xref: /freebsd/sys/i386/i386/pmap.c (revision c8a523690b1a904344662c5834b713e5fc9e9aad)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1994 John S. Dyson
7  * All rights reserved.
8  * Copyright (c) 1994 David Greenman
9  * All rights reserved.
10  * Copyright (c) 2005-2010 Alan L. Cox <alc@cs.rice.edu>
11  * All rights reserved.
12  *
13  * This code is derived from software contributed to Berkeley by
14  * the Systems Programming Group of the University of Utah Computer
15  * Science Department and William Jolitz of UUNET Technologies Inc.
16  *
17  * Redistribution and use in source and binary forms, with or without
18  * modification, are permitted provided that the following conditions
19  * are met:
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  * 3. All advertising materials mentioning features or use of this software
26  *    must display the following acknowledgement:
27  *	This product includes software developed by the University of
28  *	California, Berkeley and its contributors.
29  * 4. Neither the name of the University nor the names of its contributors
30  *    may be used to endorse or promote products derived from this software
31  *    without specific prior written permission.
32  *
33  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
34  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
35  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
36  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
37  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
38  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
39  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
41  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
42  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43  * SUCH DAMAGE.
44  */
45 /*-
46  * Copyright (c) 2003 Networks Associates Technology, Inc.
47  * All rights reserved.
48  * Copyright (c) 2018 The FreeBSD Foundation
49  * All rights reserved.
50  *
51  * This software was developed for the FreeBSD Project by Jake Burkholder,
52  * Safeport Network Services, and Network Associates Laboratories, the
53  * Security Research Division of Network Associates, Inc. under
54  * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA
55  * CHATS research program.
56  *
57  * Portions of this software were developed by
58  * Konstantin Belousov <kib@FreeBSD.org> under sponsorship from
59  * the FreeBSD Foundation.
60  *
61  * Redistribution and use in source and binary forms, with or without
62  * modification, are permitted provided that the following conditions
63  * are met:
64  * 1. Redistributions of source code must retain the above copyright
65  *    notice, this list of conditions and the following disclaimer.
66  * 2. Redistributions in binary form must reproduce the above copyright
67  *    notice, this list of conditions and the following disclaimer in the
68  *    documentation and/or other materials provided with the distribution.
69  *
70  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
71  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
72  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
73  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
74  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
75  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
76  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
77  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
78  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
79  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
80  * SUCH DAMAGE.
81  */
82 
83 #include <sys/cdefs.h>
84 /*
85  *	Manages physical address maps.
86  *
87  *	Since the information managed by this module is
88  *	also stored by the logical address mapping module,
89  *	this module may throw away valid virtual-to-physical
90  *	mappings at almost any time.  However, invalidations
91  *	of virtual-to-physical mappings must be done as
92  *	requested.
93  *
94  *	In order to cope with hardware architectures which
95  *	make virtual-to-physical map invalidates expensive,
96  *	this module may delay invalidate or reduced protection
97  *	operations until such time as they are actually
98  *	necessary.  This module is given full information as
99  *	to which processors are currently using which maps,
100  *	and to when physical maps must be made correct.
101  */
102 
103 #include "opt_apic.h"
104 #include "opt_cpu.h"
105 #include "opt_pmap.h"
106 #include "opt_smp.h"
107 #include "opt_vm.h"
108 
109 #include <sys/param.h>
110 #include <sys/systm.h>
111 #include <sys/kernel.h>
112 #include <sys/ktr.h>
113 #include <sys/lock.h>
114 #include <sys/malloc.h>
115 #include <sys/mman.h>
116 #include <sys/msgbuf.h>
117 #include <sys/mutex.h>
118 #include <sys/proc.h>
119 #include <sys/rwlock.h>
120 #include <sys/sbuf.h>
121 #include <sys/sf_buf.h>
122 #include <sys/sx.h>
123 #include <sys/vmmeter.h>
124 #include <sys/sched.h>
125 #include <sys/sysctl.h>
126 #include <sys/smp.h>
127 #include <sys/vmem.h>
128 
129 #include <vm/vm.h>
130 #include <vm/vm_param.h>
131 #include <vm/vm_kern.h>
132 #include <vm/vm_page.h>
133 #include <vm/vm_map.h>
134 #include <vm/vm_object.h>
135 #include <vm/vm_extern.h>
136 #include <vm/vm_pageout.h>
137 #include <vm/vm_pager.h>
138 #include <vm/vm_phys.h>
139 #include <vm/vm_radix.h>
140 #include <vm/vm_reserv.h>
141 #include <vm/uma.h>
142 
143 #ifdef DEV_APIC
144 #include <sys/bus.h>
145 #include <machine/intr_machdep.h>
146 #include <x86/apicvar.h>
147 #endif
148 #include <x86/ifunc.h>
149 #include <machine/bootinfo.h>
150 #include <machine/cpu.h>
151 #include <machine/cputypes.h>
152 #include <machine/md_var.h>
153 #include <machine/pcb.h>
154 #include <machine/specialreg.h>
155 #ifdef SMP
156 #include <machine/smp.h>
157 #endif
158 #include <machine/pmap_base.h>
159 
160 #ifdef PV_STATS
161 #define PV_STAT(x)	do { x ; } while (0)
162 #else
163 #define PV_STAT(x)	do { } while (0)
164 #endif
165 
166 #define	pa_index(pa)	((pa) >> PDRSHIFT)
167 #define	pa_to_pvh(pa)	(&pv_table[pa_index(pa)])
168 
169 /*
170  * PTmap is recursive pagemap at top of virtual address space.
171  * Within PTmap, the page directory can be found (third indirection).
172  */
173 #define	PTmap	((pt_entry_t *)(PTDPTDI << PDRSHIFT))
174 #define	PTD	((pd_entry_t *)((PTDPTDI << PDRSHIFT) + (PTDPTDI * PAGE_SIZE)))
175 #define	PTDpde	((pd_entry_t *)((PTDPTDI << PDRSHIFT) + (PTDPTDI * PAGE_SIZE) + \
176     (PTDPTDI * PDESIZE)))
177 
178 /*
179  * Translate a virtual address to the kernel virtual address of its page table
180  * entry (PTE).  This can be used recursively.  If the address of a PTE as
181  * previously returned by this macro is itself given as the argument, then the
182  * address of the page directory entry (PDE) that maps the PTE will be
183  * returned.
184  *
185  * This macro may be used before pmap_bootstrap() is called.
186  */
187 #define	vtopte(va)	(PTmap + i386_btop(va))
188 
189 /*
190  * Get PDEs and PTEs for user/kernel address space
191  */
192 #define	pmap_pde(m, v)	(&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
193 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
194 
195 #define pmap_pde_v(pte)		((*(int *)pte & PG_V) != 0)
196 #define pmap_pte_w(pte)		((*(int *)pte & PG_W) != 0)
197 #define pmap_pte_m(pte)		((*(int *)pte & PG_M) != 0)
198 #define pmap_pte_u(pte)		((*(int *)pte & PG_A) != 0)
199 #define pmap_pte_v(pte)		((*(int *)pte & PG_V) != 0)
200 
201 #define pmap_pte_set_w(pte, v)	((v) ? atomic_set_int((u_int *)(pte), PG_W) : \
202     atomic_clear_int((u_int *)(pte), PG_W))
203 #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v)))
204 
205 static int pgeflag = 0;		/* PG_G or-in */
206 static int pseflag = 0;		/* PG_PS or-in */
207 
208 static int nkpt = NKPT;
209 
210 #ifdef PMAP_PAE_COMP
211 pt_entry_t pg_nx;
212 static uma_zone_t pdptzone;
213 #else
214 #define	pg_nx	0
215 #endif
216 
217 _Static_assert(VM_MAXUSER_ADDRESS == VADDR(TRPTDI, 0), "VM_MAXUSER_ADDRESS");
218 _Static_assert(VM_MAX_KERNEL_ADDRESS <= VADDR(PTDPTDI, 0),
219     "VM_MAX_KERNEL_ADDRESS");
220 _Static_assert(PMAP_MAP_LOW == VADDR(LOWPTDI, 0), "PMAP_MAP_LOW");
221 _Static_assert(KERNLOAD == (KERNPTDI << PDRSHIFT), "KERNLOAD");
222 
223 extern int pat_works;
224 extern int pg_ps_enabled;
225 
226 extern int elf32_nxstack;
227 
228 #define	PAT_INDEX_SIZE	8
229 static int pat_index[PAT_INDEX_SIZE];	/* cache mode to PAT index conversion */
230 
231 /*
232  * pmap_mapdev support pre initialization (i.e. console)
233  */
234 #define	PMAP_PREINIT_MAPPING_COUNT	8
235 static struct pmap_preinit_mapping {
236 	vm_paddr_t	pa;
237 	vm_offset_t	va;
238 	vm_size_t	sz;
239 	int		mode;
240 } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT];
241 static int pmap_initialized;
242 
243 static struct rwlock_padalign pvh_global_lock;
244 
245 /*
246  * Data for the pv entry allocation mechanism
247  */
248 static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks);
249 extern int pv_entry_max, pv_entry_count;
250 static int pv_entry_high_water = 0;
251 static struct md_page *pv_table;
252 extern int shpgperproc;
253 
254 static struct pv_chunk *pv_chunkbase;	/* KVA block for pv_chunks */
255 static int pv_maxchunks;		/* How many chunks we have KVA for */
256 static vm_offset_t pv_vafree;		/* freelist stored in the PTE */
257 
258 /*
259  * All those kernel PT submaps that BSD is so fond of
260  */
261 static pt_entry_t *CMAP3;
262 static pd_entry_t *KPTD;
263 static caddr_t CADDR3;
264 
265 /*
266  * Crashdump maps.
267  */
268 static caddr_t crashdumpmap;
269 
270 static pt_entry_t *PMAP1 = NULL, *PMAP2, *PMAP3;
271 static pt_entry_t *PADDR1 = NULL, *PADDR2, *PADDR3;
272 #ifdef SMP
273 static int PMAP1cpu, PMAP3cpu;
274 extern int PMAP1changedcpu;
275 #endif
276 extern int PMAP1changed;
277 extern int PMAP1unchanged;
278 static struct mtx PMAP2mutex;
279 
280 /*
281  * Internal flags for pmap_enter()'s helper functions.
282  */
283 #define	PMAP_ENTER_NORECLAIM	0x1000000	/* Don't reclaim PV entries. */
284 #define	PMAP_ENTER_NOREPLACE	0x2000000	/* Don't replace mappings. */
285 
286 static void	free_pv_chunk(struct pv_chunk *pc);
287 static void	free_pv_entry(pmap_t pmap, pv_entry_t pv);
288 static pv_entry_t get_pv_entry(pmap_t pmap, bool try);
289 static void	pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa);
290 static bool	pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde,
291 		    u_int flags);
292 #if VM_NRESERVLEVEL > 0
293 static void	pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa);
294 #endif
295 static void	pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va);
296 static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap,
297 		    vm_offset_t va);
298 static int	pmap_pvh_wired_mappings(struct md_page *pvh, int count);
299 
300 static void	pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte);
301 static bool	pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va);
302 static int	pmap_enter_4mpage(pmap_t pmap, vm_offset_t va, vm_page_t m,
303 		    vm_prot_t prot);
304 static int	pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde,
305 		    u_int flags, vm_page_t m);
306 static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va,
307     vm_page_t m, vm_prot_t prot, vm_page_t mpte);
308 static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted,
309     bool allpte_PG_A_set);
310 static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va,
311 		    pd_entry_t pde);
312 static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte);
313 static bool pmap_is_modified_pvh(struct md_page *pvh);
314 static bool pmap_is_referenced_pvh(struct md_page *pvh);
315 static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode);
316 static void pmap_kenter_pde(vm_offset_t va, pd_entry_t newpde);
317 static void pmap_pde_attr(pd_entry_t *pde, int cache_bits);
318 #if VM_NRESERVLEVEL > 0
319 static bool pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va,
320     vm_page_t mpte);
321 #endif
322 static bool pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva,
323     vm_prot_t prot);
324 static void pmap_pte_attr(pt_entry_t *pte, int cache_bits);
325 static void pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
326     struct spglist *free);
327 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva,
328     struct spglist *free);
329 static vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va);
330 static void pmap_remove_page(pmap_t pmap, vm_offset_t va, struct spglist *free);
331 static bool	pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
332 		    struct spglist *free);
333 static void pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va);
334 static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m);
335 static bool pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va,
336     vm_page_t m);
337 static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde,
338     pd_entry_t newpde);
339 static void pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde);
340 
341 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, u_int flags);
342 
343 static vm_page_t _pmap_allocpte(pmap_t pmap, u_int ptepindex, u_int flags);
344 static void _pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free);
345 static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va);
346 static void pmap_pte_release(pt_entry_t *pte);
347 static int pmap_unuse_pt(pmap_t, vm_offset_t, struct spglist *);
348 #ifdef PMAP_PAE_COMP
349 static void *pmap_pdpt_allocf(uma_zone_t zone, vm_size_t bytes, int domain,
350     uint8_t *flags, int wait);
351 #endif
352 static void pmap_init_trm(void);
353 static void pmap_invalidate_all_int(pmap_t pmap);
354 
355 static __inline void pagezero(void *page);
356 
357 CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t));
358 CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t));
359 
360 extern char _end[];
361 extern u_long physfree;	/* phys addr of next free page */
362 extern u_long vm86phystk;/* PA of vm86/bios stack */
363 extern u_long vm86paddr;/* address of vm86 region */
364 extern int vm86pa;	/* phys addr of vm86 region */
365 extern u_long KERNend;	/* phys addr end of kernel (just after bss) */
366 #ifdef PMAP_PAE_COMP
367 pd_entry_t *IdlePTD_pae;	/* phys addr of kernel PTD */
368 pdpt_entry_t *IdlePDPT;	/* phys addr of kernel PDPT */
369 pt_entry_t *KPTmap_pae;	/* address of kernel page tables */
370 #define	IdlePTD	IdlePTD_pae
371 #define	KPTmap	KPTmap_pae
372 #else
373 pd_entry_t *IdlePTD_nopae;
374 pt_entry_t *KPTmap_nopae;
375 #define	IdlePTD	IdlePTD_nopae
376 #define	KPTmap	KPTmap_nopae
377 #endif
378 extern u_long KPTphys;	/* phys addr of kernel page tables */
379 extern u_long tramp_idleptd;
380 
381 static u_long
allocpages(u_int cnt,u_long * physfree)382 allocpages(u_int cnt, u_long *physfree)
383 {
384 	u_long res;
385 
386 	res = *physfree;
387 	*physfree += PAGE_SIZE * cnt;
388 	bzero((void *)res, PAGE_SIZE * cnt);
389 	return (res);
390 }
391 
392 static void
pmap_cold_map(u_long pa,u_long va,u_long cnt)393 pmap_cold_map(u_long pa, u_long va, u_long cnt)
394 {
395 	pt_entry_t *pt;
396 
397 	for (pt = (pt_entry_t *)KPTphys + atop(va); cnt > 0;
398 	    cnt--, pt++, va += PAGE_SIZE, pa += PAGE_SIZE)
399 		*pt = pa | PG_V | PG_RW | PG_A | PG_M;
400 }
401 
402 static void
pmap_cold_mapident(u_long pa,u_long cnt)403 pmap_cold_mapident(u_long pa, u_long cnt)
404 {
405 
406 	pmap_cold_map(pa, pa, cnt);
407 }
408 
409 _Static_assert(LOWPTDI * 2 * NBPDR == KERNBASE,
410     "Broken double-map of zero PTD");
411 
412 static void
__CONCAT(PMTYPE,remap_lower)413 __CONCAT(PMTYPE, remap_lower)(bool enable)
414 {
415 	int i;
416 
417 	for (i = 0; i < LOWPTDI; i++)
418 		IdlePTD[i] = enable ? IdlePTD[LOWPTDI + i] : 0;
419 	load_cr3(rcr3());		/* invalidate TLB */
420 }
421 
422 /*
423  * Called from locore.s before paging is enabled.  Sets up the first
424  * kernel page table.  Since kernel is mapped with PA == VA, this code
425  * does not require relocations.
426  */
427 void
__CONCAT(PMTYPE,cold)428 __CONCAT(PMTYPE, cold)(void)
429 {
430 	pt_entry_t *pt;
431 	u_long a;
432 	u_int cr3, ncr4;
433 
434 	physfree = (u_long)&_end;
435 	if (bootinfo.bi_esymtab != 0)
436 		physfree = bootinfo.bi_esymtab;
437 	if (bootinfo.bi_kernend != 0)
438 		physfree = bootinfo.bi_kernend;
439 	physfree = roundup2(physfree, NBPDR);
440 	KERNend = physfree;
441 
442 	/* Allocate Kernel Page Tables */
443 	KPTphys = allocpages(NKPT, &physfree);
444 	KPTmap = (pt_entry_t *)KPTphys;
445 
446 	/* Allocate Page Table Directory */
447 #ifdef PMAP_PAE_COMP
448 	/* XXX only need 32 bytes (easier for now) */
449 	IdlePDPT = (pdpt_entry_t *)allocpages(1, &physfree);
450 #endif
451 	IdlePTD = (pd_entry_t *)allocpages(NPGPTD, &physfree);
452 
453 	/*
454 	 * Allocate KSTACK.  Leave a guard page between IdlePTD and
455 	 * proc0kstack, to control stack overflow for thread0 and
456 	 * prevent corruption of the page table.  We leak the guard
457 	 * physical memory due to 1:1 mappings.
458 	 */
459 	allocpages(1, &physfree);
460 	proc0kstack = allocpages(TD0_KSTACK_PAGES, &physfree);
461 
462 	/* vm86/bios stack */
463 	vm86phystk = allocpages(1, &physfree);
464 
465 	/* pgtable + ext + IOPAGES */
466 	vm86paddr = vm86pa = allocpages(3, &physfree);
467 
468 	/* Install page tables into PTD.  Page table page 1 is wasted. */
469 	for (a = 0; a < NKPT; a++)
470 		IdlePTD[a] = (KPTphys + ptoa(a)) | PG_V | PG_RW | PG_A | PG_M;
471 
472 #ifdef PMAP_PAE_COMP
473 	/* PAE install PTD pointers into PDPT */
474 	for (a = 0; a < NPGPTD; a++)
475 		IdlePDPT[a] = ((u_int)IdlePTD + ptoa(a)) | PG_V;
476 #endif
477 
478 	/*
479 	 * Install recursive mapping for kernel page tables into
480 	 * itself.
481 	 */
482 	for (a = 0; a < NPGPTD; a++)
483 		IdlePTD[PTDPTDI + a] = ((u_int)IdlePTD + ptoa(a)) | PG_V |
484 		    PG_RW;
485 
486 	/*
487 	 * Initialize page table pages mapping physical address zero
488 	 * through the (physical) end of the kernel.  Many of these
489 	 * pages must be reserved, and we reserve them all and map
490 	 * them linearly for convenience.  We do this even if we've
491 	 * enabled PSE above; we'll just switch the corresponding
492 	 * kernel PDEs before we turn on paging.
493 	 *
494 	 * This and all other page table entries allow read and write
495 	 * access for various reasons.  Kernel mappings never have any
496 	 * access restrictions.
497 	 */
498 	pmap_cold_mapident(0, atop(NBPDR) * LOWPTDI);
499 	pmap_cold_map(0, NBPDR * LOWPTDI, atop(NBPDR) * LOWPTDI);
500 	pmap_cold_mapident(KERNBASE, atop(KERNend - KERNBASE));
501 
502 	/* Map page table directory */
503 #ifdef PMAP_PAE_COMP
504 	pmap_cold_mapident((u_long)IdlePDPT, 1);
505 #endif
506 	pmap_cold_mapident((u_long)IdlePTD, NPGPTD);
507 
508 	/* Map early KPTmap.  It is really pmap_cold_mapident. */
509 	pmap_cold_map(KPTphys, (u_long)KPTmap, NKPT);
510 
511 	/* Map proc0kstack */
512 	pmap_cold_mapident(proc0kstack, TD0_KSTACK_PAGES);
513 	/* ISA hole already mapped */
514 
515 	pmap_cold_mapident(vm86phystk, 1);
516 	pmap_cold_mapident(vm86pa, 3);
517 
518 	/* Map page 0 into the vm86 page table */
519 	*(pt_entry_t *)vm86pa = 0 | PG_RW | PG_U | PG_A | PG_M | PG_V;
520 
521 	/* ...likewise for the ISA hole for vm86 */
522 	for (pt = (pt_entry_t *)vm86pa + atop(ISA_HOLE_START), a = 0;
523 	    a < atop(ISA_HOLE_LENGTH); a++, pt++)
524 		*pt = (ISA_HOLE_START + ptoa(a)) | PG_RW | PG_U | PG_A |
525 		    PG_M | PG_V;
526 
527 	/* Enable PSE, PGE, VME, and PAE if configured. */
528 	ncr4 = 0;
529 	if ((cpu_feature & CPUID_PSE) != 0) {
530 		ncr4 |= CR4_PSE;
531 		pseflag = PG_PS;
532 		/*
533 		 * Superpage mapping of the kernel text.  Existing 4k
534 		 * page table pages are wasted.
535 		 */
536 		for (a = KERNBASE; a < KERNend; a += NBPDR)
537 			IdlePTD[a >> PDRSHIFT] = a | PG_PS | PG_A | PG_M |
538 			    PG_RW | PG_V;
539 	}
540 	if ((cpu_feature & CPUID_PGE) != 0) {
541 		ncr4 |= CR4_PGE;
542 		pgeflag = PG_G;
543 	}
544 	ncr4 |= (cpu_feature & CPUID_VME) != 0 ? CR4_VME : 0;
545 #ifdef PMAP_PAE_COMP
546 	ncr4 |= CR4_PAE;
547 #endif
548 	if (ncr4 != 0)
549 		load_cr4(rcr4() | ncr4);
550 
551 	/* Now enable paging */
552 #ifdef PMAP_PAE_COMP
553 	cr3 = (u_int)IdlePDPT;
554 	if ((cpu_feature & CPUID_PAT) == 0)
555 		wbinvd();
556 #else
557 	cr3 = (u_int)IdlePTD;
558 #endif
559 	tramp_idleptd = cr3;
560 	load_cr3(cr3);
561 	load_cr0(rcr0() | CR0_PG);
562 
563 	/*
564 	 * Now running relocated at KERNBASE where the system is
565 	 * linked to run.
566 	 */
567 
568 	/*
569 	 * Remove the lowest part of the double mapping of low memory
570 	 * to get some null pointer checks.
571 	 */
572 	__CONCAT(PMTYPE, remap_lower)(false);
573 
574 	kernel_vm_end = /* 0 + */ NKPT * NBPDR;
575 #ifdef PMAP_PAE_COMP
576 	i386_pmap_VM_NFREEORDER = VM_NFREEORDER_PAE;
577 	i386_pmap_VM_LEVEL_0_ORDER = VM_LEVEL_0_ORDER_PAE;
578 	i386_pmap_PDRSHIFT = PDRSHIFT_PAE;
579 #else
580 	i386_pmap_VM_NFREEORDER = VM_NFREEORDER_NOPAE;
581 	i386_pmap_VM_LEVEL_0_ORDER = VM_LEVEL_0_ORDER_NOPAE;
582 	i386_pmap_PDRSHIFT = PDRSHIFT_NOPAE;
583 #endif
584 }
585 
586 static void
__CONCAT(PMTYPE,set_nx)587 __CONCAT(PMTYPE, set_nx)(void)
588 {
589 
590 #ifdef PMAP_PAE_COMP
591 	if ((amd_feature & AMDID_NX) == 0)
592 		return;
593 	pg_nx = PG_NX;
594 	elf32_nxstack = 1;
595 	/* EFER.EFER_NXE is set in initializecpu(). */
596 #endif
597 }
598 
599 /*
600  *	Bootstrap the system enough to run with virtual memory.
601  *
602  *	On the i386 this is called after pmap_cold() created initial
603  *	kernel page table and enabled paging, and just syncs the pmap
604  *	module with what has already been done.
605  */
606 static void
__CONCAT(PMTYPE,bootstrap)607 __CONCAT(PMTYPE, bootstrap)(vm_paddr_t firstaddr)
608 {
609 	vm_offset_t va;
610 	pt_entry_t *pte, *unused __unused;
611 	struct pcpu *pc;
612 	u_long res;
613 	int i;
614 
615 	res = atop(firstaddr - (vm_paddr_t)KERNLOAD);
616 
617 	/*
618 	 * Initialize the first available kernel virtual address.
619 	 * However, using "firstaddr" may waste a few pages of the
620 	 * kernel virtual address space, because pmap_cold() may not
621 	 * have mapped every physical page that it allocated.
622 	 * Preferably, pmap_cold() would provide a first unused
623 	 * virtual address in addition to "firstaddr".
624 	 */
625 	virtual_avail = (vm_offset_t)firstaddr;
626 	virtual_end = VM_MAX_KERNEL_ADDRESS;
627 
628 	/*
629 	 * Initialize the kernel pmap (which is statically allocated).
630 	 * Count bootstrap data as being resident in case any of this data is
631 	 * later unmapped (using pmap_remove()) and freed.
632 	 */
633 	PMAP_LOCK_INIT(kernel_pmap);
634 	kernel_pmap->pm_pdir = IdlePTD;
635 #ifdef PMAP_PAE_COMP
636 	kernel_pmap->pm_pdpt = IdlePDPT;
637 #endif
638 	CPU_FILL(&kernel_pmap->pm_active);	/* don't allow deactivation */
639 	kernel_pmap->pm_stats.resident_count = res;
640 	TAILQ_INIT(&kernel_pmap->pm_pvchunk);
641 	vm_radix_init(&kernel_pmap->pm_root);
642 
643  	/*
644 	 * Initialize the global pv list lock.
645 	 */
646 	rw_init(&pvh_global_lock, "pmap pv global");
647 
648 	/*
649 	 * Reserve some special page table entries/VA space for temporary
650 	 * mapping of pages.
651 	 */
652 #define	SYSMAP(c, p, v, n)	\
653 	v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
654 
655 	va = virtual_avail;
656 	pte = vtopte(va);
657 
658 	/*
659 	 * Initialize temporary map objects on the current CPU for use
660 	 * during early boot.
661 	 * CMAP1/CMAP2 are used for zeroing and copying pages.
662 	 * CMAP3 is used for the boot-time memory test.
663 	 */
664 	pc = get_pcpu();
665 	mtx_init(&pc->pc_cmap_lock, "SYSMAPS", NULL, MTX_DEF);
666 	SYSMAP(caddr_t, pc->pc_cmap_pte1, pc->pc_cmap_addr1, 1)
667 	SYSMAP(caddr_t, pc->pc_cmap_pte2, pc->pc_cmap_addr2, 1)
668 	SYSMAP(vm_offset_t, pte, pc->pc_qmap_addr, 1)
669 
670 	SYSMAP(caddr_t, CMAP3, CADDR3, 1);
671 
672 	/*
673 	 * Crashdump maps.
674 	 */
675 	SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS)
676 
677 	/*
678 	 * ptvmmap is used for reading arbitrary physical pages via /dev/mem.
679 	 */
680 	SYSMAP(caddr_t, unused, ptvmmap, 1)
681 
682 	/*
683 	 * msgbufp is used to map the system message buffer.
684 	 */
685 	SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(msgbufsize)))
686 
687 	/*
688 	 * KPTmap is used by pmap_kextract().
689 	 *
690 	 * KPTmap is first initialized by pmap_cold().  However, that initial
691 	 * KPTmap can only support NKPT page table pages.  Here, a larger
692 	 * KPTmap is created that can support KVA_PAGES page table pages.
693 	 */
694 	SYSMAP(pt_entry_t *, KPTD, KPTmap, KVA_PAGES)
695 
696 	for (i = 0; i < NKPT; i++)
697 		KPTD[i] = (KPTphys + ptoa(i)) | PG_RW | PG_V;
698 
699 	/*
700 	 * PADDR1 and PADDR2 are used by pmap_pte_quick() and pmap_pte(),
701 	 * respectively.
702 	 */
703 	SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1)
704 	SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1)
705 	SYSMAP(pt_entry_t *, PMAP3, PADDR3, 1)
706 
707 	mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF);
708 
709 	virtual_avail = va;
710 
711 	/*
712 	 * Initialize the PAT MSR if present.
713 	 * pmap_init_pat() clears and sets CR4_PGE, which, as a
714 	 * side-effect, invalidates stale PG_G TLB entries that might
715 	 * have been created in our pre-boot environment.  We assume
716 	 * that PAT support implies PGE and in reverse, PGE presence
717 	 * comes with PAT.  Both features were added for Pentium Pro.
718 	 */
719 	pmap_init_pat();
720 }
721 
722 static void
pmap_init_reserved_pages(void)723 pmap_init_reserved_pages(void)
724 {
725 	struct pcpu *pc;
726 	vm_offset_t pages;
727 	int i;
728 
729 #ifdef PMAP_PAE_COMP
730 	if (!pae_mode)
731 		return;
732 #else
733 	if (pae_mode)
734 		return;
735 #endif
736 	CPU_FOREACH(i) {
737 		pc = pcpu_find(i);
738 		mtx_init(&pc->pc_copyout_mlock, "cpmlk", NULL, MTX_DEF |
739 		    MTX_NEW);
740 		pc->pc_copyout_maddr = kva_alloc(ptoa(2));
741 		if (pc->pc_copyout_maddr == 0)
742 			panic("unable to allocate non-sleepable copyout KVA");
743 		sx_init(&pc->pc_copyout_slock, "cpslk");
744 		pc->pc_copyout_saddr = kva_alloc(ptoa(2));
745 		if (pc->pc_copyout_saddr == 0)
746 			panic("unable to allocate sleepable copyout KVA");
747 		pc->pc_pmap_eh_va = kva_alloc(ptoa(1));
748 		if (pc->pc_pmap_eh_va == 0)
749 			panic("unable to allocate pmap_extract_and_hold KVA");
750 		pc->pc_pmap_eh_ptep = (char *)vtopte(pc->pc_pmap_eh_va);
751 
752 		/*
753 		 * Skip if the mappings have already been initialized,
754 		 * i.e. this is the BSP.
755 		 */
756 		if (pc->pc_cmap_addr1 != 0)
757 			continue;
758 
759 		mtx_init(&pc->pc_cmap_lock, "SYSMAPS", NULL, MTX_DEF);
760 		pages = kva_alloc(PAGE_SIZE * 3);
761 		if (pages == 0)
762 			panic("unable to allocate CMAP KVA");
763 		pc->pc_cmap_pte1 = vtopte(pages);
764 		pc->pc_cmap_pte2 = vtopte(pages + PAGE_SIZE);
765 		pc->pc_cmap_addr1 = (caddr_t)pages;
766 		pc->pc_cmap_addr2 = (caddr_t)(pages + PAGE_SIZE);
767 		pc->pc_qmap_addr = pages + ptoa(2);
768 	}
769 }
770 
771 SYSINIT(rpages_init, SI_SUB_CPU, SI_ORDER_ANY, pmap_init_reserved_pages, NULL);
772 
773 /*
774  * Setup the PAT MSR.
775  */
776 static void
__CONCAT(PMTYPE,init_pat)777 __CONCAT(PMTYPE, init_pat)(void)
778 {
779 	int pat_table[PAT_INDEX_SIZE];
780 	uint64_t pat_msr;
781 	u_long cr0, cr4;
782 	int i;
783 
784 	/* Set default PAT index table. */
785 	for (i = 0; i < PAT_INDEX_SIZE; i++)
786 		pat_table[i] = -1;
787 	pat_table[PAT_WRITE_BACK] = 0;
788 	pat_table[PAT_WRITE_THROUGH] = 1;
789 	pat_table[PAT_UNCACHEABLE] = 3;
790 	pat_table[PAT_WRITE_COMBINING] = 3;
791 	pat_table[PAT_WRITE_PROTECTED] = 3;
792 	pat_table[PAT_UNCACHED] = 3;
793 
794 	/*
795 	 * Bail if this CPU doesn't implement PAT.
796 	 * We assume that PAT support implies PGE.
797 	 */
798 	if ((cpu_feature & CPUID_PAT) == 0) {
799 		for (i = 0; i < PAT_INDEX_SIZE; i++)
800 			pat_index[i] = pat_table[i];
801 		pat_works = 0;
802 		return;
803 	}
804 
805 	/*
806 	 * Due to some Intel errata, we can only safely use the lower 4
807 	 * PAT entries.
