xref: /titanic_41/usr/src/uts/i86pc/vm/hat_i86.c (revision e521049d0fdbd76727f036cb77c99c98d7ee1125)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright (c) 2010, Intel Corporation.
26  * All rights reserved.
27  */
28 /*
29  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
30  */
31 
32 /*
33  * VM - Hardware Address Translation management for i386 and amd64
34  *
35  * Implementation of the interfaces described in <common/vm/hat.h>
36  *
37  * Nearly all the details of how the hardware is managed should not be
38  * visible outside this layer except for misc. machine specific functions
39  * that work in conjunction with this code.
40  *
41  * Routines used only inside of i86pc/vm start with hati_ for HAT Internal.
42  */
43 
44 #include <sys/machparam.h>
45 #include <sys/machsystm.h>
46 #include <sys/mman.h>
47 #include <sys/types.h>
48 #include <sys/systm.h>
49 #include <sys/cpuvar.h>
50 #include <sys/thread.h>
51 #include <sys/proc.h>
52 #include <sys/cpu.h>
53 #include <sys/kmem.h>
54 #include <sys/disp.h>
55 #include <sys/shm.h>
56 #include <sys/sysmacros.h>
57 #include <sys/machparam.h>
58 #include <sys/vmem.h>
59 #include <sys/vmsystm.h>
60 #include <sys/promif.h>
61 #include <sys/var.h>
62 #include <sys/x86_archext.h>
63 #include <sys/atomic.h>
64 #include <sys/bitmap.h>
65 #include <sys/controlregs.h>
66 #include <sys/bootconf.h>
67 #include <sys/bootsvcs.h>
68 #include <sys/bootinfo.h>
69 #include <sys/archsystm.h>
70 
71 #include <vm/seg_kmem.h>
72 #include <vm/hat_i86.h>
73 #include <vm/as.h>
74 #include <vm/seg.h>
75 #include <vm/page.h>
76 #include <vm/seg_kp.h>
77 #include <vm/seg_kpm.h>
78 #include <vm/vm_dep.h>
79 #ifdef __xpv
80 #include <sys/hypervisor.h>
81 #endif
82 #include <vm/kboot_mmu.h>
83 #include <vm/seg_spt.h>
84 
85 #include <sys/cmn_err.h>
86 
87 /*
88  * Basic parameters for hat operation.
89  */
90 struct hat_mmu_info mmu;
91 
92 /*
93  * The page that is the kernel's top level pagetable.
94  *
95  * For 32 bit PAE support on i86pc, the kernel hat will use the 1st 4 entries
96  * on this 4K page for its top level page table. The remaining groups of
97  * 4 entries are used for per processor copies of user VLP pagetables for
98  * running threads.  See hat_switch() and reload_pae32() for details.
99  *
100  * vlp_page[0..3] - level==2 PTEs for kernel HAT
101  * vlp_page[4..7] - level==2 PTEs for user thread on cpu 0
102  * vlp_page[8..11]  - level==2 PTE for user thread on cpu 1
103  * etc...
104  */
105 static x86pte_t *vlp_page;
106 
107 /*
108  * forward declaration of internal utility routines
109  */
110 static x86pte_t hati_update_pte(htable_t *ht, uint_t entry, x86pte_t expected,
111 	x86pte_t new);
112 
113 /*
114  * The kernel address space exists in all HATs. To implement this the
115  * kernel reserves a fixed number of entries in the topmost level(s) of page
116  * tables. The values are setup during startup and then copied to every user
117  * hat created by hat_alloc(). This means that kernelbase must be:
118  *
119  *	  4Meg aligned for 32 bit kernels
120  *	512Gig aligned for x86_64 64 bit kernel
121  *
122  * The hat_kernel_range_ts describe what needs to be copied from kernel hat
123  * to each user hat.
124  */
125 typedef struct hat_kernel_range {
126 	level_t		hkr_level;
127 	uintptr_t	hkr_start_va;
128 	uintptr_t	hkr_end_va;	/* zero means to end of memory */
129 } hat_kernel_range_t;
130 #define	NUM_KERNEL_RANGE 2
131 static hat_kernel_range_t kernel_ranges[NUM_KERNEL_RANGE];
132 static int num_kernel_ranges;
133 
134 uint_t use_boot_reserve = 1;	/* cleared after early boot process */
135 uint_t can_steal_post_boot = 0;	/* set late in boot to enable stealing */
136 
137 /*
138  * enable_1gpg: controls 1g page support for user applications.
139  * By default, 1g pages are exported to user applications. enable_1gpg can
140  * be set to 0 to not export.
141  */
142 int	enable_1gpg = 1;
143 
144 /*
145  * AMD shanghai processors provide better management of 1gb ptes in its tlb.
146  * By default, 1g page support will be disabled for pre-shanghai AMD
147  * processors that don't have optimal tlb support for the 1g page size.
148  * chk_optimal_1gtlb can be set to 0 to force 1g page support on sub-optimal
149  * processors.
150  */
151 int	chk_optimal_1gtlb = 1;
152 
153 
154 #ifdef DEBUG
155 uint_t	map1gcnt;
156 #endif
157 
158 
159 /*
160  * A cpuset for all cpus. This is used for kernel address cross calls, since
161  * the kernel addresses apply to all cpus.
162  */
163 cpuset_t khat_cpuset;
164 
165 /*
166  * management stuff for hat structures
167  */
168 kmutex_t	hat_list_lock;
169 kcondvar_t	hat_list_cv;
170 kmem_cache_t	*hat_cache;
171 kmem_cache_t	*hat_hash_cache;
172 kmem_cache_t	*vlp_hash_cache;
173 
174 /*
175  * Simple statistics
176  */
177 struct hatstats hatstat;
178 
179 /*
180  * Some earlier hypervisor versions do not emulate cmpxchg of PTEs
181  * correctly.  For such hypervisors we must set PT_USER for kernel
182  * entries ourselves (normally the emulation would set PT_USER for
183  * kernel entries and PT_USER|PT_GLOBAL for user entries).  pt_kern is
184  * thus set appropriately.  Note that dboot/kbm is OK, as only the full
185  * HAT uses cmpxchg() and the other paths (hypercall etc.) were never
186  * incorrect.
187  */
188 int pt_kern;
189 
190 /*
191  * useful stuff for atomic access/clearing/setting REF/MOD/RO bits in page_t's.
192  */
193 extern void atomic_orb(uchar_t *addr, uchar_t val);
194 extern void atomic_andb(uchar_t *addr, uchar_t val);
195 
196 #ifndef __xpv
197 extern pfn_t memseg_get_start(struct memseg *);
198 #endif
199 
200 #define	PP_GETRM(pp, rmmask)    (pp->p_nrm & rmmask)
201 #define	PP_ISMOD(pp)		PP_GETRM(pp, P_MOD)
202 #define	PP_ISREF(pp)		PP_GETRM(pp, P_REF)
203 #define	PP_ISRO(pp)		PP_GETRM(pp, P_RO)
204 
205 #define	PP_SETRM(pp, rm)	atomic_orb(&(pp->p_nrm), rm)
206 #define	PP_SETMOD(pp)		PP_SETRM(pp, P_MOD)
207 #define	PP_SETREF(pp)		PP_SETRM(pp, P_REF)
208 #define	PP_SETRO(pp)		PP_SETRM(pp, P_RO)
209 
210 #define	PP_CLRRM(pp, rm)	atomic_andb(&(pp->p_nrm), ~(rm))
211 #define	PP_CLRMOD(pp)   	PP_CLRRM(pp, P_MOD)
212 #define	PP_CLRREF(pp)   	PP_CLRRM(pp, P_REF)
213 #define	PP_CLRRO(pp)    	PP_CLRRM(pp, P_RO)
214 #define	PP_CLRALL(pp)		PP_CLRRM(pp, P_MOD | P_REF | P_RO)
215 
216 /*
217  * kmem cache constructor for struct hat
218  */
219 /*ARGSUSED*/
220 static int
221 hati_constructor(void *buf, void *handle, int kmflags)
222 {
223 	hat_t	*hat = buf;
224 
225 	mutex_init(&hat->hat_mutex, NULL, MUTEX_DEFAULT, NULL);
226 	bzero(hat->hat_pages_mapped,
227 	    sizeof (pgcnt_t) * (mmu.max_page_level + 1));
228 	hat->hat_ism_pgcnt = 0;
229 	hat->hat_stats = 0;
230 	hat->hat_flags = 0;
231 	CPUSET_ZERO(hat->hat_cpus);
232 	hat->hat_htable = NULL;
233 	hat->hat_ht_hash = NULL;
234 	return (0);
235 }
236 
237 /*
238  * Allocate a hat structure for as. We also create the top level
239  * htable and initialize it to contain the kernel hat entries.
