xref: /dflybsd-src/sys/platform/pc64/x86_64/pmap.c (revision 009414b92e8cc83dc5e26248d417bcbdde7408b1)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * Copyright (c) 1994 John S. Dyson
4  * Copyright (c) 1994 David Greenman
5  * Copyright (c) 2003 Peter Wemm
6  * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7  * Copyright (c) 2008, 2009 The DragonFly Project.
8  * Copyright (c) 2008, 2009 Jordan Gordeev.
9  * Copyright (c) 2011-2019 Matthew Dillon
10  * All rights reserved.
11  *
12  * This code is derived from software contributed to Berkeley by
13  * the Systems Programming Group of the University of Utah Computer
14  * Science Department and William Jolitz of UUNET Technologies Inc.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 3. All advertising materials mentioning features or use of this software
25  *    must display the following acknowledgement:
26  *	This product includes software developed by the University of
27  *	California, Berkeley and its contributors.
28  * 4. Neither the name of the University nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  */
44 /*
45  * Manage physical address maps for x86-64 systems.
46  *
47  * Some notes:
48  *	- The 'M'odified bit is only applicable to terminal PTEs.
49  *
50  *	- The 'U'ser access bit can be set for higher-level PTEs as
51  *	  long as it isn't set for terminal PTEs for pages we don't
52  *	  want user access to.
53  */
54 
55 #if 0 /* JG */
56 #include "opt_pmap.h"
57 #endif
58 #include "opt_msgbuf.h"
59 
60 #include <sys/param.h>
61 #include <sys/kernel.h>
62 #include <sys/proc.h>
63 #include <sys/msgbuf.h>
64 #include <sys/vmmeter.h>
65 #include <sys/mman.h>
66 #include <sys/systm.h>
67 
68 #include <vm/vm.h>
69 #include <vm/vm_param.h>
70 #include <sys/sysctl.h>
71 #include <sys/lock.h>
72 #include <vm/vm_kern.h>
73 #include <vm/vm_page.h>
74 #include <vm/vm_map.h>
75 #include <vm/vm_object.h>
76 #include <vm/vm_extern.h>
77 #include <vm/vm_pageout.h>
78 #include <vm/vm_pager.h>
79 #include <vm/vm_zone.h>
80 
81 #include <sys/thread2.h>
82 #include <sys/spinlock2.h>
83 #include <vm/vm_page2.h>
84 
85 #include <machine/cputypes.h>
86 #include <machine/cpu.h>
87 #include <machine/md_var.h>
88 #include <machine/specialreg.h>
89 #include <machine/smp.h>
90 #include <machine_base/apic/apicreg.h>
91 #include <machine/globaldata.h>
92 #include <machine/pmap.h>
93 #include <machine/pmap_inval.h>
94 
95 #include <ddb/ddb.h>
96 
97 #define PMAP_KEEP_PDIRS
98 
99 #if defined(DIAGNOSTIC)
100 #define PMAP_DIAGNOSTIC
101 #endif
102 
103 #define MINPV 2048
104 
105 /*
106  * pmap debugging will report who owns a pv lock when blocking.
107  */
108 #ifdef PMAP_DEBUG
109 
110 #define PMAP_DEBUG_DECL		,const char *func, int lineno
111 #define PMAP_DEBUG_ARGS		, __func__, __LINE__
112 #define PMAP_DEBUG_COPY		, func, lineno
113 
114 #define pv_get(pmap, pindex, pmarkp)	_pv_get(pmap, pindex, pmarkp	\
115 							PMAP_DEBUG_ARGS)
116 #define pv_lock(pv)			_pv_lock(pv			\
117 							PMAP_DEBUG_ARGS)
118 #define pv_hold_try(pv)			_pv_hold_try(pv			\
119 							PMAP_DEBUG_ARGS)
120 #define pv_alloc(pmap, pindex, isnewp)	_pv_alloc(pmap, pindex, isnewp	\
121 							PMAP_DEBUG_ARGS)
122 
123 #define pv_free(pv, pvp)		_pv_free(pv, pvp PMAP_DEBUG_ARGS)
124 
125 #else
126 
127 #define PMAP_DEBUG_DECL
128 #define PMAP_DEBUG_ARGS
129 #define PMAP_DEBUG_COPY
130 
131 #define pv_get(pmap, pindex, pmarkp)	_pv_get(pmap, pindex, pmarkp)
132 #define pv_lock(pv)			_pv_lock(pv)
133 #define pv_hold_try(pv)			_pv_hold_try(pv)
134 #define pv_alloc(pmap, pindex, isnewp)	_pv_alloc(pmap, pindex, isnewp)
135 #define pv_free(pv, pvp)		_pv_free(pv, pvp)
136 
137 #endif
138 
139 /*
140  * Get PDEs and PTEs for user/kernel address space
141  */
142 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
143 
144 #define pmap_pde_v(pmap, pte)	\
145 		((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
146 #define pmap_pte_w(pmap, pte)	\
147 		((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
148 #define pmap_pte_m(pmap, pte)	\
149 		((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
150 #define pmap_pte_u(pmap, pte)	\
151 		((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
152 #define pmap_pte_v(pmap, pte)	\
153 		((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
154 
155 /*
156  * Given a map and a machine independent protection code,
157  * convert to a vax protection code.
158  */
159 #define pte_prot(m, p)		\
160 	(m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
161 static uint64_t protection_codes[PROTECTION_CODES_SIZE];
162 
163 /*
164  * Backing scan macros.  Note that in the use case 'ipte' is only a tentitive
165  * value and must be validated by a pmap_inval_smp_cmpset*() or equivalent
166  * function.
167  *
168  * NOTE: cpu_ccfence() is required to prevent excessive optmization of
169  *	 of the (ipte) variable.
170  */
171 #define PMAP_PAGE_BACKING_SCAN(m, match_pmap, ipmap, iptep, ipte, iva)	\
172 	if (m->object) {						\
173 		vm_object_t iobj = m->object;				\
174 		vm_map_backing_t iba, next_ba;				\
175 		struct pmap *ipmap;					\
176 		pt_entry_t ipte;					\
177 		pt_entry_t *iptep;					\
178 		vm_offset_t iva;					\
179 		vm_pindex_t ipindex_start;				\
180 		vm_pindex_t ipindex_end;				\
181 									\
182 		lockmgr(&iobj->backing_lk, LK_SHARED);			\
183 		next_ba = TAILQ_FIRST(&iobj->backing_list);		\
184 		while ((iba = next_ba) != NULL) {			\
185 			next_ba = TAILQ_NEXT(iba, entry);		\
186 			ipmap = iba->pmap;				\
187 			if (match_pmap && ipmap != match_pmap)		\
188 				continue;				\
189 			ipindex_start = iba->offset >> PAGE_SHIFT;	\
190 			ipindex_end = ipindex_start +			\
191 				  ((iba->end - iba->start) >> PAGE_SHIFT); \
192 			if (m->pindex < ipindex_start ||		\
193 			    m->pindex >= ipindex_end) {			\
194 				continue;				\
195 			}						\
196 			iva = iba->start +				\
197 			      ((m->pindex - ipindex_start) << PAGE_SHIFT); \
198 			iptep = pmap_pte(ipmap, iva);			\
199 			if (iptep == NULL)				\
200 				continue;				\
201 			ipte = *iptep;					\
202 			cpu_ccfence();					\
203 			if (m->phys_addr != (ipte & PG_FRAME))		\
204 				continue;				\
205 
206 #define PMAP_PAGE_BACKING_RETRY						\
207 			{						\
208 				next_ba = iba;				\
209 				continue;				\
210 			}						\
211 
212 #define PMAP_PAGE_BACKING_DONE						\
213 		}							\
214 		lockmgr(&iobj->backing_lk, LK_RELEASE);			\
215 	}								\
216 
217 struct pmap kernel_pmap;
218 struct pmap iso_pmap;
219 
220 vm_paddr_t avail_start;		/* PA of first available physical page */
221 vm_paddr_t avail_end;		/* PA of last available physical page */
222 vm_offset_t virtual2_start;	/* cutout free area prior to kernel start */
223 vm_offset_t virtual2_end;
224 vm_offset_t virtual_start;	/* VA of first avail page (after kernel bss) */
225 vm_offset_t virtual_end;	/* VA of last avail page (end of kernel AS) */
226 vm_offset_t KvaStart;		/* VA start of KVA space */
227 vm_offset_t KvaEnd;		/* VA end of KVA space (non-inclusive) */
228 vm_offset_t KvaSize;		/* max size of kernel virtual address space */
229 vm_offset_t DMapMaxAddress;
230 static boolean_t pmap_initialized = FALSE;	/* Has pmap_init completed? */
231 //static int pgeflag;		/* PG_G or-in */
232 uint64_t PatMsr;
233 
234 static int ndmpdp;
235 static vm_paddr_t dmaplimit;
236 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
237 
238 static pt_entry_t pat_pte_index[PAT_INDEX_SIZE];	/* PAT -> PG_ bits */
239 static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];	/* PAT -> PG_ bits */
240 
241 static uint64_t KPTbase;
242 static uint64_t KPTphys;
243 static uint64_t	KPDphys;	/* phys addr of kernel level 2 */
244 static uint64_t	KPDbase;	/* phys addr of kernel level 2 @ KERNBASE */
245 uint64_t KPDPphys;		/* phys addr of kernel level 3 */
246 uint64_t KPML4phys;		/* phys addr of kernel level 4 */
247 
248 static uint64_t	DMPDphys;	/* phys addr of direct mapped level 2 */
249 static uint64_t	DMPDPphys;	/* phys addr of direct mapped level 3 */
250 
251 /*
252  * Data for the pv entry allocation mechanism
253  */
254 __read_mostly static vm_zone_t pvzone;
255 __read_mostly static int pmap_pagedaemon_waken = 0;
256 static struct vm_zone pvzone_store;
257 static struct pv_entry *pvinit;
258 
259 /*
260  * All those kernel PT submaps that BSD is so fond of
261  */
262 pt_entry_t *CMAP1 = NULL, *ptmmap;
263 caddr_t CADDR1 = NULL, ptvmmap = NULL;
264 static pt_entry_t *msgbufmap;
265 struct msgbuf *msgbufp=NULL;
266 
267 /*
268  * PMAP default PG_* bits. Needed to be able to add
269  * EPT/NPT pagetable pmap_bits for the VMM module
270  */
271 uint64_t pmap_bits_default[] = {
272 		REGULAR_PMAP,			/* TYPE_IDX		0 */
273 		X86_PG_V,			/* PG_V_IDX		1 */
274 		X86_PG_RW,			/* PG_RW_IDX		2 */
275 		X86_PG_U,			/* PG_U_IDX		3 */
276 		X86_PG_A,			/* PG_A_IDX		4 */
277 		X86_PG_M,			/* PG_M_IDX		5 */
278 		X86_PG_PS,			/* PG_PS_IDX3		6 */
279 		X86_PG_G,			/* PG_G_IDX		7 */
280 		X86_PG_AVAIL1,			/* PG_AVAIL1_IDX	8 */
281 		X86_PG_AVAIL2,			/* PG_AVAIL2_IDX	9 */
282 		X86_PG_AVAIL3,			/* PG_AVAIL3_IDX	10 */
283 		X86_PG_NC_PWT | X86_PG_NC_PCD,	/* PG_N_IDX		11 */
284 		X86_PG_NX,			/* PG_NX_IDX		12 */
285 };
286 /*
287  * Crashdump maps.
288  */
289 static pt_entry_t *pt_crashdumpmap;
290 static caddr_t crashdumpmap;
291 
292 static int pmap_debug = 0;
293 SYSCTL_INT(_machdep, OID_AUTO, pmap_debug, CTLFLAG_RW,
294     &pmap_debug, 0, "Debug pmap's");
295 #ifdef PMAP_DEBUG2
296 static int pmap_enter_debug = 0;
297 SYSCTL_INT(_machdep, OID_AUTO, pmap_enter_debug, CTLFLAG_RW,
298     &pmap_enter_debug, 0, "Debug pmap_enter's");
299 #endif
300 static int pmap_yield_count = 64;
301 SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
302     &pmap_yield_count, 0, "Yield during init_pt/release");
303 int pmap_fast_kernel_cpusync = 0;
304 SYSCTL_INT(_machdep, OID_AUTO, pmap_fast_kernel_cpusync, CTLFLAG_RW,
305     &pmap_fast_kernel_cpusync, 0, "Share page table pages when possible");
306 int pmap_dynamic_delete = 0;
307 SYSCTL_INT(_machdep, OID_AUTO, pmap_dynamic_delete, CTLFLAG_RW,
308     &pmap_dynamic_delete, 0, "Dynamically delete PT/PD/PDPs");
309 int pmap_lock_delay = 100;
310 SYSCTL_INT(_machdep, OID_AUTO, pmap_lock_delay, CTLFLAG_RW,
311     &pmap_lock_delay, 0, "Spin loops");
312 static int meltdown_mitigation = -1;
313 TUNABLE_INT("machdep.meltdown_mitigation", &meltdown_mitigation);
314 SYSCTL_INT(_machdep, OID_AUTO, meltdown_mitigation, CTLFLAG_RW,
315     &meltdown_mitigation, 0, "Userland pmap isolation");
316 
317 static int pmap_nx_enable = -1;		/* -1 = auto */
318 /* needs manual TUNABLE in early probe, see below */
319 SYSCTL_INT(_machdep, OID_AUTO, pmap_nx_enable, CTLFLAG_RD,
320     &pmap_nx_enable, 0,
321     "no-execute support (0=disabled, 1=w/READ, 2=w/READ & WRITE)");
322 
323 static int pmap_pv_debug = 50;
324 SYSCTL_INT(_machdep, OID_AUTO, pmap_pv_debug, CTLFLAG_RW,
325     &pmap_pv_debug, 0, "");
326 
327 static long vm_pmap_pv_entries;
328 SYSCTL_LONG(_vm, OID_AUTO, pmap_pv_entries, CTLFLAG_RD,
329     &vm_pmap_pv_entries, 0, "");
330 
331 /* Standard user access funtions */
332 extern int std_copyinstr (const void *udaddr, void *kaddr, size_t len,
333     size_t *lencopied);
334 extern int std_copyin (const void *udaddr, void *kaddr, size_t len);
335 extern int std_copyout (const void *kaddr, void *udaddr, size_t len);
336 extern int std_fubyte (const uint8_t *base);
337 extern int std_subyte (uint8_t *base, uint8_t byte);
338 extern int32_t std_fuword32 (const uint32_t *base);
339 extern int64_t std_fuword64 (const uint64_t *base);
340 extern int std_suword64 (uint64_t *base, uint64_t word);
341 extern int std_suword32 (uint32_t *base, int word);
342 extern uint32_t std_swapu32 (volatile uint32_t *base, uint32_t v);
343 extern uint64_t std_swapu64 (volatile uint64_t *base, uint64_t v);
344 extern uint32_t std_fuwordadd32 (volatile uint32_t *base, uint32_t v);
345 extern uint64_t std_fuwordadd64 (volatile uint64_t *base, uint64_t v);
346 
347 #if 0
348 static void pv_hold(pv_entry_t pv);
349 #endif
350 static int _pv_hold_try(pv_entry_t pv
351 				PMAP_DEBUG_DECL);
352 static void pv_drop(pv_entry_t pv);
353 static void _pv_lock(pv_entry_t pv
354 				PMAP_DEBUG_DECL);
355 static void pv_unlock(pv_entry_t pv);
356 static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
357 				PMAP_DEBUG_DECL);
358 static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp
359 				PMAP_DEBUG_DECL);
360 static void _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL);
361 static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex,
362 				vm_pindex_t **pmarkp, int *errorp);
363 static void pv_put(pv_entry_t pv);
364 static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
365 static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
366 		      pv_entry_t *pvpp);
367 static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
368 			pmap_inval_bulk_t *bulk, int destroy);
369 static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
370 static int pmap_release_pv(pv_entry_t pv, pv_entry_t pvp,
371 			pmap_inval_bulk_t *bulk);
372 
373 struct pmap_scan_info;
374 static void pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
375 		      vm_pindex_t *pte_placemark, pv_entry_t pt_pv,
376 		      vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
377 static void pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
378 		      vm_pindex_t *pte_placemark, pv_entry_t pt_pv,
379 		      vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
380 
381 static void x86_64_protection_init (void);
382 static void create_pagetables(vm_paddr_t *firstaddr);
383 static void pmap_remove_all (vm_page_t m);
384 static boolean_t pmap_testbit (vm_page_t m, int bit);
385 
386 static pt_entry_t *pmap_pte_quick (pmap_t pmap, vm_offset_t va);
387 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
388 
389 static void pmap_pinit_defaults(struct pmap *pmap);
390 static void pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark);
391 static void pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark);
392 
393 static int
394 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
395 {
396 	if (pv1->pv_pindex < pv2->pv_pindex)
397 		return(-1);
398 	if (pv1->pv_pindex > pv2->pv_pindex)
399 		return(1);
400 	return(0);
401 }
402 
403 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
404              pv_entry_compare, vm_pindex_t, pv_pindex);
405 
406 /*
407  * Keep track of pages in the pmap.  The procedure is handed
408  * the vm_page->md.pmap_count value prior to an increment or
409  * decrement.
410  *
411  *	t_arm		- Active real memory
412  *	t_avm		- Active virtual memory
413  *	t_armshr	- Active real memory that is also shared
414  *	t_avmshr	- Active virtual memory that is also shared
415  *
416  * NOTE: At the moment t_avm is effectively just the same as t_arm.
417  */
418 static __inline
419 void
420 pmap_page_stats_adding(long prev_count)
421 {
422 	globaldata_t gd = mycpu;
423 
424 	if (prev_count == 0) {
425 		++gd->gd_vmtotal.t_arm;
426 		++gd->gd_vmtotal.t_avm;
427 	} else if (prev_count == 1) {
428 		++gd->gd_vmtotal.t_armshr;
429 		++gd->gd_vmtotal.t_avmshr;
430 	} else {
431 		++gd->gd_vmtotal.t_avmshr;
432 	}
433 }
434 
435 static __inline
436 void
437 pmap_page_stats_deleting(long prev_count)
438 {
439 	globaldata_t gd = mycpu;
440 
441 	if (prev_count == 1) {
442 		--gd->gd_vmtotal.t_arm;
443 		--gd->gd_vmtotal.t_avm;
444 	} else if (prev_count == 2) {
445 		--gd->gd_vmtotal.t_armshr;
446 		--gd->gd_vmtotal.t_avmshr;
447 	} else {
448 		--gd->gd_vmtotal.t_avmshr;
449 	}
450 }
451 
452 /*
453  * Move the kernel virtual free pointer to the next
454  * 2MB.  This is used to help improve performance
455  * by using a large (2MB) page for much of the kernel
456  * (.text, .data, .bss)
457  */
458 static
459 vm_offset_t
460 pmap_kmem_choose(vm_offset_t addr)
461 {
462 	vm_offset_t newaddr = addr;
463 
464 	newaddr = roundup2(addr, NBPDR);
465 	return newaddr;
466 }
467 
468 /*
469  * Returns the pindex of a page table entry (representing a terminal page).
470  * There are NUPTE_TOTAL page table entries possible (a huge number)
471  *
472  * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
473  * We want to properly translate negative KVAs.
474  */
475 static __inline
476 vm_pindex_t
477 pmap_pte_pindex(vm_offset_t va)
478 {
479 	return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
480 }
481 
482 /*
483  * Returns the pindex of a page table.
484  */
485 static __inline
486 vm_pindex_t
487 pmap_pt_pindex(vm_offset_t va)
488 {
489 	return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
490 }
491 
492 /*
493  * Returns the pindex of a page directory.
494  */
495 static __inline
496 vm_pindex_t
497 pmap_pd_pindex(vm_offset_t va)
498 {
499 	return (NUPTE_TOTAL + NUPT_TOTAL +
500 		((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
501 }
502 
503 static __inline
504 vm_pindex_t
505 pmap_pdp_pindex(vm_offset_t va)
506 {
507 	return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
508 		((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
509 }
510 
511 static __inline
512 vm_pindex_t
513 pmap_pml4_pindex(void)
514 {
515 	return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
516 }
517 
518 /*
519  * Return various clipped indexes for a given VA
520  *
521  * Returns the index of a pt in a page directory, representing a page
522  * table.
523  */
524 static __inline
525 vm_pindex_t
526 pmap_pt_index(vm_offset_t va)
527 {
528 	return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
529 }
530 
531 /*
532  * Returns the index of a pd in a page directory page, representing a page
533  * directory.
534  */
535 static __inline
536 vm_pindex_t
537 pmap_pd_index(vm_offset_t va)
538 {
539 	return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
540 }
541 
542 /*
543  * Returns the index of a pdp in the pml4 table, representing a page
544  * directory page.
545  */
546 static __inline
547 vm_pindex_t
548 pmap_pdp_index(vm_offset_t va)
549 {
550 	return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
551 }
552 
553 /*
554  * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
555  * the PT layer.  This will speed up core pmap operations considerably.
556  * We also cache the PTE layer to (hopefully) improve relative lookup
557  * speeds.
558  *
559  * NOTE: The pmap spinlock does not need to be held but the passed-in pv
560  *	 must be in a known associated state (typically by being locked when
561  *	 the pmap spinlock isn't held).  We allow the race for that case.
562  *
563  * NOTE: pm_pvhint* is only accessed (read) with the spin-lock held, using
564  *	 cpu_ccfence() to prevent compiler optimizations from reloading the
565  *	 field.
566  */
567 static __inline
568 void
569 pv_cache(pmap_t pmap, pv_entry_t pv, vm_pindex_t pindex)
570 {
571 	if (pindex < pmap_pt_pindex(0)) {
572 		;
573 	} else if (pindex < pmap_pd_pindex(0)) {
574 		pmap->pm_pvhint_pt = pv;
575 	}
576 }
577 
578 /*
579  * Locate the requested pt_entry
580  */
581 static __inline
582 pv_entry_t
583 pv_entry_lookup(pmap_t pmap, vm_pindex_t pindex)
584 {
585 	pv_entry_t pv;
586 
587 	if (pindex < pmap_pt_pindex(0))
588 		return NULL;
589 #if 1
590 	if (pindex < pmap_pd_pindex(0))
591 		pv = pmap->pm_pvhint_pt;
592 	else
593 		pv = NULL;
594 	cpu_ccfence();
595 	if (pv == NULL || pv->pv_pmap != pmap) {
596 		pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
597 		if (pv)
598 			pv_cache(pmap, pv, pindex);
599 	} else if (pv->pv_pindex != pindex) {
600 		pv = pv_entry_rb_tree_RB_LOOKUP_REL(&pmap->pm_pvroot,
601 						    pindex, pv);
602 		if (pv)
603 			pv_cache(pmap, pv, pindex);
604 	}
605 #else
606 	pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
607 #endif
608 	return pv;
609 }
610 
611 /*
612  * pmap_pte_quick:
613  *
614  *	Super fast pmap_pte routine best used when scanning the pv lists.
615  *	This eliminates many course-grained invltlb calls.  Note that many of
616  *	the pv list scans are across different pmaps and it is very wasteful
617  *	to do an entire invltlb when checking a single mapping.
618  */
619 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
620 
621 static
622 pt_entry_t *
623 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
624 {
625 	return pmap_pte(pmap, va);
626 }
627 
628 /*
629  * The placemarker hash must be broken up into four zones so lock
630  * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
631  *
632  * Placemarkers are used to 'lock' page table indices that do not have
633  * a pv_entry.  This allows the pmap to support managed and unmanaged
634  * pages and shared page tables.
635  */
636 #define PM_PLACE_BASE	(PM_PLACEMARKS >> 2)
637 
638 static __inline
639 vm_pindex_t *
640 pmap_placemarker_hash(pmap_t pmap, vm_pindex_t pindex)
641 {
642 	int hi;
643 
644 	if (pindex < pmap_pt_pindex(0))		/* zone 0 - PTE */
645 		hi = 0;
646 	else if (pindex < pmap_pd_pindex(0))	/* zone 1 - PT */
647 		hi = PM_PLACE_BASE;
648 	else if (pindex < pmap_pdp_pindex(0))	/* zone 2 - PD */
649 		hi = PM_PLACE_BASE << 1;
650 	else					/* zone 3 - PDP (and PML4E) */
651 		hi = PM_PLACE_BASE | (PM_PLACE_BASE << 1);
652 	hi += pindex & (PM_PLACE_BASE - 1);
653 
654 	return (&pmap->pm_placemarks[hi]);
655 }
656 
657 
658 /*
659  * Generic procedure to index a pte from a pt, pd, or pdp.
660  *
661  * NOTE: Normally passed pindex as pmap_xx_index().  pmap_xx_pindex() is NOT
662  *	 a page table page index but is instead of PV lookup index.
663  */
664 static
665 void *
666 pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
667 {
668 	pt_entry_t *pte;
669 
670 	pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
671 	return(&pte[pindex]);
672 }
673 
674 /*
675  * Return pointer to PDP slot in the PML4
676  */
677 static __inline
678 pml4_entry_t *
679 pmap_pdp(pmap_t pmap, vm_offset_t va)
680 {
681 	return (&pmap->pm_pml4[pmap_pdp_index(va)]);
682 }
683 
684 /*
685  * Return pointer to PD slot in the PDP given a pointer to the PDP
686  */
687 static __inline
688 pdp_entry_t *
689 pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
690 {
691 	pdp_entry_t *pd;
692 
693 	pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
694 	return (&pd[pmap_pd_index(va)]);
695 }
696 
697 /*
698  * Return pointer to PD slot in the PDP.
699  */
700 static __inline
701 pdp_entry_t *
702 pmap_pd(pmap_t pmap, vm_offset_t va)
703 {
704 	pml4_entry_t *pdp;
705 
706 	pdp = pmap_pdp(pmap, va);
707 	if ((*pdp & pmap->pmap_bits[PG_V_IDX]) == 0)
708 		return NULL;
709 	return (pmap_pdp_to_pd(*pdp, va));
710 }
711 
712 /*
713  * Return pointer to PT slot in the PD given a pointer to the PD
714  */
715 static __inline
716 pd_entry_t *
717 pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
718 {
719 	pd_entry_t *pt;
720 
721 	pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
722 	return (&pt[pmap_pt_index(va)]);
723 }
724 
725 /*
726  * Return pointer to PT slot in the PD
727  *
728  * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
729  *		     so we cannot lookup the PD via the PDP.  Instead we
730  *		     must look it up via the pmap.
731  */
732 static __inline
733 pd_entry_t *
734 pmap_pt(pmap_t pmap, vm_offset_t va)
735 {
736 	pdp_entry_t *pd;
737 	pv_entry_t pv;
738 	vm_pindex_t pd_pindex;
739 	vm_paddr_t phys;
740 
741 	if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
742 		pd_pindex = pmap_pd_pindex(va);
743 		spin_lock_shared(&pmap->pm_spin);
744 		pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
745 		if (pv == NULL || pv->pv_m == NULL) {
746 			spin_unlock_shared(&pmap->pm_spin);
747 			return NULL;
748 		}
749 		phys = VM_PAGE_TO_PHYS(pv->pv_m);
750 		spin_unlock_shared(&pmap->pm_spin);
751 		return (pmap_pd_to_pt(phys, va));
752 	} else {
753 		pd = pmap_pd(pmap, va);
754 		if (pd == NULL || (*pd & pmap->pmap_bits[PG_V_IDX]) == 0)
755 			 return NULL;
756 		return (pmap_pd_to_pt(*pd, va));
757 	}
758 }
759 
760 /*
761  * Return pointer to PTE slot in the PT given a pointer to the PT
762  */
763 static __inline
764 pt_entry_t *
765 pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
766 {
767 	pt_entry_t *pte;
768 
769 	pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
770 	return (&pte[pmap_pte_index(va)]);
771 }
772 
773 /*
774  * Return pointer to PTE slot in the PT
775  */
776 static __inline
777 pt_entry_t *
778 pmap_pte(pmap_t pmap, vm_offset_t va)
779 {
780 	pd_entry_t *pt;
781 
782 	pt = pmap_pt(pmap, va);
783 	if (pt == NULL || (*pt & pmap->pmap_bits[PG_V_IDX]) == 0)
784 		 return NULL;
785 	if ((*pt & pmap->pmap_bits[PG_PS_IDX]) != 0)
786 		return ((pt_entry_t *)pt);
787 	return (pmap_pt_to_pte(*pt, va));
788 }
789 
790 /*
791  * Return address of PT slot in PD (KVM only)
792  *
793  * Cannot be used for user page tables because it might interfere with
794  * the shared page-table-page optimization (pmap_mmu_optimize).
795  */
796 static __inline
797 pd_entry_t *
798 vtopt(vm_offset_t va)
799 {
800 	uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
801 				  NPML4EPGSHIFT)) - 1);
802 
803 	return (PDmap + ((va >> PDRSHIFT) & mask));
804 }
805 
806 /*
807  * KVM - return address of PTE slot in PT
808  */
809 static __inline
810 pt_entry_t *
811 vtopte(vm_offset_t va)
812 {
813 	uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
814 				  NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
815 
816 	return (PTmap + ((va >> PAGE_SHIFT) & mask));
817 }
818 
819 /*
820  * Returns the physical address translation from va for a user address.
821  * (vm_paddr_t)-1 is returned on failure.