808 	 *
809 	 *   Intel Pentium III Processor Specification Update
810 	 * Errata E.27 (Upper Four PAT Entries Not Usable With Mode B
811 	 * or Mode C Paging)
812 	 *
813 	 *   Intel Pentium IV  Processor Specification Update
814 	 * Errata N46 (PAT Index MSB May Be Calculated Incorrectly)
815 	 */
816 	if (cpu_vendor_id == CPU_VENDOR_INTEL &&
817 	    !(CPUID_TO_FAMILY(cpu_id) == 6 && CPUID_TO_MODEL(cpu_id) >= 0xe))
818 		pat_works = 0;
819 
820 	/* Initialize default PAT entries. */
821 	pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |
822 	    PAT_VALUE(1, PAT_WRITE_THROUGH) |
823 	    PAT_VALUE(2, PAT_UNCACHED) |
824 	    PAT_VALUE(3, PAT_UNCACHEABLE) |
825 	    PAT_VALUE(4, PAT_WRITE_BACK) |
826 	    PAT_VALUE(5, PAT_WRITE_THROUGH) |
827 	    PAT_VALUE(6, PAT_UNCACHED) |
828 	    PAT_VALUE(7, PAT_UNCACHEABLE);
829 
830 	if (pat_works) {
831 		/*
832 		 * Leave the indices 0-3 at the default of WB, WT, UC-, and UC.
833 		 * Program 5 and 6 as WP and WC.
834 		 * Leave 4 and 7 as WB and UC.
835 		 */
836 		pat_msr &= ~(PAT_MASK(5) | PAT_MASK(6));
837 		pat_msr |= PAT_VALUE(5, PAT_WRITE_PROTECTED) |
838 		    PAT_VALUE(6, PAT_WRITE_COMBINING);
839 		pat_table[PAT_UNCACHED] = 2;
840 		pat_table[PAT_WRITE_PROTECTED] = 5;
841 		pat_table[PAT_WRITE_COMBINING] = 6;
842 	} else {
843 		/*
844 		 * Just replace PAT Index 2 with WC instead of UC-.
845 		 */
846 		pat_msr &= ~PAT_MASK(2);
847 		pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING);
848 		pat_table[PAT_WRITE_COMBINING] = 2;
849 	}
850 
851 	/* Disable PGE. */
852 	cr4 = rcr4();
853 	load_cr4(cr4 & ~CR4_PGE);
854 
855 	/* Disable caches (CD = 1, NW = 0). */
856 	cr0 = rcr0();
857 	load_cr0((cr0 & ~CR0_NW) | CR0_CD);
858 
859 	/* Flushes caches and TLBs. */
860 	wbinvd();
861 	invltlb();
862 
863 	/* Update PAT and index table. */
864 	wrmsr(MSR_PAT, pat_msr);
865 	for (i = 0; i < PAT_INDEX_SIZE; i++)
866 		pat_index[i] = pat_table[i];
867 
868 	/* Flush caches and TLBs again. */
869 	wbinvd();
870 	invltlb();
871 
872 	/* Restore caches and PGE. */
873 	load_cr0(cr0);
874 	load_cr4(cr4);
875 }
876 
877 #ifdef PMAP_PAE_COMP
878 static void *
pmap_pdpt_allocf(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * sflagsp,int flags)879 pmap_pdpt_allocf(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *sflagsp,
880     int flags)
881 {
882 
883 	/* Inform UMA that this allocator uses kernel_map/object. */
884 	*sflagsp = UMA_SLAB_KERNEL;
885 	/* contig allocations cannot be NEVERFREED */
886 	flags &= ~M_NEVERFREED;
887 	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
888 	    bytes, flags, 0x0ULL, 0xffffffffULL, 1, 0, VM_MEMATTR_DEFAULT));
889 }
890 #endif
891 
892 /*
893  * Abuse the pte nodes for unmapped kva to thread a kva freelist through.
894  * Requirements:
895  *  - Must deal with pages in order to ensure that none of the PG_* bits
896  *    are ever set, PG_V in particular.
897  *  - Assumes we can write to ptes without pte_store() atomic ops, even
898  *    on PAE systems.  This should be ok.
899  *  - Assumes nothing will ever test these addresses for 0 to indicate
900  *    no mapping instead of correctly checking PG_V.
901  *  - Assumes a vm_offset_t will fit in a pte (true for i386).
902  * Because PG_V is never set, there can be no mappings to invalidate.
903  */
904 static vm_offset_t
pmap_ptelist_alloc(vm_offset_t * head)905 pmap_ptelist_alloc(vm_offset_t *head)
906 {
907 	pt_entry_t *pte;
908 	vm_offset_t va;
909 
910 	va = *head;
911 	if (va == 0)
912 		panic("pmap_ptelist_alloc: exhausted ptelist KVA");
913 	pte = vtopte(va);
914 	*head = *pte;
915 	if (*head & PG_V)
916 		panic("pmap_ptelist_alloc: va with PG_V set!");
917 	*pte = 0;
918 	return (va);
919 }
920 
921 static void
pmap_ptelist_free(vm_offset_t * head,vm_offset_t va)922 pmap_ptelist_free(vm_offset_t *head, vm_offset_t va)
923 {
924 	pt_entry_t *pte;
925 
926 	if (va & PG_V)
927 		panic("pmap_ptelist_free: freeing va with PG_V set!");
928 	pte = vtopte(va);
929 	*pte = *head;		/* virtual! PG_V is 0 though */
930 	*head = va;
931 }
932 
933 static void
pmap_ptelist_init(vm_offset_t * head,void * base,int npages)934 pmap_ptelist_init(vm_offset_t *head, void *base, int npages)
935 {
936 	int i;
937 	vm_offset_t va;
938 
939 	*head = 0;
940 	for (i = npages - 1; i >= 0; i--) {
941 		va = (vm_offset_t)base + i * PAGE_SIZE;
942 		pmap_ptelist_free(head, va);
943 	}
944 }
945 
946 /*
947  *	Initialize the pmap module.
948  *
949  *	Called by vm_mem_init(), to initialize any structures that the pmap
950  *	system needs to map virtual memory.
951  */
952 static void
__CONCAT(PMTYPE,init)953 __CONCAT(PMTYPE, init)(void)
954 {
955 	struct pmap_preinit_mapping *ppim;
956 	vm_page_t mpte;
957 	vm_size_t s;
958 	int i, pv_npg;
959 
960 	/*
961 	 * Initialize the vm page array entries for the kernel pmap's
962 	 * page table pages.
963 	 */
964 	PMAP_LOCK(kernel_pmap);
965 	for (i = 0; i < NKPT; i++) {
966 		mpte = PHYS_TO_VM_PAGE(KPTphys + ptoa(i));
967 		KASSERT(mpte >= vm_page_array &&
968 		    mpte < &vm_page_array[vm_page_array_size],
969 		    ("pmap_init: page table page is out of range"));
970 		mpte->pindex = i + KPTDI;
971 		mpte->phys_addr = KPTphys + ptoa(i);
972 		mpte->ref_count = 1;
973 
974 		/*
975 		 * Collect the page table pages that were replaced by a 2/4MB
976 		 * page.  They are filled with equivalent 4KB page mappings.
977 		 */
978 		if (pseflag != 0 &&
979 		    KERNBASE <= i << PDRSHIFT && i << PDRSHIFT < KERNend &&
980 		    pmap_insert_pt_page(kernel_pmap, mpte, true, true))
981 			panic("pmap_init: pmap_insert_pt_page failed");
982 	}
983 	PMAP_UNLOCK(kernel_pmap);
984 	vm_wire_add(NKPT);
985 
986 	/*
987 	 * Initialize the address space (zone) for the pv entries.  Set a
988 	 * high water mark so that the system can recover from excessive
989 	 * numbers of pv entries.
990 	 */
991 	TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
992 	pv_entry_max = shpgperproc * maxproc + vm_cnt.v_page_count;
993 	TUNABLE_INT_FETCH("vm.pmap.pv_entry_max", &pv_entry_max);
994 	pv_entry_max = roundup(pv_entry_max, _NPCPV);
995 	pv_entry_high_water = 9 * (pv_entry_max / 10);
996 
997 	/*
998 	 * If the kernel is running on a virtual machine, then it must assume
999 	 * that MCA is enabled by the hypervisor.  Moreover, the kernel must
1000 	 * be prepared for the hypervisor changing the vendor and family that
1001 	 * are reported by CPUID.  Consequently, the workaround for AMD Family
1002 	 * 10h Erratum 383 is enabled if the processor's feature set does not
1003 	 * include at least one feature that is only supported by older Intel
1004 	 * or newer AMD processors.
1005 	 */
1006 	if (vm_guest != VM_GUEST_NO && (cpu_feature & CPUID_SS) == 0 &&
1007 	    (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI |
1008 	    CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP |
1009 	    AMDID2_FMA4)) == 0)
1010 		workaround_erratum383 = 1;
1011 
1012 	/*
1013 	 * Are large page mappings supported and enabled?
1014 	 */
1015 	TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled);
1016 	if (pseflag == 0)
1017 		pg_ps_enabled = 0;
1018 	else if (pg_ps_enabled) {
1019 		KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
1020 		    ("pmap_init: can't assign to pagesizes[1]"));
1021 		pagesizes[1] = NBPDR;
1022 	}
1023 
1024 	/*
1025 	 * Calculate the size of the pv head table for superpages.
1026 	 * Handle the possibility that "vm_phys_segs[...].end" is zero.
1027 	 */
1028 	pv_npg = trunc_4mpage(vm_phys_segs[vm_phys_nsegs - 1].end -
1029 	    PAGE_SIZE) / NBPDR + 1;
1030 
1031 	/*
1032 	 * Allocate memory for the pv head table for superpages.
1033 	 */
1034 	s = (vm_size_t)(pv_npg * sizeof(struct md_page));
1035 	s = round_page(s);
1036 	pv_table = kmem_malloc(s, M_WAITOK | M_ZERO);
1037 	for (i = 0; i < pv_npg; i++)
1038 		TAILQ_INIT(&pv_table[i].pv_list);
1039 
1040 	pv_maxchunks = MAX(pv_entry_max / _NPCPV, maxproc);
1041 	pv_chunkbase = (struct pv_chunk *)kva_alloc(PAGE_SIZE * pv_maxchunks);
1042 	if (pv_chunkbase == NULL)
1043 		panic("pmap_init: not enough kvm for pv chunks");
1044 	pmap_ptelist_init(&pv_vafree, pv_chunkbase, pv_maxchunks);
1045 #ifdef PMAP_PAE_COMP
1046 	pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL,
1047 	    NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1,
1048 	    UMA_ZONE_CONTIG | UMA_ZONE_VM | UMA_ZONE_NOFREE);
1049 	uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf);
1050 #endif
1051 
1052 	pmap_initialized = 1;
1053 	pmap_init_trm();
1054 
1055 	if (!bootverbose)
1056 		return;
1057 	for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
1058 		ppim = pmap_preinit_mapping + i;
1059 		if (ppim->va == 0)
1060 			continue;
1061 		printf("PPIM %u: PA=%#jx, VA=%#x, size=%#x, mode=%#x\n", i,
1062 		    (uintmax_t)ppim->pa, ppim->va, ppim->sz, ppim->mode);
1063 	}
1064 
1065 }
1066 
1067 extern u_long pmap_pde_demotions;
1068 extern u_long pmap_pde_mappings;
1069 extern u_long pmap_pde_p_failures;
1070 extern u_long pmap_pde_promotions;
1071 
1072 /***************************************************
1073  * Low level helper routines.....
1074  ***************************************************/
1075 
1076 static bool
__CONCAT(PMTYPE,is_valid_memattr)1077 __CONCAT(PMTYPE, is_valid_memattr)(pmap_t pmap __unused, vm_memattr_t mode)
1078 {
1079 
1080 	return (mode >= 0 && mode < PAT_INDEX_SIZE &&
1081 	    pat_index[(int)mode] >= 0);
1082 }
1083 
1084 /*
1085  * Determine the appropriate bits to set in a PTE or PDE for a specified
1086  * caching mode.
1087  */
1088 static int
__CONCAT(PMTYPE,cache_bits)1089 __CONCAT(PMTYPE, cache_bits)(pmap_t pmap, int mode, bool is_pde)
1090 {
1091 	int cache_bits, pat_flag, pat_idx;
1092 
1093 	if (!pmap_is_valid_memattr(pmap, mode))
1094 		panic("Unknown caching mode %d\n", mode);
1095 
1096 	/* The PAT bit is different for PTE's and PDE's. */
1097 	pat_flag = is_pde ? PG_PDE_PAT : PG_PTE_PAT;
1098 
1099 	/* Map the caching mode to a PAT index. */
1100 	pat_idx = pat_index[mode];
1101 
1102 	/* Map the 3-bit index value into the PAT, PCD, and PWT bits. */
1103 	cache_bits = 0;
1104 	if (pat_idx & 0x4)
1105 		cache_bits |= pat_flag;
1106 	if (pat_idx & 0x2)
1107 		cache_bits |= PG_NC_PCD;
1108 	if (pat_idx & 0x1)
1109 		cache_bits |= PG_NC_PWT;
1110 	return (cache_bits);
1111 }
1112 
1113 static int
pmap_pat_index(pmap_t pmap,pt_entry_t pte,bool is_pde)1114 pmap_pat_index(pmap_t pmap, pt_entry_t pte, bool is_pde)
1115 {
1116 	int pat_flag, pat_idx;
1117 
1118 	if ((cpu_feature & CPUID_PAT) == 0)
1119 		return (0);
1120 
1121 	pat_idx = 0;
1122 	/* The PAT bit is different for PTE's and PDE's. */
1123 	pat_flag = is_pde ? PG_PDE_PAT : PG_PTE_PAT;
1124 
1125 	if ((pte & pat_flag) != 0)
1126 		pat_idx |= 0x4;
1127 	if ((pte & PG_NC_PCD) != 0)
1128 		pat_idx |= 0x2;
1129 	if ((pte & PG_NC_PWT) != 0)
1130 		pat_idx |= 0x1;
1131 
1132 	/* See pmap_init_pat(). */
1133 	if (pat_works) {
1134 		if (pat_idx == 4)
1135 			pat_idx = 0;
1136 		if (pat_idx == 7)
1137 			pat_idx = 3;
1138 	} else {
1139 		/* XXXKIB */
1140 	}
1141 
1142 	return (pat_idx);
1143 }
1144 
1145 static bool
__CONCAT(PMTYPE,ps_enabled)1146 __CONCAT(PMTYPE, ps_enabled)(pmap_t pmap __unused)
1147 {
1148 
1149 	return (pg_ps_enabled);
1150 }
1151 
1152 /*
1153  * The caller is responsible for maintaining TLB consistency.
1154  */
1155 static void
pmap_kenter_pde(vm_offset_t va,pd_entry_t newpde)1156 pmap_kenter_pde(vm_offset_t va, pd_entry_t newpde)
1157 {
1158 	pd_entry_t *pde;
1159 
1160 	pde = pmap_pde(kernel_pmap, va);
1161 	pde_store(pde, newpde);
1162 }
1163 
1164 /*
1165  * After changing the page size for the specified virtual address in the page
1166  * table, flush the corresponding entries from the processor's TLB.  Only the
1167  * calling processor's TLB is affected.
1168  *
1169  * The calling thread must be pinned to a processor.
1170  */
1171 static void
pmap_update_pde_invalidate(vm_offset_t va,pd_entry_t newpde)1172 pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde)
1173 {
1174 
1175 	if ((newpde & PG_PS) == 0)
1176 		/* Demotion: flush a specific 2MB page mapping. */
1177 		invlpg(va);
1178 	else /* if ((newpde & PG_G) == 0) */
1179 		/*
1180 		 * Promotion: flush every 4KB page mapping from the TLB
1181 		 * because there are too many to flush individually.
1182 		 */
1183 		invltlb();
1184 }
1185 
1186 #ifdef SMP
1187 
1188 static void
pmap_curcpu_cb_dummy(pmap_t pmap __unused,vm_offset_t addr1 __unused,vm_offset_t addr2 __unused)1189 pmap_curcpu_cb_dummy(pmap_t pmap __unused, vm_offset_t addr1 __unused,
1190     vm_offset_t addr2 __unused)
1191 {
1192 }
1193 
1194 /*
1195  * For SMP, these functions have to use the IPI mechanism for coherence.
1196  *
1197  * N.B.: Before calling any of the following TLB invalidation functions,
1198  * the calling processor must ensure that all stores updating a non-
1199  * kernel page table are globally performed.  Otherwise, another
1200  * processor could cache an old, pre-update entry without being
1201  * invalidated.  This can happen one of two ways: (1) The pmap becomes
1202  * active on another processor after its pm_active field is checked by
1203  * one of the following functions but before a store updating the page
1204  * table is globally performed. (2) The pmap becomes active on another
1205  * processor before its pm_active field is checked but due to
1206  * speculative loads one of the following functions stills reads the
1207  * pmap as inactive on the other processor.
1208  *
1209  * The kernel page table is exempt because its pm_active field is
1210  * immutable.  The kernel page table is always active on every
1211  * processor.
1212  */
1213 static void
pmap_invalidate_page_int(pmap_t pmap,vm_offset_t va)1214 pmap_invalidate_page_int(pmap_t pmap, vm_offset_t va)
1215 {
1216 	cpuset_t *mask, other_cpus;
1217 	u_int cpuid;
1218 
1219 	sched_pin();
1220 	if (pmap == kernel_pmap) {
1221 		invlpg(va);
1222 		mask = &all_cpus;
1223 	} else if (!CPU_CMP(&pmap->pm_active, &all_cpus)) {
1224 		mask = &all_cpus;
1225 	} else {
1226 		cpuid = PCPU_GET(cpuid);
1227 		other_cpus = all_cpus;
1228 		CPU_CLR(cpuid, &other_cpus);
1229 		CPU_AND(&other_cpus, &other_cpus, &pmap->pm_active);
1230 		mask = &other_cpus;
1231 	}
1232 	smp_masked_invlpg(*mask, va, pmap, pmap_curcpu_cb_dummy);
1233 	sched_unpin();
1234 }
1235 
1236 /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */
1237 #define	PMAP_INVLPG_THRESHOLD	(4 * 1024 * PAGE_SIZE)
1238 
1239 static void
pmap_invalidate_range_int(pmap_t pmap,vm_offset_t sva,vm_offset_t eva)1240 pmap_invalidate_range_int(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1241 {
1242 	cpuset_t *mask, other_cpus;
1243 	vm_offset_t addr;
1244 	u_int cpuid;
1245 
1246 	if (eva - sva >= PMAP_INVLPG_THRESHOLD) {
1247 		pmap_invalidate_all_int(pmap);
1248 		return;
1249 	}
1250 
1251 	sched_pin();
1252 	if (pmap == kernel_pmap) {
1253 		for (addr = sva; addr < eva; addr += PAGE_SIZE)
1254 			invlpg(addr);
1255 		mask = &all_cpus;
1256 	} else  if (!CPU_CMP(&pmap->pm_active, &all_cpus)) {
1257 		mask = &all_cpus;
1258 	} else {
1259 		cpuid = PCPU_GET(cpuid);
1260 		other_cpus = all_cpus;
1261 		CPU_CLR(cpuid, &other_cpus);
1262 		CPU_AND(&other_cpus, &other_cpus, &pmap->pm_active);
1263 		mask = &other_cpus;
1264 	}
1265 	smp_masked_invlpg_range(*mask, sva, eva, pmap, pmap_curcpu_cb_dummy);
1266 	sched_unpin();
1267 }
1268 
1269 static void
pmap_invalidate_all_int(pmap_t pmap)1270 pmap_invalidate_all_int(pmap_t pmap)
1271 {
1272 	cpuset_t *mask, other_cpus;
1273 	u_int cpuid;
1274 
1275 	sched_pin();
1276 	if (pmap == kernel_pmap) {
1277 		invltlb();
1278 		mask = &all_cpus;
1279 	} else if (!CPU_CMP(&pmap->pm_active, &all_cpus)) {
1280 		mask = &all_cpus;
1281 	} else {
1282 		cpuid = PCPU_GET(cpuid);
1283 		other_cpus = all_cpus;
1284 		CPU_CLR(cpuid, &other_cpus);
1285 		CPU_AND(&other_cpus, &other_cpus, &pmap->pm_active);
1286 		mask = &other_cpus;
1287 	}
1288 	smp_masked_invltlb(*mask, pmap, pmap_curcpu_cb_dummy);
1289 	sched_unpin();
1290 }
1291 
1292 static void
pmap_invalidate_cache_curcpu_cb(pmap_t pmap __unused,vm_offset_t addr1 __unused,vm_offset_t addr2 __unused)1293 pmap_invalidate_cache_curcpu_cb(pmap_t pmap __unused,
1294     vm_offset_t addr1 __unused, vm_offset_t addr2 __unused)
1295 {
1296 	wbinvd();
1297 }
1298 
1299 static void
__CONCAT(PMTYPE,invalidate_cache)1300 __CONCAT(PMTYPE, invalidate_cache)(void)
1301 {
1302 	smp_cache_flush(pmap_invalidate_cache_curcpu_cb);
1303 }
1304 
1305 struct pde_action {
1306 	cpuset_t invalidate;	/* processors that invalidate their TLB */
1307 	vm_offset_t va;
1308 	pd_entry_t *pde;
1309 	pd_entry_t newpde;
1310 	u_int store;		/* processor that updates the PDE */
1311 };
1312 
1313 static void
pmap_update_pde_kernel(void * arg)1314 pmap_update_pde_kernel(void *arg)
1315 {
1316 	struct pde_action *act = arg;
1317 	pd_entry_t *pde;
1318 
1319 	if (act->store == PCPU_GET(cpuid)) {
1320 		pde = pmap_pde(kernel_pmap, act->va);
1321 		pde_store(pde, act->newpde);
1322 	}
1323 }
1324 
1325 static void
pmap_update_pde_user(void * arg)1326 pmap_update_pde_user(void *arg)
1327 {
1328 	struct pde_action *act = arg;
1329 
1330 	if (act->store == PCPU_GET(cpuid))
1331 		pde_store(act->pde, act->newpde);
1332 }
1333 
1334 static void
pmap_update_pde_teardown(void * arg)1335 pmap_update_pde_teardown(void *arg)
1336 {
1337 	struct pde_action *act = arg;
1338 
1339 	if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate))
1340 		pmap_update_pde_invalidate(act->va, act->newpde);
1341 }
1342 
1343 /*
1344  * Change the page size for the specified virtual address in a way that
1345  * prevents any possibility of the TLB ever having two entries that map the
1346  * same virtual address using different page sizes.  This is the recommended
1347  * workaround for Erratum 383 on AMD Family 10h processors.  It prevents a
1348  * machine check exception for a TLB state that is improperly diagnosed as a
1349  * hardware error.
1350  */
1351 static void
pmap_update_pde(pmap_t pmap,vm_offset_t va,pd_entry_t * pde,pd_entry_t newpde)1352 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
1353 {
1354 	struct pde_action act;
1355 	cpuset_t active, other_cpus;
1356 	u_int cpuid;
1357 
1358 	sched_pin();
1359 	cpuid = PCPU_GET(cpuid);
1360 	other_cpus = all_cpus;
1361 	CPU_CLR(cpuid, &other_cpus);
1362 	if (pmap == kernel_pmap)
1363 		active = all_cpus;
1364 	else
1365 		active = pmap->pm_active;
1366 	if (CPU_OVERLAP(&active, &other_cpus)) {
1367 		act.store = cpuid;
1368 		act.invalidate = active;
1369 		act.va = va;
1370 		act.pde = pde;
1371 		act.newpde = newpde;
1372 		CPU_SET(cpuid, &active);
1373 		smp_rendezvous_cpus(active,
1374 		    smp_no_rendezvous_barrier, pmap == kernel_pmap ?
1375 		    pmap_update_pde_kernel : pmap_update_pde_user,
1376 		    pmap_update_pde_teardown, &act);
1377 	} else {
1378 		if (pmap == kernel_pmap)
1379 			pmap_kenter_pde(va, newpde);
1380 		else
1381 			pde_store(pde, newpde);
1382 		if (CPU_ISSET(cpuid, &active))
1383 			pmap_update_pde_invalidate(va, newpde);
1384 	}
1385 	sched_unpin();
1386 }
1387 #else /* !SMP */
1388 /*
1389  * Normal, non-SMP, 486+ invalidation functions.
1390  * We inline these within pmap.c for speed.
1391  */
1392 static void
pmap_invalidate_page_int(pmap_t pmap,vm_offset_t va)1393 pmap_invalidate_page_int(pmap_t pmap, vm_offset_t va)
1394 {
1395 
1396 	if (pmap == kernel_pmap)
1397 		invlpg(va);
1398 }
1399 
1400 static void
pmap_invalidate_range_int(pmap_t pmap,vm_offset_t sva,vm_offset_t eva)1401 pmap_invalidate_range_int(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1402 {
1403 	vm_offset_t addr;
1404 
1405 	if (pmap == kernel_pmap)
1406 		for (addr = sva; addr < eva; addr += PAGE_SIZE)
1407 			invlpg(addr);
1408 }
1409 
1410 static void
pmap_invalidate_all_int(pmap_t pmap)1411 pmap_invalidate_all_int(pmap_t pmap)
1412 {
1413 
1414 	if (pmap == kernel_pmap)
1415 		invltlb();
1416 }
1417 
1418 static void
__CONCAT(PMTYPE,invalidate_cache)1419 __CONCAT(PMTYPE, invalidate_cache)(void)
1420 {
1421 
1422 	wbinvd();
1423 }
1424 
1425 static void
pmap_update_pde(pmap_t pmap,vm_offset_t va,pd_entry_t * pde,pd_entry_t newpde)1426 pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde)
1427 {
1428 
1429 	if (pmap == kernel_pmap)
1430 		pmap_kenter_pde(va, newpde);
1431 	else
1432 		pde_store(pde, newpde);
1433 	if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active))
1434 		pmap_update_pde_invalidate(va, newpde);
1435 }
1436 #endif /* !SMP */
1437 
1438 static void
__CONCAT(PMTYPE,invalidate_page)1439 __CONCAT(PMTYPE, invalidate_page)(pmap_t pmap, vm_offset_t va)
1440 {
1441 
1442 	pmap_invalidate_page_int(pmap, va);
1443 }
1444 
1445 static void
__CONCAT(PMTYPE,invalidate_range)1446 __CONCAT(PMTYPE, invalidate_range)(pmap_t pmap, vm_offset_t sva,
1447     vm_offset_t eva)
1448 {
1449 
1450 	pmap_invalidate_range_int(pmap, sva, eva);
1451 }
1452 
1453 static void
__CONCAT(PMTYPE,invalidate_all)1454 __CONCAT(PMTYPE, invalidate_all)(pmap_t pmap)
1455 {
1456 
1457 	pmap_invalidate_all_int(pmap);
1458 }
1459 
1460 static void
pmap_invalidate_pde_page(pmap_t pmap,vm_offset_t va,pd_entry_t pde)1461 pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde)
1462 {
1463 
1464 	/*
1465 	 * When the PDE has PG_PROMOTED set, the 2- or 4MB page mapping was
1466 	 * created by a promotion that did not invalidate the 512 or 1024 4KB
1467 	 * page mappings that might exist in the TLB.  Consequently, at this
1468 	 * point, the TLB may hold both 4KB and 2- or 4MB page mappings for
1469 	 * the address range [va, va + NBPDR).  Therefore, the entire range
1470 	 * must be invalidated here.  In contrast, when PG_PROMOTED is clear,
1471 	 * the TLB will not hold any 4KB page mappings for the address range
1472 	 * [va, va + NBPDR), and so a single INVLPG suffices to invalidate the
1473 	 * 2- or 4MB page mapping from the TLB.
1474 	 */
1475 	if ((pde & PG_PROMOTED) != 0)
1476 		pmap_invalidate_range_int(pmap, va, va + NBPDR - 1);
1477 	else
1478 		pmap_invalidate_page_int(pmap, va);
1479 }
1480 
1481 /*
1482  * Are we current address space or kernel?
1483  */
1484 static __inline int
pmap_is_current(pmap_t pmap)1485 pmap_is_current(pmap_t pmap)
1486 {
1487 
1488 	return (pmap == kernel_pmap);
1489 }
1490 
1491 /*
1492  * If the given pmap is not the current or kernel pmap, the returned pte must
1493  * be released by passing it to pmap_pte_release().
1494  */
1495 static pt_entry_t *
__CONCAT(PMTYPE,pte)1496 __CONCAT(PMTYPE, pte)(pmap_t pmap, vm_offset_t va)
1497 {
1498 	pd_entry_t newpf;
1499 	pd_entry_t *pde;
1500 
1501 	pde = pmap_pde(pmap, va);
1502 	if (*pde & PG_PS)
1503 		return (pde);
1504 	if (*pde != 0) {
1505 		/* are we current address space or kernel? */
1506 		if (pmap_is_current(pmap))
1507 			return (vtopte(va));
1508 		mtx_lock(&PMAP2mutex);
1509 		newpf = *pde & PG_FRAME;
1510 		if ((*PMAP2 & PG_FRAME) != newpf) {
1511 			*PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M;
1512 			pmap_invalidate_page_int(kernel_pmap,
1513 			    (vm_offset_t)PADDR2);
1514 		}
1515 		return (PADDR2 + (i386_btop(va) & (NPTEPG - 1)));
1516 	}
1517 	return (NULL);
1518 }
1519 
1520 /*
1521  * Releases a pte that was obtained from pmap_pte().  Be prepared for the pte
1522  * being NULL.
1523  */
1524 static __inline void
pmap_pte_release(pt_entry_t * pte)1525 pmap_pte_release(pt_entry_t *pte)
1526 {
1527 
1528 	if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2)
1529 		mtx_unlock(&PMAP2mutex);
1530 }
1531 
1532 /*
1533  * NB:  The sequence of updating a page table followed by accesses to the
1534  * corresponding pages is subject to the situation described in the "AMD64
1535  * Architecture Programmer's Manual Volume 2: System Programming" rev. 3.23,
1536  * "7.3.1 Special Coherency Considerations".  Therefore, issuing the INVLPG
1537  * right after modifying the PTE bits is crucial.
1538  */
1539 static __inline void
invlcaddr(void * caddr)1540 invlcaddr(void *caddr)
1541 {
1542 
1543 	invlpg((u_int)caddr);
1544 }
1545 
1546 /*
1547  * Super fast pmap_pte routine best used when scanning
1548  * the pv lists.  This eliminates many coarse-grained
1549  * invltlb calls.  Note that many of the pv list
1550  * scans are across different pmaps.  It is very wasteful
1551  * to do an entire invltlb for checking a single mapping.
1552  *
1553  * If the given pmap is not the current pmap, pvh_global_lock
1554  * must be held and curthread pinned to a CPU.
1555  */
1556 static pt_entry_t *
pmap_pte_quick(pmap_t pmap,vm_offset_t va)1557 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
1558 {
1559 	pd_entry_t newpf;
1560 	pd_entry_t *pde;
1561 
1562 	pde = pmap_pde(pmap, va);
1563 	if (*pde & PG_PS)
1564 		return (pde);
1565 	if (*pde != 0) {
1566 		/* are we current address space or kernel? */
1567 		if (pmap_is_current(pmap))
1568 			return (vtopte(va));
1569 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1570 		KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
1571 		newpf = *pde & PG_FRAME;
1572 		if ((*PMAP1 & PG_FRAME) != newpf) {
1573 			*PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M;
1574 #ifdef SMP
1575 			PMAP1cpu = PCPU_GET(cpuid);
1576 #endif
1577 			invlcaddr(PADDR1);
1578 			PMAP1changed++;
1579 		} else
1580 #ifdef SMP
1581 		if (PMAP1cpu != PCPU_GET(cpuid)) {
1582 			PMAP1cpu = PCPU_GET(cpuid);
1583 			invlcaddr(PADDR1);
1584 			PMAP1changedcpu++;
1585 		} else
1586 #endif
1587 			PMAP1unchanged++;
1588 		return (PADDR1 + (i386_btop(va) & (NPTEPG - 1)));
1589 	}
1590 	return (0);
1591 }
1592 
1593 static pt_entry_t *
pmap_pte_quick3(pmap_t pmap,vm_offset_t va)1594 pmap_pte_quick3(pmap_t pmap, vm_offset_t va)
1595 {
1596 	pd_entry_t newpf;
1597 	pd_entry_t *pde;
1598 
1599 	pde = pmap_pde(pmap, va);
1600 	if (*pde & PG_PS)
1601 		return (pde);
1602 	if (*pde != 0) {
1603 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1604 		KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
1605 		newpf = *pde & PG_FRAME;
1606 		if ((*PMAP3 & PG_FRAME) != newpf) {
1607 			*PMAP3 = newpf | PG_RW | PG_V | PG_A | PG_M;
1608 #ifdef SMP
1609 			PMAP3cpu = PCPU_GET(cpuid);
1610 #endif
1611 			invlcaddr(PADDR3);
1612 			PMAP1changed++;
1613 		} else
1614 #ifdef SMP
1615 		if (PMAP3cpu != PCPU_GET(cpuid)) {
1616 			PMAP3cpu = PCPU_GET(cpuid);
1617 			invlcaddr(PADDR3);
1618 			PMAP1changedcpu++;
1619 		} else
1620 #endif
1621 			PMAP1unchanged++;
1622 		return (PADDR3 + (i386_btop(va) & (NPTEPG - 1)));
1623 	}
1624 	return (0);
1625 }
1626 
1627 static pt_entry_t
pmap_pte_ufast(pmap_t pmap,vm_offset_t va,pd_entry_t pde)1628 pmap_pte_ufast(pmap_t pmap, vm_offset_t va, pd_entry_t pde)
1629 {
1630 	pt_entry_t *eh_ptep, pte, *ptep;
1631 
1632 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1633 	pde &= PG_FRAME;
1634 	critical_enter();
1635 	eh_ptep = (pt_entry_t *)PCPU_GET(pmap_eh_ptep);
1636 	if ((*eh_ptep & PG_FRAME) != pde) {
1637 		*eh_ptep = pde | PG_RW | PG_V | PG_A | PG_M;
1638 		invlcaddr((void *)PCPU_GET(pmap_eh_va));
1639 	}
1640 	ptep = (pt_entry_t *)PCPU_GET(pmap_eh_va) + (i386_btop(va) &
1641 	    (NPTEPG - 1));
1642 	pte = *ptep;
1643 	critical_exit();
1644 	return (pte);
1645 }
1646 
1647 /*
1648  * Extract from the kernel page table the physical address that is mapped by
1649  * the given virtual address "va".