240  */
241 hat_t *
242 hat_alloc(struct as *as)
243 {
244 	hat_t			*hat;
245 	htable_t		*ht;	/* top level htable */
246 	uint_t			use_vlp;
247 	uint_t			r;
248 	hat_kernel_range_t	*rp;
249 	uintptr_t		va;
250 	uintptr_t		eva;
251 	uint_t			start;
252 	uint_t			cnt;
253 	htable_t		*src;
254 
255 	/*
256 	 * Once we start creating user process HATs we can enable
257 	 * the htable_steal() code.
258 	 */
259 	if (can_steal_post_boot == 0)
260 		can_steal_post_boot = 1;
261 
262 	ASSERT(AS_WRITE_HELD(as, &as->a_lock));
263 	hat = kmem_cache_alloc(hat_cache, KM_SLEEP);
264 	hat->hat_as = as;
265 	mutex_init(&hat->hat_mutex, NULL, MUTEX_DEFAULT, NULL);
266 	ASSERT(hat->hat_flags == 0);
267 
268 #if defined(__xpv)
269 	/*
270 	 * No VLP stuff on the hypervisor due to the 64-bit split top level
271 	 * page tables.  On 32-bit it's not needed as the hypervisor takes
272 	 * care of copying the top level PTEs to a below 4Gig page.
273 	 */
274 	use_vlp = 0;
275 #else	/* __xpv */
276 	/* 32 bit processes uses a VLP style hat when running with PAE */
277 #if defined(__amd64)
278 	use_vlp = (ttoproc(curthread)->p_model == DATAMODEL_ILP32);
279 #elif defined(__i386)
280 	use_vlp = mmu.pae_hat;
281 #endif
282 #endif	/* __xpv */
283 	if (use_vlp) {
284 		hat->hat_flags = HAT_VLP;
285 		bzero(hat->hat_vlp_ptes, VLP_SIZE);
286 	}
287 
288 	/*
289 	 * Allocate the htable hash
290 	 */
291 	if ((hat->hat_flags & HAT_VLP)) {
292 		hat->hat_num_hash = mmu.vlp_hash_cnt;
293 		hat->hat_ht_hash = kmem_cache_alloc(vlp_hash_cache, KM_SLEEP);
294 	} else {
295 		hat->hat_num_hash = mmu.hash_cnt;
296 		hat->hat_ht_hash = kmem_cache_alloc(hat_hash_cache, KM_SLEEP);
297 	}
298 	bzero(hat->hat_ht_hash, hat->hat_num_hash * sizeof (htable_t *));
299 
300 	/*
301 	 * Initialize Kernel HAT entries at the top of the top level page
302 	 * tables for the new hat.
303 	 */
304 	hat->hat_htable = NULL;
305 	hat->hat_ht_cached = NULL;
306 	XPV_DISALLOW_MIGRATE();
307 	ht = htable_create(hat, (uintptr_t)0, TOP_LEVEL(hat), NULL);
308 	hat->hat_htable = ht;
309 
310 #if defined(__amd64)
311 	if (hat->hat_flags & HAT_VLP)
312 		goto init_done;
313 #endif
314 
315 	for (r = 0; r < num_kernel_ranges; ++r) {
316 		rp = &kernel_ranges[r];
317 		for (va = rp->hkr_start_va; va != rp->hkr_end_va;
318 		    va += cnt * LEVEL_SIZE(rp->hkr_level)) {
319 
320 			if (rp->hkr_level == TOP_LEVEL(hat))
321 				ht = hat->hat_htable;
322 			else
323 				ht = htable_create(hat, va, rp->hkr_level,
324 				    NULL);
325 
326 			start = htable_va2entry(va, ht);
327 			cnt = HTABLE_NUM_PTES(ht) - start;
328 			eva = va +
329 			    ((uintptr_t)cnt << LEVEL_SHIFT(rp->hkr_level));
330 			if (rp->hkr_end_va != 0 &&
331 			    (eva > rp->hkr_end_va || eva == 0))
332 				cnt = htable_va2entry(rp->hkr_end_va, ht) -
333 				    start;
334 
335 #if defined(__i386) && !defined(__xpv)
336 			if (ht->ht_flags & HTABLE_VLP) {
337 				bcopy(&vlp_page[start],
338 				    &hat->hat_vlp_ptes[start],
339 				    cnt * sizeof (x86pte_t));
340 				continue;
341 			}
342 #endif
343 			src = htable_lookup(kas.a_hat, va, rp->hkr_level);
344 			ASSERT(src != NULL);
345 			x86pte_copy(src, ht, start, cnt);
346 			htable_release(src);
347 		}
348 	}
349 
350 init_done:
351 
352 #if defined(__xpv)
353 	/*
354 	 * Pin top level page tables after initializing them
355 	 */
356 	xen_pin(hat->hat_htable->ht_pfn, mmu.max_level);
357 #if defined(__amd64)
358 	xen_pin(hat->hat_user_ptable, mmu.max_level);
359 #endif
360 #endif
361 	XPV_ALLOW_MIGRATE();
362 
363 	/*
364 	 * Put it at the start of the global list of all hats (used by stealing)
365 	 *
366 	 * kas.a_hat is not in the list but is instead used to find the
367 	 * first and last items in the list.
368 	 *
369 	 * - kas.a_hat->hat_next points to the start of the user hats.
370 	 *   The list ends where hat->hat_next == NULL
371 	 *
372 	 * - kas.a_hat->hat_prev points to the last of the user hats.
373 	 *   The list begins where hat->hat_prev == NULL
374 	 */
375 	mutex_enter(&hat_list_lock);
376 	hat->hat_prev = NULL;
377 	hat->hat_next = kas.a_hat->hat_next;
378 	if (hat->hat_next)
379 		hat->hat_next->hat_prev = hat;
380 	else
381 		kas.a_hat->hat_prev = hat;
382 	kas.a_hat->hat_next = hat;
383 	mutex_exit(&hat_list_lock);
384 
385 	return (hat);
386 }
387 
388 /*
389  * process has finished executing but as has not been cleaned up yet.
390  */
391 /*ARGSUSED*/
392 void
393 hat_free_start(hat_t *hat)
394 {
395 	ASSERT(AS_WRITE_HELD(hat->hat_as, &hat->hat_as->a_lock));
396 
397 	/*
398 	 * If the hat is currently a stealing victim, wait for the stealing
399 	 * to finish.  Once we mark it as HAT_FREEING, htable_steal()
400 	 * won't look at its pagetables anymore.
401 	 */
402 	mutex_enter(&hat_list_lock);
403 	while (hat->hat_flags & HAT_VICTIM)
404 		cv_wait(&hat_list_cv, &hat_list_lock);
405 	hat->hat_flags |= HAT_FREEING;
406 	mutex_exit(&hat_list_lock);
407 }
408 
409 /*
410  * An address space is being destroyed, so we destroy the associated hat.
411  */
412 void
413 hat_free_end(hat_t *hat)
414 {
415 	kmem_cache_t *cache;
416 
417 	ASSERT(hat->hat_flags & HAT_FREEING);
418 
419 	/*
420 	 * must not be running on the given hat
421 	 */
422 	ASSERT(CPU->cpu_current_hat != hat);
423 
424 	/*
425 	 * Remove it from the list of HATs
426 	 */
427 	mutex_enter(&hat_list_lock);
428 	if (hat->hat_prev)
429 		hat->hat_prev->hat_next = hat->hat_next;
430 	else
431 		kas.a_hat->hat_next = hat->hat_next;
432 	if (hat->hat_next)
433 		hat->hat_next->hat_prev = hat->hat_prev;
434 	else
435 		kas.a_hat->hat_prev = hat->hat_prev;
436 	mutex_exit(&hat_list_lock);
437 	hat->hat_next = hat->hat_prev = NULL;
438 
439 #if defined(__xpv)
440 	/*
441 	 * On the hypervisor, unpin top level page table(s)
442 	 */
443 	xen_unpin(hat->hat_htable->ht_pfn);
444 #if defined(__amd64)
445 	xen_unpin(hat->hat_user_ptable);
446 #endif
447 #endif
448 
449 	/*
450 	 * Make a pass through the htables freeing them all up.
451 	 */
452 	htable_purge_hat(hat);
453 
454 	/*
455 	 * Decide which kmem cache the hash table came from, then free it.
