         
4992         lwkt_gettoken(&pmap->pm_token);
4993         ptep = pmap_pte(pmap, addr);
4994
4995         if (ptep && (pte = *ptep) != 0) {
4996                 vm_offset_t pa;
4997
4998                 val = MINCORE_INCORE;
4999                 if ((pte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5000                         goto done;
5001
5002                 pa = pte & PG_FRAME;
5003
5004                 if (pte & pmap->pmap_bits[PG_DEVICE_IDX])
5005                         m = NULL;
5006                 else
5007                         m = PHYS_TO_VM_PAGE(pa);
5008
5009                 /*
5010                  * Modified by us
5011                  */
5012                 if (pte & pmap->pmap_bits[PG_M_IDX])
5013                         val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5014                 /*
5015                  * Modified by someone
5016                  */
5017                 else if (m && (m->dirty || pmap_is_modified(m)))
5018                         val |= MINCORE_MODIFIED_OTHER;
5019                 /*
5020                  * Referenced by us
5021                  */
5022                 if (pte & pmap->pmap_bits[PG_A_IDX])
5023                         val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5024
5025                 /*
5026                  * Referenced by someone
5027                  */
5028                 else if (m && ((m->flags & PG_REFERENCED) ||
5029                                 pmap_ts_referenced(m))) {
5030                         val |= MINCORE_REFERENCED_OTHER;
5031                         vm_page_flag_set(m, PG_REFERENCED);
5032                 }
5033         } 
5034 done:
5035         lwkt_reltoken(&pmap->pm_token);
5036
5037         return val;
5038 }
5039
5040 /*
5041  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5042  * vmspace will be ref'd and the old one will be deref'd.
5043  *
5044  * The vmspace for all lwps associated with the process will be adjusted
5045  * and cr3 will be reloaded if any lwp is the current lwp.
5046  *
5047  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5048  */
5049 void
5050 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5051 {
5052         struct vmspace *oldvm;
5053         struct lwp *lp;
5054
5055         oldvm = p->p_vmspace;
5056         if (oldvm != newvm) {
5057                 if (adjrefs)
5058                         sysref_get(&newvm->vm_sysref);
5059                 p->p_vmspace = newvm;
5060                 KKASSERT(p->p_nthreads == 1);
5061                 lp = RB_ROOT(&p->p_lwp_tree);
5062                 pmap_setlwpvm(lp, newvm);
5063                 if (adjrefs)
5064                         sysref_put(&oldvm->vm_sysref);
5065         }
5066 }
5067
5068 /*
5069  * Set the vmspace for a LWP.  The vmspace is almost universally set the
5070  * same as the process vmspace, but virtual kernels need to swap out contexts
5071  * on a per-lwp basis.
5072  *
5073  * Caller does not necessarily hold any vmspace tokens.  Caller must control
5074  * the lwp (typically be in the context of the lwp).  We use a critical
5075  * section to protect against statclock and hardclock (statistics collection).
5076  */
5077 void
5078 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
5079 {
5080         struct vmspace *oldvm;
5081         struct pmap *pmap;
5082
5083         oldvm = lp->lwp_vmspace;
5084
5085         if (oldvm != newvm) {
5086                 crit_enter();
5087                 lp->lwp_vmspace = newvm;
5088                 if (curthread->td_lwp == lp) {
5089                         pmap = vmspace_pmap(newvm);
5090                         atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
5091                         if (pmap->pm_active & CPUMASK_LOCK)
5092                                 pmap_interlock_wait(newvm);
5093 #if defined(SWTCH_OPTIM_STATS)
5094                         tlb_flush_count++;
5095 #endif
5096                         if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
5097                                 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
5098                         } else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
5099                                 curthread->td_pcb->pcb_cr3 = KPML4phys;
5100                         } else {
5101                                 panic("pmap_setlwpvm: unknown pmap type\n");
5102                         }
5103                         load_cr3(curthread->td_pcb->pcb_cr3);
5104                         pmap = vmspace_pmap(oldvm);
5105                         atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
5106                 }
5107                 crit_exit();
5108         }
5109 }
5110
5111 /*
5112  * Called when switching to a locked pmap, used to interlock against pmaps
5113  * undergoing modifications to prevent us from activating the MMU for the
5114  * target pmap until all such modifications have completed.  We have to do
5115  * this because the thread making the modifications has already set up its
5116  * SMP synchronization mask.
5117  *
5118  * This function cannot sleep!
5119  *
5120  * No requirements.
5121  */
5122 void
5123 pmap_interlock_wait(struct vmspace *vm)
5124 {
5125         struct pmap *pmap = &vm->vm_pmap;
5126
5127         if (pmap->pm_active & CPUMASK_LOCK) {
5128                 crit_enter();
5129                 KKASSERT(curthread->td_critcount >= 2);
5130                 DEBUG_PUSH_INFO("pmap_interlock_wait");
5131                 while (pmap->pm_active & CPUMASK_LOCK) {
5132                         cpu_ccfence();
5133                         lwkt_process_ipiq();
5134                 }
5135                 DEBUG_POP_INFO();
5136                 crit_exit();
5137         }
5138 }
5139
5140 vm_offset_t
5141 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
5142 {
5143
5144         if ((obj == NULL) || (size < NBPDR) ||
5145             ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
5146                 return addr;
5147         }
5148
5149         addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
5150         return addr;
5151 }
5152
5153 /*
5154  * Used by kmalloc/kfree, page already exists at va
5155  */
5156 vm_page_t
5157 pmap_kvtom(vm_offset_t va)
5158 {
5159         pt_entry_t *ptep = vtopte(va);
5160
5161         KKASSERT((*ptep & kernel_pmap.pmap_bits[PG_DEVICE_IDX]) == 0);
5162         return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
5163 }
5164
5165 /*
5166  * Initialize machine-specific shared page directory support.  This
5167  * is executed when a VM object is created.
5168  */
5169 void
5170 pmap_object_init(vm_object_t object)
5171 {
5172         object->md.pmap_rw = NULL;
5173         object->md.pmap_ro = NULL;
5174 }
5175
5176 /*
5177  * Clean up machine-specific shared page directory support.  This
5178  * is executed when a VM object is destroyed.
5179  */
5180 void
5181 pmap_object_free(vm_object_t object)
5182 {
5183         pmap_t pmap;
5184
5185         if ((pmap = object->md.pmap_rw) != NULL) {
5186                 object->md.pmap_rw = NULL;
5187                 pmap_remove_noinval(pmap,
5188                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5189                 pmap->pm_active = 0;
5190                 pmap_release(pmap);
5191                 pmap_puninit(pmap);
5192                 kfree(pmap, M_OBJPMAP);
5193         }
5194         if ((pmap = object->md.pmap_ro) != NULL) {
5195                 object->md.pmap_ro = NULL;
5196                 pmap_remove_noinval(pmap,
5197                                   VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
5198                 pmap->pm_active = 0;
5199                 pmap_release(pmap);
5200                 pmap_puninit(pmap);
5201                 kfree(pmap, M_OBJPMAP);
5202         }
5203 }