1650  *
1651  * This function may be used before pmap_bootstrap() is called.
1652  */
1653 static vm_paddr_t
__CONCAT(PMTYPE,kextract)1654 __CONCAT(PMTYPE, kextract)(vm_offset_t va)
1655 {
1656 	vm_paddr_t pa;
1657 
1658 	if ((pa = pte_load(&PTD[va >> PDRSHIFT])) & PG_PS) {
1659 		pa = (pa & PG_PS_FRAME) | (va & PDRMASK);
1660 	} else {
1661 		/*
1662 		 * Beware of a concurrent promotion that changes the PDE at
1663 		 * this point!  For example, vtopte() must not be used to
1664 		 * access the PTE because it would use the new PDE.  It is,
1665 		 * however, safe to use the old PDE because the page table
1666 		 * page is preserved by the promotion.
1667 		 */
1668 		pa = KPTmap[i386_btop(va)];
1669 		pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1670 	}
1671 	return (pa);
1672 }
1673 
1674 /*
1675  *	Routine:	pmap_extract
1676  *	Function:
1677  *		Extract the physical page address associated
1678  *		with the given map/virtual_address pair.
1679  */
1680 static vm_paddr_t
__CONCAT(PMTYPE,extract)1681 __CONCAT(PMTYPE, extract)(pmap_t pmap, vm_offset_t va)
1682 {
1683 	vm_paddr_t rtval;
1684 	pt_entry_t pte;
1685 	pd_entry_t pde;
1686 
1687 	rtval = 0;
1688 	PMAP_LOCK(pmap);
1689 	pde = pmap->pm_pdir[va >> PDRSHIFT];
1690 	if (pde != 0) {
1691 		if ((pde & PG_PS) != 0)
1692 			rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
1693 		else {
1694 			pte = pmap_pte_ufast(pmap, va, pde);
1695 			rtval = (pte & PG_FRAME) | (va & PAGE_MASK);
1696 		}
1697 	}
1698 	PMAP_UNLOCK(pmap);
1699 	return (rtval);
1700 }
1701 
1702 /*
1703  *	Routine:	pmap_extract_and_hold
1704  *	Function:
1705  *		Atomically extract and hold the physical page
1706  *		with the given pmap and virtual address pair
1707  *		if that mapping permits the given protection.
1708  */
1709 static vm_page_t
__CONCAT(PMTYPE,extract_and_hold)1710 __CONCAT(PMTYPE, extract_and_hold)(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1711 {
1712 	pd_entry_t pde;
1713 	pt_entry_t pte;
1714 	vm_page_t m;
1715 
1716 	m = NULL;
1717 	PMAP_LOCK(pmap);
1718 	pde = *pmap_pde(pmap, va);
1719 	if (pde != 0) {
1720 		if (pde & PG_PS) {
1721 			if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0)
1722 				m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) |
1723 				    (va & PDRMASK));
1724 		} else {
1725 			pte = pmap_pte_ufast(pmap, va, pde);
1726 			if (pte != 0 &&
1727 			    ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0))
1728 				m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
1729 		}
1730 		if (m != NULL && !vm_page_wire_mapped(m))
1731 			m = NULL;
1732 	}
1733 	PMAP_UNLOCK(pmap);
1734 	return (m);
1735 }
1736 
1737 /***************************************************
1738  * Low level mapping routines.....
1739  ***************************************************/
1740 
1741 /*
1742  * Add a wired page to the kva.
1743  * Note: not SMP coherent.
1744  *
1745  * This function may be used before pmap_bootstrap() is called.
1746  */
1747 static void
__CONCAT(PMTYPE,kenter)1748 __CONCAT(PMTYPE, kenter)(vm_offset_t va, vm_paddr_t pa)
1749 {
1750 	pt_entry_t *pte;
1751 
1752 	pte = vtopte(va);
1753 	pte_store(pte, pa | PG_RW | PG_V);
1754 }
1755 
1756 static __inline void
pmap_kenter_attr(vm_offset_t va,vm_paddr_t pa,int mode)1757 pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode)
1758 {
1759 	pt_entry_t *pte;
1760 
1761 	pte = vtopte(va);
1762 	pte_store(pte, pa | PG_RW | PG_V | pmap_cache_bits(kernel_pmap,
1763 	    mode, false));
1764 }
1765 
1766 /*
1767  * Remove a page from the kernel pagetables.
1768  * Note: not SMP coherent.
1769  *
1770  * This function may be used before pmap_bootstrap() is called.
1771  */
1772 static void
__CONCAT(PMTYPE,kremove)1773 __CONCAT(PMTYPE, kremove)(vm_offset_t va)
1774 {
1775 	pt_entry_t *pte;
1776 
1777 	pte = vtopte(va);
1778 	pte_clear(pte);
1779 }
1780 
1781 /*
1782  *	Used to map a range of physical addresses into kernel
1783  *	virtual address space.
1784  *
1785  *	The value passed in '*virt' is a suggested virtual address for
1786  *	the mapping. Architectures which can support a direct-mapped
1787  *	physical to virtual region can return the appropriate address
1788  *	within that region, leaving '*virt' unchanged. Other
1789  *	architectures should map the pages starting at '*virt' and
1790  *	update '*virt' with the first usable address after the mapped
1791  *	region.
1792  */
1793 static vm_offset_t
__CONCAT(PMTYPE,map)1794 __CONCAT(PMTYPE, map)(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end,
1795     int prot)
1796 {
1797 	vm_offset_t va, sva;
1798 	vm_paddr_t superpage_offset;
1799 	pd_entry_t newpde;
1800 
1801 	va = *virt;
1802 	/*
1803 	 * Does the physical address range's size and alignment permit at
1804 	 * least one superpage mapping to be created?
1805 	 */
1806 	superpage_offset = start & PDRMASK;
1807 	if ((end - start) - ((NBPDR - superpage_offset) & PDRMASK) >= NBPDR) {
1808 		/*
1809 		 * Increase the starting virtual address so that its alignment
1810 		 * does not preclude the use of superpage mappings.
1811 		 */
1812 		if ((va & PDRMASK) < superpage_offset)
1813 			va = (va & ~PDRMASK) + superpage_offset;
1814 		else if ((va & PDRMASK) > superpage_offset)
1815 			va = ((va + PDRMASK) & ~PDRMASK) + superpage_offset;
1816 	}
1817 	sva = va;
1818 	while (start < end) {
1819 		if ((start & PDRMASK) == 0 && end - start >= NBPDR &&
1820 		    pseflag != 0) {
1821 			KASSERT((va & PDRMASK) == 0,
1822 			    ("pmap_map: misaligned va %#x", va));
1823 			newpde = start | PG_PS | PG_RW | PG_V;
1824 			pmap_kenter_pde(va, newpde);
1825 			va += NBPDR;
1826 			start += NBPDR;
1827 		} else {
1828 			pmap_kenter(va, start);
1829 			va += PAGE_SIZE;
1830 			start += PAGE_SIZE;
1831 		}
1832 	}
1833 	pmap_invalidate_range_int(kernel_pmap, sva, va);
1834 	*virt = va;
1835 	return (sva);
1836 }
1837 
1838 /*
1839  * Add a list of wired pages to the kva
1840  * this routine is only used for temporary
1841  * kernel mappings that do not need to have
1842  * page modification or references recorded.
1843  * Note that old mappings are simply written
1844  * over.  The page *must* be wired.
1845  * Note: SMP coherent.  Uses a ranged shootdown IPI.
1846  */
1847 static void
__CONCAT(PMTYPE,qenter)1848 __CONCAT(PMTYPE, qenter)(vm_offset_t sva, vm_page_t *ma, int count)
1849 {
1850 	pt_entry_t *endpte, oldpte, pa, *pte;
1851 	vm_page_t m;
1852 
1853 	oldpte = 0;
1854 	pte = vtopte(sva);
1855 	endpte = pte + count;
1856 	while (pte < endpte) {
1857 		m = *ma++;
1858 		pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(kernel_pmap,
1859 		    m->md.pat_mode, false);
1860 		if ((*pte & (PG_FRAME | PG_PTE_CACHE)) != pa) {
1861 			oldpte |= *pte;
1862 			pte_store(pte, pa | pg_nx | PG_RW | PG_V);
1863 		}
1864 		pte++;
1865 	}
1866 	if (__predict_false((oldpte & PG_V) != 0))
1867 		pmap_invalidate_range_int(kernel_pmap, sva, sva + count *
1868 		    PAGE_SIZE);
1869 }
1870 
1871 /*
1872  * This routine tears out page mappings from the
1873  * kernel -- it is meant only for temporary mappings.
1874  * Note: SMP coherent.  Uses a ranged shootdown IPI.
1875  */
1876 static void
__CONCAT(PMTYPE,qremove)1877 __CONCAT(PMTYPE, qremove)(vm_offset_t sva, int count)
1878 {
1879 	vm_offset_t va;
1880 
1881 	va = sva;
1882 	while (count-- > 0) {
1883 		pmap_kremove(va);
1884 		va += PAGE_SIZE;
1885 	}
1886 	pmap_invalidate_range_int(kernel_pmap, sva, va);
1887 }
1888 
1889 /***************************************************
1890  * Page table page management routines.....
1891  ***************************************************/
1892 /*
1893  * Schedule the specified unused page table page to be freed.  Specifically,
1894  * add the page to the specified list of pages that will be released to the
1895  * physical memory manager after the TLB has been updated.
1896  */
1897 static __inline void
pmap_add_delayed_free_list(vm_page_t m,struct spglist * free,bool set_PG_ZERO)1898 pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, bool set_PG_ZERO)
1899 {
1900 
1901 	if (set_PG_ZERO)
1902 		m->flags |= PG_ZERO;
1903 	else
1904 		m->flags &= ~PG_ZERO;
1905 	SLIST_INSERT_HEAD(free, m, plinks.s.ss);
1906 }
1907 
1908 /*
1909  * Inserts the specified page table page into the specified pmap's collection
1910  * of idle page table pages.  Each of a pmap's page table pages is responsible
1911  * for mapping a distinct range of virtual addresses.  The pmap's collection is
1912  * ordered by this virtual address range.
1913  *
1914  * If "promoted" is false, then the page table page "mpte" must be zero filled;
1915  * "mpte"'s valid field will be set to 0.
1916  *
1917  * If "promoted" is true and "allpte_PG_A_set" is false, then "mpte" must
1918  * contain valid mappings with identical attributes except for PG_A; "mpte"'s
1919  * valid field will be set to 1.
1920  *
1921  * If "promoted" and "allpte_PG_A_set" are both true, then "mpte" must contain
1922  * valid mappings with identical attributes including PG_A; "mpte"'s valid
1923  * field will be set to VM_PAGE_BITS_ALL.
1924  */
1925 static __inline int
pmap_insert_pt_page(pmap_t pmap,vm_page_t mpte,bool promoted,bool allpte_PG_A_set)1926 pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted,
1927     bool allpte_PG_A_set)
1928 {
1929 
1930 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1931 	KASSERT(promoted || !allpte_PG_A_set,
1932 	    ("a zero-filled PTP can't have PG_A set in every PTE"));
1933 	mpte->valid = promoted ? (allpte_PG_A_set ? VM_PAGE_BITS_ALL : 1) : 0;
1934 	return (vm_radix_insert(&pmap->pm_root, mpte));
1935 }
1936 
1937 /*
1938  * Removes the page table page mapping the specified virtual address from the
1939  * specified pmap's collection of idle page table pages, and returns it.
1940  * Otherwise, returns NULL if there is no page table page corresponding to the
1941  * specified virtual address.
1942  */
1943 static __inline vm_page_t
pmap_remove_pt_page(pmap_t pmap,vm_offset_t va)1944 pmap_remove_pt_page(pmap_t pmap, vm_offset_t va)
1945 {
1946 
1947 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1948 	return (vm_radix_remove(&pmap->pm_root, va >> PDRSHIFT));
1949 }
1950 
1951 /*
1952  * Decrements a page table page's reference count, which is used to record the
1953  * number of valid page table entries within the page.  If the reference count
1954  * drops to zero, then the page table page is unmapped.  Returns true if the
1955  * page table page was unmapped and false otherwise.
1956  */
1957 static inline bool
pmap_unwire_ptp(pmap_t pmap,vm_page_t m,struct spglist * free)1958 pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free)
1959 {
1960 
1961 	--m->ref_count;
1962 	if (m->ref_count == 0) {
1963 		_pmap_unwire_ptp(pmap, m, free);
1964 		return (true);
1965 	} else
1966 		return (false);
1967 }
1968 
1969 static void
_pmap_unwire_ptp(pmap_t pmap,vm_page_t m,struct spglist * free)1970 _pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free)
1971 {
1972 
1973 	/*
1974 	 * unmap the page table page
1975 	 */
1976 	pmap->pm_pdir[m->pindex] = 0;
1977 	--pmap->pm_stats.resident_count;
1978 
1979 	/*
1980 	 * There is not need to invalidate the recursive mapping since
1981 	 * we never instantiate such mapping for the usermode pmaps,
1982 	 * and never remove page table pages from the kernel pmap.
1983 	 * Put page on a list so that it is released since all TLB
1984 	 * shootdown is done.
1985 	 */
1986 	MPASS(pmap != kernel_pmap);
1987 	pmap_add_delayed_free_list(m, free, true);
1988 }
1989 
1990 /*
1991  * After removing a page table entry, this routine is used to
1992  * conditionally free the page, and manage the reference count.
1993  */
1994 static int
pmap_unuse_pt(pmap_t pmap,vm_offset_t va,struct spglist * free)1995 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, struct spglist *free)
1996 {
1997 	pd_entry_t ptepde;
1998 	vm_page_t mpte;
1999 
2000 	if (pmap == kernel_pmap)
2001 		return (0);
2002 	ptepde = *pmap_pde(pmap, va);
2003 	mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME);
2004 	return (pmap_unwire_ptp(pmap, mpte, free));
2005 }
2006 
2007 /*
2008  * Release a page table page reference after a failed attempt to create a
2009  * mapping.
2010  */
2011 static void
pmap_abort_ptp(pmap_t pmap,vm_offset_t va,vm_page_t mpte)2012 pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
2013 {
2014 	struct spglist free;
2015 
2016 	SLIST_INIT(&free);
2017 	if (pmap_unwire_ptp(pmap, mpte, &free)) {
2018 		/*
2019 		 * Although "va" was never mapped, paging-structure caches
2020 		 * could nonetheless have entries that refer to the freed
2021 		 * page table pages.  Invalidate those entries.
2022 		 */
2023 		pmap_invalidate_page_int(pmap, va);
2024 		vm_page_free_pages_toq(&free, true);
2025 	}
2026 }
2027 
2028 /*
2029  * Initialize the pmap for the swapper process.
2030  */
2031 static void
__CONCAT(PMTYPE,pinit0)2032 __CONCAT(PMTYPE, pinit0)(pmap_t pmap)
2033 {
2034 
2035 	PMAP_LOCK_INIT(pmap);
2036 	pmap->pm_pdir = IdlePTD;
2037 #ifdef PMAP_PAE_COMP
2038 	pmap->pm_pdpt = IdlePDPT;
2039 #endif
2040 	vm_radix_init(&pmap->pm_root);
2041 	CPU_ZERO(&pmap->pm_active);
2042 	TAILQ_INIT(&pmap->pm_pvchunk);
2043 	bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
2044 	pmap_activate_boot(pmap);
2045 }
2046 
2047 /*
2048  * Initialize a preallocated and zeroed pmap structure,
2049  * such as one in a vmspace structure.
2050  */
2051 static int
__CONCAT(PMTYPE,pinit)2052 __CONCAT(PMTYPE, pinit)(pmap_t pmap)
2053 {
2054 	int i;
2055 
2056 	/*
2057 	 * No need to allocate page table space yet but we do need a valid
2058 	 * page directory table.
2059 	 */
2060 	if (pmap->pm_pdir == NULL) {
2061 		pmap->pm_pdir = (pd_entry_t *)kva_alloc(NBPTD);
2062 		if (pmap->pm_pdir == NULL)
2063 			return (0);
2064 #ifdef PMAP_PAE_COMP
2065 		pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO);
2066 		KASSERT(((vm_offset_t)pmap->pm_pdpt &
2067 		    ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0,
2068 		    ("pmap_pinit: pdpt misaligned"));
2069 		KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30),
2070 		    ("pmap_pinit: pdpt above 4g"));
2071 #endif
2072 		vm_radix_init(&pmap->pm_root);
2073 	}
2074 	KASSERT(vm_radix_is_empty(&pmap->pm_root),
2075 	    ("pmap_pinit: pmap has reserved page table page(s)"));
2076 
2077 	/*
2078 	 * allocate the page directory page(s)
2079 	 */
2080 	for (i = 0; i < NPGPTD; i++) {
2081 		pmap->pm_ptdpg[i] = vm_page_alloc_noobj(VM_ALLOC_WIRED |
2082 		    VM_ALLOC_ZERO | VM_ALLOC_WAITOK);
2083 #ifdef PMAP_PAE_COMP
2084 		pmap->pm_pdpt[i] = VM_PAGE_TO_PHYS(pmap->pm_ptdpg[i]) | PG_V;
2085 #endif
2086 	}
2087 
2088 	pmap_qenter((vm_offset_t)pmap->pm_pdir, pmap->pm_ptdpg, NPGPTD);
2089 #ifdef PMAP_PAE_COMP
2090 	if ((cpu_feature & CPUID_PAT) == 0) {
2091 		pmap_invalidate_cache_range(
2092 		    trunc_page((vm_offset_t)pmap->pm_pdpt),
2093 		    round_page((vm_offset_t)pmap->pm_pdpt +
2094 		    NPGPTD * sizeof(pdpt_entry_t)));
2095 	}
2096 #endif
2097 
2098 	/* Install the trampoline mapping. */
2099 	pmap->pm_pdir[TRPTDI] = PTD[TRPTDI];
2100 
2101 	CPU_ZERO(&pmap->pm_active);
2102 	TAILQ_INIT(&pmap->pm_pvchunk);
2103 	bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
2104 
2105 	return (1);
2106 }
2107 
2108 /*
2109  * this routine is called if the page table page is not
2110  * mapped correctly.
2111  */
2112 static vm_page_t
_pmap_allocpte(pmap_t pmap,u_int ptepindex,u_int flags)2113 _pmap_allocpte(pmap_t pmap, u_int ptepindex, u_int flags)
2114 {
2115 	vm_paddr_t ptepa;
2116 	vm_page_t m;
2117 
2118 	/*
2119 	 * Allocate a page table page.
2120 	 */
2121 	if ((m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
2122 		if ((flags & PMAP_ENTER_NOSLEEP) == 0) {
2123 			PMAP_UNLOCK(pmap);
2124 			rw_wunlock(&pvh_global_lock);
2125 			vm_wait(NULL);
2126 			rw_wlock(&pvh_global_lock);
2127 			PMAP_LOCK(pmap);
2128 		}
2129 
2130 		/*
2131 		 * Indicate the need to retry.  While waiting, the page table
2132 		 * page may have been allocated.
2133 		 */
2134 		return (NULL);
2135 	}
2136 	m->pindex = ptepindex;
2137 
2138 	/*
2139 	 * Map the pagetable page into the process address space, if
2140 	 * it isn't already there.
2141 	 */
2142 
2143 	pmap->pm_stats.resident_count++;
2144 
2145 	ptepa = VM_PAGE_TO_PHYS(m);
2146 	KASSERT((pmap->pm_pdir[ptepindex] & PG_V) == 0,
2147 	    ("%s: page directory entry %#jx is valid",
2148 	    __func__, (uintmax_t)pmap->pm_pdir[ptepindex]));
2149 	pmap->pm_pdir[ptepindex] =
2150 	    (pd_entry_t)(ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
2151 
2152 	return (m);
2153 }
2154 
2155 static vm_page_t
pmap_allocpte(pmap_t pmap,vm_offset_t va,u_int flags)2156 pmap_allocpte(pmap_t pmap, vm_offset_t va, u_int flags)
2157 {
2158 	u_int ptepindex;
2159 	pd_entry_t ptepa;
2160 	vm_page_t m;
2161 
2162 	/*
2163 	 * Calculate pagetable page index
2164 	 */
2165 	ptepindex = va >> PDRSHIFT;
2166 retry:
2167 	/*
2168 	 * Get the page directory entry
2169 	 */
2170 	ptepa = pmap->pm_pdir[ptepindex];
2171 
2172 	/*
2173 	 * This supports switching from a 4MB page to a
2174 	 * normal 4K page.
2175 	 */
2176 	if (ptepa & PG_PS) {
2177 		(void)pmap_demote_pde(pmap, &pmap->pm_pdir[ptepindex], va);
2178 		ptepa = pmap->pm_pdir[ptepindex];
2179 	}
2180 
2181 	/*
2182 	 * If the page table page is mapped, we just increment the
2183 	 * hold count, and activate it.
2184 	 */
2185 	if (ptepa) {
2186 		m = PHYS_TO_VM_PAGE(ptepa & PG_FRAME);
2187 		m->ref_count++;
2188 	} else {
2189 		/*
2190 		 * Here if the pte page isn't mapped, or if it has
2191 		 * been deallocated.
2192 		 */
2193 		m = _pmap_allocpte(pmap, ptepindex, flags);
2194 		if (m == NULL && (flags & PMAP_ENTER_NOSLEEP) == 0)
2195 			goto retry;
2196 	}
2197 	return (m);
2198 }
2199 
2200 /***************************************************
2201 * Pmap allocation/deallocation routines.
2202  ***************************************************/
2203 
2204 /*
2205  * Release any resources held by the given physical map.
2206  * Called when a pmap initialized by pmap_pinit is being released.
2207  * Should only be called if the map contains no valid mappings.
2208  */
2209 static void
__CONCAT(PMTYPE,release)2210 __CONCAT(PMTYPE, release)(pmap_t pmap)
2211 {
2212 	vm_page_t m;
2213 	int i;
2214 
2215 	KASSERT(pmap->pm_stats.resident_count == 0,
2216 	    ("pmap_release: pmap resident count %ld != 0",
2217 	    pmap->pm_stats.resident_count));
2218 	KASSERT(vm_radix_is_empty(&pmap->pm_root),
2219 	    ("pmap_release: pmap has reserved page table page(s)"));
2220 	KASSERT(CPU_EMPTY(&pmap->pm_active),
2221 	    ("releasing active pmap %p", pmap));
2222 
2223 	pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD);
2224 
2225 	for (i = 0; i < NPGPTD; i++) {
2226 		m = pmap->pm_ptdpg[i];
2227 #ifdef PMAP_PAE_COMP
2228 		KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME),
2229 		    ("pmap_release: got wrong ptd page"));
2230 #endif
2231 		vm_page_unwire_noq(m);
2232 		vm_page_free(m);
2233 	}
2234 }
2235 
2236 /*
2237  * grow the number of kernel page table entries, if needed
2238  */
2239 static int
__CONCAT(PMTYPE,growkernel)2240 __CONCAT(PMTYPE, growkernel)(vm_offset_t addr)
2241 {
2242 	vm_paddr_t ptppaddr;
2243 	vm_page_t nkpg;
2244 	pd_entry_t newpdir;
2245 
2246 	mtx_assert(&kernel_map->system_mtx, MA_OWNED);
2247 	addr = roundup2(addr, NBPDR);
2248 	if (addr - 1 >= vm_map_max(kernel_map))
2249 		addr = vm_map_max(kernel_map);
2250 	while (kernel_vm_end < addr) {
2251 		if (pdir_pde(PTD, kernel_vm_end)) {
2252 			kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK;
2253 			if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
2254 				kernel_vm_end = vm_map_max(kernel_map);
2255 				break;
2256 			}
2257 			continue;
2258 		}
2259 
2260 		nkpg = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT |
2261 		    VM_ALLOC_NOFREE | VM_ALLOC_WIRED | VM_ALLOC_ZERO);
2262 		if (nkpg == NULL)
2263 			return (KERN_RESOURCE_SHORTAGE);
2264 		nkpg->pindex = kernel_vm_end >> PDRSHIFT;
2265 		nkpt++;
2266 
2267 		ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2268 		newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2269 		pdir_pde(KPTD, kernel_vm_end) = newpdir;
2270 
2271 		pmap_kenter_pde(kernel_vm_end, newpdir);
2272 		kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK;
2273 		if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
2274 			kernel_vm_end = vm_map_max(kernel_map);
2275 			break;
2276 		}
2277 	}
2278 
2279 	return (KERN_SUCCESS);
2280 }
2281 
2282 /***************************************************
2283  * page management routines.
2284  ***************************************************/
2285 
2286 static const uint32_t pc_freemask[_NPCM] = {
2287 	[0 ... _NPCM - 2] = PC_FREEN,
2288 	[_NPCM - 1] = PC_FREEL
2289 };
2290 
2291 #ifdef PV_STATS
2292 extern int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail;
2293 extern long pv_entry_frees, pv_entry_allocs;
2294 extern int pv_entry_spare;
2295 #endif
2296 
2297 /*
2298  * We are in a serious low memory condition.  Resort to
2299  * drastic measures to free some pages so we can allocate
2300  * another pv entry chunk.
2301  */
2302 static vm_page_t
pmap_pv_reclaim(pmap_t locked_pmap)2303 pmap_pv_reclaim(pmap_t locked_pmap)
2304 {
2305 	struct pch newtail;
2306 	struct pv_chunk *pc;
2307 	struct md_page *pvh;
2308 	pd_entry_t *pde;
2309 	pmap_t pmap;
2310 	pt_entry_t *pte, tpte;
2311 	pv_entry_t pv;
2312 	vm_offset_t va;
2313 	vm_page_t m, m_pc;
2314 	struct spglist free;
2315 	uint32_t inuse;
2316 	int bit, field, freed;
2317 
2318 	PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED);
2319 	pmap = NULL;
2320 	m_pc = NULL;
2321 	SLIST_INIT(&free);
2322 	TAILQ_INIT(&newtail);
2323 	while ((pc = TAILQ_FIRST(&pv_chunks)) != NULL && (pv_vafree == 0 ||
2324 	    SLIST_EMPTY(&free))) {
2325 		TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
2326 		if (pmap != pc->pc_pmap) {
2327 			if (pmap != NULL) {
2328 				pmap_invalidate_all_int(pmap);
2329 				if (pmap != locked_pmap)
2330 					PMAP_UNLOCK(pmap);
2331 			}
2332 			pmap = pc->pc_pmap;
2333 			/* Avoid deadlock and lock recursion. */
2334 			if (pmap > locked_pmap)
2335 				PMAP_LOCK(pmap);
2336 			else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) {
2337 				pmap = NULL;
2338 				TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2339 				continue;
2340 			}
2341 		}
2342 
2343 		/*
2344 		 * Destroy every non-wired, 4 KB page mapping in the chunk.
2345 		 */
2346 		freed = 0;
2347 		for (field = 0; field < _NPCM; field++) {
2348 			for (inuse = ~pc->pc_map[field] & pc_freemask[field];
2349 			    inuse != 0; inuse &= ~(1UL << bit)) {
2350 				bit = bsfl(inuse);
2351 				pv = &pc->pc_pventry[field * 32 + bit];
2352 				va = pv->pv_va;
2353 				pde = pmap_pde(pmap, va);
2354 				if ((*pde & PG_PS) != 0)
2355 					continue;
2356 				pte = __CONCAT(PMTYPE, pte)(pmap, va);
2357 				tpte = *pte;
2358 				if ((tpte & PG_W) == 0)
2359 					tpte = pte_load_clear(pte);
2360 				pmap_pte_release(pte);
2361 				if ((tpte & PG_W) != 0)
2362 					continue;
2363 				KASSERT(tpte != 0,
2364 				    ("pmap_pv_reclaim: pmap %p va %x zero pte",
2365 				    pmap, va));
2366 				if ((tpte & PG_G) != 0)
2367 					pmap_invalidate_page_int(pmap, va);
2368 				m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
2369 				if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
2370 					vm_page_dirty(m);
2371 				if ((tpte & PG_A) != 0)
2372 					vm_page_aflag_set(m, PGA_REFERENCED);
2373 				TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
2374 				if (TAILQ_EMPTY(&m->md.pv_list) &&
2375 				    (m->flags & PG_FICTITIOUS) == 0) {
2376 					pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
2377 					if (TAILQ_EMPTY(&pvh->pv_list)) {
2378 						vm_page_aflag_clear(m,
2379 						    PGA_WRITEABLE);
2380 					}
2381 				}
2382 				pc->pc_map[field] |= 1UL << bit;
2383 				pmap_unuse_pt(pmap, va, &free);
2384 				freed++;
2385 			}
2386 		}
2387 		if (freed == 0) {
2388 			TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2389 			continue;
2390 		}
2391 		/* Every freed mapping is for a 4 KB page. */
2392 		pmap->pm_stats.resident_count -= freed;
2393 		PV_STAT(pv_entry_frees += freed);
2394 		PV_STAT(pv_entry_spare += freed);
2395 		pv_entry_count -= freed;
2396 		TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2397 		for (field = 0; field < _NPCM; field++)
2398 			if (pc->pc_map[field] != pc_freemask[field]) {
2399 				TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc,
2400 				    pc_list);
2401 				TAILQ_INSERT_TAIL(&newtail, pc, pc_lru);
2402 
2403 				/*
2404 				 * One freed pv entry in locked_pmap is
2405 				 * sufficient.
2406 				 */
2407 				if (pmap == locked_pmap)
2408 					goto out;
2409 				break;
2410 			}
2411 		if (field == _NPCM) {
2412 			PV_STAT(pv_entry_spare -= _NPCPV);
2413 			PV_STAT(pc_chunk_count--);
2414 			PV_STAT(pc_chunk_frees++);
2415 			/* Entire chunk is free; return it. */
2416 			m_pc = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc));
2417 			pmap_qremove((vm_offset_t)pc, 1);
2418 			pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc);
2419 			break;
2420 		}
2421 	}
2422 out:
2423 	TAILQ_CONCAT(&pv_chunks, &newtail, pc_lru);
2424 	if (pmap != NULL) {
2425 		pmap_invalidate_all_int(pmap);
2426 		if (pmap != locked_pmap)
2427 			PMAP_UNLOCK(pmap);
2428 	}
2429 	if (m_pc == NULL && pv_vafree != 0 && SLIST_EMPTY(&free)) {
2430 		m_pc = SLIST_FIRST(&free);
2431 		SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2432 		/* Recycle a freed page table page. */
2433 		m_pc->ref_count = 1;
2434 	}
2435 	vm_page_free_pages_toq(&free, true);
2436 	return (m_pc);
2437 }
2438 
2439 /*
2440  * free the pv_entry back to the free list
2441  */
2442 static void
free_pv_entry(pmap_t pmap,pv_entry_t pv)2443 free_pv_entry(pmap_t pmap, pv_entry_t pv)
2444 {
2445 	struct pv_chunk *pc;
2446 	int idx, field, bit;
2447 
2448 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2449 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2450 	PV_STAT(pv_entry_frees++);
2451 	PV_STAT(pv_entry_spare++);
2452 	pv_entry_count--;
2453 	pc = pv_to_chunk(pv);
2454 	idx = pv - &pc->pc_pventry[0];
2455 	field = idx / 32;
2456 	bit = idx % 32;
2457 	pc->pc_map[field] |= 1ul << bit;
2458 	for (idx = 0; idx < _NPCM; idx++)
2459 		if (pc->pc_map[idx] != pc_freemask[idx]) {
2460 			/*
2461 			 * 98% of the time, pc is already at the head of the
2462 			 * list.  If it isn't already, move it to the head.