456 	 */
457 	if (hat->hat_flags & HAT_VLP)
458 		cache = vlp_hash_cache;
459 	else
460 		cache = hat_hash_cache;
461 	kmem_cache_free(cache, hat->hat_ht_hash);
462 	hat->hat_ht_hash = NULL;
463 
464 	hat->hat_flags = 0;
465 	kmem_cache_free(hat_cache, hat);
466 }
467 
468 /*
469  * round kernelbase down to a supported value to use for _userlimit
470  *
471  * userlimit must be aligned down to an entry in the top level htable.
472  * The one exception is for 32 bit HAT's running PAE.
473  */
474 uintptr_t
475 hat_kernelbase(uintptr_t va)
476 {
477 #if defined(__i386)
478 	va &= LEVEL_MASK(1);
479 #endif
480 	if (IN_VA_HOLE(va))
481 		panic("_userlimit %p will fall in VA hole\n", (void *)va);
482 	return (va);
483 }
484 
485 /*
486  *
487  */
488 static void
489 set_max_page_level()
490 {
491 	level_t lvl;
492 
493 	if (!kbm_largepage_support) {
494 		lvl = 0;
495 	} else {
496 		if (is_x86_feature(x86_featureset, X86FSET_1GPG)) {
497 			lvl = 2;
498 			if (chk_optimal_1gtlb &&
499 			    cpuid_opteron_erratum(CPU, 6671130)) {
500 				lvl = 1;
501 			}
502 			if (plat_mnode_xcheck(LEVEL_SIZE(2) >>
503 			    LEVEL_SHIFT(0))) {
504 				lvl = 1;
505 			}
506 		} else {
507 			lvl = 1;
508 		}
509 	}
510 	mmu.max_page_level = lvl;
511 
512 	if ((lvl == 2) && (enable_1gpg == 0))
513 		mmu.umax_page_level = 1;
514 	else
515 		mmu.umax_page_level = lvl;
516 }
517 
518 /*
519  * Initialize hat data structures based on processor MMU information.
520  */
521 void
522 mmu_init(void)
523 {
524 	uint_t max_htables;
525 	uint_t pa_bits;
526 	uint_t va_bits;
527 	int i;
528 
529 	/*
530 	 * If CPU enabled the page table global bit, use it for the kernel
531 	 * This is bit 7 in CR4 (PGE - Page Global Enable).
532 	 */
533 	if (is_x86_feature(x86_featureset, X86FSET_PGE) &&
534 	    (getcr4() & CR4_PGE) != 0)
535 		mmu.pt_global = PT_GLOBAL;
536 
537 	/*
538 	 * Detect NX and PAE usage.
539 	 */
540 	mmu.pae_hat = kbm_pae_support;
541 	if (kbm_nx_support)
542 		mmu.pt_nx = PT_NX;
543 	else
544 		mmu.pt_nx = 0;
545 
546 	/*
547 	 * Use CPU info to set various MMU parameters
548 	 */
549 	cpuid_get_addrsize(CPU, &pa_bits, &va_bits);
550 
551 	if (va_bits < sizeof (void *) * NBBY) {
552 		mmu.hole_start = (1ul << (va_bits - 1));
553 		mmu.hole_end = 0ul - mmu.hole_start - 1;
554 	} else {
555 		mmu.hole_end = 0;
556 		mmu.hole_start = mmu.hole_end - 1;
557 	}
558 #if defined(OPTERON_ERRATUM_121)
559 	/*
560 	 * If erratum 121 has already been detected at this time, hole_start
561 	 * contains the value to be subtracted from mmu.hole_start.
562 	 */
563 	ASSERT(hole_start == 0 || opteron_erratum_121 != 0);
564 	hole_start = mmu.hole_start - hole_start;
565 #else
566 	hole_start = mmu.hole_start;
567 #endif
568 	hole_end = mmu.hole_end;
569 
570 	mmu.highest_pfn = mmu_btop((1ull << pa_bits) - 1);
571 	if (mmu.pae_hat == 0 && pa_bits > 32)
572 		mmu.highest_pfn = PFN_4G - 1;
573 
574 	if (mmu.pae_hat) {
575 		mmu.pte_size = 8;	/* 8 byte PTEs */
576 		mmu.pte_size_shift = 3;
577 	} else {
578 		mmu.pte_size = 4;	/* 4 byte PTEs */
579 		mmu.pte_size_shift = 2;
580 	}
581 
582 	if (mmu.pae_hat && !is_x86_feature(x86_featureset, X86FSET_PAE))
583 		panic("Processor does not support PAE");
584 
585 	if (!is_x86_feature(x86_featureset, X86FSET_CX8))
586 		panic("Processor does not support cmpxchg8b instruction");
587 
588 #if defined(__amd64)
589 
590 	mmu.num_level = 4;
591 	mmu.max_level = 3;
592 	mmu.ptes_per_table = 512;
593 	mmu.top_level_count = 512;
594 
595 	mmu.level_shift[0] = 12;
596 	mmu.level_shift[1] = 21;
597 	mmu.level_shift[2] = 30;
598 	mmu.level_shift[3] = 39;
599 
600 #elif defined(__i386)
601 
602 	if (mmu.pae_hat) {
603 		mmu.num_level = 3;
604 		mmu.max_level = 2;
605 		mmu.ptes_per_table = 512;
606 		mmu.top_level_count = 4;
607 
608 		mmu.level_shift[0] = 12;
609 		mmu.level_shift[1] = 21;
610 		mmu.level_shift[2] = 30;
611 
612 	} else {
613 		mmu.num_level = 2;
614 		mmu.max_level = 1;
615 		mmu.ptes_per_table = 1024;
616 		mmu.top_level_count = 1024;
617 
618 		mmu.level_shift[0] = 12;
619 		mmu.level_shift[1] = 22;
620 	}
621 
622 #endif	/* __i386 */
623 
624 	for (i = 0; i < mmu.num_level; ++i) {
625 		mmu.level_size[i] = 1UL << mmu.level_shift[i];
626 		mmu.level_offset[i] = mmu.level_size[i] - 1;
627 		mmu.level_mask[i] = ~mmu.level_offset[i];
628 	}
629 
630 	set_max_page_level();
631 
632 	mmu_page_sizes = mmu.max_page_level + 1;
633 	mmu_exported_page_sizes = mmu.umax_page_level + 1;
634 
635 	/* restrict legacy applications from using pagesizes 1g and above */
636 	mmu_legacy_page_sizes =
637 	    (mmu_exported_page_sizes > 2) ? 2 : mmu_exported_page_sizes;
638 
639 
640 	for (i = 0; i <= mmu.max_page_level; ++i) {
641 		mmu.pte_bits[i] = PT_VALID | pt_kern;
642 		if (i > 0)
643 			mmu.pte_bits[i] |= PT_PAGESIZE;
644 	}
645 
646 	/*
647 	 * NOTE Legacy 32 bit PAE mode only has the P_VALID bit at top level.
648 	 */
649 	for (i = 1; i < mmu.num_level; ++i)
650 		mmu.ptp_bits[i] = PT_PTPBITS;
651 
652 #if defined(__i386)
653 	mmu.ptp_bits[2] = PT_VALID;
654 #endif
655 
656 	/*
657 	 * Compute how many hash table entries to have per process for htables.
658 	 * We start with 1 page's worth of entries.
659 	 *
660 	 * If physical memory is small, reduce the amount need to cover it.
661 	 */
662 	max_htables = physmax / mmu.ptes_per_table;
663 	mmu.hash_cnt = MMU_PAGESIZE / sizeof (htable_t *);
664 	while (mmu.hash_cnt > 16 && mmu.hash_cnt >= max_htables)
665 		mmu.hash_cnt >>= 1;
666 	mmu.vlp_hash_cnt = mmu.hash_cnt;
667 
668 #if defined(__amd64)
669 	/*
670 	 * If running in 64 bits and physical memory is large,
671 	 * increase the size of the cache to cover all of memory for
672 	 * a 64 bit process.