822  */
823 vm_paddr_t
824 uservtophys(vm_offset_t va)
825 {
826 	uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
827 				  NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
828 	vm_paddr_t pa;
829 	pt_entry_t pte;
830 	pmap_t pmap;
831 
832 	pmap = vmspace_pmap(mycpu->gd_curthread->td_lwp->lwp_vmspace);
833 	pa = (vm_paddr_t)-1;
834 	if (va < VM_MAX_USER_ADDRESS) {
835 		pte = kreadmem64(PTmap + ((va >> PAGE_SHIFT) & mask));
836 		if (pte & pmap->pmap_bits[PG_V_IDX])
837 			pa = (pte & PG_FRAME) | (va & PAGE_MASK);
838 	}
839 	return pa;
840 }
841 
842 static uint64_t
843 allocpages(vm_paddr_t *firstaddr, long n)
844 {
845 	uint64_t ret;
846 
847 	ret = *firstaddr;
848 	bzero((void *)ret, n * PAGE_SIZE);
849 	*firstaddr += n * PAGE_SIZE;
850 	return (ret);
851 }
852 
853 static
854 void
855 create_pagetables(vm_paddr_t *firstaddr)
856 {
857 	long i;		/* must be 64 bits */
858 	long nkpt_base;
859 	long nkpt_phys;
860 	long nkpd_phys;
861 	int j;
862 
863 	/*
864 	 * We are running (mostly) V=P at this point
865 	 *
866 	 * Calculate how many 1GB PD entries in our PDP pages are needed
867 	 * for the DMAP.  This is only allocated if the system does not
868 	 * support 1GB pages.  Otherwise ndmpdp is simply a count of
869 	 * the number of 1G terminal entries in our PDP pages are needed.
870 	 *
871 	 * NOTE: Maxmem is in pages
872 	 */
873 	ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
874 	if (ndmpdp < 4)		/* Minimum 4GB of dirmap */
875 		ndmpdp = 4;
876 	KKASSERT(ndmpdp <= NDMPML4E * NPML4EPG);
877 	DMapMaxAddress = DMAP_MIN_ADDRESS +
878 			 ((ndmpdp * NPDEPG) << PDRSHIFT);
879 
880 	/*
881 	 * Starting at KERNBASE - map all 2G worth of page table pages.
882 	 * KERNBASE is offset -2G from the end of kvm.  This will accomodate
883 	 * all KVM allocations above KERNBASE, including the SYSMAPs below.
884 	 *
885 	 * We do this by allocating 2*512 PT pages.  Each PT page can map
886 	 * 2MB, for 2GB total.
887 	 */
888 	nkpt_base = (NPDPEPG - KPDPI) * NPTEPG;	/* typically 2 x 512 */
889 
890 	/*
891 	 * Starting at the beginning of kvm (VM_MIN_KERNEL_ADDRESS),
892 	 * Calculate how many page table pages we need to preallocate
893 	 * for early vm_map allocations.
894 	 *
895 	 * A few extra won't hurt, they will get used up in the running
896 	 * system.
897 	 *
898 	 * vm_page array
899 	 * initial pventry's
900 	 */
901 	nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
902 	nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
903 	nkpt_phys += 128;	/* a few extra */
904 
905 	/*
906 	 * The highest value nkpd_phys can be set to is
907 	 * NKPDPE - (NPDPEPG - KPDPI) (i.e. NKPDPE - 2).
908 	 *
909 	 * Doing so would cause all PD pages to be pre-populated for
910 	 * a maximal KVM space (approximately 16*512 pages, or 32MB.
911 	 * We can save memory by not doing this.
912 	 */
913 	nkpd_phys = (nkpt_phys + NPDPEPG - 1) / NPDPEPG;
914 
915 	/*
916 	 * Allocate pages
917 	 *
918 	 * Normally NKPML4E=1-16 (1-16 kernel PDP page)
919 	 * Normally NKPDPE= NKPML4E*512-1 (511 min kernel PD pages)
920 	 *
921 	 * Only allocate enough PD pages
922 	 * NOTE: We allocate all kernel PD pages up-front, typically
923 	 *	 ~511G of KVM, requiring 511 PD pages.
924 	 */
925 	KPTbase = allocpages(firstaddr, nkpt_base);	/* KERNBASE to end */
926 	KPTphys = allocpages(firstaddr, nkpt_phys);	/* KVA start */
927 	KPML4phys = allocpages(firstaddr, 1);		/* recursive PML4 map */
928 	KPDPphys = allocpages(firstaddr, NKPML4E);	/* kernel PDP pages */
929 	KPDphys = allocpages(firstaddr, nkpd_phys);	/* kernel PD pages */
930 
931 	/*
932 	 * Alloc PD pages for the area starting at KERNBASE.
933 	 */
934 	KPDbase = allocpages(firstaddr, NPDPEPG - KPDPI);
935 
936 	/*
937 	 * Stuff for our DMAP.  Use 2MB pages even when 1GB pages
938 	 * are available in order to allow APU code to adjust page
939 	 * attributes on a fixed grain (see pmap_change_attr()).
940 	 */
941 	DMPDPphys = allocpages(firstaddr, NDMPML4E);
942 #if 1
943 	DMPDphys = allocpages(firstaddr, ndmpdp);
944 #else
945 	if ((amd_feature & AMDID_PAGE1GB) == 0)
946 		DMPDphys = allocpages(firstaddr, ndmpdp);
947 #endif
948 	dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
949 
950 	/*
951 	 * Fill in the underlying page table pages for the area around
952 	 * KERNBASE.  This remaps low physical memory to KERNBASE.
953 	 *
954 	 * Read-only from zero to physfree
955 	 * XXX not fully used, underneath 2M pages
956 	 */
957 	for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
958 		((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
959 		((pt_entry_t *)KPTbase)[i] |=
960 		    pmap_bits_default[PG_RW_IDX] |
961 		    pmap_bits_default[PG_V_IDX] |
962 		    pmap_bits_default[PG_G_IDX];
963 	}
964 
965 	/*
966 	 * Now map the initial kernel page tables.  One block of page
967 	 * tables is placed at the beginning of kernel virtual memory,
968 	 * and another block is placed at KERNBASE to map the kernel binary,
969 	 * data, bss, and initial pre-allocations.
970 	 */
971 	for (i = 0; i < nkpt_base; i++) {
972 		((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
973 		((pd_entry_t *)KPDbase)[i] |=
974 		    pmap_bits_default[PG_RW_IDX] |
975 		    pmap_bits_default[PG_V_IDX];
976 	}
977 	for (i = 0; i < nkpt_phys; i++) {
978 		((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
979 		((pd_entry_t *)KPDphys)[i] |=
980 		    pmap_bits_default[PG_RW_IDX] |
981 		    pmap_bits_default[PG_V_IDX];
982 	}
983 
984 	/*
985 	 * Map from zero to end of allocations using 2M pages as an
986 	 * optimization.  This will bypass some of the KPTBase pages
987 	 * above in the KERNBASE area.
988 	 */
989 	for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
990 		((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
991 		((pd_entry_t *)KPDbase)[i] |=
992 		    pmap_bits_default[PG_RW_IDX] |
993 		    pmap_bits_default[PG_V_IDX] |
994 		    pmap_bits_default[PG_PS_IDX] |
995 		    pmap_bits_default[PG_G_IDX];
996 	}
997 
998 	/*
999 	 * Load PD addresses into the PDP pages for primary KVA space to
1000 	 * cover existing page tables.  PD's for KERNBASE are handled in
1001 	 * the next loop.
1002 	 *
1003 	 * expected to pre-populate all of its PDs.  See NKPDPE in vmparam.h.
1004 	 */
1005 	for (i = 0; i < nkpd_phys; i++) {
1006 		((pdp_entry_t *)KPDPphys)[NKPML4E * NPDPEPG - NKPDPE + i] =
1007 				KPDphys + (i << PAGE_SHIFT);
1008 		((pdp_entry_t *)KPDPphys)[NKPML4E * NPDPEPG - NKPDPE + i] |=
1009 		    pmap_bits_default[PG_RW_IDX] |
1010 		    pmap_bits_default[PG_V_IDX] |
1011 		    pmap_bits_default[PG_A_IDX];
1012 	}
1013 
1014 	/*
1015 	 * Load PDs for KERNBASE to the end
1016 	 */
1017 	i = (NKPML4E - 1) * NPDPEPG + KPDPI;
1018 	for (j = 0; j < NPDPEPG - KPDPI; ++j) {
1019 		((pdp_entry_t *)KPDPphys)[i + j] =
1020 				KPDbase + (j << PAGE_SHIFT);
1021 		((pdp_entry_t *)KPDPphys)[i + j] |=
1022 		    pmap_bits_default[PG_RW_IDX] |
1023 		    pmap_bits_default[PG_V_IDX] |
1024 		    pmap_bits_default[PG_A_IDX];
1025 	}
1026 
1027 	/*
1028 	 * Now set up the direct map space using either 2MB or 1GB pages
1029 	 * Preset PG_M and PG_A because demotion expects it.
1030 	 *
1031 	 * When filling in entries in the PD pages make sure any excess
1032 	 * entries are set to zero as we allocated enough PD pages
1033 	 *
1034 	 * Stuff for our DMAP.  Use 2MB pages even when 1GB pages
1035 	 * are available in order to allow APU code to adjust page
1036 	 * attributes on a fixed grain (see pmap_change_attr()).
1037 	 */
1038 #if 0
1039 	if ((amd_feature & AMDID_PAGE1GB) == 0)
1040 #endif
1041 	{
1042 		/*
1043 		 * Use 2MB pages
1044 		 */
1045 		for (i = 0; i < NPDEPG * ndmpdp; i++) {
1046 			((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
1047 			((pd_entry_t *)DMPDphys)[i] |=
1048 			    pmap_bits_default[PG_RW_IDX] |
1049 			    pmap_bits_default[PG_V_IDX] |
1050 			    pmap_bits_default[PG_PS_IDX] |
1051 			    pmap_bits_default[PG_G_IDX] |
1052 			    pmap_bits_default[PG_M_IDX] |
1053 			    pmap_bits_default[PG_A_IDX];
1054 		}
1055 
1056 		/*
1057 		 * And the direct map space's PDP
1058 		 */
1059 		for (i = 0; i < ndmpdp; i++) {
1060 			((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
1061 							(i << PAGE_SHIFT);
1062 			((pdp_entry_t *)DMPDPphys)[i] |=
1063 			    pmap_bits_default[PG_RW_IDX] |
1064 			    pmap_bits_default[PG_V_IDX] |
1065 			    pmap_bits_default[PG_A_IDX];
1066 		}
1067 	}
1068 #if 0
1069 	else {
1070 		/*
1071 		 * 1GB pages
1072 		 */
1073 		for (i = 0; i < ndmpdp; i++) {
1074 			((pdp_entry_t *)DMPDPphys)[i] =
1075 						(vm_paddr_t)i << PDPSHIFT;
1076 			((pdp_entry_t *)DMPDPphys)[i] |=
1077 			    pmap_bits_default[PG_RW_IDX] |
1078 			    pmap_bits_default[PG_V_IDX] |
1079 			    pmap_bits_default[PG_PS_IDX] |
1080 			    pmap_bits_default[PG_G_IDX] |
1081 			    pmap_bits_default[PG_M_IDX] |
1082 			    pmap_bits_default[PG_A_IDX];
1083 		}
1084 	}
1085 #endif
1086 
1087 	/* And recursively map PML4 to itself in order to get PTmap */
1088 	((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
1089 	((pdp_entry_t *)KPML4phys)[PML4PML4I] |=
1090 	    pmap_bits_default[PG_RW_IDX] |
1091 	    pmap_bits_default[PG_V_IDX] |
1092 	    pmap_bits_default[PG_A_IDX];
1093 
1094 	/*
1095 	 * Connect the Direct Map slots up to the PML4
1096 	 */
1097 	for (j = 0; j < NDMPML4E; ++j) {
1098 		((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
1099 		    (DMPDPphys + ((vm_paddr_t)j << PAGE_SHIFT)) |
1100 		    pmap_bits_default[PG_RW_IDX] |
1101 		    pmap_bits_default[PG_V_IDX] |
1102 		    pmap_bits_default[PG_A_IDX];
1103 	}
1104 
1105 	/*
1106 	 * Connect the KVA slot up to the PML4
1107 	 */
1108 	for (j = 0; j < NKPML4E; ++j) {
1109 		((pdp_entry_t *)KPML4phys)[KPML4I + j] =
1110 		    KPDPphys + ((vm_paddr_t)j << PAGE_SHIFT);
1111 		((pdp_entry_t *)KPML4phys)[KPML4I + j] |=
1112 		    pmap_bits_default[PG_RW_IDX] |
1113 		    pmap_bits_default[PG_V_IDX] |
1114 		    pmap_bits_default[PG_A_IDX];
1115 	}
1116 	cpu_mfence();
1117 	cpu_invltlb();
1118 }
1119 
1120 /*
1121  *	Bootstrap the system enough to run with virtual memory.
1122  *
1123  *	On x86_64 this is called after mapping has already been enabled
1124  *	and just syncs the pmap module with what has already been done.
1125  *	[We can't call it easily with mapping off since the kernel is not
1126  *	mapped with PA == VA, hence we would have to relocate every address
1127  *	from the linked base (virtual) address "KERNBASE" to the actual
1128  *	(physical) address starting relative to 0]
1129  */
1130 void
1131 pmap_bootstrap(vm_paddr_t *firstaddr)
1132 {
1133 	vm_offset_t va;
1134 	pt_entry_t *pte;
1135 	int i;
1136 
1137 	KvaStart = VM_MIN_KERNEL_ADDRESS;
1138 	KvaEnd = VM_MAX_KERNEL_ADDRESS;
1139 	KvaSize = KvaEnd - KvaStart;
1140 
1141 	avail_start = *firstaddr;
1142 
1143 	/*
1144 	 * Create an initial set of page tables to run the kernel in.
1145 	 */
1146 	create_pagetables(firstaddr);
1147 
1148 	virtual2_start = KvaStart;
1149 	virtual2_end = PTOV_OFFSET;
1150 
1151 	virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
1152 	virtual_start = pmap_kmem_choose(virtual_start);
1153 
1154 	virtual_end = VM_MAX_KERNEL_ADDRESS;
1155 
1156 	/* XXX do %cr0 as well */
1157 	load_cr4(rcr4() | CR4_PGE | CR4_PSE);
1158 	load_cr3(KPML4phys);
1159 
1160 	/*
1161 	 * Initialize protection array.
1162 	 */
1163 	x86_64_protection_init();
1164 
1165 	/*
1166 	 * The kernel's pmap is statically allocated so we don't have to use
1167 	 * pmap_create, which is unlikely to work correctly at this part of
1168 	 * the boot sequence (XXX and which no longer exists).
1169 	 */
1170 	kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
1171 	kernel_pmap.pm_count = 1;
1172 	CPUMASK_ASSALLONES(kernel_pmap.pm_active);
1173 	RB_INIT(&kernel_pmap.pm_pvroot);
1174 	spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
1175 	for (i = 0; i < PM_PLACEMARKS; ++i)
1176 		kernel_pmap.pm_placemarks[i] = PM_NOPLACEMARK;
1177 
1178 	/*
1179 	 * Reserve some special page table entries/VA space for temporary
1180 	 * mapping of pages.
1181 	 */
1182 #define	SYSMAP(c, p, v, n)	\
1183 	v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
1184 
1185 	va = virtual_start;
1186 	pte = vtopte(va);
1187 
1188 	/*
1189 	 * CMAP1/CMAP2 are used for zeroing and copying pages.
1190 	 */
1191 	SYSMAP(caddr_t, CMAP1, CADDR1, 1)
1192 
1193 	/*
1194 	 * Crashdump maps.
1195 	 */
1196 	SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
1197 
1198 	/*
1199 	 * ptvmmap is used for reading arbitrary physical pages via
1200 	 * /dev/mem.
1201 	 */
1202 	SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
1203 
1204 	/*
1205 	 * msgbufp is used to map the system message buffer.
1206 	 * XXX msgbufmap is not used.
1207 	 */
1208 	SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
1209 	       atop(round_page(MSGBUF_SIZE)))
1210 
1211 	virtual_start = va;
1212 	virtual_start = pmap_kmem_choose(virtual_start);
1213 
1214 	*CMAP1 = 0;
1215 
1216 	/*
1217 	 * PG_G is terribly broken on SMP because we IPI invltlb's in some
1218 	 * cases rather then invl1pg.  Actually, I don't even know why it
1219 	 * works under UP because self-referential page table mappings
1220 	 */
1221 //	pgeflag = 0;
1222 
1223 	cpu_invltlb();
1224 
1225 	/* Initialize the PAT MSR */
1226 	pmap_init_pat();
1227 	pmap_pinit_defaults(&kernel_pmap);
1228 
1229 	TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1230 			  &pmap_fast_kernel_cpusync);
1231 
1232 }
1233 
1234 /*
1235  * Setup the PAT MSR.
1236  */
1237 void
1238 pmap_init_pat(void)
1239 {
1240 	uint64_t pat_msr;
1241 	u_long cr0, cr4;
1242 	int i;
1243 
1244 	/*
1245 	 * Default values mapping PATi,PCD,PWT bits at system reset.
1246 	 * The default values effectively ignore the PATi bit by
1247 	 * repeating the encodings for 0-3 in 4-7, and map the PCD
1248 	 * and PWT bit combinations to the expected PAT types.
1249 	 */
1250 	pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) |	/* 000 */
1251 		  PAT_VALUE(1, PAT_WRITE_THROUGH) |	/* 001 */
1252 		  PAT_VALUE(2, PAT_UNCACHED) |		/* 010 */
1253 		  PAT_VALUE(3, PAT_UNCACHEABLE) |	/* 011 */
1254 		  PAT_VALUE(4, PAT_WRITE_BACK) |	/* 100 */
1255 		  PAT_VALUE(5, PAT_WRITE_THROUGH) |	/* 101 */
1256 		  PAT_VALUE(6, PAT_UNCACHED) |		/* 110 */
1257 		  PAT_VALUE(7, PAT_UNCACHEABLE);	/* 111 */
1258 	pat_pte_index[PAT_WRITE_BACK]	= 0;
1259 	pat_pte_index[PAT_WRITE_THROUGH]= 0         | X86_PG_NC_PWT;
1260 	pat_pte_index[PAT_UNCACHED]	= X86_PG_NC_PCD;
1261 	pat_pte_index[PAT_UNCACHEABLE]	= X86_PG_NC_PCD | X86_PG_NC_PWT;
1262 	pat_pte_index[PAT_WRITE_PROTECTED] = pat_pte_index[PAT_UNCACHEABLE];
1263 	pat_pte_index[PAT_WRITE_COMBINING] = pat_pte_index[PAT_UNCACHEABLE];
1264 
1265 	if (cpu_feature & CPUID_PAT) {
1266 		/*
1267 		 * If we support the PAT then set-up entries for
1268 		 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1269 		 * 5 and 6.
1270 		 */
1271 		pat_msr = (pat_msr & ~PAT_MASK(5)) |
1272 			  PAT_VALUE(5, PAT_WRITE_PROTECTED);
1273 		pat_msr = (pat_msr & ~PAT_MASK(6)) |
1274 			  PAT_VALUE(6, PAT_WRITE_COMBINING);
1275 		pat_pte_index[PAT_WRITE_PROTECTED] = X86_PG_PTE_PAT | X86_PG_NC_PWT;
1276 		pat_pte_index[PAT_WRITE_COMBINING] = X86_PG_PTE_PAT | X86_PG_NC_PCD;
1277 
1278 		/*
1279 		 * Then enable the PAT
1280 		 */
1281 
1282 		/* Disable PGE. */
1283 		cr4 = rcr4();
1284 		load_cr4(cr4 & ~CR4_PGE);
1285 
1286 		/* Disable caches (CD = 1, NW = 0). */
1287 		cr0 = rcr0();
1288 		load_cr0((cr0 & ~CR0_NW) | CR0_CD);
1289 
1290 		/* Flushes caches and TLBs. */
1291 		wbinvd();
1292 		cpu_invltlb();
1293 
1294 		/* Update PAT and index table. */
1295 		wrmsr(MSR_PAT, pat_msr);
1296 
1297 		/* Flush caches and TLBs again. */
1298 		wbinvd();
1299 		cpu_invltlb();
1300 
1301 		/* Restore caches and PGE. */
1302 		load_cr0(cr0);
1303 		load_cr4(cr4);
1304 		PatMsr = pat_msr;
1305 	}
1306 
1307 	for (i = 0; i < 8; ++i) {
1308 		pt_entry_t pte;
1309 
1310 		pte = pat_pte_index[i];
1311 		if (pte & X86_PG_PTE_PAT) {
1312 			pte &= ~X86_PG_PTE_PAT;
1313 			pte |= X86_PG_PDE_PAT;
1314 		}
1315 		pat_pde_index[i] = pte;
1316 	}
1317 }
1318 
1319 /*
1320  * Set 4mb pdir for mp startup
1321  */
1322 void
1323 pmap_set_opt(void)
1324 {
1325 	if (cpu_feature & CPUID_PSE) {
1326 		load_cr4(rcr4() | CR4_PSE);
1327 		if (mycpu->gd_cpuid == 0) 	/* only on BSP */
1328 			cpu_invltlb();
1329 	}
1330 
1331 	/*
1332 	 * Check for SMAP support and enable if available.  Must be done
1333 	 * after cr3 is loaded, and on all cores.
1334 	 */
1335 	if (cpu_stdext_feature & CPUID_STDEXT_SMAP) {
1336 		load_cr4(rcr4() | CR4_SMAP);
1337 	}
1338 	if (cpu_stdext_feature & CPUID_STDEXT_SMEP) {
1339 		load_cr4(rcr4() | CR4_SMEP);
1340 	}
1341 }
1342 
1343 /*
1344  * Early initialization of the pmap module.
1345  *
1346  * Called by vm_init, to initialize any structures that the pmap
1347  * system needs to map virtual memory.  pmap_init has been enhanced to
1348  * support in a fairly consistant way, discontiguous physical memory.
1349  */
1350 void
1351 pmap_init(void)
1352 {
1353 	vm_pindex_t initial_pvs;
1354 	vm_pindex_t i;
1355 
1356 	/*
1357 	 * Allocate memory for random pmap data structures.  Includes the
1358 	 * pv_head_table.
1359 	 */
1360 	for (i = 0; i < vm_page_array_size; i++) {
1361 		vm_page_t m;
1362 
1363 		m = &vm_page_array[i];
1364 		m->md.pmap_count = 0;
1365 		m->md.writeable_count = 0;
1366 	}
1367 
1368 	/*
1369 	 * init the pv free list
1370 	 */
1371 	initial_pvs = vm_page_array_size;
1372 	if (initial_pvs < MINPV)
1373 		initial_pvs = MINPV;
1374 	pvzone = &pvzone_store;
1375 	pvinit = (void *)kmem_alloc(&kernel_map,
1376 				    initial_pvs * sizeof (struct pv_entry),
1377 				    VM_SUBSYS_PVENTRY);
1378 	zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
1379 		  pvinit, initial_pvs);
1380 
1381 	/*
1382 	 * Now it is safe to enable pv_table recording.
1383 	 */
1384 	pmap_initialized = TRUE;
1385 }
1386 
1387 /*
1388  * Initialize the address space (zone) for the pv_entries.  Set a
1389  * high water mark so that the system can recover from excessive
1390  * numbers of pv entries.
1391  *
1392  * Also create the kernel page table template for isolated user
1393  * pmaps.
1394  */
1395 static void pmap_init_iso_range(vm_offset_t base, size_t bytes);
1396 static void pmap_init2_iso_pmap(void);
1397 #if 0
1398 static void dump_pmap(pmap_t pmap, pt_entry_t pte, int level, vm_offset_t base);
1399 #endif
1400 
1401 void
1402 pmap_init2(void)
1403 {
1404 	vm_pindex_t entry_max;
1405 
1406 	/*
1407 	 * We can significantly reduce pv_entry_max from historical
1408 	 * levels because pv_entry's are no longer use for PTEs at the
1409 	 * leafs.  This prevents excessive pcpu caching on many-core
1410 	 * boxes (even with the further '/ 16' done in zinitna().
1411 	 *
1412 	 * Remember, however, that processes can share physical pages
1413 	 * with each process still needing the pdp/pd/pt infrstructure
1414 	 * (which still use pv_entry's).  And don't just assume that
1415 	 * every PT will be completely filled up.  So don't make it
1416 	 * too small.
1417 	 */
1418 	entry_max = maxproc * 32 + vm_page_array_size / 16;
1419 	TUNABLE_LONG_FETCH("vm.pmap.pv_entries", &entry_max);
1420 	vm_pmap_pv_entries = entry_max;
1421 
1422 	/*
1423 	 * Subtract out pages already installed in the zone (hack)
1424 	 */
1425 	if (entry_max <= MINPV)
1426 		entry_max = MINPV;
1427 
1428 	zinitna(pvzone, NULL, 0, entry_max, ZONE_INTERRUPT);
1429 
1430 	/*
1431 	 * Enable dynamic deletion of empty higher-level page table pages
1432 	 * by default only if system memory is < 8GB (use 7GB for slop).
1433 	 * This can save a little memory, but imposes significant
1434 	 * performance overhead for things like bulk builds, and for programs
1435 	 * which do a lot of memory mapping and memory unmapping.
1436 	 */
1437 #if 0
1438 	if (pmap_dynamic_delete < 0) {
1439 		if (vmstats.v_page_count < 7LL * 1024 * 1024 * 1024 / PAGE_SIZE)
1440 			pmap_dynamic_delete = 1;
1441 		else
1442 			pmap_dynamic_delete = 0;
1443 	}
1444 #endif
1445 	/*
1446 	 * Disable so vm_map_backing iterations do not race
1447 	 */
1448 	pmap_dynamic_delete = 0;
1449 
1450 	/*
1451 	 * Automatic detection of Intel meltdown bug requiring user/kernel
1452 	 * mmap isolation.
1453 	 *
1454 	 * Currently there are so many Intel cpu's impacted that its better
1455 	 * to whitelist future Intel CPUs.  Most? AMD cpus are not impacted
1456 	 * so the default is off for AMD.
1457 	 */
1458 	if (meltdown_mitigation < 0) {
1459 		if (cpu_vendor_id == CPU_VENDOR_INTEL)
1460 			meltdown_mitigation = 1;
1461 		else
1462 			meltdown_mitigation = 0;
1463 	}
1464 	if (meltdown_mitigation) {
1465 		kprintf("machdep.meltdown_mitigation enabled to "
1466 			"protect against (mostly Intel) meltdown bug\n");
1467 		kprintf("system call performance will be impacted\n");
1468 	}
1469 
1470 	pmap_init2_iso_pmap();
1471 }
1472 
1473 /*
1474  * Create the isolation pmap template.  Once created, the template
1475  * is static and its PML4e entries are used to populate the
1476  * kernel portion of any isolated user pmaps.
1477  *
1478  * Our isolation pmap must contain:
1479  * (1) trampoline area for all cpus
1480  * (2) common_tss area for all cpus (its part of the trampoline area now)
1481  * (3) IDT for all cpus
1482  * (4) GDT for all cpus
1483  */
1484 static void
1485 pmap_init2_iso_pmap(void)
1486 {
1487 	int n;
1488 
1489 	if (bootverbose)
1490 		kprintf("Initialize isolation pmap\n");
1491 
1492 	/*
1493 	 * Try to use our normal API calls to make this easier.  We have
1494 	 * to scrap the shadowed kernel PDPs pmap_pinit() creates for our
1495 	 * iso_pmap.
1496 	 */
1497 	pmap_pinit(&iso_pmap);
1498 	bzero(iso_pmap.pm_pml4, PAGE_SIZE);
1499 
1500 	/*
1501 	 * Install areas needed by the cpu and trampoline.
1502 	 */
1503 	for (n = 0; n < ncpus; ++n) {
1504 		struct privatespace *ps;
1505 
1506 		ps = CPU_prvspace[n];
1507 		pmap_init_iso_range((vm_offset_t)&ps->trampoline,
1508 				    sizeof(ps->trampoline));
1509 		pmap_init_iso_range((vm_offset_t)&ps->dblstack,
1510 				    sizeof(ps->dblstack));
1511 		pmap_init_iso_range((vm_offset_t)&ps->dbgstack,
1512 				    sizeof(ps->dbgstack));
1513 		pmap_init_iso_range((vm_offset_t)&ps->common_tss,
1514 				    sizeof(ps->common_tss));
1515 		pmap_init_iso_range(r_idt_arr[n].rd_base,
1516 				    r_idt_arr[n].rd_limit + 1);
1517 	}
1518 	pmap_init_iso_range((register_t)gdt, sizeof(gdt));
1519 	pmap_init_iso_range((vm_offset_t)(int *)btext,
1520 			    (vm_offset_t)(int *)etext -
1521 			     (vm_offset_t)(int *)btext);
1522 
1523 #if 0
1524 	kprintf("Dump iso_pmap:\n");
1525 	dump_pmap(&iso_pmap, vtophys(iso_pmap.pm_pml4), 0, 0);
1526 	kprintf("\nDump kernel_pmap:\n");
1527 	dump_pmap(&kernel_pmap, vtophys(kernel_pmap.pm_pml4), 0, 0);
1528 #endif
1529 }
1530 
1531 /*
1532  * This adds a kernel virtual address range to the isolation pmap.