2463 			 */
2464 			if (__predict_false(TAILQ_FIRST(&pmap->pm_pvchunk) !=
2465 			    pc)) {
2466 				TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2467 				TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc,
2468 				    pc_list);
2469 			}
2470 			return;
2471 		}
2472 	TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2473 	free_pv_chunk(pc);
2474 }
2475 
2476 static void
free_pv_chunk(struct pv_chunk * pc)2477 free_pv_chunk(struct pv_chunk *pc)
2478 {
2479 	vm_page_t m;
2480 
2481  	TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
2482 	PV_STAT(pv_entry_spare -= _NPCPV);
2483 	PV_STAT(pc_chunk_count--);
2484 	PV_STAT(pc_chunk_frees++);
2485 	/* entire chunk is free, return it */
2486 	m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc));
2487 	pmap_qremove((vm_offset_t)pc, 1);
2488 	vm_page_unwire_noq(m);
2489 	vm_page_free(m);
2490 	pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc);
2491 }
2492 
2493 /*
2494  * get a new pv_entry, allocating a block from the system
2495  * when needed.
2496  */
2497 static pv_entry_t
get_pv_entry(pmap_t pmap,bool try)2498 get_pv_entry(pmap_t pmap, bool try)
2499 {
2500 	static const struct timeval printinterval = { 60, 0 };
2501 	static struct timeval lastprint;
2502 	int bit, field;
2503 	pv_entry_t pv;
2504 	struct pv_chunk *pc;
2505 	vm_page_t m;
2506 
2507 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2508 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2509 	PV_STAT(pv_entry_allocs++);
2510 	pv_entry_count++;
2511 	if (pv_entry_count > pv_entry_high_water)
2512 		if (ratecheck(&lastprint, &printinterval))
2513 			printf("Approaching the limit on PV entries, consider "
2514 			    "increasing either the vm.pmap.shpgperproc or the "
2515 			    "vm.pmap.pv_entry_max tunable.\n");
2516 retry:
2517 	pc = TAILQ_FIRST(&pmap->pm_pvchunk);
2518 	if (pc != NULL) {
2519 		for (field = 0; field < _NPCM; field++) {
2520 			if (pc->pc_map[field]) {
2521 				bit = bsfl(pc->pc_map[field]);
2522 				break;
2523 			}
2524 		}
2525 		if (field < _NPCM) {
2526 			pv = &pc->pc_pventry[field * 32 + bit];
2527 			pc->pc_map[field] &= ~(1ul << bit);
2528 			/* If this was the last item, move it to tail */
2529 			for (field = 0; field < _NPCM; field++)
2530 				if (pc->pc_map[field] != 0) {
2531 					PV_STAT(pv_entry_spare--);
2532 					return (pv);	/* not full, return */
2533 				}
2534 			TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
2535 			TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
2536 			PV_STAT(pv_entry_spare--);
2537 			return (pv);
2538 		}
2539 	}
2540 	/*
2541 	 * Access to the ptelist "pv_vafree" is synchronized by the pvh
2542 	 * global lock.  If "pv_vafree" is currently non-empty, it will
2543 	 * remain non-empty until pmap_ptelist_alloc() completes.
2544 	 */
2545 	if (pv_vafree == 0 ||
2546 	    (m = vm_page_alloc_noobj(VM_ALLOC_WIRED)) == NULL) {
2547 		if (try) {
2548 			pv_entry_count--;
2549 			PV_STAT(pc_chunk_tryfail++);
2550 			return (NULL);
2551 		}
2552 		m = pmap_pv_reclaim(pmap);
2553 		if (m == NULL)
2554 			goto retry;
2555 	}
2556 	PV_STAT(pc_chunk_count++);
2557 	PV_STAT(pc_chunk_allocs++);
2558 	pc = (struct pv_chunk *)pmap_ptelist_alloc(&pv_vafree);
2559 	pmap_qenter((vm_offset_t)pc, &m, 1);
2560 	pc->pc_pmap = pmap;
2561 	pc->pc_map[0] = pc_freemask[0] & ~1ul;	/* preallocated bit 0 */
2562 	for (field = 1; field < _NPCM; field++)
2563 		pc->pc_map[field] = pc_freemask[field];
2564 	TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru);
2565 	pv = &pc->pc_pventry[0];
2566 	TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
2567 	PV_STAT(pv_entry_spare += _NPCPV - 1);
2568 	return (pv);
2569 }
2570 
2571 static __inline pv_entry_t
pmap_pvh_remove(struct md_page * pvh,pmap_t pmap,vm_offset_t va)2572 pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
2573 {
2574 	pv_entry_t pv;
2575 
2576 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2577 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
2578 		if (pmap == PV_PMAP(pv) && va == pv->pv_va) {
2579 			TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
2580 			break;
2581 		}
2582 	}
2583 	return (pv);
2584 }
2585 
2586 static void
pmap_pv_demote_pde(pmap_t pmap,vm_offset_t va,vm_paddr_t pa)2587 pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa)
2588 {
2589 	struct md_page *pvh;
2590 	pv_entry_t pv;
2591 	vm_offset_t va_last;
2592 	vm_page_t m;
2593 
2594 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2595 	KASSERT((pa & PDRMASK) == 0,
2596 	    ("pmap_pv_demote_pde: pa is not 4mpage aligned"));
2597 
2598 	/*
2599 	 * Transfer the 4mpage's pv entry for this mapping to the first
2600 	 * page's pv list.
2601 	 */
2602 	pvh = pa_to_pvh(pa);
2603 	va = trunc_4mpage(va);
2604 	pv = pmap_pvh_remove(pvh, pmap, va);
2605 	KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found"));
2606 	m = PHYS_TO_VM_PAGE(pa);
2607 	TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
2608 	/* Instantiate the remaining NPTEPG - 1 pv entries. */
2609 	va_last = va + NBPDR - PAGE_SIZE;
2610 	do {
2611 		m++;
2612 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2613 		    ("pmap_pv_demote_pde: page %p is not managed", m));
2614 		va += PAGE_SIZE;
2615 		pmap_insert_entry(pmap, va, m);
2616 	} while (va < va_last);
2617 }
2618 
2619 #if VM_NRESERVLEVEL > 0
2620 static void
pmap_pv_promote_pde(pmap_t pmap,vm_offset_t va,vm_paddr_t pa)2621 pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa)
2622 {
2623 	struct md_page *pvh;
2624 	pv_entry_t pv;
2625 	vm_offset_t va_last;
2626 	vm_page_t m;
2627 
2628 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2629 	KASSERT((pa & PDRMASK) == 0,
2630 	    ("pmap_pv_promote_pde: pa is not 4mpage aligned"));
2631 
2632 	/*
2633 	 * Transfer the first page's pv entry for this mapping to the
2634 	 * 4mpage's pv list.  Aside from avoiding the cost of a call
2635 	 * to get_pv_entry(), a transfer avoids the possibility that
2636 	 * get_pv_entry() calls pmap_collect() and that pmap_collect()
2637 	 * removes one of the mappings that is being promoted.
2638 	 */
2639 	m = PHYS_TO_VM_PAGE(pa);
2640 	va = trunc_4mpage(va);
2641 	pv = pmap_pvh_remove(&m->md, pmap, va);
2642 	KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found"));
2643 	pvh = pa_to_pvh(pa);
2644 	TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
2645 	/* Free the remaining NPTEPG - 1 pv entries. */
2646 	va_last = va + NBPDR - PAGE_SIZE;
2647 	do {
2648 		m++;
2649 		va += PAGE_SIZE;
2650 		pmap_pvh_free(&m->md, pmap, va);
2651 	} while (va < va_last);
2652 }
2653 #endif /* VM_NRESERVLEVEL > 0 */
2654 
2655 static void
pmap_pvh_free(struct md_page * pvh,pmap_t pmap,vm_offset_t va)2656 pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
2657 {
2658 	pv_entry_t pv;
2659 
2660 	pv = pmap_pvh_remove(pvh, pmap, va);
2661 	KASSERT(pv != NULL, ("pmap_pvh_free: pv not found"));
2662 	free_pv_entry(pmap, pv);
2663 }
2664 
2665 static void
pmap_remove_entry(pmap_t pmap,vm_page_t m,vm_offset_t va)2666 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
2667 {
2668 	struct md_page *pvh;
2669 
2670 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2671 	pmap_pvh_free(&m->md, pmap, va);
2672 	if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) {
2673 		pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
2674 		if (TAILQ_EMPTY(&pvh->pv_list))
2675 			vm_page_aflag_clear(m, PGA_WRITEABLE);
2676 	}
2677 }
2678 
2679 /*
2680  * Create a pv entry for page at pa for
2681  * (pmap, va).
2682  */
2683 static void
pmap_insert_entry(pmap_t pmap,vm_offset_t va,vm_page_t m)2684 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
2685 {
2686 	pv_entry_t pv;
2687 
2688 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2689 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2690 	pv = get_pv_entry(pmap, false);
2691 	pv->pv_va = va;
2692 	TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
2693 }
2694 
2695 /*
2696  * Conditionally create a pv entry.
2697  */
2698 static bool
pmap_try_insert_pv_entry(pmap_t pmap,vm_offset_t va,vm_page_t m)2699 pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
2700 {
2701 	pv_entry_t pv;
2702 
2703 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2704 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2705 	if (pv_entry_count < pv_entry_high_water &&
2706 	    (pv = get_pv_entry(pmap, true)) != NULL) {
2707 		pv->pv_va = va;
2708 		TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
2709 		return (true);
2710 	} else
2711 		return (false);
2712 }
2713 
2714 /*
2715  * Create the pv entries for each of the pages within a superpage.
2716  */
2717 static bool
pmap_pv_insert_pde(pmap_t pmap,vm_offset_t va,pd_entry_t pde,u_int flags)2718 pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags)
2719 {
2720 	struct md_page *pvh;
2721 	pv_entry_t pv;
2722 	bool noreclaim;
2723 
2724 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2725 	noreclaim = (flags & PMAP_ENTER_NORECLAIM) != 0;
2726 	if ((noreclaim && pv_entry_count >= pv_entry_high_water) ||
2727 	    (pv = get_pv_entry(pmap, noreclaim)) == NULL)
2728 		return (false);
2729 	pv->pv_va = va;
2730 	pvh = pa_to_pvh(pde & PG_PS_FRAME);
2731 	TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
2732 	return (true);
2733 }
2734 
2735 /*
2736  * Fills a page table page with mappings to consecutive physical pages.
2737  */
2738 static void
pmap_fill_ptp(pt_entry_t * firstpte,pt_entry_t newpte)2739 pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte)
2740 {
2741 	pt_entry_t *pte;
2742 
2743 	for (pte = firstpte; pte < firstpte + NPTEPG; pte++) {
2744 		*pte = newpte;
2745 		newpte += PAGE_SIZE;
2746 	}
2747 }
2748 
2749 /*
2750  * Tries to demote a 2- or 4MB page mapping.  If demotion fails, the
2751  * 2- or 4MB page mapping is invalidated.
2752  */
2753 static bool
pmap_demote_pde(pmap_t pmap,pd_entry_t * pde,vm_offset_t va)2754 pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
2755 {
2756 	pd_entry_t newpde, oldpde;
2757 	pt_entry_t *firstpte, newpte;
2758 	vm_paddr_t mptepa;
2759 	vm_page_t mpte;
2760 	struct spglist free;
2761 	vm_offset_t sva;
2762 
2763 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2764 	oldpde = *pde;
2765 	KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V),
2766 	    ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V"));
2767 	if ((oldpde & PG_A) == 0 || (mpte = pmap_remove_pt_page(pmap, va)) ==
2768 	    NULL) {
2769 		KASSERT((oldpde & PG_W) == 0,
2770 		    ("pmap_demote_pde: page table page for a wired mapping"
2771 		    " is missing"));
2772 
2773 		/*
2774 		 * Invalidate the 2- or 4MB page mapping and return
2775 		 * "failure" if the mapping was never accessed or the
2776 		 * allocation of the new page table page fails.
2777 		 */
2778 		if ((oldpde & PG_A) == 0 ||
2779 		    (mpte = vm_page_alloc_noobj(VM_ALLOC_WIRED)) == NULL) {
2780 			SLIST_INIT(&free);
2781 			sva = trunc_4mpage(va);
2782 			pmap_remove_pde(pmap, pde, sva, &free);
2783 			if ((oldpde & PG_G) == 0)
2784 				pmap_invalidate_pde_page(pmap, sva, oldpde);
2785 			vm_page_free_pages_toq(&free, true);
2786 			CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#x"
2787 			    " in pmap %p", va, pmap);
2788 			return (false);
2789 		}
2790 		mpte->pindex = va >> PDRSHIFT;
2791 		if (pmap != kernel_pmap) {
2792 			mpte->ref_count = NPTEPG;
2793 			pmap->pm_stats.resident_count++;
2794 		}
2795 	}
2796 	mptepa = VM_PAGE_TO_PHYS(mpte);
2797 
2798 	/*
2799 	 * If the page mapping is in the kernel's address space, then the
2800 	 * KPTmap can provide access to the page table page.  Otherwise,
2801 	 * temporarily map the page table page (mpte) into the kernel's
2802 	 * address space at either PADDR1 or PADDR2.
2803 	 */
2804 	if (pmap == kernel_pmap)
2805 		firstpte = &KPTmap[i386_btop(trunc_4mpage(va))];
2806 	else if (curthread->td_pinned > 0 && rw_wowned(&pvh_global_lock)) {
2807 		if ((*PMAP1 & PG_FRAME) != mptepa) {
2808 			*PMAP1 = mptepa | PG_RW | PG_V | PG_A | PG_M;
2809 #ifdef SMP
2810 			PMAP1cpu = PCPU_GET(cpuid);
2811 #endif
2812 			invlcaddr(PADDR1);
2813 			PMAP1changed++;
2814 		} else
2815 #ifdef SMP
2816 		if (PMAP1cpu != PCPU_GET(cpuid)) {
2817 			PMAP1cpu = PCPU_GET(cpuid);
2818 			invlcaddr(PADDR1);
2819 			PMAP1changedcpu++;
2820 		} else
2821 #endif
2822 			PMAP1unchanged++;
2823 		firstpte = PADDR1;
2824 	} else {
2825 		mtx_lock(&PMAP2mutex);
2826 		if ((*PMAP2 & PG_FRAME) != mptepa) {
2827 			*PMAP2 = mptepa | PG_RW | PG_V | PG_A | PG_M;
2828 			pmap_invalidate_page_int(kernel_pmap,
2829 			    (vm_offset_t)PADDR2);
2830 		}
2831 		firstpte = PADDR2;
2832 	}
2833 	newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V;
2834 	KASSERT((oldpde & PG_A) != 0,
2835 	    ("pmap_demote_pde: oldpde is missing PG_A"));
2836 	KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW,
2837 	    ("pmap_demote_pde: oldpde is missing PG_M"));
2838 	newpte = oldpde & ~PG_PS;
2839 	if ((newpte & PG_PDE_PAT) != 0)
2840 		newpte ^= PG_PDE_PAT | PG_PTE_PAT;
2841 
2842 	/*
2843 	 * If the PTP is not leftover from an earlier promotion or it does not
2844 	 * have PG_A set in every PTE, then fill it.  The new PTEs will all
2845 	 * have PG_A set.
2846 	 */
2847 	if (!vm_page_all_valid(mpte))
2848 		pmap_fill_ptp(firstpte, newpte);
2849 
2850 	KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME),
2851 	    ("pmap_demote_pde: firstpte and newpte map different physical"
2852 	    " addresses"));
2853 
2854 	/*
2855 	 * If the mapping has changed attributes, update the PTEs.
2856 	 */
2857 	if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE))
2858 		pmap_fill_ptp(firstpte, newpte);
2859 
2860 	/*
2861 	 * Demote the mapping.  This pmap is locked.  The old PDE has
2862 	 * PG_A set.  If the old PDE has PG_RW set, it also has PG_M
2863 	 * set.  Thus, there is no danger of a race with another
2864 	 * processor changing the setting of PG_A and/or PG_M between
2865 	 * the read above and the store below.
2866 	 */
2867 	if (workaround_erratum383)
2868 		pmap_update_pde(pmap, va, pde, newpde);
2869 	else if (pmap == kernel_pmap)
2870 		pmap_kenter_pde(va, newpde);
2871 	else
2872 		pde_store(pde, newpde);
2873 	if (firstpte == PADDR2)
2874 		mtx_unlock(&PMAP2mutex);
2875 
2876 	/*
2877 	 * Invalidate the recursive mapping of the page table page.
2878 	 */
2879 	pmap_invalidate_page_int(pmap, (vm_offset_t)vtopte(va));
2880 
2881 	/*
2882 	 * Demote the pv entry.  This depends on the earlier demotion
2883 	 * of the mapping.  Specifically, the (re)creation of a per-
2884 	 * page pv entry might trigger the execution of pmap_collect(),
2885 	 * which might reclaim a newly (re)created per-page pv entry
2886 	 * and destroy the associated mapping.  In order to destroy
2887 	 * the mapping, the PDE must have already changed from mapping
2888 	 * the 2mpage to referencing the page table page.
2889 	 */
2890 	if ((oldpde & PG_MANAGED) != 0)
2891 		pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME);
2892 
2893 	pmap_pde_demotions++;
2894 	CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#x"
2895 	    " in pmap %p", va, pmap);
2896 	return (true);
2897 }
2898 
2899 /*
2900  * Removes a 2- or 4MB page mapping from the kernel pmap.
2901  */
2902 static void
pmap_remove_kernel_pde(pmap_t pmap,pd_entry_t * pde,vm_offset_t va)2903 pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va)
2904 {
2905 	pd_entry_t newpde;
2906 	vm_paddr_t mptepa;
2907 	vm_page_t mpte;
2908 
2909 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2910 	mpte = pmap_remove_pt_page(pmap, va);
2911 	if (mpte == NULL)
2912 		panic("pmap_remove_kernel_pde: Missing pt page.");
2913 
2914 	mptepa = VM_PAGE_TO_PHYS(mpte);
2915 	newpde = mptepa | PG_M | PG_A | PG_RW | PG_V;
2916 
2917 	/*
2918 	 * If this page table page was unmapped by a promotion, then it
2919 	 * contains valid mappings.  Zero it to invalidate those mappings.
2920 	 */
2921 	if (vm_page_any_valid(mpte))
2922 		pagezero((void *)&KPTmap[i386_btop(trunc_4mpage(va))]);
2923 
2924 	/*
2925 	 * Remove the mapping.
2926 	 */
2927 	if (workaround_erratum383)
2928 		pmap_update_pde(pmap, va, pde, newpde);
2929 	else
2930 		pmap_kenter_pde(va, newpde);
2931 
2932 	/*
2933 	 * Invalidate the recursive mapping of the page table page.
2934 	 */
2935 	pmap_invalidate_page_int(pmap, (vm_offset_t)vtopte(va));
2936 }
2937 
2938 /*
2939  * pmap_remove_pde: do the things to unmap a superpage in a process
2940  */
2941 static void
pmap_remove_pde(pmap_t pmap,pd_entry_t * pdq,vm_offset_t sva,struct spglist * free)2942 pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva,
2943     struct spglist *free)
2944 {
2945 	struct md_page *pvh;
2946 	pd_entry_t oldpde;
2947 	vm_offset_t eva, va;
2948 	vm_page_t m, mpte;
2949 
2950 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2951 	KASSERT((sva & PDRMASK) == 0,
2952 	    ("pmap_remove_pde: sva is not 4mpage aligned"));
2953 	oldpde = pte_load_clear(pdq);
2954 	if (oldpde & PG_W)
2955 		pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE;
2956 
2957 	/*
2958 	 * Machines that don't support invlpg, also don't support
2959 	 * PG_G.
2960 	 */
2961 	if ((oldpde & PG_G) != 0)
2962 		pmap_invalidate_pde_page(kernel_pmap, sva, oldpde);
2963 
2964 	pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2965 	if (oldpde & PG_MANAGED) {
2966 		pvh = pa_to_pvh(oldpde & PG_PS_FRAME);
2967 		pmap_pvh_free(pvh, pmap, sva);
2968 		eva = sva + NBPDR;
2969 		for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
2970 		    va < eva; va += PAGE_SIZE, m++) {
2971 			if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW))
2972 				vm_page_dirty(m);
2973 			if (oldpde & PG_A)
2974 				vm_page_aflag_set(m, PGA_REFERENCED);
2975 			if (TAILQ_EMPTY(&m->md.pv_list) &&
2976 			    TAILQ_EMPTY(&pvh->pv_list))
2977 				vm_page_aflag_clear(m, PGA_WRITEABLE);
2978 		}
2979 	}
2980 	if (pmap == kernel_pmap) {
2981 		pmap_remove_kernel_pde(pmap, pdq, sva);
2982 	} else {
2983 		mpte = pmap_remove_pt_page(pmap, sva);
2984 		if (mpte != NULL) {
2985 			KASSERT(vm_page_any_valid(mpte),
2986 			    ("pmap_remove_pde: pte page not promoted"));
2987 			pmap->pm_stats.resident_count--;
2988 			KASSERT(mpte->ref_count == NPTEPG,
2989 			    ("pmap_remove_pde: pte page ref count error"));
2990 			mpte->ref_count = 0;
2991 			pmap_add_delayed_free_list(mpte, free, false);
2992 		}
2993 	}
2994 }
2995 
2996 /*
2997  * pmap_remove_pte: do the things to unmap a page in a process
2998  */
2999 static int
pmap_remove_pte(pmap_t pmap,pt_entry_t * ptq,vm_offset_t va,struct spglist * free)3000 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va,
3001     struct spglist *free)
3002 {
3003 	pt_entry_t oldpte;
3004 	vm_page_t m;
3005 
3006 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3007 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3008 	oldpte = pte_load_clear(ptq);
3009 	KASSERT(oldpte != 0,
3010 	    ("pmap_remove_pte: pmap %p va %x zero pte", pmap, va));
3011 	if (oldpte & PG_W)
3012 		pmap->pm_stats.wired_count -= 1;
3013 	/*
3014 	 * Machines that don't support invlpg, also don't support
3015 	 * PG_G.
3016 	 */
3017 	if (oldpte & PG_G)
3018 		pmap_invalidate_page_int(kernel_pmap, va);
3019 	pmap->pm_stats.resident_count -= 1;
3020 	if (oldpte & PG_MANAGED) {
3021 		m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME);
3022 		if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
3023 			vm_page_dirty(m);
3024 		if (oldpte & PG_A)
3025 			vm_page_aflag_set(m, PGA_REFERENCED);
3026 		pmap_remove_entry(pmap, m, va);
3027 	}
3028 	return (pmap_unuse_pt(pmap, va, free));
3029 }
3030 
3031 /*
3032  * Remove a single page from a process address space
3033  */
3034 static void
pmap_remove_page(pmap_t pmap,vm_offset_t va,struct spglist * free)3035 pmap_remove_page(pmap_t pmap, vm_offset_t va, struct spglist *free)
3036 {
3037 	pt_entry_t *pte;
3038 
3039 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3040 	KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
3041 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3042 	if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0)
3043 		return;
3044 	pmap_remove_pte(pmap, pte, va, free);
3045 	pmap_invalidate_page_int(pmap, va);
3046 }
3047 
3048 /*
3049  * Removes the specified range of addresses from the page table page.
3050  */
3051 static bool
pmap_remove_ptes(pmap_t pmap,vm_offset_t sva,vm_offset_t eva,struct spglist * free)3052 pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
3053     struct spglist *free)
3054 {
3055 	pt_entry_t *pte;
3056 	bool anyvalid;
3057 
3058 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3059 	KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
3060 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3061 	anyvalid = false;
3062 	for (pte = pmap_pte_quick(pmap, sva); sva != eva; pte++,
3063 	    sva += PAGE_SIZE) {
3064 		if (*pte == 0)
3065 			continue;
3066 
3067 		/*
3068 		 * The TLB entry for a PG_G mapping is invalidated by
3069 		 * pmap_remove_pte().
3070 		 */
3071 		if ((*pte & PG_G) == 0)
3072 			anyvalid = true;
3073 
3074 		if (pmap_remove_pte(pmap, pte, sva, free))
3075 			break;
3076 	}
3077 	return (anyvalid);
3078 }
3079 
3080 /*
3081  *	Remove the given range of addresses from the specified map.
3082  *
3083  *	It is assumed that the start and end are properly
3084  *	rounded to the page size.
3085  */
3086 static void
__CONCAT(PMTYPE,remove)3087 __CONCAT(PMTYPE, remove)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3088 {
3089 	vm_offset_t pdnxt;
3090 	pd_entry_t ptpaddr;
3091 	struct spglist free;
3092 	int anyvalid;
3093 
3094 	/*
3095 	 * Perform an unsynchronized read.  This is, however, safe.
3096 	 */
3097 	if (pmap->pm_stats.resident_count == 0)
3098 		return;
3099 
3100 	anyvalid = 0;
3101 	SLIST_INIT(&free);
3102 
3103 	rw_wlock(&pvh_global_lock);
3104 	sched_pin();
3105 	PMAP_LOCK(pmap);
3106 
3107 	/*
3108 	 * special handling of removing one page.  a very
3109 	 * common operation and easy to short circuit some
3110 	 * code.
3111 	 */
3112 	if ((sva + PAGE_SIZE == eva) &&
3113 	    ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
3114 		pmap_remove_page(pmap, sva, &free);
3115 		goto out;
3116 	}
3117 
3118 	for (; sva < eva; sva = pdnxt) {
3119 		u_int pdirindex;
3120 
3121 		/*
3122 		 * Calculate index for next page table.
3123 		 */
3124 		pdnxt = (sva + NBPDR) & ~PDRMASK;
3125 		if (pdnxt < sva)
3126 			pdnxt = eva;
3127 		if (pmap->pm_stats.resident_count == 0)
3128 			break;
3129 
3130 		pdirindex = sva >> PDRSHIFT;
3131 		ptpaddr = pmap->pm_pdir[pdirindex];
3132 
3133 		/*
3134 		 * Weed out invalid mappings. Note: we assume that the page
3135 		 * directory table is always allocated, and in kernel virtual.
3136 		 */
3137 		if (ptpaddr == 0)
3138 			continue;
3139 
3140 		/*
3141 		 * Check for large page.
3142 		 */
3143 		if ((ptpaddr & PG_PS) != 0) {
3144 			/*
3145 			 * Are we removing the entire large page?  If not,
3146 			 * demote the mapping and fall through.
3147 			 */
3148 			if (sva + NBPDR == pdnxt && eva >= pdnxt) {
3149 				/*
3150 				 * The TLB entry for a PG_G mapping is
3151 				 * invalidated by pmap_remove_pde().
3152 				 */
3153 				if ((ptpaddr & PG_G) == 0)
3154 					anyvalid = 1;
3155 				pmap_remove_pde(pmap,
3156 				    &pmap->pm_pdir[pdirindex], sva, &free);
3157 				continue;
3158 			} else if (!pmap_demote_pde(pmap,
3159 			    &pmap->pm_pdir[pdirindex], sva)) {
3160 				/* The large page mapping was destroyed. */
3161 				continue;
3162 			}
3163 		}
3164 
3165 		/*
3166 		 * Limit our scan to either the end of the va represented
3167 		 * by the current page table page, or to the end of the
3168 		 * range being removed.
3169 		 */
3170 		if (pdnxt > eva)
3171 			pdnxt = eva;
3172 
3173 		if (pmap_remove_ptes(pmap, sva, pdnxt, &free))
3174 			anyvalid = 1;
3175 	}
3176 out:
3177 	sched_unpin();
3178 	if (anyvalid)
3179 		pmap_invalidate_all_int(pmap);
3180 	rw_wunlock(&pvh_global_lock);
3181 	PMAP_UNLOCK(pmap);
3182 	vm_page_free_pages_toq(&free, true);
3183 }
3184 
3185 /*
3186  *	Routine:	pmap_remove_all
3187  *	Function:
3188  *		Removes this physical page from
3189  *		all physical maps in which it resides.
3190  *		Reflects back modify bits to the pager.
3191  *
3192  *	Notes:
3193  *		Original versions of this routine were very
3194  *		inefficient because they iteratively called
3195  *		pmap_remove (slow...)
3196  */
3197 
3198 static void
__CONCAT(PMTYPE,remove_all)3199 __CONCAT(PMTYPE, remove_all)(vm_page_t m)
3200 {
3201 	struct md_page *pvh;
3202 	pv_entry_t pv;
3203 	pmap_t pmap;
3204 	pt_entry_t *pte, tpte;
3205 	pd_entry_t *pde;
3206 	vm_offset_t va;
3207 	struct spglist free;
3208 
3209 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3210 	    ("pmap_remove_all: page %p is not managed", m));
3211 	SLIST_INIT(&free);
3212 	rw_wlock(&pvh_global_lock);
3213 	sched_pin();
3214 	if ((m->flags & PG_FICTITIOUS) != 0)
3215 		goto small_mappings;
3216 	pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
3217 	while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) {
3218 		va = pv->pv_va;
3219 		pmap = PV_PMAP(pv);
3220 		PMAP_LOCK(pmap);
3221 		pde = pmap_pde(pmap, va);
3222 		(void)pmap_demote_pde(pmap, pde, va);
3223 		PMAP_UNLOCK(pmap);
3224 	}
3225 small_mappings:
3226 	while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3227 		pmap = PV_PMAP(pv);
3228 		PMAP_LOCK(pmap);
3229 		pmap->pm_stats.resident_count--;
3230 		pde = pmap_pde(pmap, pv->pv_va);
3231 		KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found"
3232 		    " a 4mpage in page %p's pv list", m));
3233 		pte = pmap_pte_quick(pmap, pv->pv_va);
3234 		tpte = pte_load_clear(pte);
3235 		KASSERT(tpte != 0, ("pmap_remove_all: pmap %p va %x zero pte",
3236 		    pmap, pv->pv_va));
3237 		if (tpte & PG_W)
3238 			pmap->pm_stats.wired_count--;
3239 		if (tpte & PG_A)
3240 			vm_page_aflag_set(m, PGA_REFERENCED);
3241 
3242 		/*
3243 		 * Update the vm_page_t clean and reference bits.
3244 		 */
3245 		if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
3246 			vm_page_dirty(m);
3247 		pmap_unuse_pt(pmap, pv->pv_va, &free);
3248 		pmap_invalidate_page_int(pmap, pv->pv_va);
3249 		TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
3250 		free_pv_entry(pmap, pv);
3251 		PMAP_UNLOCK(pmap);
3252 	}
3253 	vm_page_aflag_clear(m, PGA_WRITEABLE);
3254 	sched_unpin();
3255 	rw_wunlock(&pvh_global_lock);
3256 	vm_page_free_pages_toq(&free, true);
3257 }
3258 
3259 /*
3260  * pmap_protect_pde: do the things to protect a 4mpage in a process
3261  */
3262 static bool
pmap_protect_pde(pmap_t pmap,pd_entry_t * pde,vm_offset_t sva,vm_prot_t prot)3263 pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot)
3264 {
3265 	pd_entry_t newpde, oldpde;
3266 	vm_page_t m, mt;
3267 	bool anychanged;
3268 
3269 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3270 	KASSERT((sva & PDRMASK) == 0,
3271 	    ("pmap_protect_pde: sva is not 4mpage aligned"));
3272 	anychanged = false;
3273 retry:
3274 	oldpde = newpde = *pde;
3275 	if ((prot & VM_PROT_WRITE) == 0) {
3276 		if ((oldpde & (PG_MANAGED | PG_M | PG_RW)) ==
3277 		    (PG_MANAGED | PG_M | PG_RW)) {
3278 			m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
3279 			for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
3280 				vm_page_dirty(mt);
3281 		}
3282 		newpde &= ~(PG_RW | PG_M);
3283 	}
3284 #ifdef PMAP_PAE_COMP
3285 	if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec)
3286 		newpde |= pg_nx;
3287 #endif
3288 	if (newpde != oldpde) {
3289 		/*
3290 		 * As an optimization to future operations on this PDE, clear
3291 		 * PG_PROMOTED.  The impending invalidation will remove any
3292 		 * lingering 4KB page mappings from the TLB.
