xref: /freebsd/sys/powerpc/aim/mmu_radix.c (revision 79ac3c12a714bcd3f2354c52d948aed9575c46d6)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2018 Matthew Macy
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  *
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include <sys/param.h>
32 #include <sys/kernel.h>
33 #include <sys/systm.h>
34 #include <sys/conf.h>
35 #include <sys/bitstring.h>
36 #include <sys/queue.h>
37 #include <sys/cpuset.h>
38 #include <sys/endian.h>
39 #include <sys/kerneldump.h>
40 #include <sys/ktr.h>
41 #include <sys/lock.h>
42 #include <sys/syslog.h>
43 #include <sys/msgbuf.h>
44 #include <sys/malloc.h>
45 #include <sys/mman.h>
46 #include <sys/mutex.h>
47 #include <sys/proc.h>
48 #include <sys/rwlock.h>
49 #include <sys/sched.h>
50 #include <sys/sysctl.h>
51 #include <sys/systm.h>
52 #include <sys/vmem.h>
53 #include <sys/vmmeter.h>
54 #include <sys/smp.h>
55 
56 #include <sys/kdb.h>
57 
58 #include <dev/ofw/openfirm.h>
59 
60 #include <vm/vm.h>
61 #include <vm/pmap.h>
62 #include <vm/vm_param.h>
63 #include <vm/vm_kern.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_map.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_extern.h>
68 #include <vm/vm_pageout.h>
69 #include <vm/vm_phys.h>
70 #include <vm/vm_reserv.h>
71 #include <vm/vm_dumpset.h>
72 #include <vm/uma.h>
73 
74 #include <machine/_inttypes.h>
75 #include <machine/cpu.h>
76 #include <machine/platform.h>
77 #include <machine/frame.h>
78 #include <machine/md_var.h>
79 #include <machine/psl.h>
80 #include <machine/bat.h>
81 #include <machine/hid.h>
82 #include <machine/pte.h>
83 #include <machine/sr.h>
84 #include <machine/trap.h>
85 #include <machine/mmuvar.h>
86 
87 #ifdef INVARIANTS
88 #include <vm/uma_dbg.h>
89 #endif
90 
91 #define PPC_BITLSHIFT(bit)	(sizeof(long)*NBBY - 1 - (bit))
92 #define PPC_BIT(bit)		(1UL << PPC_BITLSHIFT(bit))
93 #define PPC_BITLSHIFT_VAL(val, bit) ((val) << PPC_BITLSHIFT(bit))
94 
95 #include "opt_ddb.h"
96 #ifdef DDB
97 static void pmap_pte_walk(pml1_entry_t *l1, vm_offset_t va);
98 #endif
99 
100 #define PG_W	RPTE_WIRED
101 #define PG_V	RPTE_VALID
102 #define PG_MANAGED	RPTE_MANAGED
103 #define PG_PROMOTED	RPTE_PROMOTED
104 #define PG_M	RPTE_C
105 #define PG_A	RPTE_R
106 #define PG_X	RPTE_EAA_X
107 #define PG_RW	RPTE_EAA_W
108 #define PG_PTE_CACHE RPTE_ATTR_MASK
109 
110 #define RPTE_SHIFT 9
111 #define NLS_MASK ((1UL<<5)-1)
112 #define RPTE_ENTRIES (1UL<<RPTE_SHIFT)
113 #define RPTE_MASK (RPTE_ENTRIES-1)
114 
115 #define NLB_SHIFT 0
116 #define NLB_MASK (((1UL<<52)-1) << 8)
117 
118 extern int nkpt;
119 extern caddr_t crashdumpmap;
120 
121 #define RIC_FLUSH_TLB 0
122 #define RIC_FLUSH_PWC 1
123 #define RIC_FLUSH_ALL 2
124 
125 #define POWER9_TLB_SETS_RADIX	128	/* # sets in POWER9 TLB Radix mode */
126 
127 #define PPC_INST_TLBIE			0x7c000264
128 #define PPC_INST_TLBIEL			0x7c000224
129 #define PPC_INST_SLBIA			0x7c0003e4
130 
131 #define ___PPC_RA(a)	(((a) & 0x1f) << 16)
132 #define ___PPC_RB(b)	(((b) & 0x1f) << 11)
133 #define ___PPC_RS(s)	(((s) & 0x1f) << 21)
134 #define ___PPC_RT(t)	___PPC_RS(t)
135 #define ___PPC_R(r)	(((r) & 0x1) << 16)
136 #define ___PPC_PRS(prs)	(((prs) & 0x1) << 17)
137 #define ___PPC_RIC(ric)	(((ric) & 0x3) << 18)
138 
139 #define PPC_SLBIA(IH)	__XSTRING(.long PPC_INST_SLBIA | \
140 				       ((IH & 0x7) << 21))
141 #define	PPC_TLBIE_5(rb,rs,ric,prs,r)				\
142 	__XSTRING(.long PPC_INST_TLBIE |			\
143 			  ___PPC_RB(rb) | ___PPC_RS(rs) |	\
144 			  ___PPC_RIC(ric) | ___PPC_PRS(prs) |	\
145 			  ___PPC_R(r))
146 
147 #define	PPC_TLBIEL(rb,rs,ric,prs,r) \
148 	 __XSTRING(.long PPC_INST_TLBIEL | \
149 			   ___PPC_RB(rb) | ___PPC_RS(rs) |	\
150 			   ___PPC_RIC(ric) | ___PPC_PRS(prs) |	\
151 			   ___PPC_R(r))
152 
153 #define PPC_INVALIDATE_ERAT		PPC_SLBIA(7)
154 
155 static __inline void
156 ttusync(void)
157 {
158 	__asm __volatile("eieio; tlbsync; ptesync" ::: "memory");
159 }
160 
161 #define TLBIEL_INVAL_SEL_MASK	0xc00	/* invalidation selector */
162 #define  TLBIEL_INVAL_PAGE	0x000	/* invalidate a single page */
163 #define  TLBIEL_INVAL_SET_PID	0x400	/* invalidate a set for the current PID */
164 #define  TLBIEL_INVAL_SET_LPID	0x800	/* invalidate a set for current LPID */
165 #define  TLBIEL_INVAL_SET	0xc00	/* invalidate a set for all LPIDs */
166 
167 #define TLBIE_ACTUAL_PAGE_MASK		0xe0
168 #define  TLBIE_ACTUAL_PAGE_4K		0x00
169 #define  TLBIE_ACTUAL_PAGE_64K		0xa0
170 #define  TLBIE_ACTUAL_PAGE_2M		0x20
171 #define  TLBIE_ACTUAL_PAGE_1G		0x40
172 
173 #define TLBIE_PRS_PARTITION_SCOPE	0x0
174 #define TLBIE_PRS_PROCESS_SCOPE	0x1
175 
176 #define TLBIE_RIC_INVALIDATE_TLB	0x0	/* Invalidate just TLB */
177 #define TLBIE_RIC_INVALIDATE_PWC	0x1	/* Invalidate just PWC */
178 #define TLBIE_RIC_INVALIDATE_ALL	0x2	/* Invalidate TLB, PWC,
179 						 * cached {proc, part}tab entries
180 						 */
181 #define TLBIE_RIC_INVALIDATE_SEQ	0x3	/* HPT - only:
182 						 * Invalidate a range of translations
183 						 */
184 
185 static __always_inline void
186 radix_tlbie(uint8_t ric, uint8_t prs, uint16_t is, uint32_t pid, uint32_t lpid,
187 			vm_offset_t va, uint16_t ap)
188 {
189 	uint64_t rb, rs;
190 
191 	MPASS((va & PAGE_MASK) == 0);
192 
193 	rs = ((uint64_t)pid << 32) | lpid;
194 	rb = va | is | ap;
195 	__asm __volatile(PPC_TLBIE_5(%0, %1, %2, %3, 1) : :
196 		"r" (rb), "r" (rs), "i" (ric), "i" (prs) : "memory");
197 }
198 
199 static __inline void
200 radix_tlbie_fixup(uint32_t pid, vm_offset_t va, int ap)
201 {
202 
203 	__asm __volatile("ptesync" ::: "memory");
204 	radix_tlbie(TLBIE_RIC_INVALIDATE_TLB, TLBIE_PRS_PROCESS_SCOPE,
205 	    TLBIEL_INVAL_PAGE, 0, 0, va, ap);
206 	__asm __volatile("ptesync" ::: "memory");
207 	radix_tlbie(TLBIE_RIC_INVALIDATE_TLB, TLBIE_PRS_PROCESS_SCOPE,
208 	    TLBIEL_INVAL_PAGE, pid, 0, va, ap);
209 }
210 
211 static __inline void
212 radix_tlbie_invlpg_user_4k(uint32_t pid, vm_offset_t va)
213 {
214 
215 	radix_tlbie(TLBIE_RIC_INVALIDATE_TLB, TLBIE_PRS_PROCESS_SCOPE,
216 		TLBIEL_INVAL_PAGE, pid, 0, va, TLBIE_ACTUAL_PAGE_4K);
217 	radix_tlbie_fixup(pid, va, TLBIE_ACTUAL_PAGE_4K);
218 }
219 
220 static __inline void
221 radix_tlbie_invlpg_user_2m(uint32_t pid, vm_offset_t va)
222 {
223 
224 	radix_tlbie(TLBIE_RIC_INVALIDATE_TLB, TLBIE_PRS_PROCESS_SCOPE,
225 		TLBIEL_INVAL_PAGE, pid, 0, va, TLBIE_ACTUAL_PAGE_2M);
226 	radix_tlbie_fixup(pid, va, TLBIE_ACTUAL_PAGE_2M);
227 }
228 
229 static __inline void
230 radix_tlbie_invlpwc_user(uint32_t pid)
231 {
232 
233 	radix_tlbie(TLBIE_RIC_INVALIDATE_PWC, TLBIE_PRS_PROCESS_SCOPE,
234 		TLBIEL_INVAL_SET_PID, pid, 0, 0, 0);
235 }
236 
237 static __inline void
238 radix_tlbie_flush_user(uint32_t pid)
239 {
240 
241 	radix_tlbie(TLBIE_RIC_INVALIDATE_ALL, TLBIE_PRS_PROCESS_SCOPE,
242 		TLBIEL_INVAL_SET_PID, pid, 0, 0, 0);
243 }
244 
245 static __inline void
246 radix_tlbie_invlpg_kernel_4k(vm_offset_t va)
247 {
248 
249 	radix_tlbie(TLBIE_RIC_INVALIDATE_TLB, TLBIE_PRS_PROCESS_SCOPE,
250 	    TLBIEL_INVAL_PAGE, 0, 0, va, TLBIE_ACTUAL_PAGE_4K);
251 	radix_tlbie_fixup(0, va, TLBIE_ACTUAL_PAGE_4K);
252 }
253 
254 static __inline void
255 radix_tlbie_invlpg_kernel_2m(vm_offset_t va)
256 {
257 
258 	radix_tlbie(TLBIE_RIC_INVALIDATE_TLB, TLBIE_PRS_PROCESS_SCOPE,
259 	    TLBIEL_INVAL_PAGE, 0, 0, va, TLBIE_ACTUAL_PAGE_2M);
260 	radix_tlbie_fixup(0, va, TLBIE_ACTUAL_PAGE_2M);
261 }
262 
263 /* 1GB pages aren't currently supported. */
264 static __inline __unused void
265 radix_tlbie_invlpg_kernel_1g(vm_offset_t va)
266 {
267 
268 	radix_tlbie(TLBIE_RIC_INVALIDATE_TLB, TLBIE_PRS_PROCESS_SCOPE,
269 	    TLBIEL_INVAL_PAGE, 0, 0, va, TLBIE_ACTUAL_PAGE_1G);
270 	radix_tlbie_fixup(0, va, TLBIE_ACTUAL_PAGE_1G);
271 }
272 
273 static __inline void
274 radix_tlbie_invlpwc_kernel(void)
275 {
276 
277 	radix_tlbie(TLBIE_RIC_INVALIDATE_PWC, TLBIE_PRS_PROCESS_SCOPE,
278 	    TLBIEL_INVAL_SET_LPID, 0, 0, 0, 0);
279 }
280 
281 static __inline void
282 radix_tlbie_flush_kernel(void)
283 {
284 
285 	radix_tlbie(TLBIE_RIC_INVALIDATE_ALL, TLBIE_PRS_PROCESS_SCOPE,
286 	    TLBIEL_INVAL_SET_LPID, 0, 0, 0, 0);
287 }
288 
289 static __inline vm_pindex_t
290 pmap_l3e_pindex(vm_offset_t va)
291 {
292 	return ((va & PG_FRAME) >> L3_PAGE_SIZE_SHIFT);
293 }
294 
295 static __inline vm_pindex_t
296 pmap_pml3e_index(vm_offset_t va)
297 {
298 
299 	return ((va >> L3_PAGE_SIZE_SHIFT) & RPTE_MASK);
300 }
301 
302 static __inline vm_pindex_t
303 pmap_pml2e_index(vm_offset_t va)
304 {
305 	return ((va >> L2_PAGE_SIZE_SHIFT) & RPTE_MASK);
306 }
307 
308 static __inline vm_pindex_t
309 pmap_pml1e_index(vm_offset_t va)
310 {
311 	return ((va & PG_FRAME) >> L1_PAGE_SIZE_SHIFT);
312 }
313 
314 /* Return various clipped indexes for a given VA */
315 static __inline vm_pindex_t
316 pmap_pte_index(vm_offset_t va)
317 {
318 
319 	return ((va >> PAGE_SHIFT) & RPTE_MASK);
320 }
321 
322 /* Return a pointer to the PT slot that corresponds to a VA */
323 static __inline pt_entry_t *
324 pmap_l3e_to_pte(pt_entry_t *l3e, vm_offset_t va)
325 {
326 	pt_entry_t *pte;
327 	vm_paddr_t ptepa;
328 
329 	ptepa = (be64toh(*l3e) & NLB_MASK);
330 	pte = (pt_entry_t *)PHYS_TO_DMAP(ptepa);
331 	return (&pte[pmap_pte_index(va)]);
332 }
333 
334 /* Return a pointer to the PD slot that corresponds to a VA */
335 static __inline pt_entry_t *
336 pmap_l2e_to_l3e(pt_entry_t *l2e, vm_offset_t va)
337 {
338 	pt_entry_t *l3e;
339 	vm_paddr_t l3pa;
340 
341 	l3pa = (be64toh(*l2e) & NLB_MASK);
342 	l3e = (pml3_entry_t *)PHYS_TO_DMAP(l3pa);
343 	return (&l3e[pmap_pml3e_index(va)]);
344 }
345 
346 /* Return a pointer to the PD slot that corresponds to a VA */
347 static __inline pt_entry_t *
348 pmap_l1e_to_l2e(pt_entry_t *l1e, vm_offset_t va)
349 {
350 	pt_entry_t *l2e;
351 	vm_paddr_t l2pa;
352 
353 	l2pa = (be64toh(*l1e) & NLB_MASK);
354 
355 	l2e = (pml2_entry_t *)PHYS_TO_DMAP(l2pa);
356 	return (&l2e[pmap_pml2e_index(va)]);
357 }
358 
359 static __inline pml1_entry_t *
360 pmap_pml1e(pmap_t pmap, vm_offset_t va)
361 {
362 
363 	return (&pmap->pm_pml1[pmap_pml1e_index(va)]);
364 }
365 
366 static pt_entry_t *
367 pmap_pml2e(pmap_t pmap, vm_offset_t va)
368 {
369 	pt_entry_t *l1e;
370 
371 	l1e = pmap_pml1e(pmap, va);
372 	if (l1e == NULL || (be64toh(*l1e) & RPTE_VALID) == 0)
373 		return (NULL);
374 	return (pmap_l1e_to_l2e(l1e, va));
375 }
376 
377 static __inline pt_entry_t *
378 pmap_pml3e(pmap_t pmap, vm_offset_t va)
379 {
380 	pt_entry_t *l2e;
381 
382 	l2e = pmap_pml2e(pmap, va);
383 	if (l2e == NULL || (be64toh(*l2e) & RPTE_VALID) == 0)
384 		return (NULL);
385 	return (pmap_l2e_to_l3e(l2e, va));
386 }
387 
388 static __inline pt_entry_t *
389 pmap_pte(pmap_t pmap, vm_offset_t va)
390 {
391 	pt_entry_t *l3e;
392 
393 	l3e = pmap_pml3e(pmap, va);
394 	if (l3e == NULL || (be64toh(*l3e) & RPTE_VALID) == 0)
395 		return (NULL);
396 	return (pmap_l3e_to_pte(l3e, va));
397 }
398 
399 int nkpt = 64;
400 SYSCTL_INT(_machdep, OID_AUTO, nkpt, CTLFLAG_RD, &nkpt, 0,
401     "Number of kernel page table pages allocated on bootup");
402 
403 vm_paddr_t dmaplimit;
404 
405 SYSCTL_DECL(_vm_pmap);
406 
407 #ifdef INVARIANTS
408 #define VERBOSE_PMAP 0
409 #define VERBOSE_PROTECT 0
410 static int pmap_logging;
411 SYSCTL_INT(_vm_pmap, OID_AUTO, pmap_logging, CTLFLAG_RWTUN,
412     &pmap_logging, 0, "verbose debug logging");
413 #endif
414 
415 static u_int64_t	KPTphys;	/* phys addr of kernel level 1 */
416 
417 //static vm_paddr_t	KERNend;	/* phys addr of end of bootstrap data */
418 
419 static vm_offset_t qframe = 0;
420 static struct mtx qframe_mtx;
421 
422 void mmu_radix_activate(struct thread *);
423 void mmu_radix_advise(pmap_t, vm_offset_t, vm_offset_t, int);
424 void mmu_radix_align_superpage(vm_object_t, vm_ooffset_t, vm_offset_t *,
425     vm_size_t);
426 void mmu_radix_clear_modify(vm_page_t);
427 void mmu_radix_copy(pmap_t, pmap_t, vm_offset_t, vm_size_t, vm_offset_t);
428 int mmu_radix_decode_kernel_ptr(vm_offset_t, int *, vm_offset_t *);
429 int mmu_radix_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int, int8_t);
430 void mmu_radix_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
431 	vm_prot_t);
432 void mmu_radix_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
433 vm_paddr_t mmu_radix_extract(pmap_t pmap, vm_offset_t va);
434 vm_page_t mmu_radix_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t);
435 void mmu_radix_kenter(vm_offset_t, vm_paddr_t);
436 vm_paddr_t mmu_radix_kextract(vm_offset_t);
437 void mmu_radix_kremove(vm_offset_t);
438 boolean_t mmu_radix_is_modified(vm_page_t);
439 boolean_t mmu_radix_is_prefaultable(pmap_t, vm_offset_t);
440 boolean_t mmu_radix_is_referenced(vm_page_t);
441 void mmu_radix_object_init_pt(pmap_t, vm_offset_t, vm_object_t,
442 	vm_pindex_t, vm_size_t);
443 boolean_t mmu_radix_page_exists_quick(pmap_t, vm_page_t);
444 void mmu_radix_page_init(vm_page_t);
445 boolean_t mmu_radix_page_is_mapped(vm_page_t m);
446 void mmu_radix_page_set_memattr(vm_page_t, vm_memattr_t);
447 int mmu_radix_page_wired_mappings(vm_page_t);
448 int mmu_radix_pinit(pmap_t);
449 void mmu_radix_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
450 bool mmu_radix_ps_enabled(pmap_t);
451 void mmu_radix_qenter(vm_offset_t, vm_page_t *, int);
452 void mmu_radix_qremove(vm_offset_t, int);
453 vm_offset_t mmu_radix_quick_enter_page(vm_page_t);
454 void mmu_radix_quick_remove_page(vm_offset_t);
455 boolean_t mmu_radix_ts_referenced(vm_page_t);
456 void mmu_radix_release(pmap_t);
457 void mmu_radix_remove(pmap_t, vm_offset_t, vm_offset_t);
458 void mmu_radix_remove_all(vm_page_t);
459 void mmu_radix_remove_pages(pmap_t);
460 void mmu_radix_remove_write(vm_page_t);
461 void mmu_radix_unwire(pmap_t, vm_offset_t, vm_offset_t);
462 void mmu_radix_zero_page(vm_page_t);
463 void mmu_radix_zero_page_area(vm_page_t, int, int);
464 int mmu_radix_change_attr(vm_offset_t, vm_size_t, vm_memattr_t);
465 void mmu_radix_page_array_startup(long pages);
466 
467 #include "mmu_oea64.h"
468 
469 /*
470  * Kernel MMU interface
471  */
472 
473 static void	mmu_radix_bootstrap(vm_offset_t, vm_offset_t);
474 
475 static void mmu_radix_copy_page(vm_page_t, vm_page_t);
476 static void mmu_radix_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
477     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
478 static void mmu_radix_growkernel(vm_offset_t);
479 static void mmu_radix_init(void);
480 static int mmu_radix_mincore(pmap_t, vm_offset_t, vm_paddr_t *);
481 static vm_offset_t mmu_radix_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
482 static void mmu_radix_pinit0(pmap_t);
483 
484 static void *mmu_radix_mapdev(vm_paddr_t, vm_size_t);
485 static void *mmu_radix_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t);
486 static void mmu_radix_unmapdev(vm_offset_t, vm_size_t);
487 static void mmu_radix_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t ma);
488 static boolean_t mmu_radix_dev_direct_mapped(vm_paddr_t, vm_size_t);
489 static void mmu_radix_dumpsys_map(vm_paddr_t pa, size_t sz, void **va);
490 static void mmu_radix_scan_init(void);
491 static void	mmu_radix_cpu_bootstrap(int ap);
492 static void	mmu_radix_tlbie_all(void);
493 
494 static struct pmap_funcs mmu_radix_methods = {
495 	.bootstrap = mmu_radix_bootstrap,
496 	.copy_page = mmu_radix_copy_page,
497 	.copy_pages = mmu_radix_copy_pages,
498 	.cpu_bootstrap = mmu_radix_cpu_bootstrap,
499 	.growkernel = mmu_radix_growkernel,
500 	.init = mmu_radix_init,
501 	.map =      		mmu_radix_map,
502 	.mincore =      	mmu_radix_mincore,
503 	.pinit = mmu_radix_pinit,
504 	.pinit0 = mmu_radix_pinit0,
505 
506 	.mapdev = mmu_radix_mapdev,
507 	.mapdev_attr = mmu_radix_mapdev_attr,
508 	.unmapdev = mmu_radix_unmapdev,
509 	.kenter_attr = mmu_radix_kenter_attr,
510 	.dev_direct_mapped = mmu_radix_dev_direct_mapped,
511 	.dumpsys_pa_init = mmu_radix_scan_init,
512 	.dumpsys_map_chunk = mmu_radix_dumpsys_map,
513 	.page_is_mapped = mmu_radix_page_is_mapped,
514 	.ps_enabled = mmu_radix_ps_enabled,
515 	.align_superpage = mmu_radix_align_superpage,
516 	.object_init_pt = mmu_radix_object_init_pt,
517 	.protect = mmu_radix_protect,
518 	/* pmap dispatcher interface */
519 	.clear_modify = mmu_radix_clear_modify,
520 	.copy = mmu_radix_copy,
521 	.enter = mmu_radix_enter,
522 	.enter_object = mmu_radix_enter_object,
523 	.enter_quick = mmu_radix_enter_quick,
524 	.extract = mmu_radix_extract,
525 	.extract_and_hold = mmu_radix_extract_and_hold,
526 	.is_modified = mmu_radix_is_modified,
527 	.is_prefaultable = mmu_radix_is_prefaultable,
528 	.is_referenced = mmu_radix_is_referenced,
529 	.ts_referenced = mmu_radix_ts_referenced,
530 	.page_exists_quick = mmu_radix_page_exists_quick,
531 	.page_init = mmu_radix_page_init,
532 	.page_wired_mappings =  mmu_radix_page_wired_mappings,
533 	.qenter = mmu_radix_qenter,
534 	.qremove = mmu_radix_qremove,
535 	.release = mmu_radix_release,
536 	.remove = mmu_radix_remove,
537 	.remove_all = mmu_radix_remove_all,
538 	.remove_write = mmu_radix_remove_write,
539 	.unwire = mmu_radix_unwire,
540 	.zero_page = mmu_radix_zero_page,
541 	.zero_page_area = mmu_radix_zero_page_area,
542 	.activate = mmu_radix_activate,
543 	.quick_enter_page =  mmu_radix_quick_enter_page,
544 	.quick_remove_page =  mmu_radix_quick_remove_page,
545 	.page_set_memattr = mmu_radix_page_set_memattr,
546 	.page_array_startup =  mmu_radix_page_array_startup,
547 
548 	/* Internal interfaces */
549 	.kenter = mmu_radix_kenter,
550 	.kextract = mmu_radix_kextract,
551 	.kremove = mmu_radix_kremove,
552 	.change_attr = mmu_radix_change_attr,
553 	.decode_kernel_ptr =  mmu_radix_decode_kernel_ptr,
554 
555 	.tlbie_all = mmu_radix_tlbie_all,
556 };
557 
558 MMU_DEF(mmu_radix, MMU_TYPE_RADIX, mmu_radix_methods);
559 
560 static boolean_t pmap_demote_l3e_locked(pmap_t pmap, pml3_entry_t *l3e, vm_offset_t va,
561 	struct rwlock **lockp);
562 static boolean_t pmap_demote_l3e(pmap_t pmap, pml3_entry_t *pde, vm_offset_t va);
563 static int pmap_unuse_pt(pmap_t, vm_offset_t, pml3_entry_t, struct spglist *);
564 static int pmap_remove_l3e(pmap_t pmap, pml3_entry_t *pdq, vm_offset_t sva,
565     struct spglist *free, struct rwlock **lockp);
566 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva,
567     pml3_entry_t ptepde, struct spglist *free, struct rwlock **lockp);
568 static vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va);
569 static bool pmap_remove_page(pmap_t pmap, vm_offset_t va, pml3_entry_t *pde,
570     struct spglist *free);
571 static bool	pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
572 	pml3_entry_t *l3e, struct spglist *free, struct rwlock **lockp);
573 
574 static bool	pmap_pv_insert_l3e(pmap_t pmap, vm_offset_t va, pml3_entry_t l3e,
575 		    u_int flags, struct rwlock **lockp);
576 #if VM_NRESERVLEVEL > 0
577 static void	pmap_pv_promote_l3e(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
578 	struct rwlock **lockp);
579 #endif
580 static void	pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va);
581 static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte);
582 static vm_page_t mmu_radix_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m,
583 	vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp, bool *invalidate);
584 
585 static bool	pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m,
586 	vm_prot_t prot, struct rwlock **lockp);
587 static int	pmap_enter_l3e(pmap_t pmap, vm_offset_t va, pml3_entry_t newpde,
588 	u_int flags, vm_page_t m, struct rwlock **lockp);
589 
590 static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp);
591 static void free_pv_chunk(struct pv_chunk *pc);
592 static vm_page_t _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp);
593 static vm_page_t pmap_allocl3e(pmap_t pmap, vm_offset_t va,
594 	struct rwlock **lockp);
595 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va,
596 	struct rwlock **lockp);
597 static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m,
598     struct spglist *free);
599 static boolean_t pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free);
600 
601 static void pmap_invalidate_page(pmap_t pmap, vm_offset_t start);
602 static void pmap_invalidate_all(pmap_t pmap);
603 static int pmap_change_attr_locked(vm_offset_t va, vm_size_t size, int mode, bool flush);
604 
605 /*
606  * Internal flags for pmap_enter()'s helper functions.
607  */
608 #define	PMAP_ENTER_NORECLAIM	0x1000000	/* Don't reclaim PV entries. */
609 #define	PMAP_ENTER_NOREPLACE	0x2000000	/* Don't replace mappings. */
610 
611 #define UNIMPLEMENTED() panic("%s not implemented", __func__)
612 #define UNTESTED() panic("%s not yet tested", __func__)
613 
614 /* Number of supported PID bits */
615 static unsigned int isa3_pid_bits;
616 
617 /* PID to start allocating from */
618 static unsigned int isa3_base_pid;
619 
620 #define PROCTAB_SIZE_SHIFT	(isa3_pid_bits + 4)
621 #define PROCTAB_ENTRIES	(1ul << isa3_pid_bits)
622 
623 /*
624  * Map of physical memory regions.
625  */
626 static struct	mem_region *regions, *pregions;
627 static struct	numa_mem_region *numa_pregions;
628 static u_int	phys_avail_count;
629 static int	regions_sz, pregions_sz, numa_pregions_sz;
630 static struct pate *isa3_parttab;
631 static struct prte *isa3_proctab;
632 static vmem_t *asid_arena;
633 
634 extern void bs_remap_earlyboot(void);
635 
636 #define	RADIX_PGD_SIZE_SHIFT	16
637 #define RADIX_PGD_SIZE	(1UL << RADIX_PGD_SIZE_SHIFT)
638 
639 #define	RADIX_PGD_INDEX_SHIFT	(RADIX_PGD_SIZE_SHIFT-3)
640 #define NL2EPG (PAGE_SIZE/sizeof(pml2_entry_t))
641 #define NL3EPG (PAGE_SIZE/sizeof(pml3_entry_t))
642 
643 #define	NUPML1E		(RADIX_PGD_SIZE/sizeof(uint64_t))	/* number of userland PML1 pages */
644 #define	NUPDPE		(NUPML1E * NL2EPG)/* number of userland PDP pages */
645 #define	NUPDE		(NUPDPE * NL3EPG)	/* number of userland PD entries */
646 
647 /* POWER9 only permits a 64k partition table size. */
648 #define	PARTTAB_SIZE_SHIFT	16
649 #define PARTTAB_SIZE	(1UL << PARTTAB_SIZE_SHIFT)
650 
651 #define PARTTAB_HR		(1UL << 63) /* host uses radix */
652 #define PARTTAB_GR		(1UL << 63) /* guest uses radix must match host */
653 
654 /* TLB flush actions. Used as argument to tlbiel_all() */
655 enum {
656 	TLB_INVAL_SCOPE_LPID = 0,	/* invalidate TLBs for current LPID */
657 	TLB_INVAL_SCOPE_GLOBAL = 1,	/* invalidate all TLBs */
658 };
659 
660 #define	NPV_LIST_LOCKS	MAXCPU
661 static int pmap_initialized;
662 static vm_paddr_t proctab0pa;
663 static vm_paddr_t parttab_phys;
664 CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE);
665 
666 /*
667  * Data for the pv entry allocation mechanism.
668  * Updates to pv_invl_gen are protected by the pv_list_locks[]
669  * elements, but reads are not.
670  */
671 static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks);
672 static struct mtx __exclusive_cache_line pv_chunks_mutex;
673 static struct rwlock __exclusive_cache_line pv_list_locks[NPV_LIST_LOCKS];
674 static struct md_page *pv_table;
675 static struct md_page pv_dummy;
676 
677 #ifdef PV_STATS
678 #define PV_STAT(x)	do { x ; } while (0)
679 #else
680 #define PV_STAT(x)	do { } while (0)
681 #endif
682 
683 #define	pa_radix_index(pa)	((pa) >> L3_PAGE_SIZE_SHIFT)
684 #define	pa_to_pvh(pa)	(&pv_table[pa_radix_index(pa)])
685 
686 #define	PHYS_TO_PV_LIST_LOCK(pa)	\
687 			(&pv_list_locks[pa_radix_index(pa) % NPV_LIST_LOCKS])
688 
689 #define	CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa)	do {	\
690 	struct rwlock **_lockp = (lockp);		\
691 	struct rwlock *_new_lock;			\
692 							\
693 	_new_lock = PHYS_TO_PV_LIST_LOCK(pa);		\
694 	if (_new_lock != *_lockp) {			\
695 		if (*_lockp != NULL)			\
696 			rw_wunlock(*_lockp);		\
697 		*_lockp = _new_lock;			\
698 		rw_wlock(*_lockp);			\
699 	}						\
700 } while (0)
701 
702 #define	CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m)	\
703 	CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, VM_PAGE_TO_PHYS(m))
704 
705 #define	RELEASE_PV_LIST_LOCK(lockp)		do {	\
706 	struct rwlock **_lockp = (lockp);		\
707 							\
708 	if (*_lockp != NULL) {				\
709 		rw_wunlock(*_lockp);			\
710 		*_lockp = NULL;				\
711 	}						\
712 } while (0)
713 
714 #define	VM_PAGE_TO_PV_LIST_LOCK(m)	\
715 	PHYS_TO_PV_LIST_LOCK(VM_PAGE_TO_PHYS(m))
716 
717 /*
718  * We support 52 bits, hence:
719  * bits 52 - 31 = 21, 0b10101
720  * RTS encoding details
721  * bits 0 - 3 of rts -> bits 6 - 8 unsigned long
722  * bits 4 - 5 of rts -> bits 62 - 63 of unsigned long
723  */
724 #define RTS_SIZE ((0x2UL << 61) | (0x5UL << 5))
725 
726 static int powernv_enabled = 1;
727 
728 static __always_inline void
729 tlbiel_radix_set_isa300(uint32_t set, uint32_t is,
730 	uint32_t pid, uint32_t ric, uint32_t prs)
731 {
732 	uint64_t rb;
733 	uint64_t rs;
734 
735 	rb = PPC_BITLSHIFT_VAL(set, 51) | PPC_BITLSHIFT_VAL(is, 53);
736 	rs = PPC_BITLSHIFT_VAL((uint64_t)pid, 31);
737 
738 	__asm __volatile(PPC_TLBIEL(%0, %1, %2, %3, 1)
739 		     : : "r"(rb), "r"(rs), "i"(ric), "i"(prs)
740 		     : "memory");
741 }
742 
743 static void
744 tlbiel_flush_isa3(uint32_t num_sets, uint32_t is)
745 {
746 	uint32_t set;
747 
748 	__asm __volatile("ptesync": : :"memory");
749 
750 	/*
751 	 * Flush the first set of the TLB, and the entire Page Walk Cache
752 	 * and partition table entries. Then flush the remaining sets of the
753 	 * TLB.