1533  */
1534 static void
1535 pmap_init_iso_range(vm_offset_t base, size_t bytes)
1536 {
1537 	pv_entry_t pv;
1538 	pv_entry_t pvp;
1539 	pt_entry_t *ptep;
1540 	pt_entry_t pte;
1541 	vm_offset_t va;
1542 
1543 	if (bootverbose) {
1544 		kprintf("isolate %016jx-%016jx (%zd)\n",
1545 			base, base + bytes, bytes);
1546 	}
1547 	va = base & ~(vm_offset_t)PAGE_MASK;
1548 	while (va < base + bytes) {
1549 		if ((va & PDRMASK) == 0 && va + NBPDR <= base + bytes &&
1550 		    (ptep = pmap_pt(&kernel_pmap, va)) != NULL &&
1551 		    (*ptep & kernel_pmap.pmap_bits[PG_V_IDX]) &&
1552 		    (*ptep & kernel_pmap.pmap_bits[PG_PS_IDX])) {
1553 			/*
1554 			 * Use 2MB pages if possible
1555 			 */
1556 			pte = *ptep;
1557 			pv = pmap_allocpte(&iso_pmap, pmap_pd_pindex(va), &pvp);
1558 			ptep = pv_pte_lookup(pv, (va >> PDRSHIFT) & 511);
1559 			*ptep = pte;
1560 			va += NBPDR;
1561 		} else {
1562 			/*
1563 			 * Otherwise use 4KB pages
1564 			 */
1565 			pv = pmap_allocpte(&iso_pmap, pmap_pt_pindex(va), &pvp);
1566 			ptep = pv_pte_lookup(pv, (va >> PAGE_SHIFT) & 511);
1567 			*ptep = vtophys(va) | kernel_pmap.pmap_bits[PG_RW_IDX] |
1568 					      kernel_pmap.pmap_bits[PG_V_IDX] |
1569 					      kernel_pmap.pmap_bits[PG_A_IDX] |
1570 					      kernel_pmap.pmap_bits[PG_M_IDX];
1571 
1572 			va += PAGE_SIZE;
1573 		}
1574 		pv_put(pv);
1575 		pv_put(pvp);
1576 	}
1577 }
1578 
1579 #if 0
1580 /*
1581  * Useful debugging pmap dumper, do not remove (#if 0 when not in use)
1582  */
1583 static
1584 void
1585 dump_pmap(pmap_t pmap, pt_entry_t pte, int level, vm_offset_t base)
1586 {
1587 	pt_entry_t *ptp;
1588 	vm_offset_t incr;
1589 	int i;
1590 
1591 	switch(level) {
1592 	case 0:					/* PML4e page, 512G entries */
1593 		incr = (1LL << 48) / 512;
1594 		break;
1595 	case 1:					/* PDP page, 1G entries */
1596 		incr = (1LL << 39) / 512;
1597 		break;
1598 	case 2:					/* PD page, 2MB entries */
1599 		incr = (1LL << 30) / 512;
1600 		break;
1601 	case 3:					/* PT page, 4KB entries */
1602 		incr = (1LL << 21) / 512;
1603 		break;
1604 	default:
1605 		incr = 0;
1606 		break;
1607 	}
1608 
1609 	if (level == 0)
1610 		kprintf("cr3 %016jx @ va=%016jx\n", pte, base);
1611 	ptp = (void *)PHYS_TO_DMAP(pte & ~(pt_entry_t)PAGE_MASK);
1612 	for (i = 0; i < 512; ++i) {
1613 		if (level == 0 && i == 128)
1614 			base += 0xFFFF000000000000LLU;
1615 		if (ptp[i]) {
1616 			kprintf("%*.*s ", level * 4, level * 4, "");
1617 			if (level == 1 && (ptp[i] & 0x180) == 0x180) {
1618 				kprintf("va=%016jx %3d term %016jx (1GB)\n",
1619 					base, i, ptp[i]);
1620 			} else if (level == 2 && (ptp[i] & 0x180) == 0x180) {
1621 				kprintf("va=%016jx %3d term %016jx (2MB)\n",
1622 					base, i, ptp[i]);
1623 			} else if (level == 3) {
1624 				kprintf("va=%016jx %3d term %016jx\n",
1625 					base, i, ptp[i]);
1626 			} else {
1627 				kprintf("va=%016jx %3d deep %016jx\n",
1628 					base, i, ptp[i]);
1629 				dump_pmap(pmap, ptp[i], level + 1, base);
1630 			}
1631 		}
1632 		base += incr;
1633 	}
1634 }
1635 
1636 #endif
1637 
1638 /*
1639  * Typically used to initialize a fictitious page by vm/device_pager.c
1640  */
1641 void
1642 pmap_page_init(struct vm_page *m)
1643 {
1644 	vm_page_init(m);
1645 	m->md.pmap_count = 0;
1646 	m->md.writeable_count = 0;
1647 }
1648 
1649 /***************************************************
1650  * Low level helper routines.....
1651  ***************************************************/
1652 
1653 /*
1654  * Extract the physical page address associated with the map/VA pair.
1655  * The page must be wired for this to work reliably.
1656  */
1657 vm_paddr_t
1658 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
1659 {
1660 	vm_paddr_t rtval;
1661 	pv_entry_t pt_pv;
1662 	pt_entry_t *ptep;
1663 
1664 	rtval = 0;
1665 	if (va >= VM_MAX_USER_ADDRESS) {
1666 		/*
1667 		 * Kernel page directories might be direct-mapped and
1668 		 * there is typically no PV tracking of pte's
1669 		 */
1670 		pd_entry_t *pt;
1671 
1672 		pt = pmap_pt(pmap, va);
1673 		if (pt && (*pt & pmap->pmap_bits[PG_V_IDX])) {
1674 			if (*pt & pmap->pmap_bits[PG_PS_IDX]) {
1675 				rtval = *pt & PG_PS_FRAME;
1676 				rtval |= va & PDRMASK;
1677 			} else {
1678 				ptep = pmap_pt_to_pte(*pt, va);
1679 				if (*pt & pmap->pmap_bits[PG_V_IDX]) {
1680 					rtval = *ptep & PG_FRAME;
1681 					rtval |= va & PAGE_MASK;
1682 				}
1683 			}
1684 		}
1685 		if (handlep)
1686 			*handlep = NULL;
1687 	} else {
1688 		/*
1689 		 * User pages currently do not direct-map the page directory
1690 		 * and some pages might not used managed PVs.  But all PT's
1691 		 * will have a PV.
1692 		 */
1693 		pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1694 		if (pt_pv) {
1695 			ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1696 			if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
1697 				rtval = *ptep & PG_FRAME;
1698 				rtval |= va & PAGE_MASK;
1699 			}
1700 			if (handlep)
1701 				*handlep = pt_pv;	/* locked until done */
1702 			else
1703 				pv_put (pt_pv);
1704 		} else if (handlep) {
1705 			*handlep = NULL;
1706 		}
1707 	}
1708 	return rtval;
1709 }
1710 
1711 void
1712 pmap_extract_done(void *handle)
1713 {
1714 	if (handle)
1715 		pv_put((pv_entry_t)handle);
1716 }
1717 
1718 /*
1719  * Similar to extract but checks protections, SMP-friendly short-cut for
1720  * vm_fault_page[_quick]().  Can return NULL to cause the caller to
1721  * fall-through to the real fault code.  Does not work with HVM page
1722  * tables.
1723  *
1724  * if busyp is NULL the returned page, if not NULL, is held (and not busied).
1725  *
1726  * If busyp is not NULL and this function sets *busyp non-zero, the returned
1727  * page is busied (and not held).
1728  *
1729  * If busyp is not NULL and this function sets *busyp to zero, the returned
1730  * page is held (and not busied).
1731  *
1732  * If VM_PROT_WRITE is set in prot, and the pte is already writable, the
1733  * returned page will be dirtied.  If the pte is not already writable NULL
1734  * is returned.  In otherwords, if the bit is set and a vm_page_t is returned,
1735  * any COW will already have happened and that page can be written by the
1736  * caller.
1737  *
1738  * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1739  *	    OR WRITING AS-IS.
1740  */
1741 vm_page_t
1742 pmap_fault_page_quick(pmap_t pmap, vm_offset_t va, vm_prot_t prot, int *busyp)
1743 {
1744 	if (pmap &&
1745 	    va < VM_MAX_USER_ADDRESS &&
1746 	    (pmap->pm_flags & PMAP_HVM) == 0) {
1747 		pv_entry_t pt_pv;
1748 		pv_entry_t pte_pv;
1749 		pt_entry_t *ptep;
1750 		pt_entry_t req;
1751 		vm_page_t m;
1752 		int error;
1753 
1754 		req = pmap->pmap_bits[PG_V_IDX] |
1755 		      pmap->pmap_bits[PG_U_IDX];
1756 		if (prot & VM_PROT_WRITE)
1757 			req |= pmap->pmap_bits[PG_RW_IDX];
1758 
1759 		pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
1760 		if (pt_pv == NULL)
1761 			return (NULL);
1762 		ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1763 		if ((*ptep & req) != req) {
1764 			pv_put(pt_pv);
1765 			return (NULL);
1766 		}
1767 		pte_pv = pv_get_try(pmap, pmap_pte_pindex(va), NULL, &error);
1768 		if (pte_pv && error == 0) {
1769 			m = pte_pv->pv_m;
1770 			if (prot & VM_PROT_WRITE) {
1771 				/* interlocked by presence of pv_entry */
1772 				vm_page_dirty(m);
1773 			}
1774 			if (busyp) {
1775 				if (prot & VM_PROT_WRITE) {
1776 					if (vm_page_busy_try(m, TRUE))
1777 						m = NULL;
1778 					*busyp = 1;
1779 				} else {
1780 					vm_page_hold(m);
1781 					*busyp = 0;
1782 				}
1783 			} else {
1784 				vm_page_hold(m);
1785 			}
1786 			pv_put(pte_pv);
1787 		} else if (pte_pv) {
1788 			pv_drop(pte_pv);
1789 			m = NULL;
1790 		} else {
1791 			/* error, since we didn't request a placemarker */
1792 			m = NULL;
1793 		}
1794 		pv_put(pt_pv);
1795 		return(m);
1796 	} else {
1797 		return(NULL);
1798 	}
1799 }
1800 
1801 /*
1802  * Extract the physical page address associated kernel virtual address.
1803  */
1804 vm_paddr_t
1805 pmap_kextract(vm_offset_t va)
1806 {
1807 	pd_entry_t pt;		/* pt entry in pd */
1808 	vm_paddr_t pa;
1809 
1810 	if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1811 		pa = DMAP_TO_PHYS(va);
1812 	} else {
1813 		pt = *vtopt(va);
1814 		if (pt & kernel_pmap.pmap_bits[PG_PS_IDX]) {
1815 			pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
1816 		} else {
1817 			/*
1818 			 * Beware of a concurrent promotion that changes the
1819 			 * PDE at this point!  For example, vtopte() must not
1820 			 * be used to access the PTE because it would use the
1821 			 * new PDE.  It is, however, safe to use the old PDE
1822 			 * because the page table page is preserved by the
1823 			 * promotion.
1824 			 */
1825 			pa = *pmap_pt_to_pte(pt, va);
1826 			pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1827 		}
1828 	}
1829 	return pa;
1830 }
1831 
1832 /***************************************************
1833  * Low level mapping routines.....
1834  ***************************************************/
1835 
1836 /*
1837  * Routine: pmap_kenter
1838  * Function:
1839  *  	Add a wired page to the KVA
1840  *  	NOTE! note that in order for the mapping to take effect -- you
1841  *  	should do an invltlb after doing the pmap_kenter().
1842  */
1843 void
1844 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1845 {
1846 	pt_entry_t *ptep;
1847 	pt_entry_t npte;
1848 
1849 	npte = pa |
1850 	       kernel_pmap.pmap_bits[PG_RW_IDX] |
1851 	       kernel_pmap.pmap_bits[PG_V_IDX];
1852 //	       pgeflag;
1853 	ptep = vtopte(va);
1854 #if 1
1855 	pmap_inval_smp(&kernel_pmap, va, 1, ptep, npte);
1856 #else
1857 	/* FUTURE */
1858 	if (*ptep)
1859 		pmap_inval_smp(&kernel_pmap, va, ptep, npte);
1860 	else
1861 		*ptep = npte;
1862 #endif
1863 }
1864 
1865 /*
1866  * Similar to pmap_kenter(), except we only invalidate the mapping on the
1867  * current CPU.  Returns 0 if the previous pte was 0, 1 if it wasn't
1868  * (caller can conditionalize calling smp_invltlb()).
1869  */
1870 int
1871 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1872 {
1873 	pt_entry_t *ptep;
1874 	pt_entry_t npte;
1875 	int res;
1876 
1877 	npte = pa | kernel_pmap.pmap_bits[PG_RW_IDX] |
1878 		    kernel_pmap.pmap_bits[PG_V_IDX];
1879 	// npte |= pgeflag;
1880 	ptep = vtopte(va);
1881 #if 1
1882 	res = 1;
1883 #else
1884 	/* FUTURE */
1885 	res = (*ptep != 0);
1886 #endif
1887 	atomic_swap_long(ptep, npte);
1888 	cpu_invlpg((void *)va);
1889 
1890 	return res;
1891 }
1892 
1893 /*
1894  * Enter addresses into the kernel pmap but don't bother
1895  * doing any tlb invalidations.  Caller will do a rollup
1896  * invalidation via pmap_rollup_inval().
1897  */
1898 int
1899 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
1900 {
1901 	pt_entry_t *ptep;
1902 	pt_entry_t npte;
1903 	int res;
1904 
1905 	npte = pa |
1906 	    kernel_pmap.pmap_bits[PG_RW_IDX] |
1907 	    kernel_pmap.pmap_bits[PG_V_IDX];
1908 //	    pgeflag;
1909 	ptep = vtopte(va);
1910 #if 1
1911 	res = 1;
1912 #else
1913 	/* FUTURE */
1914 	res = (*ptep != 0);
1915 #endif
1916 	atomic_swap_long(ptep, npte);
1917 	cpu_invlpg((void *)va);
1918 
1919 	return res;
1920 }
1921 
1922 /*
1923  * remove a page from the kernel pagetables
1924  */
1925 void
1926 pmap_kremove(vm_offset_t va)
1927 {
1928 	pt_entry_t *ptep;
1929 
1930 	ptep = vtopte(va);
1931 	pmap_inval_smp(&kernel_pmap, va, 1, ptep, 0);
1932 }
1933 
1934 void
1935 pmap_kremove_quick(vm_offset_t va)
1936 {
1937 	pt_entry_t *ptep;
1938 
1939 	ptep = vtopte(va);
1940 	(void)pte_load_clear(ptep);
1941 	cpu_invlpg((void *)va);
1942 }
1943 
1944 /*
1945  * Remove addresses from the kernel pmap but don't bother
1946  * doing any tlb invalidations.  Caller will do a rollup
1947  * invalidation via pmap_rollup_inval().
1948  */
1949 void
1950 pmap_kremove_noinval(vm_offset_t va)
1951 {
1952 	pt_entry_t *ptep;
1953 
1954 	ptep = vtopte(va);
1955 	(void)pte_load_clear(ptep);
1956 }
1957 
1958 /*
1959  * XXX these need to be recoded.  They are not used in any critical path.
1960  */
1961 void
1962 pmap_kmodify_rw(vm_offset_t va)
1963 {
1964 	atomic_set_long(vtopte(va), kernel_pmap.pmap_bits[PG_RW_IDX]);
1965 	cpu_invlpg((void *)va);
1966 }
1967 
1968 /* NOT USED
1969 void
1970 pmap_kmodify_nc(vm_offset_t va)
1971 {
1972 	atomic_set_long(vtopte(va), PG_N);
1973 	cpu_invlpg((void *)va);
1974 }
1975 */
1976 
1977 /*
1978  * Used to map a range of physical addresses into kernel virtual
1979  * address space during the low level boot, typically to map the
1980  * dump bitmap, message buffer, and vm_page_array.
1981  *
1982  * These mappings are typically made at some pointer after the end of the
1983  * kernel text+data.
1984  *
1985  * We could return PHYS_TO_DMAP(start) here and not allocate any
1986  * via (*virtp), but then kmem from userland and kernel dumps won't
1987  * have access to the related pointers.
1988  */
1989 vm_offset_t
1990 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1991 {
1992 	vm_offset_t va;
1993 	vm_offset_t va_start;
1994 
1995 	/*return PHYS_TO_DMAP(start);*/
1996 
1997 	va_start = *virtp;
1998 	va = va_start;
1999 
2000 	while (start < end) {
2001 		pmap_kenter_quick(va, start);
2002 		va += PAGE_SIZE;
2003 		start += PAGE_SIZE;
2004 	}
2005 	*virtp = va;
2006 	return va_start;
2007 }
2008 
2009 #define PMAP_CLFLUSH_THRESHOLD  (2 * 1024 * 1024)
2010 
2011 /*
2012  * Remove the specified set of pages from the data and instruction caches.
2013  *
2014  * In contrast to pmap_invalidate_cache_range(), this function does not
2015  * rely on the CPU's self-snoop feature, because it is intended for use
2016  * when moving pages into a different cache domain.
2017  */
2018 void
2019 pmap_invalidate_cache_pages(vm_page_t *pages, int count)
2020 {
2021 	vm_offset_t daddr, eva;
2022 	int i;
2023 
2024 	if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE ||
2025 	    (cpu_feature & CPUID_CLFSH) == 0)
2026 		wbinvd();
2027 	else {
2028 		cpu_mfence();
2029 		for (i = 0; i < count; i++) {
2030 			daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i]));
2031 			eva = daddr + PAGE_SIZE;
2032 			for (; daddr < eva; daddr += cpu_clflush_line_size)
2033 				clflush(daddr);
2034 		}
2035 		cpu_mfence();
2036 	}
2037 }
2038 
2039 void
2040 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
2041 {
2042 	KASSERT((sva & PAGE_MASK) == 0,
2043 	    ("pmap_invalidate_cache_range: sva not page-aligned"));
2044 	KASSERT((eva & PAGE_MASK) == 0,
2045 	    ("pmap_invalidate_cache_range: eva not page-aligned"));
2046 
2047 	if (cpu_feature & CPUID_SS) {
2048 		; /* If "Self Snoop" is supported, do nothing. */
2049 	} else {
2050 		/* Globally invalidate caches */
2051 		cpu_wbinvd_on_all_cpus();
2052 	}
2053 }
2054 
2055 /*
2056  * Invalidate the specified range of virtual memory on all cpus associated
2057  * with the pmap.
2058  */
2059 void
2060 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2061 {
2062 	pmap_inval_smp(pmap, sva, (eva - sva) >> PAGE_SHIFT, NULL, 0);
2063 }
2064 
2065 /*
2066  * Add a list of wired pages to the kva.  This routine is used for temporary
2067  * kernel mappings such as those found in buffer cache buffer.  Page
2068  * modifications and accesses are not tracked or recorded.
2069  *
2070  * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
2071  *	 semantics as previous mappings may have been zerod without any
2072  *	 invalidation.
2073  *
2074  * The page *must* be wired.
2075  */
2076 static __inline void
2077 _pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count, int doinval)
2078 {
2079 	vm_offset_t end_va;
2080 	vm_offset_t va;
2081 
2082 	end_va = beg_va + count * PAGE_SIZE;
2083 
2084 	for (va = beg_va; va < end_va; va += PAGE_SIZE) {
2085 		pt_entry_t pte;
2086 		pt_entry_t *ptep;
2087 
2088 		ptep = vtopte(va);
2089 		pte = VM_PAGE_TO_PHYS(*m) |
2090 			kernel_pmap.pmap_bits[PG_RW_IDX] |
2091 			kernel_pmap.pmap_bits[PG_V_IDX] |
2092 			kernel_pmap.pmap_cache_bits_pte[(*m)->pat_mode];
2093 //		pgeflag;
2094 		atomic_swap_long(ptep, pte);
2095 		m++;
2096 	}
2097 	if (doinval)
2098 		pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
2099 }
2100 
2101 void
2102 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
2103 {
2104 	_pmap_qenter(beg_va, m, count, 1);
2105 }
2106 
2107 void
2108 pmap_qenter_noinval(vm_offset_t beg_va, vm_page_t *m, int count)
2109 {
2110 	_pmap_qenter(beg_va, m, count, 0);
2111 }
2112 
2113 /*
2114  * This routine jerks page mappings from the kernel -- it is meant only
2115  * for temporary mappings such as those found in buffer cache buffers.
2116  * No recording modified or access status occurs.
2117  *
2118  * MPSAFE, INTERRUPT SAFE (cluster callback)
2119  */
2120 void
2121 pmap_qremove(vm_offset_t beg_va, int count)
2122 {
2123 	vm_offset_t end_va;
2124 	vm_offset_t va;
2125 
2126 	end_va = beg_va + count * PAGE_SIZE;
2127 
2128 	for (va = beg_va; va < end_va; va += PAGE_SIZE) {
2129 		pt_entry_t *pte;
2130 
2131 		pte = vtopte(va);
2132 		(void)pte_load_clear(pte);
2133 		cpu_invlpg((void *)va);
2134 	}
2135 	pmap_invalidate_range(&kernel_pmap, beg_va, end_va);
2136 }
2137 
2138 /*
2139  * This routine removes temporary kernel mappings, only invalidating them
2140  * on the current cpu.  It should only be used under carefully controlled
2141  * conditions.
2142  */
2143 void
2144 pmap_qremove_quick(vm_offset_t beg_va, int count)
2145 {
2146 	vm_offset_t end_va;
2147 	vm_offset_t va;
2148 
2149 	end_va = beg_va + count * PAGE_SIZE;
2150 
2151 	for (va = beg_va; va < end_va; va += PAGE_SIZE) {
2152 		pt_entry_t *pte;
2153 
2154 		pte = vtopte(va);
2155 		(void)pte_load_clear(pte);
2156 		cpu_invlpg((void *)va);
2157 	}
2158 }
2159 
2160 /*
2161  * This routine removes temporary kernel mappings *without* invalidating
2162  * the TLB.  It can only be used on permanent kva reservations such as those
2163  * found in buffer cache buffers, under carefully controlled circumstances.
2164  *
2165  * NOTE: Repopulating these KVAs requires unconditional invalidation.
2166  *	 (pmap_qenter() does unconditional invalidation).
2167  */
2168 void
2169 pmap_qremove_noinval(vm_offset_t beg_va, int count)
2170 {
2171 	vm_offset_t end_va;
2172 	vm_offset_t va;
2173 
2174 	end_va = beg_va + count * PAGE_SIZE;
2175 
2176 	for (va = beg_va; va < end_va; va += PAGE_SIZE) {
2177 		pt_entry_t *pte;
2178 
2179 		pte = vtopte(va);
2180 		(void)pte_load_clear(pte);
2181 	}
2182 }
2183 
2184 /*
2185  * Create a new thread and optionally associate it with a (new) process.
2186  * NOTE! the new thread's cpu may not equal the current cpu.
2187  */
2188 void
2189 pmap_init_thread(thread_t td)
2190 {
2191 	/* enforce pcb placement & alignment */
2192 	td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
2193 	td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
2194 	td->td_savefpu = &td->td_pcb->pcb_save;
2195 	td->td_sp = (char *)td->td_pcb;	/* no -16 */
2196 }
2197 
2198 /*
2199  * This routine directly affects the fork perf for a process.
2200  */
2201 void
2202 pmap_init_proc(struct proc *p)
2203 {
2204 }
2205 
2206 static void
2207 pmap_pinit_defaults(struct pmap *pmap)
2208 {
2209 	bcopy(pmap_bits_default, pmap->pmap_bits,
2210 	      sizeof(pmap_bits_default));
2211 	bcopy(protection_codes, pmap->protection_codes,
2212 	      sizeof(protection_codes));
2213 	bcopy(pat_pte_index, pmap->pmap_cache_bits_pte,
2214 	      sizeof(pat_pte_index));
2215 	bcopy(pat_pde_index, pmap->pmap_cache_bits_pde,
2216 	      sizeof(pat_pte_index));
2217 	pmap->pmap_cache_mask_pte = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PTE_PAT;
2218 	pmap->pmap_cache_mask_pde = X86_PG_NC_PWT | X86_PG_NC_PCD | X86_PG_PDE_PAT;
2219 	pmap->copyinstr = std_copyinstr;
2220 	pmap->copyin = std_copyin;
2221 	pmap->copyout = std_copyout;
2222 	pmap->fubyte = std_fubyte;
2223 	pmap->subyte = std_subyte;
2224 	pmap->fuword32 = std_fuword32;
2225 	pmap->fuword64 = std_fuword64;
2226 	pmap->suword32 = std_suword32;
2227 	pmap->suword64 = std_suword64;
2228 	pmap->swapu32 = std_swapu32;
2229 	pmap->swapu64 = std_swapu64;
2230 	pmap->fuwordadd32 = std_fuwordadd32;
2231 	pmap->fuwordadd64 = std_fuwordadd64;
2232 }
2233 /*
2234  * Initialize pmap0/vmspace0.
2235  *
2236  * On architectures where the kernel pmap is not integrated into the user
2237  * process pmap, this pmap represents the process pmap, not the kernel pmap.
2238  * kernel_pmap should be used to directly access the kernel_pmap.
2239  */
2240 void
2241 pmap_pinit0(struct pmap *pmap)
2242 {
2243 	int i;
2244 
2245 	pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
2246 	pmap->pm_count = 1;
2247 	CPUMASK_ASSZERO(pmap->pm_active);
2248 	pmap->pm_pvhint_pt = NULL;
2249 	pmap->pm_pvhint_unused = NULL;
2250 	RB_INIT(&pmap->pm_pvroot);
2251 	spin_init(&pmap->pm_spin, "pmapinit0");
2252 	for (i = 0; i < PM_PLACEMARKS; ++i)
2253 		pmap->pm_placemarks[i] = PM_NOPLACEMARK;
2254 	bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
2255 	pmap_pinit_defaults(pmap);
2256 }
2257 
2258 /*
2259  * Initialize a preallocated and zeroed pmap structure,
2260  * such as one in a vmspace structure.
2261  */
2262 static void
2263 pmap_pinit_simple(struct pmap *pmap)
2264 {
2265 	int i;
2266 
2267 	/*
2268 	 * Misc initialization
2269 	 */
2270 	pmap->pm_count = 1;
2271 	CPUMASK_ASSZERO(pmap->pm_active);
2272 	pmap->pm_pvhint_pt = NULL;
2273 	pmap->pm_pvhint_unused = NULL;
2274 	pmap->pm_flags = PMAP_FLAG_SIMPLE;
2275 
2276 	pmap_pinit_defaults(pmap);
2277 
2278 	/*
2279 	 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
2280 	 * for this).
2281 	 */
2282 	if (pmap->pm_pmlpv == NULL) {
2283 		RB_INIT(&pmap->pm_pvroot);
2284 		bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
2285 		spin_init(&pmap->pm_spin, "pmapinitsimple");
2286 		for (i = 0; i < PM_PLACEMARKS; ++i)
2287 			pmap->pm_placemarks[i] = PM_NOPLACEMARK;
2288 	}
2289 }
2290 
2291 void
2292 pmap_pinit(struct pmap *pmap)
2293 {
2294 	pv_entry_t pv;
2295 	int j;
2296 
2297 	if (pmap->pm_pmlpv) {
2298 		if (pmap->pmap_bits[TYPE_IDX] != REGULAR_PMAP) {
2299 			pmap_puninit(pmap);
2300 		}
2301 	}
2302 
2303 	pmap_pinit_simple(pmap);
2304 	pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
2305 
2306 	/*
2307 	 * No need to allocate page table space yet but we do need a valid
2308 	 * page directory table.
2309 	 */
2310 	if (pmap->pm_pml4 == NULL) {
2311 		pmap->pm_pml4 =
2312 		    (pml4_entry_t *)kmem_alloc_pageable(&kernel_map,
2313 							PAGE_SIZE * 2,
2314 							VM_SUBSYS_PML4);
2315 		pmap->pm_pml4_iso = (void *)((char *)pmap->pm_pml4 + PAGE_SIZE);
2316 	}
2317 
2318 	/*
2319 	 * Allocate the PML4e table, which wires it even though it isn't
2320 	 * being entered into some higher level page table (it being the
2321 	 * highest level).  If one is already cached we don't have to do
2322 	 * anything.
2323 	 */
2324 	if ((pv = pmap->pm_pmlpv) == NULL) {
2325 		pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2326 		pmap->pm_pmlpv = pv;
2327 		pmap_kenter((vm_offset_t)pmap->pm_pml4,
2328 			    VM_PAGE_TO_PHYS(pv->pv_m));
2329 		pv_put(pv);
2330 
2331 		/*
2332 		 * Install DMAP and KMAP.
2333 		 */
2334 		for (j = 0; j < NDMPML4E; ++j) {
2335 			pmap->pm_pml4[DMPML4I + j] =
2336 			    (DMPDPphys + ((vm_paddr_t)j << PAGE_SHIFT)) |
2337 			    pmap->pmap_bits[PG_RW_IDX] |
2338 			    pmap->pmap_bits[PG_V_IDX] |
2339 			    pmap->pmap_bits[PG_A_IDX];
2340 		}
2341 		for (j = 0; j < NKPML4E; ++j) {
2342 			pmap->pm_pml4[KPML4I + j] =
2343 			    (KPDPphys + ((vm_paddr_t)j << PAGE_SHIFT)) |
2344 			    pmap->pmap_bits[PG_RW_IDX] |
2345 			    pmap->pmap_bits[PG_V_IDX] |
2346 			    pmap->pmap_bits[PG_A_IDX];
2347 		}
2348 
2349 		/*
2350 		 * install self-referential address mapping entry
2351 		 */
2352 		pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
2353 		    pmap->pmap_bits[PG_V_IDX] |
2354 		    pmap->pmap_bits[PG_RW_IDX] |
2355 		    pmap->pmap_bits[PG_A_IDX];
2356 	} else {
2357 		KKASSERT(pv->pv_m->flags & PG_MAPPED);
2358 		KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
2359 	}
2360 	KKASSERT(pmap->pm_pml4[255] == 0);
2361 
2362 	/*
2363 	 * When implementing an isolated userland pmap, a second PML4e table
2364 	 * is needed.  We use pmap_pml4_pindex() + 1 for convenience, but
2365 	 * note that we do not operate on this table using our API functions
2366 	 * so handling of the + 1 case is mostly just to prevent implosions.