3293 		 */
3294 		if (!pde_cmpset(pde, oldpde, newpde & ~PG_PROMOTED))
3295 			goto retry;
3296 		if ((oldpde & PG_G) != 0)
3297 			pmap_invalidate_pde_page(kernel_pmap, sva, oldpde);
3298 		else
3299 			anychanged = true;
3300 	}
3301 	return (anychanged);
3302 }
3303 
3304 /*
3305  *	Set the physical protection on the
3306  *	specified range of this map as requested.
3307  */
3308 static void
__CONCAT(PMTYPE,protect)3309 __CONCAT(PMTYPE, protect)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
3310     vm_prot_t prot)
3311 {
3312 	vm_offset_t pdnxt;
3313 	pd_entry_t ptpaddr;
3314 	pt_entry_t *pte;
3315 	bool anychanged, pv_lists_locked;
3316 
3317 	KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot));
3318 	if (prot == VM_PROT_NONE) {
3319 		pmap_remove(pmap, sva, eva);
3320 		return;
3321 	}
3322 
3323 #ifdef PMAP_PAE_COMP
3324 	if ((prot & (VM_PROT_WRITE | VM_PROT_EXECUTE)) ==
3325 	    (VM_PROT_WRITE | VM_PROT_EXECUTE))
3326 		return;
3327 #else
3328 	if (prot & VM_PROT_WRITE)
3329 		return;
3330 #endif
3331 
3332 	if (pmap_is_current(pmap))
3333 		pv_lists_locked = false;
3334 	else {
3335 		pv_lists_locked = true;
3336 resume:
3337 		rw_wlock(&pvh_global_lock);
3338 		sched_pin();
3339 	}
3340 	anychanged = false;
3341 
3342 	PMAP_LOCK(pmap);
3343 	for (; sva < eva; sva = pdnxt) {
3344 		pt_entry_t obits, pbits;
3345 		u_int pdirindex;
3346 
3347 		pdnxt = (sva + NBPDR) & ~PDRMASK;
3348 		if (pdnxt < sva)
3349 			pdnxt = eva;
3350 
3351 		pdirindex = sva >> PDRSHIFT;
3352 		ptpaddr = pmap->pm_pdir[pdirindex];
3353 
3354 		/*
3355 		 * Weed out invalid mappings. Note: we assume that the page
3356 		 * directory table is always allocated, and in kernel virtual.
3357 		 */
3358 		if (ptpaddr == 0)
3359 			continue;
3360 
3361 		/*
3362 		 * Check for large page.
3363 		 */
3364 		if ((ptpaddr & PG_PS) != 0) {
3365 			/*
3366 			 * Are we protecting the entire large page?  If not,
3367 			 * demote the mapping and fall through.
3368 			 */
3369 			if (sva + NBPDR == pdnxt && eva >= pdnxt) {
3370 				/*
3371 				 * The TLB entry for a PG_G mapping is
3372 				 * invalidated by pmap_protect_pde().
3373 				 */
3374 				if (pmap_protect_pde(pmap,
3375 				    &pmap->pm_pdir[pdirindex], sva, prot))
3376 					anychanged = true;
3377 				continue;
3378 			} else {
3379 				if (!pv_lists_locked) {
3380 					pv_lists_locked = true;
3381 					if (!rw_try_wlock(&pvh_global_lock)) {
3382 						if (anychanged)
3383 							pmap_invalidate_all_int(
3384 							    pmap);
3385 						PMAP_UNLOCK(pmap);
3386 						goto resume;
3387 					}
3388 					sched_pin();
3389 				}
3390 				if (!pmap_demote_pde(pmap,
3391 				    &pmap->pm_pdir[pdirindex], sva)) {
3392 					/*
3393 					 * The large page mapping was
3394 					 * destroyed.
3395 					 */
3396 					continue;
3397 				}
3398 			}
3399 		}
3400 
3401 		if (pdnxt > eva)
3402 			pdnxt = eva;
3403 
3404 		for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
3405 		    sva += PAGE_SIZE) {
3406 			vm_page_t m;
3407 
3408 retry:
3409 			/*
3410 			 * Regardless of whether a pte is 32 or 64 bits in
3411 			 * size, PG_RW, PG_A, and PG_M are among the least
3412 			 * significant 32 bits.
3413 			 */
3414 			obits = pbits = *pte;
3415 			if ((pbits & PG_V) == 0)
3416 				continue;
3417 
3418 			if ((prot & VM_PROT_WRITE) == 0) {
3419 				if ((pbits & (PG_MANAGED | PG_M | PG_RW)) ==
3420 				    (PG_MANAGED | PG_M | PG_RW)) {
3421 					m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3422 					vm_page_dirty(m);
3423 				}
3424 				pbits &= ~(PG_RW | PG_M);
3425 			}
3426 #ifdef PMAP_PAE_COMP
3427 			if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec)
3428 				pbits |= pg_nx;
3429 #endif
3430 
3431 			if (pbits != obits) {
3432 #ifdef PMAP_PAE_COMP
3433 				if (!atomic_cmpset_64(pte, obits, pbits))
3434 					goto retry;
3435 #else
3436 				if (!atomic_cmpset_int((u_int *)pte, obits,
3437 				    pbits))
3438 					goto retry;
3439 #endif
3440 				if (obits & PG_G)
3441 					pmap_invalidate_page_int(pmap, sva);
3442 				else
3443 					anychanged = true;
3444 			}
3445 		}
3446 	}
3447 	if (anychanged)
3448 		pmap_invalidate_all_int(pmap);
3449 	if (pv_lists_locked) {
3450 		sched_unpin();
3451 		rw_wunlock(&pvh_global_lock);
3452 	}
3453 	PMAP_UNLOCK(pmap);
3454 }
3455 
3456 #if VM_NRESERVLEVEL > 0
3457 /*
3458  * Tries to promote the 512 or 1024, contiguous 4KB page mappings that are
3459  * within a single page table page (PTP) to a single 2- or 4MB page mapping.
3460  * For promotion to occur, two conditions must be met: (1) the 4KB page
3461  * mappings must map aligned, contiguous physical memory and (2) the 4KB page
3462  * mappings must have identical characteristics.
3463  *
3464  * Managed (PG_MANAGED) mappings within the kernel address space are not
3465  * promoted.  The reason is that kernel PDEs are replicated in each pmap but
3466  * pmap_clear_ptes() and pmap_ts_referenced() only read the PDE from the kernel
3467  * pmap.
3468  */
3469 static bool
pmap_promote_pde(pmap_t pmap,pd_entry_t * pde,vm_offset_t va,vm_page_t mpte)3470 pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, vm_page_t mpte)
3471 {
3472 	pd_entry_t newpde;
3473 	pt_entry_t allpte_PG_A, *firstpte, oldpte, pa, *pte;
3474 #ifdef KTR
3475 	vm_offset_t oldpteva;
3476 #endif
3477 
3478 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3479 	if (!pg_ps_enabled)
3480 		return (false);
3481 
3482 	/*
3483 	 * Examine the first PTE in the specified PTP.  Abort if this PTE is
3484 	 * either invalid or does not map the first 4KB physical page
3485 	 * within a 2- or 4MB page.
3486 	 */
3487 	firstpte = pmap_pte_quick(pmap, trunc_4mpage(va));
3488 setpde:
3489 	newpde = *firstpte;
3490 	if ((newpde & ((PG_FRAME & PDRMASK) | PG_V)) != PG_V) {
3491 		pmap_pde_p_failures++;
3492 		CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x"
3493 		    " in pmap %p", va, pmap);
3494 		return (false);
3495 	}
3496 	if ((*firstpte & PG_MANAGED) != 0 && pmap == kernel_pmap) {
3497 		pmap_pde_p_failures++;
3498 		CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x"
3499 		    " in pmap %p", va, pmap);
3500 		return (false);
3501 	}
3502 
3503 	/*
3504 	 * Both here and in the below "for" loop, to allow for repromotion
3505 	 * after MADV_FREE, conditionally write protect a clean PTE before
3506 	 * possibly aborting the promotion due to other PTE attributes.  Why?
3507 	 * Suppose that MADV_FREE is applied to a part of a superpage, the
3508 	 * address range [S, E).  pmap_advise() will demote the superpage
3509 	 * mapping, destroy the 4KB page mapping at the end of [S, E), and
3510 	 * clear PG_M and PG_A in the PTEs for the rest of [S, E).  Later,
3511 	 * imagine that the memory in [S, E) is recycled, but the last 4KB
3512 	 * page in [S, E) is not the last to be rewritten, or simply accessed.
3513 	 * In other words, there is still a 4KB page in [S, E), call it P,
3514 	 * that is writeable but PG_M and PG_A are clear in P's PTE.  Unless
3515 	 * we write protect P before aborting the promotion, if and when P is
3516 	 * finally rewritten, there won't be a page fault to trigger
3517 	 * repromotion.
3518 	 */
3519 	if ((newpde & (PG_M | PG_RW)) == PG_RW) {
3520 		/*
3521 		 * When PG_M is already clear, PG_RW can be cleared without
3522 		 * a TLB invalidation.
3523 		 */
3524 		if (!atomic_cmpset_int((u_int *)firstpte, newpde, newpde &
3525 		    ~PG_RW))
3526 			goto setpde;
3527 		newpde &= ~PG_RW;
3528 		CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#lx"
3529 		    " in pmap %p", va & ~PDRMASK, pmap);
3530 	}
3531 
3532 	/*
3533 	 * Examine each of the other PTEs in the specified PTP.  Abort if this
3534 	 * PTE maps an unexpected 4KB physical page or does not have identical
3535 	 * characteristics to the first PTE.
3536 	 */
3537 	allpte_PG_A = newpde & PG_A;
3538 	pa = (newpde & (PG_PS_FRAME | PG_V)) + NBPDR - PAGE_SIZE;
3539 	for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) {
3540 setpte:
3541 		oldpte = *pte;
3542 		if ((oldpte & (PG_FRAME | PG_V)) != pa) {
3543 			pmap_pde_p_failures++;
3544 			CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x"
3545 			    " in pmap %p", va, pmap);
3546 			return (false);
3547 		}
3548 		if ((oldpte & (PG_M | PG_RW)) == PG_RW) {
3549 			/*
3550 			 * When PG_M is already clear, PG_RW can be cleared
3551 			 * without a TLB invalidation.
3552 			 */
3553 			if (!atomic_cmpset_int((u_int *)pte, oldpte,
3554 			    oldpte & ~PG_RW))
3555 				goto setpte;
3556 			oldpte &= ~PG_RW;
3557 #ifdef KTR
3558 			oldpteva = (oldpte & PG_FRAME & PDRMASK) |
3559 			    (va & ~PDRMASK);
3560 #endif
3561 			CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#x"
3562 			    " in pmap %p", oldpteva, pmap);
3563 		}
3564 		if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) {
3565 			pmap_pde_p_failures++;
3566 			CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x"
3567 			    " in pmap %p", va, pmap);
3568 			return (false);
3569 		}
3570 		allpte_PG_A &= oldpte;
3571 		pa -= PAGE_SIZE;
3572 	}
3573 
3574 	/*
3575 	 * Unless all PTEs have PG_A set, clear it from the superpage mapping,
3576 	 * so that promotions triggered by speculative mappings, such as
3577 	 * pmap_enter_quick(), don't automatically mark the underlying pages
3578 	 * as referenced.
3579 	 */
3580 	newpde &= ~PG_A | allpte_PG_A;
3581 
3582 	/*
3583 	 * Save the PTP in its current state until the PDE mapping the
3584 	 * superpage is demoted by pmap_demote_pde() or destroyed by
3585 	 * pmap_remove_pde().  If PG_A is not set in every PTE, then request
3586 	 * that the PTP be refilled on demotion.
3587 	 */
3588 	if (mpte == NULL)
3589 		mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
3590 	KASSERT(mpte >= vm_page_array &&
3591 	    mpte < &vm_page_array[vm_page_array_size],
3592 	    ("pmap_promote_pde: page table page is out of range"));
3593 	KASSERT(mpte->pindex == va >> PDRSHIFT,
3594 	    ("pmap_promote_pde: page table page's pindex is wrong"));
3595 	if (pmap_insert_pt_page(pmap, mpte, true, allpte_PG_A != 0)) {
3596 		pmap_pde_p_failures++;
3597 		CTR2(KTR_PMAP,
3598 		    "pmap_promote_pde: failure for va %#x in pmap %p", va,
3599 		    pmap);
3600 		return (false);
3601 	}
3602 
3603 	/*
3604 	 * Promote the pv entries.
3605 	 */
3606 	if ((newpde & PG_MANAGED) != 0)
3607 		pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME);
3608 
3609 	/*
3610 	 * Propagate the PAT index to its proper position.
3611 	 */
3612 	if ((newpde & PG_PTE_PAT) != 0)
3613 		newpde ^= PG_PDE_PAT | PG_PTE_PAT;
3614 
3615 	/*
3616 	 * Map the superpage.
3617 	 */
3618 	if (workaround_erratum383)
3619 		pmap_update_pde(pmap, va, pde, PG_PS | newpde);
3620 	else if (pmap == kernel_pmap)
3621 		pmap_kenter_pde(va, PG_PROMOTED | PG_PS | newpde);
3622 	else
3623 		pde_store(pde, PG_PROMOTED | PG_PS | newpde);
3624 
3625 	pmap_pde_promotions++;
3626 	CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#x"
3627 	    " in pmap %p", va, pmap);
3628 	return (true);
3629 }
3630 #endif /* VM_NRESERVLEVEL > 0 */
3631 
3632 /*
3633  *	Insert the given physical page (p) at
3634  *	the specified virtual address (v) in the
3635  *	target physical map with the protection requested.
3636  *
3637  *	If specified, the page will be wired down, meaning
3638  *	that the related pte can not be reclaimed.
3639  *
3640  *	NB:  This is the only routine which MAY NOT lazy-evaluate
3641  *	or lose information.  That is, this routine must actually
3642  *	insert this page into the given map NOW.
3643  */
3644 static int
__CONCAT(PMTYPE,enter)3645 __CONCAT(PMTYPE, enter)(pmap_t pmap, vm_offset_t va, vm_page_t m,
3646     vm_prot_t prot, u_int flags, int8_t psind)
3647 {
3648 	pd_entry_t *pde;
3649 	pt_entry_t *pte;
3650 	pt_entry_t newpte, origpte;
3651 	pv_entry_t pv;
3652 	vm_paddr_t opa, pa;
3653 	vm_page_t mpte, om;
3654 	int rv;
3655 
3656 	va = trunc_page(va);
3657 	KASSERT((pmap == kernel_pmap && va < VM_MAX_KERNEL_ADDRESS) ||
3658 	    (pmap != kernel_pmap && va < VM_MAXUSER_ADDRESS),
3659 	    ("pmap_enter: toobig k%d %#x", pmap == kernel_pmap, va));
3660 	KASSERT(va < PMAP_TRM_MIN_ADDRESS,
3661 	    ("pmap_enter: invalid to pmap_enter into trampoline (va: 0x%x)",
3662 	    va));
3663 	KASSERT(pmap != kernel_pmap || (m->oflags & VPO_UNMANAGED) != 0 ||
3664 	    !VA_IS_CLEANMAP(va),
3665 	    ("pmap_enter: managed mapping within the clean submap"));
3666 	if ((m->oflags & VPO_UNMANAGED) == 0)
3667 		VM_PAGE_OBJECT_BUSY_ASSERT(m);
3668 	KASSERT((flags & PMAP_ENTER_RESERVED) == 0,
3669 	    ("pmap_enter: flags %u has reserved bits set", flags));
3670 	pa = VM_PAGE_TO_PHYS(m);
3671 	newpte = (pt_entry_t)(pa | PG_A | PG_V);
3672 	if ((flags & VM_PROT_WRITE) != 0)
3673 		newpte |= PG_M;
3674 	if ((prot & VM_PROT_WRITE) != 0)
3675 		newpte |= PG_RW;
3676 	KASSERT((newpte & (PG_M | PG_RW)) != PG_M,
3677 	    ("pmap_enter: flags includes VM_PROT_WRITE but prot doesn't"));
3678 #ifdef PMAP_PAE_COMP
3679 	if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec)
3680 		newpte |= pg_nx;
3681 #endif
3682 	if ((flags & PMAP_ENTER_WIRED) != 0)
3683 		newpte |= PG_W;
3684 	if (pmap != kernel_pmap)
3685 		newpte |= PG_U;
3686 	newpte |= pmap_cache_bits(pmap, m->md.pat_mode, psind > 0);
3687 	if ((m->oflags & VPO_UNMANAGED) == 0)
3688 		newpte |= PG_MANAGED;
3689 
3690 	rw_wlock(&pvh_global_lock);
3691 	PMAP_LOCK(pmap);
3692 	sched_pin();
3693 	if (psind == 1) {
3694 		/* Assert the required virtual and physical alignment. */
3695 		KASSERT((va & PDRMASK) == 0, ("pmap_enter: va unaligned"));
3696 		KASSERT(m->psind > 0, ("pmap_enter: m->psind < psind"));
3697 		rv = pmap_enter_pde(pmap, va, newpte | PG_PS, flags, m);
3698 		goto out;
3699 	}
3700 
3701 	pde = pmap_pde(pmap, va);
3702 	if (pmap != kernel_pmap) {
3703 		/*
3704 		 * va is for UVA.
3705 		 * In the case that a page table page is not resident,
3706 		 * we are creating it here.  pmap_allocpte() handles
3707 		 * demotion.
3708 		 */
3709 		mpte = pmap_allocpte(pmap, va, flags);
3710 		if (mpte == NULL) {
3711 			KASSERT((flags & PMAP_ENTER_NOSLEEP) != 0,
3712 			    ("pmap_allocpte failed with sleep allowed"));
3713 			rv = KERN_RESOURCE_SHORTAGE;
3714 			goto out;
3715 		}
3716 	} else {
3717 		/*
3718 		 * va is for KVA, so pmap_demote_pde() will never fail
3719 		 * to install a page table page.  PG_V is also
3720 		 * asserted by pmap_demote_pde().
3721 		 */
3722 		mpte = NULL;
3723 		KASSERT(pde != NULL && (*pde & PG_V) != 0,
3724 		    ("KVA %#x invalid pde pdir %#jx", va,
3725 		    (uintmax_t)pmap->pm_pdir[PTDPTDI]));
3726 		if ((*pde & PG_PS) != 0)
3727 			pmap_demote_pde(pmap, pde, va);
3728 	}
3729 	pte = pmap_pte_quick(pmap, va);
3730 
3731 	/*
3732 	 * Page Directory table entry is not valid, which should not
3733 	 * happen.  We should have either allocated the page table
3734 	 * page or demoted the existing mapping above.
3735 	 */
3736 	if (pte == NULL) {
3737 		panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x",
3738 		    (uintmax_t)pmap->pm_pdir[PTDPTDI], va);
3739 	}
3740 
3741 	origpte = *pte;
3742 	pv = NULL;
3743 
3744 	/*
3745 	 * Is the specified virtual address already mapped?
3746 	 */
3747 	if ((origpte & PG_V) != 0) {
3748 		/*
3749 		 * Wiring change, just update stats. We don't worry about
3750 		 * wiring PT pages as they remain resident as long as there
3751 		 * are valid mappings in them. Hence, if a user page is wired,
3752 		 * the PT page will be also.
3753 		 */
3754 		if ((newpte & PG_W) != 0 && (origpte & PG_W) == 0)
3755 			pmap->pm_stats.wired_count++;
3756 		else if ((newpte & PG_W) == 0 && (origpte & PG_W) != 0)
3757 			pmap->pm_stats.wired_count--;
3758 
3759 		/*
3760 		 * Remove the extra PT page reference.
3761 		 */
3762 		if (mpte != NULL) {
3763 			mpte->ref_count--;
3764 			KASSERT(mpte->ref_count > 0,
3765 			    ("pmap_enter: missing reference to page table page,"
3766 			     " va: 0x%x", va));
3767 		}
3768 
3769 		/*
3770 		 * Has the physical page changed?
3771 		 */
3772 		opa = origpte & PG_FRAME;
3773 		if (opa == pa) {
3774 			/*
3775 			 * No, might be a protection or wiring change.
3776 			 */
3777 			if ((origpte & PG_MANAGED) != 0 &&
3778 			    (newpte & PG_RW) != 0)
3779 				vm_page_aflag_set(m, PGA_WRITEABLE);
3780 			if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0)
3781 				goto unchanged;
3782 			goto validate;
3783 		}
3784 
3785 		/*
3786 		 * The physical page has changed.  Temporarily invalidate
3787 		 * the mapping.  This ensures that all threads sharing the
3788 		 * pmap keep a consistent view of the mapping, which is
3789 		 * necessary for the correct handling of COW faults.  It
3790 		 * also permits reuse of the old mapping's PV entry,
3791 		 * avoiding an allocation.
3792 		 *
3793 		 * For consistency, handle unmanaged mappings the same way.
3794 		 */
3795 		origpte = pte_load_clear(pte);
3796 		KASSERT((origpte & PG_FRAME) == opa,
3797 		    ("pmap_enter: unexpected pa update for %#x", va));
3798 		if ((origpte & PG_MANAGED) != 0) {
3799 			om = PHYS_TO_VM_PAGE(opa);
3800 
3801 			/*
3802 			 * The pmap lock is sufficient to synchronize with
3803 			 * concurrent calls to pmap_page_test_mappings() and
3804 			 * pmap_ts_referenced().
3805 			 */
3806 			if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
3807 				vm_page_dirty(om);
3808 			if ((origpte & PG_A) != 0) {
3809 				pmap_invalidate_page_int(pmap, va);
3810 				vm_page_aflag_set(om, PGA_REFERENCED);
3811 			}
3812 			pv = pmap_pvh_remove(&om->md, pmap, va);
3813 			KASSERT(pv != NULL,
3814 			    ("pmap_enter: no PV entry for %#x", va));
3815 			if ((newpte & PG_MANAGED) == 0)
3816 				free_pv_entry(pmap, pv);
3817 			if ((om->a.flags & PGA_WRITEABLE) != 0 &&
3818 			    TAILQ_EMPTY(&om->md.pv_list) &&
3819 			    ((om->flags & PG_FICTITIOUS) != 0 ||
3820 			    TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list)))
3821 				vm_page_aflag_clear(om, PGA_WRITEABLE);
3822 		} else {
3823 			/*
3824 			 * Since this mapping is unmanaged, assume that PG_A
3825 			 * is set.
3826 			 */
3827 			pmap_invalidate_page_int(pmap, va);
3828 		}
3829 		origpte = 0;
3830 	} else {
3831 		/*
3832 		 * Increment the counters.
3833 		 */
3834 		if ((newpte & PG_W) != 0)
3835 			pmap->pm_stats.wired_count++;
3836 		pmap->pm_stats.resident_count++;
3837 	}
3838 
3839 	/*
3840 	 * Enter on the PV list if part of our managed memory.
3841 	 */
3842 	if ((newpte & PG_MANAGED) != 0) {
3843 		if (pv == NULL) {
3844 			pv = get_pv_entry(pmap, false);
3845 			pv->pv_va = va;
3846 		}
3847 		TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
3848 		if ((newpte & PG_RW) != 0)
3849 			vm_page_aflag_set(m, PGA_WRITEABLE);
3850 	}
3851 
3852 	/*
3853 	 * Update the PTE.
3854 	 */
3855 	if ((origpte & PG_V) != 0) {
3856 validate:
3857 		origpte = pte_load_store(pte, newpte);
3858 		KASSERT((origpte & PG_FRAME) == pa,
3859 		    ("pmap_enter: unexpected pa update for %#x", va));
3860 		if ((newpte & PG_M) == 0 && (origpte & (PG_M | PG_RW)) ==
3861 		    (PG_M | PG_RW)) {
3862 			if ((origpte & PG_MANAGED) != 0)
3863 				vm_page_dirty(m);
3864 
3865 			/*
3866 			 * Although the PTE may still have PG_RW set, TLB
3867 			 * invalidation may nonetheless be required because
3868 			 * the PTE no longer has PG_M set.
3869 			 */
3870 		}
3871 #ifdef PMAP_PAE_COMP
3872 		else if ((origpte & PG_NX) != 0 || (newpte & PG_NX) == 0) {
3873 			/*
3874 			 * This PTE change does not require TLB invalidation.
3875 			 */
3876 			goto unchanged;
3877 		}
3878 #endif
3879 		if ((origpte & PG_A) != 0)
3880 			pmap_invalidate_page_int(pmap, va);
3881 	} else
3882 		pte_store_zero(pte, newpte);
3883 
3884 unchanged:
3885 
3886 #if VM_NRESERVLEVEL > 0
3887 	/*
3888 	 * If both the page table page and the reservation are fully
3889 	 * populated, then attempt promotion.
3890 	 */
3891 	if ((mpte == NULL || mpte->ref_count == NPTEPG) &&
3892 	    (m->flags & PG_FICTITIOUS) == 0 &&
3893 	    vm_reserv_level_iffullpop(m) == 0)
3894 		(void)pmap_promote_pde(pmap, pde, va, mpte);
3895 #endif
3896 
3897 	rv = KERN_SUCCESS;
3898 out:
3899 	sched_unpin();
3900 	rw_wunlock(&pvh_global_lock);
3901 	PMAP_UNLOCK(pmap);
3902 	return (rv);
3903 }
3904 
3905 /*
3906  * Tries to create a read- and/or execute-only 2 or 4 MB page mapping.  Returns
3907  * KERN_SUCCESS if the mapping was created.  Otherwise, returns an error
3908  * value.  See pmap_enter_pde() for the possible error values when "no sleep",
3909  * "no replace", and "no reclaim" are specified.
3910  */
3911 static int
pmap_enter_4mpage(pmap_t pmap,vm_offset_t va,vm_page_t m,vm_prot_t prot)3912 pmap_enter_4mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot)
3913 {
3914 	pd_entry_t newpde;
3915 
3916 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3917 	newpde = VM_PAGE_TO_PHYS(m) |
3918 	    pmap_cache_bits(pmap, m->md.pat_mode, true) | PG_PS | PG_V;
3919 	if ((m->oflags & VPO_UNMANAGED) == 0)
3920 		newpde |= PG_MANAGED;
3921 #ifdef PMAP_PAE_COMP
3922 	if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec)
3923 		newpde |= pg_nx;
3924 #endif
3925 	if (pmap != kernel_pmap)
3926 		newpde |= PG_U;
3927 	return (pmap_enter_pde(pmap, va, newpde, PMAP_ENTER_NOSLEEP |
3928 	    PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, NULL));
3929 }
3930 
3931 /*
3932  * Returns true if every page table entry in the page table page that maps
3933  * the specified kernel virtual address is zero.
3934  */
3935 static bool
pmap_every_pte_zero(vm_offset_t va)3936 pmap_every_pte_zero(vm_offset_t va)
3937 {
3938 	pt_entry_t *pt_end, *pte;
3939 
3940 	KASSERT((va & PDRMASK) == 0, ("va is misaligned"));
3941 	pte = vtopte(va);
3942 	for (pt_end = pte + NPTEPG; pte < pt_end; pte++) {
3943 		if (*pte != 0)
3944 			return (false);
3945 	}
3946 	return (true);
3947 }
3948 
3949 /*
3950  * Tries to create the specified 2 or 4 MB page mapping.  Returns KERN_SUCCESS
3951  * if the mapping was created, and one of KERN_FAILURE, KERN_NO_SPACE,
3952  * or KERN_RESOURCE_SHORTAGE otherwise.  Returns KERN_FAILURE if
3953  * PMAP_ENTER_NOREPLACE was specified and a 4 KB page mapping already exists
3954  * within the 2 or 4 MB virtual address range starting at the specified virtual
3955  * address.  Returns KERN_NO_SPACE if PMAP_ENTER_NOREPLACE was specified and a
3956  * 2 or 4 MB page mapping already exists at the specified virtual address.
3957  * Returns KERN_RESOURCE_SHORTAGE if PMAP_ENTER_NORECLAIM was specified and a
3958  * PV entry allocation failed.
3959  *
3960  * The parameter "m" is only used when creating a managed, writeable mapping.
3961  */
3962 static int
pmap_enter_pde(pmap_t pmap,vm_offset_t va,pd_entry_t newpde,u_int flags,vm_page_t m)3963 pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags,
3964     vm_page_t m)
3965 {
3966 	struct spglist free;
3967 	pd_entry_t oldpde, *pde;
3968 	vm_page_t mt;
3969 	vm_page_t uwptpg;
3970 
3971 	rw_assert(&pvh_global_lock, RA_WLOCKED);
3972 	KASSERT((newpde & (PG_M | PG_RW)) != PG_RW,
3973 	    ("pmap_enter_pde: newpde is missing PG_M"));
3974 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3975 	pde = pmap_pde(pmap, va);
3976 	oldpde = *pde;
3977 	if ((oldpde & PG_V) != 0) {
3978 		if ((flags & PMAP_ENTER_NOREPLACE) != 0) {
3979 			if ((oldpde & PG_PS) != 0) {
3980 				CTR2(KTR_PMAP,
3981 				    "pmap_enter_pde: no space for va %#lx"
3982 				    " in pmap %p", va, pmap);
3983 				return (KERN_NO_SPACE);
3984 			} else if (pmap != kernel_pmap ||
3985 			    !pmap_every_pte_zero(va)) {
3986 				CTR2(KTR_PMAP,
3987 				    "pmap_enter_pde: failure for va %#lx"
3988 				    " in pmap %p", va, pmap);
3989 				return (KERN_FAILURE);
3990 			}
3991 		}
3992 		/* Break the existing mapping(s). */
3993 		SLIST_INIT(&free);
3994 		if ((oldpde & PG_PS) != 0) {
3995 			/*
3996 			 * If the PDE resulted from a promotion, then a
3997 			 * reserved PT page could be freed.
3998 			 */
3999 			(void)pmap_remove_pde(pmap, pde, va, &free);
4000 			if ((oldpde & PG_G) == 0)
4001 				pmap_invalidate_pde_page(pmap, va, oldpde);
4002 		} else {
4003 			if (pmap_remove_ptes(pmap, va, va + NBPDR, &free))
4004 		               pmap_invalidate_all_int(pmap);
4005 		}
4006 		if (pmap != kernel_pmap) {
4007 			vm_page_free_pages_toq(&free, true);
4008 			KASSERT(*pde == 0, ("pmap_enter_pde: non-zero pde %p",
4009 			    pde));
4010 		} else {
4011 			KASSERT(SLIST_EMPTY(&free),
4012 			    ("pmap_enter_pde: freed kernel page table page"));
4013 
4014 			/*
4015 			 * Both pmap_remove_pde() and pmap_remove_ptes() will
4016 			 * leave the kernel page table page zero filled.
4017 			 */
4018 			mt = PHYS_TO_VM_PAGE(*pde & PG_FRAME);
4019 			if (pmap_insert_pt_page(pmap, mt, false, false))
4020 				panic("pmap_enter_pde: trie insert failed");
4021 		}
4022 	}
4023 
4024 	/*
4025 	 * Allocate a leaf ptpage for wired userspace pages.
4026 	 */
4027 	uwptpg = NULL;
4028 	if ((newpde & PG_W) != 0 && pmap != kernel_pmap) {
4029 		uwptpg = vm_page_alloc_noobj(VM_ALLOC_WIRED);
4030 		if (uwptpg == NULL) {
4031 			return (KERN_RESOURCE_SHORTAGE);
4032 		}
4033 		uwptpg->pindex = va >> PDRSHIFT;
4034 		if (pmap_insert_pt_page(pmap, uwptpg, true, false)) {
4035 			vm_page_unwire_noq(uwptpg);
4036 			vm_page_free(uwptpg);
4037 			return (KERN_RESOURCE_SHORTAGE);
4038 		}
4039 		pmap->pm_stats.resident_count++;
4040 		uwptpg->ref_count = NPTEPG;
4041 	}
4042 	if ((newpde & PG_MANAGED) != 0) {
4043 		/*
4044 		 * Abort this mapping if its PV entry could not be created.
4045 		 */
4046 		if (!pmap_pv_insert_pde(pmap, va, newpde, flags)) {
4047 			if (uwptpg != NULL) {
4048 				mt = pmap_remove_pt_page(pmap, va);
4049 				KASSERT(mt == uwptpg,
4050 				    ("removed pt page %p, expected %p", mt,
4051 				    uwptpg));
4052 				pmap->pm_stats.resident_count--;
4053 				uwptpg->ref_count = 1;
4054 				vm_page_unwire_noq(uwptpg);
4055 				vm_page_free(uwptpg);
4056 			}
4057 			CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
4058 			    " in pmap %p", va, pmap);
4059 			return (KERN_RESOURCE_SHORTAGE);
4060 		}
4061 		if ((newpde & PG_RW) != 0) {
4062 			for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
4063 				vm_page_aflag_set(mt, PGA_WRITEABLE);
4064 		}
4065 	}
4066 
4067 	/*
4068 	 * Increment counters.