673 	 */
674 #define	HASH_MAX_LENGTH 4
675 	while (mmu.hash_cnt * HASH_MAX_LENGTH < max_htables)
676 		mmu.hash_cnt <<= 1;
677 #endif
678 }
679 
680 
681 /*
682  * initialize hat data structures
683  */
684 void
685 hat_init()
686 {
687 #if defined(__i386)
688 	/*
689 	 * _userlimit must be aligned correctly
690 	 */
691 	if ((_userlimit & LEVEL_MASK(1)) != _userlimit) {
692 		prom_printf("hat_init(): _userlimit=%p, not aligned at %p\n",
693 		    (void *)_userlimit, (void *)LEVEL_SIZE(1));
694 		halt("hat_init(): Unable to continue");
695 	}
696 #endif
697 
698 	cv_init(&hat_list_cv, NULL, CV_DEFAULT, NULL);
699 
700 	/*
701 	 * initialize kmem caches
702 	 */
703 	htable_init();
704 	hment_init();
705 
706 	hat_cache = kmem_cache_create("hat_t",
707 	    sizeof (hat_t), 0, hati_constructor, NULL, NULL,
708 	    NULL, 0, 0);
709 
710 	hat_hash_cache = kmem_cache_create("HatHash",
711 	    mmu.hash_cnt * sizeof (htable_t *), 0, NULL, NULL, NULL,
712 	    NULL, 0, 0);
713 
714 	/*
715 	 * VLP hats can use a smaller hash table size on large memroy machines
716 	 */
717 	if (mmu.hash_cnt == mmu.vlp_hash_cnt) {
718 		vlp_hash_cache = hat_hash_cache;
719 	} else {
720 		vlp_hash_cache = kmem_cache_create("HatVlpHash",
721 		    mmu.vlp_hash_cnt * sizeof (htable_t *), 0, NULL, NULL, NULL,
722 		    NULL, 0, 0);
723 	}
724 
725 	/*
726 	 * Set up the kernel's hat
727 	 */
728 	AS_LOCK_ENTER(&kas, &kas.a_lock, RW_WRITER);
729 	kas.a_hat = kmem_cache_alloc(hat_cache, KM_NOSLEEP);
730 	mutex_init(&kas.a_hat->hat_mutex, NULL, MUTEX_DEFAULT, NULL);
731 	kas.a_hat->hat_as = &kas;
732 	kas.a_hat->hat_flags = 0;
733 	AS_LOCK_EXIT(&kas, &kas.a_lock);
734 
735 	CPUSET_ZERO(khat_cpuset);
736 	CPUSET_ADD(khat_cpuset, CPU->cpu_id);
737 
738 	/*
739 	 * The kernel hat's next pointer serves as the head of the hat list .
740 	 * The kernel hat's prev pointer tracks the last hat on the list for
741 	 * htable_steal() to use.
742 	 */
743 	kas.a_hat->hat_next = NULL;
744 	kas.a_hat->hat_prev = NULL;
745 
746 	/*
747 	 * Allocate an htable hash bucket for the kernel
748 	 * XX64 - tune for 64 bit procs
749 	 */
750 	kas.a_hat->hat_num_hash = mmu.hash_cnt;
751 	kas.a_hat->hat_ht_hash = kmem_cache_alloc(hat_hash_cache, KM_NOSLEEP);
752 	bzero(kas.a_hat->hat_ht_hash, mmu.hash_cnt * sizeof (htable_t *));
753 
754 	/*
755 	 * zero out the top level and cached htable pointers
756 	 */
757 	kas.a_hat->hat_ht_cached = NULL;
758 	kas.a_hat->hat_htable = NULL;
759 
760 	/*
761 	 * Pre-allocate hrm_hashtab before enabling the collection of
762 	 * refmod statistics.  Allocating on the fly would mean us
763 	 * running the risk of suffering recursive mutex enters or
764 	 * deadlocks.
765 	 */
766 	hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *),
767 	    KM_SLEEP);
768 }
769 
770 /*
771  * Prepare CPU specific pagetables for VLP processes on 64 bit kernels.
772  *
773  * Each CPU has a set of 2 pagetables that are reused for any 32 bit
774  * process it runs. They are the top level pagetable, hci_vlp_l3ptes, and
775  * the next to top level table for the bottom 512 Gig, hci_vlp_l2ptes.
776  */
777 /*ARGSUSED*/
778 static void
779 hat_vlp_setup(struct cpu *cpu)
780 {
781 #if defined(__amd64) && !defined(__xpv)
782 	struct hat_cpu_info *hci = cpu->cpu_hat_info;
783 	pfn_t pfn;
784 
785 	/*
786 	 * allocate the level==2 page table for the bottom most
787 	 * 512Gig of address space (this is where 32 bit apps live)
788 	 */
789 	ASSERT(hci != NULL);
790 	hci->hci_vlp_l2ptes = kmem_zalloc(MMU_PAGESIZE, KM_SLEEP);
791 
792 	/*
793 	 * Allocate a top level pagetable and copy the kernel's
794 	 * entries into it. Then link in hci_vlp_l2ptes in the 1st entry.
795 	 */
796 	hci->hci_vlp_l3ptes = kmem_zalloc(MMU_PAGESIZE, KM_SLEEP);
797 	hci->hci_vlp_pfn =
798 	    hat_getpfnum(kas.a_hat, (caddr_t)hci->hci_vlp_l3ptes);
799 	ASSERT(hci->hci_vlp_pfn != PFN_INVALID);
800 	bcopy(vlp_page, hci->hci_vlp_l3ptes, MMU_PAGESIZE);
801 
802 	pfn = hat_getpfnum(kas.a_hat, (caddr_t)hci->hci_vlp_l2ptes);
803 	ASSERT(pfn != PFN_INVALID);
804 	hci->hci_vlp_l3ptes[0] = MAKEPTP(pfn, 2);
805 #endif /* __amd64 && !__xpv */
806 }
807 
808 /*ARGSUSED*/
809 static void
810 hat_vlp_teardown(cpu_t *cpu)
811 {
812 #if defined(__amd64) && !defined(__xpv)
813 	struct hat_cpu_info *hci;
814 
815 	if ((hci = cpu->cpu_hat_info) == NULL)
816 		return;
817 	if (hci->hci_vlp_l2ptes)
818 		kmem_free(hci->hci_vlp_l2ptes, MMU_PAGESIZE);
819 	if (hci->hci_vlp_l3ptes)
820 		kmem_free(hci->hci_vlp_l3ptes, MMU_PAGESIZE);
821 #endif
822 }
823 
824 #define	NEXT_HKR(r, l, s, e) {			\
825 	kernel_ranges[r].hkr_level = l;		\
826 	kernel_ranges[r].hkr_start_va = s;	\
827 	kernel_ranges[r].hkr_end_va = e;	\
828 	++r;					\
829 }
830 
831 /*
832  * Finish filling in the kernel hat.
833  * Pre fill in all top level kernel page table entries for the kernel's
834  * part of the address range.  From this point on we can't use any new
835  * kernel large pages if they need PTE's at max_level
836  *
837  * create the kmap mappings.
838  */
839 void
840 hat_init_finish(void)
841 {
842 	size_t		size;
843 	uint_t		r = 0;
844 	uintptr_t	va;
845 	hat_kernel_range_t *rp;
846 
847 
848 	/*
849 	 * We are now effectively running on the kernel hat.
850 	 * Clearing use_boot_reserve shuts off using the pre-allocated boot
851 	 * reserve for all HAT allocations.  From here on, the reserves are
852 	 * only used when avoiding recursion in kmem_alloc().
853 	 */
854 	use_boot_reserve = 0;
855 	htable_adjust_reserve();
856 
857 	/*
858 	 * User HATs are initialized with copies of all kernel mappings in
859 	 * higher level page tables. Ensure that those entries exist.
860 	 */
861 #if defined(__amd64)
862 
863 	NEXT_HKR(r, 3, kernelbase, 0);
864 #if defined(__xpv)
865 	NEXT_HKR(r, 3, HYPERVISOR_VIRT_START, HYPERVISOR_VIRT_END);
866 #endif
867 
868 #elif defined(__i386)
869 
870 #if !defined(__xpv)
871 	if (mmu.pae_hat) {
872 		va = kernelbase;
873 		if ((va & LEVEL_MASK(2)) != va) {
874 			va = P2ROUNDUP(va, LEVEL_SIZE(2));
875 			NEXT_HKR(r, 1, kernelbase, va);
876 		}
877 		if (va != 0)
878 			NEXT_HKR(r, 2, va, 0);
879 	} else
880 #endif /* __xpv */
881 		NEXT_HKR(r, 1, kernelbase, 0);
882 
883 #endif /* __i386 */
884 
885 	num_kernel_ranges = r;
886 
887 	/*
888 	 * Create all the kernel pagetables that will have entries
889 	 * shared to user HATs.