754 	 */
755 	tlbiel_radix_set_isa300(0, is, 0, RIC_FLUSH_ALL, 0);
756 	for (set = 1; set < num_sets; set++)
757 		tlbiel_radix_set_isa300(set, is, 0, RIC_FLUSH_TLB, 0);
758 
759 	/* Do the same for process scoped entries. */
760 	tlbiel_radix_set_isa300(0, is, 0, RIC_FLUSH_ALL, 1);
761 	for (set = 1; set < num_sets; set++)
762 		tlbiel_radix_set_isa300(set, is, 0, RIC_FLUSH_TLB, 1);
763 
764 	__asm __volatile("ptesync": : :"memory");
765 }
766 
767 static void
768 mmu_radix_tlbiel_flush(int scope)
769 {
770 	int is;
771 
772 	MPASS(scope == TLB_INVAL_SCOPE_LPID ||
773 		  scope == TLB_INVAL_SCOPE_GLOBAL);
774 	is = scope + 2;
775 
776 	tlbiel_flush_isa3(POWER9_TLB_SETS_RADIX, is);
777 	__asm __volatile(PPC_INVALIDATE_ERAT "; isync" : : :"memory");
778 }
779 
780 static void
781 mmu_radix_tlbie_all()
782 {
783 	/* TODO: LPID invalidate */
784 	mmu_radix_tlbiel_flush(TLB_INVAL_SCOPE_GLOBAL);
785 }
786 
787 static void
788 mmu_radix_init_amor(void)
789 {
790 	/*
791 	* In HV mode, we init AMOR (Authority Mask Override Register) so that
792 	* the hypervisor and guest can setup IAMR (Instruction Authority Mask
793 	* Register), enable key 0 and set it to 1.
794 	*
795 	* AMOR = 0b1100 .... 0000 (Mask for key 0 is 11)
796 	*/
797 	mtspr(SPR_AMOR, (3ul << 62));
798 }
799 
800 static void
801 mmu_radix_init_iamr(void)
802 {
803 	/*
804 	 * Radix always uses key0 of the IAMR to determine if an access is
805 	 * allowed. We set bit 0 (IBM bit 1) of key0, to prevent instruction
806 	 * fetch.
807 	 */
808 	mtspr(SPR_IAMR, (1ul << 62));
809 }
810 
811 static void
812 mmu_radix_pid_set(pmap_t pmap)
813 {
814 
815 	mtspr(SPR_PID, pmap->pm_pid);
816 	isync();
817 }
818 
819 /* Quick sort callout for comparing physical addresses. */
820 static int
821 pa_cmp(const void *a, const void *b)
822 {
823 	const vm_paddr_t *pa = a, *pb = b;
824 
825 	if (*pa < *pb)
826 		return (-1);
827 	else if (*pa > *pb)
828 		return (1);
829 	else
830 		return (0);
831 }
832 
833 #define	pte_load_store(ptep, pte)	atomic_swap_long(ptep, pte)
834 #define	pte_load_clear(ptep)		atomic_swap_long(ptep, 0)
835 #define	pte_store(ptep, pte) do {	   \
836 	MPASS((pte) & (RPTE_EAA_R | RPTE_EAA_W | RPTE_EAA_X));	\
837 	*(u_long *)(ptep) = htobe64((u_long)((pte) | PG_V | RPTE_LEAF)); \
838 } while (0)
839 /*
840  * NB: should only be used for adding directories - not for direct mappings
841  */
842 #define	pde_store(ptep, pa) do {				\
843 	*(u_long *)(ptep) = htobe64((u_long)(pa|RPTE_VALID|RPTE_SHIFT)); \
844 } while (0)
845 
846 #define	pte_clear(ptep) do {					\
847 		*(u_long *)(ptep) = (u_long)(0);		\
848 } while (0)
849 
850 #define	PMAP_PDE_SUPERPAGE	(1 << 8)	/* supports 2MB superpages */
851 
852 /*
853  * Promotion to a 2MB (PDE) page mapping requires that the corresponding 4KB
854  * (PTE) page mappings have identical settings for the following fields:
855  */
856 #define	PG_PTE_PROMOTE	(PG_X | PG_MANAGED | PG_W | PG_PTE_CACHE | \
857 	    PG_M | PG_A | RPTE_EAA_MASK | PG_V)
858 
859 static __inline void
860 pmap_resident_count_inc(pmap_t pmap, int count)
861 {
862 
863 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
864 	pmap->pm_stats.resident_count += count;
865 }
866 
867 static __inline void
868 pmap_resident_count_dec(pmap_t pmap, int count)
869 {
870 
871 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
872 	KASSERT(pmap->pm_stats.resident_count >= count,
873 	    ("pmap %p resident count underflow %ld %d", pmap,
874 	    pmap->pm_stats.resident_count, count));
875 	pmap->pm_stats.resident_count -= count;
876 }
877 
878 static void
879 pagezero(vm_offset_t va)
880 {
881 	va = trunc_page(va);
882 
883 	bzero((void *)va, PAGE_SIZE);
884 }
885 
886 static uint64_t
887 allocpages(int n)
888 {
889 	u_int64_t ret;
890 
891 	ret = moea64_bootstrap_alloc(n * PAGE_SIZE, PAGE_SIZE);
892 	for (int i = 0; i < n; i++)
893 		pagezero(PHYS_TO_DMAP(ret + i * PAGE_SIZE));
894 	return (ret);
895 }
896 
897 static pt_entry_t *
898 kvtopte(vm_offset_t va)
899 {
900 	pt_entry_t *l3e;
901 
902 	l3e = pmap_pml3e(kernel_pmap, va);
903 	if ((be64toh(*l3e) & RPTE_VALID) == 0)
904 		return (NULL);
905 	return (pmap_l3e_to_pte(l3e, va));
906 }
907 
908 void
909 mmu_radix_kenter(vm_offset_t va, vm_paddr_t pa)
910 {
911 	pt_entry_t *pte;
912 
913 	pte = kvtopte(va);
914 	MPASS(pte != NULL);
915 	*pte = htobe64(pa | RPTE_VALID | RPTE_LEAF | RPTE_EAA_R | \
916 	    RPTE_EAA_W | RPTE_EAA_P | PG_M | PG_A);
917 }
918 
919 bool
920 mmu_radix_ps_enabled(pmap_t pmap)
921 {
922 	return (superpages_enabled && (pmap->pm_flags & PMAP_PDE_SUPERPAGE) != 0);
923 }
924 
925 static pt_entry_t *
926 pmap_nofault_pte(pmap_t pmap, vm_offset_t va, int *is_l3e)
927 {
928 	pml3_entry_t *l3e;
929 	pt_entry_t *pte;
930 
931 	va &= PG_PS_FRAME;
932 	l3e = pmap_pml3e(pmap, va);
933 	if (l3e == NULL || (be64toh(*l3e) & PG_V) == 0)
934 		return (NULL);
935 
936 	if (be64toh(*l3e) & RPTE_LEAF) {
937 		*is_l3e = 1;
938 		return (l3e);
939 	}
940 	*is_l3e = 0;
941 	va &= PG_FRAME;
942 	pte = pmap_l3e_to_pte(l3e, va);
943 	if (pte == NULL || (be64toh(*pte) & PG_V) == 0)
944 		return (NULL);
945 	return (pte);
946 }
947 
948 int
949 pmap_nofault(pmap_t pmap, vm_offset_t va, vm_prot_t flags)
950 {
951 	pt_entry_t *pte;
952 	pt_entry_t startpte, origpte, newpte;
953 	vm_page_t m;
954 	int is_l3e;
955 
956 	startpte = 0;
957  retry:
958 	if ((pte = pmap_nofault_pte(pmap, va, &is_l3e)) == NULL)
959 		return (KERN_INVALID_ADDRESS);
960 	origpte = newpte = be64toh(*pte);
961 	if (startpte == 0) {
962 		startpte = origpte;
963 		if (((flags & VM_PROT_WRITE) && (startpte & PG_M)) ||
964 		    ((flags & VM_PROT_READ) && (startpte & PG_A))) {
965 			pmap_invalidate_all(pmap);
966 #ifdef INVARIANTS
967 			if (VERBOSE_PMAP || pmap_logging)
968 				printf("%s(%p, %#lx, %#x) (%#lx) -- invalidate all\n",
969 				    __func__, pmap, va, flags, origpte);
970 #endif
971 			return (KERN_FAILURE);
972 		}
973 	}
974 #ifdef INVARIANTS
975 	if (VERBOSE_PMAP || pmap_logging)
976 		printf("%s(%p, %#lx, %#x) (%#lx)\n", __func__, pmap, va,
977 		    flags, origpte);
978 #endif
979 	PMAP_LOCK(pmap);
980 	if ((pte = pmap_nofault_pte(pmap, va, &is_l3e)) == NULL ||
981 	    be64toh(*pte) != origpte) {
982 		PMAP_UNLOCK(pmap);
983 		return (KERN_FAILURE);
984 	}
985 	m = PHYS_TO_VM_PAGE(newpte & PG_FRAME);
986 	MPASS(m != NULL);
987 	switch (flags) {
988 	case VM_PROT_READ:
989 		if ((newpte & (RPTE_EAA_R|RPTE_EAA_X)) == 0)
990 			goto protfail;
991 		newpte |= PG_A;
992 		vm_page_aflag_set(m, PGA_REFERENCED);
993 		break;
994 	case VM_PROT_WRITE:
995 		if ((newpte & RPTE_EAA_W) == 0)
996 			goto protfail;
997 		if (is_l3e)
998 			goto protfail;
999 		newpte |= PG_M;
1000 		vm_page_dirty(m);
1001 		break;
1002 	case VM_PROT_EXECUTE:
1003 		if ((newpte & RPTE_EAA_X) == 0)
1004 			goto protfail;
1005 		newpte |= PG_A;
1006 		vm_page_aflag_set(m, PGA_REFERENCED);
1007 		break;
1008 	}
1009 
1010 	if (!atomic_cmpset_long(pte, htobe64(origpte), htobe64(newpte)))
1011 		goto retry;
1012 	ptesync();
1013 	PMAP_UNLOCK(pmap);
1014 	if (startpte == newpte)
1015 		return (KERN_FAILURE);
1016 	return (0);
1017  protfail:
1018 	PMAP_UNLOCK(pmap);
1019 	return (KERN_PROTECTION_FAILURE);
1020 }
1021 
1022 /*
1023  * Returns TRUE if the given page is mapped individually or as part of
1024  * a 2mpage.  Otherwise, returns FALSE.
1025  */
1026 boolean_t
1027 mmu_radix_page_is_mapped(vm_page_t m)
1028 {
1029 	struct rwlock *lock;
1030 	boolean_t rv;
1031 
1032 	if ((m->oflags & VPO_UNMANAGED) != 0)
1033 		return (FALSE);
1034 	lock = VM_PAGE_TO_PV_LIST_LOCK(m);
1035 	rw_rlock(lock);
1036 	rv = !TAILQ_EMPTY(&m->md.pv_list) ||
1037 	    ((m->flags & PG_FICTITIOUS) == 0 &&
1038 	    !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list));
1039 	rw_runlock(lock);
1040 	return (rv);
1041 }
1042 
1043 /*
1044  * Determine the appropriate bits to set in a PTE or PDE for a specified
1045  * caching mode.
1046  */
1047 static int
1048 pmap_cache_bits(vm_memattr_t ma)
1049 {
1050 	if (ma != VM_MEMATTR_DEFAULT) {
1051 		switch (ma) {
1052 		case VM_MEMATTR_UNCACHEABLE:
1053 			return (RPTE_ATTR_GUARDEDIO);
1054 		case VM_MEMATTR_CACHEABLE:
1055 			return (RPTE_ATTR_MEM);
1056 		case VM_MEMATTR_WRITE_BACK:
1057 		case VM_MEMATTR_PREFETCHABLE:
1058 		case VM_MEMATTR_WRITE_COMBINING:
1059 			return (RPTE_ATTR_UNGUARDEDIO);
1060 		}
1061 	}
1062 	return (0);
1063 }
1064 
1065 static void
1066 pmap_invalidate_page(pmap_t pmap, vm_offset_t start)
1067 {
1068 	ptesync();
1069 	if (pmap == kernel_pmap)
1070 		radix_tlbie_invlpg_kernel_4k(start);
1071 	else
1072 		radix_tlbie_invlpg_user_4k(pmap->pm_pid, start);
1073 	ttusync();
1074 }
1075 
1076 static void
1077 pmap_invalidate_page_2m(pmap_t pmap, vm_offset_t start)
1078 {
1079 	ptesync();
1080 	if (pmap == kernel_pmap)
1081 		radix_tlbie_invlpg_kernel_2m(start);
1082 	else
1083 		radix_tlbie_invlpg_user_2m(pmap->pm_pid, start);
1084 	ttusync();
1085 }
1086 
1087 static void
1088 pmap_invalidate_pwc(pmap_t pmap)
1089 {
1090 	ptesync();
1091 	if (pmap == kernel_pmap)
1092 		radix_tlbie_invlpwc_kernel();
1093 	else
1094 		radix_tlbie_invlpwc_user(pmap->pm_pid);
1095 	ttusync();
1096 }
1097 
1098 static void
1099 pmap_invalidate_range(pmap_t pmap, vm_offset_t start, vm_offset_t end)
1100 {
1101 	if (((start - end) >> PAGE_SHIFT) > 8) {
1102 		pmap_invalidate_all(pmap);
1103 		return;
1104 	}
1105 	ptesync();
1106 	if (pmap == kernel_pmap) {
1107 		while (start < end) {
1108 			radix_tlbie_invlpg_kernel_4k(start);
1109 			start += PAGE_SIZE;
1110 		}
1111 	} else {
1112 		while (start < end) {
1113 			radix_tlbie_invlpg_user_4k(pmap->pm_pid, start);
1114 			start += PAGE_SIZE;
1115 		}
1116 	}
1117 	ttusync();
1118 }
1119 
1120 static void
1121 pmap_invalidate_all(pmap_t pmap)
1122 {
1123 	ptesync();
1124 	if (pmap == kernel_pmap)
1125 		radix_tlbie_flush_kernel();
1126 	else
1127 		radix_tlbie_flush_user(pmap->pm_pid);
1128 	ttusync();
1129 }
1130 
1131 static void
1132 pmap_invalidate_l3e_page(pmap_t pmap, vm_offset_t va, pml3_entry_t l3e)
1133 {
1134 
1135 	/*
1136 	 * When the PDE has PG_PROMOTED set, the 2MB page mapping was created
1137 	 * by a promotion that did not invalidate the 512 4KB page mappings
1138 	 * that might exist in the TLB.  Consequently, at this point, the TLB
1139 	 * may hold both 4KB and 2MB page mappings for the address range [va,
1140 	 * va + L3_PAGE_SIZE).  Therefore, the entire range must be invalidated here.
1141 	 * In contrast, when PG_PROMOTED is clear, the TLB will not hold any
1142 	 * 4KB page mappings for the address range [va, va + L3_PAGE_SIZE), and so a
1143 	 * single INVLPG suffices to invalidate the 2MB page mapping from the
1144 	 * TLB.
1145 	 */
1146 	ptesync();
1147 	if ((l3e & PG_PROMOTED) != 0)
1148 		pmap_invalidate_range(pmap, va, va + L3_PAGE_SIZE - 1);
1149 	else
1150 		pmap_invalidate_page_2m(pmap, va);
1151 
1152 	pmap_invalidate_pwc(pmap);
1153 }
1154 
1155 static __inline struct pv_chunk *
1156 pv_to_chunk(pv_entry_t pv)
1157 {
1158 
1159 	return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK));
1160 }
1161 
1162 #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap)
1163 
1164 #define	PC_FREE0	0xfffffffffffffffful
1165 #define	PC_FREE1	0x3ffffffffffffffful
1166 
1167 static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1 };
1168 
1169 /*
1170  * Ensure that the number of spare PV entries in the specified pmap meets or
1171  * exceeds the given count, "needed".
1172  *
1173  * The given PV list lock may be released.
1174  */
1175 static void
1176 reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp)
1177 {
1178 	struct pch new_tail;
1179 	struct pv_chunk *pc;
1180 	vm_page_t m;
1181 	int avail, free;
1182 	bool reclaimed;
1183 
1184 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1185 	KASSERT(lockp != NULL, ("reserve_pv_entries: lockp is NULL"));
1186 
1187 	/*
1188 	 * Newly allocated PV chunks must be stored in a private list until
1189 	 * the required number of PV chunks have been allocated.  Otherwise,
1190 	 * reclaim_pv_chunk() could recycle one of these chunks.  In
1191 	 * contrast, these chunks must be added to the pmap upon allocation.
1192 	 */
1193 	TAILQ_INIT(&new_tail);
1194 retry:
1195 	avail = 0;
1196 	TAILQ_FOREACH(pc, &pmap->pm_pvchunk, pc_list) {
1197 		//		if ((cpu_feature2 & CPUID2_POPCNT) == 0)
1198 		bit_count((bitstr_t *)pc->pc_map, 0,
1199 				  sizeof(pc->pc_map) * NBBY, &free);
1200 #if 0
1201 		free = popcnt_pc_map_pq(pc->pc_map);
1202 #endif
1203 		if (free == 0)
1204 			break;
1205 		avail += free;
1206 		if (avail >= needed)
1207 			break;
1208 	}
1209 	for (reclaimed = false; avail < needed; avail += _NPCPV) {
1210 		m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
1211 		    VM_ALLOC_WIRED);
1212 		if (m == NULL) {
1213 			m = reclaim_pv_chunk(pmap, lockp);
1214 			if (m == NULL)
1215 				goto retry;
1216 			reclaimed = true;
1217 		}
1218 		PV_STAT(atomic_add_int(&pc_chunk_count, 1));
1219 		PV_STAT(atomic_add_int(&pc_chunk_allocs, 1));
1220 		dump_add_page(m->phys_addr);
1221 		pc = (void *)PHYS_TO_DMAP(m->phys_addr);
1222 		pc->pc_pmap = pmap;
1223 		pc->pc_map[0] = PC_FREE0;
1224 		pc->pc_map[1] = PC_FREE1;
1225 		TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
1226 		TAILQ_INSERT_TAIL(&new_tail, pc, pc_lru);
1227 		PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV));
1228 
1229 		/*
1230 		 * The reclaim might have freed a chunk from the current pmap.
1231 		 * If that chunk contained available entries, we need to
1232 		 * re-count the number of available entries.
1233 		 */
1234 		if (reclaimed)
1235 			goto retry;
1236 	}
1237 	if (!TAILQ_EMPTY(&new_tail)) {
1238 		mtx_lock(&pv_chunks_mutex);
1239 		TAILQ_CONCAT(&pv_chunks, &new_tail, pc_lru);
1240 		mtx_unlock(&pv_chunks_mutex);
1241 	}
1242 }
1243 
1244 /*
1245  * First find and then remove the pv entry for the specified pmap and virtual
1246  * address from the specified pv list.  Returns the pv entry if found and NULL
1247  * otherwise.  This operation can be performed on pv lists for either 4KB or
1248  * 2MB page mappings.
1249  */
1250 static __inline pv_entry_t
1251 pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
1252 {
1253 	pv_entry_t pv;
1254 
1255 	TAILQ_FOREACH(pv, &pvh->pv_list, pv_link) {
1256 #ifdef INVARIANTS
1257 		if (PV_PMAP(pv) == NULL) {
1258 			printf("corrupted pv_chunk/pv %p\n", pv);
1259 			printf("pv_chunk: %64D\n", pv_to_chunk(pv), ":");
1260 		}
1261 		MPASS(PV_PMAP(pv) != NULL);
1262 		MPASS(pv->pv_va != 0);
1263 #endif
1264 		if (pmap == PV_PMAP(pv) && va == pv->pv_va) {
1265 			TAILQ_REMOVE(&pvh->pv_list, pv, pv_link);
1266 			pvh->pv_gen++;
1267 			break;
1268 		}
1269 	}
1270 	return (pv);
1271 }
1272 
1273 /*
1274  * After demotion from a 2MB page mapping to 512 4KB page mappings,
1275  * destroy the pv entry for the 2MB page mapping and reinstantiate the pv
1276  * entries for each of the 4KB page mappings.
1277  */
1278 static void
1279 pmap_pv_demote_l3e(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1280     struct rwlock **lockp)
1281 {
1282 	struct md_page *pvh;
1283 	struct pv_chunk *pc;
1284 	pv_entry_t pv;
1285 	vm_offset_t va_last;
1286 	vm_page_t m;
1287 	int bit, field;
1288 
1289 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1290 	KASSERT((pa & L3_PAGE_MASK) == 0,
1291 	    ("pmap_pv_demote_pde: pa is not 2mpage aligned"));
1292 	CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
1293 
1294 	/*
1295 	 * Transfer the 2mpage's pv entry for this mapping to the first
1296 	 * page's pv list.  Once this transfer begins, the pv list lock
1297 	 * must not be released until the last pv entry is reinstantiated.
1298 	 */
1299 	pvh = pa_to_pvh(pa);
1300 	va = trunc_2mpage(va);
1301 	pv = pmap_pvh_remove(pvh, pmap, va);
1302 	KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found"));
1303 	m = PHYS_TO_VM_PAGE(pa);
1304 	TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_link);
1305 
1306 	m->md.pv_gen++;
1307 	/* Instantiate the remaining NPTEPG - 1 pv entries. */
1308 	PV_STAT(atomic_add_long(&pv_entry_allocs, NPTEPG - 1));
1309 	va_last = va + L3_PAGE_SIZE - PAGE_SIZE;
1310 	for (;;) {
1311 		pc = TAILQ_FIRST(&pmap->pm_pvchunk);
1312 		KASSERT(pc->pc_map[0] != 0 || pc->pc_map[1] != 0
1313 		    , ("pmap_pv_demote_pde: missing spare"));
1314 		for (field = 0; field < _NPCM; field++) {
1315 			while (pc->pc_map[field]) {
1316 				bit = cnttzd(pc->pc_map[field]);
1317 				pc->pc_map[field] &= ~(1ul << bit);
1318 				pv = &pc->pc_pventry[field * 64 + bit];
1319 				va += PAGE_SIZE;
1320 				pv->pv_va = va;
1321 				m++;
1322 				KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1323 			    ("pmap_pv_demote_pde: page %p is not managed", m));
1324 				TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_link);
1325 
1326 				m->md.pv_gen++;
1327 				if (va == va_last)
1328 					goto out;
1329 			}
1330 		}
1331 		TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
1332 		TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
1333 	}
1334 out:
1335 	if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0) {
1336 		TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
1337 		TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list);
1338 	}
1339 	PV_STAT(atomic_add_long(&pv_entry_count, NPTEPG - 1));
1340 	PV_STAT(atomic_subtract_int(&pv_entry_spare, NPTEPG - 1));
1341 }
1342 
1343 static void
1344 reclaim_pv_chunk_leave_pmap(pmap_t pmap, pmap_t locked_pmap)
1345 {
1346 
1347 	if (pmap == NULL)
1348 		return;
1349 	pmap_invalidate_all(pmap);
1350 	if (pmap != locked_pmap)
1351 		PMAP_UNLOCK(pmap);
1352 }
1353 
1354 /*
1355  * We are in a serious low memory condition.  Resort to
1356  * drastic measures to free some pages so we can allocate
1357  * another pv entry chunk.
1358  *
1359  * Returns NULL if PV entries were reclaimed from the specified pmap.
1360  *
1361  * We do not, however, unmap 2mpages because subsequent accesses will
1362  * allocate per-page pv entries until repromotion occurs, thereby
1363  * exacerbating the shortage of free pv entries.
1364  */
1365 static int active_reclaims = 0;
1366 static vm_page_t
1367 reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp)
1368 {
1369 	struct pv_chunk *pc, *pc_marker, *pc_marker_end;
1370 	struct pv_chunk_header pc_marker_b, pc_marker_end_b;
1371 	struct md_page *pvh;
1372 	pml3_entry_t *l3e;
1373 	pmap_t next_pmap, pmap;
1374 	pt_entry_t *pte, tpte;
1375 	pv_entry_t pv;
1376 	vm_offset_t va;
1377 	vm_page_t m, m_pc;
1378 	struct spglist free;
1379 	uint64_t inuse;
1380 	int bit, field, freed;
1381 
1382 	PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED);
1383 	KASSERT(lockp != NULL, ("reclaim_pv_chunk: lockp is NULL"));
1384 	pmap = NULL;
1385 	m_pc = NULL;
1386 	SLIST_INIT(&free);
1387 	bzero(&pc_marker_b, sizeof(pc_marker_b));
1388 	bzero(&pc_marker_end_b, sizeof(pc_marker_end_b));
1389 	pc_marker = (struct pv_chunk *)&pc_marker_b;
1390 	pc_marker_end = (struct pv_chunk *)&pc_marker_end_b;
1391 
1392 	mtx_lock(&pv_chunks_mutex);
1393 	active_reclaims++;
1394 	TAILQ_INSERT_HEAD(&pv_chunks, pc_marker, pc_lru);
1395 	TAILQ_INSERT_TAIL(&pv_chunks, pc_marker_end, pc_lru);
1396 	while ((pc = TAILQ_NEXT(pc_marker, pc_lru)) != pc_marker_end &&
1397 	    SLIST_EMPTY(&free)) {
1398 		next_pmap = pc->pc_pmap;
1399 		if (next_pmap == NULL) {
1400 			/*
1401 			 * The next chunk is a marker.  However, it is
1402 			 * not our marker, so active_reclaims must be
1403 			 * > 1.  Consequently, the next_chunk code
1404 			 * will not rotate the pv_chunks list.
1405 			 */
1406 			goto next_chunk;
1407 		}
1408 		mtx_unlock(&pv_chunks_mutex);
1409 
1410 		/*
1411 		 * A pv_chunk can only be removed from the pc_lru list
1412 		 * when both pc_chunks_mutex is owned and the
1413 		 * corresponding pmap is locked.
1414 		 */
1415 		if (pmap != next_pmap) {
1416 			reclaim_pv_chunk_leave_pmap(pmap, locked_pmap);
1417 			pmap = next_pmap;
1418 			/* Avoid deadlock and lock recursion. */
1419 			if (pmap > locked_pmap) {
1420 				RELEASE_PV_LIST_LOCK(lockp);
1421 				PMAP_LOCK(pmap);
1422 				mtx_lock(&pv_chunks_mutex);
1423 				continue;
1424 			} else if (pmap != locked_pmap) {
1425 				if (PMAP_TRYLOCK(pmap)) {
1426 					mtx_lock(&pv_chunks_mutex);
1427 					continue;
1428 				} else {
1429 					pmap = NULL; /* pmap is not locked */
1430 					mtx_lock(&pv_chunks_mutex);
1431 					pc = TAILQ_NEXT(pc_marker, pc_lru);
1432 					if (pc == NULL ||
1433 					    pc->pc_pmap != next_pmap)
1434 						continue;
1435 					goto next_chunk;
1436 				}
1437 			}
1438 		}
1439 
1440 		/*
1441 		 * Destroy every non-wired, 4 KB page mapping in the chunk.
1442 		 */
1443 		freed = 0;
1444 		for (field = 0; field < _NPCM; field++) {
1445 			for (inuse = ~pc->pc_map[field] & pc_freemask[field];
1446 			    inuse != 0; inuse &= ~(1UL << bit)) {
1447 				bit = cnttzd(inuse);
1448 				pv = &pc->pc_pventry[field * 64 + bit];
1449 				va = pv->pv_va;
1450 				l3e = pmap_pml3e(pmap, va);
1451 				if ((be64toh(*l3e) & RPTE_LEAF) != 0)
1452 					continue;
1453 				pte = pmap_l3e_to_pte(l3e, va);
1454 				if ((be64toh(*pte) & PG_W) != 0)
1455 					continue;
1456 				tpte = be64toh(pte_load_clear(pte));
1457 				m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
1458 				if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
1459 					vm_page_dirty(m);
1460 				if ((tpte & PG_A) != 0)
1461 					vm_page_aflag_set(m, PGA_REFERENCED);
1462 				CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
1463 				TAILQ_REMOVE(&m->md.pv_list, pv, pv_link);
1464 
1465 				m->md.pv_gen++;
1466 				if (TAILQ_EMPTY(&m->md.pv_list) &&
1467 				    (m->flags & PG_FICTITIOUS) == 0) {
1468 					pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
1469 					if (TAILQ_EMPTY(&pvh->pv_list)) {
1470 						vm_page_aflag_clear(m,
1471 						    PGA_WRITEABLE);
1472 					}
1473 				}
1474 				pc->pc_map[field] |= 1UL << bit;
1475 				pmap_unuse_pt(pmap, va, be64toh(*l3e), &free);
1476 				freed++;
1477 			}
1478 		}
1479 		if (freed == 0) {
1480 			mtx_lock(&pv_chunks_mutex);
1481 			goto next_chunk;
1482 		}
1483 		/* Every freed mapping is for a 4 KB page. */
1484 		pmap_resident_count_dec(pmap, freed);
1485 		PV_STAT(atomic_add_long(&pv_entry_frees, freed));
1486 		PV_STAT(atomic_add_int(&pv_entry_spare, freed));
1487 		PV_STAT(atomic_subtract_long(&pv_entry_count, freed));
1488 		TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
1489 		if (pc->pc_map[0] == PC_FREE0 && pc->pc_map[1] == PC_FREE1) {
1490 			PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV));
1491 			PV_STAT(atomic_subtract_int(&pc_chunk_count, 1));
1492 			PV_STAT(atomic_add_int(&pc_chunk_frees, 1));
1493 			/* Entire chunk is free; return it. */
1494 			m_pc = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc));
1495 			dump_drop_page(m_pc->phys_addr);
1496 			mtx_lock(&pv_chunks_mutex);
1497 			TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
1498 			break;
1499 		}
1500 		TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
1501 		mtx_lock(&pv_chunks_mutex);
1502 		/* One freed pv entry in locked_pmap is sufficient. */
1503 		if (pmap == locked_pmap)
1504 			break;
1505 next_chunk:
1506 		TAILQ_REMOVE(&pv_chunks, pc_marker, pc_lru);
1507 		TAILQ_INSERT_AFTER(&pv_chunks, pc, pc_marker, pc_lru);
1508 		if (active_reclaims == 1 && pmap != NULL) {
1509 			/*
1510 			 * Rotate the pv chunks list so that we do not
1511 			 * scan the same pv chunks that could not be
1512 			 * freed (because they contained a wired
1513 			 * and/or superpage mapping) on every
1514 			 * invocation of reclaim_pv_chunk().
1515 			 */
1516 			while ((pc = TAILQ_FIRST(&pv_chunks)) != pc_marker) {
1517 				MPASS(pc->pc_pmap != NULL);
1518 				TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
1519 				TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru);
1520 			}
1521 		}
1522 	}
1523 	TAILQ_REMOVE(&pv_chunks, pc_marker, pc_lru);
1524 	TAILQ_REMOVE(&pv_chunks, pc_marker_end, pc_lru);
1525 	active_reclaims--;
1526 	mtx_unlock(&pv_chunks_mutex);
1527 	reclaim_pv_chunk_leave_pmap(pmap, locked_pmap);
1528 	if (m_pc == NULL && !SLIST_EMPTY(&free)) {
1529 		m_pc = SLIST_FIRST(&free);
1530 		SLIST_REMOVE_HEAD(&free, plinks.s.ss);
1531 		/* Recycle a freed page table page. */
1532 		m_pc->ref_count = 1;
1533 	}
1534 	vm_page_free_pages_toq(&free, true);
1535 	return (m_pc);
1536 }
1537 
1538 /*
1539  * free the pv_entry back to the free list
1540  */
1541 static void
1542 free_pv_entry(pmap_t pmap, pv_entry_t pv)
1543 {
1544 	struct pv_chunk *pc;
1545 	int idx, field, bit;
1546 
1547 #ifdef VERBOSE_PV
1548 	if (pmap != kernel_pmap)
1549 		printf("%s(%p, %p)\n", __func__, pmap, pv);
1550 #endif
1551 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1552 	PV_STAT(atomic_add_long(&pv_entry_frees, 1));
1553 	PV_STAT(atomic_add_int(&pv_entry_spare, 1));
1554 	PV_STAT(atomic_subtract_long(&pv_entry_count, 1));
1555 	pc = pv_to_chunk(pv);
1556 	idx = pv - &pc->pc_pventry[0];
1557 	field = idx / 64;
1558 	bit = idx % 64;
1559 	pc->pc_map[field] |= 1ul << bit;
1560 	if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1) {
1561 		/* 98% of the time, pc is already at the head of the list. */
1562 		if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) {
1563 			TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
1564 			TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
1565 		}
1566 		return;
1567 	}
1568 	TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
1569 	free_pv_chunk(pc);
1570 }
1571 
1572 static void
1573 free_pv_chunk(struct pv_chunk *pc)
1574 {
1575 	vm_page_t m;
1576 
1577 	mtx_lock(&pv_chunks_mutex);
1578  	TAILQ_REMOVE(&pv_chunks, pc, pc_lru);
1579 	mtx_unlock(&pv_chunks_mutex);
1580 	PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV));
1581 	PV_STAT(atomic_subtract_int(&pc_chunk_count, 1));
1582 	PV_STAT(atomic_add_int(&pc_chunk_frees, 1));
1583 	/* entire chunk is free, return it */
1584 	m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc));
1585 	dump_drop_page(m->phys_addr);
1586 	vm_page_unwire_noq(m);
1587 	vm_page_free(m);
1588 }
1589 
1590 /*
1591  * Returns a new PV entry, allocating a new PV chunk from the system when
1592  * needed.  If this PV chunk allocation fails and a PV list lock pointer was
1593  * given, a PV chunk is reclaimed from an arbitrary pmap.  Otherwise, NULL is
1594  * returned.
1595  *
1596  * The given PV list lock may be released.