4069 	 */
4070 	if ((newpde & PG_W) != 0)
4071 		pmap->pm_stats.wired_count += NBPDR / PAGE_SIZE;
4072 	pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
4073 
4074 	/*
4075 	 * Map the superpage.  (This is not a promoted mapping; there will not
4076 	 * be any lingering 4KB page mappings in the TLB.)
4077 	 */
4078 	pde_store(pde, newpde);
4079 
4080 	pmap_pde_mappings++;
4081 	CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx in pmap %p",
4082 	    va, pmap);
4083 	return (KERN_SUCCESS);
4084 }
4085 
4086 /*
4087  * Maps a sequence of resident pages belonging to the same object.
4088  * The sequence begins with the given page m_start.  This page is
4089  * mapped at the given virtual address start.  Each subsequent page is
4090  * mapped at a virtual address that is offset from start by the same
4091  * amount as the page is offset from m_start within the object.  The
4092  * last page in the sequence is the page with the largest offset from
4093  * m_start that can be mapped at a virtual address less than the given
4094  * virtual address end.  Not every virtual page between start and end
4095  * is mapped; only those for which a resident page exists with the
4096  * corresponding offset from m_start are mapped.
4097  */
4098 static void
__CONCAT(PMTYPE,enter_object)4099 __CONCAT(PMTYPE, enter_object)(pmap_t pmap, vm_offset_t start, vm_offset_t end,
4100     vm_page_t m_start, vm_prot_t prot)
4101 {
4102 	struct pctrie_iter pages;
4103 	vm_offset_t va;
4104 	vm_page_t m, mpte;
4105 	int rv;
4106 
4107 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
4108 
4109 	mpte = NULL;
4110 	vm_page_iter_limit_init(&pages, m_start->object,
4111 	    m_start->pindex + atop(end - start));
4112 	m = vm_radix_iter_lookup(&pages, m_start->pindex);
4113 	rw_wlock(&pvh_global_lock);
4114 	PMAP_LOCK(pmap);
4115 	while (m != NULL) {
4116 		va = start + ptoa(m->pindex - m_start->pindex);
4117 		if ((va & PDRMASK) == 0 && va + NBPDR <= end &&
4118 		    m->psind == 1 && pg_ps_enabled &&
4119 		    ((rv = pmap_enter_4mpage(pmap, va, m, prot)) ==
4120 		    KERN_SUCCESS || rv == KERN_NO_SPACE)) {
4121 			m = vm_radix_iter_jump(&pages, NBPDR / PAGE_SIZE);
4122 		} else {
4123 			mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte);
4124 			m = vm_radix_iter_step(&pages);
4125 		}
4126 	}
4127 	rw_wunlock(&pvh_global_lock);
4128 	PMAP_UNLOCK(pmap);
4129 }
4130 
4131 /*
4132  * this code makes some *MAJOR* assumptions:
4133  * 1. Current pmap & pmap exists.
4134  * 2. Not wired.
4135  * 3. Read access.
4136  * 4. No page table pages.
4137  * but is *MUCH* faster than pmap_enter...
4138  */
4139 
4140 static void
__CONCAT(PMTYPE,enter_quick)4141 __CONCAT(PMTYPE, enter_quick)(pmap_t pmap, vm_offset_t va, vm_page_t m,
4142     vm_prot_t prot)
4143 {
4144 
4145 	rw_wlock(&pvh_global_lock);
4146 	PMAP_LOCK(pmap);
4147 	(void)pmap_enter_quick_locked(pmap, va, m, prot, NULL);
4148 	rw_wunlock(&pvh_global_lock);
4149 	PMAP_UNLOCK(pmap);
4150 }
4151 
4152 static vm_page_t
pmap_enter_quick_locked(pmap_t pmap,vm_offset_t va,vm_page_t m,vm_prot_t prot,vm_page_t mpte)4153 pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m,
4154     vm_prot_t prot, vm_page_t mpte)
4155 {
4156 	pt_entry_t newpte, *pte;
4157 	pd_entry_t *pde;
4158 
4159 	KASSERT(pmap != kernel_pmap || !VA_IS_CLEANMAP(va) ||
4160 	    (m->oflags & VPO_UNMANAGED) != 0,
4161 	    ("pmap_enter_quick_locked: managed mapping within the clean submap"));
4162 	rw_assert(&pvh_global_lock, RA_WLOCKED);
4163 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4164 	pde = NULL;
4165 
4166 	/*
4167 	 * In the case that a page table page is not
4168 	 * resident, we are creating it here.
4169 	 */
4170 	if (pmap != kernel_pmap) {
4171 		u_int ptepindex;
4172 		pd_entry_t ptepa;
4173 
4174 		/*
4175 		 * Calculate pagetable page index
4176 		 */
4177 		ptepindex = va >> PDRSHIFT;
4178 		if (mpte && (mpte->pindex == ptepindex)) {
4179 			mpte->ref_count++;
4180 		} else {
4181 			/*
4182 			 * Get the page directory entry
4183 			 */
4184 			pde = &pmap->pm_pdir[ptepindex];
4185 			ptepa = *pde;
4186 
4187 			/*
4188 			 * If the page table page is mapped, we just increment
4189 			 * the hold count, and activate it.
4190 			 */
4191 			if (ptepa) {
4192 				if (ptepa & PG_PS)
4193 					return (NULL);
4194 				mpte = PHYS_TO_VM_PAGE(ptepa & PG_FRAME);
4195 				mpte->ref_count++;
4196 			} else {
4197 				mpte = _pmap_allocpte(pmap, ptepindex,
4198 				    PMAP_ENTER_NOSLEEP);
4199 				if (mpte == NULL)
4200 					return (mpte);
4201 			}
4202 		}
4203 	} else {
4204 		mpte = NULL;
4205 	}
4206 
4207 	sched_pin();
4208 	pte = pmap_pte_quick(pmap, va);
4209 	if (*pte) {
4210 		if (mpte != NULL)
4211 			mpte->ref_count--;
4212 		sched_unpin();
4213 		return (NULL);
4214 	}
4215 
4216 	/*
4217 	 * Enter on the PV list if part of our managed memory.
4218 	 */
4219 	if ((m->oflags & VPO_UNMANAGED) == 0 &&
4220 	    !pmap_try_insert_pv_entry(pmap, va, m)) {
4221 		if (mpte != NULL)
4222 			pmap_abort_ptp(pmap, va, mpte);
4223 		sched_unpin();
4224 		return (NULL);
4225 	}
4226 
4227 	/*
4228 	 * Increment counters
4229 	 */
4230 	pmap->pm_stats.resident_count++;
4231 
4232 	newpte = VM_PAGE_TO_PHYS(m) | PG_V |
4233 	    pmap_cache_bits(pmap, m->md.pat_mode, false);
4234 	if ((m->oflags & VPO_UNMANAGED) == 0)
4235 		newpte |= PG_MANAGED;
4236 #ifdef PMAP_PAE_COMP
4237 	if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec)
4238 		newpte |= pg_nx;
4239 #endif
4240 	if (pmap != kernel_pmap)
4241 		newpte |= PG_U;
4242 	pte_store_zero(pte, newpte);
4243 
4244 #if VM_NRESERVLEVEL > 0
4245 	/*
4246 	 * If both the PTP and the reservation are fully populated, then
4247 	 * attempt promotion.
4248 	 */
4249 	if ((prot & VM_PROT_NO_PROMOTE) == 0 &&
4250 	    (mpte == NULL || mpte->ref_count == NPTEPG) &&
4251 	    (m->flags & PG_FICTITIOUS) == 0 &&
4252 	    vm_reserv_level_iffullpop(m) == 0) {
4253 		if (pde == NULL)
4254 			pde = pmap_pde(pmap, va);
4255 
4256 		/*
4257 		 * If promotion succeeds, then the next call to this function
4258 		 * should not be given the unmapped PTP as a hint.
4259 		 */
4260 		if (pmap_promote_pde(pmap, pde, va, mpte))
4261 			mpte = NULL;
4262 	}
4263 #endif
4264 
4265 	sched_unpin();
4266 	return (mpte);
4267 }
4268 
4269 /*
4270  * Make a temporary mapping for a physical address.  This is only intended
4271  * to be used for panic dumps.
4272  */
4273 static void *
__CONCAT(PMTYPE,kenter_temporary)4274 __CONCAT(PMTYPE, kenter_temporary)(vm_paddr_t pa, int i)
4275 {
4276 	vm_offset_t va;
4277 
4278 	va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
4279 	pmap_kenter(va, pa);
4280 	invlpg(va);
4281 	return ((void *)crashdumpmap);
4282 }
4283 
4284 /*
4285  * This code maps large physical mmap regions into the
4286  * processor address space.  Note that some shortcuts
4287  * are taken, but the code works.
4288  */
4289 static void
__CONCAT(PMTYPE,object_init_pt)4290 __CONCAT(PMTYPE, object_init_pt)(pmap_t pmap, vm_offset_t addr,
4291     vm_object_t object, vm_pindex_t pindex, vm_size_t size)
4292 {
4293 	struct pctrie_iter pages;
4294 	pd_entry_t *pde;
4295 	vm_paddr_t pa, ptepa;
4296 	vm_page_t p;
4297 	int pat_mode;
4298 
4299 	VM_OBJECT_ASSERT_WLOCKED(object);
4300 	KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG,
4301 	    ("pmap_object_init_pt: non-device object"));
4302 	if (pg_ps_enabled &&
4303 	    (addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) {
4304 		if (!vm_object_populate(object, pindex, pindex + atop(size)))
4305 			return;
4306 		vm_page_iter_init(&pages, object);
4307 		p = vm_radix_iter_lookup(&pages, pindex);
4308 		KASSERT(vm_page_all_valid(p),
4309 		    ("pmap_object_init_pt: invalid page %p", p));
4310 		pat_mode = p->md.pat_mode;
4311 
4312 		/*
4313 		 * Abort the mapping if the first page is not physically
4314 		 * aligned to a 2/4MB page boundary.
4315 		 */
4316 		ptepa = VM_PAGE_TO_PHYS(p);
4317 		if (ptepa & (NBPDR - 1))
4318 			return;
4319 
4320 		/*
4321 		 * Skip the first page.  Abort the mapping if the rest of
4322 		 * the pages are not physically contiguous or have differing
4323 		 * memory attributes.
4324 		 */
4325 		for (pa = ptepa + PAGE_SIZE; pa < ptepa + size;
4326 		    pa += PAGE_SIZE) {
4327 			p = vm_radix_iter_next(&pages);
4328 			KASSERT(vm_page_all_valid(p),
4329 			    ("pmap_object_init_pt: invalid page %p", p));
4330 			if (pa != VM_PAGE_TO_PHYS(p) ||
4331 			    pat_mode != p->md.pat_mode)
4332 				return;
4333 		}
4334 
4335 		/*
4336 		 * Map using 2/4MB pages.  Since "ptepa" is 2/4M aligned and
4337 		 * "size" is a multiple of 2/4M, adding the PAT setting to
4338 		 * "pa" will not affect the termination of this loop.
4339 		 */
4340 		PMAP_LOCK(pmap);
4341 		for (pa = ptepa | pmap_cache_bits(pmap, pat_mode, true);
4342 		    pa < ptepa + size; pa += NBPDR) {
4343 			pde = pmap_pde(pmap, addr);
4344 			if (*pde == 0) {
4345 				pde_store(pde, pa | PG_PS | PG_M | PG_A |
4346 				    PG_U | PG_RW | PG_V);
4347 				pmap->pm_stats.resident_count += NBPDR /
4348 				    PAGE_SIZE;
4349 				pmap_pde_mappings++;
4350 			}
4351 			/* Else continue on if the PDE is already valid. */
4352 			addr += NBPDR;
4353 		}
4354 		PMAP_UNLOCK(pmap);
4355 	}
4356 }
4357 
4358 /*
4359  *	Clear the wired attribute from the mappings for the specified range of
4360  *	addresses in the given pmap.  Every valid mapping within that range
4361  *	must have the wired attribute set.  In contrast, invalid mappings
4362  *	cannot have the wired attribute set, so they are ignored.
4363  *
4364  *	The wired attribute of the page table entry is not a hardware feature,
4365  *	so there is no need to invalidate any TLB entries.
4366  */
4367 static void
__CONCAT(PMTYPE,unwire)4368 __CONCAT(PMTYPE, unwire)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
4369 {
4370 	vm_offset_t pdnxt;
4371 	pd_entry_t *pde;
4372 	pt_entry_t *pte;
4373 	bool pv_lists_locked;
4374 
4375 	if (pmap_is_current(pmap))
4376 		pv_lists_locked = false;
4377 	else {
4378 		pv_lists_locked = true;
4379 resume:
4380 		rw_wlock(&pvh_global_lock);
4381 		sched_pin();
4382 	}
4383 	PMAP_LOCK(pmap);
4384 	for (; sva < eva; sva = pdnxt) {
4385 		pdnxt = (sva + NBPDR) & ~PDRMASK;
4386 		if (pdnxt < sva)
4387 			pdnxt = eva;
4388 		pde = pmap_pde(pmap, sva);
4389 		if ((*pde & PG_V) == 0)
4390 			continue;
4391 		if ((*pde & PG_PS) != 0) {
4392 			if ((*pde & PG_W) == 0)
4393 				panic("pmap_unwire: pde %#jx is missing PG_W",
4394 				    (uintmax_t)*pde);
4395 
4396 			/*
4397 			 * Are we unwiring the entire large page?  If not,
4398 			 * demote the mapping and fall through.
4399 			 */
4400 			if (sva + NBPDR == pdnxt && eva >= pdnxt) {
4401 				/*
4402 				 * Regardless of whether a pde (or pte) is 32
4403 				 * or 64 bits in size, PG_W is among the least
4404 				 * significant 32 bits.
4405 				 */
4406 				atomic_clear_int((u_int *)pde, PG_W);
4407 				pmap->pm_stats.wired_count -= NBPDR /
4408 				    PAGE_SIZE;
4409 				continue;
4410 			} else {
4411 				if (!pv_lists_locked) {
4412 					pv_lists_locked = true;
4413 					if (!rw_try_wlock(&pvh_global_lock)) {
4414 						PMAP_UNLOCK(pmap);
4415 						/* Repeat sva. */
4416 						goto resume;
4417 					}
4418 					sched_pin();
4419 				}
4420 				if (!pmap_demote_pde(pmap, pde, sva))
4421 					panic("pmap_unwire: demotion failed");
4422 			}
4423 		}
4424 		if (pdnxt > eva)
4425 			pdnxt = eva;
4426 		for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
4427 		    sva += PAGE_SIZE) {
4428 			if ((*pte & PG_V) == 0)
4429 				continue;
4430 			if ((*pte & PG_W) == 0)
4431 				panic("pmap_unwire: pte %#jx is missing PG_W",
4432 				    (uintmax_t)*pte);
4433 
4434 			/*
4435 			 * PG_W must be cleared atomically.  Although the pmap
4436 			 * lock synchronizes access to PG_W, another processor
4437 			 * could be setting PG_M and/or PG_A concurrently.
4438 			 *
4439 			 * PG_W is among the least significant 32 bits.
4440 			 */
4441 			atomic_clear_int((u_int *)pte, PG_W);
4442 			pmap->pm_stats.wired_count--;
4443 		}
4444 	}
4445 	if (pv_lists_locked) {
4446 		sched_unpin();
4447 		rw_wunlock(&pvh_global_lock);
4448 	}
4449 	PMAP_UNLOCK(pmap);
4450 }
4451 
4452 /*
4453  *	Copy the range specified by src_addr/len
4454  *	from the source map to the range dst_addr/len
4455  *	in the destination map.
4456  *
4457  *	This routine is only advisory and need not do anything.  Since
4458  *	current pmap is always the kernel pmap when executing in
4459  *	kernel, and we do not copy from the kernel pmap to a user
4460  *	pmap, this optimization is not usable in 4/4G full split i386
4461  *	world.
4462  */
4463 
4464 static void
__CONCAT(PMTYPE,copy)4465 __CONCAT(PMTYPE, copy)(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4466     vm_size_t len, vm_offset_t src_addr)
4467 {
4468 	pt_entry_t *src_pte, *dst_pte, ptetemp;
4469 	pd_entry_t srcptepaddr;
4470 	vm_page_t dstmpte, srcmpte;
4471 	vm_offset_t addr, end_addr, pdnxt;
4472 	u_int ptepindex;
4473 
4474 	if (dst_addr != src_addr)
4475 		return;
4476 
4477 	end_addr = src_addr + len;
4478 
4479 	rw_wlock(&pvh_global_lock);
4480 	if (dst_pmap < src_pmap) {
4481 		PMAP_LOCK(dst_pmap);
4482 		PMAP_LOCK(src_pmap);
4483 	} else {
4484 		PMAP_LOCK(src_pmap);
4485 		PMAP_LOCK(dst_pmap);
4486 	}
4487 	sched_pin();
4488 	for (addr = src_addr; addr < end_addr; addr = pdnxt) {
4489 		KASSERT(addr < PMAP_TRM_MIN_ADDRESS,
4490 		    ("pmap_copy: invalid to pmap_copy the trampoline"));
4491 
4492 		pdnxt = (addr + NBPDR) & ~PDRMASK;
4493 		if (pdnxt < addr)
4494 			pdnxt = end_addr;
4495 		ptepindex = addr >> PDRSHIFT;
4496 
4497 		srcptepaddr = src_pmap->pm_pdir[ptepindex];
4498 		if (srcptepaddr == 0)
4499 			continue;
4500 
4501 		if (srcptepaddr & PG_PS) {
4502 			if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr)
4503 				continue;
4504 			if (dst_pmap->pm_pdir[ptepindex] == 0 &&
4505 			    ((srcptepaddr & PG_MANAGED) == 0 ||
4506 			    pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr,
4507 			    PMAP_ENTER_NORECLAIM))) {
4508 				dst_pmap->pm_pdir[ptepindex] = srcptepaddr &
4509 				    ~PG_W;
4510 				dst_pmap->pm_stats.resident_count +=
4511 				    NBPDR / PAGE_SIZE;
4512 				pmap_pde_mappings++;
4513 			}
4514 			continue;
4515 		}
4516 
4517 		srcmpte = PHYS_TO_VM_PAGE(srcptepaddr & PG_FRAME);
4518 		KASSERT(srcmpte->ref_count > 0,
4519 		    ("pmap_copy: source page table page is unused"));
4520 
4521 		if (pdnxt > end_addr)
4522 			pdnxt = end_addr;
4523 
4524 		src_pte = pmap_pte_quick3(src_pmap, addr);
4525 		while (addr < pdnxt) {
4526 			ptetemp = *src_pte;
4527 			/*
4528 			 * we only virtual copy managed pages
4529 			 */
4530 			if ((ptetemp & PG_MANAGED) != 0) {
4531 				dstmpte = pmap_allocpte(dst_pmap, addr,
4532 				    PMAP_ENTER_NOSLEEP);
4533 				if (dstmpte == NULL)
4534 					goto out;
4535 				dst_pte = pmap_pte_quick(dst_pmap, addr);
4536 				if (*dst_pte == 0 &&
4537 				    pmap_try_insert_pv_entry(dst_pmap, addr,
4538 				    PHYS_TO_VM_PAGE(ptetemp & PG_FRAME))) {
4539 					/*
4540 					 * Clear the wired, modified, and
4541 					 * accessed (referenced) bits
4542 					 * during the copy.
4543 					 */
4544 					*dst_pte = ptetemp & ~(PG_W | PG_M |
4545 					    PG_A);
4546 					dst_pmap->pm_stats.resident_count++;
4547 				} else {
4548 					pmap_abort_ptp(dst_pmap, addr, dstmpte);
4549 					goto out;
4550 				}
4551 				if (dstmpte->ref_count >= srcmpte->ref_count)
4552 					break;
4553 			}
4554 			addr += PAGE_SIZE;
4555 			src_pte++;
4556 		}
4557 	}
4558 out:
4559 	sched_unpin();
4560 	rw_wunlock(&pvh_global_lock);
4561 	PMAP_UNLOCK(src_pmap);
4562 	PMAP_UNLOCK(dst_pmap);
4563 }
4564 
4565 /*
4566  * Zero 1 page of virtual memory mapped from a hardware page by the caller.
4567  */
4568 static __inline void
pagezero(void * page)4569 pagezero(void *page)
4570 {
4571 #if defined(I686_CPU)
4572 	if (cpu_class == CPUCLASS_686) {
4573 		if (cpu_feature & CPUID_SSE2)
4574 			sse2_pagezero(page);
4575 		else
4576 			i686_pagezero(page);
4577 	} else
4578 #endif
4579 		bzero(page, PAGE_SIZE);
4580 }
4581 
4582 /*
4583  * Zero the specified hardware page.
4584  */
4585 static void
__CONCAT(PMTYPE,zero_page)4586 __CONCAT(PMTYPE, zero_page)(vm_page_t m)
4587 {
4588 	pt_entry_t *cmap_pte2;
4589 	struct pcpu *pc;
4590 
4591 	sched_pin();
4592 	pc = get_pcpu();
4593 	cmap_pte2 = pc->pc_cmap_pte2;
4594 	mtx_lock(&pc->pc_cmap_lock);
4595 	if (*cmap_pte2)
4596 		panic("pmap_zero_page: CMAP2 busy");
4597 	*cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M |
4598 	    pmap_cache_bits(kernel_pmap, m->md.pat_mode, false);
4599 	invlcaddr(pc->pc_cmap_addr2);
4600 	pagezero(pc->pc_cmap_addr2);
4601 	*cmap_pte2 = 0;
4602 
4603 	/*
4604 	 * Unpin the thread before releasing the lock.  Otherwise the thread
4605 	 * could be rescheduled while still bound to the current CPU, only
4606 	 * to unpin itself immediately upon resuming execution.
4607 	 */
4608 	sched_unpin();
4609 	mtx_unlock(&pc->pc_cmap_lock);
4610 }
4611 
4612 /*
4613  * Zero an area within a single hardware page.  off and size must not
4614  * cover an area beyond a single hardware page.
4615  */
4616 static void
__CONCAT(PMTYPE,zero_page_area)4617 __CONCAT(PMTYPE, zero_page_area)(vm_page_t m, int off, int size)
4618 {
4619 	pt_entry_t *cmap_pte2;
4620 	struct pcpu *pc;
4621 
4622 	sched_pin();
4623 	pc = get_pcpu();
4624 	cmap_pte2 = pc->pc_cmap_pte2;
4625 	mtx_lock(&pc->pc_cmap_lock);
4626 	if (*cmap_pte2)
4627 		panic("pmap_zero_page_area: CMAP2 busy");
4628 	*cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M |
4629 	    pmap_cache_bits(kernel_pmap, m->md.pat_mode, false);
4630 	invlcaddr(pc->pc_cmap_addr2);
4631 	if (off == 0 && size == PAGE_SIZE)
4632 		pagezero(pc->pc_cmap_addr2);
4633 	else
4634 		bzero(pc->pc_cmap_addr2 + off, size);
4635 	*cmap_pte2 = 0;
4636 	sched_unpin();
4637 	mtx_unlock(&pc->pc_cmap_lock);
4638 }
4639 
4640 /*
4641  * Copy 1 specified hardware page to another.
4642  */
4643 static void
__CONCAT(PMTYPE,copy_page)4644 __CONCAT(PMTYPE, copy_page)(vm_page_t src, vm_page_t dst)
4645 {
4646 	pt_entry_t *cmap_pte1, *cmap_pte2;
4647 	struct pcpu *pc;
4648 
4649 	sched_pin();
4650 	pc = get_pcpu();
4651 	cmap_pte1 = pc->pc_cmap_pte1;
4652 	cmap_pte2 = pc->pc_cmap_pte2;
4653 	mtx_lock(&pc->pc_cmap_lock);
4654 	if (*cmap_pte1)
4655 		panic("pmap_copy_page: CMAP1 busy");
4656 	if (*cmap_pte2)
4657 		panic("pmap_copy_page: CMAP2 busy");
4658 	*cmap_pte1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A |
4659 	    pmap_cache_bits(kernel_pmap, src->md.pat_mode, false);
4660 	invlcaddr(pc->pc_cmap_addr1);
4661 	*cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M |
4662 	    pmap_cache_bits(kernel_pmap, dst->md.pat_mode, false);
4663 	invlcaddr(pc->pc_cmap_addr2);
4664 	bcopy(pc->pc_cmap_addr1, pc->pc_cmap_addr2, PAGE_SIZE);
4665 	*cmap_pte1 = 0;
4666 	*cmap_pte2 = 0;
4667 	sched_unpin();
4668 	mtx_unlock(&pc->pc_cmap_lock);
4669 }
4670 
4671 static void
__CONCAT(PMTYPE,copy_pages)4672 __CONCAT(PMTYPE, copy_pages)(vm_page_t ma[], vm_offset_t a_offset,
4673     vm_page_t mb[], vm_offset_t b_offset, int xfersize)
4674 {
4675 	vm_page_t a_pg, b_pg;
4676 	char *a_cp, *b_cp;
4677 	vm_offset_t a_pg_offset, b_pg_offset;
4678 	pt_entry_t *cmap_pte1, *cmap_pte2;
4679 	struct pcpu *pc;
4680 	int cnt;
4681 
4682 	sched_pin();
4683 	pc = get_pcpu();
4684 	cmap_pte1 = pc->pc_cmap_pte1;
4685 	cmap_pte2 = pc->pc_cmap_pte2;
4686 	mtx_lock(&pc->pc_cmap_lock);
4687 	if (*cmap_pte1 != 0)
4688 		panic("pmap_copy_pages: CMAP1 busy");
4689 	if (*cmap_pte2 != 0)
4690 		panic("pmap_copy_pages: CMAP2 busy");
4691 	while (xfersize > 0) {
4692 		a_pg = ma[a_offset >> PAGE_SHIFT];
4693 		a_pg_offset = a_offset & PAGE_MASK;
4694 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
4695 		b_pg = mb[b_offset >> PAGE_SHIFT];
4696 		b_pg_offset = b_offset & PAGE_MASK;
4697 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
4698 		*cmap_pte1 = PG_V | VM_PAGE_TO_PHYS(a_pg) | PG_A |
4699 		    pmap_cache_bits(kernel_pmap, a_pg->md.pat_mode, false);
4700 		invlcaddr(pc->pc_cmap_addr1);
4701 		*cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(b_pg) | PG_A |
4702 		    PG_M | pmap_cache_bits(kernel_pmap, b_pg->md.pat_mode,
4703 		    false);
4704 		invlcaddr(pc->pc_cmap_addr2);
4705 		a_cp = pc->pc_cmap_addr1 + a_pg_offset;
4706 		b_cp = pc->pc_cmap_addr2 + b_pg_offset;
4707 		bcopy(a_cp, b_cp, cnt);
4708 		a_offset += cnt;
4709 		b_offset += cnt;
4710 		xfersize -= cnt;
4711 	}
4712 	*cmap_pte1 = 0;
4713 	*cmap_pte2 = 0;
4714 	sched_unpin();
4715 	mtx_unlock(&pc->pc_cmap_lock);
4716 }
4717 
4718 /*
4719  * Returns true if the pmap's pv is one of the first
4720  * 16 pvs linked to from this page.  This count may
4721  * be changed upwards or downwards in the future; it
4722  * is only necessary that true be returned for a small
4723  * subset of pmaps for proper page aging.
4724  */
4725 static bool
__CONCAT(PMTYPE,page_exists_quick)4726 __CONCAT(PMTYPE, page_exists_quick)(pmap_t pmap, vm_page_t m)
4727 {
4728 	struct md_page *pvh;
4729 	pv_entry_t pv;
4730 	int loops = 0;
4731 	bool rv;
4732 
4733 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4734 	    ("pmap_page_exists_quick: page %p is not managed", m));
4735 	rv = false;
4736 	rw_wlock(&pvh_global_lock);
4737 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
4738 		if (PV_PMAP(pv) == pmap) {
4739 			rv = true;
4740 			break;
4741 		}
4742 		loops++;
4743 		if (loops >= 16)
4744 			break;
4745 	}
4746 	if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) {
4747 		pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4748 		TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
4749 			if (PV_PMAP(pv) == pmap) {
4750 				rv = true;
4751 				break;
4752 			}
4753 			loops++;
4754 			if (loops >= 16)
4755 				break;
4756 		}
4757 	}
4758 	rw_wunlock(&pvh_global_lock);
4759 	return (rv);
4760 }
4761 
4762 /*
4763  *	pmap_page_wired_mappings:
4764  *
4765  *	Return the number of managed mappings to the given physical page
4766  *	that are wired.
4767  */
4768 static int
__CONCAT(PMTYPE,page_wired_mappings)4769 __CONCAT(PMTYPE, page_wired_mappings)(vm_page_t m)
4770 {
4771 	int count;
4772 
4773 	count = 0;
4774 	if ((m->oflags & VPO_UNMANAGED) != 0)
4775 		return (count);
4776 	rw_wlock(&pvh_global_lock);
4777 	count = pmap_pvh_wired_mappings(&m->md, count);
4778 	if ((m->flags & PG_FICTITIOUS) == 0) {
4779 	    count = pmap_pvh_wired_mappings(pa_to_pvh(VM_PAGE_TO_PHYS(m)),
4780 	        count);
4781 	}
4782 	rw_wunlock(&pvh_global_lock);
4783 	return (count);
4784 }
4785 
4786 /*
4787  *	pmap_pvh_wired_mappings:
4788  *
4789  *	Return the updated number "count" of managed mappings that are wired.
4790  */
4791 static int
pmap_pvh_wired_mappings(struct md_page * pvh,int count)4792 pmap_pvh_wired_mappings(struct md_page *pvh, int count)
4793 {
4794 	pmap_t pmap;
4795 	pt_entry_t *pte;
4796 	pv_entry_t pv;
4797 
4798 	rw_assert(&pvh_global_lock, RA_WLOCKED);
4799 	sched_pin();
4800 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
4801 		pmap = PV_PMAP(pv);
4802 		PMAP_LOCK(pmap);
4803 		pte = pmap_pte_quick(pmap, pv->pv_va);
4804 		if ((*pte & PG_W) != 0)
4805 			count++;
4806 		PMAP_UNLOCK(pmap);
4807 	}
4808 	sched_unpin();
4809 	return (count);
4810 }
4811 
4812 /*
4813  * Returns true if the given page is mapped individually or as part of
4814  * a 4mpage.  Otherwise, returns false.
4815  */
4816 static bool
__CONCAT(PMTYPE,page_is_mapped)4817 __CONCAT(PMTYPE, page_is_mapped)(vm_page_t m)
4818 {
4819 	bool rv;
4820 
4821 	if ((m->oflags & VPO_UNMANAGED) != 0)
4822 		return (false);
4823 	rw_wlock(&pvh_global_lock);
4824 	rv = !TAILQ_EMPTY(&m->md.pv_list) ||
4825 	    ((m->flags & PG_FICTITIOUS) == 0 &&
4826 	    !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list));
4827 	rw_wunlock(&pvh_global_lock);
4828 	return (rv);
4829 }
4830 
4831 /*
4832  * Remove all pages from specified address space
4833  * this aids process exit speeds.  Also, this code
4834  * is special cased for current process only, but
4835  * can have the more generic (and slightly slower)
4836  * mode enabled.  This is much faster than pmap_remove
4837  * in the case of running down an entire address space.