890 	 */
891 	for (r = 0; r < num_kernel_ranges; ++r) {
892 		rp = &kernel_ranges[r];
893 		for (va = rp->hkr_start_va; va != rp->hkr_end_va;
894 		    va += LEVEL_SIZE(rp->hkr_level)) {
895 			htable_t *ht;
896 
897 			if (IN_HYPERVISOR_VA(va))
898 				continue;
899 
900 			/* can/must skip if a page mapping already exists */
901 			if (rp->hkr_level <= mmu.max_page_level &&
902 			    (ht = htable_getpage(kas.a_hat, va, NULL)) !=
903 			    NULL) {
904 				htable_release(ht);
905 				continue;
906 			}
907 
908 			(void) htable_create(kas.a_hat, va, rp->hkr_level - 1,
909 			    NULL);
910 		}
911 	}
912 
913 	/*
914 	 * 32 bit PAE metal kernels use only 4 of the 512 entries in the
915 	 * page holding the top level pagetable. We use the remainder for
916 	 * the "per CPU" page tables for VLP processes.
917 	 * Map the top level kernel pagetable into the kernel to make
918 	 * it easy to use bcopy access these tables.
919 	 */
920 	if (mmu.pae_hat) {
921 		vlp_page = vmem_alloc(heap_arena, MMU_PAGESIZE, VM_SLEEP);
922 		hat_devload(kas.a_hat, (caddr_t)vlp_page, MMU_PAGESIZE,
923 		    kas.a_hat->hat_htable->ht_pfn,
924 #if !defined(__xpv)
925 		    PROT_WRITE |
926 #endif
927 		    PROT_READ | HAT_NOSYNC | HAT_UNORDERED_OK,
928 		    HAT_LOAD | HAT_LOAD_NOCONSIST);
929 	}
930 	hat_vlp_setup(CPU);
931 
932 	/*
933 	 * Create kmap (cached mappings of kernel PTEs)
934 	 * for 32 bit we map from segmap_start .. ekernelheap
935 	 * for 64 bit we map from segmap_start .. segmap_start + segmapsize;
936 	 */
937 #if defined(__i386)
938 	size = (uintptr_t)ekernelheap - segmap_start;
939 #elif defined(__amd64)
940 	size = segmapsize;
941 #endif
942 	hat_kmap_init((uintptr_t)segmap_start, size);
943 }
944 
945 /*
946  * On 32 bit PAE mode, PTE's are 64 bits, but ordinary atomic memory references
947  * are 32 bit, so for safety we must use atomic_cas_64() to install these.
948  */
949 #ifdef __i386
950 static void
951 reload_pae32(hat_t *hat, cpu_t *cpu)
952 {
953 	x86pte_t *src;
954 	x86pte_t *dest;
955 	x86pte_t pte;
956 	int i;
957 
958 	/*
959 	 * Load the 4 entries of the level 2 page table into this
960 	 * cpu's range of the vlp_page and point cr3 at them.
961 	 */
962 	ASSERT(mmu.pae_hat);
963 	src = hat->hat_vlp_ptes;
964 	dest = vlp_page + (cpu->cpu_id + 1) * VLP_NUM_PTES;
965 	for (i = 0; i < VLP_NUM_PTES; ++i) {
966 		for (;;) {
967 			pte = dest[i];
968 			if (pte == src[i])
969 				break;
970 			if (atomic_cas_64(dest + i, pte, src[i]) != src[i])
971 				break;
972 		}
973 	}
974 }
975 #endif
976 
977 /*
978  * Switch to a new active hat, maintaining bit masks to track active CPUs.
979  *
980  * On the 32-bit PAE hypervisor, %cr3 is a 64-bit value, on metal it
981  * remains a 32-bit value.
982  */
983 void
984 hat_switch(hat_t *hat)
985 {
986 	uint64_t	newcr3;
987 	cpu_t		*cpu = CPU;
988 	hat_t		*old = cpu->cpu_current_hat;
989 
990 	/*
991 	 * set up this information first, so we don't miss any cross calls
992 	 */
993 	if (old != NULL) {
994 		if (old == hat)
995 			return;
996 		if (old != kas.a_hat)
997 			CPUSET_ATOMIC_DEL(old->hat_cpus, cpu->cpu_id);
998 	}
999 
1000 	/*
1001 	 * Add this CPU to the active set for this HAT.
1002 	 */
1003 	if (hat != kas.a_hat) {
1004 		CPUSET_ATOMIC_ADD(hat->hat_cpus, cpu->cpu_id);
1005 	}
1006 	cpu->cpu_current_hat = hat;
1007 
1008 	/*
1009 	 * now go ahead and load cr3
1010 	 */
1011 	if (hat->hat_flags & HAT_VLP) {
1012 #if defined(__amd64)
1013 		x86pte_t *vlpptep = cpu->cpu_hat_info->hci_vlp_l2ptes;
1014 
1015 		VLP_COPY(hat->hat_vlp_ptes, vlpptep);
1016 		newcr3 = MAKECR3(cpu->cpu_hat_info->hci_vlp_pfn);
1017 #elif defined(__i386)
1018 		reload_pae32(hat, cpu);
1019 		newcr3 = MAKECR3(kas.a_hat->hat_htable->ht_pfn) +
1020 		    (cpu->cpu_id + 1) * VLP_SIZE;
1021 #endif
1022 	} else {
1023 		newcr3 = MAKECR3((uint64_t)hat->hat_htable->ht_pfn);
1024 	}
1025 #ifdef __xpv
1026 	{
1027 		struct mmuext_op t[2];
1028 		uint_t retcnt;
1029 		uint_t opcnt = 1;
1030 
1031 		t[0].cmd = MMUEXT_NEW_BASEPTR;
1032 		t[0].arg1.mfn = mmu_btop(pa_to_ma(newcr3));
1033 #if defined(__amd64)
1034 		/*
1035 		 * There's an interesting problem here, as to what to
1036 		 * actually specify when switching to the kernel hat.
1037 		 * For now we'll reuse the kernel hat again.
1038 		 */
1039 		t[1].cmd = MMUEXT_NEW_USER_BASEPTR;
1040 		if (hat == kas.a_hat)
1041 			t[1].arg1.mfn = mmu_btop(pa_to_ma(newcr3));
1042 		else
1043 			t[1].arg1.mfn = pfn_to_mfn(hat->hat_user_ptable);
1044 		++opcnt;
1045 #endif	/* __amd64 */
1046 		if (HYPERVISOR_mmuext_op(t, opcnt, &retcnt, DOMID_SELF) < 0)
1047 			panic("HYPERVISOR_mmu_update() failed");
1048 		ASSERT(retcnt == opcnt);
1049 
1050 	}
1051 #else
1052 	setcr3(newcr3);
1053 #endif
1054 	ASSERT(cpu == CPU);
1055 }
1056 
1057 /*
1058  * Utility to return a valid x86pte_t from protections, pfn, and level number
1059  */
1060 static x86pte_t
1061 hati_mkpte(pfn_t pfn, uint_t attr, level_t level, uint_t flags)
1062 {
1063 	x86pte_t	pte;
1064 	uint_t		cache_attr = attr & HAT_ORDER_MASK;
1065 
1066 	pte = MAKEPTE(pfn, level);
1067 
1068 	if (attr & PROT_WRITE)
1069 		PTE_SET(pte, PT_WRITABLE);
1070 
1071 	if (attr & PROT_USER)
1072 		PTE_SET(pte, PT_USER);
1073 
1074 	if (!(attr & PROT_EXEC))
1075 		PTE_SET(pte, mmu.pt_nx);
1076 
1077 	/*
1078 	 * Set the software bits used track ref/mod sync's and hments.
1079 	 * If not using REF/MOD, set them to avoid h/w rewriting PTEs.
1080 	 */
1081 	if (flags & HAT_LOAD_NOCONSIST)
1082 		PTE_SET(pte, PT_NOCONSIST | PT_REF | PT_MOD);
1083 	else if (attr & HAT_NOSYNC)
1084 		PTE_SET(pte, PT_NOSYNC | PT_REF | PT_MOD);
1085 
1086 	/*
1087 	 * Set the caching attributes in the PTE. The combination
1088 	 * of attributes are poorly defined, so we pay attention
1089 	 * to them in the given order.
1090 	 *
1091 	 * The test for HAT_STRICTORDER is different because it's defined
1092 	 * as "0" - which was a stupid thing to do, but is too late to change!