1597  */
1598 static pv_entry_t
1599 get_pv_entry(pmap_t pmap, struct rwlock **lockp)
1600 {
1601 	int bit, field;
1602 	pv_entry_t pv;
1603 	struct pv_chunk *pc;
1604 	vm_page_t m;
1605 
1606 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1607 	PV_STAT(atomic_add_long(&pv_entry_allocs, 1));
1608 retry:
1609 	pc = TAILQ_FIRST(&pmap->pm_pvchunk);
1610 	if (pc != NULL) {
1611 		for (field = 0; field < _NPCM; field++) {
1612 			if (pc->pc_map[field]) {
1613 				bit = cnttzd(pc->pc_map[field]);
1614 				break;
1615 			}
1616 		}
1617 		if (field < _NPCM) {
1618 			pv = &pc->pc_pventry[field * 64 + bit];
1619 			pc->pc_map[field] &= ~(1ul << bit);
1620 			/* If this was the last item, move it to tail */
1621 			if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0) {
1622 				TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
1623 				TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc,
1624 				    pc_list);
1625 			}
1626 			PV_STAT(atomic_add_long(&pv_entry_count, 1));
1627 			PV_STAT(atomic_subtract_int(&pv_entry_spare, 1));
1628 			MPASS(PV_PMAP(pv) != NULL);
1629 			return (pv);
1630 		}
1631 	}
1632 	/* No free items, allocate another chunk */
1633 	m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ |
1634 	    VM_ALLOC_WIRED);
1635 	if (m == NULL) {
1636 		if (lockp == NULL) {
1637 			PV_STAT(pc_chunk_tryfail++);
1638 			return (NULL);
1639 		}
1640 		m = reclaim_pv_chunk(pmap, lockp);
1641 		if (m == NULL)
1642 			goto retry;
1643 	}
1644 	PV_STAT(atomic_add_int(&pc_chunk_count, 1));
1645 	PV_STAT(atomic_add_int(&pc_chunk_allocs, 1));
1646 	dump_add_page(m->phys_addr);
1647 	pc = (void *)PHYS_TO_DMAP(m->phys_addr);
1648 	pc->pc_pmap = pmap;
1649 	pc->pc_map[0] = PC_FREE0 & ~1ul;	/* preallocated bit 0 */
1650 	pc->pc_map[1] = PC_FREE1;
1651 	mtx_lock(&pv_chunks_mutex);
1652 	TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru);
1653 	mtx_unlock(&pv_chunks_mutex);
1654 	pv = &pc->pc_pventry[0];
1655 	TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list);
1656 	PV_STAT(atomic_add_long(&pv_entry_count, 1));
1657 	PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV - 1));
1658 	MPASS(PV_PMAP(pv) != NULL);
1659 	return (pv);
1660 }
1661 
1662 #if VM_NRESERVLEVEL > 0
1663 /*
1664  * After promotion from 512 4KB page mappings to a single 2MB page mapping,
1665  * replace the many pv entries for the 4KB page mappings by a single pv entry
1666  * for the 2MB page mapping.
1667  */
1668 static void
1669 pmap_pv_promote_l3e(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1670     struct rwlock **lockp)
1671 {
1672 	struct md_page *pvh;
1673 	pv_entry_t pv;
1674 	vm_offset_t va_last;
1675 	vm_page_t m;
1676 
1677 	KASSERT((pa & L3_PAGE_MASK) == 0,
1678 	    ("pmap_pv_promote_pde: pa is not 2mpage aligned"));
1679 	CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
1680 
1681 	/*
1682 	 * Transfer the first page's pv entry for this mapping to the 2mpage's
1683 	 * pv list.  Aside from avoiding the cost of a call to get_pv_entry(),
1684 	 * a transfer avoids the possibility that get_pv_entry() calls
1685 	 * reclaim_pv_chunk() and that reclaim_pv_chunk() removes one of the
1686 	 * mappings that is being promoted.
1687 	 */
1688 	m = PHYS_TO_VM_PAGE(pa);
1689 	va = trunc_2mpage(va);
1690 	pv = pmap_pvh_remove(&m->md, pmap, va);
1691 	KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found"));
1692 	pvh = pa_to_pvh(pa);
1693 	TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_link);
1694 	pvh->pv_gen++;
1695 	/* Free the remaining NPTEPG - 1 pv entries. */
1696 	va_last = va + L3_PAGE_SIZE - PAGE_SIZE;
1697 	do {
1698 		m++;
1699 		va += PAGE_SIZE;
1700 		pmap_pvh_free(&m->md, pmap, va);
1701 	} while (va < va_last);
1702 }
1703 #endif /* VM_NRESERVLEVEL > 0 */
1704 
1705 /*
1706  * First find and then destroy the pv entry for the specified pmap and virtual
1707  * address.  This operation can be performed on pv lists for either 4KB or 2MB
1708  * page mappings.
1709  */
1710 static void
1711 pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va)
1712 {
1713 	pv_entry_t pv;
1714 
1715 	pv = pmap_pvh_remove(pvh, pmap, va);
1716 	KASSERT(pv != NULL, ("pmap_pvh_free: pv not found"));
1717 	free_pv_entry(pmap, pv);
1718 }
1719 
1720 /*
1721  * Conditionally create the PV entry for a 4KB page mapping if the required
1722  * memory can be allocated without resorting to reclamation.
1723  */
1724 static boolean_t
1725 pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m,
1726     struct rwlock **lockp)
1727 {
1728 	pv_entry_t pv;
1729 
1730 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1731 	/* Pass NULL instead of the lock pointer to disable reclamation. */
1732 	if ((pv = get_pv_entry(pmap, NULL)) != NULL) {
1733 		pv->pv_va = va;
1734 		CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
1735 		TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_link);
1736 		m->md.pv_gen++;
1737 		return (TRUE);
1738 	} else
1739 		return (FALSE);
1740 }
1741 
1742 vm_paddr_t phys_avail_debug[2 * VM_PHYSSEG_MAX];
1743 #ifdef INVARIANTS
1744 static void
1745 validate_addr(vm_paddr_t addr, vm_size_t size)
1746 {
1747 	vm_paddr_t end = addr + size;
1748 	bool found = false;
1749 
1750 	for (int i = 0; i < 2 * phys_avail_count; i += 2) {
1751 		if (addr >= phys_avail_debug[i] &&
1752 			end <= phys_avail_debug[i + 1]) {
1753 			found = true;
1754 			break;
1755 		}
1756 	}
1757 	KASSERT(found, ("%#lx-%#lx outside of initial phys_avail array",
1758 					addr, end));
1759 }
1760 #else
1761 static void validate_addr(vm_paddr_t addr, vm_size_t size) {}
1762 #endif
1763 #define DMAP_PAGE_BITS (RPTE_VALID | RPTE_LEAF | RPTE_EAA_MASK | PG_M | PG_A)
1764 
1765 static vm_paddr_t
1766 alloc_pt_page(void)
1767 {
1768 	vm_paddr_t page;
1769 
1770 	page = allocpages(1);
1771 	pagezero(PHYS_TO_DMAP(page));
1772 	return (page);
1773 }
1774 
1775 static void
1776 mmu_radix_dmap_range(vm_paddr_t start, vm_paddr_t end)
1777 {
1778 	pt_entry_t *pte, pteval;
1779 	vm_paddr_t page;
1780 
1781 	if (bootverbose)
1782 		printf("%s %lx -> %lx\n", __func__, start, end);
1783 	while (start < end) {
1784 		pteval = start | DMAP_PAGE_BITS;
1785 		pte = pmap_pml1e(kernel_pmap, PHYS_TO_DMAP(start));
1786 		if ((be64toh(*pte) & RPTE_VALID) == 0) {
1787 			page = alloc_pt_page();
1788 			pde_store(pte, page);
1789 		}
1790 		pte = pmap_l1e_to_l2e(pte, PHYS_TO_DMAP(start));
1791 		if ((start & L2_PAGE_MASK) == 0 &&
1792 			end - start >= L2_PAGE_SIZE) {
1793 			start += L2_PAGE_SIZE;
1794 			goto done;
1795 		} else if ((be64toh(*pte) & RPTE_VALID) == 0) {
1796 			page = alloc_pt_page();
1797 			pde_store(pte, page);
1798 		}
1799 
1800 		pte = pmap_l2e_to_l3e(pte, PHYS_TO_DMAP(start));
1801 		if ((start & L3_PAGE_MASK) == 0 &&
1802 			end - start >= L3_PAGE_SIZE) {
1803 			start += L3_PAGE_SIZE;
1804 			goto done;
1805 		} else if ((be64toh(*pte) & RPTE_VALID) == 0) {
1806 			page = alloc_pt_page();
1807 			pde_store(pte, page);
1808 		}
1809 		pte = pmap_l3e_to_pte(pte, PHYS_TO_DMAP(start));
1810 		start += PAGE_SIZE;
1811 	done:
1812 		pte_store(pte, pteval);
1813 	}
1814 }
1815 
1816 static void
1817 mmu_radix_dmap_populate(vm_size_t hwphyssz)
1818 {
1819 	vm_paddr_t start, end;
1820 
1821 	for (int i = 0; i < pregions_sz; i++) {
1822 		start = pregions[i].mr_start;
1823 		end = start + pregions[i].mr_size;
1824 		if (hwphyssz && start >= hwphyssz)
1825 			break;
1826 		if (hwphyssz && hwphyssz < end)
1827 			end = hwphyssz;
1828 		mmu_radix_dmap_range(start, end);
1829 	}
1830 }
1831 
1832 static void
1833 mmu_radix_setup_pagetables(vm_size_t hwphyssz)
1834 {
1835 	vm_paddr_t ptpages, pages;
1836 	pt_entry_t *pte;
1837 	vm_paddr_t l1phys;
1838 
1839 	bzero(kernel_pmap, sizeof(struct pmap));
1840 	PMAP_LOCK_INIT(kernel_pmap);
1841 
1842 	ptpages = allocpages(3);
1843 	l1phys = moea64_bootstrap_alloc(RADIX_PGD_SIZE, RADIX_PGD_SIZE);
1844 	validate_addr(l1phys, RADIX_PGD_SIZE);
1845 	if (bootverbose)
1846 		printf("l1phys=%lx\n", l1phys);
1847 	MPASS((l1phys & (RADIX_PGD_SIZE-1)) == 0);
1848 	for (int i = 0; i < RADIX_PGD_SIZE/PAGE_SIZE; i++)
1849 		pagezero(PHYS_TO_DMAP(l1phys + i * PAGE_SIZE));
1850 	kernel_pmap->pm_pml1 = (pml1_entry_t *)PHYS_TO_DMAP(l1phys);
1851 
1852 	mmu_radix_dmap_populate(hwphyssz);
1853 
1854 	/*
1855 	 * Create page tables for first 128MB of KVA
1856 	 */
1857 	pages = ptpages;
1858 	pte = pmap_pml1e(kernel_pmap, VM_MIN_KERNEL_ADDRESS);
1859 	*pte = htobe64(pages | RPTE_VALID | RPTE_SHIFT);
1860 	pages += PAGE_SIZE;
1861 	pte = pmap_l1e_to_l2e(pte, VM_MIN_KERNEL_ADDRESS);
1862 	*pte = htobe64(pages | RPTE_VALID | RPTE_SHIFT);
1863 	pages += PAGE_SIZE;
1864 	pte = pmap_l2e_to_l3e(pte, VM_MIN_KERNEL_ADDRESS);
1865 	/*
1866 	 * the kernel page table pages need to be preserved in
1867 	 * phys_avail and not overlap with previous  allocations
1868 	 */
1869 	pages = allocpages(nkpt);
1870 	if (bootverbose) {
1871 		printf("phys_avail after dmap populate and nkpt allocation\n");
1872 		for (int j = 0; j < 2 * phys_avail_count; j+=2)
1873 			printf("phys_avail[%d]=%08lx - phys_avail[%d]=%08lx\n",
1874 				   j, phys_avail[j], j + 1, phys_avail[j + 1]);
1875 	}
1876 	KPTphys = pages;
1877 	for (int i = 0; i < nkpt; i++, pte++, pages += PAGE_SIZE)
1878 		*pte = htobe64(pages | RPTE_VALID | RPTE_SHIFT);
1879 	kernel_vm_end = VM_MIN_KERNEL_ADDRESS + nkpt * L3_PAGE_SIZE;
1880 	if (bootverbose)
1881 		printf("kernel_pmap pml1 %p\n", kernel_pmap->pm_pml1);
1882 	/*
1883 	 * Add a physical memory segment (vm_phys_seg) corresponding to the
1884 	 * preallocated kernel page table pages so that vm_page structures
1885 	 * representing these pages will be created.  The vm_page structures
1886 	 * are required for promotion of the corresponding kernel virtual
1887 	 * addresses to superpage mappings.
1888 	 */
1889 	vm_phys_add_seg(KPTphys, KPTphys + ptoa(nkpt));
1890 }
1891 
1892 static void
1893 mmu_radix_early_bootstrap(vm_offset_t start, vm_offset_t end)
1894 {
1895 	vm_paddr_t	kpstart, kpend;
1896 	vm_size_t	physsz, hwphyssz;
1897 	//uint64_t	l2virt;
1898 	int		rm_pavail, proctab_size;
1899 	int		i, j;
1900 
1901 	kpstart = start & ~DMAP_BASE_ADDRESS;
1902 	kpend = end & ~DMAP_BASE_ADDRESS;
1903 
1904 	/* Get physical memory regions from firmware */
1905 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
1906 	CTR0(KTR_PMAP, "mmu_radix_early_bootstrap: physical memory");
1907 
1908 	if (2 * VM_PHYSSEG_MAX < regions_sz)
1909 		panic("mmu_radix_early_bootstrap: phys_avail too small");
1910 
1911 	if (bootverbose)
1912 		for (int i = 0; i < regions_sz; i++)
1913 			printf("regions[%d].mr_start=%lx regions[%d].mr_size=%lx\n",
1914 			    i, regions[i].mr_start, i, regions[i].mr_size);
1915 	/*
1916 	 * XXX workaround a simulator bug
1917 	 */
1918 	for (int i = 0; i < regions_sz; i++)
1919 		if (regions[i].mr_start & PAGE_MASK) {
1920 			regions[i].mr_start += PAGE_MASK;
1921 			regions[i].mr_start &= ~PAGE_MASK;
1922 			regions[i].mr_size &= ~PAGE_MASK;
1923 		}
1924 	if (bootverbose)
1925 		for (int i = 0; i < pregions_sz; i++)
1926 			printf("pregions[%d].mr_start=%lx pregions[%d].mr_size=%lx\n",
1927 			    i, pregions[i].mr_start, i, pregions[i].mr_size);
1928 
1929 	phys_avail_count = 0;
1930 	physsz = 0;
1931 	hwphyssz = 0;
1932 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
1933 	for (i = 0, j = 0; i < regions_sz; i++) {
1934 		if (bootverbose)
1935 			printf("regions[%d].mr_start=%016lx regions[%d].mr_size=%016lx\n",
1936 			    i, regions[i].mr_start, i, regions[i].mr_size);
1937 
1938 		if (regions[i].mr_size < PAGE_SIZE)
1939 			continue;
1940 
1941 		if (hwphyssz != 0 &&
1942 		    (physsz + regions[i].mr_size) >= hwphyssz) {
1943 			if (physsz < hwphyssz) {
1944 				phys_avail[j] = regions[i].mr_start;
1945 				phys_avail[j + 1] = regions[i].mr_start +
1946 				    (hwphyssz - physsz);
1947 				physsz = hwphyssz;
1948 				phys_avail_count++;
1949 				dump_avail[j] = phys_avail[j];
1950 				dump_avail[j + 1] = phys_avail[j + 1];
1951 			}
1952 			break;
1953 		}
1954 		phys_avail[j] = regions[i].mr_start;
1955 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
1956 		dump_avail[j] = phys_avail[j];
1957 		dump_avail[j + 1] = phys_avail[j + 1];
1958 
1959 		phys_avail_count++;
1960 		physsz += regions[i].mr_size;
1961 		j += 2;
1962 	}
1963 
1964 	/* Check for overlap with the kernel and exception vectors */
1965 	rm_pavail = 0;
1966 	for (j = 0; j < 2 * phys_avail_count; j+=2) {
1967 		if (phys_avail[j] < EXC_LAST)
1968 			phys_avail[j] += EXC_LAST;
1969 
1970 		if (phys_avail[j] >= kpstart &&
1971 		    phys_avail[j + 1] <= kpend) {
1972 			phys_avail[j] = phys_avail[j + 1] = ~0;
1973 			rm_pavail++;
1974 			continue;
1975 		}
1976 
1977 		if (kpstart >= phys_avail[j] &&
1978 		    kpstart < phys_avail[j + 1]) {
1979 			if (kpend < phys_avail[j + 1]) {
1980 				phys_avail[2 * phys_avail_count] =
1981 				    (kpend & ~PAGE_MASK) + PAGE_SIZE;
1982 				phys_avail[2 * phys_avail_count + 1] =
1983 				    phys_avail[j + 1];
1984 				phys_avail_count++;
1985 			}
1986 
1987 			phys_avail[j + 1] = kpstart & ~PAGE_MASK;
1988 		}
1989 
1990 		if (kpend >= phys_avail[j] &&
1991 		    kpend < phys_avail[j + 1]) {
1992 			if (kpstart > phys_avail[j]) {
1993 				phys_avail[2 * phys_avail_count] = phys_avail[j];
1994 				phys_avail[2 * phys_avail_count + 1] =
1995 				    kpstart & ~PAGE_MASK;
1996 				phys_avail_count++;
1997 			}
1998 
1999 			phys_avail[j] = (kpend & ~PAGE_MASK) +
2000 			    PAGE_SIZE;
2001 		}
2002 	}
2003 	qsort(phys_avail, 2 * phys_avail_count, sizeof(phys_avail[0]), pa_cmp);
2004 	for (i = 0; i < 2 * phys_avail_count; i++)
2005 		phys_avail_debug[i] = phys_avail[i];
2006 
2007 	/* Remove physical available regions marked for removal (~0) */
2008 	if (rm_pavail) {
2009 		phys_avail_count -= rm_pavail;
2010 		for (i = 2 * phys_avail_count;
2011 		     i < 2*(phys_avail_count + rm_pavail); i+=2)
2012 			phys_avail[i] = phys_avail[i + 1] = 0;
2013 	}
2014 	if (bootverbose) {
2015 		printf("phys_avail ranges after filtering:\n");
2016 		for (j = 0; j < 2 * phys_avail_count; j+=2)
2017 			printf("phys_avail[%d]=%08lx - phys_avail[%d]=%08lx\n",
2018 				   j, phys_avail[j], j + 1, phys_avail[j + 1]);
2019 	}
2020 	physmem = btoc(physsz);
2021 
2022 	/* XXX assume we're running non-virtualized and
2023 	 * we don't support BHYVE
2024 	 */
2025 	if (isa3_pid_bits == 0)
2026 		isa3_pid_bits = 20;
2027 	parttab_phys = moea64_bootstrap_alloc(PARTTAB_SIZE, PARTTAB_SIZE);
2028 	validate_addr(parttab_phys, PARTTAB_SIZE);
2029 	for (int i = 0; i < PARTTAB_SIZE/PAGE_SIZE; i++)
2030 		pagezero(PHYS_TO_DMAP(parttab_phys + i * PAGE_SIZE));
2031 
2032 	proctab_size = 1UL << PROCTAB_SIZE_SHIFT;
2033 	proctab0pa = moea64_bootstrap_alloc(proctab_size, proctab_size);
2034 	validate_addr(proctab0pa, proctab_size);
2035 	for (int i = 0; i < proctab_size/PAGE_SIZE; i++)
2036 		pagezero(PHYS_TO_DMAP(proctab0pa + i * PAGE_SIZE));
2037 
2038 	mmu_radix_setup_pagetables(hwphyssz);
2039 }
2040 
2041 static void
2042 mmu_radix_late_bootstrap(vm_offset_t start, vm_offset_t end)
2043 {
2044 	int		i;
2045 	vm_paddr_t	pa;
2046 	void		*dpcpu;
2047 	vm_offset_t va;
2048 
2049 	/*
2050 	 * Set up the Open Firmware pmap and add its mappings if not in real
2051 	 * mode.
2052 	 */
2053 	if (bootverbose)
2054 		printf("%s enter\n", __func__);
2055 
2056 	/*
2057 	 * Calculate the last available physical address, and reserve the
2058 	 * vm_page_array (upper bound).
2059 	 */
2060 	Maxmem = 0;
2061 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
2062 		Maxmem = MAX(Maxmem, powerpc_btop(phys_avail[i + 1]));
2063 
2064 	/*
2065 	 * Set the start and end of kva.
2066 	 */
2067 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
2068 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
2069 
2070 	/*
2071 	 * Remap any early IO mappings (console framebuffer, etc.)
2072 	 */
2073 	bs_remap_earlyboot();
2074 
2075 	/*
2076 	 * Allocate a kernel stack with a guard page for thread0 and map it
2077 	 * into the kernel page map.
2078 	 */
2079 	pa = allocpages(kstack_pages);
2080 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
2081 	virtual_avail = va + kstack_pages * PAGE_SIZE;
2082 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
2083 	thread0.td_kstack = va;
2084 	for (i = 0; i < kstack_pages; i++) {
2085 		mmu_radix_kenter(va, pa);
2086 		pa += PAGE_SIZE;
2087 		va += PAGE_SIZE;
2088 	}
2089 	thread0.td_kstack_pages = kstack_pages;
2090 
2091 	/*
2092 	 * Allocate virtual address space for the message buffer.
2093 	 */
2094 	pa = msgbuf_phys = allocpages((msgbufsize + PAGE_MASK)  >> PAGE_SHIFT);
2095 	msgbufp = (struct msgbuf *)PHYS_TO_DMAP(pa);
2096 
2097 	/*
2098 	 * Allocate virtual address space for the dynamic percpu area.
2099 	 */
2100 	pa = allocpages(DPCPU_SIZE >> PAGE_SHIFT);
2101 	dpcpu = (void *)PHYS_TO_DMAP(pa);
2102 	dpcpu_init(dpcpu, curcpu);
2103 
2104 	crashdumpmap = (caddr_t)virtual_avail;
2105 	virtual_avail += MAXDUMPPGS * PAGE_SIZE;
2106 
2107 	/*
2108 	 * Reserve some special page table entries/VA space for temporary
2109 	 * mapping of pages.
2110 	 */
2111 }
2112 
2113 static void
2114 mmu_parttab_init(void)
2115 {
2116 	uint64_t ptcr;
2117 
2118 	isa3_parttab = (struct pate *)PHYS_TO_DMAP(parttab_phys);
2119 
2120 	if (bootverbose)
2121 		printf("%s parttab: %p\n", __func__, isa3_parttab);
2122 	ptcr = parttab_phys | (PARTTAB_SIZE_SHIFT-12);
2123 	if (bootverbose)
2124 		printf("setting ptcr %lx\n", ptcr);
2125 	mtspr(SPR_PTCR, ptcr);
2126 }
2127 
2128 static void
2129 mmu_parttab_update(uint64_t lpid, uint64_t pagetab, uint64_t proctab)
2130 {
2131 	uint64_t prev;
2132 
2133 	if (bootverbose)
2134 		printf("%s isa3_parttab %p lpid %lx pagetab %lx proctab %lx\n", __func__, isa3_parttab,
2135 			   lpid, pagetab, proctab);
2136 	prev = be64toh(isa3_parttab[lpid].pagetab);
2137 	isa3_parttab[lpid].pagetab = htobe64(pagetab);
2138 	isa3_parttab[lpid].proctab = htobe64(proctab);
2139 
2140 	if (prev & PARTTAB_HR) {
2141 		__asm __volatile(PPC_TLBIE_5(%0,%1,2,0,1) : :
2142 			     "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
2143 		__asm __volatile(PPC_TLBIE_5(%0,%1,2,1,1) : :
2144 			     "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
2145 	} else {
2146 		__asm __volatile(PPC_TLBIE_5(%0,%1,2,0,0) : :
2147 			     "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
2148 	}
2149 	ttusync();
2150 }
2151 
2152 static void
2153 mmu_radix_parttab_init(void)
2154 {
2155 	uint64_t pagetab;
2156 
2157 	mmu_parttab_init();
2158 	pagetab = RTS_SIZE | DMAP_TO_PHYS((vm_offset_t)kernel_pmap->pm_pml1) | \
2159 		         RADIX_PGD_INDEX_SHIFT | PARTTAB_HR;
2160 	mmu_parttab_update(0, pagetab, 0);
2161 }
2162 
2163 static void
2164 mmu_radix_proctab_register(vm_paddr_t proctabpa, uint64_t table_size)
2165 {
2166 	uint64_t pagetab, proctab;
2167 
2168 	pagetab = be64toh(isa3_parttab[0].pagetab);
2169 	proctab = proctabpa | table_size | PARTTAB_GR;
2170 	mmu_parttab_update(0, pagetab, proctab);
2171 }
2172 
2173 static void
2174 mmu_radix_proctab_init(void)
2175 {
2176 
2177 	isa3_base_pid = 1;
2178 
2179 	isa3_proctab = (void*)PHYS_TO_DMAP(proctab0pa);
2180 	isa3_proctab->proctab0 =
2181 	    htobe64(RTS_SIZE | DMAP_TO_PHYS((vm_offset_t)kernel_pmap->pm_pml1) |
2182 		RADIX_PGD_INDEX_SHIFT);
2183 
2184 	mmu_radix_proctab_register(proctab0pa, PROCTAB_SIZE_SHIFT - 12);
2185 
2186 	__asm __volatile("ptesync" : : : "memory");
2187 	__asm __volatile(PPC_TLBIE_5(%0,%1,2,1,1) : :
2188 		     "r" (TLBIEL_INVAL_SET_LPID), "r" (0));
2189 	__asm __volatile("eieio; tlbsync; ptesync" : : : "memory");
2190 	if (bootverbose)
2191 		printf("process table %p and kernel radix PDE: %p\n",
2192 			   isa3_proctab, kernel_pmap->pm_pml1);
2193 	mtmsr(mfmsr() | PSL_DR );
2194 	mtmsr(mfmsr() &  ~PSL_DR);
2195 	kernel_pmap->pm_pid = isa3_base_pid;
2196 	isa3_base_pid++;
2197 }
2198 
2199 void
2200 mmu_radix_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
2201     int advice)
2202 {
2203 	struct rwlock *lock;
2204 	pml1_entry_t *l1e;
2205 	pml2_entry_t *l2e;
2206 	pml3_entry_t oldl3e, *l3e;
2207 	pt_entry_t *pte;
2208 	vm_offset_t va, va_next;
2209 	vm_page_t m;
2210 	bool anychanged;
2211 
2212 	if (advice != MADV_DONTNEED && advice != MADV_FREE)
2213 		return;
2214 	anychanged = false;
2215 	PMAP_LOCK(pmap);
2216 	for (; sva < eva; sva = va_next) {
2217 		l1e = pmap_pml1e(pmap, sva);
2218 		if ((be64toh(*l1e) & PG_V) == 0) {
2219 			va_next = (sva + L1_PAGE_SIZE) & ~L1_PAGE_MASK;
2220 			if (va_next < sva)
2221 				va_next = eva;
2222 			continue;
2223 		}
2224 		l2e = pmap_l1e_to_l2e(l1e, sva);
2225 		if ((be64toh(*l2e) & PG_V) == 0) {
2226 			va_next = (sva + L2_PAGE_SIZE) & ~L2_PAGE_MASK;
2227 			if (va_next < sva)
2228 				va_next = eva;
2229 			continue;
2230 		}
2231 		va_next = (sva + L3_PAGE_SIZE) & ~L3_PAGE_MASK;
2232 		if (va_next < sva)
2233 			va_next = eva;
2234 		l3e = pmap_l2e_to_l3e(l2e, sva);
2235 		oldl3e = be64toh(*l3e);
2236 		if ((oldl3e & PG_V) == 0)
2237 			continue;
2238 		else if ((oldl3e & RPTE_LEAF) != 0) {
2239 			if ((oldl3e & PG_MANAGED) == 0)
2240 				continue;
2241 			lock = NULL;
2242 			if (!pmap_demote_l3e_locked(pmap, l3e, sva, &lock)) {
2243 				if (lock != NULL)
2244 					rw_wunlock(lock);
2245 
2246 				/*
2247 				 * The large page mapping was destroyed.
2248 				 */
2249 				continue;
2250 			}
2251 
2252 			/*
2253 			 * Unless the page mappings are wired, remove the
2254 			 * mapping to a single page so that a subsequent
2255 			 * access may repromote.  Choosing the last page
2256 			 * within the address range [sva, min(va_next, eva))
2257 			 * generally results in more repromotions.  Since the
2258 			 * underlying page table page is fully populated, this
2259 			 * removal never frees a page table page.
2260 			 */
2261 			if ((oldl3e & PG_W) == 0) {
2262 				va = eva;
2263 				if (va > va_next)
2264 					va = va_next;
2265 				va -= PAGE_SIZE;
2266 				KASSERT(va >= sva,
2267 				    ("mmu_radix_advise: no address gap"));
2268 				pte = pmap_l3e_to_pte(l3e, va);
2269 				KASSERT((be64toh(*pte) & PG_V) != 0,
2270 				    ("pmap_advise: invalid PTE"));
2271 				pmap_remove_pte(pmap, pte, va, be64toh(*l3e), NULL,
2272 				    &lock);
2273 				anychanged = true;
2274 			}
2275 			if (lock != NULL)
2276 				rw_wunlock(lock);
2277 		}
2278 		if (va_next > eva)
2279 			va_next = eva;
2280 		va = va_next;
2281 		for (pte = pmap_l3e_to_pte(l3e, sva); sva != va_next;
2282 			 pte++, sva += PAGE_SIZE) {
2283 			MPASS(pte == pmap_pte(pmap, sva));
2284 
2285 			if ((be64toh(*pte) & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V))
2286 				goto maybe_invlrng;
2287 			else if ((be64toh(*pte) & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
2288 				if (advice == MADV_DONTNEED) {
2289 					/*
2290 					 * Future calls to pmap_is_modified()
2291 					 * can be avoided by making the page
2292 					 * dirty now.
2293 					 */
2294 					m = PHYS_TO_VM_PAGE(be64toh(*pte) & PG_FRAME);
2295 					vm_page_dirty(m);
2296 				}
2297 				atomic_clear_long(pte, htobe64(PG_M | PG_A));
2298 			} else if ((be64toh(*pte) & PG_A) != 0)
2299 				atomic_clear_long(pte, htobe64(PG_A));
2300 			else
2301 				goto maybe_invlrng;
2302 			anychanged = true;
2303 			continue;
2304 maybe_invlrng:
2305 			if (va != va_next) {
2306 				anychanged = true;
2307 				va = va_next;
2308 			}
2309 		}
2310 		if (va != va_next)
2311 			anychanged = true;
2312 	}
2313 	if (anychanged)
2314 		pmap_invalidate_all(pmap);
2315 	PMAP_UNLOCK(pmap);
2316 }
2317 
2318 /*
2319  * Routines used in machine-dependent code
2320  */
2321 static void
2322 mmu_radix_bootstrap(vm_offset_t start, vm_offset_t end)
2323 {
2324 	uint64_t lpcr;
2325 
2326 	if (bootverbose)
2327 		printf("%s\n", __func__);
2328 	hw_direct_map = 1;
2329 	mmu_radix_early_bootstrap(start, end);
2330 	if (bootverbose)
2331 		printf("early bootstrap complete\n");
2332 	if (powernv_enabled) {
2333 		lpcr = mfspr(SPR_LPCR);
2334 		mtspr(SPR_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
2335 		mmu_radix_parttab_init();
2336 		mmu_radix_init_amor();
2337 		if (bootverbose)
2338 			printf("powernv init complete\n");
2339 	}
2340 	mmu_radix_init_iamr();
2341 	mmu_radix_proctab_init();
2342 	mmu_radix_pid_set(kernel_pmap);
2343 	/* XXX assume CPU_FTR_HVMODE */
2344 	mmu_radix_tlbiel_flush(TLB_INVAL_SCOPE_GLOBAL);
2345 
2346 	mmu_radix_late_bootstrap(start, end);
2347 	numa_mem_regions(&numa_pregions, &numa_pregions_sz);
2348 	if (bootverbose)
2349 		printf("%s done\n", __func__);
2350 	pmap_bootstrapped = 1;
2351 	dmaplimit = roundup2(powerpc_ptob(Maxmem), L2_PAGE_SIZE);
2352 	PCPU_SET(flags, PCPU_GET(flags) | PC_FLAG_NOSRS);
2353 }
2354 
2355 static void
2356 mmu_radix_cpu_bootstrap(int ap)
2357 {
2358 	uint64_t lpcr;
2359 	uint64_t ptcr;
2360 
2361 	if (powernv_enabled) {
2362 		lpcr = mfspr(SPR_LPCR);
2363 		mtspr(SPR_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
2364 
2365 		ptcr = parttab_phys | (PARTTAB_SIZE_SHIFT-12);
2366 		mtspr(SPR_PTCR, ptcr);
2367 		mmu_radix_init_amor();
2368 	}
2369 	mmu_radix_init_iamr();
2370 	mmu_radix_pid_set(kernel_pmap);
2371 	mmu_radix_tlbiel_flush(TLB_INVAL_SCOPE_GLOBAL);
2372 }
2373 
2374 static SYSCTL_NODE(_vm_pmap, OID_AUTO, l3e, CTLFLAG_RD, 0,
2375     "2MB page mapping counters");
2376 
2377 static COUNTER_U64_DEFINE_EARLY(pmap_l3e_demotions);
2378 SYSCTL_COUNTER_U64(_vm_pmap_l3e, OID_AUTO, demotions, CTLFLAG_RD,
2379     &pmap_l3e_demotions, "2MB page demotions");
2380 
2381 static COUNTER_U64_DEFINE_EARLY(pmap_l3e_mappings);
2382 SYSCTL_COUNTER_U64(_vm_pmap_l3e, OID_AUTO, mappings, CTLFLAG_RD,
2383     &pmap_l3e_mappings, "2MB page mappings");
2384 
2385 static COUNTER_U64_DEFINE_EARLY(pmap_l3e_p_failures);
2386 SYSCTL_COUNTER_U64(_vm_pmap_l3e, OID_AUTO, p_failures, CTLFLAG_RD,
2387     &pmap_l3e_p_failures, "2MB page promotion failures");
2388 
2389 static COUNTER_U64_DEFINE_EARLY(pmap_l3e_promotions);
2390 SYSCTL_COUNTER_U64(_vm_pmap_l3e, OID_AUTO, promotions, CTLFLAG_RD,
2391     &pmap_l3e_promotions, "2MB page promotions");
2392 
2393 static SYSCTL_NODE(_vm_pmap, OID_AUTO, l2e, CTLFLAG_RD, 0,
2394     "1GB page mapping counters");
2395 
2396 static COUNTER_U64_DEFINE_EARLY(pmap_l2e_demotions);
2397 SYSCTL_COUNTER_U64(_vm_pmap_l2e, OID_AUTO, demotions, CTLFLAG_RD,
2398     &pmap_l2e_demotions, "1GB page demotions");
2399 
2400 void
2401 mmu_radix_clear_modify(vm_page_t m)
2402 {
2403 	struct md_page *pvh;
2404 	pmap_t pmap;
2405 	pv_entry_t next_pv, pv;
2406 	pml3_entry_t oldl3e, *l3e;
2407 	pt_entry_t oldpte, *pte;
2408 	struct rwlock *lock;
2409 	vm_offset_t va;
2410 	int md_gen, pvh_gen;
2411 
2412 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2413 	    ("pmap_clear_modify: page %p is not managed", m));
2414 	vm_page_assert_busied(m);
2415 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
2416 
2417 	/*
2418 	 * If the page is not PGA_WRITEABLE, then no PTEs can have PG_M set.