4838  */
4839 static void
__CONCAT(PMTYPE,remove_pages)4840 __CONCAT(PMTYPE, remove_pages)(pmap_t pmap)
4841 {
4842 	pt_entry_t *pte, tpte;
4843 	vm_page_t m, mpte, mt;
4844 	pv_entry_t pv;
4845 	struct md_page *pvh;
4846 	struct pv_chunk *pc, *npc;
4847 	struct spglist free;
4848 	int field, idx;
4849 	int32_t bit;
4850 	uint32_t inuse, bitmask;
4851 	int allfree;
4852 
4853 	if (pmap != PCPU_GET(curpmap)) {
4854 		printf("warning: pmap_remove_pages called with non-current pmap\n");
4855 		return;
4856 	}
4857 	SLIST_INIT(&free);
4858 	rw_wlock(&pvh_global_lock);
4859 	PMAP_LOCK(pmap);
4860 	sched_pin();
4861 	TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) {
4862 		KASSERT(pc->pc_pmap == pmap, ("Wrong pmap %p %p", pmap,
4863 		    pc->pc_pmap));
4864 		allfree = 1;
4865 		for (field = 0; field < _NPCM; field++) {
4866 			inuse = ~pc->pc_map[field] & pc_freemask[field];
4867 			while (inuse != 0) {
4868 				bit = bsfl(inuse);
4869 				bitmask = 1UL << bit;
4870 				idx = field * 32 + bit;
4871 				pv = &pc->pc_pventry[idx];
4872 				inuse &= ~bitmask;
4873 
4874 				pte = pmap_pde(pmap, pv->pv_va);
4875 				tpte = *pte;
4876 				if ((tpte & PG_PS) == 0) {
4877 					pte = pmap_pte_quick(pmap, pv->pv_va);
4878 					tpte = *pte & ~PG_PTE_PAT;
4879 				}
4880 
4881 				if (tpte == 0) {
4882 					printf(
4883 					    "TPTE at %p  IS ZERO @ VA %08x\n",
4884 					    pte, pv->pv_va);
4885 					panic("bad pte");
4886 				}
4887 
4888 /*
4889  * We cannot remove wired pages from a process' mapping at this time
4890  */
4891 				if (tpte & PG_W) {
4892 					allfree = 0;
4893 					continue;
4894 				}
4895 
4896 				m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
4897 				KASSERT(m->phys_addr == (tpte & PG_FRAME),
4898 				    ("vm_page_t %p phys_addr mismatch %016jx %016jx",
4899 				    m, (uintmax_t)m->phys_addr,
4900 				    (uintmax_t)tpte));
4901 
4902 				KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
4903 				    m < &vm_page_array[vm_page_array_size],
4904 				    ("pmap_remove_pages: bad tpte %#jx",
4905 				    (uintmax_t)tpte));
4906 
4907 				pte_clear(pte);
4908 
4909 				/*
4910 				 * Update the vm_page_t clean/reference bits.
4911 				 */
4912 				if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
4913 					if ((tpte & PG_PS) != 0) {
4914 						for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
4915 							vm_page_dirty(mt);
4916 					} else
4917 						vm_page_dirty(m);
4918 				}
4919 
4920 				/* Mark free */
4921 				PV_STAT(pv_entry_frees++);
4922 				PV_STAT(pv_entry_spare++);
4923 				pv_entry_count--;
4924 				pc->pc_map[field] |= bitmask;
4925 				if ((tpte & PG_PS) != 0) {
4926 					pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
4927 					pvh = pa_to_pvh(tpte & PG_PS_FRAME);
4928 					TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
4929 					if (TAILQ_EMPTY(&pvh->pv_list)) {
4930 						for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++)
4931 							if (TAILQ_EMPTY(&mt->md.pv_list))
4932 								vm_page_aflag_clear(mt, PGA_WRITEABLE);
4933 					}
4934 					mpte = pmap_remove_pt_page(pmap, pv->pv_va);
4935 					if (mpte != NULL) {
4936 						KASSERT(vm_page_any_valid(mpte),
4937 						    ("pmap_remove_pages: pte page not promoted"));
4938 						pmap->pm_stats.resident_count--;
4939 						KASSERT(mpte->ref_count == NPTEPG,
4940 						    ("pmap_remove_pages: pte page ref count error"));
4941 						mpte->ref_count = 0;
4942 						pmap_add_delayed_free_list(mpte, &free, false);
4943 					}
4944 				} else {
4945 					pmap->pm_stats.resident_count--;
4946 					TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
4947 					if (TAILQ_EMPTY(&m->md.pv_list) &&
4948 					    (m->flags & PG_FICTITIOUS) == 0) {
4949 						pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4950 						if (TAILQ_EMPTY(&pvh->pv_list))
4951 							vm_page_aflag_clear(m, PGA_WRITEABLE);
4952 					}
4953 					pmap_unuse_pt(pmap, pv->pv_va, &free);
4954 				}
4955 			}
4956 		}
4957 		if (allfree) {
4958 			TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
4959 			free_pv_chunk(pc);
4960 		}
4961 	}
4962 	sched_unpin();
4963 	pmap_invalidate_all_int(pmap);
4964 	rw_wunlock(&pvh_global_lock);
4965 	PMAP_UNLOCK(pmap);
4966 	vm_page_free_pages_toq(&free, true);
4967 }
4968 
4969 /*
4970  *	pmap_is_modified:
4971  *
4972  *	Return whether or not the specified physical page was modified
4973  *	in any physical maps.
4974  */
4975 static bool
__CONCAT(PMTYPE,is_modified)4976 __CONCAT(PMTYPE, is_modified)(vm_page_t m)
4977 {
4978 	bool rv;
4979 
4980 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4981 	    ("pmap_is_modified: page %p is not managed", m));
4982 
4983 	/*
4984 	 * If the page is not busied then this check is racy.
4985 	 */
4986 	if (!pmap_page_is_write_mapped(m))
4987 		return (false);
4988 	rw_wlock(&pvh_global_lock);
4989 	rv = pmap_is_modified_pvh(&m->md) ||
4990 	    ((m->flags & PG_FICTITIOUS) == 0 &&
4991 	    pmap_is_modified_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m))));
4992 	rw_wunlock(&pvh_global_lock);
4993 	return (rv);
4994 }
4995 
4996 /*
4997  * Returns true if any of the given mappings were used to modify
4998  * physical memory.  Otherwise, returns false.  Both page and 2mpage
4999  * mappings are supported.
5000  */
5001 static bool
pmap_is_modified_pvh(struct md_page * pvh)5002 pmap_is_modified_pvh(struct md_page *pvh)
5003 {
5004 	pv_entry_t pv;
5005 	pt_entry_t *pte;
5006 	pmap_t pmap;
5007 	bool rv;
5008 
5009 	rw_assert(&pvh_global_lock, RA_WLOCKED);
5010 	rv = false;
5011 	sched_pin();
5012 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
5013 		pmap = PV_PMAP(pv);
5014 		PMAP_LOCK(pmap);
5015 		pte = pmap_pte_quick(pmap, pv->pv_va);
5016 		rv = (*pte & (PG_M | PG_RW)) == (PG_M | PG_RW);
5017 		PMAP_UNLOCK(pmap);
5018 		if (rv)
5019 			break;
5020 	}
5021 	sched_unpin();
5022 	return (rv);
5023 }
5024 
5025 /*
5026  *	pmap_is_prefaultable:
5027  *
5028  *	Return whether or not the specified virtual address is elgible
5029  *	for prefault.
5030  */
5031 static bool
__CONCAT(PMTYPE,is_prefaultable)5032 __CONCAT(PMTYPE, is_prefaultable)(pmap_t pmap, vm_offset_t addr)
5033 {
5034 	pd_entry_t pde;
5035 	bool rv;
5036 
5037 	rv = false;
5038 	PMAP_LOCK(pmap);
5039 	pde = *pmap_pde(pmap, addr);
5040 	if (pde != 0 && (pde & PG_PS) == 0)
5041 		rv = pmap_pte_ufast(pmap, addr, pde) == 0;
5042 	PMAP_UNLOCK(pmap);
5043 	return (rv);
5044 }
5045 
5046 /*
5047  *	pmap_is_referenced:
5048  *
5049  *	Return whether or not the specified physical page was referenced
5050  *	in any physical maps.
5051  */
5052 static bool
__CONCAT(PMTYPE,is_referenced)5053 __CONCAT(PMTYPE, is_referenced)(vm_page_t m)
5054 {
5055 	bool rv;
5056 
5057 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5058 	    ("pmap_is_referenced: page %p is not managed", m));
5059 	rw_wlock(&pvh_global_lock);
5060 	rv = pmap_is_referenced_pvh(&m->md) ||
5061 	    ((m->flags & PG_FICTITIOUS) == 0 &&
5062 	    pmap_is_referenced_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m))));
5063 	rw_wunlock(&pvh_global_lock);
5064 	return (rv);
5065 }
5066 
5067 /*
5068  * Returns true if any of the given mappings were referenced and false
5069  * otherwise.  Both page and 4mpage mappings are supported.
5070  */
5071 static bool
pmap_is_referenced_pvh(struct md_page * pvh)5072 pmap_is_referenced_pvh(struct md_page *pvh)
5073 {
5074 	pv_entry_t pv;
5075 	pt_entry_t *pte;
5076 	pmap_t pmap;
5077 	bool rv;
5078 
5079 	rw_assert(&pvh_global_lock, RA_WLOCKED);
5080 	rv = false;
5081 	sched_pin();
5082 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) {
5083 		pmap = PV_PMAP(pv);
5084 		PMAP_LOCK(pmap);
5085 		pte = pmap_pte_quick(pmap, pv->pv_va);
5086 		rv = (*pte & (PG_A | PG_V)) == (PG_A | PG_V);
5087 		PMAP_UNLOCK(pmap);
5088 		if (rv)
5089 			break;
5090 	}
5091 	sched_unpin();
5092 	return (rv);
5093 }
5094 
5095 /*
5096  * Clear the write and modified bits in each of the given page's mappings.
5097  */
5098 static void
__CONCAT(PMTYPE,remove_write)5099 __CONCAT(PMTYPE, remove_write)(vm_page_t m)
5100 {
5101 	struct md_page *pvh;
5102 	pv_entry_t next_pv, pv;
5103 	pmap_t pmap;
5104 	pd_entry_t *pde;
5105 	pt_entry_t oldpte, *pte;
5106 	vm_offset_t va;
5107 
5108 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5109 	    ("pmap_remove_write: page %p is not managed", m));
5110 	vm_page_assert_busied(m);
5111 
5112 	if (!pmap_page_is_write_mapped(m))
5113 		return;
5114 	rw_wlock(&pvh_global_lock);
5115 	sched_pin();
5116 	if ((m->flags & PG_FICTITIOUS) != 0)
5117 		goto small_mappings;
5118 	pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
5119 	TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) {
5120 		va = pv->pv_va;
5121 		pmap = PV_PMAP(pv);
5122 		PMAP_LOCK(pmap);
5123 		pde = pmap_pde(pmap, va);
5124 		if ((*pde & PG_RW) != 0)
5125 			(void)pmap_demote_pde(pmap, pde, va);
5126 		PMAP_UNLOCK(pmap);
5127 	}
5128 small_mappings:
5129 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
5130 		pmap = PV_PMAP(pv);
5131 		PMAP_LOCK(pmap);
5132 		pde = pmap_pde(pmap, pv->pv_va);
5133 		KASSERT((*pde & PG_PS) == 0, ("pmap_clear_write: found"
5134 		    " a 4mpage in page %p's pv list", m));
5135 		pte = pmap_pte_quick(pmap, pv->pv_va);
5136 retry:
5137 		oldpte = *pte;
5138 		if ((oldpte & PG_RW) != 0) {
5139 			/*
5140 			 * Regardless of whether a pte is 32 or 64 bits
5141 			 * in size, PG_RW and PG_M are among the least
5142 			 * significant 32 bits.
5143 			 */
5144 			if (!atomic_cmpset_int((u_int *)pte, oldpte,
5145 			    oldpte & ~(PG_RW | PG_M)))
5146 				goto retry;
5147 			if ((oldpte & PG_M) != 0)
5148 				vm_page_dirty(m);
5149 			pmap_invalidate_page_int(pmap, pv->pv_va);
5150 		}
5151 		PMAP_UNLOCK(pmap);
5152 	}
5153 	vm_page_aflag_clear(m, PGA_WRITEABLE);
5154 	sched_unpin();
5155 	rw_wunlock(&pvh_global_lock);
5156 }
5157 
5158 /*
5159  *	pmap_ts_referenced:
5160  *
5161  *	Return a count of reference bits for a page, clearing those bits.
5162  *	It is not necessary for every reference bit to be cleared, but it
5163  *	is necessary that 0 only be returned when there are truly no
5164  *	reference bits set.
5165  *
5166  *	As an optimization, update the page's dirty field if a modified bit is
5167  *	found while counting reference bits.  This opportunistic update can be
5168  *	performed at low cost and can eliminate the need for some future calls
5169  *	to pmap_is_modified().  However, since this function stops after
5170  *	finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some
5171  *	dirty pages.  Those dirty pages will only be detected by a future call
5172  *	to pmap_is_modified().
5173  */
5174 static int
__CONCAT(PMTYPE,ts_referenced)5175 __CONCAT(PMTYPE, ts_referenced)(vm_page_t m)
5176 {
5177 	struct md_page *pvh;
5178 	pv_entry_t pv, pvf;
5179 	pmap_t pmap;
5180 	pd_entry_t *pde;
5181 	pt_entry_t *pte;
5182 	vm_paddr_t pa;
5183 	int rtval = 0;
5184 
5185 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5186 	    ("pmap_ts_referenced: page %p is not managed", m));
5187 	pa = VM_PAGE_TO_PHYS(m);
5188 	pvh = pa_to_pvh(pa);
5189 	rw_wlock(&pvh_global_lock);
5190 	sched_pin();
5191 	if ((m->flags & PG_FICTITIOUS) != 0 ||
5192 	    (pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL)
5193 		goto small_mappings;
5194 	pv = pvf;
5195 	do {
5196 		pmap = PV_PMAP(pv);
5197 		PMAP_LOCK(pmap);
5198 		pde = pmap_pde(pmap, pv->pv_va);
5199 		if ((*pde & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
5200 			/*
5201 			 * Although "*pde" is mapping a 2/4MB page, because
5202 			 * this function is called at a 4KB page granularity,
5203 			 * we only update the 4KB page under test.
5204 			 */
5205 			vm_page_dirty(m);
5206 		}
5207 		if ((*pde & PG_A) != 0) {
5208 			/*
5209 			 * Since this reference bit is shared by either 1024
5210 			 * or 512 4KB pages, it should not be cleared every
5211 			 * time it is tested.  Apply a simple "hash" function
5212 			 * on the physical page number, the virtual superpage
5213 			 * number, and the pmap address to select one 4KB page
5214 			 * out of the 1024 or 512 on which testing the
5215 			 * reference bit will result in clearing that bit.
5216 			 * This function is designed to avoid the selection of
5217 			 * the same 4KB page for every 2- or 4MB page mapping.
5218 			 *
5219 			 * On demotion, a mapping that hasn't been referenced
5220 			 * is simply destroyed.  To avoid the possibility of a
5221 			 * subsequent page fault on a demoted wired mapping,
5222 			 * always leave its reference bit set.  Moreover,
5223 			 * since the superpage is wired, the current state of
5224 			 * its reference bit won't affect page replacement.
5225 			 */
5226 			if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PDRSHIFT) ^
5227 			    (uintptr_t)pmap) & (NPTEPG - 1)) == 0 &&
5228 			    (*pde & PG_W) == 0) {
5229 				atomic_clear_int((u_int *)pde, PG_A);
5230 				pmap_invalidate_page_int(pmap, pv->pv_va);
5231 			}
5232 			rtval++;
5233 		}
5234 		PMAP_UNLOCK(pmap);
5235 		/* Rotate the PV list if it has more than one entry. */
5236 		if (TAILQ_NEXT(pv, pv_next) != NULL) {
5237 			TAILQ_REMOVE(&pvh->pv_list, pv, pv_next);
5238 			TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next);
5239 		}
5240 		if (rtval >= PMAP_TS_REFERENCED_MAX)
5241 			goto out;
5242 	} while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf);
5243 small_mappings:
5244 	if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL)
5245 		goto out;
5246 	pv = pvf;
5247 	do {
5248 		pmap = PV_PMAP(pv);
5249 		PMAP_LOCK(pmap);
5250 		pde = pmap_pde(pmap, pv->pv_va);
5251 		KASSERT((*pde & PG_PS) == 0,
5252 		    ("pmap_ts_referenced: found a 4mpage in page %p's pv list",
5253 		    m));
5254 		pte = pmap_pte_quick(pmap, pv->pv_va);
5255 		if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5256 			vm_page_dirty(m);
5257 		if ((*pte & PG_A) != 0) {
5258 			atomic_clear_int((u_int *)pte, PG_A);
5259 			pmap_invalidate_page_int(pmap, pv->pv_va);
5260 			rtval++;
5261 		}
5262 		PMAP_UNLOCK(pmap);
5263 		/* Rotate the PV list if it has more than one entry. */
5264 		if (TAILQ_NEXT(pv, pv_next) != NULL) {
5265 			TAILQ_REMOVE(&m->md.pv_list, pv, pv_next);
5266 			TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next);
5267 		}
5268 	} while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && rtval <
5269 	    PMAP_TS_REFERENCED_MAX);
5270 out:
5271 	sched_unpin();
5272 	rw_wunlock(&pvh_global_lock);
5273 	return (rtval);
5274 }
5275 
5276 /*
5277  *	Apply the given advice to the specified range of addresses within the
5278  *	given pmap.  Depending on the advice, clear the referenced and/or
5279  *	modified flags in each mapping and set the mapped page's dirty field.
5280  */
5281 static void
__CONCAT(PMTYPE,advise)5282 __CONCAT(PMTYPE, advise)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
5283     int advice)
5284 {
5285 	pd_entry_t oldpde, *pde;
5286 	pt_entry_t *pte;
5287 	vm_offset_t va, pdnxt;
5288 	vm_page_t m;
5289 	bool anychanged, pv_lists_locked;
5290 
5291 	if (advice != MADV_DONTNEED && advice != MADV_FREE)
5292 		return;
5293 	if (pmap_is_current(pmap))
5294 		pv_lists_locked = false;
5295 	else {
5296 		pv_lists_locked = true;
5297 resume:
5298 		rw_wlock(&pvh_global_lock);
5299 		sched_pin();
5300 	}
5301 	anychanged = false;
5302 	PMAP_LOCK(pmap);
5303 	for (; sva < eva; sva = pdnxt) {
5304 		pdnxt = (sva + NBPDR) & ~PDRMASK;
5305 		if (pdnxt < sva)
5306 			pdnxt = eva;
5307 		pde = pmap_pde(pmap, sva);
5308 		oldpde = *pde;
5309 		if ((oldpde & PG_V) == 0)
5310 			continue;
5311 		else if ((oldpde & PG_PS) != 0) {
5312 			if ((oldpde & PG_MANAGED) == 0)
5313 				continue;
5314 			if (!pv_lists_locked) {
5315 				pv_lists_locked = true;
5316 				if (!rw_try_wlock(&pvh_global_lock)) {
5317 					if (anychanged)
5318 						pmap_invalidate_all_int(pmap);
5319 					PMAP_UNLOCK(pmap);
5320 					goto resume;
5321 				}
5322 				sched_pin();
5323 			}
5324 			if (!pmap_demote_pde(pmap, pde, sva)) {
5325 				/*
5326 				 * The large page mapping was destroyed.
5327 				 */
5328 				continue;
5329 			}
5330 
5331 			/*
5332 			 * Unless the page mappings are wired, remove the
5333 			 * mapping to a single page so that a subsequent
5334 			 * access may repromote.  Choosing the last page
5335 			 * within the address range [sva, min(pdnxt, eva))
5336 			 * generally results in more repromotions.  Since the
5337 			 * underlying page table page is fully populated, this
5338 			 * removal never frees a page table page.
5339 			 */
5340 			if ((oldpde & PG_W) == 0) {
5341 				va = eva;
5342 				if (va > pdnxt)
5343 					va = pdnxt;
5344 				va -= PAGE_SIZE;
5345 				KASSERT(va >= sva,
5346 				    ("pmap_advise: no address gap"));
5347 				pte = pmap_pte_quick(pmap, va);
5348 				KASSERT((*pte & PG_V) != 0,
5349 				    ("pmap_advise: invalid PTE"));
5350 				pmap_remove_pte(pmap, pte, va, NULL);
5351 				anychanged = true;
5352 			}
5353 		}
5354 		if (pdnxt > eva)
5355 			pdnxt = eva;
5356 		va = pdnxt;
5357 		for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
5358 		    sva += PAGE_SIZE) {
5359 			if ((*pte & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V))
5360 				goto maybe_invlrng;
5361 			else if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
5362 				if (advice == MADV_DONTNEED) {
5363 					/*
5364 					 * Future calls to pmap_is_modified()
5365 					 * can be avoided by making the page
5366 					 * dirty now.
5367 					 */
5368 					m = PHYS_TO_VM_PAGE(*pte & PG_FRAME);
5369 					vm_page_dirty(m);
5370 				}
5371 				atomic_clear_int((u_int *)pte, PG_M | PG_A);
5372 			} else if ((*pte & PG_A) != 0)
5373 				atomic_clear_int((u_int *)pte, PG_A);
5374 			else
5375 				goto maybe_invlrng;
5376 			if ((*pte & PG_G) != 0) {
5377 				if (va == pdnxt)
5378 					va = sva;
5379 			} else
5380 				anychanged = true;
5381 			continue;
5382 maybe_invlrng:
5383 			if (va != pdnxt) {
5384 				pmap_invalidate_range_int(pmap, va, sva);
5385 				va = pdnxt;
5386 			}
5387 		}
5388 		if (va != pdnxt)
5389 			pmap_invalidate_range_int(pmap, va, sva);
5390 	}
5391 	if (anychanged)
5392 		pmap_invalidate_all_int(pmap);
5393 	if (pv_lists_locked) {
5394 		sched_unpin();
5395 		rw_wunlock(&pvh_global_lock);
5396 	}
5397 	PMAP_UNLOCK(pmap);
5398 }
5399 
5400 /*
5401  *	Clear the modify bits on the specified physical page.
5402  */
5403 static void
__CONCAT(PMTYPE,clear_modify)5404 __CONCAT(PMTYPE, clear_modify)(vm_page_t m)
5405 {
5406 	struct md_page *pvh;
5407 	pv_entry_t next_pv, pv;
5408 	pmap_t pmap;
5409 	pd_entry_t oldpde, *pde;
5410 	pt_entry_t *pte;
5411 	vm_offset_t va;
5412 
5413 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5414 	    ("pmap_clear_modify: page %p is not managed", m));
5415 	vm_page_assert_busied(m);
5416 
5417 	if (!pmap_page_is_write_mapped(m))
5418 		return;
5419 	rw_wlock(&pvh_global_lock);
5420 	sched_pin();
5421 	if ((m->flags & PG_FICTITIOUS) != 0)
5422 		goto small_mappings;
5423 	pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
5424 	TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) {
5425 		va = pv->pv_va;
5426 		pmap = PV_PMAP(pv);
5427 		PMAP_LOCK(pmap);
5428 		pde = pmap_pde(pmap, va);
5429 		oldpde = *pde;
5430 		/* If oldpde has PG_RW set, then it also has PG_M set. */
5431 		if ((oldpde & PG_RW) != 0 &&
5432 		    pmap_demote_pde(pmap, pde, va) &&
5433 		    (oldpde & PG_W) == 0) {
5434 			/*
5435 			 * Write protect the mapping to a single page so that
5436 			 * a subsequent write access may repromote.
5437 			 */
5438 			va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME);
5439 			pte = pmap_pte_quick(pmap, va);
5440 			/*
5441 			 * Regardless of whether a pte is 32 or 64 bits
5442 			 * in size, PG_RW and PG_M are among the least
5443 			 * significant 32 bits.
5444 			 */
5445 			atomic_clear_int((u_int *)pte, PG_M | PG_RW);
5446 			vm_page_dirty(m);
5447 			pmap_invalidate_page_int(pmap, va);
5448 		}
5449 		PMAP_UNLOCK(pmap);
5450 	}
5451 small_mappings:
5452 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) {
5453 		pmap = PV_PMAP(pv);
5454 		PMAP_LOCK(pmap);
5455 		pde = pmap_pde(pmap, pv->pv_va);
5456 		KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found"
5457 		    " a 4mpage in page %p's pv list", m));
5458 		pte = pmap_pte_quick(pmap, pv->pv_va);
5459 		if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
5460 			/*
5461 			 * Regardless of whether a pte is 32 or 64 bits
5462 			 * in size, PG_M is among the least significant
5463 			 * 32 bits.
5464 			 */
5465 			atomic_clear_int((u_int *)pte, PG_M);
5466 			pmap_invalidate_page_int(pmap, pv->pv_va);
5467 		}
5468 		PMAP_UNLOCK(pmap);
5469 	}
5470 	sched_unpin();
5471 	rw_wunlock(&pvh_global_lock);
5472 }
5473 
5474 /*
5475  * Miscellaneous support routines follow
5476  */
5477 
5478 /* Adjust the cache mode for a 4KB page mapped via a PTE. */
5479 static __inline void
pmap_pte_attr(pt_entry_t * pte,int cache_bits)5480 pmap_pte_attr(pt_entry_t *pte, int cache_bits)
5481 {
5482 	u_int opte, npte;
5483 
5484 	/*
5485 	 * The cache mode bits are all in the low 32-bits of the
5486 	 * PTE, so we can just spin on updating the low 32-bits.
5487 	 */
5488 	do {
5489 		opte = *(u_int *)pte;
5490 		npte = opte & ~PG_PTE_CACHE;
5491 		npte |= cache_bits;
5492 	} while (npte != opte && !atomic_cmpset_int((u_int *)pte, opte, npte));
5493 }
5494 
5495 /* Adjust the cache mode for a 2/4MB page mapped via a PDE. */
5496 static __inline void
pmap_pde_attr(pd_entry_t * pde,int cache_bits)5497 pmap_pde_attr(pd_entry_t *pde, int cache_bits)
5498 {
5499 	u_int opde, npde;
5500 
5501 	/*
5502 	 * The cache mode bits are all in the low 32-bits of the
5503 	 * PDE, so we can just spin on updating the low 32-bits.
5504 	 */
5505 	do {
5506 		opde = *(u_int *)pde;
5507 		npde = opde & ~PG_PDE_CACHE;
5508 		npde |= cache_bits;
5509 	} while (npde != opde && !atomic_cmpset_int((u_int *)pde, opde, npde));
5510 }
5511 
5512 /*
5513  * Map a set of physical memory pages into the kernel virtual
5514  * address space. Return a pointer to where it is mapped. This
5515  * routine is intended to be used for mapping device memory,
5516  * NOT real memory.
5517  */
5518 static void *
__CONCAT(PMTYPE,mapdev_attr)5519 __CONCAT(PMTYPE, mapdev_attr)(vm_paddr_t pa, vm_size_t size, int mode,
5520     int flags)
5521 {
5522 	struct pmap_preinit_mapping *ppim;
5523 	vm_offset_t va, offset;
5524 	vm_page_t m;
5525 	vm_size_t tmpsize;
5526 	int i;
5527 
5528 	offset = pa & PAGE_MASK;
5529 	size = round_page(offset + size);
5530 	pa = pa & PG_FRAME;
5531 
5532 	if (pa < PMAP_MAP_LOW && pa + size <= PMAP_MAP_LOW) {
5533 		va = pa + PMAP_MAP_LOW;
5534 		if ((flags & MAPDEV_SETATTR) == 0)
5535 			return ((void *)(va + offset));
5536 	} else if (!pmap_initialized) {
5537 		va = 0;
5538 		for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
5539 			ppim = pmap_preinit_mapping + i;
5540 			if (ppim->va == 0) {
5541 				ppim->pa = pa;
5542 				ppim->sz = size;
5543 				ppim->mode = mode;
5544 				ppim->va = virtual_avail;
5545 				virtual_avail += size;
5546 				va = ppim->va;
5547 				break;
5548 			}
5549 		}
5550 		if (va == 0)
5551 			panic("%s: too many preinit mappings", __func__);
5552 	} else {
5553 		/*
5554 		 * If we have a preinit mapping, re-use it.
5555 		 */
5556 		for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
5557 			ppim = pmap_preinit_mapping + i;
5558 			if (ppim->pa == pa && ppim->sz == size &&
5559 			    (ppim->mode == mode ||
5560 			    (flags & MAPDEV_SETATTR) == 0))
5561 				return ((void *)(ppim->va + offset));
5562 		}
5563 		va = kva_alloc(size);
5564 		if (va == 0)
5565 			panic("%s: Couldn't allocate KVA", __func__);
5566 	}
5567 	for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE) {
5568 		if ((flags & MAPDEV_SETATTR) == 0 && pmap_initialized) {
5569 			m = PHYS_TO_VM_PAGE(pa);
5570 			if (m != NULL && VM_PAGE_TO_PHYS(m) == pa) {
5571 				pmap_kenter_attr(va + tmpsize, pa + tmpsize,
5572 				    m->md.pat_mode);
5573 				continue;
5574 			}
5575 		}
5576 		pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode);
5577 	}
5578 	pmap_invalidate_range_int(kernel_pmap, va, va + tmpsize);
5579 	pmap_invalidate_cache_range(va, va + size);
5580 	return ((void *)(va + offset));
5581 }
5582 
5583 static void
__CONCAT(PMTYPE,unmapdev)5584 __CONCAT(PMTYPE, unmapdev)(void *p, vm_size_t size)
5585 {
5586 	struct pmap_preinit_mapping *ppim;
5587 	vm_offset_t offset, va;
5588 	int i;
5589 
5590 	va = (vm_offset_t)p;
5591 	if (va >= PMAP_MAP_LOW && va <= KERNBASE && va + size <= KERNBASE)
5592 		return;
5593 	offset = va & PAGE_MASK;
5594 	size = round_page(offset + size);
5595 	va = trunc_page(va);
5596 	for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) {
5597 		ppim = pmap_preinit_mapping + i;
5598 		if (ppim->va == va && ppim->sz == size) {
5599 			if (pmap_initialized)
5600 				return;
5601 			ppim->pa = 0;
5602 			ppim->va = 0;
5603 			ppim->sz = 0;
5604 			ppim->mode = 0;
5605 			if (va + size == virtual_avail)
5606 				virtual_avail = va;
5607 			return;
5608 		}
5609 	}
5610 	if (pmap_initialized) {
5611 		pmap_qremove(va, atop(size));
5612 		kva_free(va, size);
5613 	}
5614 }
5615 
5616 /*
5617  * Sets the memory attribute for the specified page.
5618  */
5619 static void
__CONCAT(PMTYPE,page_set_memattr)5620 __CONCAT(PMTYPE, page_set_memattr)(vm_page_t m, vm_memattr_t ma)
5621 {
5622 	if (m->md.pat_mode == ma)
5623 		return;
5624 
5625 	m->md.pat_mode = ma;
5626 	if ((m->flags & PG_FICTITIOUS) != 0)
5627 		return;
5628 
5629 	/*
5630 	 * If "m" is a normal page, flush it from the cache.
5631 	 * See pmap_invalidate_cache_range().
5632 	 *
5633 	 * First, try to find an existing mapping of the page by sf
5634 	 * buffer. sf_buf_invalidate_cache() modifies mapping and
5635 	 * flushes the cache.
5636 	 */
5637 	if (sf_buf_invalidate_cache(m))
5638 		return;
5639 
5640 	/*
5641 	 * If page is not mapped by sf buffer, but CPU does not
5642 	 * support self snoop, map the page transient and do
5643 	 * invalidation. In the worst case, whole cache is flushed by
5644 	 * pmap_invalidate_cache_range().
5645 	 */
5646 	if ((cpu_feature & CPUID_SS) == 0)
5647 		pmap_flush_page(m);
5648 }
5649 
5650 static void
__CONCAT(PMTYPE,flush_page)5651 __CONCAT(PMTYPE, flush_page)(vm_page_t m)
5652 {
5653 	pt_entry_t *cmap_pte2;
5654 	struct pcpu *pc;
5655 	vm_offset_t sva, eva;
5656 	bool useclflushopt;
5657 
5658 	useclflushopt = (cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0;
5659 	if (useclflushopt || (cpu_feature & CPUID_CLFSH) != 0) {
5660 		sched_pin();
5661 		pc = get_pcpu();
5662 		cmap_pte2 = pc->pc_cmap_pte2;
5663 		mtx_lock(&pc->pc_cmap_lock);
5664 		if (*cmap_pte2)
5665 			panic("pmap_flush_page: CMAP2 busy");
5666 		*cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) |
5667 		    PG_A | PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode,
5668 		    false);
5669 		invlcaddr(pc->pc_cmap_addr2);
5670 		sva = (vm_offset_t)pc->pc_cmap_addr2;
5671 		eva = sva + PAGE_SIZE;
5672 
5673 		/*
5674 		 * Use mfence or sfence despite the ordering implied by
5675 		 * mtx_{un,}lock() because clflush on non-Intel CPUs
5676 		 * and clflushopt are not guaranteed to be ordered by
5677 		 * any other instruction.