2367 	 *
2368 	 * We install an isolated version of the kernel PDPs into this
2369 	 * second PML4e table.  The pmap code will mirror all user PDPs
2370 	 * between the primary and secondary PML4e table.
2371 	 */
2372 	if ((pv = pmap->pm_pmlpv_iso) == NULL && meltdown_mitigation &&
2373 	    pmap != &iso_pmap) {
2374 		pv = pmap_allocpte(pmap, pmap_pml4_pindex() + 1, NULL);
2375 		pmap->pm_pmlpv_iso = pv;
2376 		pmap_kenter((vm_offset_t)pmap->pm_pml4_iso,
2377 			    VM_PAGE_TO_PHYS(pv->pv_m));
2378 		pv_put(pv);
2379 
2380 		/*
2381 		 * Install an isolated version of the kernel pmap for
2382 		 * user consumption, using PDPs constructed in iso_pmap.
2383 		 */
2384 		for (j = 0; j < NKPML4E; ++j) {
2385 			pmap->pm_pml4_iso[KPML4I + j] =
2386 				iso_pmap.pm_pml4[KPML4I + j];
2387 		}
2388 	} else if (pv) {
2389 		KKASSERT(pv->pv_m->flags & PG_MAPPED);
2390 		KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
2391 	}
2392 }
2393 
2394 /*
2395  * Clean up a pmap structure so it can be physically freed.  This routine
2396  * is called by the vmspace dtor function.  A great deal of pmap data is
2397  * left passively mapped to improve vmspace management so we have a bit
2398  * of cleanup work to do here.
2399  */
2400 void
2401 pmap_puninit(pmap_t pmap)
2402 {
2403 	pv_entry_t pv;
2404 	vm_page_t p;
2405 
2406 	KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
2407 	if ((pv = pmap->pm_pmlpv) != NULL) {
2408 		if (pv_hold_try(pv) == 0)
2409 			pv_lock(pv);
2410 		KKASSERT(pv == pmap->pm_pmlpv);
2411 		p = pmap_remove_pv_page(pv);
2412 		pv_free(pv, NULL);
2413 		pv = NULL;	/* safety */
2414 		pmap_kremove((vm_offset_t)pmap->pm_pml4);
2415 		vm_page_busy_wait(p, FALSE, "pgpun");
2416 		KKASSERT(p->flags & PG_UNQUEUED);
2417 		vm_page_unwire(p, 0);
2418 		vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2419 		vm_page_free(p);
2420 		pmap->pm_pmlpv = NULL;
2421 	}
2422 	if ((pv = pmap->pm_pmlpv_iso) != NULL) {
2423 		if (pv_hold_try(pv) == 0)
2424 			pv_lock(pv);
2425 		KKASSERT(pv == pmap->pm_pmlpv_iso);
2426 		p = pmap_remove_pv_page(pv);
2427 		pv_free(pv, NULL);
2428 		pv = NULL;	/* safety */
2429 		pmap_kremove((vm_offset_t)pmap->pm_pml4_iso);
2430 		vm_page_busy_wait(p, FALSE, "pgpun");
2431 		KKASSERT(p->flags & PG_UNQUEUED);
2432 		vm_page_unwire(p, 0);
2433 		vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
2434 		vm_page_free(p);
2435 		pmap->pm_pmlpv_iso = NULL;
2436 	}
2437 	if (pmap->pm_pml4) {
2438 		KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
2439 		kmem_free(&kernel_map,
2440 			  (vm_offset_t)pmap->pm_pml4, PAGE_SIZE * 2);
2441 		pmap->pm_pml4 = NULL;
2442 		pmap->pm_pml4_iso = NULL;
2443 	}
2444 	KKASSERT(pmap->pm_stats.resident_count == 0);
2445 	KKASSERT(pmap->pm_stats.wired_count == 0);
2446 }
2447 
2448 /*
2449  * This function is now unused (used to add the pmap to the pmap_list)
2450  */
2451 void
2452 pmap_pinit2(struct pmap *pmap)
2453 {
2454 }
2455 
2456 /*
2457  * This routine is called when various levels in the page table need to
2458  * be populated.  This routine cannot fail.
2459  *
2460  * This function returns two locked pv_entry's, one representing the
2461  * requested pv and one representing the requested pv's parent pv.  If
2462  * an intermediate page table does not exist it will be created, mapped,
2463  * wired, and the parent page table will be given an additional hold
2464  * count representing the presence of the child pv_entry.
2465  */
2466 static
2467 pv_entry_t
2468 pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
2469 {
2470 	pt_entry_t *ptep;
2471 	pt_entry_t *ptep_iso;
2472 	pv_entry_t pv;
2473 	pv_entry_t pvp;
2474 	pt_entry_t v;
2475 	vm_page_t m;
2476 	int isnew;
2477 	int ispt;
2478 
2479 	/*
2480 	 * If the pv already exists and we aren't being asked for the
2481 	 * parent page table page we can just return it.  A locked+held pv
2482 	 * is returned.  The pv will also have a second hold related to the
2483 	 * pmap association that we don't have to worry about.
2484 	 */
2485 	ispt = 0;
2486 	pv = pv_alloc(pmap, ptepindex, &isnew);
2487 	if (isnew == 0 && pvpp == NULL)
2488 		return(pv);
2489 
2490 	/*
2491 	 * DragonFly doesn't use PV's to represent terminal PTEs any more.
2492 	 * The index range is still used for placemarkers, but not for
2493 	 * actual pv_entry's.
2494 	 */
2495 	KKASSERT(ptepindex >= pmap_pt_pindex(0));
2496 
2497 	/*
2498 	 * Note that pt_pv's are only returned for user VAs. We assert that
2499 	 * a pt_pv is not being requested for kernel VAs.  The kernel
2500 	 * pre-wires all higher-level page tables so don't overload managed
2501 	 * higher-level page tables on top of it!
2502 	 *
2503 	 * However, its convenient for us to allow the case when creating
2504 	 * iso_pmap.  This is a bit of a hack but it simplifies iso_pmap
2505 	 * a lot.
2506 	 */
2507 
2508 	/*
2509 	 * The kernel never uses managed PT/PD/PDP pages.
2510 	 */
2511 	KKASSERT(pmap != &kernel_pmap);
2512 
2513 	/*
2514 	 * Non-terminal PVs allocate a VM page to represent the page table,
2515 	 * so we have to resolve pvp and calculate ptepindex for the pvp
2516 	 * and then for the page table entry index in the pvp for
2517 	 * fall-through.
2518 	 */
2519 	if (ptepindex < pmap_pd_pindex(0)) {
2520 		/*
2521 		 * pv is PT, pvp is PD
2522 		 */
2523 		ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
2524 		ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
2525 		pvp = pmap_allocpte(pmap, ptepindex, NULL);
2526 
2527 		/*
2528 		 * PT index in PD
2529 		 */
2530 		ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
2531 		ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
2532 		ispt = 1;
2533 	} else if (ptepindex < pmap_pdp_pindex(0)) {
2534 		/*
2535 		 * pv is PD, pvp is PDP
2536 		 *
2537 		 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2538 		 *		     the PD.
2539 		 */
2540 		ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
2541 		ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2542 
2543 		if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
2544 			KKASSERT(pvpp == NULL);
2545 			pvp = NULL;
2546 		} else {
2547 			pvp = pmap_allocpte(pmap, ptepindex, NULL);
2548 		}
2549 
2550 		/*
2551 		 * PD index in PDP
2552 		 */
2553 		ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
2554 		ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
2555 	} else if (ptepindex < pmap_pml4_pindex()) {
2556 		/*
2557 		 * pv is PDP, pvp is the root pml4 table
2558 		 */
2559 		pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
2560 
2561 		/*
2562 		 * PDP index in PML4
2563 		 */
2564 		ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
2565 		ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
2566 	} else {
2567 		/*
2568 		 * pv represents the top-level PML4, there is no parent.
2569 		 */
2570 		pvp = NULL;
2571 	}
2572 
2573 	if (isnew == 0)
2574 		goto notnew;
2575 
2576 	/*
2577 	 * (isnew) is TRUE, pv is not terminal.
2578 	 *
2579 	 * (1) Add a wire count to the parent page table (pvp).
2580 	 * (2) Allocate a VM page for the page table.
2581 	 * (3) Enter the VM page into the parent page table.
2582 	 *
2583 	 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2584 	 */
2585 	if (pvp)
2586 		vm_page_wire_quick(pvp->pv_m);
2587 
2588 	for (;;) {
2589 		m = vm_page_alloc(NULL, pv->pv_pindex,
2590 				  VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2591 				  VM_ALLOC_INTERRUPT);
2592 		if (m)
2593 			break;
2594 		vm_wait(0);
2595 	}
2596 	vm_page_wire(m);	/* wire for mapping in parent */
2597 	pmap_zero_page(VM_PAGE_TO_PHYS(m));
2598 	m->valid = VM_PAGE_BITS_ALL;
2599 	vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_UNQUEUED);
2600 	KKASSERT(m->queue == PQ_NONE);
2601 
2602 	pv->pv_m = m;
2603 
2604 	/*
2605 	 * (isnew) is TRUE, pv is not terminal.
2606 	 *
2607 	 * Wire the page into pvp.  Bump the resident_count for the pmap.
2608 	 * There is no pvp for the top level, address the pm_pml4[] array
2609 	 * directly.
2610 	 *
2611 	 * If the caller wants the parent we return it, otherwise
2612 	 * we just put it away.
2613 	 *
2614 	 * No interlock is needed for pte 0 -> non-zero.
2615 	 *
2616 	 * In the situation where *ptep is valid we might have an unmanaged
2617 	 * page table page shared from another page table which we need to
2618 	 * unshare before installing our private page table page.
2619 	 */
2620 	if (pvp) {
2621 		v = VM_PAGE_TO_PHYS(m) |
2622 		    (pmap->pmap_bits[PG_RW_IDX] |
2623 		     pmap->pmap_bits[PG_V_IDX] |
2624 		     pmap->pmap_bits[PG_A_IDX]);
2625 		if (ptepindex < NUPTE_USER)
2626 			v |= pmap->pmap_bits[PG_U_IDX];
2627 		if (ptepindex < pmap_pt_pindex(0))
2628 			v |= pmap->pmap_bits[PG_M_IDX];
2629 
2630 		ptep = pv_pte_lookup(pvp, ptepindex);
2631 		if (pvp == pmap->pm_pmlpv && pmap->pm_pmlpv_iso)
2632 			ptep_iso = pv_pte_lookup(pmap->pm_pmlpv_iso, ptepindex);
2633 		else
2634 			ptep_iso  = NULL;
2635 		if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
2636 			panic("pmap_allocpte: ptpte present without pv_entry!");
2637 		} else {
2638 			pt_entry_t pte;
2639 
2640 			pte = atomic_swap_long(ptep, v);
2641 			if (ptep_iso)
2642 				atomic_swap_long(ptep_iso, v);
2643 			if (pte != 0) {
2644 				kprintf("install pgtbl mixup 0x%016jx "
2645 					"old/new 0x%016jx/0x%016jx\n",
2646 					(intmax_t)ptepindex, pte, v);
2647 			}
2648 		}
2649 	}
2650 	vm_page_wakeup(m);
2651 
2652 	/*
2653 	 * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2654 	 */
2655 notnew:
2656 	if (pvp) {
2657 		KKASSERT(pvp->pv_m != NULL);
2658 		ptep = pv_pte_lookup(pvp, ptepindex);
2659 		v = VM_PAGE_TO_PHYS(pv->pv_m) |
2660 		    (pmap->pmap_bits[PG_RW_IDX] |
2661 		     pmap->pmap_bits[PG_V_IDX] |
2662 		     pmap->pmap_bits[PG_A_IDX]);
2663 		if (ptepindex < NUPTE_USER)
2664 			v |= pmap->pmap_bits[PG_U_IDX];
2665 		if (ptepindex < pmap_pt_pindex(0))
2666 			v |= pmap->pmap_bits[PG_M_IDX];
2667 		if (*ptep != v) {
2668 			kprintf("mismatched upper level pt %016jx/%016jx\n",
2669 				*ptep, v);
2670 		}
2671 	}
2672 	if (pvpp)
2673 		*pvpp = pvp;
2674 	else if (pvp)
2675 		pv_put(pvp);
2676 	return (pv);
2677 }
2678 
2679 /*
2680  * Release any resources held by the given physical map.
2681  *
2682  * Called when a pmap initialized by pmap_pinit is being released.  Should
2683  * only be called if the map contains no valid mappings.
2684  */
2685 struct pmap_release_info {
2686 	pmap_t	pmap;
2687 	int	retry;
2688 	pv_entry_t pvp;
2689 };
2690 
2691 static int pmap_release_callback(pv_entry_t pv, void *data);
2692 
2693 void
2694 pmap_release(struct pmap *pmap)
2695 {
2696 	struct pmap_release_info info;
2697 
2698 	KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
2699 		("pmap still active! %016jx",
2700 		(uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
2701 
2702 	/*
2703 	 * There is no longer a pmap_list, if there were we would remove the
2704 	 * pmap from it here.
2705 	 */
2706 
2707 	/*
2708 	 * Pull pv's off the RB tree in order from low to high and release
2709 	 * each page.
2710 	 */
2711 	info.pmap = pmap;
2712 	do {
2713 		info.retry = 0;
2714 		info.pvp = NULL;
2715 
2716 		spin_lock(&pmap->pm_spin);
2717 		RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
2718 			pmap_release_callback, &info);
2719 		spin_unlock(&pmap->pm_spin);
2720 
2721 		if (info.pvp)
2722 			pv_put(info.pvp);
2723 	} while (info.retry);
2724 
2725 
2726 	/*
2727 	 * One resident page (the pml4 page) should remain.  Two if
2728 	 * the pmap has implemented an isolated userland PML4E table.
2729 	 * No wired pages should remain.
2730 	 */
2731 	int expected_res = 0;
2732 
2733 	if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0)
2734 		++expected_res;
2735 	if (pmap->pm_pmlpv_iso)
2736 		++expected_res;
2737 
2738 #if 1
2739 	if (pmap->pm_stats.resident_count != expected_res ||
2740 	    pmap->pm_stats.wired_count != 0) {
2741 		kprintf("fatal pmap problem - pmap %p flags %08x "
2742 			"rescnt=%jd wirecnt=%jd\n",
2743 			pmap,
2744 			pmap->pm_flags,
2745 			pmap->pm_stats.resident_count,
2746 			pmap->pm_stats.wired_count);
2747 		tsleep(pmap, 0, "DEAD", 0);
2748 	}
2749 #else
2750 	KKASSERT(pmap->pm_stats.resident_count == expected_res);
2751 	KKASSERT(pmap->pm_stats.wired_count == 0);
2752 #endif
2753 }
2754 
2755 /*
2756  * Called from low to high.  We must cache the proper parent pv so we
2757  * can adjust its wired count.
2758  */
2759 static int
2760 pmap_release_callback(pv_entry_t pv, void *data)
2761 {
2762 	struct pmap_release_info *info = data;
2763 	pmap_t pmap = info->pmap;
2764 	vm_pindex_t pindex;
2765 	int r;
2766 
2767 	/*
2768 	 * Acquire a held and locked pv, check for release race
2769 	 */
2770 	pindex = pv->pv_pindex;
2771 	if (info->pvp == pv) {
2772 		spin_unlock(&pmap->pm_spin);
2773 		info->pvp = NULL;
2774 	} else if (pv_hold_try(pv)) {
2775 		spin_unlock(&pmap->pm_spin);
2776 	} else {
2777 		spin_unlock(&pmap->pm_spin);
2778 		pv_lock(pv);
2779 		pv_put(pv);
2780 		info->retry = 1;
2781 		spin_lock(&pmap->pm_spin);
2782 
2783 		return -1;
2784 	}
2785 	KKASSERT(pv->pv_pmap == pmap && pindex == pv->pv_pindex);
2786 
2787 	if (pv->pv_pindex < pmap_pt_pindex(0)) {
2788 		/*
2789 		 * I am PTE, parent is PT
2790 		 */
2791 		pindex = pv->pv_pindex >> NPTEPGSHIFT;
2792 		pindex += NUPTE_TOTAL;
2793 	} else if (pv->pv_pindex < pmap_pd_pindex(0)) {
2794 		/*
2795 		 * I am PT, parent is PD
2796 		 */
2797 		pindex = (pv->pv_pindex - NUPTE_TOTAL) >> NPDEPGSHIFT;
2798 		pindex += NUPTE_TOTAL + NUPT_TOTAL;
2799 	} else if (pv->pv_pindex < pmap_pdp_pindex(0)) {
2800 		/*
2801 		 * I am PD, parent is PDP
2802 		 */
2803 		pindex = (pv->pv_pindex - NUPTE_TOTAL - NUPT_TOTAL) >>
2804 			 NPDPEPGSHIFT;
2805 		pindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
2806 	} else if (pv->pv_pindex < pmap_pml4_pindex()) {
2807 		/*
2808 		 * I am PDP, parent is PML4.  We always calculate the
2809 		 * normal PML4 here, not the isolated PML4.
2810 		 */
2811 		pindex = pmap_pml4_pindex();
2812 	} else {
2813 		/*
2814 		 * parent is NULL
2815 		 */
2816 		if (info->pvp) {
2817 			pv_put(info->pvp);
2818 			info->pvp = NULL;
2819 		}
2820 		pindex = 0;
2821 	}
2822 	if (pindex) {
2823 		if (info->pvp && info->pvp->pv_pindex != pindex) {
2824 			pv_put(info->pvp);
2825 			info->pvp = NULL;
2826 		}
2827 		if (info->pvp == NULL)
2828 			info->pvp = pv_get(pmap, pindex, NULL);
2829 	} else {
2830 		if (info->pvp) {
2831 			pv_put(info->pvp);
2832 			info->pvp = NULL;
2833 		}
2834 	}
2835 	r = pmap_release_pv(pv, info->pvp, NULL);
2836 	spin_lock(&pmap->pm_spin);
2837 
2838 	return(r);
2839 }
2840 
2841 /*
2842  * Called with held (i.e. also locked) pv.  This function will dispose of
2843  * the lock along with the pv.
2844  *
2845  * If the caller already holds the locked parent page table for pv it
2846  * must pass it as pvp, allowing us to avoid a deadlock, else it can
2847  * pass NULL for pvp.
2848  */
2849 static int
2850 pmap_release_pv(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk)
2851 {
2852 	vm_page_t p;
2853 
2854 	/*
2855 	 * The pmap is currently not spinlocked, pv is held+locked.
2856 	 * Remove the pv's page from its parent's page table.  The
2857 	 * parent's page table page's wire_count will be decremented.
2858 	 *
2859 	 * This will clean out the pte at any level of the page table.
2860 	 * If smp != 0 all cpus are affected.
2861 	 *
2862 	 * Do not tear-down recursively, its faster to just let the
2863 	 * release run its course.
2864 	 */
2865 	pmap_remove_pv_pte(pv, pvp, bulk, 0);
2866 
2867 	/*
2868 	 * Terminal pvs are unhooked from their vm_pages.  Because
2869 	 * terminal pages aren't page table pages they aren't wired
2870 	 * by us, so we have to be sure not to unwire them either.
2871 	 */
2872 	if (pv->pv_pindex < pmap_pt_pindex(0)) {
2873 		pmap_remove_pv_page(pv);
2874 		goto skip;
2875 	}
2876 
2877 	/*
2878 	 * We leave the top-level page table page cached, wired, and
2879 	 * mapped in the pmap until the dtor function (pmap_puninit())
2880 	 * gets called.
2881 	 *
2882 	 * Since we are leaving the top-level pv intact we need
2883 	 * to break out of what would otherwise be an infinite loop.
2884 	 *
2885 	 * This covers both the normal and the isolated PML4 page.
2886 	 */
2887 	if (pv->pv_pindex >= pmap_pml4_pindex()) {
2888 		pv_put(pv);
2889 		return(-1);
2890 	}
2891 
2892 	/*
2893 	 * For page table pages (other than the top-level page),
2894 	 * remove and free the vm_page.  The representitive mapping
2895 	 * removed above by pmap_remove_pv_pte() did not undo the
2896 	 * last wire_count so we have to do that as well.
2897 	 */
2898 	p = pmap_remove_pv_page(pv);
2899 	vm_page_busy_wait(p, FALSE, "pmaprl");
2900 	if (p->wire_count != 1) {
2901 		const char *tstr;
2902 
2903 		if (pv->pv_pindex >= pmap_pdp_pindex(0))
2904 			tstr = "PDP";
2905 		else if (pv->pv_pindex >= pmap_pd_pindex(0))
2906 			tstr = "PD";
2907 		else if (pv->pv_pindex >= pmap_pt_pindex(0))
2908 			tstr = "PT";
2909 		else
2910 			tstr = "PTE";
2911 
2912 		kprintf("p(%s) p->wire_count was %016lx %d\n",
2913 			tstr, pv->pv_pindex, p->wire_count);
2914 	}
2915 	KKASSERT(p->wire_count == 1);
2916 	KKASSERT(p->flags & PG_UNQUEUED);
2917 
2918 	vm_page_unwire(p, 0);
2919 	KKASSERT(p->wire_count == 0);
2920 
2921 	vm_page_free(p);
2922 skip:
2923 	pv_free(pv, pvp);
2924 
2925 	return 0;
2926 }
2927 
2928 /*
2929  * This function will remove the pte associated with a pv from its parent.
2930  * Terminal pv's are supported.  All cpus specified by (bulk) are properly
2931  * invalidated.
2932  *
2933  * The wire count will be dropped on the parent page table.  The wire
2934  * count on the page being removed (pv->pv_m) from the parent page table
2935  * is NOT touched.  Note that terminal pages will not have any additional
2936  * wire counts while page table pages will have at least one representing
2937  * the mapping, plus others representing sub-mappings.
2938  *
2939  * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2940  *	 pages and user page table and terminal pages.
2941  *
2942  * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2943  *	 be freshly allocated and not imply that the pte is managed.  In this
2944  *	 case pv->pv_m should be NULL.
2945  *
2946  * The pv must be locked.  The pvp, if supplied, must be locked.  All
2947  * supplied pv's will remain locked on return.
2948  *
2949  * XXX must lock parent pv's if they exist to remove pte XXX
2950  */
2951 static
2952 void
2953 pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, pmap_inval_bulk_t *bulk,
2954 		   int destroy)
2955 {
2956 	vm_pindex_t ptepindex = pv->pv_pindex;
2957 	pmap_t pmap = pv->pv_pmap;
2958 	vm_page_t p;
2959 	int gotpvp = 0;
2960 
2961 	KKASSERT(pmap);
2962 
2963 	if (ptepindex >= pmap_pml4_pindex()) {
2964 		/*
2965 		 * We are the top level PML4E table, there is no parent.
2966 		 *
2967 		 * This is either the normal or isolated PML4E table.
2968 		 * Only the normal is used in regular operation, the isolated
2969 		 * is only passed in when breaking down the whole pmap.
2970 		 */
2971 		p = pmap->pm_pmlpv->pv_m;
2972 		KKASSERT(pv->pv_m == p);	/* debugging */
2973 	} else if (ptepindex >= pmap_pdp_pindex(0)) {
2974 		/*
2975 		 * Remove a PDP page from the PML4E.  This can only occur
2976 		 * with user page tables.  We do not have to lock the
2977 		 * pml4 PV so just ignore pvp.
2978 		 */
2979 		vm_pindex_t pml4_pindex;
2980 		vm_pindex_t pdp_index;
2981 		pml4_entry_t *pdp;
2982 		pml4_entry_t *pdp_iso;
2983 
2984 		pdp_index = ptepindex - pmap_pdp_pindex(0);
2985 		if (pvp == NULL) {
2986 			pml4_pindex = pmap_pml4_pindex();
2987 			pvp = pv_get(pv->pv_pmap, pml4_pindex, NULL);
2988 			KKASSERT(pvp);
2989 			gotpvp = 1;
2990 		}
2991 
2992 		pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2993 		KKASSERT((*pdp & pmap->pmap_bits[PG_V_IDX]) != 0);
2994 		p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2995 		pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp, 0);
2996 
2997 		/*
2998 		 * Also remove the PDP from the isolated PML4E if the
2999 		 * process uses one.
3000 		 */
3001 		if (pvp == pmap->pm_pmlpv && pmap->pm_pmlpv_iso) {
3002 			pdp_iso = &pmap->pm_pml4_iso[pdp_index &
3003 						((1ul << NPML4EPGSHIFT) - 1)];
3004 			pmap_inval_bulk(bulk, (vm_offset_t)-1, pdp_iso, 0);
3005 		}
3006 		KKASSERT(pv->pv_m == p);	/* debugging */
3007 	} else if (ptepindex >= pmap_pd_pindex(0)) {
3008 		/*
3009 		 * Remove a PD page from the PDP
3010 		 *
3011 		 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
3012 		 *		     of a simple pmap because it stops at
3013 		 *		     the PD page.
3014 		 */
3015 		vm_pindex_t pdp_pindex;
3016 		vm_pindex_t pd_index;
3017 		pdp_entry_t *pd;
3018 
3019 		pd_index = ptepindex - pmap_pd_pindex(0);
3020 
3021 		if (pvp == NULL) {
3022 			pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
3023 				     (pd_index >> NPML4EPGSHIFT);
3024 			pvp = pv_get(pv->pv_pmap, pdp_pindex, NULL);
3025 			gotpvp = 1;
3026 		}
3027 
3028 		if (pvp) {
3029 			pd = pv_pte_lookup(pvp, pd_index &
3030 						((1ul << NPDPEPGSHIFT) - 1));
3031 			KKASSERT((*pd & pmap->pmap_bits[PG_V_IDX]) != 0);
3032 			p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
3033 			pmap_inval_bulk(bulk, (vm_offset_t)-1, pd, 0);
3034 		} else {
3035 			KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
3036 			p = pv->pv_m;		/* degenerate test later */
3037 		}
3038 		KKASSERT(pv->pv_m == p);	/* debugging */
3039 	} else if (ptepindex >= pmap_pt_pindex(0)) {
3040 		/*
3041 		 *  Remove a PT page from the PD
3042 		 */
3043 		vm_pindex_t pd_pindex;
3044 		vm_pindex_t pt_index;
3045 		pd_entry_t *pt;
3046 
3047 		pt_index = ptepindex - pmap_pt_pindex(0);
3048 
3049 		if (pvp == NULL) {
3050 			pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
3051 				    (pt_index >> NPDPEPGSHIFT);
3052 			pvp = pv_get(pv->pv_pmap, pd_pindex, NULL);
3053 			KKASSERT(pvp);
3054 			gotpvp = 1;
3055 		}
3056 
3057 		pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
3058 #if 0
3059 		KASSERT((*pt & pmap->pmap_bits[PG_V_IDX]) != 0,
3060 			("*pt unexpectedly invalid %016jx "
3061 			 "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
3062 			*pt, gotpvp, ptepindex, pt_index, pv, pvp));
3063 		p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
3064 #else
3065 		if ((*pt & pmap->pmap_bits[PG_V_IDX]) == 0) {
3066 			kprintf("*pt unexpectedly invalid %016jx "
3067 			        "gotpvp=%d ptepindex=%ld ptindex=%ld "
3068 				"pv=%p pvp=%p\n",
3069 				*pt, gotpvp, ptepindex, pt_index, pv, pvp);
3070 			tsleep(pt, 0, "DEAD", 0);
3071 			p = pv->pv_m;
3072 		} else {
3073 			p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
3074 		}
3075 #endif
3076 		pmap_inval_bulk(bulk, (vm_offset_t)-1, pt, 0);
3077 		KKASSERT(pv->pv_m == p);	/* debugging */
3078 	} else {
3079 		KKASSERT(0);
3080 	}
3081 
3082 	/*
3083 	 * If requested, scrap the underlying pv->pv_m and the underlying
3084 	 * pv.  If this is a page-table-page we must also free the page.
3085 	 *
3086 	 * pvp must be returned locked.
3087 	 */
3088 	if (destroy == 1) {
3089 		/*
3090 		 * page table page (PT, PD, PDP, PML4), caller was responsible
3091 		 * for testing wired_count.
3092 		 */
3093 		KKASSERT(pv->pv_m->wire_count == 1);
3094 		p = pmap_remove_pv_page(pv);
3095 		pv_free(pv, pvp);
3096 		pv = NULL;
3097 
3098 		vm_page_busy_wait(p, FALSE, "pgpun");
3099 		vm_page_unwire(p, 0);
3100 		vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
3101 		vm_page_free(p);
3102 	} else if (destroy == 2) {
3103 		/*
3104 		 * Normal page, remove from pmap and leave the underlying
3105 		 * page untouched.