2419 	 * If the object containing the page is locked and the page is not
2420 	 * exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
2421 	 */
2422 	if ((m->a.flags & PGA_WRITEABLE) == 0)
2423 		return;
2424 	pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
2425 	    pa_to_pvh(VM_PAGE_TO_PHYS(m));
2426 	lock = VM_PAGE_TO_PV_LIST_LOCK(m);
2427 	rw_wlock(lock);
2428 restart:
2429 	TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_link, next_pv) {
2430 		pmap = PV_PMAP(pv);
2431 		if (!PMAP_TRYLOCK(pmap)) {
2432 			pvh_gen = pvh->pv_gen;
2433 			rw_wunlock(lock);
2434 			PMAP_LOCK(pmap);
2435 			rw_wlock(lock);
2436 			if (pvh_gen != pvh->pv_gen) {
2437 				PMAP_UNLOCK(pmap);
2438 				goto restart;
2439 			}
2440 		}
2441 		va = pv->pv_va;
2442 		l3e = pmap_pml3e(pmap, va);
2443 		oldl3e = be64toh(*l3e);
2444 		if ((oldl3e & PG_RW) != 0 &&
2445 		    pmap_demote_l3e_locked(pmap, l3e, va, &lock) &&
2446 		    (oldl3e & PG_W) == 0) {
2447 			/*
2448 			 * Write protect the mapping to a
2449 			 * single page so that a subsequent
2450 			 * write access may repromote.
2451 			 */
2452 			va += VM_PAGE_TO_PHYS(m) - (oldl3e &
2453 			    PG_PS_FRAME);
2454 			pte = pmap_l3e_to_pte(l3e, va);
2455 			oldpte = be64toh(*pte);
2456 			while (!atomic_cmpset_long(pte,
2457 			    htobe64(oldpte),
2458 				htobe64((oldpte | RPTE_EAA_R) & ~(PG_M | PG_RW))))
2459 				   oldpte = be64toh(*pte);
2460 			vm_page_dirty(m);
2461 			pmap_invalidate_page(pmap, va);
2462 		}
2463 		PMAP_UNLOCK(pmap);
2464 	}
2465 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) {
2466 		pmap = PV_PMAP(pv);
2467 		if (!PMAP_TRYLOCK(pmap)) {
2468 			md_gen = m->md.pv_gen;
2469 			pvh_gen = pvh->pv_gen;
2470 			rw_wunlock(lock);
2471 			PMAP_LOCK(pmap);
2472 			rw_wlock(lock);
2473 			if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
2474 				PMAP_UNLOCK(pmap);
2475 				goto restart;
2476 			}
2477 		}
2478 		l3e = pmap_pml3e(pmap, pv->pv_va);
2479 		KASSERT((be64toh(*l3e) & RPTE_LEAF) == 0, ("pmap_clear_modify: found"
2480 		    " a 2mpage in page %p's pv list", m));
2481 		pte = pmap_l3e_to_pte(l3e, pv->pv_va);
2482 		if ((be64toh(*pte) & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
2483 			atomic_clear_long(pte, htobe64(PG_M));
2484 			pmap_invalidate_page(pmap, pv->pv_va);
2485 		}
2486 		PMAP_UNLOCK(pmap);
2487 	}
2488 	rw_wunlock(lock);
2489 }
2490 
2491 void
2492 mmu_radix_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2493     vm_size_t len, vm_offset_t src_addr)
2494 {
2495 	struct rwlock *lock;
2496 	struct spglist free;
2497 	vm_offset_t addr;
2498 	vm_offset_t end_addr = src_addr + len;
2499 	vm_offset_t va_next;
2500 	vm_page_t dst_pdpg, dstmpte, srcmpte;
2501 	bool invalidate_all;
2502 
2503 	CTR6(KTR_PMAP,
2504 	    "%s(dst_pmap=%p, src_pmap=%p, dst_addr=%lx, len=%lu, src_addr=%lx)\n",
2505 	    __func__, dst_pmap, src_pmap, dst_addr, len, src_addr);
2506 
2507 	if (dst_addr != src_addr)
2508 		return;
2509 	lock = NULL;
2510 	invalidate_all = false;
2511 	if (dst_pmap < src_pmap) {
2512 		PMAP_LOCK(dst_pmap);
2513 		PMAP_LOCK(src_pmap);
2514 	} else {
2515 		PMAP_LOCK(src_pmap);
2516 		PMAP_LOCK(dst_pmap);
2517 	}
2518 
2519 	for (addr = src_addr; addr < end_addr; addr = va_next) {
2520 		pml1_entry_t *l1e;
2521 		pml2_entry_t *l2e;
2522 		pml3_entry_t srcptepaddr, *l3e;
2523 		pt_entry_t *src_pte, *dst_pte;
2524 
2525 		l1e = pmap_pml1e(src_pmap, addr);
2526 		if ((be64toh(*l1e) & PG_V) == 0) {
2527 			va_next = (addr + L1_PAGE_SIZE) & ~L1_PAGE_MASK;
2528 			if (va_next < addr)
2529 				va_next = end_addr;
2530 			continue;
2531 		}
2532 
2533 		l2e = pmap_l1e_to_l2e(l1e, addr);
2534 		if ((be64toh(*l2e) & PG_V) == 0) {
2535 			va_next = (addr + L2_PAGE_SIZE) & ~L2_PAGE_MASK;
2536 			if (va_next < addr)
2537 				va_next = end_addr;
2538 			continue;
2539 		}
2540 
2541 		va_next = (addr + L3_PAGE_SIZE) & ~L3_PAGE_MASK;
2542 		if (va_next < addr)
2543 			va_next = end_addr;
2544 
2545 		l3e = pmap_l2e_to_l3e(l2e, addr);
2546 		srcptepaddr = be64toh(*l3e);
2547 		if (srcptepaddr == 0)
2548 			continue;
2549 
2550 		if (srcptepaddr & RPTE_LEAF) {
2551 			if ((addr & L3_PAGE_MASK) != 0 ||
2552 			    addr + L3_PAGE_SIZE > end_addr)
2553 				continue;
2554 			dst_pdpg = pmap_allocl3e(dst_pmap, addr, NULL);
2555 			if (dst_pdpg == NULL)
2556 				break;
2557 			l3e = (pml3_entry_t *)
2558 			    PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dst_pdpg));
2559 			l3e = &l3e[pmap_pml3e_index(addr)];
2560 			if (be64toh(*l3e) == 0 && ((srcptepaddr & PG_MANAGED) == 0 ||
2561 			    pmap_pv_insert_l3e(dst_pmap, addr, srcptepaddr,
2562 			    PMAP_ENTER_NORECLAIM, &lock))) {
2563 				*l3e = htobe64(srcptepaddr & ~PG_W);
2564 				pmap_resident_count_inc(dst_pmap,
2565 				    L3_PAGE_SIZE / PAGE_SIZE);
2566 				counter_u64_add(pmap_l3e_mappings, 1);
2567 			} else
2568 				dst_pdpg->ref_count--;
2569 			continue;
2570 		}
2571 
2572 		srcptepaddr &= PG_FRAME;
2573 		srcmpte = PHYS_TO_VM_PAGE(srcptepaddr);
2574 		KASSERT(srcmpte->ref_count > 0,
2575 		    ("pmap_copy: source page table page is unused"));
2576 
2577 		if (va_next > end_addr)
2578 			va_next = end_addr;
2579 
2580 		src_pte = (pt_entry_t *)PHYS_TO_DMAP(srcptepaddr);
2581 		src_pte = &src_pte[pmap_pte_index(addr)];
2582 		dstmpte = NULL;
2583 		while (addr < va_next) {
2584 			pt_entry_t ptetemp;
2585 			ptetemp = be64toh(*src_pte);
2586 			/*
2587 			 * we only virtual copy managed pages
2588 			 */
2589 			if ((ptetemp & PG_MANAGED) != 0) {
2590 				if (dstmpte != NULL &&
2591 				    dstmpte->pindex == pmap_l3e_pindex(addr))
2592 					dstmpte->ref_count++;
2593 				else if ((dstmpte = pmap_allocpte(dst_pmap,
2594 				    addr, NULL)) == NULL)
2595 					goto out;
2596 				dst_pte = (pt_entry_t *)
2597 				    PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dstmpte));
2598 				dst_pte = &dst_pte[pmap_pte_index(addr)];
2599 				if (be64toh(*dst_pte) == 0 &&
2600 				    pmap_try_insert_pv_entry(dst_pmap, addr,
2601 				    PHYS_TO_VM_PAGE(ptetemp & PG_FRAME),
2602 				    &lock)) {
2603 					/*
2604 					 * Clear the wired, modified, and
2605 					 * accessed (referenced) bits
2606 					 * during the copy.
2607 					 */
2608 					*dst_pte = htobe64(ptetemp & ~(PG_W | PG_M |
2609 					    PG_A));
2610 					pmap_resident_count_inc(dst_pmap, 1);
2611 				} else {
2612 					SLIST_INIT(&free);
2613 					if (pmap_unwire_ptp(dst_pmap, addr,
2614 					    dstmpte, &free)) {
2615 						/*
2616 						 * Although "addr" is not
2617 						 * mapped, paging-structure
2618 						 * caches could nonetheless
2619 						 * have entries that refer to
2620 						 * the freed page table pages.
2621 						 * Invalidate those entries.
2622 						 */
2623 						invalidate_all = true;
2624 						vm_page_free_pages_toq(&free,
2625 						    true);
2626 					}
2627 					goto out;
2628 				}
2629 				if (dstmpte->ref_count >= srcmpte->ref_count)
2630 					break;
2631 			}
2632 			addr += PAGE_SIZE;
2633 			if (__predict_false((addr & L3_PAGE_MASK) == 0))
2634 				src_pte = pmap_pte(src_pmap, addr);
2635 			else
2636 				src_pte++;
2637 		}
2638 	}
2639 out:
2640 	if (invalidate_all)
2641 		pmap_invalidate_all(dst_pmap);
2642 	if (lock != NULL)
2643 		rw_wunlock(lock);
2644 	PMAP_UNLOCK(src_pmap);
2645 	PMAP_UNLOCK(dst_pmap);
2646 }
2647 
2648 static void
2649 mmu_radix_copy_page(vm_page_t msrc, vm_page_t mdst)
2650 {
2651 	vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc));
2652 	vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst));
2653 
2654 	CTR3(KTR_PMAP, "%s(%p, %p)", __func__, src, dst);
2655 	/*
2656 	 * XXX slow
2657 	 */
2658 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
2659 }
2660 
2661 static void
2662 mmu_radix_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[],
2663     vm_offset_t b_offset, int xfersize)
2664 {
2665         void *a_cp, *b_cp;
2666         vm_offset_t a_pg_offset, b_pg_offset;
2667         int cnt;
2668 
2669 	CTR6(KTR_PMAP, "%s(%p, %#x, %p, %#x, %#x)", __func__, ma,
2670 	    a_offset, mb, b_offset, xfersize);
2671 
2672         while (xfersize > 0) {
2673                 a_pg_offset = a_offset & PAGE_MASK;
2674                 cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
2675                 a_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
2676                     VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])) +
2677                     a_pg_offset;
2678                 b_pg_offset = b_offset & PAGE_MASK;
2679                 cnt = min(cnt, PAGE_SIZE - b_pg_offset);
2680                 b_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
2681                     VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])) +
2682                     b_pg_offset;
2683                 bcopy(a_cp, b_cp, cnt);
2684                 a_offset += cnt;
2685                 b_offset += cnt;
2686                 xfersize -= cnt;
2687         }
2688 }
2689 
2690 #if VM_NRESERVLEVEL > 0
2691 /*
2692  * Tries to promote the 512, contiguous 4KB page mappings that are within a
2693  * single page table page (PTP) to a single 2MB page mapping.  For promotion
2694  * to occur, two conditions must be met: (1) the 4KB page mappings must map
2695  * aligned, contiguous physical memory and (2) the 4KB page mappings must have
2696  * identical characteristics.
2697  */
2698 static int
2699 pmap_promote_l3e(pmap_t pmap, pml3_entry_t *pde, vm_offset_t va,
2700     struct rwlock **lockp)
2701 {
2702 	pml3_entry_t newpde;
2703 	pt_entry_t *firstpte, oldpte, pa, *pte;
2704 	vm_page_t mpte;
2705 
2706 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
2707 
2708 	/*
2709 	 * Examine the first PTE in the specified PTP.  Abort if this PTE is
2710 	 * either invalid, unused, or does not map the first 4KB physical page
2711 	 * within a 2MB page.
2712 	 */
2713 	firstpte = (pt_entry_t *)PHYS_TO_DMAP(be64toh(*pde) & PG_FRAME);
2714 setpde:
2715 	newpde = be64toh(*firstpte);
2716 	if ((newpde & ((PG_FRAME & L3_PAGE_MASK) | PG_A | PG_V)) != (PG_A | PG_V)) {
2717 		CTR2(KTR_PMAP, "pmap_promote_l3e: failure for va %#lx"
2718 		    " in pmap %p", va, pmap);
2719 		goto fail;
2720 	}
2721 	if ((newpde & (PG_M | PG_RW)) == PG_RW) {
2722 		/*
2723 		 * When PG_M is already clear, PG_RW can be cleared without
2724 		 * a TLB invalidation.
2725 		 */
2726 		if (!atomic_cmpset_long(firstpte, htobe64(newpde), htobe64((newpde | RPTE_EAA_R) & ~RPTE_EAA_W)))
2727 			goto setpde;
2728 		newpde &= ~RPTE_EAA_W;
2729 	}
2730 
2731 	/*
2732 	 * Examine each of the other PTEs in the specified PTP.  Abort if this
2733 	 * PTE maps an unexpected 4KB physical page or does not have identical
2734 	 * characteristics to the first PTE.
2735 	 */
2736 	pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + L3_PAGE_SIZE - PAGE_SIZE;
2737 	for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) {
2738 setpte:
2739 		oldpte = be64toh(*pte);
2740 		if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) {
2741 			CTR2(KTR_PMAP, "pmap_promote_l3e: failure for va %#lx"
2742 			    " in pmap %p", va, pmap);
2743 			goto fail;
2744 		}
2745 		if ((oldpte & (PG_M | PG_RW)) == PG_RW) {
2746 			/*
2747 			 * When PG_M is already clear, PG_RW can be cleared
2748 			 * without a TLB invalidation.
2749 			 */
2750 			if (!atomic_cmpset_long(pte, htobe64(oldpte), htobe64((oldpte | RPTE_EAA_R) & ~RPTE_EAA_W)))
2751 				goto setpte;
2752 			oldpte &= ~RPTE_EAA_W;
2753 			CTR2(KTR_PMAP, "pmap_promote_l3e: protect for va %#lx"
2754 			    " in pmap %p", (oldpte & PG_FRAME & L3_PAGE_MASK) |
2755 			    (va & ~L3_PAGE_MASK), pmap);
2756 		}
2757 		if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) {
2758 			CTR2(KTR_PMAP, "pmap_promote_l3e: failure for va %#lx"
2759 			    " in pmap %p", va, pmap);
2760 			goto fail;
2761 		}
2762 		pa -= PAGE_SIZE;
2763 	}
2764 
2765 	/*
2766 	 * Save the page table page in its current state until the PDE
2767 	 * mapping the superpage is demoted by pmap_demote_pde() or
2768 	 * destroyed by pmap_remove_pde().
2769 	 */
2770 	mpte = PHYS_TO_VM_PAGE(be64toh(*pde) & PG_FRAME);
2771 	KASSERT(mpte >= vm_page_array &&
2772 	    mpte < &vm_page_array[vm_page_array_size],
2773 	    ("pmap_promote_l3e: page table page is out of range"));
2774 	KASSERT(mpte->pindex == pmap_l3e_pindex(va),
2775 	    ("pmap_promote_l3e: page table page's pindex is wrong"));
2776 	if (pmap_insert_pt_page(pmap, mpte)) {
2777 		CTR2(KTR_PMAP,
2778 		    "pmap_promote_l3e: failure for va %#lx in pmap %p", va,
2779 		    pmap);
2780 		goto fail;
2781 	}
2782 
2783 	/*
2784 	 * Promote the pv entries.
2785 	 */
2786 	if ((newpde & PG_MANAGED) != 0)
2787 		pmap_pv_promote_l3e(pmap, va, newpde & PG_PS_FRAME, lockp);
2788 
2789 	pte_store(pde, PG_PROMOTED | newpde);
2790 	ptesync();
2791 	counter_u64_add(pmap_l3e_promotions, 1);
2792 	CTR2(KTR_PMAP, "pmap_promote_l3e: success for va %#lx"
2793 	    " in pmap %p", va, pmap);
2794 	return (0);
2795  fail:
2796 	counter_u64_add(pmap_l3e_p_failures, 1);
2797 	return (KERN_FAILURE);
2798 }
2799 #endif /* VM_NRESERVLEVEL > 0 */
2800 
2801 int
2802 mmu_radix_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
2803     vm_prot_t prot, u_int flags, int8_t psind)
2804 {
2805 	struct rwlock *lock;
2806 	pml3_entry_t *l3e;
2807 	pt_entry_t *pte;
2808 	pt_entry_t newpte, origpte;
2809 	pv_entry_t pv;
2810 	vm_paddr_t opa, pa;
2811 	vm_page_t mpte, om;
2812 	int rv, retrycount;
2813 	boolean_t nosleep, invalidate_all, invalidate_page;
2814 
2815 	va = trunc_page(va);
2816 	retrycount = 0;
2817 	invalidate_page = invalidate_all = false;
2818 	CTR6(KTR_PMAP, "pmap_enter(%p, %#lx, %p, %#x, %#x, %d)", pmap, va,
2819 	    m, prot, flags, psind);
2820 	KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig"));
2821 	KASSERT((m->oflags & VPO_UNMANAGED) != 0 || !VA_IS_CLEANMAP(va),
2822 	    ("pmap_enter: managed mapping within the clean submap"));
2823 	if ((m->oflags & VPO_UNMANAGED) == 0)
2824 		VM_PAGE_OBJECT_BUSY_ASSERT(m);
2825 
2826 	KASSERT((flags & PMAP_ENTER_RESERVED) == 0,
2827 	    ("pmap_enter: flags %u has reserved bits set", flags));
2828 	pa = VM_PAGE_TO_PHYS(m);
2829 	newpte = (pt_entry_t)(pa | PG_A | PG_V | RPTE_LEAF);
2830 	if ((flags & VM_PROT_WRITE) != 0)
2831 		newpte |= PG_M;
2832 	if ((flags & VM_PROT_READ) != 0)
2833 		newpte |= PG_A;
2834 	if (prot & VM_PROT_READ)
2835 		newpte |= RPTE_EAA_R;
2836 	if ((prot & VM_PROT_WRITE) != 0)
2837 		newpte |= RPTE_EAA_W;
2838 	KASSERT((newpte & (PG_M | PG_RW)) != PG_M,
2839 	    ("pmap_enter: flags includes VM_PROT_WRITE but prot doesn't"));
2840 
2841 	if (prot & VM_PROT_EXECUTE)
2842 		newpte |= PG_X;
2843 	if ((flags & PMAP_ENTER_WIRED) != 0)
2844 		newpte |= PG_W;
2845 	if (va >= DMAP_MIN_ADDRESS)
2846 		newpte |= RPTE_EAA_P;
2847 	newpte |= pmap_cache_bits(m->md.mdpg_cache_attrs);
2848 	/*
2849 	 * Set modified bit gratuitously for writeable mappings if
2850 	 * the page is unmanaged. We do not want to take a fault
2851 	 * to do the dirty bit accounting for these mappings.
2852 	 */
2853 	if ((m->oflags & VPO_UNMANAGED) != 0) {
2854 		if ((newpte & PG_RW) != 0)
2855 			newpte |= PG_M;
2856 	} else
2857 		newpte |= PG_MANAGED;
2858 
2859 	lock = NULL;
2860 	PMAP_LOCK(pmap);
2861 	if (psind == 1) {
2862 		/* Assert the required virtual and physical alignment. */
2863 		KASSERT((va & L3_PAGE_MASK) == 0, ("pmap_enter: va unaligned"));
2864 		KASSERT(m->psind > 0, ("pmap_enter: m->psind < psind"));
2865 		rv = pmap_enter_l3e(pmap, va, newpte | RPTE_LEAF, flags, m, &lock);
2866 		goto out;
2867 	}
2868 	mpte = NULL;
2869 
2870 	/*
2871 	 * In the case that a page table page is not
2872 	 * resident, we are creating it here.
2873 	 */
2874 retry:
2875 	l3e = pmap_pml3e(pmap, va);
2876 	if (l3e != NULL && (be64toh(*l3e) & PG_V) != 0 && ((be64toh(*l3e) & RPTE_LEAF) == 0 ||
2877 	    pmap_demote_l3e_locked(pmap, l3e, va, &lock))) {
2878 		pte = pmap_l3e_to_pte(l3e, va);
2879 		if (va < VM_MAXUSER_ADDRESS && mpte == NULL) {
2880 			mpte = PHYS_TO_VM_PAGE(be64toh(*l3e) & PG_FRAME);
2881 			mpte->ref_count++;
2882 		}
2883 	} else if (va < VM_MAXUSER_ADDRESS) {
2884 		/*
2885 		 * Here if the pte page isn't mapped, or if it has been
2886 		 * deallocated.
2887 		 */
2888 		nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0;
2889 		mpte = _pmap_allocpte(pmap, pmap_l3e_pindex(va),
2890 		    nosleep ? NULL : &lock);
2891 		if (mpte == NULL && nosleep) {
2892 			rv = KERN_RESOURCE_SHORTAGE;
2893 			goto out;
2894 		}
2895 		if (__predict_false(retrycount++ == 6))
2896 			panic("too many retries");
2897 		invalidate_all = true;
2898 		goto retry;
2899 	} else
2900 		panic("pmap_enter: invalid page directory va=%#lx", va);
2901 
2902 	origpte = be64toh(*pte);
2903 	pv = NULL;
2904 
2905 	/*
2906 	 * Is the specified virtual address already mapped?
2907 	 */
2908 	if ((origpte & PG_V) != 0) {
2909 #ifdef INVARIANTS
2910 		if (VERBOSE_PMAP || pmap_logging) {
2911 			printf("cow fault pmap_enter(%p, %#lx, %p, %#x, %x, %d) --"
2912 			    " asid=%lu curpid=%d name=%s origpte0x%lx\n",
2913 			    pmap, va, m, prot, flags, psind, pmap->pm_pid,
2914 			    curproc->p_pid, curproc->p_comm, origpte);
2915 			pmap_pte_walk(pmap->pm_pml1, va);
2916 		}
2917 #endif
2918 		/*
2919 		 * Wiring change, just update stats. We don't worry about
2920 		 * wiring PT pages as they remain resident as long as there
2921 		 * are valid mappings in them. Hence, if a user page is wired,
2922 		 * the PT page will be also.
2923 		 */
2924 		if ((newpte & PG_W) != 0 && (origpte & PG_W) == 0)
2925 			pmap->pm_stats.wired_count++;
2926 		else if ((newpte & PG_W) == 0 && (origpte & PG_W) != 0)
2927 			pmap->pm_stats.wired_count--;
2928 
2929 		/*
2930 		 * Remove the extra PT page reference.
2931 		 */
2932 		if (mpte != NULL) {
2933 			mpte->ref_count--;
2934 			KASSERT(mpte->ref_count > 0,
2935 			    ("pmap_enter: missing reference to page table page,"
2936 			     " va: 0x%lx", va));
2937 		}
2938 
2939 		/*
2940 		 * Has the physical page changed?
2941 		 */
2942 		opa = origpte & PG_FRAME;
2943 		if (opa == pa) {
2944 			/*
2945 			 * No, might be a protection or wiring change.
2946 			 */
2947 			if ((origpte & PG_MANAGED) != 0 &&
2948 			    (newpte & PG_RW) != 0)
2949 				vm_page_aflag_set(m, PGA_WRITEABLE);
2950 			if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0) {
2951 				if ((newpte & (PG_A|PG_M)) != (origpte & (PG_A|PG_M))) {
2952 					if (!atomic_cmpset_long(pte, htobe64(origpte), htobe64(newpte)))
2953 						goto retry;
2954 					if ((newpte & PG_M) != (origpte & PG_M))
2955 						vm_page_dirty(m);
2956 					if ((newpte & PG_A) != (origpte & PG_A))
2957 						vm_page_aflag_set(m, PGA_REFERENCED);
2958 					ptesync();
2959 				} else
2960 					invalidate_all = true;
2961 				if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0)
2962 					goto unchanged;
2963 			}
2964 			goto validate;
2965 		}
2966 
2967 		/*
2968 		 * The physical page has changed.  Temporarily invalidate
2969 		 * the mapping.  This ensures that all threads sharing the
2970 		 * pmap keep a consistent view of the mapping, which is
2971 		 * necessary for the correct handling of COW faults.  It
2972 		 * also permits reuse of the old mapping's PV entry,
2973 		 * avoiding an allocation.
2974 		 *
2975 		 * For consistency, handle unmanaged mappings the same way.
2976 		 */
2977 		origpte = be64toh(pte_load_clear(pte));
2978 		KASSERT((origpte & PG_FRAME) == opa,
2979 		    ("pmap_enter: unexpected pa update for %#lx", va));
2980 		if ((origpte & PG_MANAGED) != 0) {
2981 			om = PHYS_TO_VM_PAGE(opa);
2982 
2983 			/*
2984 			 * The pmap lock is sufficient to synchronize with
2985 			 * concurrent calls to pmap_page_test_mappings() and
2986 			 * pmap_ts_referenced().
2987 			 */
2988 			if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
2989 				vm_page_dirty(om);
2990 			if ((origpte & PG_A) != 0)
2991 				vm_page_aflag_set(om, PGA_REFERENCED);
2992 			CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, opa);
2993 			pv = pmap_pvh_remove(&om->md, pmap, va);
2994 			if ((newpte & PG_MANAGED) == 0)
2995 				free_pv_entry(pmap, pv);
2996 #ifdef INVARIANTS
2997 			else if (origpte & PG_MANAGED) {
2998 				if (pv == NULL) {
2999 					pmap_page_print_mappings(om);
3000 					MPASS(pv != NULL);
3001 				}
3002 			}
3003 #endif
3004 			if ((om->a.flags & PGA_WRITEABLE) != 0 &&
3005 			    TAILQ_EMPTY(&om->md.pv_list) &&
3006 			    ((om->flags & PG_FICTITIOUS) != 0 ||
3007 			    TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list)))
3008 				vm_page_aflag_clear(om, PGA_WRITEABLE);
3009 		}
3010 		if ((origpte & PG_A) != 0)
3011 			invalidate_page = true;
3012 		origpte = 0;
3013 	} else {
3014 		if (pmap != kernel_pmap) {
3015 #ifdef INVARIANTS
3016 			if (VERBOSE_PMAP || pmap_logging)
3017 				printf("pmap_enter(%p, %#lx, %p, %#x, %x, %d) -- asid=%lu curpid=%d name=%s\n",
3018 				    pmap, va, m, prot, flags, psind,
3019 				    pmap->pm_pid, curproc->p_pid,
3020 				    curproc->p_comm);
3021 #endif
3022 		}
3023 
3024 		/*
3025 		 * Increment the counters.
3026 		 */
3027 		if ((newpte & PG_W) != 0)
3028 			pmap->pm_stats.wired_count++;
3029 		pmap_resident_count_inc(pmap, 1);
3030 	}
3031 
3032 	/*
3033 	 * Enter on the PV list if part of our managed memory.
3034 	 */
3035 	if ((newpte & PG_MANAGED) != 0) {
3036 		if (pv == NULL) {
3037 			pv = get_pv_entry(pmap, &lock);
3038 			pv->pv_va = va;
3039 		}
3040 #ifdef VERBOSE_PV
3041 		else
3042 			printf("reassigning pv: %p to pmap: %p\n",
3043 				   pv, pmap);
3044 #endif
3045 		CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, pa);
3046 		TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_link);
3047 		m->md.pv_gen++;
3048 		if ((newpte & PG_RW) != 0)
3049 			vm_page_aflag_set(m, PGA_WRITEABLE);
3050 	}
3051 
3052 	/*
3053 	 * Update the PTE.
3054 	 */
3055 	if ((origpte & PG_V) != 0) {
3056 validate:
3057 		origpte = be64toh(pte_load_store(pte, htobe64(newpte)));
3058 		KASSERT((origpte & PG_FRAME) == pa,
3059 		    ("pmap_enter: unexpected pa update for %#lx", va));
3060 		if ((newpte & PG_M) == 0 && (origpte & (PG_M | PG_RW)) ==
3061 		    (PG_M | PG_RW)) {
3062 			if ((origpte & PG_MANAGED) != 0)
3063 				vm_page_dirty(m);
3064 			invalidate_page = true;
3065 
3066 			/*
3067 			 * Although the PTE may still have PG_RW set, TLB
3068 			 * invalidation may nonetheless be required because
3069 			 * the PTE no longer has PG_M set.
3070 			 */
3071 		} else if ((origpte & PG_X) != 0 || (newpte & PG_X) == 0) {
3072 			/*
3073 			 * Removing capabilities requires invalidation on POWER
3074 			 */
3075 			invalidate_page = true;
3076 			goto unchanged;
3077 		}
3078 		if ((origpte & PG_A) != 0)
3079 			invalidate_page = true;
3080 	} else {
3081 		pte_store(pte, newpte);
3082 		ptesync();
3083 	}
3084 unchanged:
3085 
3086 #if VM_NRESERVLEVEL > 0
3087 	/*
3088 	 * If both the page table page and the reservation are fully
3089 	 * populated, then attempt promotion.
3090 	 */
3091 	if ((mpte == NULL || mpte->ref_count == NPTEPG) &&
3092 	    mmu_radix_ps_enabled(pmap) &&
3093 	    (m->flags & PG_FICTITIOUS) == 0 &&
3094 	    vm_reserv_level_iffullpop(m) == 0 &&
3095 		pmap_promote_l3e(pmap, l3e, va, &lock) == 0)
3096 		invalidate_all = true;
3097 #endif
3098 	if (invalidate_all)
3099 		pmap_invalidate_all(pmap);
3100 	else if (invalidate_page)
3101 		pmap_invalidate_page(pmap, va);
3102 
3103 	rv = KERN_SUCCESS;
3104 out:
3105 	if (lock != NULL)
3106 		rw_wunlock(lock);
3107 	PMAP_UNLOCK(pmap);
3108 
3109 	return (rv);
3110 }
3111 
3112 /*
3113  * Tries to create a read- and/or execute-only 2MB page mapping.  Returns true
3114  * if successful.  Returns false if (1) a page table page cannot be allocated
3115  * without sleeping, (2) a mapping already exists at the specified virtual
3116  * address, or (3) a PV entry cannot be allocated without reclaiming another
3117  * PV entry.
3118  */
3119 static bool
3120 pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
3121     struct rwlock **lockp)
3122 {
3123 	pml3_entry_t newpde;
3124 
3125 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3126 	newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.mdpg_cache_attrs) |
3127 	    RPTE_LEAF | PG_V;
3128 	if ((m->oflags & VPO_UNMANAGED) == 0)
3129 		newpde |= PG_MANAGED;
3130 	if (prot & VM_PROT_EXECUTE)
3131 		newpde |= PG_X;
3132 	if (prot & VM_PROT_READ)
3133 		newpde |= RPTE_EAA_R;
3134 	if (va >= DMAP_MIN_ADDRESS)
3135 		newpde |= RPTE_EAA_P;
3136 	return (pmap_enter_l3e(pmap, va, newpde, PMAP_ENTER_NOSLEEP |
3137 	    PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, NULL, lockp) ==
3138 	    KERN_SUCCESS);
3139 }
3140 
3141 /*
3142  * Tries to create the specified 2MB page mapping.  Returns KERN_SUCCESS if
3143  * the mapping was created, and either KERN_FAILURE or KERN_RESOURCE_SHORTAGE
3144  * otherwise.  Returns KERN_FAILURE if PMAP_ENTER_NOREPLACE was specified and
3145  * a mapping already exists at the specified virtual address.  Returns
3146  * KERN_RESOURCE_SHORTAGE if PMAP_ENTER_NOSLEEP was specified and a page table
3147  * page allocation failed.  Returns KERN_RESOURCE_SHORTAGE if
3148  * PMAP_ENTER_NORECLAIM was specified and a PV entry allocation failed.