1093 	 */
1094 	if (cache_attr == HAT_STRICTORDER) {
1095 		PTE_SET(pte, PT_NOCACHE);
1096 	/*LINTED [Lint hates empty ifs, but it's the obvious way to do this] */
1097 	} else if (cache_attr & (HAT_UNORDERED_OK | HAT_STORECACHING_OK)) {
1098 		/* nothing to set */;
1099 	} else if (cache_attr & (HAT_MERGING_OK | HAT_LOADCACHING_OK)) {
1100 		PTE_SET(pte, PT_NOCACHE);
1101 		if (is_x86_feature(x86_featureset, X86FSET_PAT))
1102 			PTE_SET(pte, (level == 0) ? PT_PAT_4K : PT_PAT_LARGE);
1103 		else
1104 			PTE_SET(pte, PT_WRITETHRU);
1105 	} else {
1106 		panic("hati_mkpte(): bad caching attributes: %x\n", cache_attr);
1107 	}
1108 
1109 	return (pte);
1110 }
1111 
1112 /*
1113  * Duplicate address translations of the parent to the child.
1114  * This function really isn't used anymore.
1115  */
1116 /*ARGSUSED*/
1117 int
1118 hat_dup(hat_t *old, hat_t *new, caddr_t addr, size_t len, uint_t flag)
1119 {
1120 	ASSERT((uintptr_t)addr < kernelbase);
1121 	ASSERT(new != kas.a_hat);
1122 	ASSERT(old != kas.a_hat);
1123 	return (0);
1124 }
1125 
1126 /*
1127  * Allocate any hat resources required for a process being swapped in.
1128  */
1129 /*ARGSUSED*/
1130 void
1131 hat_swapin(hat_t *hat)
1132 {
1133 	/* do nothing - we let everything fault back in */
1134 }
1135 
1136 /*
1137  * Unload all translations associated with an address space of a process
1138  * that is being swapped out.
1139  */
1140 void
1141 hat_swapout(hat_t *hat)
1142 {
1143 	uintptr_t	vaddr = (uintptr_t)0;
1144 	uintptr_t	eaddr = _userlimit;
1145 	htable_t	*ht = NULL;
1146 	level_t		l;
1147 
1148 	XPV_DISALLOW_MIGRATE();
1149 	/*
1150 	 * We can't just call hat_unload(hat, 0, _userlimit...)  here, because
1151 	 * seg_spt and shared pagetables can't be swapped out.
1152 	 * Take a look at segspt_shmswapout() - it's a big no-op.
1153 	 *
1154 	 * Instead we'll walk through all the address space and unload
1155 	 * any mappings which we are sure are not shared, not locked.
1156 	 */
1157 	ASSERT(IS_PAGEALIGNED(vaddr));
1158 	ASSERT(IS_PAGEALIGNED(eaddr));
1159 	ASSERT(AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1160 	if ((uintptr_t)hat->hat_as->a_userlimit < eaddr)
1161 		eaddr = (uintptr_t)hat->hat_as->a_userlimit;
1162 
1163 	while (vaddr < eaddr) {
1164 		(void) htable_walk(hat, &ht, &vaddr, eaddr);
1165 		if (ht == NULL)
1166 			break;
1167 
1168 		ASSERT(!IN_VA_HOLE(vaddr));
1169 
1170 		/*
1171 		 * If the page table is shared skip its entire range.
1172 		 */
1173 		l = ht->ht_level;
1174 		if (ht->ht_flags & HTABLE_SHARED_PFN) {
1175 			vaddr = ht->ht_vaddr + LEVEL_SIZE(l + 1);
1176 			htable_release(ht);
1177 			ht = NULL;
1178 			continue;
1179 		}
1180 
1181 		/*
1182 		 * If the page table has no locked entries, unload this one.
1183 		 */
1184 		if (ht->ht_lock_cnt == 0)
1185 			hat_unload(hat, (caddr_t)vaddr, LEVEL_SIZE(l),
1186 			    HAT_UNLOAD_UNMAP);
1187 
1188 		/*
1189 		 * If we have a level 0 page table with locked entries,
1190 		 * skip the entire page table, otherwise skip just one entry.
1191 		 */
1192 		if (ht->ht_lock_cnt > 0 && l == 0)
1193 			vaddr = ht->ht_vaddr + LEVEL_SIZE(1);
1194 		else
1195 			vaddr += LEVEL_SIZE(l);
1196 	}
1197 	if (ht)
1198 		htable_release(ht);
1199 
1200 	/*
1201 	 * We're in swapout because the system is low on memory, so
1202 	 * go back and flush all the htables off the cached list.
1203 	 */
1204 	htable_purge_hat(hat);
1205 	XPV_ALLOW_MIGRATE();
1206 }
1207 
1208 /*
1209  * returns number of bytes that have valid mappings in hat.
1210  */
1211 size_t
1212 hat_get_mapped_size(hat_t *hat)
1213 {
1214 	size_t total = 0;
1215 	int l;
1216 
1217 	for (l = 0; l <= mmu.max_page_level; l++)
1218 		total += (hat->hat_pages_mapped[l] << LEVEL_SHIFT(l));
1219 	total += hat->hat_ism_pgcnt;
1220 
1221 	return (total);
1222 }
1223 
1224 /*
1225  * enable/disable collection of stats for hat.
1226  */
1227 int
1228 hat_stats_enable(hat_t *hat)
1229 {
1230 	atomic_add_32(&hat->hat_stats, 1);
1231 	return (1);
1232 }
1233 
1234 void
1235 hat_stats_disable(hat_t *hat)
1236 {
1237 	atomic_add_32(&hat->hat_stats, -1);
1238 }
1239 
1240 /*
1241  * Utility to sync the ref/mod bits from a page table entry to the page_t
1242  * We must be holding the mapping list lock when this is called.
1243  */
1244 static void
1245 hati_sync_pte_to_page(page_t *pp, x86pte_t pte, level_t level)
1246 {
1247 	uint_t	rm = 0;
1248 	pgcnt_t	pgcnt;
1249 
1250 	if (PTE_GET(pte, PT_SOFTWARE) >= PT_NOSYNC)
1251 		return;
1252 
1253 	if (PTE_GET(pte, PT_REF))
1254 		rm |= P_REF;
1255 
1256 	if (PTE_GET(pte, PT_MOD))
1257 		rm |= P_MOD;
1258 
1259 	if (rm == 0)
1260 		return;
1261 
1262 	/*
1263 	 * sync to all constituent pages of a large page
1264 	 */
1265 	ASSERT(x86_hm_held(pp));
1266 	pgcnt = page_get_pagecnt(level);
1267 	ASSERT(IS_P2ALIGNED(pp->p_pagenum, pgcnt));
1268 	for (; pgcnt > 0; --pgcnt) {
1269 		/*
1270 		 * hat_page_demote() can't decrease
1271 		 * pszc below this mapping size
1272 		 * since this large mapping existed after we
1273 		 * took mlist lock.
1274 		 */
1275 		ASSERT(pp->p_szc >= level);
1276 		hat_page_setattr(pp, rm);
1277 		++pp;
1278 	}
1279 }
1280 
1281 /*
1282  * This the set of PTE bits for PFN, permissions and caching
1283  * that are allowed to change on a HAT_LOAD_REMAP
1284  */
1285 #define	PT_REMAP_BITS							\
1286 	(PT_PADDR | PT_NX | PT_WRITABLE | PT_WRITETHRU |		\
1287 	PT_NOCACHE | PT_PAT_4K | PT_PAT_LARGE | PT_IGNORE | PT_REF | PT_MOD)
1288 
1289 #define	REMAPASSERT(EX)	if (!(EX)) panic("hati_pte_map: " #EX)
1290 /*
1291  * Do the low-level work to get a mapping entered into a HAT's pagetables
1292  * and in the mapping list of the associated page_t.
1293  */
1294 static int
1295 hati_pte_map(
1296 	htable_t	*ht,
1297 	uint_t		entry,
1298 	page_t		*pp,
1299 	x86pte_t	pte,
1300 	int		flags,
1301 	void		*pte_ptr)
1302 {
1303 	hat_t		*hat = ht->ht_hat;
1304 	x86pte_t	old_pte;
1305 	level_t		l = ht->ht_level;
1306 	hment_t		*hm;
1307 	uint_t		is_consist;
1308 	uint_t		is_locked;
1309 	int		rv = 0;
1310 
1311 	/*
1312 	 * Is this a consistent (ie. need mapping list lock) mapping?
1313 	 */
1314 	is_consist = (pp != NULL && (flags & HAT_LOAD_NOCONSIST) == 0);
1315 
1316 	/*
1317 	 * Track locked mapping count in the htable.  Do this first,
1318 	 * as we track locking even if there already is a mapping present.