3106 		 */
3107 		pmap_remove_pv_page(pv);
3108 		pv_free(pv, pvp);
3109 		pv = NULL;		/* safety */
3110 	}
3111 
3112 	/*
3113 	 * If we acquired pvp ourselves then we are responsible for
3114 	 * recursively deleting it.
3115 	 */
3116 	if (pvp && gotpvp) {
3117 		/*
3118 		 * Recursively destroy higher-level page tables.
3119 		 *
3120 		 * This is optional.  If we do not, they will still
3121 		 * be destroyed when the process exits.
3122 		 *
3123 		 * NOTE: Do not destroy pv_entry's with extra hold refs,
3124 		 *	 a caller may have unlocked it and intends to
3125 		 *	 continue to use it.
3126 		 */
3127 		if (pmap_dynamic_delete &&
3128 		    pvp->pv_m &&
3129 		    pvp->pv_m->wire_count == 1 &&
3130 		    (pvp->pv_hold & PV_HOLD_MASK) == 2 &&
3131 		    pvp->pv_pindex < pmap_pml4_pindex()) {
3132 			if (pmap != &kernel_pmap) {
3133 				pmap_remove_pv_pte(pvp, NULL, bulk, 1);
3134 				pvp = NULL;	/* safety */
3135 			} else {
3136 				kprintf("Attempt to remove kernel_pmap pindex "
3137 					"%jd\n", pvp->pv_pindex);
3138 				pv_put(pvp);
3139 			}
3140 		} else {
3141 			pv_put(pvp);
3142 		}
3143 	}
3144 }
3145 
3146 /*
3147  * Remove the vm_page association to a pv.  The pv must be locked.
3148  */
3149 static
3150 vm_page_t
3151 pmap_remove_pv_page(pv_entry_t pv)
3152 {
3153 	vm_page_t m;
3154 
3155 	m = pv->pv_m;
3156 	pv->pv_m = NULL;
3157 	vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3158 
3159 	return(m);
3160 }
3161 
3162 /*
3163  * Grow the number of kernel page table entries, if needed.
3164  *
3165  * This routine is always called to validate any address space
3166  * beyond KERNBASE (for kldloads).  kernel_vm_end only governs the address
3167  * space below KERNBASE.
3168  *
3169  * kernel_map must be locked exclusively by the caller.
3170  */
3171 void
3172 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
3173 {
3174 	vm_paddr_t paddr;
3175 	vm_offset_t ptppaddr;
3176 	vm_page_t nkpg;
3177 	pd_entry_t *pt, newpt;
3178 	pdp_entry_t *pd, newpd;
3179 	int update_kernel_vm_end;
3180 
3181 	/*
3182 	 * bootstrap kernel_vm_end on first real VM use
3183 	 */
3184 	if (kernel_vm_end == 0) {
3185 		kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
3186 
3187 		for (;;) {
3188 			pt = pmap_pt(&kernel_pmap, kernel_vm_end);
3189 			if (pt == NULL)
3190 				break;
3191 			if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) == 0)
3192 				break;
3193 			kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
3194 					~(vm_offset_t)(PAGE_SIZE * NPTEPG - 1);
3195 			if (kernel_vm_end - 1 >= vm_map_max(&kernel_map)) {
3196 				kernel_vm_end = vm_map_max(&kernel_map);
3197 				break;
3198 			}
3199 		}
3200 	}
3201 
3202 	/*
3203 	 * Fill in the gaps.  kernel_vm_end is only adjusted for ranges
3204 	 * below KERNBASE.  Ranges above KERNBASE are kldloaded and we
3205 	 * do not want to force-fill 128G worth of page tables.
3206 	 */
3207 	if (kstart < KERNBASE) {
3208 		if (kstart > kernel_vm_end)
3209 			kstart = kernel_vm_end;
3210 		KKASSERT(kend <= KERNBASE);
3211 		update_kernel_vm_end = 1;
3212 	} else {
3213 		update_kernel_vm_end = 0;
3214 	}
3215 
3216 	kstart = rounddown2(kstart, (vm_offset_t)(PAGE_SIZE * NPTEPG));
3217 	kend = roundup2(kend, (vm_offset_t)(PAGE_SIZE * NPTEPG));
3218 
3219 	if (kend - 1 >= vm_map_max(&kernel_map))
3220 		kend = vm_map_max(&kernel_map);
3221 
3222 	while (kstart < kend) {
3223 		pt = pmap_pt(&kernel_pmap, kstart);
3224 		if (pt == NULL) {
3225 			/*
3226 			 * We need a new PD entry
3227 			 */
3228 			nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3229 			                     VM_ALLOC_NORMAL |
3230 					     VM_ALLOC_SYSTEM |
3231 					     VM_ALLOC_INTERRUPT);
3232 			if (nkpg == NULL) {
3233 				panic("pmap_growkernel: no memory to grow "
3234 				      "kernel");
3235 			}
3236 			paddr = VM_PAGE_TO_PHYS(nkpg);
3237 			pmap_zero_page(paddr);
3238 			pd = pmap_pd(&kernel_pmap, kstart);
3239 
3240 			newpd = (pdp_entry_t)
3241 			    (paddr |
3242 			    kernel_pmap.pmap_bits[PG_V_IDX] |
3243 			    kernel_pmap.pmap_bits[PG_RW_IDX] |
3244 			    kernel_pmap.pmap_bits[PG_A_IDX]);
3245 			atomic_swap_long(pd, newpd);
3246 
3247 #if 0
3248 			kprintf("NEWPD pd=%p pde=%016jx phys=%016jx\n",
3249 				pd, newpd, paddr);
3250 #endif
3251 
3252 			continue; /* try again */
3253 		}
3254 
3255 		if ((*pt & kernel_pmap.pmap_bits[PG_V_IDX]) != 0) {
3256 			kstart = (kstart + PAGE_SIZE * NPTEPG) &
3257 				 ~(vm_offset_t)(PAGE_SIZE * NPTEPG - 1);
3258 			if (kstart - 1 >= vm_map_max(&kernel_map)) {
3259 				kstart = vm_map_max(&kernel_map);
3260 				break;
3261 			}
3262 			continue;
3263 		}
3264 
3265 		/*
3266 		 * We need a new PT
3267 		 *
3268 		 * This index is bogus, but out of the way
3269 		 */
3270 		nkpg = vm_page_alloc(NULL, mycpu->gd_rand_incr++,
3271 				     VM_ALLOC_NORMAL |
3272 				     VM_ALLOC_SYSTEM |
3273 				     VM_ALLOC_INTERRUPT);
3274 		if (nkpg == NULL)
3275 			panic("pmap_growkernel: no memory to grow kernel");
3276 
3277 		vm_page_wire(nkpg);
3278 		ptppaddr = VM_PAGE_TO_PHYS(nkpg);
3279 		pmap_zero_page(ptppaddr);
3280 		newpt = (pd_entry_t)(ptppaddr |
3281 				     kernel_pmap.pmap_bits[PG_V_IDX] |
3282 				     kernel_pmap.pmap_bits[PG_RW_IDX] |
3283 				     kernel_pmap.pmap_bits[PG_A_IDX]);
3284 		atomic_swap_long(pt, newpt);
3285 
3286 		kstart = (kstart + PAGE_SIZE * NPTEPG) &
3287 			  ~(vm_offset_t)(PAGE_SIZE * NPTEPG - 1);
3288 
3289 		if (kstart - 1 >= vm_map_max(&kernel_map)) {
3290 			kstart = vm_map_max(&kernel_map);
3291 			break;
3292 		}
3293 	}
3294 
3295 	/*
3296 	 * Only update kernel_vm_end for areas below KERNBASE.
3297 	 */
3298 	if (update_kernel_vm_end && kernel_vm_end < kstart)
3299 		kernel_vm_end = kstart;
3300 }
3301 
3302 /*
3303  *	Add a reference to the specified pmap.
3304  */
3305 void
3306 pmap_reference(pmap_t pmap)
3307 {
3308 	if (pmap != NULL)
3309 		atomic_add_int(&pmap->pm_count, 1);
3310 }
3311 
3312 void
3313 pmap_maybethreaded(pmap_t pmap)
3314 {
3315 	atomic_set_int(&pmap->pm_flags, PMAP_MULTI);
3316 }
3317 
3318 /*
3319  * Called while page is hard-busied to clear the PG_MAPPED and PG_WRITEABLE
3320  * flags if able.
3321  */
3322 int
3323 pmap_mapped_sync(vm_page_t m)
3324 {
3325 	if (m->md.pmap_count == 0)
3326 		vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3327 	return (m->flags);
3328 }
3329 
3330 /***************************************************
3331  * page management routines.
3332  ***************************************************/
3333 
3334 /*
3335  * Hold a pv without locking it
3336  */
3337 #if 0
3338 static void
3339 pv_hold(pv_entry_t pv)
3340 {
3341 	atomic_add_int(&pv->pv_hold, 1);
3342 }
3343 #endif
3344 
3345 /*
3346  * Hold a pv_entry, preventing its destruction.  TRUE is returned if the pv
3347  * was successfully locked, FALSE if it wasn't.  The caller must dispose of
3348  * the pv properly.
3349  *
3350  * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3351  * pv list via its page) must be held by the caller in order to stabilize
3352  * the pv.
3353  */
3354 static int
3355 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
3356 {
3357 	u_int count;
3358 
3359 	/*
3360 	 * Critical path shortcut expects pv to already have one ref
3361 	 * (for the pv->pv_pmap).
3362 	 */
3363 	count = pv->pv_hold;
3364 	cpu_ccfence();
3365 	for (;;) {
3366 		if ((count & PV_HOLD_LOCKED) == 0) {
3367 			if (atomic_fcmpset_int(&pv->pv_hold, &count,
3368 					      (count + 1) | PV_HOLD_LOCKED)) {
3369 #ifdef PMAP_DEBUG
3370 				pv->pv_func = func;
3371 				pv->pv_line = lineno;
3372 #endif
3373 				return TRUE;
3374 			}
3375 		} else {
3376 			if (atomic_fcmpset_int(&pv->pv_hold, &count, count + 1))
3377 				return FALSE;
3378 		}
3379 		/* retry */
3380 	}
3381 }
3382 
3383 /*
3384  * Drop a previously held pv_entry which could not be locked, allowing its
3385  * destruction.
3386  *
3387  * Must not be called with a spinlock held as we might zfree() the pv if it
3388  * is no longer associated with a pmap and this was the last hold count.
3389  */
3390 static void
3391 pv_drop(pv_entry_t pv)
3392 {
3393 	u_int count;
3394 
3395 	for (;;) {
3396 		count = pv->pv_hold;
3397 		cpu_ccfence();
3398 		KKASSERT((count & PV_HOLD_MASK) > 0);
3399 		KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
3400 			 (PV_HOLD_LOCKED | 1));
3401 		if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
3402 			if ((count & PV_HOLD_MASK) == 1) {
3403 #ifdef PMAP_DEBUG2
3404 				if (pmap_enter_debug > 0) {
3405 					--pmap_enter_debug;
3406 					kprintf("pv_drop: free pv %p\n", pv);
3407 				}
3408 #endif
3409 				KKASSERT(count == 1);
3410 				KKASSERT(pv->pv_pmap == NULL);
3411 				zfree(pvzone, pv);
3412 			}
3413 			return;
3414 		}
3415 		/* retry */
3416 	}
3417 }
3418 
3419 /*
3420  * Find or allocate the requested PV entry, returning a locked, held pv.
3421  *
3422  * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3423  * for the caller and one representing the pmap and vm_page association.
3424  *
3425  * If (*isnew) is zero, the returned pv will have only one hold count.
3426  *
3427  * Since both associations can only be adjusted while the pv is locked,
3428  * together they represent just one additional hold.
3429  */
3430 static
3431 pv_entry_t
3432 _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
3433 {
3434 	struct mdglobaldata *md = mdcpu;
3435 	pv_entry_t pv;
3436 	pv_entry_t pnew;
3437 	int pmap_excl = 0;
3438 
3439 	pnew = NULL;
3440 	if (md->gd_newpv) {
3441 #if 1
3442 		pnew = atomic_swap_ptr((void *)&md->gd_newpv, NULL);
3443 #else
3444 		crit_enter();
3445 		pnew = md->gd_newpv;	/* might race NULL */
3446 		md->gd_newpv = NULL;
3447 		crit_exit();
3448 #endif
3449 	}
3450 	if (pnew == NULL)
3451 		pnew = zalloc(pvzone);
3452 
3453 	spin_lock_shared(&pmap->pm_spin);
3454 	for (;;) {
3455 		/*
3456 		 * Shortcut cache
3457 		 */
3458 		pv = pv_entry_lookup(pmap, pindex);
3459 		if (pv == NULL) {
3460 			vm_pindex_t *pmark;
3461 
3462 			/*
3463 			 * Requires exclusive pmap spinlock
3464 			 */
3465 			if (pmap_excl == 0) {
3466 				pmap_excl = 1;
3467 				if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3468 					spin_unlock_shared(&pmap->pm_spin);
3469 					spin_lock(&pmap->pm_spin);
3470 					continue;
3471 				}
3472 			}
3473 
3474 			/*
3475 			 * We need to block if someone is holding our
3476 			 * placemarker.  As long as we determine the
3477 			 * placemarker has not been aquired we do not
3478 			 * need to get it as acquision also requires
3479 			 * the pmap spin lock.
3480 			 *
3481 			 * However, we can race the wakeup.
3482 			 */
3483 			pmark = pmap_placemarker_hash(pmap, pindex);
3484 
3485 			if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3486 				tsleep_interlock(pmark, 0);
3487 				atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3488 				if (((*pmark ^ pindex) &
3489 				     ~PM_PLACEMARK_WAKEUP) == 0) {
3490 					spin_unlock(&pmap->pm_spin);
3491 					tsleep(pmark, PINTERLOCKED, "pvplc", 0);
3492 					spin_lock(&pmap->pm_spin);
3493 				}
3494 				continue;
3495 			}
3496 
3497 			/*
3498 			 * Setup the new entry
3499 			 */
3500 			pnew->pv_pmap = pmap;
3501 			pnew->pv_pindex = pindex;
3502 			pnew->pv_hold = PV_HOLD_LOCKED | 2;
3503 			pnew->pv_flags = 0;
3504 #ifdef PMAP_DEBUG
3505 			pnew->pv_func = func;
3506 			pnew->pv_line = lineno;
3507 			if (pnew->pv_line_lastfree > 0) {
3508 				pnew->pv_line_lastfree =
3509 						-pnew->pv_line_lastfree;
3510 			}
3511 #endif
3512 			pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
3513 			atomic_add_long(&pmap->pm_stats.resident_count, 1);
3514 			spin_unlock(&pmap->pm_spin);
3515 			*isnew = 1;
3516 
3517 			KASSERT(pv == NULL, ("pv insert failed %p->%p", pnew, pv));
3518 			return(pnew);
3519 		}
3520 
3521 		/*
3522 		 * We already have an entry, cleanup the staged pnew if
3523 		 * we can get the lock, otherwise block and retry.
3524 		 */
3525 		if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY))) {
3526 			if (pmap_excl)
3527 				spin_unlock(&pmap->pm_spin);
3528 			else
3529 				spin_unlock_shared(&pmap->pm_spin);
3530 #if 1
3531 			pnew = atomic_swap_ptr((void *)&md->gd_newpv, pnew);
3532 			if (pnew)
3533 				zfree(pvzone, pnew);
3534 #else
3535 			crit_enter();
3536 			if (md->gd_newpv == NULL)
3537 				md->gd_newpv = pnew;
3538 			else
3539 				zfree(pvzone, pnew);
3540 			crit_exit();
3541 #endif
3542 			KKASSERT(pv->pv_pmap == pmap &&
3543 				 pv->pv_pindex == pindex);
3544 			*isnew = 0;
3545 			return(pv);
3546 		}
3547 		if (pmap_excl) {
3548 			spin_unlock(&pmap->pm_spin);
3549 			_pv_lock(pv PMAP_DEBUG_COPY);
3550 			pv_put(pv);
3551 			spin_lock(&pmap->pm_spin);
3552 		} else {
3553 			spin_unlock_shared(&pmap->pm_spin);
3554 			_pv_lock(pv PMAP_DEBUG_COPY);
3555 			pv_put(pv);
3556 			spin_lock_shared(&pmap->pm_spin);
3557 		}
3558 	}
3559 	/* NOT REACHED */
3560 }
3561 
3562 /*
3563  * Find the requested PV entry, returning a locked+held pv or NULL
3564  */
3565 static
3566 pv_entry_t
3567 _pv_get(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp PMAP_DEBUG_DECL)
3568 {
3569 	pv_entry_t pv;
3570 	int pmap_excl = 0;
3571 
3572 	spin_lock_shared(&pmap->pm_spin);
3573 	for (;;) {
3574 		/*
3575 		 * Shortcut cache
3576 		 */
3577 		pv = pv_entry_lookup(pmap, pindex);
3578 		if (pv == NULL) {
3579 			/*
3580 			 * Block if there is ANY placemarker.  If we are to
3581 			 * return it, we must also aquire the spot, so we
3582 			 * have to block even if the placemarker is held on
3583 			 * a different address.
3584 			 *
3585 			 * OPTIMIZATION: If pmarkp is passed as NULL the
3586 			 * caller is just probing (or looking for a real
3587 			 * pv_entry), and in this case we only need to check
3588 			 * to see if the placemarker matches pindex.
3589 			 */
3590 			vm_pindex_t *pmark;
3591 
3592 			/*
3593 			 * Requires exclusive pmap spinlock
3594 			 */
3595 			if (pmap_excl == 0) {
3596 				pmap_excl = 1;
3597 				if (!spin_lock_upgrade_try(&pmap->pm_spin)) {
3598 					spin_unlock_shared(&pmap->pm_spin);
3599 					spin_lock(&pmap->pm_spin);
3600 					continue;
3601 				}
3602 			}
3603 
3604 			pmark = pmap_placemarker_hash(pmap, pindex);
3605 
3606 			if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3607 			    ((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3608 				tsleep_interlock(pmark, 0);
3609 				atomic_set_long(pmark, PM_PLACEMARK_WAKEUP);
3610 				if ((pmarkp && *pmark != PM_NOPLACEMARK) ||
3611 				    ((*pmark ^ pindex) &
3612 				     ~PM_PLACEMARK_WAKEUP) == 0) {
3613 					spin_unlock(&pmap->pm_spin);
3614 					tsleep(pmark, PINTERLOCKED, "pvpld", 0);
3615 					spin_lock(&pmap->pm_spin);
3616 				}
3617 				continue;
3618 			}
3619 			if (pmarkp) {
3620 				if (atomic_swap_long(pmark, pindex) !=
3621 				    PM_NOPLACEMARK) {
3622 					panic("_pv_get: pmark race");
3623 				}
3624 				*pmarkp = pmark;
3625 			}
3626 			spin_unlock(&pmap->pm_spin);
3627 			return NULL;
3628 		}
3629 		if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
3630 			if (pmap_excl)
3631 				spin_unlock(&pmap->pm_spin);
3632 			else
3633 				spin_unlock_shared(&pmap->pm_spin);
3634 			KKASSERT(pv->pv_pmap == pmap &&
3635 				 pv->pv_pindex == pindex);
3636 			return(pv);
3637 		}
3638 		if (pmap_excl) {
3639 			spin_unlock(&pmap->pm_spin);
3640 			_pv_lock(pv PMAP_DEBUG_COPY);
3641 			pv_put(pv);
3642 			spin_lock(&pmap->pm_spin);
3643 		} else {
3644 			spin_unlock_shared(&pmap->pm_spin);
3645 			_pv_lock(pv PMAP_DEBUG_COPY);
3646 			pv_put(pv);
3647 			spin_lock_shared(&pmap->pm_spin);
3648 		}
3649 	}
3650 }
3651 
3652 /*
3653  * Lookup, hold, and attempt to lock (pmap,pindex).
3654  *
3655  * If the entry does not exist NULL is returned and *errorp is set to 0
3656  *
3657  * If the entry exists and could be successfully locked it is returned and
3658  * errorp is set to 0.
3659  *
3660  * If the entry exists but could NOT be successfully locked it is returned
3661  * held and *errorp is set to 1.
3662  *
3663  * If the entry is placemarked by someone else NULL is returned and *errorp
3664  * is set to 1.
3665  */
3666 static
3667 pv_entry_t
3668 pv_get_try(pmap_t pmap, vm_pindex_t pindex, vm_pindex_t **pmarkp, int *errorp)
3669 {
3670 	pv_entry_t pv;
3671 
3672 	spin_lock_shared(&pmap->pm_spin);
3673 
3674 	pv = pv_entry_lookup(pmap, pindex);
3675 	if (pv == NULL) {
3676 		vm_pindex_t *pmark;
3677 
3678 		pmark = pmap_placemarker_hash(pmap, pindex);
3679 
3680 		if (((*pmark ^ pindex) & ~PM_PLACEMARK_WAKEUP) == 0) {
3681 			*errorp = 1;
3682 		} else if (pmarkp &&
3683 			   atomic_cmpset_long(pmark, PM_NOPLACEMARK, pindex)) {
3684 			*errorp = 0;
3685 		} else {
3686 			/*
3687 			 * Can't set a placemark with a NULL pmarkp, or if
3688 			 * pmarkp is non-NULL but we failed to set our
3689 			 * placemark.
3690 			 */
3691 			*errorp = 1;
3692 		}
3693 		if (pmarkp)
3694 			*pmarkp = pmark;
3695 		spin_unlock_shared(&pmap->pm_spin);
3696 
3697 		return NULL;
3698 	}
3699 
3700 	/*
3701 	 * XXX This has problems if the lock is shared, why?
3702 	 */
3703 	if (pv_hold_try(pv)) {
3704 		spin_unlock_shared(&pmap->pm_spin);
3705 		*errorp = 0;
3706 		KKASSERT(pv->pv_pmap == pmap && pv->pv_pindex == pindex);
3707 		return(pv);	/* lock succeeded */
3708 	}
3709 	spin_unlock_shared(&pmap->pm_spin);
3710 	*errorp = 1;
3711 
3712 	return (pv);		/* lock failed */
3713 }
3714 
3715 /*
3716  * Lock a held pv, keeping the hold count
3717  */
3718 static
3719 void
3720 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
3721 {
3722 	u_int count;
3723 
3724 	for (;;) {
3725 		count = pv->pv_hold;
3726 		cpu_ccfence();
3727 		if ((count & PV_HOLD_LOCKED) == 0) {
3728 			if (atomic_cmpset_int(&pv->pv_hold, count,
3729 					      count | PV_HOLD_LOCKED)) {
3730 #ifdef PMAP_DEBUG
3731 				pv->pv_func = func;
3732 				pv->pv_line = lineno;
3733 #endif
3734 				return;
3735 			}
3736 			continue;
3737 		}
3738 		tsleep_interlock(pv, 0);
3739 		if (atomic_cmpset_int(&pv->pv_hold, count,
3740 				      count | PV_HOLD_WAITING)) {
3741 #ifdef PMAP_DEBUG2
3742 			if (pmap_enter_debug > 0) {
3743 				--pmap_enter_debug;
3744 				kprintf("pv waiting on %s:%d\n",
3745 					pv->pv_func, pv->pv_line);
3746 			}
3747 #endif
3748 			tsleep(pv, PINTERLOCKED, "pvwait", hz);
3749 		}
3750 		/* retry */
3751 	}
3752 }
3753 
3754 /*
3755  * Unlock a held and locked pv, keeping the hold count.
3756  */
3757 static
3758 void
3759 pv_unlock(pv_entry_t pv)
3760 {
3761 	u_int count;
3762 
3763 	for (;;) {
3764 		count = pv->pv_hold;
3765 		cpu_ccfence();
3766 		KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) >=
3767 			 (PV_HOLD_LOCKED | 1));
3768 		if (atomic_cmpset_int(&pv->pv_hold, count,
3769 				      count &
3770 				      ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
3771 			if (count & PV_HOLD_WAITING)
3772 				wakeup(pv);
3773 			break;
3774 		}
3775 	}
3776 }
3777 
3778 /*
3779  * Unlock and drop a pv.  If the pv is no longer associated with a pmap
3780  * and the hold count drops to zero we will free it.
3781  *
3782  * Caller should not hold any spin locks.  We are protected from hold races
3783  * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3784  * lock held.  A pv cannot be located otherwise.
3785  */
3786 static
3787 void
3788 pv_put(pv_entry_t pv)
3789 {
3790 #ifdef PMAP_DEBUG2
3791 	if (pmap_enter_debug > 0) {
3792 		--pmap_enter_debug;
3793 		kprintf("pv_put pv=%p hold=%08x\n", pv, pv->pv_hold);
3794 	}
3795 #endif
3796 
3797 	/*
3798 	 * Normal put-aways must have a pv_m associated with the pv,
3799 	 * but allow the case where the pv has been destructed due
3800 	 * to pmap_dynamic_delete.
3801 	 */
3802 	KKASSERT(pv->pv_pmap == NULL || pv->pv_m != NULL);
3803 
3804 	/*
3805 	 * Fast - shortcut most common condition
3806 	 */
3807 	if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 1))
3808 		return;
3809 
3810 	/*
3811 	 * Slow
3812 	 */
3813 	pv_unlock(pv);
3814 	pv_drop(pv);
3815 }
3816 
3817 /*
3818  * Remove the pmap association from a pv, require that pv_m already be removed,
3819  * then unlock and drop the pv.  Any pte operations must have already been
3820  * completed.  This call may result in a last-drop which will physically free
3821  * the pv.
3822  *
3823  * Removing the pmap association entails an additional drop.
3824  *
3825  * pv must be exclusively locked on call and will be disposed of on return.
3826  */
3827 static
3828 void
3829 _pv_free(pv_entry_t pv, pv_entry_t pvp PMAP_DEBUG_DECL)
3830 {
3831 	pmap_t pmap;
3832 
3833 #ifdef PMAP_DEBUG
3834 	pv->pv_func_lastfree = func;
3835 	pv->pv_line_lastfree = lineno;
3836 #endif
3837 	KKASSERT(pv->pv_m == NULL);
3838 	KKASSERT((pv->pv_hold & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
3839 		  (PV_HOLD_LOCKED|1));
3840 	if ((pmap = pv->pv_pmap) != NULL) {
3841 		spin_lock(&pmap->pm_spin);
3842 		KKASSERT(pv->pv_pmap == pmap);
3843 		if (pmap->pm_pvhint_pt == pv)
3844 			pmap->pm_pvhint_pt = NULL;
3845 		if (pmap->pm_pvhint_unused == pv)
3846 			pmap->pm_pvhint_unused = NULL;
3847 		pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
3848 		atomic_add_long(&pmap->pm_stats.resident_count, -1);
3849 		pv->pv_pmap = NULL;
3850 		pv->pv_pindex = 0;
3851 		spin_unlock(&pmap->pm_spin);
3852 
3853 		/*
3854 		 * Try to shortcut three atomic ops, otherwise fall through
3855 		 * and do it normally.  Drop two refs and the lock all in
3856 		 * one go.
3857 		 */
3858 		if (pvp) {
3859 			if (vm_page_unwire_quick(pvp->pv_m))
3860 				panic("_pv_free: bad wirecount on pvp");
3861 		}
3862 		if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 2, 0)) {
3863 #ifdef PMAP_DEBUG2
3864 			if (pmap_enter_debug > 0) {
3865 				--pmap_enter_debug;
3866 				kprintf("pv_free: free pv %p\n", pv);
3867 			}
3868 #endif
3869 			zfree(pvzone, pv);
3870 			return;
3871 		}
3872 		pv_drop(pv);	/* ref for pv_pmap */
3873 	}
3874 	pv_unlock(pv);
3875 	pv_drop(pv);
3876 }
3877 
3878 /*
3879  * This routine is very drastic, but can save the system
3880  * in a pinch.
3881  */
3882 void
3883 pmap_collect(void)
3884 {
3885 	int i;
3886 	vm_page_t m;
3887 	static int warningdone=0;
3888 
3889 	if (pmap_pagedaemon_waken == 0)
3890 		return;
3891 	pmap_pagedaemon_waken = 0;
3892 	if (warningdone < 5) {
3893 		kprintf("pmap_collect: pv_entries exhausted -- "
3894 			"suggest increasing vm.pmap_pv_entries above %ld\n",
3895 			vm_pmap_pv_entries);
3896 		warningdone++;
3897 	}
3898 
3899 	for (i = 0; i < vm_page_array_size; i++) {
3900 		m = &vm_page_array[i];
3901 		if (m->wire_count || m->hold_count)
3902 			continue;
3903 		if (vm_page_busy_try(m, TRUE) == 0) {
3904 			if (m->wire_count == 0 && m->hold_count == 0) {
3905 				pmap_remove_all(m);
3906 			}
3907 			vm_page_wakeup(m);
3908 		}
3909 	}
3910 }
3911 
3912 /*
3913  * Scan the pmap for active page table entries and issue a callback.
3914  * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3915  * its parent page table.
3916  *
3917  * pte_pv will be NULL if the page or page table is unmanaged.
3918  * pt_pv will point to the page table page containing the pte for the page.
3919  *
3920  * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3921  *	 we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3922  *	 process pmap's PD and page to the callback function.  This can be
3923  *	 confusing because the pt_pv is really a pd_pv, and the target page
3924  *	 table page is simply aliased by the pmap and not owned by it.
3925  *
3926  * It is assumed that the start and end are properly rounded to the page size.