3149  *
3150  * The parameter "m" is only used when creating a managed, writeable mapping.
3151  */
3152 static int
3153 pmap_enter_l3e(pmap_t pmap, vm_offset_t va, pml3_entry_t newpde, u_int flags,
3154     vm_page_t m, struct rwlock **lockp)
3155 {
3156 	struct spglist free;
3157 	pml3_entry_t oldl3e, *l3e;
3158 	vm_page_t mt, pdpg;
3159 
3160 	KASSERT((newpde & (PG_M | PG_RW)) != PG_RW,
3161 	    ("pmap_enter_pde: newpde is missing PG_M"));
3162 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3163 
3164 	if ((pdpg = pmap_allocl3e(pmap, va, (flags & PMAP_ENTER_NOSLEEP) != 0 ?
3165 	    NULL : lockp)) == NULL) {
3166 		CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
3167 		    " in pmap %p", va, pmap);
3168 		return (KERN_RESOURCE_SHORTAGE);
3169 	}
3170 	l3e = (pml3_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pdpg));
3171 	l3e = &l3e[pmap_pml3e_index(va)];
3172 	oldl3e = be64toh(*l3e);
3173 	if ((oldl3e & PG_V) != 0) {
3174 		KASSERT(pdpg->ref_count > 1,
3175 		    ("pmap_enter_pde: pdpg's wire count is too low"));
3176 		if ((flags & PMAP_ENTER_NOREPLACE) != 0) {
3177 			pdpg->ref_count--;
3178 			CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
3179 			    " in pmap %p", va, pmap);
3180 			return (KERN_FAILURE);
3181 		}
3182 		/* Break the existing mapping(s). */
3183 		SLIST_INIT(&free);
3184 		if ((oldl3e & RPTE_LEAF) != 0) {
3185 			/*
3186 			 * The reference to the PD page that was acquired by
3187 			 * pmap_allocl3e() ensures that it won't be freed.
3188 			 * However, if the PDE resulted from a promotion, then
3189 			 * a reserved PT page could be freed.
3190 			 */
3191 			(void)pmap_remove_l3e(pmap, l3e, va, &free, lockp);
3192 			pmap_invalidate_l3e_page(pmap, va, oldl3e);
3193 		} else {
3194 			if (pmap_remove_ptes(pmap, va, va + L3_PAGE_SIZE, l3e,
3195 			    &free, lockp))
3196 		               pmap_invalidate_all(pmap);
3197 		}
3198 		vm_page_free_pages_toq(&free, true);
3199 		if (va >= VM_MAXUSER_ADDRESS) {
3200 			mt = PHYS_TO_VM_PAGE(be64toh(*l3e) & PG_FRAME);
3201 			if (pmap_insert_pt_page(pmap, mt)) {
3202 				/*
3203 				 * XXX Currently, this can't happen because
3204 				 * we do not perform pmap_enter(psind == 1)
3205 				 * on the kernel pmap.
3206 				 */
3207 				panic("pmap_enter_pde: trie insert failed");
3208 			}
3209 		} else
3210 			KASSERT(be64toh(*l3e) == 0, ("pmap_enter_pde: non-zero pde %p",
3211 			    l3e));
3212 	}
3213 	if ((newpde & PG_MANAGED) != 0) {
3214 		/*
3215 		 * Abort this mapping if its PV entry could not be created.
3216 		 */
3217 		if (!pmap_pv_insert_l3e(pmap, va, newpde, flags, lockp)) {
3218 			SLIST_INIT(&free);
3219 			if (pmap_unwire_ptp(pmap, va, pdpg, &free)) {
3220 				/*
3221 				 * Although "va" is not mapped, paging-
3222 				 * structure caches could nonetheless have
3223 				 * entries that refer to the freed page table
3224 				 * pages.  Invalidate those entries.
3225 				 */
3226 				pmap_invalidate_page(pmap, va);
3227 				vm_page_free_pages_toq(&free, true);
3228 			}
3229 			CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx"
3230 			    " in pmap %p", va, pmap);
3231 			return (KERN_RESOURCE_SHORTAGE);
3232 		}
3233 		if ((newpde & PG_RW) != 0) {
3234 			for (mt = m; mt < &m[L3_PAGE_SIZE / PAGE_SIZE]; mt++)
3235 				vm_page_aflag_set(mt, PGA_WRITEABLE);
3236 		}
3237 	}
3238 
3239 	/*
3240 	 * Increment counters.
3241 	 */
3242 	if ((newpde & PG_W) != 0)
3243 		pmap->pm_stats.wired_count += L3_PAGE_SIZE / PAGE_SIZE;
3244 	pmap_resident_count_inc(pmap, L3_PAGE_SIZE / PAGE_SIZE);
3245 
3246 	/*
3247 	 * Map the superpage.  (This is not a promoted mapping; there will not
3248 	 * be any lingering 4KB page mappings in the TLB.)
3249 	 */
3250 	pte_store(l3e, newpde);
3251 	ptesync();
3252 
3253 	counter_u64_add(pmap_l3e_mappings, 1);
3254 	CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx"
3255 	    " in pmap %p", va, pmap);
3256 	return (KERN_SUCCESS);
3257 }
3258 
3259 void
3260 mmu_radix_enter_object(pmap_t pmap, vm_offset_t start,
3261     vm_offset_t end, vm_page_t m_start, vm_prot_t prot)
3262 {
3263 
3264 	struct rwlock *lock;
3265 	vm_offset_t va;
3266 	vm_page_t m, mpte;
3267 	vm_pindex_t diff, psize;
3268 	bool invalidate;
3269 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
3270 
3271 	CTR6(KTR_PMAP, "%s(%p, %#x, %#x, %p, %#x)", __func__, pmap, start,
3272 	    end, m_start, prot);
3273 
3274 	invalidate = false;
3275 	psize = atop(end - start);
3276 	mpte = NULL;
3277 	m = m_start;
3278 	lock = NULL;
3279 	PMAP_LOCK(pmap);
3280 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
3281 		va = start + ptoa(diff);
3282 		if ((va & L3_PAGE_MASK) == 0 && va + L3_PAGE_SIZE <= end &&
3283 		    m->psind == 1 && mmu_radix_ps_enabled(pmap) &&
3284 		    pmap_enter_2mpage(pmap, va, m, prot, &lock))
3285 			m = &m[L3_PAGE_SIZE / PAGE_SIZE - 1];
3286 		else
3287 			mpte = mmu_radix_enter_quick_locked(pmap, va, m, prot,
3288 			    mpte, &lock, &invalidate);
3289 		m = TAILQ_NEXT(m, listq);
3290 	}
3291 	ptesync();
3292 	if (lock != NULL)
3293 		rw_wunlock(lock);
3294 	if (invalidate)
3295 		pmap_invalidate_all(pmap);
3296 	PMAP_UNLOCK(pmap);
3297 }
3298 
3299 static vm_page_t
3300 mmu_radix_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m,
3301     vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp, bool *invalidate)
3302 {
3303 	struct spglist free;
3304 	pt_entry_t *pte;
3305 	vm_paddr_t pa;
3306 
3307 	KASSERT(!VA_IS_CLEANMAP(va) ||
3308 	    (m->oflags & VPO_UNMANAGED) != 0,
3309 	    ("mmu_radix_enter_quick_locked: managed mapping within the clean submap"));
3310 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3311 
3312 	/*
3313 	 * In the case that a page table page is not
3314 	 * resident, we are creating it here.
3315 	 */
3316 	if (va < VM_MAXUSER_ADDRESS) {
3317 		vm_pindex_t ptepindex;
3318 		pml3_entry_t *ptepa;
3319 
3320 		/*
3321 		 * Calculate pagetable page index
3322 		 */
3323 		ptepindex = pmap_l3e_pindex(va);
3324 		if (mpte && (mpte->pindex == ptepindex)) {
3325 			mpte->ref_count++;
3326 		} else {
3327 			/*
3328 			 * Get the page directory entry
3329 			 */
3330 			ptepa = pmap_pml3e(pmap, va);
3331 
3332 			/*
3333 			 * If the page table page is mapped, we just increment
3334 			 * the hold count, and activate it.  Otherwise, we
3335 			 * attempt to allocate a page table page.  If this
3336 			 * attempt fails, we don't retry.  Instead, we give up.
3337 			 */
3338 			if (ptepa && (be64toh(*ptepa) & PG_V) != 0) {
3339 				if (be64toh(*ptepa) & RPTE_LEAF)
3340 					return (NULL);
3341 				mpte = PHYS_TO_VM_PAGE(be64toh(*ptepa) & PG_FRAME);
3342 				mpte->ref_count++;
3343 			} else {
3344 				/*
3345 				 * Pass NULL instead of the PV list lock
3346 				 * pointer, because we don't intend to sleep.
3347 				 */
3348 				mpte = _pmap_allocpte(pmap, ptepindex, NULL);
3349 				if (mpte == NULL)
3350 					return (mpte);
3351 			}
3352 		}
3353 		pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte));
3354 		pte = &pte[pmap_pte_index(va)];
3355 	} else {
3356 		mpte = NULL;
3357 		pte = pmap_pte(pmap, va);
3358 	}
3359 	if (be64toh(*pte)) {
3360 		if (mpte != NULL) {
3361 			mpte->ref_count--;
3362 			mpte = NULL;
3363 		}
3364 		return (mpte);
3365 	}
3366 
3367 	/*
3368 	 * Enter on the PV list if part of our managed memory.
3369 	 */
3370 	if ((m->oflags & VPO_UNMANAGED) == 0 &&
3371 	    !pmap_try_insert_pv_entry(pmap, va, m, lockp)) {
3372 		if (mpte != NULL) {
3373 			SLIST_INIT(&free);
3374 			if (pmap_unwire_ptp(pmap, va, mpte, &free)) {
3375 				/*
3376 				 * Although "va" is not mapped, paging-
3377 				 * structure caches could nonetheless have
3378 				 * entries that refer to the freed page table
3379 				 * pages.  Invalidate those entries.
3380 				 */
3381 				*invalidate = true;
3382 				vm_page_free_pages_toq(&free, true);
3383 			}
3384 			mpte = NULL;
3385 		}
3386 		return (mpte);
3387 	}
3388 
3389 	/*
3390 	 * Increment counters
3391 	 */
3392 	pmap_resident_count_inc(pmap, 1);
3393 
3394 	pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.mdpg_cache_attrs);
3395 	if (prot & VM_PROT_EXECUTE)
3396 		pa |= PG_X;
3397 	else
3398 		pa |= RPTE_EAA_R;
3399 	if ((m->oflags & VPO_UNMANAGED) == 0)
3400 		pa |= PG_MANAGED;
3401 
3402 	pte_store(pte, pa);
3403 	return (mpte);
3404 }
3405 
3406 void
3407 mmu_radix_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m,
3408     vm_prot_t prot)
3409 {
3410 	struct rwlock *lock;
3411 	bool invalidate;
3412 
3413 	lock = NULL;
3414 	invalidate = false;
3415 	PMAP_LOCK(pmap);
3416 	mmu_radix_enter_quick_locked(pmap, va, m, prot, NULL, &lock,
3417 	    &invalidate);
3418 	ptesync();
3419 	if (lock != NULL)
3420 		rw_wunlock(lock);
3421 	if (invalidate)
3422 		pmap_invalidate_all(pmap);
3423 	PMAP_UNLOCK(pmap);
3424 }
3425 
3426 vm_paddr_t
3427 mmu_radix_extract(pmap_t pmap, vm_offset_t va)
3428 {
3429 	pml3_entry_t *l3e;
3430 	pt_entry_t *pte;
3431 	vm_paddr_t pa;
3432 
3433 	l3e = pmap_pml3e(pmap, va);
3434 	if (__predict_false(l3e == NULL))
3435 		return (0);
3436 	if (be64toh(*l3e) & RPTE_LEAF) {
3437 		pa = (be64toh(*l3e) & PG_PS_FRAME) | (va & L3_PAGE_MASK);
3438 		pa |= (va & L3_PAGE_MASK);
3439 	} else {
3440 		/*
3441 		 * Beware of a concurrent promotion that changes the
3442 		 * PDE at this point!  For example, vtopte() must not
3443 		 * be used to access the PTE because it would use the
3444 		 * new PDE.  It is, however, safe to use the old PDE
3445 		 * because the page table page is preserved by the
3446 		 * promotion.
3447 		 */
3448 		pte = pmap_l3e_to_pte(l3e, va);
3449 		if (__predict_false(pte == NULL))
3450 			return (0);
3451 		pa = be64toh(*pte);
3452 		pa = (pa & PG_FRAME) | (va & PAGE_MASK);
3453 		pa |= (va & PAGE_MASK);
3454 	}
3455 	return (pa);
3456 }
3457 
3458 vm_page_t
3459 mmu_radix_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
3460 {
3461 	pml3_entry_t l3e, *l3ep;
3462 	pt_entry_t pte;
3463 	vm_page_t m;
3464 
3465 	m = NULL;
3466 	CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, pmap, va, prot);
3467 	PMAP_LOCK(pmap);
3468 	l3ep = pmap_pml3e(pmap, va);
3469 	if (l3ep != NULL && (l3e = be64toh(*l3ep))) {
3470 		if (l3e & RPTE_LEAF) {
3471 			if ((l3e & PG_RW) || (prot & VM_PROT_WRITE) == 0)
3472 				m = PHYS_TO_VM_PAGE((l3e & PG_PS_FRAME) |
3473 				    (va & L3_PAGE_MASK));
3474 		} else {
3475 			/* Native endian PTE, do not pass to pmap functions */
3476 			pte = be64toh(*pmap_l3e_to_pte(l3ep, va));
3477 			if ((pte & PG_V) &&
3478 			    ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0))
3479 				m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
3480 		}
3481 		if (m != NULL && !vm_page_wire_mapped(m))
3482 			m = NULL;
3483 	}
3484 	PMAP_UNLOCK(pmap);
3485 	return (m);
3486 }
3487 
3488 static void
3489 mmu_radix_growkernel(vm_offset_t addr)
3490 {
3491 	vm_paddr_t paddr;
3492 	vm_page_t nkpg;
3493 	pml3_entry_t *l3e;
3494 	pml2_entry_t *l2e;
3495 
3496 	CTR2(KTR_PMAP, "%s(%#x)", __func__, addr);
3497 	if (VM_MIN_KERNEL_ADDRESS < addr &&
3498 		addr < (VM_MIN_KERNEL_ADDRESS + nkpt * L3_PAGE_SIZE))
3499 		return;
3500 
3501 	addr = roundup2(addr, L3_PAGE_SIZE);
3502 	if (addr - 1 >= vm_map_max(kernel_map))
3503 		addr = vm_map_max(kernel_map);
3504 	while (kernel_vm_end < addr) {
3505 		l2e = pmap_pml2e(kernel_pmap, kernel_vm_end);
3506 		if ((be64toh(*l2e) & PG_V) == 0) {
3507 			/* We need a new PDP entry */
3508 			nkpg = vm_page_alloc(NULL, kernel_vm_end >> L2_PAGE_SIZE_SHIFT,
3509 			    VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ |
3510 			    VM_ALLOC_WIRED | VM_ALLOC_ZERO);
3511 			if (nkpg == NULL)
3512 				panic("pmap_growkernel: no memory to grow kernel");
3513 			if ((nkpg->flags & PG_ZERO) == 0)
3514 				mmu_radix_zero_page(nkpg);
3515 			paddr = VM_PAGE_TO_PHYS(nkpg);
3516 			pde_store(l2e, paddr);
3517 			continue; /* try again */
3518 		}
3519 		l3e = pmap_l2e_to_l3e(l2e, kernel_vm_end);
3520 		if ((be64toh(*l3e) & PG_V) != 0) {
3521 			kernel_vm_end = (kernel_vm_end + L3_PAGE_SIZE) & ~L3_PAGE_MASK;
3522 			if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
3523 				kernel_vm_end = vm_map_max(kernel_map);
3524 				break;
3525 			}
3526 			continue;
3527 		}
3528 
3529 		nkpg = vm_page_alloc(NULL, pmap_l3e_pindex(kernel_vm_end),
3530 		    VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
3531 		    VM_ALLOC_ZERO);
3532 		if (nkpg == NULL)
3533 			panic("pmap_growkernel: no memory to grow kernel");
3534 		if ((nkpg->flags & PG_ZERO) == 0)
3535 			mmu_radix_zero_page(nkpg);
3536 		paddr = VM_PAGE_TO_PHYS(nkpg);
3537 		pde_store(l3e, paddr);
3538 
3539 		kernel_vm_end = (kernel_vm_end + L3_PAGE_SIZE) & ~L3_PAGE_MASK;
3540 		if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
3541 			kernel_vm_end = vm_map_max(kernel_map);
3542 			break;
3543 		}
3544 	}
3545 	ptesync();
3546 }
3547 
3548 static MALLOC_DEFINE(M_RADIX_PGD, "radix_pgd", "radix page table root directory");
3549 static uma_zone_t zone_radix_pgd;
3550 
3551 static int
3552 radix_pgd_import(void *arg __unused, void **store, int count, int domain __unused,
3553     int flags)
3554 {
3555 
3556 	for (int i = 0; i < count; i++) {
3557 		vm_page_t m = vm_page_alloc_contig(NULL, 0,
3558 		    VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
3559 		    VM_ALLOC_ZERO | VM_ALLOC_WAITOK, RADIX_PGD_SIZE/PAGE_SIZE,
3560 		    0, (vm_paddr_t)-1, RADIX_PGD_SIZE, L1_PAGE_SIZE,
3561 		    VM_MEMATTR_DEFAULT);
3562 		/* XXX zero on alloc here so we don't have to later */
3563 		store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3564 	}
3565 	return (count);
3566 }
3567 
3568 static void
3569 radix_pgd_release(void *arg __unused, void **store, int count)
3570 {
3571 	vm_page_t m;
3572 	struct spglist free;
3573 	int page_count;
3574 
3575 	SLIST_INIT(&free);
3576 	page_count = RADIX_PGD_SIZE/PAGE_SIZE;
3577 
3578 	for (int i = 0; i < count; i++) {
3579 		/*
3580 		 * XXX selectively remove dmap and KVA entries so we don't
3581 		 * need to bzero
3582 		 */
3583 		m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i]));
3584 		for (int j = page_count-1; j >= 0; j--) {
3585 			vm_page_unwire_noq(&m[j]);
3586 			SLIST_INSERT_HEAD(&free, &m[j], plinks.s.ss);
3587 		}
3588 		vm_page_free_pages_toq(&free, false);
3589 	}
3590 }
3591 
3592 static void
3593 mmu_radix_init()
3594 {
3595 	vm_page_t mpte;
3596 	vm_size_t s;
3597 	int error, i, pv_npg;
3598 
3599 	/* XXX is this really needed for POWER? */
3600 	/* L1TF, reserve page @0 unconditionally */
3601 	vm_page_blacklist_add(0, bootverbose);
3602 
3603 	zone_radix_pgd = uma_zcache_create("radix_pgd_cache",
3604 		RADIX_PGD_SIZE, NULL, NULL,
3605 #ifdef INVARIANTS
3606 	    trash_init, trash_fini,
3607 #else
3608 	    NULL, NULL,
3609 #endif
3610 		radix_pgd_import, radix_pgd_release,
3611 		NULL, UMA_ZONE_NOBUCKET);
3612 
3613 	/*
3614 	 * Initialize the vm page array entries for the kernel pmap's
3615 	 * page table pages.
3616 	 */
3617 	PMAP_LOCK(kernel_pmap);
3618 	for (i = 0; i < nkpt; i++) {
3619 		mpte = PHYS_TO_VM_PAGE(KPTphys + (i << PAGE_SHIFT));
3620 		KASSERT(mpte >= vm_page_array &&
3621 		    mpte < &vm_page_array[vm_page_array_size],
3622 		    ("pmap_init: page table page is out of range size: %lu",
3623 		     vm_page_array_size));
3624 		mpte->pindex = pmap_l3e_pindex(VM_MIN_KERNEL_ADDRESS) + i;
3625 		mpte->phys_addr = KPTphys + (i << PAGE_SHIFT);
3626 		MPASS(PHYS_TO_VM_PAGE(mpte->phys_addr) == mpte);
3627 		//pmap_insert_pt_page(kernel_pmap, mpte);
3628 		mpte->ref_count = 1;
3629 	}
3630 	PMAP_UNLOCK(kernel_pmap);
3631 	vm_wire_add(nkpt);
3632 
3633 	CTR1(KTR_PMAP, "%s()", __func__);
3634 	TAILQ_INIT(&pv_dummy.pv_list);
3635 
3636 	/*
3637 	 * Are large page mappings enabled?
3638 	 */
3639 	TUNABLE_INT_FETCH("vm.pmap.superpages_enabled", &superpages_enabled);
3640 	if (superpages_enabled) {
3641 		KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
3642 		    ("pmap_init: can't assign to pagesizes[1]"));
3643 		pagesizes[1] = L3_PAGE_SIZE;
3644 	}
3645 
3646 	/*
3647 	 * Initialize the pv chunk list mutex.
3648 	 */
3649 	mtx_init(&pv_chunks_mutex, "pmap pv chunk list", NULL, MTX_DEF);
3650 
3651 	/*
3652 	 * Initialize the pool of pv list locks.
3653 	 */
3654 	for (i = 0; i < NPV_LIST_LOCKS; i++)
3655 		rw_init(&pv_list_locks[i], "pmap pv list");
3656 
3657 	/*
3658 	 * Calculate the size of the pv head table for superpages.
3659 	 */
3660 	pv_npg = howmany(vm_phys_segs[vm_phys_nsegs - 1].end, L3_PAGE_SIZE);
3661 
3662 	/*
3663 	 * Allocate memory for the pv head table for superpages.
3664 	 */
3665 	s = (vm_size_t)(pv_npg * sizeof(struct md_page));
3666 	s = round_page(s);
3667 	pv_table = (struct md_page *)kmem_malloc(s, M_WAITOK | M_ZERO);
3668 	for (i = 0; i < pv_npg; i++)
3669 		TAILQ_INIT(&pv_table[i].pv_list);
3670 	TAILQ_INIT(&pv_dummy.pv_list);
3671 
3672 	pmap_initialized = 1;
3673 	mtx_init(&qframe_mtx, "qfrmlk", NULL, MTX_SPIN);
3674 	error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK,
3675 	    (vmem_addr_t *)&qframe);
3676 
3677 	if (error != 0)
3678 		panic("qframe allocation failed");
3679 	asid_arena = vmem_create("ASID", isa3_base_pid + 1, (1<<isa3_pid_bits),
3680 	    1, 1, M_WAITOK);
3681 }
3682 
3683 static boolean_t
3684 pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified)
3685 {
3686 	struct rwlock *lock;
3687 	pv_entry_t pv;
3688 	struct md_page *pvh;
3689 	pt_entry_t *pte, mask;
3690 	pmap_t pmap;
3691 	int md_gen, pvh_gen;
3692 	boolean_t rv;
3693 
3694 	rv = FALSE;
3695 	lock = VM_PAGE_TO_PV_LIST_LOCK(m);
3696 	rw_rlock(lock);
3697 restart:
3698 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) {
3699 		pmap = PV_PMAP(pv);
3700 		if (!PMAP_TRYLOCK(pmap)) {
3701 			md_gen = m->md.pv_gen;
3702 			rw_runlock(lock);
3703 			PMAP_LOCK(pmap);
3704 			rw_rlock(lock);
3705 			if (md_gen != m->md.pv_gen) {
3706 				PMAP_UNLOCK(pmap);
3707 				goto restart;
3708 			}
3709 		}
3710 		pte = pmap_pte(pmap, pv->pv_va);
3711 		mask = 0;
3712 		if (modified)
3713 			mask |= PG_RW | PG_M;
3714 		if (accessed)
3715 			mask |= PG_V | PG_A;
3716 		rv = (be64toh(*pte) & mask) == mask;
3717 		PMAP_UNLOCK(pmap);
3718 		if (rv)
3719 			goto out;
3720 	}
3721 	if ((m->flags & PG_FICTITIOUS) == 0) {
3722 		pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
3723 		TAILQ_FOREACH(pv, &pvh->pv_list, pv_link) {
3724 			pmap = PV_PMAP(pv);
3725 			if (!PMAP_TRYLOCK(pmap)) {
3726 				md_gen = m->md.pv_gen;
3727 				pvh_gen = pvh->pv_gen;
3728 				rw_runlock(lock);
3729 				PMAP_LOCK(pmap);
3730 				rw_rlock(lock);
3731 				if (md_gen != m->md.pv_gen ||
3732 				    pvh_gen != pvh->pv_gen) {
3733 					PMAP_UNLOCK(pmap);
3734 					goto restart;
3735 				}
3736 			}
3737 			pte = pmap_pml3e(pmap, pv->pv_va);
3738 			mask = 0;
3739 			if (modified)
3740 				mask |= PG_RW | PG_M;
3741 			if (accessed)
3742 				mask |= PG_V | PG_A;
3743 			rv = (be64toh(*pte) & mask) == mask;
3744 			PMAP_UNLOCK(pmap);
3745 			if (rv)
3746 				goto out;
3747 		}
3748 	}
3749 out:
3750 	rw_runlock(lock);
3751 	return (rv);
3752 }
3753 
3754 /*
3755  *	pmap_is_modified:
3756  *
3757  *	Return whether or not the specified physical page was modified
3758  *	in any physical maps.
3759  */
3760 boolean_t
3761 mmu_radix_is_modified(vm_page_t m)
3762 {
3763 
3764 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3765 	    ("pmap_is_modified: page %p is not managed", m));
3766 
3767 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
3768 	/*
3769 	 * If the page is not busied then this check is racy.
3770 	 */
3771 	if (!pmap_page_is_write_mapped(m))
3772 		return (FALSE);
3773 	return (pmap_page_test_mappings(m, FALSE, TRUE));
3774 }
3775 
3776 boolean_t
3777 mmu_radix_is_prefaultable(pmap_t pmap, vm_offset_t addr)
3778 {
3779 	pml3_entry_t *l3e;
3780 	pt_entry_t *pte;
3781 	boolean_t rv;
3782 
3783 	CTR3(KTR_PMAP, "%s(%p, %#x)", __func__, pmap, addr);
3784 	rv = FALSE;
3785 	PMAP_LOCK(pmap);
3786 	l3e = pmap_pml3e(pmap, addr);
3787 	if (l3e != NULL && (be64toh(*l3e) & (RPTE_LEAF | PG_V)) == PG_V) {
3788 		pte = pmap_l3e_to_pte(l3e, addr);
3789 		rv = (be64toh(*pte) & PG_V) == 0;
3790 	}
3791 	PMAP_UNLOCK(pmap);
3792 	return (rv);
3793 }
3794 
3795 boolean_t
3796 mmu_radix_is_referenced(vm_page_t m)
3797 {
3798 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3799 	    ("pmap_is_referenced: page %p is not managed", m));
3800 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
3801 	return (pmap_page_test_mappings(m, TRUE, FALSE));
3802 }
3803 
3804 /*
3805  *	pmap_ts_referenced:
3806  *
3807  *	Return a count of reference bits for a page, clearing those bits.
3808  *	It is not necessary for every reference bit to be cleared, but it
3809  *	is necessary that 0 only be returned when there are truly no
3810  *	reference bits set.
3811  *
3812  *	As an optimization, update the page's dirty field if a modified bit is
3813  *	found while counting reference bits.  This opportunistic update can be
3814  *	performed at low cost and can eliminate the need for some future calls
3815  *	to pmap_is_modified().  However, since this function stops after
3816  *	finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some
3817  *	dirty pages.  Those dirty pages will only be detected by a future call
3818  *	to pmap_is_modified().
3819  *
3820  *	A DI block is not needed within this function, because
3821  *	invalidations are performed before the PV list lock is
3822  *	released.
3823  */
3824 boolean_t
3825 mmu_radix_ts_referenced(vm_page_t m)
3826 {
3827 	struct md_page *pvh;
3828 	pv_entry_t pv, pvf;
3829 	pmap_t pmap;
3830 	struct rwlock *lock;
3831 	pml3_entry_t oldl3e, *l3e;
3832 	pt_entry_t *pte;
3833 	vm_paddr_t pa;
3834 	int cleared, md_gen, not_cleared, pvh_gen;
3835 	struct spglist free;
3836 
3837 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
3838 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3839 	    ("pmap_ts_referenced: page %p is not managed", m));
3840 	SLIST_INIT(&free);
3841 	cleared = 0;
3842 	pa = VM_PAGE_TO_PHYS(m);
3843 	lock = PHYS_TO_PV_LIST_LOCK(pa);
3844 	pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(pa);
3845 	rw_wlock(lock);
3846 retry:
3847 	not_cleared = 0;
3848 	if ((pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL)
3849 		goto small_mappings;
3850 	pv = pvf;
3851 	do {
3852 		if (pvf == NULL)
3853 			pvf = pv;
3854 		pmap = PV_PMAP(pv);
3855 		if (!PMAP_TRYLOCK(pmap)) {
3856 			pvh_gen = pvh->pv_gen;
3857 			rw_wunlock(lock);
3858 			PMAP_LOCK(pmap);
3859 			rw_wlock(lock);
3860 			if (pvh_gen != pvh->pv_gen) {
3861 				PMAP_UNLOCK(pmap);
3862 				goto retry;
3863 			}
3864 		}
3865 		l3e = pmap_pml3e(pmap, pv->pv_va);
3866 		oldl3e = be64toh(*l3e);
3867 		if ((oldl3e & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
3868 			/*
3869 			 * Although "oldpde" is mapping a 2MB page, because
3870 			 * this function is called at a 4KB page granularity,
3871 			 * we only update the 4KB page under test.
3872 			 */
3873 			vm_page_dirty(m);
3874 		}
3875 		if ((oldl3e & PG_A) != 0) {
3876 			/*
3877 			 * Since this reference bit is shared by 512 4KB
3878 			 * pages, it should not be cleared every time it is
3879 			 * tested.  Apply a simple "hash" function on the
3880 			 * physical page number, the virtual superpage number,
3881 			 * and the pmap address to select one 4KB page out of
3882 			 * the 512 on which testing the reference bit will
3883 			 * result in clearing that reference bit.  This
3884 			 * function is designed to avoid the selection of the
3885 			 * same 4KB page for every 2MB page mapping.
3886 			 *
3887 			 * On demotion, a mapping that hasn't been referenced
3888 			 * is simply destroyed.  To avoid the possibility of a
3889 			 * subsequent page fault on a demoted wired mapping,
3890 			 * always leave its reference bit set.  Moreover,
3891 			 * since the superpage is wired, the current state of
3892 			 * its reference bit won't affect page replacement.
3893 			 */
3894 			if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> L3_PAGE_SIZE_SHIFT) ^
3895 			    (uintptr_t)pmap) & (NPTEPG - 1)) == 0 &&
3896 			    (oldl3e & PG_W) == 0) {
3897 				atomic_clear_long(l3e, htobe64(PG_A));
3898 				pmap_invalidate_page(pmap, pv->pv_va);
3899 				cleared++;
3900 				KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m),
3901 				    ("inconsistent pv lock %p %p for page %p",
3902 				    lock, VM_PAGE_TO_PV_LIST_LOCK(m), m));
3903 			} else
3904 				not_cleared++;
3905 		}
3906 		PMAP_UNLOCK(pmap);
3907 		/* Rotate the PV list if it has more than one entry. */
3908 		if (pv != NULL && TAILQ_NEXT(pv, pv_link) != NULL) {
3909 			TAILQ_REMOVE(&pvh->pv_list, pv, pv_link);
3910 			TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_link);
3911 			pvh->pv_gen++;
3912 		}
3913 		if (cleared + not_cleared >= PMAP_TS_REFERENCED_MAX)
3914 			goto out;
3915 	} while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf);
3916 small_mappings:
3917 	if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL)
3918 		goto out;
3919 	pv = pvf;
3920 	do {
3921 		if (pvf == NULL)
3922 			pvf = pv;
3923 		pmap = PV_PMAP(pv);
3924 		if (!PMAP_TRYLOCK(pmap)) {
3925 			pvh_gen = pvh->pv_gen;
3926 			md_gen = m->md.pv_gen;
3927 			rw_wunlock(lock);
3928 			PMAP_LOCK(pmap);
3929 			rw_wlock(lock);
3930 			if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
3931 				PMAP_UNLOCK(pmap);
3932 				goto retry;
3933 			}
3934 		}
3935 		l3e = pmap_pml3e(pmap, pv->pv_va);
3936 		KASSERT((be64toh(*l3e) & RPTE_LEAF) == 0,
3937 		    ("pmap_ts_referenced: found a 2mpage in page %p's pv list",
3938 		    m));
3939 		pte = pmap_l3e_to_pte(l3e, pv->pv_va);
3940 		if ((be64toh(*pte) & (PG_M | PG_RW)) == (PG_M | PG_RW))
3941 			vm_page_dirty(m);
3942 		if ((be64toh(*pte) & PG_A) != 0) {
3943 			atomic_clear_long(pte, htobe64(PG_A));
3944 			pmap_invalidate_page(pmap, pv->pv_va);
3945 			cleared++;
3946 		}
3947 		PMAP_UNLOCK(pmap);
3948 		/* Rotate the PV list if it has more than one entry. */
3949 		if (pv != NULL && TAILQ_NEXT(pv, pv_link) != NULL) {
3950 			TAILQ_REMOVE(&m->md.pv_list, pv, pv_link);
3951 			TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_link);
3952 			m->md.pv_gen++;
3953 		}
3954 	} while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared +
3955 	    not_cleared < PMAP_TS_REFERENCED_MAX);
3956 out:
3957 	rw_wunlock(lock);
3958 	vm_page_free_pages_toq(&free, true);
3959 	return (cleared + not_cleared);
3960 }
3961 
3962 static vm_offset_t
3963 mmu_radix_map(vm_offset_t *virt __unused, vm_paddr_t start,
3964     vm_paddr_t end, int prot __unused)
3965 {
3966 
3967 	CTR5(KTR_PMAP, "%s(%p, %#x, %#x, %#x)", __func__, virt, start, end,
3968 		 prot);
3969 	return (PHYS_TO_DMAP(start));
3970 }
3971 
3972 void
3973 mmu_radix_object_init_pt(pmap_t pmap, vm_offset_t addr,
3974     vm_object_t object, vm_pindex_t pindex, vm_size_t size)
3975 {
3976 	pml3_entry_t *l3e;
3977 	vm_paddr_t pa, ptepa;
3978 	vm_page_t p, pdpg;
3979 	vm_memattr_t ma;
3980 
3981 	CTR6(KTR_PMAP, "%s(%p, %#x, %p, %u, %#x)", __func__, pmap, addr,
3982 	    object, pindex, size);
3983 	VM_OBJECT_ASSERT_WLOCKED(object);
3984 	KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG,
3985 			("pmap_object_init_pt: non-device object"));
3986 	/* NB: size can be logically ored with addr here */
3987 	if ((addr & L3_PAGE_MASK) == 0 && (size & L3_PAGE_MASK) == 0) {
3988 		if (!mmu_radix_ps_enabled(pmap))
3989 			return;
3990 		if (!vm_object_populate(object, pindex, pindex + atop(size)))
3991 			return;
3992 		p = vm_page_lookup(object, pindex);
3993 		KASSERT(p->valid == VM_PAGE_BITS_ALL,
3994 		    ("pmap_object_init_pt: invalid page %p", p));
3995 		ma = p->md.mdpg_cache_attrs;
3996 
3997 		/*
3998 		 * Abort the mapping if the first page is not physically
3999 		 * aligned to a 2MB page boundary.