5678 		 */
5679 		if (useclflushopt)
5680 			sfence();
5681 		else if (cpu_vendor_id != CPU_VENDOR_INTEL)
5682 			mfence();
5683 		for (; sva < eva; sva += cpu_clflush_line_size) {
5684 			if (useclflushopt)
5685 				clflushopt(sva);
5686 			else
5687 				clflush(sva);
5688 		}
5689 		if (useclflushopt)
5690 			sfence();
5691 		else if (cpu_vendor_id != CPU_VENDOR_INTEL)
5692 			mfence();
5693 		*cmap_pte2 = 0;
5694 		sched_unpin();
5695 		mtx_unlock(&pc->pc_cmap_lock);
5696 	} else
5697 		pmap_invalidate_cache();
5698 }
5699 
5700 /*
5701  * Changes the specified virtual address range's memory type to that given by
5702  * the parameter "mode".  The specified virtual address range must be
5703  * completely contained within either the kernel map.
5704  *
5705  * Returns zero if the change completed successfully, and either EINVAL or
5706  * ENOMEM if the change failed.  Specifically, EINVAL is returned if some part
5707  * of the virtual address range was not mapped, and ENOMEM is returned if
5708  * there was insufficient memory available to complete the change.
5709  */
5710 static int
__CONCAT(PMTYPE,change_attr)5711 __CONCAT(PMTYPE, change_attr)(vm_offset_t va, vm_size_t size, int mode)
5712 {
5713 	vm_offset_t base, offset, tmpva;
5714 	pd_entry_t *pde;
5715 	pt_entry_t *pte;
5716 	int cache_bits_pte, cache_bits_pde;
5717 	bool changed;
5718 
5719 	base = trunc_page(va);
5720 	offset = va & PAGE_MASK;
5721 	size = round_page(offset + size);
5722 
5723 	/*
5724 	 * Only supported on kernel virtual addresses above the recursive map.
5725 	 */
5726 	if (base < VM_MIN_KERNEL_ADDRESS)
5727 		return (EINVAL);
5728 
5729 	cache_bits_pde = pmap_cache_bits(kernel_pmap, mode, true);
5730 	cache_bits_pte = pmap_cache_bits(kernel_pmap, mode, false);
5731 	changed = false;
5732 
5733 	/*
5734 	 * Pages that aren't mapped aren't supported.  Also break down
5735 	 * 2/4MB pages into 4KB pages if required.
5736 	 */
5737 	PMAP_LOCK(kernel_pmap);
5738 	for (tmpva = base; tmpva < base + size; ) {
5739 		pde = pmap_pde(kernel_pmap, tmpva);
5740 		if (*pde == 0) {
5741 			PMAP_UNLOCK(kernel_pmap);
5742 			return (EINVAL);
5743 		}
5744 		if (*pde & PG_PS) {
5745 			/*
5746 			 * If the current 2/4MB page already has
5747 			 * the required memory type, then we need not
5748 			 * demote this page.  Just increment tmpva to
5749 			 * the next 2/4MB page frame.
5750 			 */
5751 			if ((*pde & PG_PDE_CACHE) == cache_bits_pde) {
5752 				tmpva = trunc_4mpage(tmpva) + NBPDR;
5753 				continue;
5754 			}
5755 
5756 			/*
5757 			 * If the current offset aligns with a 2/4MB
5758 			 * page frame and there is at least 2/4MB left
5759 			 * within the range, then we need not break
5760 			 * down this page into 4KB pages.
5761 			 */
5762 			if ((tmpva & PDRMASK) == 0 &&
5763 			    tmpva + PDRMASK < base + size) {
5764 				tmpva += NBPDR;
5765 				continue;
5766 			}
5767 			if (!pmap_demote_pde(kernel_pmap, pde, tmpva)) {
5768 				PMAP_UNLOCK(kernel_pmap);
5769 				return (ENOMEM);
5770 			}
5771 		}
5772 		pte = vtopte(tmpva);
5773 		if (*pte == 0) {
5774 			PMAP_UNLOCK(kernel_pmap);
5775 			return (EINVAL);
5776 		}
5777 		tmpva += PAGE_SIZE;
5778 	}
5779 	PMAP_UNLOCK(kernel_pmap);
5780 
5781 	/*
5782 	 * Ok, all the pages exist, so run through them updating their
5783 	 * cache mode if required.
5784 	 */
5785 	for (tmpva = base; tmpva < base + size; ) {
5786 		pde = pmap_pde(kernel_pmap, tmpva);
5787 		if (*pde & PG_PS) {
5788 			if ((*pde & PG_PDE_CACHE) != cache_bits_pde) {
5789 				pmap_pde_attr(pde, cache_bits_pde);
5790 				changed = true;
5791 			}
5792 			tmpva = trunc_4mpage(tmpva) + NBPDR;
5793 		} else {
5794 			pte = vtopte(tmpva);
5795 			if ((*pte & PG_PTE_CACHE) != cache_bits_pte) {
5796 				pmap_pte_attr(pte, cache_bits_pte);
5797 				changed = true;
5798 			}
5799 			tmpva += PAGE_SIZE;
5800 		}
5801 	}
5802 
5803 	/*
5804 	 * Flush CPU caches to make sure any data isn't cached that
5805 	 * shouldn't be, etc.
5806 	 */
5807 	if (changed) {
5808 		pmap_invalidate_range_int(kernel_pmap, base, tmpva);
5809 		pmap_invalidate_cache_range(base, tmpva);
5810 	}
5811 	return (0);
5812 }
5813 
5814 /*
5815  * Perform the pmap work for mincore(2).  If the page is not both referenced and
5816  * modified by this pmap, returns its physical address so that the caller can
5817  * find other mappings.
5818  */
5819 static int
__CONCAT(PMTYPE,mincore)5820 __CONCAT(PMTYPE, mincore)(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap)
5821 {
5822 	pd_entry_t pde;
5823 	pt_entry_t pte;
5824 	vm_paddr_t pa;
5825 	int val;
5826 
5827 	PMAP_LOCK(pmap);
5828 	pde = *pmap_pde(pmap, addr);
5829 	if (pde != 0) {
5830 		if ((pde & PG_PS) != 0) {
5831 			pte = pde;
5832 			/* Compute the physical address of the 4KB page. */
5833 			pa = ((pde & PG_PS_FRAME) | (addr & PDRMASK)) &
5834 			    PG_FRAME;
5835 			val = MINCORE_PSIND(1);
5836 		} else {
5837 			pte = pmap_pte_ufast(pmap, addr, pde);
5838 			pa = pte & PG_FRAME;
5839 			val = 0;
5840 		}
5841 	} else {
5842 		pte = 0;
5843 		pa = 0;
5844 		val = 0;
5845 	}
5846 	if ((pte & PG_V) != 0) {
5847 		val |= MINCORE_INCORE;
5848 		if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5849 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
5850 		if ((pte & PG_A) != 0)
5851 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
5852 	}
5853 	if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
5854 	    (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
5855 	    (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) {
5856 		*pap = pa;
5857 	}
5858 	PMAP_UNLOCK(pmap);
5859 	return (val);
5860 }
5861 
5862 static void
__CONCAT(PMTYPE,activate)5863 __CONCAT(PMTYPE, activate)(struct thread *td)
5864 {
5865 	pmap_t	pmap, oldpmap;
5866 	u_int	cpuid;
5867 	u_int32_t  cr3;
5868 
5869 	critical_enter();
5870 	pmap = vmspace_pmap(td->td_proc->p_vmspace);
5871 	oldpmap = PCPU_GET(curpmap);
5872 	cpuid = PCPU_GET(cpuid);
5873 #if defined(SMP)
5874 	CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active);
5875 	CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
5876 #else
5877 	CPU_CLR(cpuid, &oldpmap->pm_active);
5878 	CPU_SET(cpuid, &pmap->pm_active);
5879 #endif
5880 #ifdef PMAP_PAE_COMP
5881 	cr3 = vtophys(pmap->pm_pdpt);
5882 #else
5883 	cr3 = vtophys(pmap->pm_pdir);
5884 #endif
5885 	/*
5886 	 * pmap_activate is for the current thread on the current cpu
5887 	 */
5888 	td->td_pcb->pcb_cr3 = cr3;
5889 	PCPU_SET(curpmap, pmap);
5890 	critical_exit();
5891 }
5892 
5893 static void
__CONCAT(PMTYPE,activate_boot)5894 __CONCAT(PMTYPE, activate_boot)(pmap_t pmap)
5895 {
5896 	u_int cpuid;
5897 
5898 	cpuid = PCPU_GET(cpuid);
5899 #if defined(SMP)
5900 	CPU_SET_ATOMIC(cpuid, &pmap->pm_active);
5901 #else
5902 	CPU_SET(cpuid, &pmap->pm_active);
5903 #endif
5904 	PCPU_SET(curpmap, pmap);
5905 }
5906 
5907 /*
5908  *	Increase the starting virtual address of the given mapping if a
5909  *	different alignment might result in more superpage mappings.
5910  */
5911 static void
__CONCAT(PMTYPE,align_superpage)5912 __CONCAT(PMTYPE, align_superpage)(vm_object_t object, vm_ooffset_t offset,
5913     vm_offset_t *addr, vm_size_t size)
5914 {
5915 	vm_offset_t superpage_offset;
5916 
5917 	if (size < NBPDR)
5918 		return;
5919 	if (object != NULL && (object->flags & OBJ_COLORED) != 0)
5920 		offset += ptoa(object->pg_color);
5921 	superpage_offset = offset & PDRMASK;
5922 	if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR ||
5923 	    (*addr & PDRMASK) == superpage_offset)
5924 		return;
5925 	if ((*addr & PDRMASK) < superpage_offset)
5926 		*addr = (*addr & ~PDRMASK) + superpage_offset;
5927 	else
5928 		*addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset;
5929 }
5930 
5931 static vm_offset_t
__CONCAT(PMTYPE,quick_enter_page)5932 __CONCAT(PMTYPE, quick_enter_page)(vm_page_t m)
5933 {
5934 	vm_offset_t qaddr;
5935 	pt_entry_t *pte;
5936 
5937 	critical_enter();
5938 	qaddr = PCPU_GET(qmap_addr);
5939 	pte = vtopte(qaddr);
5940 
5941 	KASSERT(*pte == 0,
5942 	    ("pmap_quick_enter_page: PTE busy %#jx", (uintmax_t)*pte));
5943 	*pte = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M |
5944 	    pmap_cache_bits(kernel_pmap, pmap_page_get_memattr(m), false);
5945 	invlpg(qaddr);
5946 
5947 	return (qaddr);
5948 }
5949 
5950 static void
__CONCAT(PMTYPE,quick_remove_page)5951 __CONCAT(PMTYPE, quick_remove_page)(vm_offset_t addr)
5952 {
5953 	vm_offset_t qaddr;
5954 	pt_entry_t *pte;
5955 
5956 	qaddr = PCPU_GET(qmap_addr);
5957 	pte = vtopte(qaddr);
5958 
5959 	KASSERT(*pte != 0, ("pmap_quick_remove_page: PTE not in use"));
5960 	KASSERT(addr == qaddr, ("pmap_quick_remove_page: invalid address"));
5961 
5962 	*pte = 0;
5963 	critical_exit();
5964 }
5965 
5966 static vmem_t *pmap_trm_arena;
5967 static vmem_addr_t pmap_trm_arena_last = PMAP_TRM_MIN_ADDRESS;
5968 static int trm_guard = PAGE_SIZE;
5969 
5970 static int
pmap_trm_import(void * unused __unused,vmem_size_t size,int flags,vmem_addr_t * addrp)5971 pmap_trm_import(void *unused __unused, vmem_size_t size, int flags,
5972     vmem_addr_t *addrp)
5973 {
5974 	vm_page_t m;
5975 	vmem_addr_t af, addr, prev_addr;
5976 	pt_entry_t *trm_pte;
5977 
5978 	prev_addr = atomic_load_int(&pmap_trm_arena_last);
5979 	size = round_page(size) + trm_guard;
5980 	for (;;) {
5981 		if (prev_addr + size < prev_addr || prev_addr + size < size ||
5982 		    prev_addr + size > PMAP_TRM_MAX_ADDRESS)
5983 			return (ENOMEM);
5984 		addr = prev_addr + size;
5985 		if (atomic_fcmpset_int(&pmap_trm_arena_last, &prev_addr, addr))
5986 			break;
5987 	}
5988 	prev_addr += trm_guard;
5989 	trm_pte = PTmap + atop(prev_addr);
5990 	for (af = prev_addr; af < addr; af += PAGE_SIZE) {
5991 		m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_WAITOK);
5992 		pte_store(&trm_pte[atop(af - prev_addr)], VM_PAGE_TO_PHYS(m) |
5993 		    PG_M | PG_A | PG_RW | PG_V | pgeflag |
5994 		    pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, false));
5995 	}
5996 	*addrp = prev_addr;
5997 	return (0);
5998 }
5999 
6000 void
pmap_init_trm(void)6001 pmap_init_trm(void)
6002 {
6003 	vm_page_t pd_m;
6004 
6005 	TUNABLE_INT_FETCH("machdep.trm_guard", &trm_guard);
6006 	if ((trm_guard & PAGE_MASK) != 0)
6007 		trm_guard = 0;
6008 	pmap_trm_arena = vmem_create("i386trampoline", 0, 0, 1, 0, M_WAITOK);
6009 	vmem_set_import(pmap_trm_arena, pmap_trm_import, NULL, NULL, PAGE_SIZE);
6010 	pd_m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_WAITOK |
6011 	    VM_ALLOC_ZERO);
6012 	PTD[TRPTDI] = VM_PAGE_TO_PHYS(pd_m) | PG_M | PG_A | PG_RW | PG_V |
6013 	    pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, true);
6014 }
6015 
6016 static void *
__CONCAT(PMTYPE,trm_alloc)6017 __CONCAT(PMTYPE, trm_alloc)(size_t size, int flags)
6018 {
6019 	vmem_addr_t res;
6020 	int error;
6021 
6022 	MPASS((flags & ~(M_WAITOK | M_NOWAIT | M_ZERO)) == 0);
6023 	error = vmem_xalloc(pmap_trm_arena, roundup2(size, 4), sizeof(int),
6024 	    0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags | M_FIRSTFIT, &res);
6025 	if (error != 0)
6026 		return (NULL);
6027 	if ((flags & M_ZERO) != 0)
6028 		bzero((void *)res, size);
6029 	return ((void *)res);
6030 }
6031 
6032 static void
__CONCAT(PMTYPE,trm_free)6033 __CONCAT(PMTYPE, trm_free)(void *addr, size_t size)
6034 {
6035 
6036 	vmem_free(pmap_trm_arena, (uintptr_t)addr, roundup2(size, 4));
6037 }
6038 
6039 static void
__CONCAT(PMTYPE,ksetrw)6040 __CONCAT(PMTYPE, ksetrw)(vm_offset_t va)
6041 {
6042 
6043 	*vtopte(va) |= PG_RW;
6044 }
6045 
6046 static void
__CONCAT(PMTYPE,remap_lowptdi)6047 __CONCAT(PMTYPE, remap_lowptdi)(bool enable)
6048 {
6049 
6050 	PTD[KPTDI] = enable ? PTD[LOWPTDI] : 0;
6051 	invltlb_glob();
6052 }
6053 
6054 static vm_offset_t
__CONCAT(PMTYPE,get_map_low)6055 __CONCAT(PMTYPE, get_map_low)(void)
6056 {
6057 
6058 	return (PMAP_MAP_LOW);
6059 }
6060 
6061 static vm_offset_t
__CONCAT(PMTYPE,get_vm_maxuser_address)6062 __CONCAT(PMTYPE, get_vm_maxuser_address)(void)
6063 {
6064 
6065 	return (VM_MAXUSER_ADDRESS);
6066 }
6067 
6068 static vm_paddr_t
__CONCAT(PMTYPE,pg_frame)6069 __CONCAT(PMTYPE, pg_frame)(vm_paddr_t pa)
6070 {
6071 
6072 	return (pa & PG_FRAME);
6073 }
6074 
6075 static void
__CONCAT(PMTYPE,sf_buf_map)6076 __CONCAT(PMTYPE, sf_buf_map)(struct sf_buf *sf)
6077 {
6078 	pt_entry_t opte, *ptep;
6079 
6080 	/*
6081 	 * Update the sf_buf's virtual-to-physical mapping, flushing the
6082 	 * virtual address from the TLB.  Since the reference count for
6083 	 * the sf_buf's old mapping was zero, that mapping is not
6084 	 * currently in use.  Consequently, there is no need to exchange
6085 	 * the old and new PTEs atomically, even under PAE.
6086 	 */
6087 	ptep = vtopte(sf->kva);
6088 	opte = *ptep;
6089 	*ptep = VM_PAGE_TO_PHYS(sf->m) | PG_RW | PG_V |
6090 	    pmap_cache_bits(kernel_pmap, sf->m->md.pat_mode, false);
6091 
6092 	/*
6093 	 * Avoid unnecessary TLB invalidations: If the sf_buf's old
6094 	 * virtual-to-physical mapping was not used, then any processor
6095 	 * that has invalidated the sf_buf's virtual address from its TLB
6096 	 * since the last used mapping need not invalidate again.
6097 	 */
6098 #ifdef SMP
6099 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
6100 		CPU_ZERO(&sf->cpumask);
6101 #else
6102 	if ((opte & (PG_V | PG_A)) ==  (PG_V | PG_A))
6103 		pmap_invalidate_page_int(kernel_pmap, sf->kva);
6104 #endif
6105 }
6106 
6107 static void
__CONCAT(PMTYPE,cp_slow0_map)6108 __CONCAT(PMTYPE, cp_slow0_map)(vm_offset_t kaddr, int plen, vm_page_t *ma)
6109 {
6110 	pt_entry_t *pte;
6111 	int i;
6112 
6113 	for (i = 0, pte = vtopte(kaddr); i < plen; i++, pte++) {
6114 		*pte = PG_V | PG_RW | PG_A | PG_M | VM_PAGE_TO_PHYS(ma[i]) |
6115 		    pmap_cache_bits(kernel_pmap, pmap_page_get_memattr(ma[i]),
6116 		    false);
6117 		invlpg(kaddr + ptoa(i));
6118 	}
6119 }
6120 
6121 static u_int
__CONCAT(PMTYPE,get_kcr3)6122 __CONCAT(PMTYPE, get_kcr3)(void)
6123 {
6124 
6125 #ifdef PMAP_PAE_COMP
6126 	return ((u_int)IdlePDPT);
6127 #else
6128 	return ((u_int)IdlePTD);
6129 #endif
6130 }
6131 
6132 static u_int
__CONCAT(PMTYPE,get_cr3)6133 __CONCAT(PMTYPE, get_cr3)(pmap_t pmap)
6134 {
6135 
6136 #ifdef PMAP_PAE_COMP
6137 	return ((u_int)vtophys(pmap->pm_pdpt));
6138 #else
6139 	return ((u_int)vtophys(pmap->pm_pdir));
6140 #endif
6141 }
6142 
6143 static caddr_t
__CONCAT(PMTYPE,cmap3)6144 __CONCAT(PMTYPE, cmap3)(vm_paddr_t pa, u_int pte_bits)
6145 {
6146 	pt_entry_t *pte;
6147 
6148 	pte = CMAP3;
6149 	*pte = pa | pte_bits;
6150 	invltlb();
6151 	return (CADDR3);
6152 }
6153 
6154 static void
__CONCAT(PMTYPE,basemem_setup)6155 __CONCAT(PMTYPE, basemem_setup)(u_int basemem)
6156 {
6157 	pt_entry_t *pte;
6158 	int i;
6159 
6160 	/*
6161 	 * Map pages between basemem and ISA_HOLE_START, if any, r/w into
6162 	 * the vm86 page table so that vm86 can scribble on them using
6163 	 * the vm86 map too.  XXX: why 2 ways for this and only 1 way for
6164 	 * page 0, at least as initialized here?
6165 	 */
6166 	pte = (pt_entry_t *)vm86paddr;
6167 	for (i = basemem / 4; i < 160; i++)
6168 		pte[i] = (i << PAGE_SHIFT) | PG_V | PG_RW | PG_U;
6169 }
6170 
6171 struct bios16_pmap_handle {
6172 	pt_entry_t	*pte;
6173 	pd_entry_t	*ptd;
6174 	pt_entry_t	orig_ptd;
6175 };
6176 
6177 static void *
__CONCAT(PMTYPE,bios16_enter)6178 __CONCAT(PMTYPE, bios16_enter)(void)
6179 {
6180 	struct bios16_pmap_handle *h;
6181 
6182 	/*
6183 	 * no page table, so create one and install it.
6184 	 */
6185 	h = malloc(sizeof(struct bios16_pmap_handle), M_TEMP, M_WAITOK);
6186 	h->pte = (pt_entry_t *)malloc(PAGE_SIZE, M_TEMP, M_WAITOK);
6187 	h->ptd = IdlePTD;
6188 	*h->pte = vm86phystk | PG_RW | PG_V;
6189 	h->orig_ptd = *h->ptd;
6190 	*h->ptd = vtophys(h->pte) | PG_RW | PG_V;
6191 	pmap_invalidate_all_int(kernel_pmap);	/* XXX insurance for now */
6192 	return (h);
6193 }
6194 
6195 static void
__CONCAT(PMTYPE,bios16_leave)6196 __CONCAT(PMTYPE, bios16_leave)(void *arg)
6197 {
6198 	struct bios16_pmap_handle *h;
6199 
6200 	h = arg;
6201 	*h->ptd = h->orig_ptd;		/* remove page table */
6202 	/*
6203 	 * XXX only needs to be invlpg(0) but that doesn't work on the 386
6204 	 */
6205 	pmap_invalidate_all_int(kernel_pmap);
6206 	free(h->pte, M_TEMP);		/* ... and free it */
6207 }
6208 
6209 struct pmap_kernel_map_range {
6210 	vm_offset_t sva;
6211 	pt_entry_t attrs;
6212 	int ptes;
6213 	int pdes;
6214 	int pdpes;
6215 };
6216 
6217 static void
sysctl_kmaps_dump(struct sbuf * sb,struct pmap_kernel_map_range * range,vm_offset_t eva)6218 sysctl_kmaps_dump(struct sbuf *sb, struct pmap_kernel_map_range *range,
6219     vm_offset_t eva)
6220 {
6221 	const char *mode;
6222 	int i, pat_idx;
6223 
6224 	if (eva <= range->sva)
6225 		return;
6226 
6227 	pat_idx = pmap_pat_index(kernel_pmap, range->attrs, true);
6228 	for (i = 0; i < PAT_INDEX_SIZE; i++)
6229 		if (pat_index[i] == pat_idx)
6230 			break;
6231 
6232 	switch (i) {
6233 	case PAT_WRITE_BACK:
6234 		mode = "WB";
6235 		break;
6236 	case PAT_WRITE_THROUGH:
6237 		mode = "WT";
6238 		break;
6239 	case PAT_UNCACHEABLE:
6240 		mode = "UC";
6241 		break;
6242 	case PAT_UNCACHED:
6243 		mode = "U-";
6244 		break;
6245 	case PAT_WRITE_PROTECTED:
6246 		mode = "WP";
6247 		break;
6248 	case PAT_WRITE_COMBINING:
6249 		mode = "WC";
6250 		break;
6251 	default:
6252 		printf("%s: unknown PAT mode %#x for range 0x%08x-0x%08x\n",
6253 		    __func__, pat_idx, range->sva, eva);
6254 		mode = "??";
6255 		break;
6256 	}
6257 
6258 	sbuf_printf(sb, "0x%08x-0x%08x r%c%c%c%c %s %d %d %d\n",
6259 	    range->sva, eva,
6260 	    (range->attrs & PG_RW) != 0 ? 'w' : '-',
6261 	    (range->attrs & pg_nx) != 0 ? '-' : 'x',
6262 	    (range->attrs & PG_U) != 0 ? 'u' : 's',
6263 	    (range->attrs & PG_G) != 0 ? 'g' : '-',
6264 	    mode, range->pdpes, range->pdes, range->ptes);
6265 
6266 	/* Reset to sentinel value. */
6267 	range->sva = 0xffffffff;
6268 }
6269 
6270 /*
6271  * Determine whether the attributes specified by a page table entry match those
6272  * being tracked by the current range.  This is not quite as simple as a direct
6273  * flag comparison since some PAT modes have multiple representations.
6274  */
6275 static bool
sysctl_kmaps_match(struct pmap_kernel_map_range * range,pt_entry_t attrs)6276 sysctl_kmaps_match(struct pmap_kernel_map_range *range, pt_entry_t attrs)
6277 {
6278 	pt_entry_t diff, mask;
6279 
6280 	mask = pg_nx | PG_G | PG_RW | PG_U | PG_PDE_CACHE;
6281 	diff = (range->attrs ^ attrs) & mask;
6282 	if (diff == 0)
6283 		return (true);
6284 	if ((diff & ~PG_PDE_PAT) == 0 &&
6285 	    pmap_pat_index(kernel_pmap, range->attrs, true) ==
6286 	    pmap_pat_index(kernel_pmap, attrs, true))
6287 		return (true);
6288 	return (false);
6289 }
6290 
6291 static void
sysctl_kmaps_reinit(struct pmap_kernel_map_range * range,vm_offset_t va,pt_entry_t attrs)6292 sysctl_kmaps_reinit(struct pmap_kernel_map_range *range, vm_offset_t va,
6293     pt_entry_t attrs)
6294 {
6295 
6296 	memset(range, 0, sizeof(*range));
6297 	range->sva = va;
6298 	range->attrs = attrs;
6299 }
6300 
6301 /*
6302  * Given a leaf PTE, derive the mapping's attributes.  If they do not match
6303  * those of the current run, dump the address range and its attributes, and
6304  * begin a new run.
6305  */
6306 static void
sysctl_kmaps_check(struct sbuf * sb,struct pmap_kernel_map_range * range,vm_offset_t va,pd_entry_t pde,pt_entry_t pte)6307 sysctl_kmaps_check(struct sbuf *sb, struct pmap_kernel_map_range *range,
6308     vm_offset_t va, pd_entry_t pde, pt_entry_t pte)
6309 {
6310 	pt_entry_t attrs;
6311 
6312 	attrs = pde & (PG_RW | PG_U | pg_nx);
6313 
6314 	if ((pde & PG_PS) != 0) {
6315 		attrs |= pde & (PG_G | PG_PDE_CACHE);
6316 	} else if (pte != 0) {
6317 		attrs |= pte & pg_nx;
6318 		attrs &= pg_nx | (pte & (PG_RW | PG_U));
6319 		attrs |= pte & (PG_G | PG_PTE_CACHE);
6320 
6321 		/* Canonicalize by always using the PDE PAT bit. */
6322 		if ((attrs & PG_PTE_PAT) != 0)
6323 			attrs ^= PG_PDE_PAT | PG_PTE_PAT;
6324 	}
6325 
6326 	if (range->sva > va || !sysctl_kmaps_match(range, attrs)) {
6327 		sysctl_kmaps_dump(sb, range, va);
6328 		sysctl_kmaps_reinit(range, va, attrs);
6329 	}
6330 }
6331 
6332 static int
__CONCAT(PMTYPE,sysctl_kmaps)6333 __CONCAT(PMTYPE, sysctl_kmaps)(SYSCTL_HANDLER_ARGS)
6334 {
6335 	struct pmap_kernel_map_range range;
6336 	struct sbuf sbuf, *sb;
6337 	pd_entry_t pde;
6338 	pt_entry_t *pt, pte;
6339 	vm_offset_t sva;
6340 	int error;
6341 	u_int i, k;
6342 
6343 	error = sysctl_wire_old_buffer(req, 0);
6344 	if (error != 0)
6345 		return (error);
6346 	sb = &sbuf;
6347 	sbuf_new_for_sysctl(sb, NULL, PAGE_SIZE, req);
6348 
6349 	/* Sentinel value. */
6350 	range.sva = 0xffffffff;
6351 
6352 	/*
6353 	 * Iterate over the kernel page tables without holding the
6354 	 * kernel pmap lock.  Kernel page table pages are never freed,
6355 	 * so at worst we will observe inconsistencies in the output.
6356 	 */
6357 	for (sva = 0, i = 0; i < NPTEPG * NPGPTD * NPDEPG ;) {
6358 		if (i == 0)
6359 			sbuf_printf(sb, "\nLow PDE:\n");
6360 		else if (i == LOWPTDI * NPTEPG)
6361 			sbuf_printf(sb, "Low PDE dup:\n");
6362 		else if (i == PTDPTDI * NPTEPG)
6363 			sbuf_printf(sb, "Recursive map:\n");
6364 		else if (i == KERNPTDI * NPTEPG)
6365 			sbuf_printf(sb, "Kernel base:\n");
6366 		else if (i == TRPTDI * NPTEPG)
6367 			sbuf_printf(sb, "Trampoline:\n");
6368 		pde = IdlePTD[sva >> PDRSHIFT];
6369 		if ((pde & PG_V) == 0) {
6370 			sva = rounddown2(sva, NBPDR);
6371 			sysctl_kmaps_dump(sb, &range, sva);
6372 			sva += NBPDR;
6373 			i += NPTEPG;
6374 			continue;
6375 		}
6376 		if ((pde & PG_PS) != 0) {
6377 			sysctl_kmaps_check(sb, &range, sva, pde, 0);
6378 			range.pdes++;
6379 			sva += NBPDR;
6380 			i += NPTEPG;
6381 			continue;
6382 		}
6383 		for (pt = vtopte(sva), k = 0; k < NPTEPG; i++, k++, pt++,
6384 		    sva += PAGE_SIZE) {
6385 			pte = *pt;
6386 			if ((pte & PG_V) == 0) {
6387 				sysctl_kmaps_dump(sb, &range, sva);
6388 				continue;
6389 			}
6390 			sysctl_kmaps_check(sb, &range, sva, pde, pte);
6391 			range.ptes++;
6392 		}
6393 	}
6394 
6395 	error = sbuf_finish(sb);
6396 	sbuf_delete(sb);
6397 	return (error);
6398 }
6399 
6400 #define	PMM(a)					\
6401 	.pm_##a = __CONCAT(PMTYPE, a),
6402 
6403 struct pmap_methods __CONCAT(PMTYPE, methods) = {
6404 	PMM(ksetrw)
6405 	PMM(remap_lower)
6406 	PMM(remap_lowptdi)
6407 	PMM(align_superpage)
6408 	PMM(quick_enter_page)
6409 	PMM(quick_remove_page)
6410 	PMM(trm_alloc)
6411 	PMM(trm_free)
6412 	PMM(get_map_low)
6413 	PMM(get_vm_maxuser_address)
6414 	PMM(kextract)
6415 	PMM(pg_frame)
6416 	PMM(sf_buf_map)
6417 	PMM(cp_slow0_map)
6418 	PMM(get_kcr3)
6419 	PMM(get_cr3)
6420 	PMM(cmap3)
6421 	PMM(basemem_setup)
6422 	PMM(set_nx)
6423 	PMM(bios16_enter)
6424 	PMM(bios16_leave)
6425 	PMM(bootstrap)
6426 	PMM(is_valid_memattr)
6427 	PMM(cache_bits)
6428 	PMM(ps_enabled)
6429 	PMM(pinit0)
6430 	PMM(pinit)
6431 	PMM(activate)
6432 	PMM(activate_boot)
6433 	PMM(advise)
6434 	PMM(clear_modify)
6435 	PMM(change_attr)
6436 	PMM(mincore)
6437 	PMM(copy)
6438 	PMM(copy_page)
6439 	PMM(copy_pages)
6440 	PMM(zero_page)
6441 	PMM(zero_page_area)
6442 	PMM(enter)
6443 	PMM(enter_object)
6444 	PMM(enter_quick)
6445 	PMM(kenter_temporary)
6446 	PMM(object_init_pt)
6447 	PMM(unwire)
6448 	PMM(page_exists_quick)
6449 	PMM(page_wired_mappings)
6450 	PMM(page_is_mapped)
6451 	PMM(remove_pages)
6452 	PMM(is_modified)
6453 	PMM(is_prefaultable)
6454 	PMM(is_referenced)
6455 	PMM(remove_write)
6456 	PMM(ts_referenced)
6457 	PMM(mapdev_attr)
6458 	PMM(unmapdev)
6459 	PMM(page_set_memattr)
6460 	PMM(extract)
6461 	PMM(extract_and_hold)
6462 	PMM(map)
6463 	PMM(qenter)
6464 	PMM(qremove)
6465 	PMM(release)
6466 	PMM(remove)
6467 	PMM(protect)
6468 	PMM(remove_all)
6469 	PMM(init)
6470 	PMM(init_pat)
6471 	PMM(growkernel)
6472 	PMM(invalidate_page)
6473 	PMM(invalidate_range)
6474 	PMM(invalidate_all)
6475 	PMM(invalidate_cache)
6476 	PMM(flush_page)
6477 	PMM(kenter)
6478 	PMM(kremove)
6479 	PMM(sysctl_kmaps)
6480 };
6481