3927  *
3928  * It is assumed that PD pages and above are managed and thus in the RB tree,
3929  * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3930  */
3931 struct pmap_scan_info {
3932 	struct pmap *pmap;
3933 	vm_offset_t sva;
3934 	vm_offset_t eva;
3935 	vm_pindex_t sva_pd_pindex;
3936 	vm_pindex_t eva_pd_pindex;
3937 	void (*func)(pmap_t, struct pmap_scan_info *,
3938 		     vm_pindex_t *, pv_entry_t, vm_offset_t,
3939 		     pt_entry_t *, void *);
3940 	void *arg;
3941 	pmap_inval_bulk_t bulk_core;
3942 	pmap_inval_bulk_t *bulk;
3943 	int count;
3944 	int stop;
3945 };
3946 
3947 static int pmap_scan_cmp(pv_entry_t pv, void *data);
3948 static int pmap_scan_callback(pv_entry_t pv, void *data);
3949 
3950 static void
3951 pmap_scan(struct pmap_scan_info *info, int smp_inval)
3952 {
3953 	struct pmap *pmap = info->pmap;
3954 	pv_entry_t pt_pv;	/* A page table PV */
3955 	pv_entry_t pte_pv;	/* A page table entry PV */
3956 	vm_pindex_t *pte_placemark;
3957 	vm_pindex_t *pt_placemark;
3958 	pt_entry_t *ptep;
3959 	pt_entry_t oldpte;
3960 	struct pv_entry dummy_pv;
3961 
3962 	info->stop = 0;
3963 	if (pmap == NULL)
3964 		return;
3965 	if (info->sva == info->eva)
3966 		return;
3967 	if (smp_inval) {
3968 		info->bulk = &info->bulk_core;
3969 		pmap_inval_bulk_init(&info->bulk_core, pmap);
3970 	} else {
3971 		info->bulk = NULL;
3972 	}
3973 
3974 	/*
3975 	 * Hold the token for stability; if the pmap is empty we have nothing
3976 	 * to do.
3977 	 */
3978 #if 0
3979 	if (pmap->pm_stats.resident_count == 0) {
3980 		return;
3981 	}
3982 #endif
3983 
3984 	info->count = 0;
3985 
3986 	/*
3987 	 * Special handling for scanning one page, which is a very common
3988 	 * operation (it is?).
3989 	 *
3990 	 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3991 	 */
3992 	if (info->sva + PAGE_SIZE == info->eva) {
3993 		if (info->sva >= VM_MAX_USER_ADDRESS) {
3994 			/*
3995 			 * Kernel mappings do not track wire counts on
3996 			 * page table pages and only maintain pd_pv and
3997 			 * pte_pv levels so pmap_scan() works.
3998 			 */
3999 			pt_pv = NULL;
4000 			pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
4001 					&pte_placemark);
4002 			KKASSERT(pte_pv == NULL);
4003 			ptep = vtopte(info->sva);
4004 		} else {
4005 			/*
4006 			 * We hold pte_placemark across the operation for
4007 			 * unmanaged pages.
4008 			 *
4009 			 * WARNING!  We must hold pt_placemark across the
4010 			 *	     *ptep test to prevent misintepreting
4011 			 *	     a non-zero *ptep as a shared page
4012 			 *	     table page.  Hold it across the function
4013 			 *	     callback as well for SMP safety.
4014 			 */
4015 			pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva),
4016 					&pte_placemark);
4017 			KKASSERT(pte_pv == NULL);
4018 			pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva),
4019 				       &pt_placemark);
4020 			if (pt_pv == NULL) {
4021 #if 0
4022 				KKASSERT(0);
4023 				pd_pv = pv_get(pmap,
4024 					       pmap_pd_pindex(info->sva),
4025 					       NULL);
4026 				if (pd_pv) {
4027 					ptep = pv_pte_lookup(pd_pv,
4028 						    pmap_pt_index(info->sva));
4029 					if (*ptep) {
4030 						info->func(pmap, info,
4031 						     pt_placemark, pd_pv,
4032 						     info->sva, ptep,
4033 						     info->arg);
4034 					} else {
4035 						pv_placemarker_wakeup(pmap,
4036 								  pt_placemark);
4037 					}
4038 					pv_put(pd_pv);
4039 				} else {
4040 					pv_placemarker_wakeup(pmap,
4041 							      pt_placemark);
4042 				}
4043 #else
4044 				pv_placemarker_wakeup(pmap, pt_placemark);
4045 #endif
4046 				pv_placemarker_wakeup(pmap, pte_placemark);
4047 				goto fast_skip;
4048 			}
4049 			ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
4050 		}
4051 
4052 		/*
4053 		 * NOTE: *ptep can't be ripped out from under us if we hold
4054 		 *	 pte_pv (or pte_placemark) locked, but bits can
4055 		 *	 change.
4056 		 */
4057 		oldpte = *ptep;
4058 		cpu_ccfence();
4059 		if (oldpte == 0) {
4060 			KKASSERT(pte_pv == NULL);
4061 			pv_placemarker_wakeup(pmap, pte_placemark);
4062 		} else {
4063 			KASSERT((oldpte & pmap->pmap_bits[PG_V_IDX]) ==
4064 				pmap->pmap_bits[PG_V_IDX],
4065 			    ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
4066 			    *ptep, oldpte, info->sva));
4067 			info->func(pmap, info, pte_placemark, pt_pv,
4068 				   info->sva, ptep, info->arg);
4069 		}
4070 		if (pt_pv)
4071 			pv_put(pt_pv);
4072 fast_skip:
4073 		pmap_inval_bulk_flush(info->bulk);
4074 		return;
4075 	}
4076 
4077 	/*
4078 	 * Nominal scan case, RB_SCAN() for PD pages and iterate from
4079 	 * there.
4080 	 *
4081 	 * WARNING! eva can overflow our standard ((N + mask) >> bits)
4082 	 *	    bounds, resulting in a pd_pindex of 0.  To solve the
4083 	 *	    problem we use an inclusive range.
4084 	 */
4085 	info->sva_pd_pindex = pmap_pd_pindex(info->sva);
4086 	info->eva_pd_pindex = pmap_pd_pindex(info->eva - PAGE_SIZE);
4087 
4088 	if (info->sva >= VM_MAX_USER_ADDRESS) {
4089 		/*
4090 		 * The kernel does not currently maintain any pv_entry's for
4091 		 * higher-level page tables.
4092 		 */
4093 		bzero(&dummy_pv, sizeof(dummy_pv));
4094 		dummy_pv.pv_pindex = info->sva_pd_pindex;
4095 		spin_lock(&pmap->pm_spin);
4096 		while (dummy_pv.pv_pindex <= info->eva_pd_pindex) {
4097 			pmap_scan_callback(&dummy_pv, info);
4098 			++dummy_pv.pv_pindex;
4099 			if (dummy_pv.pv_pindex < info->sva_pd_pindex) /*wrap*/
4100 				break;
4101 		}
4102 		spin_unlock(&pmap->pm_spin);
4103 	} else {
4104 		/*
4105 		 * User page tables maintain local PML4, PDP, PD, and PT
4106 		 * pv_entry's.  pv_entry's are not used for PTEs.
4107 		 */
4108 		spin_lock(&pmap->pm_spin);
4109 		pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot, pmap_scan_cmp,
4110 					 pmap_scan_callback, info);
4111 		spin_unlock(&pmap->pm_spin);
4112 	}
4113 	pmap_inval_bulk_flush(info->bulk);
4114 }
4115 
4116 /*
4117  * WARNING! pmap->pm_spin held
4118  *
4119  * WARNING! eva can overflow our standard ((N + mask) >> bits)
4120  *	    bounds, resulting in a pd_pindex of 0.  To solve the
4121  *	    problem we use an inclusive range.
4122  */
4123 static int
4124 pmap_scan_cmp(pv_entry_t pv, void *data)
4125 {
4126 	struct pmap_scan_info *info = data;
4127 	if (pv->pv_pindex < info->sva_pd_pindex)
4128 		return(-1);
4129 	if (pv->pv_pindex > info->eva_pd_pindex)
4130 		return(1);
4131 	return(0);
4132 }
4133 
4134 /*
4135  * pmap_scan() by PDs
4136  *
4137  * WARNING! pmap->pm_spin held
4138  */
4139 static int
4140 pmap_scan_callback(pv_entry_t pv, void *data)
4141 {
4142 	struct pmap_scan_info *info = data;
4143 	struct pmap *pmap = info->pmap;
4144 	pv_entry_t pd_pv;	/* A page directory PV */
4145 	pv_entry_t pt_pv;	/* A page table PV */
4146 	vm_pindex_t *pt_placemark;
4147 	pt_entry_t *ptep;
4148 	pt_entry_t oldpte;
4149 	vm_offset_t sva;
4150 	vm_offset_t eva;
4151 	vm_offset_t va_next;
4152 	vm_pindex_t pd_pindex;
4153 	int error;
4154 
4155 	/*
4156 	 * Stop if requested
4157 	 */
4158 	if (info->stop)
4159 		return -1;
4160 
4161 	/*
4162 	 * Pull the PD pindex from the pv before releasing the spinlock.
4163 	 *
4164 	 * WARNING: pv is faked for kernel pmap scans.
4165 	 */
4166 	pd_pindex = pv->pv_pindex;
4167 	spin_unlock(&pmap->pm_spin);
4168 	pv = NULL;	/* invalid after spinlock unlocked */
4169 
4170 	/*
4171 	 * Calculate the page range within the PD.  SIMPLE pmaps are
4172 	 * direct-mapped for the entire 2^64 address space.  Normal pmaps
4173 	 * reflect the user and kernel address space which requires
4174 	 * cannonicalization w/regards to converting pd_pindex's back
4175 	 * into addresses.
4176 	 */
4177 	sva = (pd_pindex - pmap_pd_pindex(0)) << PDPSHIFT;
4178 	if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
4179 	    (sva & PML4_SIGNMASK)) {
4180 		sva |= PML4_SIGNMASK;
4181 	}
4182 	eva = sva + NBPDP;	/* can overflow */
4183 	if (sva < info->sva)
4184 		sva = info->sva;
4185 	if (eva < info->sva || eva > info->eva)
4186 		eva = info->eva;
4187 
4188 	/*
4189 	 * NOTE: kernel mappings do not track page table pages, only
4190 	 * 	 terminal pages.
4191 	 *
4192 	 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
4193 	 *	 However, for the scan to be efficient we try to
4194 	 *	 cache items top-down.
4195 	 */
4196 	pd_pv = NULL;
4197 	pt_pv = NULL;
4198 
4199 	for (; sva < eva; sva = va_next) {
4200 		if (info->stop)
4201 			break;
4202 		if (sva >= VM_MAX_USER_ADDRESS) {
4203 			if (pt_pv) {
4204 				pv_put(pt_pv);
4205 				pt_pv = NULL;
4206 			}
4207 			goto kernel_skip;
4208 		}
4209 
4210 		/*
4211 		 * PD cache, scan shortcut if it doesn't exist.
4212 		 */
4213 		if (pd_pv == NULL) {
4214 			pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4215 		} else if (pd_pv->pv_pmap != pmap ||
4216 			   pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
4217 			pv_put(pd_pv);
4218 			pd_pv = pv_get(pmap, pmap_pd_pindex(sva), NULL);
4219 		}
4220 		if (pd_pv == NULL) {
4221 			va_next = (sva + NBPDP) & ~PDPMASK;
4222 			if (va_next < sva)
4223 				va_next = eva;
4224 			continue;
4225 		}
4226 
4227 		/*
4228 		 * PT cache
4229 		 *
4230 		 * NOTE: The cached pt_pv can be removed from the pmap when
4231 		 *	 pmap_dynamic_delete is enabled.
4232 		 */
4233 		if (pt_pv && (pt_pv->pv_pmap != pmap ||
4234 			      pt_pv->pv_pindex != pmap_pt_pindex(sva))) {
4235 			pv_put(pt_pv);
4236 			pt_pv = NULL;
4237 		}
4238 		if (pt_pv == NULL) {
4239 			pt_pv = pv_get_try(pmap, pmap_pt_pindex(sva),
4240 					   &pt_placemark, &error);
4241 			if (error) {
4242 				pv_put(pd_pv);	/* lock order */
4243 				pd_pv = NULL;
4244 				if (pt_pv) {
4245 					pv_lock(pt_pv);
4246 					pv_put(pt_pv);
4247 					pt_pv = NULL;
4248 				} else {
4249 					pv_placemarker_wait(pmap, pt_placemark);
4250 				}
4251 				va_next = sva;
4252 				continue;
4253 			}
4254 			/* may have to re-check later if pt_pv is NULL here */
4255 		}
4256 
4257 		/*
4258 		 * If pt_pv is NULL we either have a shared page table
4259 		 * page (NOT IMPLEMENTED XXX) and must issue a callback
4260 		 * specific to that case, or there is no page table page.
4261 		 *
4262 		 * Either way we can skip the page table page.
4263 		 *
4264 		 * WARNING! pt_pv can also be NULL due to a pv creation
4265 		 *	    race where we find it to be NULL and then
4266 		 *	    later see a pte_pv.  But its possible the pt_pv
4267 		 *	    got created inbetween the two operations, so
4268 		 *	    we must check.
4269 		 *
4270 		 *	    XXX This should no longer be the case because
4271 		 *	    we have pt_placemark.
4272 		 */
4273 		if (pt_pv == NULL) {
4274 #if 0
4275 			/* XXX REMOVED */
4276 			/*
4277 			 * Possible unmanaged (shared from another pmap)
4278 			 * page table page.
4279 			 *
4280 			 * WARNING!  We must hold pt_placemark across the
4281 			 *	     *ptep test to prevent misintepreting
4282 			 *	     a non-zero *ptep as a shared page
4283 			 *	     table page.  Hold it across the function
4284 			 *	     callback as well for SMP safety.
4285 			 */
4286 			KKASSERT(0);
4287 			ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
4288 			if (*ptep & pmap->pmap_bits[PG_V_IDX]) {
4289 				info->func(pmap, info, pt_placemark, pd_pv,
4290 					   sva, ptep, info->arg);
4291 			} else {
4292 				pv_placemarker_wakeup(pmap, pt_placemark);
4293 			}
4294 #else
4295 			pv_placemarker_wakeup(pmap, pt_placemark);
4296 #endif
4297 
4298 			/*
4299 			 * Done, move to next page table page.
4300 			 */
4301 			va_next = (sva + NBPDR) & ~PDRMASK;
4302 			if (va_next < sva)
4303 				va_next = eva;
4304 			continue;
4305 		}
4306 
4307 		/*
4308 		 * From this point in the loop testing pt_pv for non-NULL
4309 		 * means we are in UVM, else if it is NULL we are in KVM.
4310 		 *
4311 		 * Limit our scan to either the end of the va represented
4312 		 * by the current page table page, or to the end of the
4313 		 * range being removed.
4314 		 */
4315 kernel_skip:
4316 		va_next = (sva + NBPDR) & ~PDRMASK;
4317 		if (va_next < sva)
4318 			va_next = eva;
4319 		if (va_next > eva)
4320 			va_next = eva;
4321 
4322 		/*
4323 		 * Scan the page table for pages.  Some pages may not be
4324 		 * managed (might not have a pv_entry).
4325 		 *
4326 		 * There is no page table management for kernel pages so
4327 		 * pt_pv will be NULL in that case, but otherwise pt_pv
4328 		 * is non-NULL, locked, and referenced.
4329 		 */
4330 
4331 		/*
4332 		 * At this point a non-NULL pt_pv means a UVA, and a NULL
4333 		 * pt_pv means a KVA.
4334 		 */
4335 		if (pt_pv)
4336 			ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
4337 		else
4338 			ptep = vtopte(sva);
4339 
4340 		while (sva < va_next) {
4341 			vm_pindex_t *pte_placemark;
4342 			pv_entry_t pte_pv;
4343 
4344 			/*
4345 			 * Yield every 64 pages, stop if requested.
4346 			 */
4347 			if ((++info->count & 63) == 0)
4348 				lwkt_user_yield();
4349 			if (info->stop)
4350 				break;
4351 
4352 			/*
4353 			 * We can shortcut our scan if *ptep == 0.  This is
4354 			 * an unlocked check.
4355 			 */
4356 			if (*ptep == 0) {
4357 				sva += PAGE_SIZE;
4358 				++ptep;
4359 				continue;
4360 			}
4361 			cpu_ccfence();
4362 
4363 			/*
4364 			 * Acquire the pte_placemark.  pte_pv's won't exist
4365 			 * for leaf pages.
4366 			 *
4367 			 * A multitude of races are possible here so if we
4368 			 * cannot lock definite state we clean out our cache
4369 			 * and break the inner while() loop to force a loop
4370 			 * up to the top of the for().
4371 			 *
4372 			 * XXX unlock/relock pd_pv, pt_pv, and re-test their
4373 			 *     validity instead of looping up?
4374 			 */
4375 			pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
4376 					    &pte_placemark, &error);
4377 			KKASSERT(pte_pv == NULL);
4378 			if (error) {
4379 				if (pd_pv) {
4380 					pv_put(pd_pv);	/* lock order */
4381 					pd_pv = NULL;
4382 				}
4383 				if (pt_pv) {
4384 					pv_put(pt_pv);	/* lock order */
4385 					pt_pv = NULL;
4386 				}
4387 				pv_placemarker_wait(pmap, pte_placemark);
4388 				va_next = sva;		/* retry */
4389 				break;
4390 			}
4391 
4392 			/*
4393 			 * Reload *ptep after successfully locking the
4394 			 * pindex.
4395 			 */
4396 			cpu_ccfence();
4397 			oldpte = *ptep;
4398 			if (oldpte == 0) {
4399 				pv_placemarker_wakeup(pmap, pte_placemark);
4400 				sva += PAGE_SIZE;
4401 				++ptep;
4402 				continue;
4403 			}
4404 
4405 			/*
4406 			 * We can't hold pd_pv across the callback (because
4407 			 * we don't pass it to the callback and the callback
4408 			 * might deadlock)
4409 			 */
4410 			if (pd_pv) {
4411 				vm_page_wire_quick(pd_pv->pv_m);
4412 				pv_unlock(pd_pv);
4413 			}
4414 
4415 			/*
4416 			 * Ready for the callback.  The locked placemarker
4417 			 * is consumed by the callback.
4418 			 */
4419 			if (oldpte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4420 				/*
4421 				 * Managed pte
4422 				 */
4423 				KASSERT((oldpte & pmap->pmap_bits[PG_V_IDX]),
4424 				    ("badC *ptep %016lx/%016lx sva %016lx",
4425 				    *ptep, oldpte, sva));
4426 				/*
4427 				 * We must unlock pd_pv across the callback
4428 				 * to avoid deadlocks on any recursive
4429 				 * disposal.  Re-check that it still exists
4430 				 * after re-locking.
4431 				 *
4432 				 * Call target disposes of pte_placemark
4433 				 * and may destroy but will not dispose
4434 				 * of pt_pv.
4435 				 */
4436 				info->func(pmap, info, pte_placemark, pt_pv,
4437 					   sva, ptep, info->arg);
4438 			} else {
4439 				/*
4440 				 * Unmanaged pte
4441 				 *
4442 				 * We must unlock pd_pv across the callback
4443 				 * to avoid deadlocks on any recursive
4444 				 * disposal.  Re-check that it still exists
4445 				 * after re-locking.
4446 				 *
4447 				 * Call target disposes of pte_placemark
4448 				 * and may destroy but will not dispose
4449 				 * of pt_pv.
4450 				 */
4451 				KASSERT((oldpte & pmap->pmap_bits[PG_V_IDX]),
4452 				    ("badD *ptep %016lx/%016lx sva %016lx ",
4453 				     *ptep, oldpte, sva));
4454 				info->func(pmap, info, pte_placemark, pt_pv,
4455 					   sva, ptep, info->arg);
4456 			}
4457 			if (pd_pv) {
4458 				pv_lock(pd_pv);
4459 				if (vm_page_unwire_quick(pd_pv->pv_m)) {
4460 					panic("pmap_scan_callback: "
4461 					      "bad wirecount on pd_pv");
4462 				}
4463 				if (pd_pv->pv_pmap == NULL) {
4464 					va_next = sva;		/* retry */
4465 					break;
4466 				}
4467 			}
4468 
4469 			/*
4470 			 * NOTE: The cached pt_pv can be removed from the
4471 			 *	 pmap when pmap_dynamic_delete is enabled,
4472 			 *	 which will cause ptep to become stale.
4473 			 *
4474 			 *	 This also means that no pages remain under
4475 			 *	 the PT, so we can just break out of the inner
4476 			 *	 loop and let the outer loop clean everything
4477 			 *	 up.
4478 			 */
4479 			if (pt_pv && pt_pv->pv_pmap != pmap)
4480 				break;
4481 			sva += PAGE_SIZE;
4482 			++ptep;
4483 		}
4484 	}
4485 	if (pd_pv) {
4486 		pv_put(pd_pv);
4487 		pd_pv = NULL;
4488 	}
4489 	if (pt_pv) {
4490 		pv_put(pt_pv);
4491 		pt_pv = NULL;
4492 	}
4493 	if ((++info->count & 7) == 0)
4494 		lwkt_user_yield();
4495 
4496 	/*
4497 	 * Relock before returning.
4498 	 */
4499 	spin_lock(&pmap->pm_spin);
4500 	return (0);
4501 }
4502 
4503 void
4504 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4505 {
4506 	struct pmap_scan_info info;
4507 
4508 	info.pmap = pmap;
4509 	info.sva = sva;
4510 	info.eva = eva;
4511 	info.func = pmap_remove_callback;
4512 	info.arg = NULL;
4513 	pmap_scan(&info, 1);
4514 #if 0
4515 	cpu_invltlb();
4516 	if (eva - sva < 1024*1024) {
4517 		while (sva < eva) {
4518 			cpu_invlpg((void *)sva);
4519 			sva += PAGE_SIZE;
4520 		}
4521 	}
4522 #endif
4523 }
4524 
4525 static void
4526 pmap_remove_noinval(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
4527 {
4528 	struct pmap_scan_info info;
4529 
4530 	info.pmap = pmap;
4531 	info.sva = sva;
4532 	info.eva = eva;
4533 	info.func = pmap_remove_callback;
4534 	info.arg = NULL;
4535 	pmap_scan(&info, 0);
4536 }
4537 
4538 static void
4539 pmap_remove_callback(pmap_t pmap, struct pmap_scan_info *info,
4540 		     vm_pindex_t *pte_placemark, pv_entry_t pt_pv,
4541 		     vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
4542 {
4543 	pt_entry_t pte;
4544 
4545 	/*
4546 	 * Managed or unmanaged pte (pte_placemark is non-NULL)
4547 	 *
4548 	 * pt_pv's wire_count is still bumped by unmanaged pages
4549 	 * so we must decrement it manually.
4550 	 *
4551 	 * We have to unwire the target page table page.
4552 	 */
4553 	pte = pmap_inval_bulk(info->bulk, va, ptep, 0);
4554 	if (pte & pmap->pmap_bits[PG_MANAGED_IDX]) {
4555 		vm_page_t p;
4556 
4557 		p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
4558 		KKASSERT(pte & pmap->pmap_bits[PG_V_IDX]);
4559 		if (pte & pmap->pmap_bits[PG_M_IDX])
4560 			vm_page_dirty(p);
4561 		if (pte & pmap->pmap_bits[PG_A_IDX])
4562 			vm_page_flag_set(p, PG_REFERENCED);
4563 
4564 		/*
4565 		 * NOTE: p is not hard-busied so it is not safe to
4566 		 *	 clear PG_MAPPED and PG_WRITEABLE on the 1->0
4567 		 *	 transition against them being set in
4568 		 *	 pmap_enter().
4569 		 */
4570 		if (pte & pmap->pmap_bits[PG_RW_IDX])
4571 			atomic_add_long(&p->md.writeable_count, -1);
4572 		pmap_page_stats_deleting(
4573 			atomic_fetchadd_long(&p->md.pmap_count, -1));
4574 	}
4575 	if (pte & pmap->pmap_bits[PG_V_IDX]) {
4576 		atomic_add_long(&pmap->pm_stats.resident_count, -1);
4577 		if (pt_pv && vm_page_unwire_quick(pt_pv->pv_m))
4578 			panic("pmap_remove: insufficient wirecount");
4579 	}
4580 	if (pte & pmap->pmap_bits[PG_W_IDX])
4581 		atomic_add_long(&pmap->pm_stats.wired_count, -1);
4582 	if (pte & pmap->pmap_bits[PG_G_IDX])
4583 		cpu_invlpg((void *)va);
4584 	pv_placemarker_wakeup(pmap, pte_placemark);
4585 }
4586 
4587 /*
4588  * Removes this physical page from all physical maps in which it resides.
4589  * Reflects back modify bits to the pager.
4590  *
4591  * This routine may not be called from an interrupt.
4592  *
4593  * The page must be busied by its caller, preventing new ptes from being
4594  * installed.  This allows us to assert that pmap_count is zero and safely
4595  * clear the MAPPED and WRITEABLE bits upon completion.
4596  */
4597 static
4598 void
4599 pmap_remove_all(vm_page_t m)
4600 {
4601 	int retry;
4602 
4603 	if (!pmap_initialized)
4604 		return;
4605 
4606 	/*
4607 	 * pmap_count doesn't cover fictitious pages, but PG_MAPPED does
4608 	 * (albeit without certain race protections).
4609 	 */
4610 #if 0
4611 	if (m->md.pmap_count == 0)
4612 		return;
4613 #endif
4614 	if ((m->flags & PG_MAPPED) == 0)
4615 		return;
4616 
4617 	retry = ticks + hz * 60;
4618 again:
4619 	PMAP_PAGE_BACKING_SCAN(m, NULL, ipmap, iptep, ipte, iva) {
4620 		if (!pmap_inval_smp_cmpset(ipmap, iva, iptep, ipte, 0))
4621 			PMAP_PAGE_BACKING_RETRY;
4622 		if (ipte & ipmap->pmap_bits[PG_MANAGED_IDX]) {
4623 			if (ipte & ipmap->pmap_bits[PG_M_IDX])
4624 				vm_page_dirty(m);
4625 			if (ipte & ipmap->pmap_bits[PG_A_IDX])
4626 				vm_page_flag_set(m, PG_REFERENCED);
4627 
4628 			/*
4629 			 * NOTE: m is not hard-busied so it is not safe to
4630 			 *	 clear PG_MAPPED and PG_WRITEABLE on the 1->0
4631 			 *	 transition against them being set in
4632 			 *	 pmap_enter().
4633 			 */
4634 			if (ipte & ipmap->pmap_bits[PG_RW_IDX])
4635 				atomic_add_long(&m->md.writeable_count, -1);
4636 			pmap_page_stats_deleting(
4637 				atomic_fetchadd_long(&m->md.pmap_count, -1));
4638 		}
4639 
4640 		/*
4641 		 * Cleanup various tracking counters.  pt_pv can't go away
4642 		 * due to our wired ref.
4643 		 */
4644 		if (ipmap != &kernel_pmap) {
4645 			pv_entry_t pt_pv;
4646 
4647 			spin_lock_shared(&ipmap->pm_spin);
4648 			pt_pv = pv_entry_lookup(ipmap, pmap_pt_pindex(iva));
4649 			spin_unlock_shared(&ipmap->pm_spin);
4650 
4651 			if (pt_pv) {
4652 				if (vm_page_unwire_quick(pt_pv->pv_m)) {
4653 					panic("pmap_remove_all: bad "
4654 					      "wire_count on pt_pv");
4655 				}
4656 				atomic_add_long(
4657 					&ipmap->pm_stats.resident_count, -1);
4658 			}
4659 		}
4660 		if (ipte & ipmap->pmap_bits[PG_W_IDX])
4661 			atomic_add_long(&ipmap->pm_stats.wired_count, -1);
4662 		if (ipte & ipmap->pmap_bits[PG_G_IDX])
4663 			cpu_invlpg((void *)iva);
4664 	} PMAP_PAGE_BACKING_DONE;
4665 
4666 	/*
4667 	 * pmap_count should be zero but it is possible to race a pmap_enter()
4668 	 * replacement (see 'oldm').  Once it is zero it cannot become
4669 	 * non-zero because the page is hard-busied.
4670 	 */
4671 	if (m->md.pmap_count || m->md.writeable_count) {
4672 		tsleep(&m->md.pmap_count, 0, "pgunm", 1);
4673 		if (retry - ticks > 0)
4674 			goto again;
4675 		panic("pmap_remove_all: cannot return pmap_count "
4676 		      "to 0 (%p, %ld, %ld)",
4677 		      m, m->md.pmap_count, m->md.writeable_count);
4678 	}
4679 	vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
4680 }
4681 
4682 /*
4683  * Removes the page from a particular pmap.
4684  *
4685  * The page must be busied by the caller.
4686  */
4687 void
4688 pmap_remove_specific(pmap_t pmap_match, vm_page_t m)
4689 {
4690 	if (!pmap_initialized)
4691 		return;
4692 
4693 	/*
4694 	 * PG_MAPPED test works for both non-fictitious and fictitious pages.
4695 	 */
4696 	if ((m->flags & PG_MAPPED) == 0)
4697 		return;
4698 
4699 	PMAP_PAGE_BACKING_SCAN(m, pmap_match, ipmap, iptep, ipte, iva) {
4700 		if (!pmap_inval_smp_cmpset(ipmap, iva, iptep, ipte, 0))
4701 			PMAP_PAGE_BACKING_RETRY;
4702 		if (ipte & ipmap->pmap_bits[PG_MANAGED_IDX]) {
4703 			if (ipte & ipmap->pmap_bits[PG_M_IDX])
4704 				vm_page_dirty(m);
4705 			if (ipte & ipmap->pmap_bits[PG_A_IDX])
4706 				vm_page_flag_set(m, PG_REFERENCED);
4707 
4708 			/*
4709 			 * NOTE: m is not hard-busied so it is not safe to
4710 			 *	 clear PG_MAPPED and PG_WRITEABLE on the 1->0
4711 			 *	 transition against them being set in
4712 			 *	 pmap_enter().