4000 		 */
4001 		ptepa = VM_PAGE_TO_PHYS(p);
4002 		if (ptepa & L3_PAGE_MASK)
4003 			return;
4004 
4005 		/*
4006 		 * Skip the first page.  Abort the mapping if the rest of
4007 		 * the pages are not physically contiguous or have differing
4008 		 * memory attributes.
4009 		 */
4010 		p = TAILQ_NEXT(p, listq);
4011 		for (pa = ptepa + PAGE_SIZE; pa < ptepa + size;
4012 		    pa += PAGE_SIZE) {
4013 			KASSERT(p->valid == VM_PAGE_BITS_ALL,
4014 			    ("pmap_object_init_pt: invalid page %p", p));
4015 			if (pa != VM_PAGE_TO_PHYS(p) ||
4016 			    ma != p->md.mdpg_cache_attrs)
4017 				return;
4018 			p = TAILQ_NEXT(p, listq);
4019 		}
4020 
4021 		PMAP_LOCK(pmap);
4022 		for (pa = ptepa | pmap_cache_bits(ma);
4023 		    pa < ptepa + size; pa += L3_PAGE_SIZE) {
4024 			pdpg = pmap_allocl3e(pmap, addr, NULL);
4025 			if (pdpg == NULL) {
4026 				/*
4027 				 * The creation of mappings below is only an
4028 				 * optimization.  If a page directory page
4029 				 * cannot be allocated without blocking,
4030 				 * continue on to the next mapping rather than
4031 				 * blocking.
4032 				 */
4033 				addr += L3_PAGE_SIZE;
4034 				continue;
4035 			}
4036 			l3e = (pml3_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pdpg));
4037 			l3e = &l3e[pmap_pml3e_index(addr)];
4038 			if ((be64toh(*l3e) & PG_V) == 0) {
4039 				pa |= PG_M | PG_A | PG_RW;
4040 				pte_store(l3e, pa);
4041 				pmap_resident_count_inc(pmap, L3_PAGE_SIZE / PAGE_SIZE);
4042 				counter_u64_add(pmap_l3e_mappings, 1);
4043 			} else {
4044 				/* Continue on if the PDE is already valid. */
4045 				pdpg->ref_count--;
4046 				KASSERT(pdpg->ref_count > 0,
4047 				    ("pmap_object_init_pt: missing reference "
4048 				    "to page directory page, va: 0x%lx", addr));
4049 			}
4050 			addr += L3_PAGE_SIZE;
4051 		}
4052 		ptesync();
4053 		PMAP_UNLOCK(pmap);
4054 	}
4055 }
4056 
4057 boolean_t
4058 mmu_radix_page_exists_quick(pmap_t pmap, vm_page_t m)
4059 {
4060 	struct md_page *pvh;
4061 	struct rwlock *lock;
4062 	pv_entry_t pv;
4063 	int loops = 0;
4064 	boolean_t rv;
4065 
4066 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4067 	    ("pmap_page_exists_quick: page %p is not managed", m));
4068 	CTR3(KTR_PMAP, "%s(%p, %p)", __func__, pmap, m);
4069 	rv = FALSE;
4070 	lock = VM_PAGE_TO_PV_LIST_LOCK(m);
4071 	rw_rlock(lock);
4072 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) {
4073 		if (PV_PMAP(pv) == pmap) {
4074 			rv = TRUE;
4075 			break;
4076 		}
4077 		loops++;
4078 		if (loops >= 16)
4079 			break;
4080 	}
4081 	if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) {
4082 		pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4083 		TAILQ_FOREACH(pv, &pvh->pv_list, pv_link) {
4084 			if (PV_PMAP(pv) == pmap) {
4085 				rv = TRUE;
4086 				break;
4087 			}
4088 			loops++;
4089 			if (loops >= 16)
4090 				break;
4091 		}
4092 	}
4093 	rw_runlock(lock);
4094 	return (rv);
4095 }
4096 
4097 void
4098 mmu_radix_page_init(vm_page_t m)
4099 {
4100 
4101 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
4102 	TAILQ_INIT(&m->md.pv_list);
4103 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
4104 }
4105 
4106 int
4107 mmu_radix_page_wired_mappings(vm_page_t m)
4108 {
4109 	struct rwlock *lock;
4110 	struct md_page *pvh;
4111 	pmap_t pmap;
4112 	pt_entry_t *pte;
4113 	pv_entry_t pv;
4114 	int count, md_gen, pvh_gen;
4115 
4116 	if ((m->oflags & VPO_UNMANAGED) != 0)
4117 		return (0);
4118 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
4119 	lock = VM_PAGE_TO_PV_LIST_LOCK(m);
4120 	rw_rlock(lock);
4121 restart:
4122 	count = 0;
4123 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) {
4124 		pmap = PV_PMAP(pv);
4125 		if (!PMAP_TRYLOCK(pmap)) {
4126 			md_gen = m->md.pv_gen;
4127 			rw_runlock(lock);
4128 			PMAP_LOCK(pmap);
4129 			rw_rlock(lock);
4130 			if (md_gen != m->md.pv_gen) {
4131 				PMAP_UNLOCK(pmap);
4132 				goto restart;
4133 			}
4134 		}
4135 		pte = pmap_pte(pmap, pv->pv_va);
4136 		if ((be64toh(*pte) & PG_W) != 0)
4137 			count++;
4138 		PMAP_UNLOCK(pmap);
4139 	}
4140 	if ((m->flags & PG_FICTITIOUS) == 0) {
4141 		pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
4142 		TAILQ_FOREACH(pv, &pvh->pv_list, pv_link) {
4143 			pmap = PV_PMAP(pv);
4144 			if (!PMAP_TRYLOCK(pmap)) {
4145 				md_gen = m->md.pv_gen;
4146 				pvh_gen = pvh->pv_gen;
4147 				rw_runlock(lock);
4148 				PMAP_LOCK(pmap);
4149 				rw_rlock(lock);
4150 				if (md_gen != m->md.pv_gen ||
4151 				    pvh_gen != pvh->pv_gen) {
4152 					PMAP_UNLOCK(pmap);
4153 					goto restart;
4154 				}
4155 			}
4156 			pte = pmap_pml3e(pmap, pv->pv_va);
4157 			if ((be64toh(*pte) & PG_W) != 0)
4158 				count++;
4159 			PMAP_UNLOCK(pmap);
4160 		}
4161 	}
4162 	rw_runlock(lock);
4163 	return (count);
4164 }
4165 
4166 static void
4167 mmu_radix_update_proctab(int pid, pml1_entry_t l1pa)
4168 {
4169 	isa3_proctab[pid].proctab0 = htobe64(RTS_SIZE |  l1pa | RADIX_PGD_INDEX_SHIFT);
4170 }
4171 
4172 int
4173 mmu_radix_pinit(pmap_t pmap)
4174 {
4175 	vmem_addr_t pid;
4176 	vm_paddr_t l1pa;
4177 
4178 	CTR2(KTR_PMAP, "%s(%p)", __func__, pmap);
4179 
4180 	/*
4181 	 * allocate the page directory page
4182 	 */
4183 	pmap->pm_pml1 = uma_zalloc(zone_radix_pgd, M_WAITOK);
4184 
4185 	for (int j = 0; j <  RADIX_PGD_SIZE_SHIFT; j++)
4186 		pagezero((vm_offset_t)pmap->pm_pml1 + j * PAGE_SIZE);
4187 	pmap->pm_radix.rt_root = 0;
4188 	TAILQ_INIT(&pmap->pm_pvchunk);
4189 	bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
4190 	pmap->pm_flags = PMAP_PDE_SUPERPAGE;
4191 	vmem_alloc(asid_arena, 1, M_FIRSTFIT|M_WAITOK, &pid);
4192 
4193 	pmap->pm_pid = pid;
4194 	l1pa = DMAP_TO_PHYS((vm_offset_t)pmap->pm_pml1);
4195 	mmu_radix_update_proctab(pid, l1pa);
4196 	__asm __volatile("ptesync;isync" : : : "memory");
4197 
4198 	return (1);
4199 }
4200 
4201 /*
4202  * This routine is called if the desired page table page does not exist.
4203  *
4204  * If page table page allocation fails, this routine may sleep before
4205  * returning NULL.  It sleeps only if a lock pointer was given.
4206  *
4207  * Note: If a page allocation fails at page table level two or three,
4208  * one or two pages may be held during the wait, only to be released
4209  * afterwards.  This conservative approach is easily argued to avoid
4210  * race conditions.
4211  */
4212 static vm_page_t
4213 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp)
4214 {
4215 	vm_page_t m, pdppg, pdpg;
4216 
4217 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4218 
4219 	/*
4220 	 * Allocate a page table page.
4221 	 */
4222 	if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ |
4223 	    VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
4224 		if (lockp != NULL) {
4225 			RELEASE_PV_LIST_LOCK(lockp);
4226 			PMAP_UNLOCK(pmap);
4227 			vm_wait(NULL);
4228 			PMAP_LOCK(pmap);
4229 		}
4230 		/*
4231 		 * Indicate the need to retry.  While waiting, the page table
4232 		 * page may have been allocated.
4233 		 */
4234 		return (NULL);
4235 	}
4236 	if ((m->flags & PG_ZERO) == 0)
4237 		mmu_radix_zero_page(m);
4238 
4239 	/*
4240 	 * Map the pagetable page into the process address space, if
4241 	 * it isn't already there.
4242 	 */
4243 
4244 	if (ptepindex >= (NUPDE + NUPDPE)) {
4245 		pml1_entry_t *l1e;
4246 		vm_pindex_t pml1index;
4247 
4248 		/* Wire up a new PDPE page */
4249 		pml1index = ptepindex - (NUPDE + NUPDPE);
4250 		l1e = &pmap->pm_pml1[pml1index];
4251 		pde_store(l1e, VM_PAGE_TO_PHYS(m));
4252 
4253 	} else if (ptepindex >= NUPDE) {
4254 		vm_pindex_t pml1index;
4255 		vm_pindex_t pdpindex;
4256 		pml1_entry_t *l1e;
4257 		pml2_entry_t *l2e;
4258 
4259 		/* Wire up a new l2e page */
4260 		pdpindex = ptepindex - NUPDE;
4261 		pml1index = pdpindex >> RPTE_SHIFT;
4262 
4263 		l1e = &pmap->pm_pml1[pml1index];
4264 		if ((be64toh(*l1e) & PG_V) == 0) {
4265 			/* Have to allocate a new pdp, recurse */
4266 			if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml1index,
4267 				lockp) == NULL) {
4268 				vm_page_unwire_noq(m);
4269 				vm_page_free_zero(m);
4270 				return (NULL);
4271 			}
4272 		} else {
4273 			/* Add reference to l2e page */
4274 			pdppg = PHYS_TO_VM_PAGE(be64toh(*l1e) & PG_FRAME);
4275 			pdppg->ref_count++;
4276 		}
4277 		l2e = (pml2_entry_t *)PHYS_TO_DMAP(be64toh(*l1e) & PG_FRAME);
4278 
4279 		/* Now find the pdp page */
4280 		l2e = &l2e[pdpindex & RPTE_MASK];
4281 		pde_store(l2e, VM_PAGE_TO_PHYS(m));
4282 
4283 	} else {
4284 		vm_pindex_t pml1index;
4285 		vm_pindex_t pdpindex;
4286 		pml1_entry_t *l1e;
4287 		pml2_entry_t *l2e;
4288 		pml3_entry_t *l3e;
4289 
4290 		/* Wire up a new PTE page */
4291 		pdpindex = ptepindex >> RPTE_SHIFT;
4292 		pml1index = pdpindex >> RPTE_SHIFT;
4293 
4294 		/* First, find the pdp and check that its valid. */
4295 		l1e = &pmap->pm_pml1[pml1index];
4296 		if ((be64toh(*l1e) & PG_V) == 0) {
4297 			/* Have to allocate a new pd, recurse */
4298 			if (_pmap_allocpte(pmap, NUPDE + pdpindex,
4299 			    lockp) == NULL) {
4300 				vm_page_unwire_noq(m);
4301 				vm_page_free_zero(m);
4302 				return (NULL);
4303 			}
4304 			l2e = (pml2_entry_t *)PHYS_TO_DMAP(be64toh(*l1e) & PG_FRAME);
4305 			l2e = &l2e[pdpindex & RPTE_MASK];
4306 		} else {
4307 			l2e = (pml2_entry_t *)PHYS_TO_DMAP(be64toh(*l1e) & PG_FRAME);
4308 			l2e = &l2e[pdpindex & RPTE_MASK];
4309 			if ((be64toh(*l2e) & PG_V) == 0) {
4310 				/* Have to allocate a new pd, recurse */
4311 				if (_pmap_allocpte(pmap, NUPDE + pdpindex,
4312 				    lockp) == NULL) {
4313 					vm_page_unwire_noq(m);
4314 					vm_page_free_zero(m);
4315 					return (NULL);
4316 				}
4317 			} else {
4318 				/* Add reference to the pd page */
4319 				pdpg = PHYS_TO_VM_PAGE(be64toh(*l2e) & PG_FRAME);
4320 				pdpg->ref_count++;
4321 			}
4322 		}
4323 		l3e = (pml3_entry_t *)PHYS_TO_DMAP(be64toh(*l2e) & PG_FRAME);
4324 
4325 		/* Now we know where the page directory page is */
4326 		l3e = &l3e[ptepindex & RPTE_MASK];
4327 		pde_store(l3e, VM_PAGE_TO_PHYS(m));
4328 	}
4329 
4330 	pmap_resident_count_inc(pmap, 1);
4331 	return (m);
4332 }
4333 static vm_page_t
4334 pmap_allocl3e(pmap_t pmap, vm_offset_t va, struct rwlock **lockp)
4335 {
4336 	vm_pindex_t pdpindex, ptepindex;
4337 	pml2_entry_t *pdpe;
4338 	vm_page_t pdpg;
4339 
4340 retry:
4341 	pdpe = pmap_pml2e(pmap, va);
4342 	if (pdpe != NULL && (be64toh(*pdpe) & PG_V) != 0) {
4343 		/* Add a reference to the pd page. */
4344 		pdpg = PHYS_TO_VM_PAGE(be64toh(*pdpe) & PG_FRAME);
4345 		pdpg->ref_count++;
4346 	} else {
4347 		/* Allocate a pd page. */
4348 		ptepindex = pmap_l3e_pindex(va);
4349 		pdpindex = ptepindex >> RPTE_SHIFT;
4350 		pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex, lockp);
4351 		if (pdpg == NULL && lockp != NULL)
4352 			goto retry;
4353 	}
4354 	return (pdpg);
4355 }
4356 
4357 static vm_page_t
4358 pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp)
4359 {
4360 	vm_pindex_t ptepindex;
4361 	pml3_entry_t *pd;
4362 	vm_page_t m;
4363 
4364 	/*
4365 	 * Calculate pagetable page index
4366 	 */
4367 	ptepindex = pmap_l3e_pindex(va);
4368 retry:
4369 	/*
4370 	 * Get the page directory entry
4371 	 */
4372 	pd = pmap_pml3e(pmap, va);
4373 
4374 	/*
4375 	 * This supports switching from a 2MB page to a
4376 	 * normal 4K page.
4377 	 */
4378 	if (pd != NULL && (be64toh(*pd) & (RPTE_LEAF | PG_V)) == (RPTE_LEAF | PG_V)) {
4379 		if (!pmap_demote_l3e_locked(pmap, pd, va, lockp)) {
4380 			/*
4381 			 * Invalidation of the 2MB page mapping may have caused
4382 			 * the deallocation of the underlying PD page.
4383 			 */
4384 			pd = NULL;
4385 		}
4386 	}
4387 
4388 	/*
4389 	 * If the page table page is mapped, we just increment the
4390 	 * hold count, and activate it.
4391 	 */
4392 	if (pd != NULL && (be64toh(*pd) & PG_V) != 0) {
4393 		m = PHYS_TO_VM_PAGE(be64toh(*pd) & PG_FRAME);
4394 		m->ref_count++;
4395 	} else {
4396 		/*
4397 		 * Here if the pte page isn't mapped, or if it has been
4398 		 * deallocated.
4399 		 */
4400 		m = _pmap_allocpte(pmap, ptepindex, lockp);
4401 		if (m == NULL && lockp != NULL)
4402 			goto retry;
4403 	}
4404 	return (m);
4405 }
4406 
4407 static void
4408 mmu_radix_pinit0(pmap_t pmap)
4409 {
4410 
4411 	CTR2(KTR_PMAP, "%s(%p)", __func__, pmap);
4412 	PMAP_LOCK_INIT(pmap);
4413 	pmap->pm_pml1 = kernel_pmap->pm_pml1;
4414 	pmap->pm_pid = kernel_pmap->pm_pid;
4415 
4416 	pmap->pm_radix.rt_root = 0;
4417 	TAILQ_INIT(&pmap->pm_pvchunk);
4418 	bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
4419 	kernel_pmap->pm_flags =
4420 		pmap->pm_flags = PMAP_PDE_SUPERPAGE;
4421 }
4422 /*
4423  * pmap_protect_l3e: do the things to protect a 2mpage in a process
4424  */
4425 static boolean_t
4426 pmap_protect_l3e(pmap_t pmap, pt_entry_t *l3e, vm_offset_t sva, vm_prot_t prot)
4427 {
4428 	pt_entry_t newpde, oldpde;
4429 	vm_offset_t eva, va;
4430 	vm_page_t m;
4431 	boolean_t anychanged;
4432 
4433 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4434 	KASSERT((sva & L3_PAGE_MASK) == 0,
4435 	    ("pmap_protect_l3e: sva is not 2mpage aligned"));
4436 	anychanged = FALSE;
4437 retry:
4438 	oldpde = newpde = be64toh(*l3e);
4439 	if ((oldpde & (PG_MANAGED | PG_M | PG_RW)) ==
4440 	    (PG_MANAGED | PG_M | PG_RW)) {
4441 		eva = sva + L3_PAGE_SIZE;
4442 		for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
4443 		    va < eva; va += PAGE_SIZE, m++)
4444 			vm_page_dirty(m);
4445 	}
4446 	if ((prot & VM_PROT_WRITE) == 0) {
4447 		newpde &= ~(PG_RW | PG_M);
4448 		newpde |= RPTE_EAA_R;
4449 	}
4450 	if (prot & VM_PROT_EXECUTE)
4451 		newpde |= PG_X;
4452 	if (newpde != oldpde) {
4453 		/*
4454 		 * As an optimization to future operations on this PDE, clear
4455 		 * PG_PROMOTED.  The impending invalidation will remove any
4456 		 * lingering 4KB page mappings from the TLB.
4457 		 */
4458 		if (!atomic_cmpset_long(l3e, htobe64(oldpde), htobe64(newpde & ~PG_PROMOTED)))
4459 			goto retry;
4460 		anychanged = TRUE;
4461 	}
4462 	return (anychanged);
4463 }
4464 
4465 void
4466 mmu_radix_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
4467     vm_prot_t prot)
4468 {
4469 	vm_offset_t va_next;
4470 	pml1_entry_t *l1e;
4471 	pml2_entry_t *l2e;
4472 	pml3_entry_t ptpaddr, *l3e;
4473 	pt_entry_t *pte;
4474 	boolean_t anychanged;
4475 
4476 	CTR5(KTR_PMAP, "%s(%p, %#x, %#x, %#x)", __func__, pmap, sva, eva,
4477 	    prot);
4478 
4479 	KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot));
4480 	if (prot == VM_PROT_NONE) {
4481 		mmu_radix_remove(pmap, sva, eva);
4482 		return;
4483 	}
4484 
4485 	if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) ==
4486 	    (VM_PROT_WRITE|VM_PROT_EXECUTE))
4487 		return;
4488 
4489 #ifdef INVARIANTS
4490 	if (VERBOSE_PROTECT || pmap_logging)
4491 		printf("pmap_protect(%p, %#lx, %#lx, %x) - asid: %lu\n",
4492 			   pmap, sva, eva, prot, pmap->pm_pid);
4493 #endif
4494 	anychanged = FALSE;
4495 
4496 	PMAP_LOCK(pmap);
4497 	for (; sva < eva; sva = va_next) {
4498 		l1e = pmap_pml1e(pmap, sva);
4499 		if ((be64toh(*l1e) & PG_V) == 0) {
4500 			va_next = (sva + L1_PAGE_SIZE) & ~L1_PAGE_MASK;
4501 			if (va_next < sva)
4502 				va_next = eva;
4503 			continue;
4504 		}
4505 
4506 		l2e = pmap_l1e_to_l2e(l1e, sva);
4507 		if ((be64toh(*l2e) & PG_V) == 0) {
4508 			va_next = (sva + L2_PAGE_SIZE) & ~L2_PAGE_MASK;
4509 			if (va_next < sva)
4510 				va_next = eva;
4511 			continue;
4512 		}
4513 
4514 		va_next = (sva + L3_PAGE_SIZE) & ~L3_PAGE_MASK;
4515 		if (va_next < sva)
4516 			va_next = eva;
4517 
4518 		l3e = pmap_l2e_to_l3e(l2e, sva);
4519 		ptpaddr = be64toh(*l3e);
4520 
4521 		/*
4522 		 * Weed out invalid mappings.
4523 		 */
4524 		if (ptpaddr == 0)
4525 			continue;
4526 
4527 		/*
4528 		 * Check for large page.
4529 		 */
4530 		if ((ptpaddr & RPTE_LEAF) != 0) {
4531 			/*
4532 			 * Are we protecting the entire large page?  If not,
4533 			 * demote the mapping and fall through.
4534 			 */
4535 			if (sva + L3_PAGE_SIZE == va_next && eva >= va_next) {
4536 				if (pmap_protect_l3e(pmap, l3e, sva, prot))
4537 					anychanged = TRUE;
4538 				continue;
4539 			} else if (!pmap_demote_l3e(pmap, l3e, sva)) {
4540 				/*
4541 				 * The large page mapping was destroyed.
4542 				 */
4543 				continue;
4544 			}
4545 		}
4546 
4547 		if (va_next > eva)
4548 			va_next = eva;
4549 
4550 		for (pte = pmap_l3e_to_pte(l3e, sva); sva != va_next; pte++,
4551 		    sva += PAGE_SIZE) {
4552 			pt_entry_t obits, pbits;
4553 			vm_page_t m;
4554 
4555 retry:
4556 			MPASS(pte == pmap_pte(pmap, sva));
4557 			obits = pbits = be64toh(*pte);
4558 			if ((pbits & PG_V) == 0)
4559 				continue;
4560 
4561 			if ((prot & VM_PROT_WRITE) == 0) {
4562 				if ((pbits & (PG_MANAGED | PG_M | PG_RW)) ==
4563 				    (PG_MANAGED | PG_M | PG_RW)) {
4564 					m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
4565 					vm_page_dirty(m);
4566 				}
4567 				pbits &= ~(PG_RW | PG_M);
4568 				pbits |= RPTE_EAA_R;
4569 			}
4570 			if (prot & VM_PROT_EXECUTE)
4571 				pbits |= PG_X;
4572 
4573 			if (pbits != obits) {
4574 				if (!atomic_cmpset_long(pte, htobe64(obits), htobe64(pbits)))
4575 					goto retry;
4576 				if (obits & (PG_A|PG_M)) {
4577 					anychanged = TRUE;
4578 #ifdef INVARIANTS
4579 					if (VERBOSE_PROTECT || pmap_logging)
4580 						printf("%#lx %#lx -> %#lx\n",
4581 						    sva, obits, pbits);
4582 #endif
4583 				}
4584 			}
4585 		}
4586 	}
4587 	if (anychanged)
4588 		pmap_invalidate_all(pmap);
4589 	PMAP_UNLOCK(pmap);
4590 }
4591 
4592 void
4593 mmu_radix_qenter(vm_offset_t sva, vm_page_t *ma, int count)
4594 {
4595 
4596 	CTR4(KTR_PMAP, "%s(%#x, %p, %d)", __func__, sva, ma, count);
4597 	pt_entry_t oldpte, pa, *pte;
4598 	vm_page_t m;
4599 	uint64_t cache_bits, attr_bits;
4600 	vm_offset_t va;
4601 
4602 	oldpte = 0;
4603 	attr_bits = RPTE_EAA_R | RPTE_EAA_W | RPTE_EAA_P | PG_M | PG_A;
4604 	va = sva;
4605 	pte = kvtopte(va);
4606 	while (va < sva + PAGE_SIZE * count) {
4607 		if (__predict_false((va & L3_PAGE_MASK) == 0))
4608 			pte = kvtopte(va);
4609 		MPASS(pte == pmap_pte(kernel_pmap, va));
4610 
4611 		/*
4612 		 * XXX there has to be a more efficient way than traversing
4613 		 * the page table every time - but go for correctness for
4614 		 * today
4615 		 */
4616 
4617 		m = *ma++;
4618 		cache_bits = pmap_cache_bits(m->md.mdpg_cache_attrs);
4619 		pa = VM_PAGE_TO_PHYS(m) | cache_bits | attr_bits;
4620 		if (be64toh(*pte) != pa) {
4621 			oldpte |= be64toh(*pte);
4622 			pte_store(pte, pa);
4623 		}
4624 		va += PAGE_SIZE;
4625 		pte++;
4626 	}
4627 	if (__predict_false((oldpte & RPTE_VALID) != 0))
4628 		pmap_invalidate_range(kernel_pmap, sva, sva + count *
4629 		    PAGE_SIZE);
4630 	else
4631 		ptesync();
4632 }
4633 
4634 void
4635 mmu_radix_qremove(vm_offset_t sva, int count)
4636 {
4637 	vm_offset_t va;
4638 	pt_entry_t *pte;
4639 
4640 	CTR3(KTR_PMAP, "%s(%#x, %d)", __func__, sva, count);
4641 	KASSERT(sva >= VM_MIN_KERNEL_ADDRESS, ("usermode or dmap va %lx", sva));
4642 
4643 	va = sva;
4644 	pte = kvtopte(va);
4645 	while (va < sva + PAGE_SIZE * count) {
4646 		if (__predict_false((va & L3_PAGE_MASK) == 0))
4647 			pte = kvtopte(va);
4648 		pte_clear(pte);
4649 		pte++;
4650 		va += PAGE_SIZE;
4651 	}
4652 	pmap_invalidate_range(kernel_pmap, sva, va);
4653 }
4654 
4655 /***************************************************
4656  * Page table page management routines.....
4657  ***************************************************/
4658 /*
4659  * Schedule the specified unused page table page to be freed.  Specifically,
4660  * add the page to the specified list of pages that will be released to the
4661  * physical memory manager after the TLB has been updated.
4662  */
4663 static __inline void
4664 pmap_add_delayed_free_list(vm_page_t m, struct spglist *free,
4665     boolean_t set_PG_ZERO)
4666 {
4667 
4668 	if (set_PG_ZERO)
4669 		m->flags |= PG_ZERO;
4670 	else
4671 		m->flags &= ~PG_ZERO;
4672 	SLIST_INSERT_HEAD(free, m, plinks.s.ss);
4673 }
4674 
4675 /*
4676  * Inserts the specified page table page into the specified pmap's collection
4677  * of idle page table pages.  Each of a pmap's page table pages is responsible
4678  * for mapping a distinct range of virtual addresses.  The pmap's collection is
4679  * ordered by this virtual address range.
4680  */
4681 static __inline int
4682 pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte)
4683 {
4684 
4685 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4686 	return (vm_radix_insert(&pmap->pm_radix, mpte));
4687 }
4688 
4689 /*
4690  * Removes the page table page mapping the specified virtual address from the
4691  * specified pmap's collection of idle page table pages, and returns it.
4692  * Otherwise, returns NULL if there is no page table page corresponding to the
4693  * specified virtual address.
4694  */
4695 static __inline vm_page_t
4696 pmap_remove_pt_page(pmap_t pmap, vm_offset_t va)
4697 {
4698 
4699 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4700 	return (vm_radix_remove(&pmap->pm_radix, pmap_l3e_pindex(va)));
4701 }
4702 
4703 /*
4704  * Decrements a page table page's wire count, which is used to record the
4705  * number of valid page table entries within the page.  If the wire count
4706  * drops to zero, then the page table page is unmapped.  Returns TRUE if the
4707  * page table page was unmapped and FALSE otherwise.
4708  */
4709 static inline boolean_t
4710 pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
4711 {
4712 
4713 	--m->ref_count;
4714 	if (m->ref_count == 0) {
4715 		_pmap_unwire_ptp(pmap, va, m, free);
4716 		return (TRUE);
4717 	} else
4718 		return (FALSE);
4719 }
4720 
4721 static void
4722 _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
4723 {
4724 
4725 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4726 	/*
4727 	 * unmap the page table page
4728 	 */
4729 	if (m->pindex >= NUPDE + NUPDPE) {
4730 		/* PDP page */
4731 		pml1_entry_t *pml1;
4732 		pml1 = pmap_pml1e(pmap, va);
4733 		*pml1 = 0;
4734 	} else if (m->pindex >= NUPDE) {
4735 		/* PD page */
4736 		pml2_entry_t *l2e;
4737 		l2e = pmap_pml2e(pmap, va);
4738 		*l2e = 0;
4739 	} else {
4740 		/* PTE page */
4741 		pml3_entry_t *l3e;
4742 		l3e = pmap_pml3e(pmap, va);
4743 		*l3e = 0;
4744 	}
4745 	pmap_resident_count_dec(pmap, 1);
4746 	if (m->pindex < NUPDE) {
4747 		/* We just released a PT, unhold the matching PD */
4748 		vm_page_t pdpg;
4749 
4750 		pdpg = PHYS_TO_VM_PAGE(be64toh(*pmap_pml2e(pmap, va)) & PG_FRAME);
4751 		pmap_unwire_ptp(pmap, va, pdpg, free);
4752 	}
4753 	else if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
4754 		/* We just released a PD, unhold the matching PDP */
4755 		vm_page_t pdppg;
4756 
4757 		pdppg = PHYS_TO_VM_PAGE(be64toh(*pmap_pml1e(pmap, va)) & PG_FRAME);
4758 		pmap_unwire_ptp(pmap, va, pdppg, free);
4759 	}
4760 
4761 	/*
4762 	 * Put page on a list so that it is released after
4763 	 * *ALL* TLB shootdown is done
4764 	 */
4765 	pmap_add_delayed_free_list(m, free, TRUE);
4766 }
4767 
4768 /*
4769  * After removing a page table entry, this routine is used to
4770  * conditionally free the page, and manage the hold/wire counts.
4771  */
4772 static int
4773 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, pml3_entry_t ptepde,
4774     struct spglist *free)
4775 {
4776 	vm_page_t mpte;
4777 
4778 	if (va >= VM_MAXUSER_ADDRESS)
4779 		return (0);
4780 	KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0"));
4781 	mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME);
4782 	return (pmap_unwire_ptp(pmap, va, mpte, free));
4783 }
4784 
4785 void
4786 mmu_radix_release(pmap_t pmap)
4787 {
4788 
4789 	CTR2(KTR_PMAP, "%s(%p)", __func__, pmap);
4790 	KASSERT(pmap->pm_stats.resident_count == 0,
4791 	    ("pmap_release: pmap resident count %ld != 0",
4792 	    pmap->pm_stats.resident_count));
4793 	KASSERT(vm_radix_is_empty(&pmap->pm_radix),
4794 	    ("pmap_release: pmap has reserved page table page(s)"));
4795 
4796 	pmap_invalidate_all(pmap);
4797 	isa3_proctab[pmap->pm_pid].proctab0 = 0;
4798 	uma_zfree(zone_radix_pgd, pmap->pm_pml1);
4799 	vmem_free(asid_arena, pmap->pm_pid, 1);
4800 }
4801 
4802 /*
4803  * Create the PV entry for a 2MB page mapping.  Always returns true unless the
4804  * flag PMAP_ENTER_NORECLAIM is specified.  If that flag is specified, returns
4805  * false if the PV entry cannot be allocated without resorting to reclamation.
4806  */
4807 static bool
4808 pmap_pv_insert_l3e(pmap_t pmap, vm_offset_t va, pml3_entry_t pde, u_int flags,
4809     struct rwlock **lockp)
4810 {
4811 	struct md_page *pvh;
4812 	pv_entry_t pv;
4813 	vm_paddr_t pa;
4814 
4815 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4816 	/* Pass NULL instead of the lock pointer to disable reclamation. */
4817 	if ((pv = get_pv_entry(pmap, (flags & PMAP_ENTER_NORECLAIM) != 0 ?