1319 	 */
1320 	is_locked = (flags & HAT_LOAD_LOCK) != 0 && hat != kas.a_hat;
1321 	if (is_locked)
1322 		HTABLE_LOCK_INC(ht);
1323 
1324 	/*
1325 	 * Acquire the page's mapping list lock and get an hment to use.
1326 	 * Note that hment_prepare() might return NULL.
1327 	 */
1328 	if (is_consist) {
1329 		x86_hm_enter(pp);
1330 		hm = hment_prepare(ht, entry, pp);
1331 	}
1332 
1333 	/*
1334 	 * Set the new pte, retrieving the old one at the same time.
1335 	 */
1336 	old_pte = x86pte_set(ht, entry, pte, pte_ptr);
1337 
1338 	/*
1339 	 * Did we get a large page / page table collision?
1340 	 */
1341 	if (old_pte == LPAGE_ERROR) {
1342 		if (is_locked)
1343 			HTABLE_LOCK_DEC(ht);
1344 		rv = -1;
1345 		goto done;
1346 	}
1347 
1348 	/*
1349 	 * If the mapping didn't change there is nothing more to do.
1350 	 */
1351 	if (PTE_EQUIV(pte, old_pte))
1352 		goto done;
1353 
1354 	/*
1355 	 * Install a new mapping in the page's mapping list
1356 	 */
1357 	if (!PTE_ISVALID(old_pte)) {
1358 		if (is_consist) {
1359 			hment_assign(ht, entry, pp, hm);
1360 			x86_hm_exit(pp);
1361 		} else {
1362 			ASSERT(flags & HAT_LOAD_NOCONSIST);
1363 		}
1364 #if defined(__amd64)
1365 		if (ht->ht_flags & HTABLE_VLP) {
1366 			cpu_t *cpu = CPU;
1367 			x86pte_t *vlpptep = cpu->cpu_hat_info->hci_vlp_l2ptes;
1368 			VLP_COPY(hat->hat_vlp_ptes, vlpptep);
1369 		}
1370 #endif
1371 		HTABLE_INC(ht->ht_valid_cnt);
1372 		PGCNT_INC(hat, l);
1373 		return (rv);
1374 	}
1375 
1376 	/*
1377 	 * Remap's are more complicated:
1378 	 *  - HAT_LOAD_REMAP must be specified if changing the pfn.
1379 	 *    We also require that NOCONSIST be specified.
1380 	 *  - Otherwise only permission or caching bits may change.
1381 	 */
1382 	if (!PTE_ISPAGE(old_pte, l))
1383 		panic("non-null/page mapping pte=" FMT_PTE, old_pte);
1384 
1385 	if (PTE2PFN(old_pte, l) != PTE2PFN(pte, l)) {
1386 		REMAPASSERT(flags & HAT_LOAD_REMAP);
1387 		REMAPASSERT(flags & HAT_LOAD_NOCONSIST);
1388 		REMAPASSERT(PTE_GET(old_pte, PT_SOFTWARE) >= PT_NOCONSIST);
1389 		REMAPASSERT(pf_is_memory(PTE2PFN(old_pte, l)) ==
1390 		    pf_is_memory(PTE2PFN(pte, l)));
1391 		REMAPASSERT(!is_consist);
1392 	}
1393 
1394 	/*
1395 	 * We only let remaps change the certain bits in the PTE.
1396 	 */
1397 	if (PTE_GET(old_pte, ~PT_REMAP_BITS) != PTE_GET(pte, ~PT_REMAP_BITS))
1398 		panic("remap bits changed: old_pte="FMT_PTE", pte="FMT_PTE"\n",
1399 		    old_pte, pte);
1400 
1401 	/*
1402 	 * We don't create any mapping list entries on a remap, so release
1403 	 * any allocated hment after we drop the mapping list lock.
1404 	 */
1405 done:
1406 	if (is_consist) {
1407 		x86_hm_exit(pp);
1408 		if (hm != NULL)
1409 			hment_free(hm);
1410 	}
1411 	return (rv);
1412 }
1413 
1414 /*
1415  * Internal routine to load a single page table entry. This only fails if
1416  * we attempt to overwrite a page table link with a large page.
1417  */
1418 static int
1419 hati_load_common(
1420 	hat_t		*hat,
1421 	uintptr_t	va,
1422 	page_t		*pp,
1423 	uint_t		attr,
1424 	uint_t		flags,
1425 	level_t		level,
1426 	pfn_t		pfn)
1427 {
1428 	htable_t	*ht;
1429 	uint_t		entry;
1430 	x86pte_t	pte;
1431 	int		rv = 0;
1432 
1433 	/*
1434 	 * The number 16 is arbitrary and here to catch a recursion problem
1435 	 * early before we blow out the kernel stack.
1436 	 */
1437 	++curthread->t_hatdepth;
1438 	ASSERT(curthread->t_hatdepth < 16);
1439 
1440 	ASSERT(hat == kas.a_hat ||
1441 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1442 
1443 	if (flags & HAT_LOAD_SHARE)
1444 		hat->hat_flags |= HAT_SHARED;
1445 
1446 	/*
1447 	 * Find the page table that maps this page if it already exists.
1448 	 */
1449 	ht = htable_lookup(hat, va, level);
1450 
1451 	/*
1452 	 * We must have HAT_LOAD_NOCONSIST if page_t is NULL.
1453 	 */
1454 	if (pp == NULL)
1455 		flags |= HAT_LOAD_NOCONSIST;
1456 
1457 	if (ht == NULL) {
1458 		ht = htable_create(hat, va, level, NULL);
1459 		ASSERT(ht != NULL);
1460 	}
1461 	entry = htable_va2entry(va, ht);
1462 
1463 	/*
1464 	 * a bunch of paranoid error checking
1465 	 */
1466 	ASSERT(ht->ht_busy > 0);
1467 	if (ht->ht_vaddr > va || va > HTABLE_LAST_PAGE(ht))
1468 		panic("hati_load_common: bad htable %p, va %p",
1469 		    (void *)ht, (void *)va);
1470 	ASSERT(ht->ht_level == level);
1471 
1472 	/*
1473 	 * construct the new PTE
1474 	 */
1475 	if (hat == kas.a_hat)
1476 		attr &= ~PROT_USER;
1477 	pte = hati_mkpte(pfn, attr, level, flags);
1478 	if (hat == kas.a_hat && va >= kernelbase)
1479 		PTE_SET(pte, mmu.pt_global);
1480 
1481 	/*
1482 	 * establish the mapping
1483 	 */
1484 	rv = hati_pte_map(ht, entry, pp, pte, flags, NULL);
1485 
1486 	/*
1487 	 * release the htable and any reserves
1488 	 */
1489 	htable_release(ht);
1490 	--curthread->t_hatdepth;
1491 	return (rv);
1492 }
1493 
1494 /*
1495  * special case of hat_memload to deal with some kernel addrs for performance
1496  */
1497 static void
1498 hat_kmap_load(
1499 	caddr_t		addr,
1500 	page_t		*pp,
1501 	uint_t		attr,
1502 	uint_t		flags)
1503 {
1504 	uintptr_t	va = (uintptr_t)addr;
1505 	x86pte_t	pte;
1506 	pfn_t		pfn = page_pptonum(pp);
1507 	pgcnt_t		pg_off = mmu_btop(va - mmu.kmap_addr);
1508 	htable_t	*ht;
1509 	uint_t		entry;
1510 	void		*pte_ptr;
1511 
1512 	/*
1513 	 * construct the requested PTE
1514 	 */
1515 	attr &= ~PROT_USER;
1516 	attr |= HAT_STORECACHING_OK;
1517 	pte = hati_mkpte(pfn, attr, 0, flags);
1518 	PTE_SET(pte, mmu.pt_global);
1519 
1520 	/*
1521 	 * Figure out the pte_ptr and htable and use common code to finish up
1522 	 */
1523 	if (mmu.pae_hat)
1524 		pte_ptr = mmu.kmap_ptes + pg_off;
1525 	else
1526 		pte_ptr = (x86pte32_t *)mmu.kmap_ptes + pg_off;
1527 	ht = mmu.kmap_htables[(va - mmu.kmap_htables[0]->ht_vaddr) >>
1528 	    LEVEL_SHIFT(1)];
1529 	entry = htable_va2entry(va, ht);
1530 	++curthread->t_hatdepth;
1531 	ASSERT(curthread->t_hatdepth < 16);
1532 	(void) hati_pte_map(ht, entry, pp, pte, flags, pte_ptr);
1533 	--curthread->t_hatdepth;
1534 }
1535 
1536 /*
1537  * hat_memload() - load a translation to the given page struct
1538  *
1539  * Flags for hat_memload/hat_devload/hat_*attr.