4713 			 */
4714 			if (ipte & ipmap->pmap_bits[PG_RW_IDX])
4715 				atomic_add_long(&m->md.writeable_count, -1);
4716 			pmap_page_stats_deleting(
4717 				atomic_fetchadd_long(&m->md.pmap_count, -1));
4718 		}
4719 
4720 		/*
4721 		 * Cleanup various tracking counters.  pt_pv can't go away
4722 		 * due to our wired ref.
4723 		 */
4724 		if (ipmap != &kernel_pmap) {
4725 			pv_entry_t pt_pv;
4726 
4727 			spin_lock_shared(&ipmap->pm_spin);
4728 			pt_pv = pv_entry_lookup(ipmap, pmap_pt_pindex(iva));
4729 			spin_unlock_shared(&ipmap->pm_spin);
4730 
4731 			if (pt_pv) {
4732 				atomic_add_long(
4733 					&ipmap->pm_stats.resident_count, -1);
4734 				if (vm_page_unwire_quick(pt_pv->pv_m)) {
4735 					panic("pmap_remove_specific: bad "
4736 					      "wire_count on pt_pv");
4737 				}
4738 			}
4739 		}
4740 		if (ipte & ipmap->pmap_bits[PG_W_IDX])
4741 			atomic_add_long(&ipmap->pm_stats.wired_count, -1);
4742 		if (ipte & ipmap->pmap_bits[PG_G_IDX])
4743 			cpu_invlpg((void *)iva);
4744 	} PMAP_PAGE_BACKING_DONE;
4745 }
4746 
4747 /*
4748  * Set the physical protection on the specified range of this map
4749  * as requested.  This function is typically only used for debug watchpoints
4750  * and COW pages.
4751  *
4752  * This function may not be called from an interrupt if the map is
4753  * not the kernel_pmap.
4754  *
4755  * NOTE!  For shared page table pages we just unmap the page.
4756  */
4757 void
4758 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
4759 {
4760 	struct pmap_scan_info info;
4761 	/* JG review for NX */
4762 
4763 	if (pmap == NULL)
4764 		return;
4765 	if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == VM_PROT_NONE) {
4766 		pmap_remove(pmap, sva, eva);
4767 		return;
4768 	}
4769 	if (prot & VM_PROT_WRITE)
4770 		return;
4771 	info.pmap = pmap;
4772 	info.sva = sva;
4773 	info.eva = eva;
4774 	info.func = pmap_protect_callback;
4775 	info.arg = &prot;
4776 	pmap_scan(&info, 1);
4777 }
4778 
4779 static
4780 void
4781 pmap_protect_callback(pmap_t pmap, struct pmap_scan_info *info,
4782 		      vm_pindex_t *pte_placemark,
4783 		      pv_entry_t pt_pv, vm_offset_t va,
4784 		      pt_entry_t *ptep, void *arg __unused)
4785 {
4786 	pt_entry_t pbits;
4787 	pt_entry_t cbits;
4788 	vm_page_t m;
4789 
4790 again:
4791 	pbits = *ptep;
4792 	cpu_ccfence();
4793 	cbits = pbits;
4794 	if (pbits & pmap->pmap_bits[PG_MANAGED_IDX]) {
4795 		cbits &= ~pmap->pmap_bits[PG_A_IDX];
4796 		cbits &= ~pmap->pmap_bits[PG_M_IDX];
4797 	}
4798 	/* else unmanaged page, adjust bits, no wire changes */
4799 
4800 	if (ptep) {
4801 		cbits &= ~pmap->pmap_bits[PG_RW_IDX];
4802 #ifdef PMAP_DEBUG2
4803 		if (pmap_enter_debug > 0) {
4804 			--pmap_enter_debug;
4805 			kprintf("pmap_protect va=%lx ptep=%p "
4806 				"pt_pv=%p cbits=%08lx\n",
4807 				va, ptep, pt_pv, cbits
4808 			);
4809 		}
4810 #endif
4811 		if (pbits != cbits) {
4812 			if (!pmap_inval_smp_cmpset(pmap, va,
4813 						   ptep, pbits, cbits)) {
4814 				goto again;
4815 			}
4816 		}
4817 		if (pbits & pmap->pmap_bits[PG_MANAGED_IDX]) {
4818 			m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4819 			if (pbits & pmap->pmap_bits[PG_A_IDX])
4820 				vm_page_flag_set(m, PG_REFERENCED);
4821 			if (pbits & pmap->pmap_bits[PG_M_IDX])
4822 				vm_page_dirty(m);
4823 			if (pbits & pmap->pmap_bits[PG_RW_IDX])
4824 				atomic_add_long(&m->md.writeable_count, -1);
4825 
4826 		}
4827 	}
4828 	pv_placemarker_wakeup(pmap, pte_placemark);
4829 }
4830 
4831 /*
4832  * Insert the vm_page (m) at the virtual address (va), replacing any prior
4833  * mapping at that address.  Set protection and wiring as requested.
4834  *
4835  * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4836  * possible.  If it is we enter the page into the appropriate shared pmap
4837  * hanging off the related VM object instead of the passed pmap, then we
4838  * share the page table page from the VM object's pmap into the current pmap.
4839  *
4840  * NOTE: This routine MUST insert the page into the pmap now, it cannot
4841  *	 lazy-evaluate.
4842  */
4843 void
4844 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
4845 	   boolean_t wired, vm_map_entry_t entry)
4846 {
4847 	pv_entry_t pt_pv;	/* page table */
4848 	pv_entry_t pte_pv;	/* page table entry */
4849 	vm_pindex_t *pte_placemark;
4850 	pt_entry_t *ptep;
4851 	pt_entry_t origpte;
4852 	vm_paddr_t opa;
4853 	vm_page_t oldm;
4854 	pt_entry_t newpte;
4855 	vm_paddr_t pa;
4856 
4857 	if (pmap == NULL)
4858 		return;
4859 	va = trunc_page(va);
4860 #ifdef PMAP_DIAGNOSTIC
4861 	if (va >= KvaEnd)
4862 		panic("pmap_enter: toobig");
4863 	if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
4864 		panic("pmap_enter: invalid to pmap_enter page table "
4865 		      "pages (va: 0x%lx)", va);
4866 #endif
4867 	if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
4868 		kprintf("Warning: pmap_enter called on UVA with "
4869 			"kernel_pmap\n");
4870 #ifdef DDB
4871 		db_print_backtrace();
4872 #endif
4873 	}
4874 	if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
4875 		kprintf("Warning: pmap_enter called on KVA without"
4876 			"kernel_pmap\n");
4877 #ifdef DDB
4878 		db_print_backtrace();
4879 #endif
4880 	}
4881 
4882 	/*
4883 	 * Get the locked page table page (pt_pv) for our new page table
4884 	 * entry, allocating it if necessary.
4885 	 *
4886 	 * There is no pte_pv for a terminal pte so the terminal pte will
4887 	 * be locked via pte_placemark.
4888 	 *
4889 	 * Only MMU actions by the CPU itself can modify the ptep out from
4890 	 * under us.
4891 	 *
4892 	 * If the pmap is still being initialized we assume existing
4893 	 * page tables.
4894 	 *
4895 	 * NOTE: Kernel mapppings do not track page table pages
4896 	 *	 (i.e. there is no pt_pv pt_pv structure).
4897 	 *
4898 	 * NOTE: origpte here is 'tentative', used only to check for
4899 	 *	 the degenerate case where the entry already exists and
4900 	 *	 matches.
4901 	 */
4902 	if (pmap_initialized == FALSE) {
4903 		pte_pv = NULL;
4904 		pt_pv = NULL;
4905 		pte_placemark = NULL;
4906 		ptep = vtopte(va);
4907 		origpte = *ptep;
4908 	} else {
4909 		pte_pv = pv_get(pmap, pmap_pte_pindex(va), &pte_placemark);
4910 		KKASSERT(pte_pv == NULL);
4911 		if (va >= VM_MAX_USER_ADDRESS) {
4912 			pt_pv = NULL;
4913 			ptep = vtopte(va);
4914 		} else {
4915 			pt_pv = pmap_allocpte(pmap, pmap_pt_pindex(va), NULL);
4916 			ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
4917 		}
4918 		origpte = *ptep;
4919 		cpu_ccfence();
4920 	}
4921 
4922 	pa = VM_PAGE_TO_PHYS(m);
4923 
4924 	/*
4925 	 * Calculate the new PTE.
4926 	 */
4927 	newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) |
4928 		 pmap->pmap_bits[PG_V_IDX] | pmap->pmap_bits[PG_A_IDX]);
4929 	if (wired)
4930 		newpte |= pmap->pmap_bits[PG_W_IDX];
4931 	if (va < VM_MAX_USER_ADDRESS)
4932 		newpte |= pmap->pmap_bits[PG_U_IDX];
4933 	if ((m->flags & PG_FICTITIOUS) == 0)
4934 		newpte |= pmap->pmap_bits[PG_MANAGED_IDX];
4935 //	if (pmap == &kernel_pmap)
4936 //		newpte |= pgeflag;
4937 	newpte |= pmap->pmap_cache_bits_pte[m->pat_mode];
4938 
4939 	/*
4940 	 * It is possible for multiple faults to occur in threaded
4941 	 * environments, the existing pte might be correct.
4942 	 */
4943 	if (((origpte ^ newpte) &
4944 	    ~(pt_entry_t)(pmap->pmap_bits[PG_M_IDX] |
4945 			  pmap->pmap_bits[PG_A_IDX])) == 0) {
4946 		goto done;
4947 	}
4948 
4949 	/*
4950 	 * Adjust page flags.  The page is soft-busied or hard-busied, we
4951 	 * should be able to safely set PG_* flag bits even with the (shared)
4952 	 * soft-busy.
4953 	 *
4954 	 * The pmap_count and writeable_count is only tracked for
4955 	 * non-fictitious pages.  As a bit of a safety, bump pmap_count
4956 	 * and set the PG_* bits before mapping the page.  If another part
4957 	 * of the system does not properly hard-busy the page (against our
4958 	 * soft-busy or hard-busy) in order to remove mappings it might not
4959 	 * see the pte that we are about to add and thus will not be able to
4960 	 * drop pmap_count to 0.
4961 	 *
4962 	 * The PG_MAPPED and PG_WRITEABLE flags are set for any type of page.
4963 	 *
4964 	 * NOTE! PG_MAPPED and PG_WRITEABLE can only be cleared when
4965 	 *	 the page is hard-busied AND pmap_count is 0.  This
4966 	 *	 interlocks our setting of the flags here.
4967 	 */
4968 	/*vm_page_spin_lock(m);*/
4969 	if ((m->flags & PG_FICTITIOUS) == 0) {
4970 		pmap_page_stats_adding(
4971 			atomic_fetchadd_long(&m->md.pmap_count, 1));
4972 		if (newpte & pmap->pmap_bits[PG_RW_IDX])
4973 			atomic_add_long(&m->md.writeable_count, 1);
4974 	}
4975 	if (newpte & pmap->pmap_bits[PG_RW_IDX]) {
4976 		if ((m->flags & (PG_MAPPED | PG_WRITEABLE)) == 0)
4977 			vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
4978 	} else {
4979 		if ((m->flags & PG_MAPPED) == 0)
4980 			vm_page_flag_set(m, PG_MAPPED);
4981 	}
4982 	/*vm_page_spin_unlock(m);*/
4983 
4984 	/*
4985 	 * A race can develop when replacing an existing mapping.  The new
4986 	 * page has been busied and the pte is placemark-locked, but the
4987 	 * old page is could be ripped out from under us at any time by
4988 	 * a backing scan.
4989 	 *
4990 	 * The race is handled by having the backing scans check pmap_count
4991 	 * writeable_count when doing operations that should ensure one
4992 	 * becomes 0.
4993 	 */
4994 	opa = origpte & PG_FRAME;
4995 	if (opa && (origpte & pmap->pmap_bits[PG_MANAGED_IDX])) {
4996 		oldm = PHYS_TO_VM_PAGE(opa);
4997 		KKASSERT(opa == oldm->phys_addr);
4998 		KKASSERT(entry != NULL);
4999 	} else {
5000 		oldm = NULL;
5001 	}
5002 
5003 	/*
5004 	 * Swap the new and old PTEs and perform any necessary SMP
5005 	 * synchronization.
5006 	 */
5007 	if ((prot & VM_PROT_NOSYNC) || (opa == 0 && pt_pv != NULL)) {
5008 		/*
5009 		 * Explicitly permitted to avoid pmap cpu mask synchronization
5010 		 * or the prior content of a non-kernel-related pmap was
5011 		 * invalid.
5012 		 */
5013 		origpte = atomic_swap_long(ptep, newpte);
5014 		if (opa)
5015 			cpu_invlpg((void *)va);
5016 	} else {
5017 		/*
5018 		 * Not permitted to avoid pmap cpu mask synchronization
5019 		 * or there prior content being replaced or this is a kernel
5020 		 * related pmap.
5021 		 *
5022 		 * Due to other kernel optimizations, we cannot assume a
5023 		 * 0->non_zero transition of *ptep can be done with a swap.
5024 		 */
5025 		origpte = pmap_inval_smp(pmap, va, 1, ptep, newpte);
5026 	}
5027 	opa = origpte & PG_FRAME;
5028 
5029 #ifdef PMAP_DEBUG2
5030 	if (pmap_enter_debug > 0) {
5031 		--pmap_enter_debug;
5032 		kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
5033 			" pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
5034 			va, m,
5035 			origpte, newpte, ptep,
5036 			pte_pv, pt_pv, opa, prot);
5037 	}
5038 #endif
5039 
5040 	/*
5041 	 * Account for the changes in the pt_pv and pmap.
5042 	 *
5043 	 * Retain the same wiring count due to replacing an existing page,
5044 	 * or bump the wiring count for a new page.
5045 	 */
5046 	if (pt_pv && opa == 0) {
5047 		vm_page_wire_quick(pt_pv->pv_m);
5048 		atomic_add_long(&pt_pv->pv_pmap->pm_stats.resident_count, 1);
5049 	}
5050 	if (wired && (origpte & pmap->pmap_bits[PG_W_IDX]) == 0)
5051 		atomic_add_long(&pmap->pm_stats.wired_count, 1);
5052 
5053 	/*
5054 	 * Account for the removal of the old page.  pmap and pt_pv stats
5055 	 * have already been fully adjusted for both.
5056 	 *
5057 	 * WARNING! oldm is not soft or hard-busied.  The pte at worst can
5058 	 *	    only be removed out from under us since we hold the
5059 	 *	    placemarker.  So if it is still there, it must not have
5060 	 *	    changed.
5061 	 */
5062 	if (opa && (origpte & pmap->pmap_bits[PG_MANAGED_IDX])) {
5063 		KKASSERT(oldm == PHYS_TO_VM_PAGE(opa));
5064 		if (origpte & pmap->pmap_bits[PG_M_IDX])
5065 			vm_page_dirty(oldm);
5066 		if (origpte & pmap->pmap_bits[PG_A_IDX])
5067 			vm_page_flag_set(oldm, PG_REFERENCED);
5068 
5069 		/*
5070 		 * NOTE: oldm is not hard-busied so it is not safe to
5071 		 *	 clear PG_MAPPED and PG_WRITEABLE on the 1->0
5072 		 *	 transition against them being set in
5073 		 *	 pmap_enter().
5074 		 */
5075 		if (origpte & pmap->pmap_bits[PG_RW_IDX])
5076 			atomic_add_long(&oldm->md.writeable_count, -1);
5077 		pmap_page_stats_deleting(
5078 			atomic_fetchadd_long(&oldm->md.pmap_count, -1));
5079 	}
5080 
5081 done:
5082 	KKASSERT((newpte & pmap->pmap_bits[PG_MANAGED_IDX]) == 0 ||
5083 		 (m->flags & PG_MAPPED));
5084 
5085 	/*
5086 	 * Cleanup the pv entry, allowing other accessors.  If the new page
5087 	 * is not managed but we have a pte_pv (which was locking our
5088 	 * operation), we can free it now.  pte_pv->pv_m should be NULL.
5089 	 */
5090 	if (pte_placemark)
5091 		pv_placemarker_wakeup(pmap, pte_placemark);
5092 	if (pt_pv)
5093 		pv_put(pt_pv);
5094 }
5095 
5096 /*
5097  * Make a temporary mapping for a physical address.  This is only intended
5098  * to be used for panic dumps.
5099  *
5100  * The caller is responsible for calling smp_invltlb().
5101  */
5102 void *
5103 pmap_kenter_temporary(vm_paddr_t pa, long i)
5104 {
5105 	pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
5106 	return ((void *)crashdumpmap);
5107 }
5108 
5109 #if 0
5110 #define MAX_INIT_PT (96)
5111 
5112 /*
5113  * This routine preloads the ptes for a given object into the specified pmap.
5114  * This eliminates the blast of soft faults on process startup and
5115  * immediately after an mmap.
5116  */
5117 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
5118 #endif
5119 
5120 void
5121 pmap_object_init_pt(pmap_t pmap, vm_map_entry_t entry,
5122 		    vm_offset_t addr, vm_size_t size, int limit)
5123 {
5124 #if 0
5125 	vm_prot_t prot = entry->protection;
5126 	vm_object_t object = entry->ba.object;
5127 	vm_pindex_t pindex = atop(entry->ba.offset + (addr - entry->ba.start));
5128 	struct rb_vm_page_scan_info info;
5129 	struct lwp *lp;
5130 	vm_size_t psize;
5131 
5132 	/*
5133 	 * We can't preinit if read access isn't set or there is no pmap
5134 	 * or object.
5135 	 */
5136 	if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
5137 		return;
5138 
5139 	/*
5140 	 * We can't preinit if the pmap is not the current pmap
5141 	 */
5142 	lp = curthread->td_lwp;
5143 	if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
5144 		return;
5145 
5146 	/*
5147 	 * Misc additional checks
5148 	 */
5149 	psize = x86_64_btop(size);
5150 
5151 	if ((object->type != OBJT_VNODE) ||
5152 		((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
5153 			(object->resident_page_count > MAX_INIT_PT))) {
5154 		return;
5155 	}
5156 
5157 	if (pindex + psize > object->size) {
5158 		if (object->size < pindex)
5159 			return;
5160 		psize = object->size - pindex;
5161 	}
5162 
5163 	if (psize == 0)
5164 		return;
5165 
5166 	/*
5167 	 * If everything is segment-aligned do not pre-init here.  Instead
5168 	 * allow the normal vm_fault path to pass a segment hint to
5169 	 * pmap_enter() which will then use an object-referenced shared
5170 	 * page table page.
5171 	 */
5172 	if ((addr & SEG_MASK) == 0 &&
5173 	    (ctob(psize) & SEG_MASK) == 0 &&
5174 	    (ctob(pindex) & SEG_MASK) == 0) {
5175 		return;
5176 	}
5177 
5178 	/*
5179 	 * Use a red-black scan to traverse the requested range and load
5180 	 * any valid pages found into the pmap.
5181 	 *
5182 	 * We cannot safely scan the object's memq without holding the
5183 	 * object token.
5184 	 */
5185 	info.start_pindex = pindex;
5186 	info.end_pindex = pindex + psize - 1;
5187 	info.limit = limit;
5188 	info.mpte = NULL;
5189 	info.addr = addr;
5190 	info.pmap = pmap;
5191 	info.object = object;
5192 	info.entry = entry;
5193 
5194 	/*
5195 	 * By using the NOLK scan, the callback function must be sure
5196 	 * to return -1 if the VM page falls out of the object.
5197 	 */
5198 	vm_object_hold_shared(object);
5199 	vm_page_rb_tree_RB_SCAN_NOLK(&object->rb_memq, rb_vm_page_scancmp,
5200 				     pmap_object_init_pt_callback, &info);
5201 	vm_object_drop(object);
5202 #endif
5203 }
5204 
5205 #if 0
5206 
5207 static
5208 int
5209 pmap_object_init_pt_callback(vm_page_t p, void *data)
5210 {
5211 	struct rb_vm_page_scan_info *info = data;
5212 	vm_pindex_t rel_index;
5213 	int hard_busy;
5214 
5215 	/*
5216 	 * don't allow an madvise to blow away our really
5217 	 * free pages allocating pv entries.
5218 	 */
5219 	if ((info->limit & MAP_PREFAULT_MADVISE) &&
5220 		vmstats.v_free_count < vmstats.v_free_reserved) {
5221 		    return(-1);
5222 	}
5223 
5224 	/*
5225 	 * Ignore list markers and ignore pages we cannot instantly
5226 	 * busy (while holding the object token).
5227 	 */
5228 	if (p->flags & PG_MARKER)
5229 		return 0;
5230 	hard_busy = 0;
5231 again:
5232 	if (hard_busy) {
5233 		if (vm_page_busy_try(p, TRUE))
5234 			return 0;
5235 	} else {
5236 		if (vm_page_sbusy_try(p))
5237 			return 0;
5238 	}
5239 	if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
5240 	    (p->flags & PG_FICTITIOUS) == 0) {
5241 		if ((p->queue - p->pc) == PQ_CACHE) {
5242 			if (hard_busy == 0) {
5243 				vm_page_sbusy_drop(p);
5244 				hard_busy = 1;
5245 				goto again;
5246 			}
5247 			vm_page_deactivate(p);
5248 		}
5249 		rel_index = p->pindex - info->start_pindex;
5250 		pmap_enter(info->pmap, info->addr + x86_64_ptob(rel_index), p,
5251 			   VM_PROT_READ, FALSE, info->entry);
5252 	}
5253 	if (hard_busy)
5254 		vm_page_wakeup(p);
5255 	else
5256 		vm_page_sbusy_drop(p);
5257 
5258 	/*
5259 	 * We are using an unlocked scan (that is, the scan expects its
5260 	 * current element to remain in the tree on return).  So we have
5261 	 * to check here and abort the scan if it isn't.
5262 	 */
5263 	if (p->object != info->object)
5264 		return -1;
5265 	lwkt_yield();
5266 	return(0);
5267 }
5268 
5269 #endif
5270 
5271 /*
5272  * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5273  * address.
5274  *
5275  * Returns FALSE if it would be non-trivial or if a pte is already loaded
5276  * into the slot.
5277  *
5278  * The address must reside within a vm_map mapped range to ensure that the
5279  * page table doesn't get ripped out from under us.
5280  *
5281  * XXX This is safe only because page table pages are not freed.
5282  */
5283 int
5284 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
5285 {
5286 	pt_entry_t *pte;
5287 
5288 	/*spin_lock(&pmap->pm_spin);*/
5289 	if ((pte = pmap_pte(pmap, addr)) != NULL) {
5290 		if (*pte & pmap->pmap_bits[PG_V_IDX]) {
5291 			/*spin_unlock(&pmap->pm_spin);*/
5292 			return FALSE;
5293 		}
5294 	}
5295 	/*spin_unlock(&pmap->pm_spin);*/
5296 	return TRUE;
5297 }
5298 
5299 /*
5300  * Change the wiring attribute for a pmap/va pair.  The mapping must already
5301  * exist in the pmap.  The mapping may or may not be managed.  The wiring in
5302  * the page is not changed, the page is returned so the caller can adjust
5303  * its wiring (the page is not locked in any way).
5304  *
5305  * Wiring is not a hardware characteristic so there is no need to invalidate
5306  * TLB.  However, in an SMP environment we must use a locked bus cycle to
5307  * update the pte (if we are not using the pmap_inval_*() API that is)...
5308  * it's ok to do this for simple wiring changes.
5309  */
5310 vm_page_t
5311 pmap_unwire(pmap_t pmap, vm_offset_t va)
5312 {
5313 	pt_entry_t *ptep;
5314 	pv_entry_t pt_pv;
5315 	vm_paddr_t pa;
5316 	vm_page_t m;
5317 
5318 	if (pmap == NULL)
5319 		return NULL;
5320 
5321 	/*
5322 	 * Assume elements in the kernel pmap are stable
5323 	 */
5324 	if (pmap == &kernel_pmap) {
5325 		if (pmap_pt(pmap, va) == 0)
5326 			return NULL;
5327 		ptep = pmap_pte_quick(pmap, va);
5328 		if (pmap_pte_v(pmap, ptep)) {
5329 			if (pmap_pte_w(pmap, ptep))
5330 				atomic_add_long(&pmap->pm_stats.wired_count,-1);
5331 			atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5332 			pa = *ptep & PG_FRAME;
5333 			m = PHYS_TO_VM_PAGE(pa);
5334 		} else {
5335 			m = NULL;
5336 		}
5337 	} else {
5338 		/*
5339 		 * We can only [un]wire pmap-local pages (we cannot wire
5340 		 * shared pages)
5341 		 */
5342 		pt_pv = pv_get(pmap, pmap_pt_pindex(va), NULL);
5343 		if (pt_pv == NULL)
5344 			return NULL;
5345 
5346 		ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
5347 		if ((*ptep & pmap->pmap_bits[PG_V_IDX]) == 0) {
5348 			pv_put(pt_pv);
5349 			return NULL;
5350 		}
5351 
5352 		if (pmap_pte_w(pmap, ptep)) {
5353 			atomic_add_long(&pt_pv->pv_pmap->pm_stats.wired_count,
5354 					-1);
5355 		}
5356 		/* XXX else return NULL so caller doesn't unwire m ? */
5357 
5358 		atomic_clear_long(ptep, pmap->pmap_bits[PG_W_IDX]);
5359 
5360 		pa = *ptep & PG_FRAME;
5361 		m = PHYS_TO_VM_PAGE(pa);	/* held by wired count */
5362 		pv_put(pt_pv);
5363 	}
5364 	return m;
5365 }
5366 
5367 /*
5368  * Copy the range specified by src_addr/len from the source map to
5369  * the range dst_addr/len in the destination map.
5370  *
5371  * This routine is only advisory and need not do anything.
5372  */
5373 void
5374 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
5375 	  vm_size_t len, vm_offset_t src_addr)
5376 {
5377 }
5378 
5379 /*
5380  * pmap_zero_page:
5381  *
5382  *	Zero the specified physical page.
5383  *
5384  *	This function may be called from an interrupt and no locking is
5385  *	required.
5386  */
5387 void
5388 pmap_zero_page(vm_paddr_t phys)
5389 {
5390 	vm_offset_t va = PHYS_TO_DMAP(phys);
5391 
5392 	pagezero((void *)va);
5393 }
5394 
5395 /*
5396  * pmap_zero_page:
5397  *
5398  *	Zero part of a physical page by mapping it into memory and clearing
5399  *	its contents with bzero.
5400  *
5401  *	off and size may not cover an area beyond a single hardware page.
5402  */
5403 void
5404 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
5405 {
5406 	vm_offset_t virt = PHYS_TO_DMAP(phys);
5407 
5408 	bzero((char *)virt + off, size);
5409 }
5410 
5411 /*
5412  * pmap_copy_page:
5413  *
5414  *	Copy the physical page from the source PA to the target PA.
5415  *	This function may be called from an interrupt.  No locking
5416  *	is required.
5417  */
5418 void
5419 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
5420 {
5421 	vm_offset_t src_virt, dst_virt;
5422 
5423 	src_virt = PHYS_TO_DMAP(src);
5424 	dst_virt = PHYS_TO_DMAP(dst);
5425 	bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
5426 }
5427 
5428 /*
5429  * pmap_copy_page_frag:
5430  *
5431  *	Copy the physical page from the source PA to the target PA.
5432  *	This function may be called from an interrupt.  No locking
5433  *	is required.
5434  */
5435 void
5436 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
5437 {
5438 	vm_offset_t src_virt, dst_virt;
5439 
5440 	src_virt = PHYS_TO_DMAP(src);
5441 	dst_virt = PHYS_TO_DMAP(dst);
5442 
5443 	bcopy((char *)src_virt + (src & PAGE_MASK),
5444 	      (char *)dst_virt + (dst & PAGE_MASK),
5445 	      bytes);
5446 }
5447 
5448 /*
5449  * Remove all pages from specified address space this aids process exit
5450  * speeds.  Also, this code may be special cased for the current process
5451  * only.
5452  */
5453 void
5454 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5455 {
5456 	pmap_remove_noinval(pmap, sva, eva);
5457 	cpu_invltlb();
5458 }
5459 
5460 /*
5461  * pmap_testbit tests bits in pte's note that the testbit/clearbit
5462  * routines are inline, and a lot of things compile-time evaluate.
5463  */
5464 static
5465 boolean_t
5466 pmap_testbit(vm_page_t m, int bit)
5467 {
5468 	int res = FALSE;
5469 
5470 	if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5471 		return FALSE;
5472 	/*
5473 	 * Nothing to do if all the mappings are already read-only.
5474 	 * The page's [M]odify bits have already been synchronized
5475 	 * to the vm_page_t and cleaned out.