4818 	    NULL : lockp)) == NULL)
4819 		return (false);
4820 	pv->pv_va = va;
4821 	pa = pde & PG_PS_FRAME;
4822 	CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa);
4823 	pvh = pa_to_pvh(pa);
4824 	TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_link);
4825 	pvh->pv_gen++;
4826 	return (true);
4827 }
4828 
4829 /*
4830  * Fills a page table page with mappings to consecutive physical pages.
4831  */
4832 static void
4833 pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte)
4834 {
4835 	pt_entry_t *pte;
4836 
4837 	for (pte = firstpte; pte < firstpte + NPTEPG; pte++) {
4838 		*pte = htobe64(newpte);
4839 		newpte += PAGE_SIZE;
4840 	}
4841 }
4842 
4843 static boolean_t
4844 pmap_demote_l3e(pmap_t pmap, pml3_entry_t *pde, vm_offset_t va)
4845 {
4846 	struct rwlock *lock;
4847 	boolean_t rv;
4848 
4849 	lock = NULL;
4850 	rv = pmap_demote_l3e_locked(pmap, pde, va, &lock);
4851 	if (lock != NULL)
4852 		rw_wunlock(lock);
4853 	return (rv);
4854 }
4855 
4856 static boolean_t
4857 pmap_demote_l3e_locked(pmap_t pmap, pml3_entry_t *l3e, vm_offset_t va,
4858     struct rwlock **lockp)
4859 {
4860 	pml3_entry_t oldpde;
4861 	pt_entry_t *firstpte;
4862 	vm_paddr_t mptepa;
4863 	vm_page_t mpte;
4864 	struct spglist free;
4865 	vm_offset_t sva;
4866 
4867 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4868 	oldpde = be64toh(*l3e);
4869 	KASSERT((oldpde & (RPTE_LEAF | PG_V)) == (RPTE_LEAF | PG_V),
4870 	    ("pmap_demote_l3e: oldpde is missing RPTE_LEAF and/or PG_V %lx",
4871 	    oldpde));
4872 	if ((oldpde & PG_A) == 0 || (mpte = pmap_remove_pt_page(pmap, va)) ==
4873 	    NULL) {
4874 		KASSERT((oldpde & PG_W) == 0,
4875 		    ("pmap_demote_l3e: page table page for a wired mapping"
4876 		    " is missing"));
4877 
4878 		/*
4879 		 * Invalidate the 2MB page mapping and return "failure" if the
4880 		 * mapping was never accessed or the allocation of the new
4881 		 * page table page fails.  If the 2MB page mapping belongs to
4882 		 * the direct map region of the kernel's address space, then
4883 		 * the page allocation request specifies the highest possible
4884 		 * priority (VM_ALLOC_INTERRUPT).  Otherwise, the priority is
4885 		 * normal.  Page table pages are preallocated for every other
4886 		 * part of the kernel address space, so the direct map region
4887 		 * is the only part of the kernel address space that must be
4888 		 * handled here.
4889 		 */
4890 		if ((oldpde & PG_A) == 0 || (mpte = vm_page_alloc(NULL,
4891 		    pmap_l3e_pindex(va), (va >= DMAP_MIN_ADDRESS && va <
4892 		    DMAP_MAX_ADDRESS ? VM_ALLOC_INTERRUPT : VM_ALLOC_NORMAL) |
4893 		    VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) {
4894 			SLIST_INIT(&free);
4895 			sva = trunc_2mpage(va);
4896 			pmap_remove_l3e(pmap, l3e, sva, &free, lockp);
4897 			pmap_invalidate_l3e_page(pmap, sva, oldpde);
4898 			vm_page_free_pages_toq(&free, true);
4899 			CTR2(KTR_PMAP, "pmap_demote_l3e: failure for va %#lx"
4900 			    " in pmap %p", va, pmap);
4901 			return (FALSE);
4902 		}
4903 		if (va < VM_MAXUSER_ADDRESS)
4904 			pmap_resident_count_inc(pmap, 1);
4905 	}
4906 	mptepa = VM_PAGE_TO_PHYS(mpte);
4907 	firstpte = (pt_entry_t *)PHYS_TO_DMAP(mptepa);
4908 	KASSERT((oldpde & PG_A) != 0,
4909 	    ("pmap_demote_l3e: oldpde is missing PG_A"));
4910 	KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW,
4911 	    ("pmap_demote_l3e: oldpde is missing PG_M"));
4912 
4913 	/*
4914 	 * If the page table page is new, initialize it.
4915 	 */
4916 	if (mpte->ref_count == 1) {
4917 		mpte->ref_count = NPTEPG;
4918 		pmap_fill_ptp(firstpte, oldpde);
4919 	}
4920 
4921 	KASSERT((be64toh(*firstpte) & PG_FRAME) == (oldpde & PG_FRAME),
4922 	    ("pmap_demote_l3e: firstpte and newpte map different physical"
4923 	    " addresses"));
4924 
4925 	/*
4926 	 * If the mapping has changed attributes, update the page table
4927 	 * entries.
4928 	 */
4929 	if ((be64toh(*firstpte) & PG_PTE_PROMOTE) != (oldpde & PG_PTE_PROMOTE))
4930 		pmap_fill_ptp(firstpte, oldpde);
4931 
4932 	/*
4933 	 * The spare PV entries must be reserved prior to demoting the
4934 	 * mapping, that is, prior to changing the PDE.  Otherwise, the state
4935 	 * of the PDE and the PV lists will be inconsistent, which can result
4936 	 * in reclaim_pv_chunk() attempting to remove a PV entry from the
4937 	 * wrong PV list and pmap_pv_demote_l3e() failing to find the expected
4938 	 * PV entry for the 2MB page mapping that is being demoted.
4939 	 */
4940 	if ((oldpde & PG_MANAGED) != 0)
4941 		reserve_pv_entries(pmap, NPTEPG - 1, lockp);
4942 
4943 	/*
4944 	 * Demote the mapping.  This pmap is locked.  The old PDE has
4945 	 * PG_A set.  If the old PDE has PG_RW set, it also has PG_M
4946 	 * set.  Thus, there is no danger of a race with another
4947 	 * processor changing the setting of PG_A and/or PG_M between
4948 	 * the read above and the store below.
4949 	 */
4950 	pde_store(l3e, mptepa);
4951 	pmap_invalidate_l3e_page(pmap, trunc_2mpage(va), oldpde);
4952 	/*
4953 	 * Demote the PV entry.
4954 	 */
4955 	if ((oldpde & PG_MANAGED) != 0)
4956 		pmap_pv_demote_l3e(pmap, va, oldpde & PG_PS_FRAME, lockp);
4957 
4958 	counter_u64_add(pmap_l3e_demotions, 1);
4959 	CTR2(KTR_PMAP, "pmap_demote_l3e: success for va %#lx"
4960 	    " in pmap %p", va, pmap);
4961 	return (TRUE);
4962 }
4963 
4964 /*
4965  * pmap_remove_kernel_pde: Remove a kernel superpage mapping.
4966  */
4967 static void
4968 pmap_remove_kernel_l3e(pmap_t pmap, pml3_entry_t *l3e, vm_offset_t va)
4969 {
4970 	vm_paddr_t mptepa;
4971 	vm_page_t mpte;
4972 
4973 	KASSERT(pmap == kernel_pmap, ("pmap %p is not kernel_pmap", pmap));
4974 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4975 	mpte = pmap_remove_pt_page(pmap, va);
4976 	if (mpte == NULL)
4977 		panic("pmap_remove_kernel_pde: Missing pt page.");
4978 
4979 	mptepa = VM_PAGE_TO_PHYS(mpte);
4980 
4981 	/*
4982 	 * Initialize the page table page.
4983 	 */
4984 	pagezero(PHYS_TO_DMAP(mptepa));
4985 
4986 	/*
4987 	 * Demote the mapping.
4988 	 */
4989 	pde_store(l3e, mptepa);
4990 	ptesync();
4991 }
4992 
4993 /*
4994  * pmap_remove_l3e: do the things to unmap a superpage in a process
4995  */
4996 static int
4997 pmap_remove_l3e(pmap_t pmap, pml3_entry_t *pdq, vm_offset_t sva,
4998     struct spglist *free, struct rwlock **lockp)
4999 {
5000 	struct md_page *pvh;
5001 	pml3_entry_t oldpde;
5002 	vm_offset_t eva, va;
5003 	vm_page_t m, mpte;
5004 
5005 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5006 	KASSERT((sva & L3_PAGE_MASK) == 0,
5007 	    ("pmap_remove_l3e: sva is not 2mpage aligned"));
5008 	oldpde = be64toh(pte_load_clear(pdq));
5009 	if (oldpde & PG_W)
5010 		pmap->pm_stats.wired_count -= (L3_PAGE_SIZE / PAGE_SIZE);
5011 	pmap_resident_count_dec(pmap, L3_PAGE_SIZE / PAGE_SIZE);
5012 	if (oldpde & PG_MANAGED) {
5013 		CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, oldpde & PG_PS_FRAME);
5014 		pvh = pa_to_pvh(oldpde & PG_PS_FRAME);
5015 		pmap_pvh_free(pvh, pmap, sva);
5016 		eva = sva + L3_PAGE_SIZE;
5017 		for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME);
5018 		    va < eva; va += PAGE_SIZE, m++) {
5019 			if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW))
5020 				vm_page_dirty(m);
5021 			if (oldpde & PG_A)
5022 				vm_page_aflag_set(m, PGA_REFERENCED);
5023 			if (TAILQ_EMPTY(&m->md.pv_list) &&
5024 			    TAILQ_EMPTY(&pvh->pv_list))
5025 				vm_page_aflag_clear(m, PGA_WRITEABLE);
5026 		}
5027 	}
5028 	if (pmap == kernel_pmap) {
5029 		pmap_remove_kernel_l3e(pmap, pdq, sva);
5030 	} else {
5031 		mpte = pmap_remove_pt_page(pmap, sva);
5032 		if (mpte != NULL) {
5033 			pmap_resident_count_dec(pmap, 1);
5034 			KASSERT(mpte->ref_count == NPTEPG,
5035 			    ("pmap_remove_l3e: pte page wire count error"));
5036 			mpte->ref_count = 0;
5037 			pmap_add_delayed_free_list(mpte, free, FALSE);
5038 		}
5039 	}
5040 	return (pmap_unuse_pt(pmap, sva, be64toh(*pmap_pml2e(pmap, sva)), free));
5041 }
5042 
5043 /*
5044  * pmap_remove_pte: do the things to unmap a page in a process
5045  */
5046 static int
5047 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va,
5048     pml3_entry_t ptepde, struct spglist *free, struct rwlock **lockp)
5049 {
5050 	struct md_page *pvh;
5051 	pt_entry_t oldpte;
5052 	vm_page_t m;
5053 
5054 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5055 	oldpte = be64toh(pte_load_clear(ptq));
5056 	if (oldpte & RPTE_WIRED)
5057 		pmap->pm_stats.wired_count -= 1;
5058 	pmap_resident_count_dec(pmap, 1);
5059 	if (oldpte & RPTE_MANAGED) {
5060 		m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME);
5061 		if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5062 			vm_page_dirty(m);
5063 		if (oldpte & PG_A)
5064 			vm_page_aflag_set(m, PGA_REFERENCED);
5065 		CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m);
5066 		pmap_pvh_free(&m->md, pmap, va);
5067 		if (TAILQ_EMPTY(&m->md.pv_list) &&
5068 		    (m->flags & PG_FICTITIOUS) == 0) {
5069 			pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
5070 			if (TAILQ_EMPTY(&pvh->pv_list))
5071 				vm_page_aflag_clear(m, PGA_WRITEABLE);
5072 		}
5073 	}
5074 	return (pmap_unuse_pt(pmap, va, ptepde, free));
5075 }
5076 
5077 /*
5078  * Remove a single page from a process address space
5079  */
5080 static bool
5081 pmap_remove_page(pmap_t pmap, vm_offset_t va, pml3_entry_t *l3e,
5082     struct spglist *free)
5083 {
5084 	struct rwlock *lock;
5085 	pt_entry_t *pte;
5086 	bool invalidate_all;
5087 
5088 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5089 	if ((be64toh(*l3e) & RPTE_VALID) == 0) {
5090 		return (false);
5091 	}
5092 	pte = pmap_l3e_to_pte(l3e, va);
5093 	if ((be64toh(*pte) & RPTE_VALID) == 0) {
5094 		return (false);
5095 	}
5096 	lock = NULL;
5097 
5098 	invalidate_all = pmap_remove_pte(pmap, pte, va, be64toh(*l3e), free, &lock);
5099 	if (lock != NULL)
5100 		rw_wunlock(lock);
5101 	if (!invalidate_all)
5102 		pmap_invalidate_page(pmap, va);
5103 	return (invalidate_all);
5104 }
5105 
5106 /*
5107  * Removes the specified range of addresses from the page table page.
5108  */
5109 static bool
5110 pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
5111     pml3_entry_t *l3e, struct spglist *free, struct rwlock **lockp)
5112 {
5113 	pt_entry_t *pte;
5114 	vm_offset_t va;
5115 	bool anyvalid;
5116 
5117 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5118 	anyvalid = false;
5119 	va = eva;
5120 	for (pte = pmap_l3e_to_pte(l3e, sva); sva != eva; pte++,
5121 	    sva += PAGE_SIZE) {
5122 		MPASS(pte == pmap_pte(pmap, sva));
5123 		if (*pte == 0) {
5124 			if (va != eva) {
5125 				anyvalid = true;
5126 				va = eva;
5127 			}
5128 			continue;
5129 		}
5130 		if (va == eva)
5131 			va = sva;
5132 		if (pmap_remove_pte(pmap, pte, sva, be64toh(*l3e), free, lockp)) {
5133 			anyvalid = true;
5134 			sva += PAGE_SIZE;
5135 			break;
5136 		}
5137 	}
5138 	if (anyvalid)
5139 		pmap_invalidate_all(pmap);
5140 	else if (va != eva)
5141 		pmap_invalidate_range(pmap, va, sva);
5142 	return (anyvalid);
5143 }
5144 
5145 void
5146 mmu_radix_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5147 {
5148 	struct rwlock *lock;
5149 	vm_offset_t va_next;
5150 	pml1_entry_t *l1e;
5151 	pml2_entry_t *l2e;
5152 	pml3_entry_t ptpaddr, *l3e;
5153 	struct spglist free;
5154 	bool anyvalid;
5155 
5156 	CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, pmap, sva, eva);
5157 
5158 	/*
5159 	 * Perform an unsynchronized read.  This is, however, safe.
5160 	 */
5161 	if (pmap->pm_stats.resident_count == 0)
5162 		return;
5163 
5164 	anyvalid = false;
5165 	SLIST_INIT(&free);
5166 
5167 	/* XXX something fishy here */
5168 	sva = (sva + PAGE_MASK) & ~PAGE_MASK;
5169 	eva = (eva + PAGE_MASK) & ~PAGE_MASK;
5170 
5171 	PMAP_LOCK(pmap);
5172 
5173 	/*
5174 	 * special handling of removing one page.  a very
5175 	 * common operation and easy to short circuit some
5176 	 * code.
5177 	 */
5178 	if (sva + PAGE_SIZE == eva) {
5179 		l3e = pmap_pml3e(pmap, sva);
5180 		if (l3e && (be64toh(*l3e) & RPTE_LEAF) == 0) {
5181 			anyvalid = pmap_remove_page(pmap, sva, l3e, &free);
5182 			goto out;
5183 		}
5184 	}
5185 
5186 	lock = NULL;
5187 	for (; sva < eva; sva = va_next) {
5188 		if (pmap->pm_stats.resident_count == 0)
5189 			break;
5190 		l1e = pmap_pml1e(pmap, sva);
5191 		if (l1e == NULL || (be64toh(*l1e) & PG_V) == 0) {
5192 			va_next = (sva + L1_PAGE_SIZE) & ~L1_PAGE_MASK;
5193 			if (va_next < sva)
5194 				va_next = eva;
5195 			continue;
5196 		}
5197 
5198 		l2e = pmap_l1e_to_l2e(l1e, sva);
5199 		if (l2e == NULL || (be64toh(*l2e) & PG_V) == 0) {
5200 			va_next = (sva + L2_PAGE_SIZE) & ~L2_PAGE_MASK;
5201 			if (va_next < sva)
5202 				va_next = eva;
5203 			continue;
5204 		}
5205 
5206 		/*
5207 		 * Calculate index for next page table.
5208 		 */
5209 		va_next = (sva + L3_PAGE_SIZE) & ~L3_PAGE_MASK;
5210 		if (va_next < sva)
5211 			va_next = eva;
5212 
5213 		l3e = pmap_l2e_to_l3e(l2e, sva);
5214 		ptpaddr = be64toh(*l3e);
5215 
5216 		/*
5217 		 * Weed out invalid mappings.
5218 		 */
5219 		if (ptpaddr == 0)
5220 			continue;
5221 
5222 		/*
5223 		 * Check for large page.
5224 		 */
5225 		if ((ptpaddr & RPTE_LEAF) != 0) {
5226 			/*
5227 			 * Are we removing the entire large page?  If not,
5228 			 * demote the mapping and fall through.
5229 			 */
5230 			if (sva + L3_PAGE_SIZE == va_next && eva >= va_next) {
5231 				pmap_remove_l3e(pmap, l3e, sva, &free, &lock);
5232 				anyvalid = true;
5233 				continue;
5234 			} else if (!pmap_demote_l3e_locked(pmap, l3e, sva,
5235 			    &lock)) {
5236 				/* The large page mapping was destroyed. */
5237 				continue;
5238 			} else
5239 				ptpaddr = be64toh(*l3e);
5240 		}
5241 
5242 		/*
5243 		 * Limit our scan to either the end of the va represented
5244 		 * by the current page table page, or to the end of the
5245 		 * range being removed.
5246 		 */
5247 		if (va_next > eva)
5248 			va_next = eva;
5249 
5250 		if (pmap_remove_ptes(pmap, sva, va_next, l3e, &free, &lock))
5251 			anyvalid = true;
5252 	}
5253 	if (lock != NULL)
5254 		rw_wunlock(lock);
5255 out:
5256 	if (anyvalid)
5257 		pmap_invalidate_all(pmap);
5258 	PMAP_UNLOCK(pmap);
5259 	vm_page_free_pages_toq(&free, true);
5260 }
5261 
5262 void
5263 mmu_radix_remove_all(vm_page_t m)
5264 {
5265 	struct md_page *pvh;
5266 	pv_entry_t pv;
5267 	pmap_t pmap;
5268 	struct rwlock *lock;
5269 	pt_entry_t *pte, tpte;
5270 	pml3_entry_t *l3e;
5271 	vm_offset_t va;
5272 	struct spglist free;
5273 	int pvh_gen, md_gen;
5274 
5275 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
5276 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5277 	    ("pmap_remove_all: page %p is not managed", m));
5278 	SLIST_INIT(&free);
5279 	lock = VM_PAGE_TO_PV_LIST_LOCK(m);
5280 	pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
5281 	    pa_to_pvh(VM_PAGE_TO_PHYS(m));
5282 retry:
5283 	rw_wlock(lock);
5284 	while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) {
5285 		pmap = PV_PMAP(pv);
5286 		if (!PMAP_TRYLOCK(pmap)) {
5287 			pvh_gen = pvh->pv_gen;
5288 			rw_wunlock(lock);
5289 			PMAP_LOCK(pmap);
5290 			rw_wlock(lock);
5291 			if (pvh_gen != pvh->pv_gen) {
5292 				rw_wunlock(lock);
5293 				PMAP_UNLOCK(pmap);
5294 				goto retry;
5295 			}
5296 		}
5297 		va = pv->pv_va;
5298 		l3e = pmap_pml3e(pmap, va);
5299 		(void)pmap_demote_l3e_locked(pmap, l3e, va, &lock);
5300 		PMAP_UNLOCK(pmap);
5301 	}
5302 	while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
5303 		pmap = PV_PMAP(pv);
5304 		if (!PMAP_TRYLOCK(pmap)) {
5305 			pvh_gen = pvh->pv_gen;
5306 			md_gen = m->md.pv_gen;
5307 			rw_wunlock(lock);
5308 			PMAP_LOCK(pmap);
5309 			rw_wlock(lock);
5310 			if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) {
5311 				rw_wunlock(lock);
5312 				PMAP_UNLOCK(pmap);
5313 				goto retry;
5314 			}
5315 		}
5316 		pmap_resident_count_dec(pmap, 1);
5317 		l3e = pmap_pml3e(pmap, pv->pv_va);
5318 		KASSERT((be64toh(*l3e) & RPTE_LEAF) == 0, ("pmap_remove_all: found"
5319 		    " a 2mpage in page %p's pv list", m));
5320 		pte = pmap_l3e_to_pte(l3e, pv->pv_va);
5321 		tpte = be64toh(pte_load_clear(pte));
5322 		if (tpte & PG_W)
5323 			pmap->pm_stats.wired_count--;
5324 		if (tpte & PG_A)
5325 			vm_page_aflag_set(m, PGA_REFERENCED);
5326 
5327 		/*
5328 		 * Update the vm_page_t clean and reference bits.
5329 		 */
5330 		if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5331 			vm_page_dirty(m);
5332 		pmap_unuse_pt(pmap, pv->pv_va, be64toh(*l3e), &free);
5333 		pmap_invalidate_page(pmap, pv->pv_va);
5334 		TAILQ_REMOVE(&m->md.pv_list, pv, pv_link);
5335 		m->md.pv_gen++;
5336 		free_pv_entry(pmap, pv);
5337 		PMAP_UNLOCK(pmap);
5338 	}
5339 	vm_page_aflag_clear(m, PGA_WRITEABLE);
5340 	rw_wunlock(lock);
5341 	vm_page_free_pages_toq(&free, true);
5342 }
5343 
5344 /*
5345  * Destroy all managed, non-wired mappings in the given user-space
5346  * pmap.  This pmap cannot be active on any processor besides the
5347  * caller.
5348  *
5349  * This function cannot be applied to the kernel pmap.  Moreover, it
5350  * is not intended for general use.  It is only to be used during
5351  * process termination.  Consequently, it can be implemented in ways
5352  * that make it faster than pmap_remove().  First, it can more quickly
5353  * destroy mappings by iterating over the pmap's collection of PV
5354  * entries, rather than searching the page table.  Second, it doesn't
5355  * have to test and clear the page table entries atomically, because
5356  * no processor is currently accessing the user address space.  In
5357  * particular, a page table entry's dirty bit won't change state once
5358  * this function starts.
5359  *
5360  * Although this function destroys all of the pmap's managed,
5361  * non-wired mappings, it can delay and batch the invalidation of TLB
5362  * entries without calling pmap_delayed_invl_started() and
5363  * pmap_delayed_invl_finished().  Because the pmap is not active on
5364  * any other processor, none of these TLB entries will ever be used
5365  * before their eventual invalidation.  Consequently, there is no need
5366  * for either pmap_remove_all() or pmap_remove_write() to wait for
5367  * that eventual TLB invalidation.
5368  */
5369 
5370 void
5371 mmu_radix_remove_pages(pmap_t pmap)
5372 {
5373 
5374 	CTR2(KTR_PMAP, "%s(%p)", __func__, pmap);
5375 	pml3_entry_t ptel3e;
5376 	pt_entry_t *pte, tpte;
5377 	struct spglist free;
5378 	vm_page_t m, mpte, mt;
5379 	pv_entry_t pv;
5380 	struct md_page *pvh;
5381 	struct pv_chunk *pc, *npc;
5382 	struct rwlock *lock;
5383 	int64_t bit;
5384 	uint64_t inuse, bitmask;
5385 	int allfree, field, freed, idx;
5386 	boolean_t superpage;
5387 	vm_paddr_t pa;
5388 
5389 	/*
5390 	 * Assert that the given pmap is only active on the current
5391 	 * CPU.  Unfortunately, we cannot block another CPU from
5392 	 * activating the pmap while this function is executing.
5393 	 */
5394 	KASSERT(pmap->pm_pid == mfspr(SPR_PID),
5395 	    ("non-current asid %lu - expected %lu", pmap->pm_pid,
5396 	    mfspr(SPR_PID)));
5397 
5398 	lock = NULL;
5399 
5400 	SLIST_INIT(&free);
5401 	PMAP_LOCK(pmap);
5402 	TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) {
5403 		allfree = 1;
5404 		freed = 0;
5405 		for (field = 0; field < _NPCM; field++) {
5406 			inuse = ~pc->pc_map[field] & pc_freemask[field];
5407 			while (inuse != 0) {
5408 				bit = cnttzd(inuse);
5409 				bitmask = 1UL << bit;
5410 				idx = field * 64 + bit;
5411 				pv = &pc->pc_pventry[idx];
5412 				inuse &= ~bitmask;
5413 
5414 				pte = pmap_pml2e(pmap, pv->pv_va);
5415 				ptel3e = be64toh(*pte);
5416 				pte = pmap_l2e_to_l3e(pte, pv->pv_va);
5417 				tpte = be64toh(*pte);
5418 				if ((tpte & (RPTE_LEAF | PG_V)) == PG_V) {
5419 					superpage = FALSE;
5420 					ptel3e = tpte;
5421 					pte = (pt_entry_t *)PHYS_TO_DMAP(tpte &
5422 					    PG_FRAME);
5423 					pte = &pte[pmap_pte_index(pv->pv_va)];
5424 					tpte = be64toh(*pte);
5425 				} else {
5426 					/*
5427 					 * Keep track whether 'tpte' is a
5428 					 * superpage explicitly instead of
5429 					 * relying on RPTE_LEAF being set.
5430 					 *
5431 					 * This is because RPTE_LEAF is numerically
5432 					 * identical to PG_PTE_PAT and thus a
5433 					 * regular page could be mistaken for
5434 					 * a superpage.
5435 					 */
5436 					superpage = TRUE;
5437 				}
5438 
5439 				if ((tpte & PG_V) == 0) {
5440 					panic("bad pte va %lx pte %lx",
5441 					    pv->pv_va, tpte);
5442 				}
5443 
5444 /*
5445  * We cannot remove wired pages from a process' mapping at this time
5446  */
5447 				if (tpte & PG_W) {
5448 					allfree = 0;
5449 					continue;
5450 				}
5451 
5452 				if (superpage)
5453 					pa = tpte & PG_PS_FRAME;
5454 				else
5455 					pa = tpte & PG_FRAME;
5456 
5457 				m = PHYS_TO_VM_PAGE(pa);
5458 				KASSERT(m->phys_addr == pa,
5459 				    ("vm_page_t %p phys_addr mismatch %016jx %016jx",
5460 				    m, (uintmax_t)m->phys_addr,
5461 				    (uintmax_t)tpte));
5462 
5463 				KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
5464 				    m < &vm_page_array[vm_page_array_size],
5465 				    ("pmap_remove_pages: bad tpte %#jx",
5466 				    (uintmax_t)tpte));
5467 
5468 				pte_clear(pte);
5469 
5470 				/*
5471 				 * Update the vm_page_t clean/reference bits.
5472 				 */
5473 				if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
5474 					if (superpage) {
5475 						for (mt = m; mt < &m[L3_PAGE_SIZE / PAGE_SIZE]; mt++)
5476 							vm_page_dirty(mt);
5477 					} else
5478 						vm_page_dirty(m);
5479 				}
5480 
5481 				CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m);
5482 
5483 				/* Mark free */
5484 				pc->pc_map[field] |= bitmask;
5485 				if (superpage) {
5486 					pmap_resident_count_dec(pmap, L3_PAGE_SIZE / PAGE_SIZE);
5487 					pvh = pa_to_pvh(tpte & PG_PS_FRAME);
5488 					TAILQ_REMOVE(&pvh->pv_list, pv, pv_link);
5489 					pvh->pv_gen++;
5490 					if (TAILQ_EMPTY(&pvh->pv_list)) {
5491 						for (mt = m; mt < &m[L3_PAGE_SIZE / PAGE_SIZE]; mt++)
5492 							if ((mt->a.flags & PGA_WRITEABLE) != 0 &&
5493 							    TAILQ_EMPTY(&mt->md.pv_list))
5494 								vm_page_aflag_clear(mt, PGA_WRITEABLE);
5495 					}
5496 					mpte = pmap_remove_pt_page(pmap, pv->pv_va);
5497 					if (mpte != NULL) {
5498 						pmap_resident_count_dec(pmap, 1);
5499 						KASSERT(mpte->ref_count == NPTEPG,
5500 						    ("pmap_remove_pages: pte page wire count error"));
5501 						mpte->ref_count = 0;
5502 						pmap_add_delayed_free_list(mpte, &free, FALSE);
5503 					}
5504 				} else {
5505 					pmap_resident_count_dec(pmap, 1);
5506 #ifdef VERBOSE_PV
5507 					printf("freeing pv (%p, %p)\n",
5508 						   pmap, pv);
5509 #endif
5510 					TAILQ_REMOVE(&m->md.pv_list, pv, pv_link);
5511 					m->md.pv_gen++;
5512 					if ((m->a.flags & PGA_WRITEABLE) != 0 &&
5513 					    TAILQ_EMPTY(&m->md.pv_list) &&
5514 					    (m->flags & PG_FICTITIOUS) == 0) {
5515 						pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m));
5516 						if (TAILQ_EMPTY(&pvh->pv_list))
5517 							vm_page_aflag_clear(m, PGA_WRITEABLE);
5518 					}
5519 				}
5520 				pmap_unuse_pt(pmap, pv->pv_va, ptel3e, &free);
5521 				freed++;
5522 			}
5523 		}
5524 		PV_STAT(atomic_add_long(&pv_entry_frees, freed));
5525 		PV_STAT(atomic_add_int(&pv_entry_spare, freed));
5526 		PV_STAT(atomic_subtract_long(&pv_entry_count, freed));
5527 		if (allfree) {
5528 			TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list);
5529 			free_pv_chunk(pc);
5530 		}
5531 	}
5532 	if (lock != NULL)
5533 		rw_wunlock(lock);
5534 	pmap_invalidate_all(pmap);
5535 	PMAP_UNLOCK(pmap);
5536 	vm_page_free_pages_toq(&free, true);
5537 }
5538 
5539 void
5540 mmu_radix_remove_write(vm_page_t m)
5541 {
5542 	struct md_page *pvh;
5543 	pmap_t pmap;
5544 	struct rwlock *lock;
5545 	pv_entry_t next_pv, pv;
5546 	pml3_entry_t *l3e;
5547 	pt_entry_t oldpte, *pte;
5548 	int pvh_gen, md_gen;
5549 
5550 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
5551 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5552 	    ("pmap_remove_write: page %p is not managed", m));
5553 	vm_page_assert_busied(m);
5554 
5555 	if (!pmap_page_is_write_mapped(m))
5556 		return;
5557 	lock = VM_PAGE_TO_PV_LIST_LOCK(m);
5558 	pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy :
5559 	    pa_to_pvh(VM_PAGE_TO_PHYS(m));
5560 retry_pv_loop:
5561 	rw_wlock(lock);
5562 	TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_link, next_pv) {
5563 		pmap = PV_PMAP(pv);
5564 		if (!PMAP_TRYLOCK(pmap)) {
5565 			pvh_gen = pvh->pv_gen;
5566 			rw_wunlock(lock);
5567 			PMAP_LOCK(pmap);
5568 			rw_wlock(lock);
5569 			if (pvh_gen != pvh->pv_gen) {
5570 				PMAP_UNLOCK(pmap);
5571 				rw_wunlock(lock);
5572 				goto retry_pv_loop;
5573 			}
5574 		}
5575 		l3e = pmap_pml3e(pmap, pv->pv_va);
5576 		if ((be64toh(*l3e) & PG_RW) != 0)
5577 			(void)pmap_demote_l3e_locked(pmap, l3e, pv->pv_va, &lock);
5578 		KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m),
5579 		    ("inconsistent pv lock %p %p for page %p",
5580 		    lock, VM_PAGE_TO_PV_LIST_LOCK(m), m));
5581 		PMAP_UNLOCK(pmap);
5582 	}
5583 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) {
5584 		pmap = PV_PMAP(pv);
5585 		if (!PMAP_TRYLOCK(pmap)) {
5586 			pvh_gen = pvh->pv_gen;
5587 			md_gen = m->md.pv_gen;
5588 			rw_wunlock(lock);
5589 			PMAP_LOCK(pmap);
5590 			rw_wlock(lock);
5591 			if (pvh_gen != pvh->pv_gen ||
5592 			    md_gen != m->md.pv_gen) {
5593 				PMAP_UNLOCK(pmap);
5594 				rw_wunlock(lock);
5595 				goto retry_pv_loop;
5596 			}
5597 		}
5598 		l3e = pmap_pml3e(pmap, pv->pv_va);
5599 		KASSERT((be64toh(*l3e) & RPTE_LEAF) == 0,
5600 		    ("pmap_remove_write: found a 2mpage in page %p's pv list",
5601 		    m));
5602 		pte = pmap_l3e_to_pte(l3e, pv->pv_va);
5603 retry:
5604 		oldpte = be64toh(*pte);
5605 		if (oldpte & PG_RW) {
5606 			if (!atomic_cmpset_long(pte, htobe64(oldpte),
5607 			    htobe64((oldpte | RPTE_EAA_R) & ~(PG_RW | PG_M))))
5608 				goto retry;
5609 			if ((oldpte & PG_M) != 0)
5610 				vm_page_dirty(m);
5611 			pmap_invalidate_page(pmap, pv->pv_va);
5612 		}
5613 		PMAP_UNLOCK(pmap);
5614 	}
5615 	rw_wunlock(lock);
5616 	vm_page_aflag_clear(m, PGA_WRITEABLE);
5617 }
5618 
5619 /*
5620  *	Clear the wired attribute from the mappings for the specified range of
5621  *	addresses in the given pmap.  Every valid mapping within that range
5622  *	must have the wired attribute set.  In contrast, invalid mappings
5623  *	cannot have the wired attribute set, so they are ignored.
5624  *
5625  *	The wired attribute of the page table entry is not a hardware
5626  *	feature, so there is no need to invalidate any TLB entries.