5476 	 */
5477 	if (bit == PG_M_IDX && m->md.writeable_count == 0)
5478 		return FALSE;
5479 
5480 	/*
5481 	 * Iterate the mapping
5482 	 */
5483 	PMAP_PAGE_BACKING_SCAN(m, NULL, ipmap, iptep, ipte, iva) {
5484 		if (ipte & ipmap->pmap_bits[bit]) {
5485 			res = TRUE;
5486 			break;
5487 		}
5488 	} PMAP_PAGE_BACKING_DONE;
5489 	return res;
5490 }
5491 
5492 /*
5493  * This routine is used to modify bits in ptes.  Only one bit should be
5494  * specified.  PG_RW requires special handling.  This call works with
5495  * any sort of mapped page.  PG_FICTITIOUS pages might not be optimal.
5496  *
5497  * Caller must NOT hold any spin locks
5498  * Caller must hold (m) hard-busied
5499  *
5500  * NOTE: When clearing PG_M we could also (not implemented) drop
5501  *       through to the PG_RW code and clear PG_RW too, forcing
5502  *       a fault on write to redetect PG_M for virtual kernels, but
5503  *       it isn't necessary since virtual kernels invalidate the
5504  *       pte when they clear the VPTE_M bit in their virtual page
5505  *       tables.
5506  *
5507  * NOTE: Does not re-dirty the page when clearing only PG_M.
5508  *
5509  * NOTE: Because we do not lock the pv, *pte can be in a state of
5510  *       flux.  Despite this the value of *pte is still somewhat
5511  *       related while we hold the vm_page spin lock.
5512  *
5513  *       *pte can be zero due to this race.  Since we are clearing
5514  *       bits we basically do no harm when this race occurs.
5515  */
5516 static __inline
5517 void
5518 pmap_clearbit(vm_page_t m, int bit_index)
5519 {
5520 	pt_entry_t npte;
5521 	int retry;
5522 
5523 	/*
5524 	 * Too early in the boot
5525 	 */
5526 	if (!pmap_initialized) {
5527 		if (bit_index == PG_RW_IDX)
5528 			vm_page_flag_clear(m, PG_WRITEABLE);
5529 		return;
5530 	}
5531 
5532 	/*
5533 	 * Being asked to clear other random bits, we don't track them
5534 	 * so we have to iterate.
5535 	 */
5536 	if (bit_index != PG_RW_IDX) {
5537 		PMAP_PAGE_BACKING_SCAN(m, NULL, ipmap, iptep, ipte, iva) {
5538 			if (ipte & ipmap->pmap_bits[bit_index]) {
5539 				atomic_clear_long(iptep,
5540 						  ipmap->pmap_bits[bit_index]);
5541 			}
5542 		} PMAP_PAGE_BACKING_DONE;
5543 		return;
5544 	}
5545 
5546 	/*
5547 	 * Being asked to clear the RW bit.
5548 	 *
5549 	 * Nothing to do if all the mappings are already read-only
5550 	 */
5551 	if (m->md.writeable_count == 0)
5552 		return;
5553 
5554 	/*
5555 	 * Iterate the mappings and check.
5556 	 */
5557 	retry = ticks + hz * 60;
5558 again:
5559 	/*
5560 	 * Clear PG_RW. This also clears PG_M and marks the page dirty if
5561 	 * PG_M was set.
5562 	 *
5563 	 * Since the caller holds the page hard-busied we can safely clear
5564 	 * PG_WRITEABLE, and callers expect us to for the PG_RW_IDX path.
5565 	 */
5566 	PMAP_PAGE_BACKING_SCAN(m, NULL, ipmap, iptep, ipte, iva) {
5567 #if 0
5568 		if ((ipte & ipmap->pmap_bits[PG_MANAGED_IDX]) == 0)
5569 			continue;
5570 #endif
5571 		if ((ipte & ipmap->pmap_bits[PG_RW_IDX]) == 0)
5572 			continue;
5573 		npte = ipte & ~(ipmap->pmap_bits[PG_RW_IDX] |
5574 				ipmap->pmap_bits[PG_M_IDX]);
5575 		if (!pmap_inval_smp_cmpset(ipmap, iva, iptep, ipte, npte))
5576 			PMAP_PAGE_BACKING_RETRY;
5577 		if (ipte & ipmap->pmap_bits[PG_M_IDX])
5578 			vm_page_dirty(m);
5579 
5580 		/*
5581 		 * NOTE: m is not hard-busied so it is not safe to
5582 		 *	 clear PG_WRITEABLE on the 1->0 transition
5583 		 *	 against it being set in pmap_enter().
5584 		 *
5585 		 *	 pmap_count and writeable_count are only applicable
5586 		 *	 to non-fictitious pages (PG_MANAGED_IDX from pte)
5587 		 */
5588 		if (ipte & ipmap->pmap_bits[PG_MANAGED_IDX])
5589 			atomic_add_long(&m->md.writeable_count, -1);
5590 	} PMAP_PAGE_BACKING_DONE;
5591 
5592 	/*
5593 	 * writeable_count should be zero but it is possible to race
5594 	 * a pmap_enter() replacement (see 'oldm').  Once it is zero
5595 	 * it cannot become non-zero because the page is hard-busied.
5596 	 */
5597 	if (m->md.writeable_count != 0) {
5598 		tsleep(&m->md.writeable_count, 0, "pgwab", 1);
5599 		if (retry - ticks > 0)
5600 			goto again;
5601 		panic("pmap_remove_all: cannot return writeable_count "
5602 		      "to 0 (%ld)",
5603 		      m->md.writeable_count);
5604 	}
5605 	vm_page_flag_clear(m, PG_WRITEABLE);
5606 }
5607 
5608 /*
5609  * Lower the permission for all mappings to a given page.
5610  *
5611  * Page must be hard-busied by caller.  Because the page is busied by the
5612  * caller, this should not be able to race a pmap_enter().
5613  */
5614 void
5615 pmap_page_protect(vm_page_t m, vm_prot_t prot)
5616 {
5617 	/* JG NX support? */
5618 	if ((prot & VM_PROT_WRITE) == 0) {
5619 		if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5620 			/*
5621 			 * NOTE: pmap_clearbit(.. PG_RW) also clears
5622 			 *	 the PG_WRITEABLE flag in (m).
5623 			 */
5624 			pmap_clearbit(m, PG_RW_IDX);
5625 		} else {
5626 			pmap_remove_all(m);
5627 		}
5628 	}
5629 }
5630 
5631 vm_paddr_t
5632 pmap_phys_address(vm_pindex_t ppn)
5633 {
5634 	return (x86_64_ptob(ppn));
5635 }
5636 
5637 /*
5638  * Return a count of reference bits for a page, clearing those bits.
5639  * It is not necessary for every reference bit to be cleared, but it
5640  * is necessary that 0 only be returned when there are truly no
5641  * reference bits set.
5642  *
5643  * XXX: The exact number of bits to check and clear is a matter that
5644  * should be tested and standardized at some point in the future for
5645  * optimal aging of shared pages.
5646  *
5647  * This routine may not block.
5648  */
5649 int
5650 pmap_ts_referenced(vm_page_t m)
5651 {
5652 	int rval = 0;
5653 	pt_entry_t npte;
5654 
5655 	if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
5656 		return rval;
5657 	PMAP_PAGE_BACKING_SCAN(m, NULL, ipmap, iptep, ipte, iva) {
5658 		if (ipte & ipmap->pmap_bits[PG_A_IDX]) {
5659 			npte = ipte & ~ipmap->pmap_bits[PG_A_IDX];
5660 			if (!atomic_cmpset_long(iptep, ipte, npte))
5661 				PMAP_PAGE_BACKING_RETRY;
5662 			++rval;
5663 			if (rval > 4)
5664 				break;
5665 		}
5666 	} PMAP_PAGE_BACKING_DONE;
5667 	return rval;
5668 }
5669 
5670 /*
5671  *	pmap_is_modified:
5672  *
5673  *	Return whether or not the specified physical page was modified
5674  *	in any physical maps.
5675  */
5676 boolean_t
5677 pmap_is_modified(vm_page_t m)
5678 {
5679 	boolean_t res;
5680 
5681 	res = pmap_testbit(m, PG_M_IDX);
5682 	return (res);
5683 }
5684 
5685 /*
5686  * Clear the modify bit on the vm_page.
5687  *
5688  * The page must be hard-busied.
5689  */
5690 void
5691 pmap_clear_modify(vm_page_t m)
5692 {
5693 	pmap_clearbit(m, PG_M_IDX);
5694 }
5695 
5696 /*
5697  *	pmap_clear_reference:
5698  *
5699  *	Clear the reference bit on the specified physical page.
5700  */
5701 void
5702 pmap_clear_reference(vm_page_t m)
5703 {
5704 	pmap_clearbit(m, PG_A_IDX);
5705 }
5706 
5707 /*
5708  * Miscellaneous support routines follow
5709  */
5710 
5711 static
5712 void
5713 x86_64_protection_init(void)
5714 {
5715 	uint64_t *kp;
5716 	int prot;
5717 
5718 	/*
5719 	 * NX supported? (boot time loader.conf override only)
5720 	 *
5721 	 * -1	Automatic (sets mode 1)
5722 	 *  0	Disabled
5723 	 *  1	NX implemented, differentiates PROT_READ vs PROT_READ|PROT_EXEC
5724 	 *  2	NX implemented for all cases
5725 	 */
5726 	TUNABLE_INT_FETCH("machdep.pmap_nx_enable", &pmap_nx_enable);
5727 	if ((amd_feature & AMDID_NX) == 0) {
5728 		pmap_bits_default[PG_NX_IDX] = 0;
5729 		pmap_nx_enable = 0;
5730 	} else if (pmap_nx_enable < 0) {
5731 		pmap_nx_enable = 1;		/* default to mode 1 (READ) */
5732 	}
5733 
5734 	/*
5735 	 * 0 is basically read-only access, but also set the NX (no-execute)
5736 	 * bit when VM_PROT_EXECUTE is not specified.
5737 	 */
5738 	kp = protection_codes;
5739 	for (prot = 0; prot < PROTECTION_CODES_SIZE; prot++) {
5740 		switch (prot) {
5741 		case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
5742 			/*
5743 			 * This case handled elsewhere
5744 			 */
5745 			*kp = 0;
5746 			break;
5747 		case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
5748 			/*
5749 			 * Read-only is 0|NX	(pmap_nx_enable mode >= 1)
5750 			 */
5751 			if (pmap_nx_enable >= 1)
5752 				*kp = pmap_bits_default[PG_NX_IDX];
5753 			break;
5754 		case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
5755 		case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
5756 			/*
5757 			 * Execute requires read access
5758 			 */
5759 			*kp = 0;
5760 			break;
5761 		case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
5762 		case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
5763 			/*
5764 			 * Write without execute is RW|NX
5765 			 *			(pmap_nx_enable mode >= 2)
5766 			 */
5767 			*kp = pmap_bits_default[PG_RW_IDX];
5768 			if (pmap_nx_enable >= 2)
5769 				*kp |= pmap_bits_default[PG_NX_IDX];
5770 			break;
5771 		case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
5772 		case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
5773 			/*
5774 			 * Write with execute is RW
5775 			 */
5776 			*kp = pmap_bits_default[PG_RW_IDX];
5777 			break;
5778 		}
5779 		++kp;
5780 	}
5781 }
5782 
5783 /*
5784  * Map a set of physical memory pages into the kernel virtual
5785  * address space. Return a pointer to where it is mapped. This
5786  * routine is intended to be used for mapping device memory,
5787  * NOT real memory.
5788  *
5789  * NOTE: We can't use pgeflag unless we invalidate the pages one at
5790  *	 a time.
5791  *
5792  * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5793  *	 work whether the cpu supports PAT or not.  The remaining PAT
5794  *	 attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5795  *	 supports PAT.
5796  */
5797 void *
5798 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
5799 {
5800 	return(pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5801 }
5802 
5803 void *
5804 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
5805 {
5806 	return(pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE));
5807 }
5808 
5809 void *
5810 pmap_mapbios(vm_paddr_t pa, vm_size_t size)
5811 {
5812 	return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK));
5813 }
5814 
5815 /*
5816  * Map a set of physical memory pages into the kernel virtual
5817  * address space. Return a pointer to where it is mapped. This
5818  * routine is intended to be used for mapping device memory,
5819  * NOT real memory.
5820  */
5821 void *
5822 pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode)
5823 {
5824 	vm_offset_t va, tmpva, offset;
5825 	pt_entry_t *pte;
5826 	vm_size_t tmpsize;
5827 
5828 	offset = pa & PAGE_MASK;
5829 	size = roundup(offset + size, PAGE_SIZE);
5830 
5831 	va = kmem_alloc_nofault(&kernel_map, size, VM_SUBSYS_MAPDEV, PAGE_SIZE);
5832 	if (va == 0)
5833 		panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5834 
5835 	pa = pa & ~PAGE_MASK;
5836 	for (tmpva = va, tmpsize = size; tmpsize > 0;) {
5837 		pte = vtopte(tmpva);
5838 		*pte = pa |
5839 		    kernel_pmap.pmap_bits[PG_RW_IDX] |
5840 		    kernel_pmap.pmap_bits[PG_V_IDX] | /* pgeflag | */
5841 		    kernel_pmap.pmap_cache_bits_pte[mode];
5842 		tmpsize -= PAGE_SIZE;
5843 		tmpva += PAGE_SIZE;
5844 		pa += PAGE_SIZE;
5845 	}
5846 	pmap_invalidate_range(&kernel_pmap, va, va + size);
5847 	pmap_invalidate_cache_range(va, va + size);
5848 
5849 	return ((void *)(va + offset));
5850 }
5851 
5852 void
5853 pmap_unmapdev(vm_offset_t va, vm_size_t size)
5854 {
5855 	vm_offset_t base, offset;
5856 
5857 	base = va & ~PAGE_MASK;
5858 	offset = va & PAGE_MASK;
5859 	size = roundup(offset + size, PAGE_SIZE);
5860 	pmap_qremove(va, size >> PAGE_SHIFT);
5861 	kmem_free(&kernel_map, base, size);
5862 }
5863 
5864 /*
5865  * Sets the memory attribute for the specified page.
5866  */
5867 void
5868 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5869 {
5870 
5871     m->pat_mode = ma;
5872 
5873     /*
5874      * If "m" is a normal page, update its direct mapping.  This update
5875      * can be relied upon to perform any cache operations that are
5876      * required for data coherence.
5877      */
5878     if ((m->flags & PG_FICTITIOUS) == 0)
5879         pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), 1, m->pat_mode);
5880 }
5881 
5882 /*
5883  * Change the PAT attribute on an existing kernel memory map.  Caller
5884  * must ensure that the virtual memory in question is not accessed
5885  * during the adjustment.
5886  *
5887  * If the va is within the DMAP we cannot use vtopte() because the DMAP
5888  * utilizes 2MB or 1GB pages.  2MB is forced atm so calculate the pd_entry
5889  * pointer based on that.
5890  */
5891 void
5892 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
5893 {
5894 	pt_entry_t *pte;
5895 	vm_offset_t base;
5896 	int changed = 0;
5897 
5898 	if (va == 0)
5899 		panic("pmap_change_attr: va is NULL");
5900 	base = trunc_page(va);
5901 
5902 	if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
5903 		pd_entry_t *pd;
5904 
5905 		KKASSERT(va < DMapMaxAddress);
5906 		pd = (pd_entry_t *)PHYS_TO_DMAP(DMPDphys);
5907 		pd += (va - DMAP_MIN_ADDRESS) >> PDRSHIFT;
5908 
5909 		while ((long)count > 0) {
5910 			*pd =
5911 			   (*pd & ~(pd_entry_t)(kernel_pmap.pmap_cache_mask_pde)) |
5912 			   kernel_pmap.pmap_cache_bits_pde[mode];
5913 			count -= NBPDR / PAGE_SIZE;
5914 			va += NBPDR;
5915 			++pd;
5916 		}
5917 	} else {
5918 		while (count) {
5919 			pte = vtopte(va);
5920 			*pte =
5921 			   (*pte & ~(pt_entry_t)(kernel_pmap.pmap_cache_mask_pte)) |
5922 			   kernel_pmap.pmap_cache_bits_pte[mode];
5923 			--count;
5924 			va += PAGE_SIZE;
5925 		}
5926 	}
5927 
5928 	changed = 1;	/* XXX: not optimal */
5929 
5930 	/*
5931 	 * Flush CPU caches if required to make sure any data isn't cached that
5932 	 * shouldn't be, etc.
5933 	 */
5934 	if (changed) {
5935 		pmap_invalidate_range(&kernel_pmap, base, va);
5936 		pmap_invalidate_cache_range(base, va);
5937 	}
5938 }
5939 
5940 /*
5941  * perform the pmap work for mincore
5942  */
5943 int
5944 pmap_mincore(pmap_t pmap, vm_offset_t addr)
5945 {
5946 	pt_entry_t *ptep, pte;
5947 	vm_page_t m;
5948 	int val = 0;
5949 
5950 	ptep = pmap_pte(pmap, addr);
5951 
5952 	if (ptep && (pte = *ptep) != 0) {
5953 		vm_offset_t pa;
5954 
5955 		val = MINCORE_INCORE;
5956 		pa = pte & PG_FRAME;
5957 		if (pte & pmap->pmap_bits[PG_MANAGED_IDX])
5958 			m = PHYS_TO_VM_PAGE(pa);
5959 		else
5960 			m = NULL;
5961 
5962 		/*
5963 		 * Modified by us
5964 		 */
5965 		if (pte & pmap->pmap_bits[PG_M_IDX])
5966 			val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
5967 
5968 		/*
5969 		 * Modified by someone
5970 		 */
5971 		else if (m && (m->dirty || pmap_is_modified(m)))
5972 			val |= MINCORE_MODIFIED_OTHER;
5973 
5974 		/*
5975 		 * Referenced by us, or someone else.
5976 		 */
5977 		if (pte & pmap->pmap_bits[PG_A_IDX]) {
5978 			val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
5979 		} else if (m && ((m->flags & PG_REFERENCED) ||
5980 				 pmap_ts_referenced(m))) {
5981 			val |= MINCORE_REFERENCED_OTHER;
5982 			vm_page_flag_set(m, PG_REFERENCED);
5983 		}
5984 	}
5985 	return val;
5986 }
5987 
5988 /*
5989  * Replace p->p_vmspace with a new one.  If adjrefs is non-zero the new
5990  * vmspace will be ref'd and the old one will be deref'd.
5991  *
5992  * The vmspace for all lwps associated with the process will be adjusted
5993  * and cr3 will be reloaded if any lwp is the current lwp.
5994  *
5995  * The process must hold the vmspace->vm_map.token for oldvm and newvm
5996  */
5997 void
5998 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
5999 {
6000 	struct vmspace *oldvm;
6001 	struct lwp *lp;
6002 
6003 	oldvm = p->p_vmspace;
6004 	if (oldvm != newvm) {
6005 		if (adjrefs)
6006 			vmspace_ref(newvm);
6007 		p->p_vmspace = newvm;
6008 		KKASSERT(p->p_nthreads == 1);
6009 		lp = RB_ROOT(&p->p_lwp_tree);
6010 		pmap_setlwpvm(lp, newvm);
6011 		if (adjrefs)
6012 			vmspace_rel(oldvm);
6013 	}
6014 }
6015 
6016 /*
6017  * Set the vmspace for a LWP.  The vmspace is almost universally set the
6018  * same as the process vmspace, but virtual kernels need to swap out contexts
6019  * on a per-lwp basis.
6020  *
6021  * Caller does not necessarily hold any vmspace tokens.  Caller must control
6022  * the lwp (typically be in the context of the lwp).  We use a critical
6023  * section to protect against statclock and hardclock (statistics collection).
6024  */
6025 void
6026 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
6027 {
6028 	struct vmspace *oldvm;
6029 	struct pmap *pmap;
6030 	thread_t td;
6031 
6032 	oldvm = lp->lwp_vmspace;
6033 
6034 	if (oldvm != newvm) {
6035 		crit_enter();
6036 		td = curthread;
6037 		KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
6038 		lp->lwp_vmspace = newvm;
6039 		if (td->td_lwp == lp) {
6040 			pmap = vmspace_pmap(newvm);
6041 			ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
6042 			if (pmap->pm_active_lock & CPULOCK_EXCL)
6043 				pmap_interlock_wait(newvm);
6044 #if defined(SWTCH_OPTIM_STATS)
6045 			tlb_flush_count++;
6046 #endif
6047 			if (pmap->pmap_bits[TYPE_IDX] == REGULAR_PMAP) {
6048 				td->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
6049 				if (meltdown_mitigation && pmap->pm_pmlpv_iso) {
6050 					td->td_pcb->pcb_cr3_iso =
6051 						vtophys(pmap->pm_pml4_iso);
6052 					td->td_pcb->pcb_flags |= PCB_ISOMMU;
6053 				} else {
6054 					td->td_pcb->pcb_cr3_iso = 0;
6055 					td->td_pcb->pcb_flags &= ~PCB_ISOMMU;
6056 				}
6057 			} else if (pmap->pmap_bits[TYPE_IDX] == EPT_PMAP) {
6058 				td->td_pcb->pcb_cr3 = KPML4phys;
6059 				td->td_pcb->pcb_cr3_iso = 0;
6060 				td->td_pcb->pcb_flags &= ~PCB_ISOMMU;
6061 			} else {
6062 				panic("pmap_setlwpvm: unknown pmap type\n");
6063 			}
6064 
6065 			/*
6066 			 * The MMU separation fields needs to be updated.
6067 			 * (it can't access the pcb directly from the
6068 			 * restricted user pmap).
6069 			 */
6070 			{
6071 				struct trampframe *tramp;
6072 
6073 				tramp = &pscpu->trampoline;
6074 				tramp->tr_pcb_cr3 = td->td_pcb->pcb_cr3;
6075 				tramp->tr_pcb_cr3_iso = td->td_pcb->pcb_cr3_iso;
6076 				tramp->tr_pcb_flags = td->td_pcb->pcb_flags;
6077 				tramp->tr_pcb_rsp = (register_t)td->td_pcb;
6078 				/* tr_pcb_rsp doesn't change */
6079 			}
6080 
6081 			/*
6082 			 * In kernel-land we always use the normal PML4E
6083 			 * so the kernel is fully mapped and can also access
6084 			 * user memory.
6085 			 */
6086 			load_cr3(td->td_pcb->pcb_cr3);
6087 			pmap = vmspace_pmap(oldvm);
6088 			ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
6089 					       mycpu->gd_cpuid);
6090 		}
6091 		crit_exit();
6092 	}
6093 }
6094 
6095 /*
6096  * Called when switching to a locked pmap, used to interlock against pmaps
6097  * undergoing modifications to prevent us from activating the MMU for the
6098  * target pmap until all such modifications have completed.  We have to do
6099  * this because the thread making the modifications has already set up its
6100  * SMP synchronization mask.
6101  *
6102  * This function cannot sleep!
6103  *
6104  * No requirements.
6105  */
6106 void
6107 pmap_interlock_wait(struct vmspace *vm)
6108 {
6109 	struct pmap *pmap = &vm->vm_pmap;
6110 
6111 	if (pmap->pm_active_lock & CPULOCK_EXCL) {
6112 		crit_enter();
6113 		KKASSERT(curthread->td_critcount >= 2);
6114 		DEBUG_PUSH_INFO("pmap_interlock_wait");
6115 		while (pmap->pm_active_lock & CPULOCK_EXCL) {
6116 			cpu_ccfence();
6117 			lwkt_process_ipiq();
6118 		}
6119 		DEBUG_POP_INFO();
6120 		crit_exit();
6121 	}
6122 }
6123 
6124 vm_offset_t
6125 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
6126 {
6127 
6128 	if ((obj == NULL) || (size < NBPDR) ||
6129 	    ((obj->type != OBJT_DEVICE) && (obj->type != OBJT_MGTDEVICE))) {
6130 		return addr;
6131 	}
6132 
6133 	addr = roundup2(addr, NBPDR);
6134 	return addr;
6135 }
6136 
6137 /*
6138  * Used by kmalloc/kfree, page already exists at va
6139  */
6140 vm_page_t
6141 pmap_kvtom(vm_offset_t va)
6142 {
6143 	pt_entry_t *ptep = vtopte(va);
6144 
6145 	return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
6146 }
6147 
6148 /*
6149  * Initialize machine-specific shared page directory support.  This
6150  * is executed when a VM object is created.
6151  */
6152 void
6153 pmap_object_init(vm_object_t object)
6154 {
6155 }
6156 
6157 /*
6158  * Clean up machine-specific shared page directory support.  This
6159  * is executed when a VM object is destroyed.
6160  */
6161 void
6162 pmap_object_free(vm_object_t object)
6163 {
6164 }
6165 
6166 /*
6167  * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6168  * VM page and issue a pginfo->callback.
6169  */
6170 static
6171 void
6172 pmap_pgscan_callback(pmap_t pmap, struct pmap_scan_info *info,
6173 		      vm_pindex_t *pte_placemark,
6174 		      pv_entry_t pt_pv, vm_offset_t va,
6175 		      pt_entry_t *ptep, void *arg)
6176 {
6177 	struct pmap_pgscan_info *pginfo = arg;
6178 	vm_page_t m;
6179 	pt_entry_t pte;
6180 
6181 	pte = *ptep;
6182 	cpu_ccfence();
6183 
6184 	if (pte & pmap->pmap_bits[PG_MANAGED_IDX]) {
6185 		/*
6186 		 * Try to busy the page while we hold the pte_placemark locked.
6187 		 */
6188 		m = PHYS_TO_VM_PAGE(*ptep & PG_FRAME);
6189 		if (vm_page_busy_try(m, TRUE) == 0) {
6190 			if (m == PHYS_TO_VM_PAGE(*ptep & PG_FRAME)) {
6191 				/*
6192 				 * The callback is issued with the pt_pv
6193 				 * unlocked.
6194 				 */
6195 				pv_placemarker_wakeup(pmap, pte_placemark);
6196 				if (pt_pv) {
6197 					vm_page_wire_quick(pt_pv->pv_m);
6198 					pv_unlock(pt_pv);
6199 				}
6200 				if (pginfo->callback(pginfo, va, m) < 0)
6201 					info->stop = 1;
6202 				if (pt_pv) {
6203 					pv_lock(pt_pv);
6204 					if (vm_page_unwire_quick(pt_pv->pv_m)) {
6205 						panic("pmap_pgscan: bad wire_"
6206 						      "count on pt_pv");
6207 					}
6208 				}
6209 			} else {
6210 				vm_page_wakeup(m);
6211 				pv_placemarker_wakeup(pmap, pte_placemark);
6212 			}
6213 		} else {
6214 			++pginfo->busycount;
6215 			pv_placemarker_wakeup(pmap, pte_placemark);
6216 		}
6217 	} else {
6218 		/*
6219 		 * Shared page table or unmanaged page (sharept or !sharept)
6220 		 */
6221 		pv_placemarker_wakeup(pmap, pte_placemark);
6222 	}
6223 }
6224 
6225 void
6226 pmap_pgscan(struct pmap_pgscan_info *pginfo)
6227 {
6228 	struct pmap_scan_info info;
6229 
6230 	pginfo->offset = pginfo->beg_addr;
6231 	info.pmap = pginfo->pmap;
6232 	info.sva = pginfo->beg_addr;
6233 	info.eva = pginfo->end_addr;
6234 	info.func = pmap_pgscan_callback;
6235 	info.arg = pginfo;
6236 	pmap_scan(&info, 0);
6237 	if (info.stop == 0)
6238 		pginfo->offset = pginfo->end_addr;
6239 }
6240 
6241 /*
6242  * Wait for a placemarker that we do not own to clear.  The placemarker
6243  * in question is not necessarily set to the pindex we want, we may have
6244  * to wait on the element because we want to reserve it ourselves.
6245  *
6246  * NOTE: PM_PLACEMARK_WAKEUP sets a bit which is already set in
6247  *	 PM_NOPLACEMARK, so it does not interfere with placemarks
6248  *	 which have already been woken up.
6249  *
6250  * NOTE: This routine is called without the pmap spin-lock and so can
6251  *	 race changes to *pmark.  Due to the sensitivity of the routine
6252  *	 to possible MULTIPLE interactions from other cpus, and the
6253  *	 overloading of the WAKEUP bit on PM_NOPLACEMARK, we have to
6254  *	 use a cmpset loop to avoid a race that might cause the WAKEUP
6255  *	 bit to be lost.
6256  *
6257  * Caller is expected to retry its operation upon return.
6258  */
6259 static
6260 void
6261 pv_placemarker_wait(pmap_t pmap, vm_pindex_t *pmark)
6262 {
6263 	vm_pindex_t mark;
6264 
6265 	mark = *pmark;
6266 	cpu_ccfence();
6267 	while (mark != PM_NOPLACEMARK) {
6268 		tsleep_interlock(pmark, 0);
6269 		if (atomic_fcmpset_long(pmark, &mark,
6270 				       mark | PM_PLACEMARK_WAKEUP)) {
6271 			tsleep(pmark, PINTERLOCKED, "pvplw", 0);
6272 			break;
6273 		}
6274 	}
6275 }
6276 
6277 /*
6278  * Wakeup a placemarker that we own.  Replace the entry with
6279  * PM_NOPLACEMARK and issue a wakeup() if necessary.
6280  */
6281 static
6282 void
6283 pv_placemarker_wakeup(pmap_t pmap, vm_pindex_t *pmark)
6284 {
6285 	vm_pindex_t pindex;
6286 
6287 	pindex = atomic_swap_long(pmark, PM_NOPLACEMARK);
6288 	KKASSERT(pindex != PM_NOPLACEMARK);
6289 	if (pindex & PM_PLACEMARK_WAKEUP)
6290 		wakeup(pmark);
6291 }
6292