5627  *	Since pmap_demote_l3e() for the wired entry must never fail,
5628  *	pmap_delayed_invl_started()/finished() calls around the
5629  *	function are not needed.
5630  */
5631 void
5632 mmu_radix_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
5633 {
5634 	vm_offset_t va_next;
5635 	pml1_entry_t *l1e;
5636 	pml2_entry_t *l2e;
5637 	pml3_entry_t *l3e;
5638 	pt_entry_t *pte;
5639 
5640 	CTR4(KTR_PMAP, "%s(%p, %#x, %#x)", __func__, pmap, sva, eva);
5641 	PMAP_LOCK(pmap);
5642 	for (; sva < eva; sva = va_next) {
5643 		l1e = pmap_pml1e(pmap, sva);
5644 		if ((be64toh(*l1e) & PG_V) == 0) {
5645 			va_next = (sva + L1_PAGE_SIZE) & ~L1_PAGE_MASK;
5646 			if (va_next < sva)
5647 				va_next = eva;
5648 			continue;
5649 		}
5650 		l2e = pmap_l1e_to_l2e(l1e, sva);
5651 		if ((be64toh(*l2e) & PG_V) == 0) {
5652 			va_next = (sva + L2_PAGE_SIZE) & ~L2_PAGE_MASK;
5653 			if (va_next < sva)
5654 				va_next = eva;
5655 			continue;
5656 		}
5657 		va_next = (sva + L3_PAGE_SIZE) & ~L3_PAGE_MASK;
5658 		if (va_next < sva)
5659 			va_next = eva;
5660 		l3e = pmap_l2e_to_l3e(l2e, sva);
5661 		if ((be64toh(*l3e) & PG_V) == 0)
5662 			continue;
5663 		if ((be64toh(*l3e) & RPTE_LEAF) != 0) {
5664 			if ((be64toh(*l3e) & PG_W) == 0)
5665 				panic("pmap_unwire: pde %#jx is missing PG_W",
5666 				    (uintmax_t)(be64toh(*l3e)));
5667 
5668 			/*
5669 			 * Are we unwiring the entire large page?  If not,
5670 			 * demote the mapping and fall through.
5671 			 */
5672 			if (sva + L3_PAGE_SIZE == va_next && eva >= va_next) {
5673 				atomic_clear_long(l3e, htobe64(PG_W));
5674 				pmap->pm_stats.wired_count -= L3_PAGE_SIZE /
5675 				    PAGE_SIZE;
5676 				continue;
5677 			} else if (!pmap_demote_l3e(pmap, l3e, sva))
5678 				panic("pmap_unwire: demotion failed");
5679 		}
5680 		if (va_next > eva)
5681 			va_next = eva;
5682 		for (pte = pmap_l3e_to_pte(l3e, sva); sva != va_next; pte++,
5683 		    sva += PAGE_SIZE) {
5684 			MPASS(pte == pmap_pte(pmap, sva));
5685 			if ((be64toh(*pte) & PG_V) == 0)
5686 				continue;
5687 			if ((be64toh(*pte) & PG_W) == 0)
5688 				panic("pmap_unwire: pte %#jx is missing PG_W",
5689 				    (uintmax_t)(be64toh(*pte)));
5690 
5691 			/*
5692 			 * PG_W must be cleared atomically.  Although the pmap
5693 			 * lock synchronizes access to PG_W, another processor
5694 			 * could be setting PG_M and/or PG_A concurrently.
5695 			 */
5696 			atomic_clear_long(pte, htobe64(PG_W));
5697 			pmap->pm_stats.wired_count--;
5698 		}
5699 	}
5700 	PMAP_UNLOCK(pmap);
5701 }
5702 
5703 void
5704 mmu_radix_zero_page(vm_page_t m)
5705 {
5706 	vm_offset_t addr;
5707 
5708 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
5709 	addr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
5710 	pagezero(addr);
5711 }
5712 
5713 void
5714 mmu_radix_zero_page_area(vm_page_t m, int off, int size)
5715 {
5716 	caddr_t addr;
5717 
5718 	CTR4(KTR_PMAP, "%s(%p, %d, %d)", __func__, m, off, size);
5719 	MPASS(off + size <= PAGE_SIZE);
5720 	addr = (caddr_t)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
5721 	memset(addr + off, 0, size);
5722 }
5723 
5724 static int
5725 mmu_radix_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa)
5726 {
5727 	pml3_entry_t *l3ep;
5728 	pt_entry_t pte;
5729 	vm_paddr_t pa;
5730 	int val;
5731 
5732 	CTR3(KTR_PMAP, "%s(%p, %#x)", __func__, pmap, addr);
5733 	PMAP_LOCK(pmap);
5734 
5735 	l3ep = pmap_pml3e(pmap, addr);
5736 	if (l3ep != NULL && (be64toh(*l3ep) & PG_V)) {
5737 		if (be64toh(*l3ep) & RPTE_LEAF) {
5738 			pte = be64toh(*l3ep);
5739 			/* Compute the physical address of the 4KB page. */
5740 			pa = ((be64toh(*l3ep) & PG_PS_FRAME) | (addr & L3_PAGE_MASK)) &
5741 			    PG_FRAME;
5742 			val = MINCORE_PSIND(1);
5743 		} else {
5744 			/* Native endian PTE, do not pass to functions */
5745 			pte = be64toh(*pmap_l3e_to_pte(l3ep, addr));
5746 			pa = pte & PG_FRAME;
5747 			val = 0;
5748 		}
5749 	} else {
5750 		pte = 0;
5751 		pa = 0;
5752 		val = 0;
5753 	}
5754 	if ((pte & PG_V) != 0) {
5755 		val |= MINCORE_INCORE;
5756 		if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW))
5757 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
5758 		if ((pte & PG_A) != 0)
5759 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
5760 	}
5761 	if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
5762 	    (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
5763 	    (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) {
5764 		*locked_pa = pa;
5765 	}
5766 	PMAP_UNLOCK(pmap);
5767 	return (val);
5768 }
5769 
5770 void
5771 mmu_radix_activate(struct thread *td)
5772 {
5773 	pmap_t pmap;
5774 	uint32_t curpid;
5775 
5776 	CTR2(KTR_PMAP, "%s(%p)", __func__, td);
5777 	critical_enter();
5778 	pmap = vmspace_pmap(td->td_proc->p_vmspace);
5779 	curpid = mfspr(SPR_PID);
5780 	if (pmap->pm_pid > isa3_base_pid &&
5781 		curpid != pmap->pm_pid) {
5782 		mmu_radix_pid_set(pmap);
5783 	}
5784 	critical_exit();
5785 }
5786 
5787 /*
5788  *	Increase the starting virtual address of the given mapping if a
5789  *	different alignment might result in more superpage mappings.
5790  */
5791 void
5792 mmu_radix_align_superpage(vm_object_t object, vm_ooffset_t offset,
5793     vm_offset_t *addr, vm_size_t size)
5794 {
5795 
5796 	CTR5(KTR_PMAP, "%s(%p, %#x, %p, %#x)", __func__, object, offset, addr,
5797 	    size);
5798 	vm_offset_t superpage_offset;
5799 
5800 	if (size < L3_PAGE_SIZE)
5801 		return;
5802 	if (object != NULL && (object->flags & OBJ_COLORED) != 0)
5803 		offset += ptoa(object->pg_color);
5804 	superpage_offset = offset & L3_PAGE_MASK;
5805 	if (size - ((L3_PAGE_SIZE - superpage_offset) & L3_PAGE_MASK) < L3_PAGE_SIZE ||
5806 	    (*addr & L3_PAGE_MASK) == superpage_offset)
5807 		return;
5808 	if ((*addr & L3_PAGE_MASK) < superpage_offset)
5809 		*addr = (*addr & ~L3_PAGE_MASK) + superpage_offset;
5810 	else
5811 		*addr = ((*addr + L3_PAGE_MASK) & ~L3_PAGE_MASK) + superpage_offset;
5812 }
5813 
5814 static void *
5815 mmu_radix_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t attr)
5816 {
5817 	vm_offset_t va, tmpva, ppa, offset;
5818 
5819 	ppa = trunc_page(pa);
5820 	offset = pa & PAGE_MASK;
5821 	size = roundup2(offset + size, PAGE_SIZE);
5822 	if (pa < powerpc_ptob(Maxmem))
5823 		panic("bad pa: %#lx less than Maxmem %#lx\n",
5824 			  pa, powerpc_ptob(Maxmem));
5825 	va = kva_alloc(size);
5826 	if (bootverbose)
5827 		printf("%s(%#lx, %lu, %d)\n", __func__, pa, size, attr);
5828 	KASSERT(size > 0, ("%s(%#lx, %lu, %d)", __func__, pa, size, attr));
5829 
5830 	if (!va)
5831 		panic("%s: Couldn't alloc kernel virtual memory", __func__);
5832 
5833 	for (tmpva = va; size > 0;) {
5834 		mmu_radix_kenter_attr(tmpva, ppa, attr);
5835 		size -= PAGE_SIZE;
5836 		tmpva += PAGE_SIZE;
5837 		ppa += PAGE_SIZE;
5838 	}
5839 	ptesync();
5840 
5841 	return ((void *)(va + offset));
5842 }
5843 
5844 static void *
5845 mmu_radix_mapdev(vm_paddr_t pa, vm_size_t size)
5846 {
5847 
5848 	CTR3(KTR_PMAP, "%s(%#x, %#x)", __func__, pa, size);
5849 
5850 	return (mmu_radix_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT));
5851 }
5852 
5853 void
5854 mmu_radix_page_set_memattr(vm_page_t m, vm_memattr_t ma)
5855 {
5856 
5857 	CTR3(KTR_PMAP, "%s(%p, %#x)", __func__, m, ma);
5858 	m->md.mdpg_cache_attrs = ma;
5859 
5860 	/*
5861 	 * If "m" is a normal page, update its direct mapping.  This update
5862 	 * can be relied upon to perform any cache operations that are
5863 	 * required for data coherence.
5864 	 */
5865 	if ((m->flags & PG_FICTITIOUS) == 0 &&
5866 	    mmu_radix_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)),
5867 	    PAGE_SIZE, m->md.mdpg_cache_attrs))
5868 		panic("memory attribute change on the direct map failed");
5869 }
5870 
5871 static void
5872 mmu_radix_unmapdev(vm_offset_t va, vm_size_t size)
5873 {
5874 	vm_offset_t offset;
5875 
5876 	CTR3(KTR_PMAP, "%s(%#x, %#x)", __func__, va, size);
5877 	/* If we gave a direct map region in pmap_mapdev, do nothing */
5878 	if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS)
5879 		return;
5880 
5881 	offset = va & PAGE_MASK;
5882 	size = round_page(offset + size);
5883 	va = trunc_page(va);
5884 
5885 	if (pmap_initialized) {
5886 		mmu_radix_qremove(va, atop(size));
5887 		kva_free(va, size);
5888 	}
5889 }
5890 
5891 static __inline void
5892 pmap_pte_attr(pt_entry_t *pte, uint64_t cache_bits, uint64_t mask)
5893 {
5894 	uint64_t opte, npte;
5895 
5896 	/*
5897 	 * The cache mode bits are all in the low 32-bits of the
5898 	 * PTE, so we can just spin on updating the low 32-bits.
5899 	 */
5900 	do {
5901 		opte = be64toh(*pte);
5902 		npte = opte & ~mask;
5903 		npte |= cache_bits;
5904 	} while (npte != opte && !atomic_cmpset_long(pte, htobe64(opte), htobe64(npte)));
5905 }
5906 
5907 /*
5908  * Tries to demote a 1GB page mapping.
5909  */
5910 static boolean_t
5911 pmap_demote_l2e(pmap_t pmap, pml2_entry_t *l2e, vm_offset_t va)
5912 {
5913 	pml2_entry_t oldpdpe;
5914 	pml3_entry_t *firstpde, newpde, *pde;
5915 	vm_paddr_t pdpgpa;
5916 	vm_page_t pdpg;
5917 
5918 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
5919 	oldpdpe = be64toh(*l2e);
5920 	KASSERT((oldpdpe & (RPTE_LEAF | PG_V)) == (RPTE_LEAF | PG_V),
5921 	    ("pmap_demote_pdpe: oldpdpe is missing PG_PS and/or PG_V"));
5922 	pdpg = vm_page_alloc(NULL, va >> L2_PAGE_SIZE_SHIFT,
5923 	    VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED);
5924 	if (pdpg == NULL) {
5925 		CTR2(KTR_PMAP, "pmap_demote_pdpe: failure for va %#lx"
5926 		    " in pmap %p", va, pmap);
5927 		return (FALSE);
5928 	}
5929 	pdpgpa = VM_PAGE_TO_PHYS(pdpg);
5930 	firstpde = (pml3_entry_t *)PHYS_TO_DMAP(pdpgpa);
5931 	KASSERT((oldpdpe & PG_A) != 0,
5932 	    ("pmap_demote_pdpe: oldpdpe is missing PG_A"));
5933 	KASSERT((oldpdpe & (PG_M | PG_RW)) != PG_RW,
5934 	    ("pmap_demote_pdpe: oldpdpe is missing PG_M"));
5935 	newpde = oldpdpe;
5936 
5937 	/*
5938 	 * Initialize the page directory page.
5939 	 */
5940 	for (pde = firstpde; pde < firstpde + NPDEPG; pde++) {
5941 		*pde = htobe64(newpde);
5942 		newpde += L3_PAGE_SIZE;
5943 	}
5944 
5945 	/*
5946 	 * Demote the mapping.
5947 	 */
5948 	pde_store(l2e, pdpgpa);
5949 
5950 	/*
5951 	 * Flush PWC --- XXX revisit
5952 	 */
5953 	pmap_invalidate_all(pmap);
5954 
5955 	counter_u64_add(pmap_l2e_demotions, 1);
5956 	CTR2(KTR_PMAP, "pmap_demote_pdpe: success for va %#lx"
5957 	    " in pmap %p", va, pmap);
5958 	return (TRUE);
5959 }
5960 
5961 vm_paddr_t
5962 mmu_radix_kextract(vm_offset_t va)
5963 {
5964 	pml3_entry_t l3e;
5965 	vm_paddr_t pa;
5966 
5967 	CTR2(KTR_PMAP, "%s(%#x)", __func__, va);
5968 	if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
5969 		pa = DMAP_TO_PHYS(va);
5970 	} else {
5971 		/* Big-endian PTE on stack */
5972 		l3e = *pmap_pml3e(kernel_pmap, va);
5973 		if (be64toh(l3e) & RPTE_LEAF) {
5974 			pa = (be64toh(l3e) & PG_PS_FRAME) | (va & L3_PAGE_MASK);
5975 			pa |= (va & L3_PAGE_MASK);
5976 		} else {
5977 			/*
5978 			 * Beware of a concurrent promotion that changes the
5979 			 * PDE at this point!  For example, vtopte() must not
5980 			 * be used to access the PTE because it would use the
5981 			 * new PDE.  It is, however, safe to use the old PDE
5982 			 * because the page table page is preserved by the
5983 			 * promotion.
5984 			 */
5985 			pa = be64toh(*pmap_l3e_to_pte(&l3e, va));
5986 			pa = (pa & PG_FRAME) | (va & PAGE_MASK);
5987 			pa |= (va & PAGE_MASK);
5988 		}
5989 	}
5990 	return (pa);
5991 }
5992 
5993 static pt_entry_t
5994 mmu_radix_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
5995 {
5996 
5997 	if (ma != VM_MEMATTR_DEFAULT) {
5998 		return pmap_cache_bits(ma);
5999 	}
6000 
6001 	/*
6002 	 * Assume the page is cache inhibited and access is guarded unless
6003 	 * it's in our available memory array.
6004 	 */
6005 	for (int i = 0; i < pregions_sz; i++) {
6006 		if ((pa >= pregions[i].mr_start) &&
6007 		    (pa < (pregions[i].mr_start + pregions[i].mr_size)))
6008 			return (RPTE_ATTR_MEM);
6009 	}
6010 	return (RPTE_ATTR_GUARDEDIO);
6011 }
6012 
6013 static void
6014 mmu_radix_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
6015 {
6016 	pt_entry_t *pte, pteval;
6017 	uint64_t cache_bits;
6018 
6019 	pte = kvtopte(va);
6020 	MPASS(pte != NULL);
6021 	pteval = pa | RPTE_EAA_R | RPTE_EAA_W | RPTE_EAA_P | PG_M | PG_A;
6022 	cache_bits = mmu_radix_calc_wimg(pa, ma);
6023 	pte_store(pte, pteval | cache_bits);
6024 }
6025 
6026 void
6027 mmu_radix_kremove(vm_offset_t va)
6028 {
6029 	pt_entry_t *pte;
6030 
6031 	CTR2(KTR_PMAP, "%s(%#x)", __func__, va);
6032 
6033 	pte = kvtopte(va);
6034 	pte_clear(pte);
6035 }
6036 
6037 int
6038 mmu_radix_decode_kernel_ptr(vm_offset_t addr,
6039     int *is_user, vm_offset_t *decoded)
6040 {
6041 
6042 	CTR2(KTR_PMAP, "%s(%#jx)", __func__, (uintmax_t)addr);
6043 	*decoded = addr;
6044 	*is_user = (addr < VM_MAXUSER_ADDRESS);
6045 	return (0);
6046 }
6047 
6048 static boolean_t
6049 mmu_radix_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
6050 {
6051 
6052 	CTR3(KTR_PMAP, "%s(%#x, %#x)", __func__, pa, size);
6053 	return (mem_valid(pa, size));
6054 }
6055 
6056 static void
6057 mmu_radix_scan_init()
6058 {
6059 
6060 	CTR1(KTR_PMAP, "%s()", __func__);
6061 	UNIMPLEMENTED();
6062 }
6063 
6064 static void
6065 mmu_radix_dumpsys_map(vm_paddr_t pa, size_t sz,
6066 	void **va)
6067 {
6068 	CTR4(KTR_PMAP, "%s(%#jx, %#zx, %p)", __func__, (uintmax_t)pa, sz, va);
6069 	UNIMPLEMENTED();
6070 }
6071 
6072 vm_offset_t
6073 mmu_radix_quick_enter_page(vm_page_t m)
6074 {
6075 	vm_paddr_t paddr;
6076 
6077 	CTR2(KTR_PMAP, "%s(%p)", __func__, m);
6078 	paddr = VM_PAGE_TO_PHYS(m);
6079 	return (PHYS_TO_DMAP(paddr));
6080 }
6081 
6082 void
6083 mmu_radix_quick_remove_page(vm_offset_t addr __unused)
6084 {
6085 	/* no work to do here */
6086 	CTR2(KTR_PMAP, "%s(%#x)", __func__, addr);
6087 }
6088 
6089 static void
6090 pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva)
6091 {
6092 	cpu_flush_dcache((void *)sva, eva - sva);
6093 }
6094 
6095 int
6096 mmu_radix_change_attr(vm_offset_t va, vm_size_t size,
6097     vm_memattr_t mode)
6098 {
6099 	int error;
6100 
6101 	CTR4(KTR_PMAP, "%s(%#x, %#zx, %d)", __func__, va, size, mode);
6102 	PMAP_LOCK(kernel_pmap);
6103 	error = pmap_change_attr_locked(va, size, mode, true);
6104 	PMAP_UNLOCK(kernel_pmap);
6105 	return (error);
6106 }
6107 
6108 static int
6109 pmap_change_attr_locked(vm_offset_t va, vm_size_t size, int mode, bool flush)
6110 {
6111 	vm_offset_t base, offset, tmpva;
6112 	vm_paddr_t pa_start, pa_end, pa_end1;
6113 	pml2_entry_t *l2e;
6114 	pml3_entry_t *l3e;
6115 	pt_entry_t *pte;
6116 	int cache_bits, error;
6117 	boolean_t changed;
6118 
6119 	PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED);
6120 	base = trunc_page(va);
6121 	offset = va & PAGE_MASK;
6122 	size = round_page(offset + size);
6123 
6124 	/*
6125 	 * Only supported on kernel virtual addresses, including the direct
6126 	 * map but excluding the recursive map.
6127 	 */
6128 	if (base < DMAP_MIN_ADDRESS)
6129 		return (EINVAL);
6130 
6131 	cache_bits = pmap_cache_bits(mode);
6132 	changed = FALSE;
6133 
6134 	/*
6135 	 * Pages that aren't mapped aren't supported.  Also break down 2MB pages
6136 	 * into 4KB pages if required.
6137 	 */
6138 	for (tmpva = base; tmpva < base + size; ) {
6139 		l2e = pmap_pml2e(kernel_pmap, tmpva);
6140 		if (l2e == NULL || *l2e == 0)
6141 			return (EINVAL);
6142 		if (be64toh(*l2e) & RPTE_LEAF) {
6143 			/*
6144 			 * If the current 1GB page already has the required
6145 			 * memory type, then we need not demote this page. Just
6146 			 * increment tmpva to the next 1GB page frame.
6147 			 */
6148 			if ((be64toh(*l2e) & RPTE_ATTR_MASK) == cache_bits) {
6149 				tmpva = trunc_1gpage(tmpva) + L2_PAGE_SIZE;
6150 				continue;
6151 			}
6152 
6153 			/*
6154 			 * If the current offset aligns with a 1GB page frame
6155 			 * and there is at least 1GB left within the range, then
6156 			 * we need not break down this page into 2MB pages.
6157 			 */
6158 			if ((tmpva & L2_PAGE_MASK) == 0 &&
6159 			    tmpva + L2_PAGE_MASK < base + size) {
6160 				tmpva += L2_PAGE_MASK;
6161 				continue;
6162 			}
6163 			if (!pmap_demote_l2e(kernel_pmap, l2e, tmpva))
6164 				return (ENOMEM);
6165 		}
6166 		l3e = pmap_l2e_to_l3e(l2e, tmpva);
6167 		KASSERT(l3e != NULL, ("no l3e entry for %#lx in %p\n",
6168 		    tmpva, l2e));
6169 		if (*l3e == 0)
6170 			return (EINVAL);
6171 		if (be64toh(*l3e) & RPTE_LEAF) {
6172 			/*
6173 			 * If the current 2MB page already has the required
6174 			 * memory type, then we need not demote this page. Just
6175 			 * increment tmpva to the next 2MB page frame.
6176 			 */
6177 			if ((be64toh(*l3e) & RPTE_ATTR_MASK) == cache_bits) {
6178 				tmpva = trunc_2mpage(tmpva) + L3_PAGE_SIZE;
6179 				continue;
6180 			}
6181 
6182 			/*
6183 			 * If the current offset aligns with a 2MB page frame
6184 			 * and there is at least 2MB left within the range, then
6185 			 * we need not break down this page into 4KB pages.
6186 			 */
6187 			if ((tmpva & L3_PAGE_MASK) == 0 &&
6188 			    tmpva + L3_PAGE_MASK < base + size) {
6189 				tmpva += L3_PAGE_SIZE;
6190 				continue;
6191 			}
6192 			if (!pmap_demote_l3e(kernel_pmap, l3e, tmpva))
6193 				return (ENOMEM);
6194 		}
6195 		pte = pmap_l3e_to_pte(l3e, tmpva);
6196 		if (*pte == 0)
6197 			return (EINVAL);
6198 		tmpva += PAGE_SIZE;
6199 	}
6200 	error = 0;
6201 
6202 	/*
6203 	 * Ok, all the pages exist, so run through them updating their
6204 	 * cache mode if required.
6205 	 */
6206 	pa_start = pa_end = 0;
6207 	for (tmpva = base; tmpva < base + size; ) {
6208 		l2e = pmap_pml2e(kernel_pmap, tmpva);
6209 		if (be64toh(*l2e) & RPTE_LEAF) {
6210 			if ((be64toh(*l2e) & RPTE_ATTR_MASK) != cache_bits) {
6211 				pmap_pte_attr(l2e, cache_bits,
6212 				    RPTE_ATTR_MASK);
6213 				changed = TRUE;
6214 			}
6215 			if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
6216 			    (*l2e & PG_PS_FRAME) < dmaplimit) {
6217 				if (pa_start == pa_end) {
6218 					/* Start physical address run. */
6219 					pa_start = be64toh(*l2e) & PG_PS_FRAME;
6220 					pa_end = pa_start + L2_PAGE_SIZE;
6221 				} else if (pa_end == (be64toh(*l2e) & PG_PS_FRAME))
6222 					pa_end += L2_PAGE_SIZE;
6223 				else {
6224 					/* Run ended, update direct map. */
6225 					error = pmap_change_attr_locked(
6226 					    PHYS_TO_DMAP(pa_start),
6227 					    pa_end - pa_start, mode, flush);
6228 					if (error != 0)
6229 						break;
6230 					/* Start physical address run. */
6231 					pa_start = be64toh(*l2e) & PG_PS_FRAME;
6232 					pa_end = pa_start + L2_PAGE_SIZE;
6233 				}
6234 			}
6235 			tmpva = trunc_1gpage(tmpva) + L2_PAGE_SIZE;
6236 			continue;
6237 		}
6238 		l3e = pmap_l2e_to_l3e(l2e, tmpva);
6239 		if (be64toh(*l3e) & RPTE_LEAF) {
6240 			if ((be64toh(*l3e) & RPTE_ATTR_MASK) != cache_bits) {
6241 				pmap_pte_attr(l3e, cache_bits,
6242 				    RPTE_ATTR_MASK);
6243 				changed = TRUE;
6244 			}
6245 			if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
6246 			    (be64toh(*l3e) & PG_PS_FRAME) < dmaplimit) {
6247 				if (pa_start == pa_end) {
6248 					/* Start physical address run. */
6249 					pa_start = be64toh(*l3e) & PG_PS_FRAME;
6250 					pa_end = pa_start + L3_PAGE_SIZE;
6251 				} else if (pa_end == (be64toh(*l3e) & PG_PS_FRAME))
6252 					pa_end += L3_PAGE_SIZE;
6253 				else {
6254 					/* Run ended, update direct map. */
6255 					error = pmap_change_attr_locked(
6256 					    PHYS_TO_DMAP(pa_start),
6257 					    pa_end - pa_start, mode, flush);
6258 					if (error != 0)
6259 						break;
6260 					/* Start physical address run. */
6261 					pa_start = be64toh(*l3e) & PG_PS_FRAME;
6262 					pa_end = pa_start + L3_PAGE_SIZE;
6263 				}
6264 			}
6265 			tmpva = trunc_2mpage(tmpva) + L3_PAGE_SIZE;
6266 		} else {
6267 			pte = pmap_l3e_to_pte(l3e, tmpva);
6268 			if ((be64toh(*pte) & RPTE_ATTR_MASK) != cache_bits) {
6269 				pmap_pte_attr(pte, cache_bits,
6270 				    RPTE_ATTR_MASK);
6271 				changed = TRUE;
6272 			}
6273 			if (tmpva >= VM_MIN_KERNEL_ADDRESS &&
6274 			    (be64toh(*pte) & PG_FRAME) < dmaplimit) {
6275 				if (pa_start == pa_end) {
6276 					/* Start physical address run. */
6277 					pa_start = be64toh(*pte) & PG_FRAME;
6278 					pa_end = pa_start + PAGE_SIZE;
6279 				} else if (pa_end == (be64toh(*pte) & PG_FRAME))
6280 					pa_end += PAGE_SIZE;
6281 				else {
6282 					/* Run ended, update direct map. */
6283 					error = pmap_change_attr_locked(
6284 					    PHYS_TO_DMAP(pa_start),
6285 					    pa_end - pa_start, mode, flush);
6286 					if (error != 0)
6287 						break;
6288 					/* Start physical address run. */
6289 					pa_start = be64toh(*pte) & PG_FRAME;
6290 					pa_end = pa_start + PAGE_SIZE;
6291 				}
6292 			}
6293 			tmpva += PAGE_SIZE;
6294 		}
6295 	}
6296 	if (error == 0 && pa_start != pa_end && pa_start < dmaplimit) {
6297 		pa_end1 = MIN(pa_end, dmaplimit);
6298 		if (pa_start != pa_end1)
6299 			error = pmap_change_attr_locked(PHYS_TO_DMAP(pa_start),
6300 			    pa_end1 - pa_start, mode, flush);
6301 	}
6302 
6303 	/*
6304 	 * Flush CPU caches if required to make sure any data isn't cached that
6305 	 * shouldn't be, etc.
6306 	 */
6307 	if (changed) {
6308 		pmap_invalidate_all(kernel_pmap);
6309 
6310 		if (flush)
6311 			pmap_invalidate_cache_range(base, tmpva);
6312 	}
6313 	return (error);
6314 }
6315 
6316 /*
6317  * Allocate physical memory for the vm_page array and map it into KVA,
6318  * attempting to back the vm_pages with domain-local memory.
6319  */
6320 void
6321 mmu_radix_page_array_startup(long pages)
6322 {
6323 #ifdef notyet
6324 	pml2_entry_t *l2e;
6325 	pml3_entry_t *pde;
6326 	pml3_entry_t newl3;
6327 	vm_offset_t va;
6328 	long pfn;
6329 	int domain, i;
6330 #endif
6331 	vm_paddr_t pa;
6332 	vm_offset_t start, end;
6333 
6334 	vm_page_array_size = pages;
6335 
6336 	start = VM_MIN_KERNEL_ADDRESS;
6337 	end = start + pages * sizeof(struct vm_page);
6338 
6339 	pa = vm_phys_early_alloc(0, end - start);
6340 
6341 	start = mmu_radix_map(&start, pa, end - start, VM_MEMATTR_DEFAULT);
6342 #ifdef notyet
6343 	/* TODO: NUMA vm_page_array.  Blocked out until then (copied from amd64). */
6344 	for (va = start; va < end; va += L3_PAGE_SIZE) {
6345 		pfn = first_page + (va - start) / sizeof(struct vm_page);
6346 		domain = vm_phys_domain(ptoa(pfn));
6347 		l2e = pmap_pml2e(kernel_pmap, va);
6348 		if ((be64toh(*l2e) & PG_V) == 0) {
6349 			pa = vm_phys_early_alloc(domain, PAGE_SIZE);
6350 			dump_add_page(pa);
6351 			pagezero(PHYS_TO_DMAP(pa));
6352 			pde_store(l2e, (pml2_entry_t)pa);
6353 		}
6354 		pde = pmap_l2e_to_l3e(l2e, va);
6355 		if ((be64toh(*pde) & PG_V) != 0)
6356 			panic("Unexpected pde %p", pde);
6357 		pa = vm_phys_early_alloc(domain, L3_PAGE_SIZE);
6358 		for (i = 0; i < NPDEPG; i++)
6359 			dump_add_page(pa + i * PAGE_SIZE);
6360 		newl3 = (pml3_entry_t)(pa | RPTE_EAA_P | RPTE_EAA_R | RPTE_EAA_W);
6361 		pte_store(pde, newl3);
6362 	}
6363 #endif
6364 	vm_page_array = (vm_page_t)start;
6365 }
6366 
6367 #ifdef DDB
6368 #include <sys/kdb.h>
6369 #include <ddb/ddb.h>
6370 
6371 static void
6372 pmap_pte_walk(pml1_entry_t *l1, vm_offset_t va)
6373 {
6374 	pml1_entry_t *l1e;
6375 	pml2_entry_t *l2e;
6376 	pml3_entry_t *l3e;
6377 	pt_entry_t *pte;
6378 
6379 	l1e = &l1[pmap_pml1e_index(va)];
6380 	db_printf("VA %#016lx l1e %#016lx", va, be64toh(*l1e));
6381 	if ((be64toh(*l1e) & PG_V) == 0) {
6382 		db_printf("\n");
6383 		return;
6384 	}
6385 	l2e = pmap_l1e_to_l2e(l1e, va);
6386 	db_printf(" l2e %#016lx", be64toh(*l2e));
6387 	if ((be64toh(*l2e) & PG_V) == 0 || (be64toh(*l2e) & RPTE_LEAF) != 0) {
6388 		db_printf("\n");
6389 		return;
6390 	}
6391 	l3e = pmap_l2e_to_l3e(l2e, va);
6392 	db_printf(" l3e %#016lx", be64toh(*l3e));
6393 	if ((be64toh(*l3e) & PG_V) == 0 || (be64toh(*l3e) & RPTE_LEAF) != 0) {
6394 		db_printf("\n");
6395 		return;
6396 	}
6397 	pte = pmap_l3e_to_pte(l3e, va);
6398 	db_printf(" pte %#016lx\n", be64toh(*pte));
6399 }
6400 
6401 void
6402 pmap_page_print_mappings(vm_page_t m)
6403 {
6404 	pmap_t pmap;
6405 	pv_entry_t pv;
6406 
6407 	db_printf("page %p(%lx)\n", m, m->phys_addr);
6408 	/* need to elide locks if running in ddb */
6409 	TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) {
6410 		db_printf("pv: %p ", pv);
6411 		db_printf("va: %#016lx ", pv->pv_va);
6412 		pmap = PV_PMAP(pv);
6413 		db_printf("pmap %p  ", pmap);
6414 		if (pmap != NULL) {
6415 			db_printf("asid: %lu\n", pmap->pm_pid);
6416 			pmap_pte_walk(pmap->pm_pml1, pv->pv_va);
6417 		}
6418 	}
6419 }
6420 
6421 DB_SHOW_COMMAND(pte, pmap_print_pte)
6422 {
6423 	vm_offset_t va;
6424 	pmap_t pmap;
6425 
6426 	if (!have_addr) {
6427 		db_printf("show pte addr\n");
6428 		return;
6429 	}
6430 	va = (vm_offset_t)addr;
6431 
6432 	if (va >= DMAP_MIN_ADDRESS)
6433 		pmap = kernel_pmap;
6434 	else if (kdb_thread != NULL)
6435 		pmap = vmspace_pmap(kdb_thread->td_proc->p_vmspace);
6436 	else
6437 		pmap = vmspace_pmap(curthread->td_proc->p_vmspace);
6438 
6439 	pmap_pte_walk(pmap->pm_pml1, va);
6440 }
6441 
6442 #endif
6443