1540  *
1541  * 	HAT_LOAD	Default flags to load a translation to the page.
1542  *
1543  * 	HAT_LOAD_LOCK	Lock down mapping resources; hat_map(), hat_memload(),
1544  *			and hat_devload().
1545  *
1546  *	HAT_LOAD_NOCONSIST Do not add mapping to page_t mapping list.
1547  *			sets PT_NOCONSIST
1548  *
1549  *	HAT_LOAD_SHARE	A flag to hat_memload() to indicate h/w page tables
1550  *			that map some user pages (not kas) is shared by more
1551  *			than one process (eg. ISM).
1552  *
1553  *	HAT_LOAD_REMAP	Reload a valid pte with a different page frame.
1554  *
1555  *	HAT_NO_KALLOC	Do not kmem_alloc while creating the mapping; at this
1556  *			point, it's setting up mapping to allocate internal
1557  *			hat layer data structures.  This flag forces hat layer
1558  *			to tap its reserves in order to prevent infinite
1559  *			recursion.
1560  *
1561  * The following is a protection attribute (like PROT_READ, etc.)
1562  *
1563  *	HAT_NOSYNC	set PT_NOSYNC - this mapping's ref/mod bits
1564  *			are never cleared.
1565  *
1566  * Installing new valid PTE's and creation of the mapping list
1567  * entry are controlled under the same lock. It's derived from the
1568  * page_t being mapped.
1569  */
1570 static uint_t supported_memload_flags =
1571 	HAT_LOAD | HAT_LOAD_LOCK | HAT_LOAD_ADV | HAT_LOAD_NOCONSIST |
1572 	HAT_LOAD_SHARE | HAT_NO_KALLOC | HAT_LOAD_REMAP | HAT_LOAD_TEXT;
1573 
1574 void
1575 hat_memload(
1576 	hat_t		*hat,
1577 	caddr_t		addr,
1578 	page_t		*pp,
1579 	uint_t		attr,
1580 	uint_t		flags)
1581 {
1582 	uintptr_t	va = (uintptr_t)addr;
1583 	level_t		level = 0;
1584 	pfn_t		pfn = page_pptonum(pp);
1585 
1586 	XPV_DISALLOW_MIGRATE();
1587 	ASSERT(IS_PAGEALIGNED(va));
1588 	ASSERT(hat == kas.a_hat || va < _userlimit);
1589 	ASSERT(hat == kas.a_hat ||
1590 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1591 	ASSERT((flags & supported_memload_flags) == flags);
1592 
1593 	ASSERT(!IN_VA_HOLE(va));
1594 	ASSERT(!PP_ISFREE(pp));
1595 
1596 	/*
1597 	 * kernel address special case for performance.
1598 	 */
1599 	if (mmu.kmap_addr <= va && va < mmu.kmap_eaddr) {
1600 		ASSERT(hat == kas.a_hat);
1601 		hat_kmap_load(addr, pp, attr, flags);
1602 		XPV_ALLOW_MIGRATE();
1603 		return;
1604 	}
1605 
1606 	/*
1607 	 * This is used for memory with normal caching enabled, so
1608 	 * always set HAT_STORECACHING_OK.
1609 	 */
1610 	attr |= HAT_STORECACHING_OK;
1611 	if (hati_load_common(hat, va, pp, attr, flags, level, pfn) != 0)
1612 		panic("unexpected hati_load_common() failure");
1613 	XPV_ALLOW_MIGRATE();
1614 }
1615 
1616 /* ARGSUSED */
1617 void
1618 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp,
1619     uint_t attr, uint_t flags, hat_region_cookie_t rcookie)
1620 {
1621 	hat_memload(hat, addr, pp, attr, flags);
1622 }
1623 
1624 /*
1625  * Load the given array of page structs using large pages when possible
1626  */
1627 void
1628 hat_memload_array(
1629 	hat_t		*hat,
1630 	caddr_t		addr,
1631 	size_t		len,
1632 	page_t		**pages,
1633 	uint_t		attr,
1634 	uint_t		flags)
1635 {
1636 	uintptr_t	va = (uintptr_t)addr;
1637 	uintptr_t	eaddr = va + len;
1638 	level_t		level;
1639 	size_t		pgsize;
1640 	pgcnt_t		pgindx = 0;
1641 	pfn_t		pfn;
1642 	pgcnt_t		i;
1643 
1644 	XPV_DISALLOW_MIGRATE();
1645 	ASSERT(IS_PAGEALIGNED(va));
1646 	ASSERT(hat == kas.a_hat || va + len <= _userlimit);
1647 	ASSERT(hat == kas.a_hat ||
1648 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1649 	ASSERT((flags & supported_memload_flags) == flags);
1650 
1651 	/*
1652 	 * memload is used for memory with full caching enabled, so
1653 	 * set HAT_STORECACHING_OK.
1654 	 */
1655 	attr |= HAT_STORECACHING_OK;
1656 
1657 	/*
1658 	 * handle all pages using largest possible pagesize
1659 	 */
1660 	while (va < eaddr) {
1661 		/*
1662 		 * decide what level mapping to use (ie. pagesize)
1663 		 */
1664 		pfn = page_pptonum(pages[pgindx]);
1665 		for (level = mmu.max_page_level; ; --level) {
1666 			pgsize = LEVEL_SIZE(level);
1667 			if (level == 0)
1668 				break;
1669 
1670 			if (!IS_P2ALIGNED(va, pgsize) ||
1671 			    (eaddr - va) < pgsize ||
1672 			    !IS_P2ALIGNED(pfn_to_pa(pfn), pgsize))
1673 				continue;
1674 
1675 			/*
1676 			 * To use a large mapping of this size, all the
1677 			 * pages we are passed must be sequential subpages
1678 			 * of the large page.
1679 			 * hat_page_demote() can't change p_szc because
1680 			 * all pages are locked.
1681 			 */
1682 			if (pages[pgindx]->p_szc >= level) {
1683 				for (i = 0; i < mmu_btop(pgsize); ++i) {
1684 					if (pfn + i !=
1685 					    page_pptonum(pages[pgindx + i]))
1686 						break;
1687 					ASSERT(pages[pgindx + i]->p_szc >=
1688 					    level);
1689 					ASSERT(pages[pgindx] + i ==
1690 					    pages[pgindx + i]);
1691 				}
1692 				if (i == mmu_btop(pgsize)) {
1693 #ifdef DEBUG
1694 					if (level == 2)
1695 						map1gcnt++;
1696 #endif
1697 					break;
1698 				}
1699 			}
1700 		}
1701 
1702 		/*
1703 		 * Load this page mapping. If the load fails, try a smaller
1704 		 * pagesize.
1705 		 */
1706 		ASSERT(!IN_VA_HOLE(va));
1707 		while (hati_load_common(hat, va, pages[pgindx], attr,
1708 		    flags, level, pfn) != 0) {
1709 			if (level == 0)
1710 				panic("unexpected hati_load_common() failure");
1711 			--level;
1712 			pgsize = LEVEL_SIZE(level);
1713 		}
1714 
1715 		/*
1716 		 * move to next page
1717 		 */
1718 		va += pgsize;
1719 		pgindx += mmu_btop(pgsize);
1720 	}
1721 	XPV_ALLOW_MIGRATE();
1722 }
1723 
1724 /* ARGSUSED */
1725 void
1726 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len,
1727     struct page **pps, uint_t attr, uint_t flags,
1728     hat_region_cookie_t rcookie)
1729 {
1730 	hat_memload_array(hat, addr, len, pps, attr, flags);
1731 }
1732 
1733 /*
1734  * void hat_devload(hat, addr, len, pf, attr, flags)
1735  *	load/lock the given page frame number
1736  *
1737  * Advisory ordering attributes. Apply only to device mappings.
1738  *
1739  * HAT_STRICTORDER: the CPU must issue the references in order, as the
1740  *	programmer specified.  This is the default.
1741  * HAT_UNORDERED_OK: the CPU may reorder the references (this is all kinds
1742  *	of reordering; store or load with store or load).
1743  * HAT_MERGING_OK: merging and batching: the CPU may merge individual stores
1744  *	to consecutive locations (for example, turn two consecutive byte
1745  *	stores into one halfword store), and it may batch individual loads
1746  *	(for example, turn two consecutive byte loads into one halfword load).
1747  *	This also implies re-ordering.
1748  * HAT_LOADCACHING_OK: the CPU may cache the data it fetches and reuse it
1749  *	until another store occurs.  The default is to fetch new data
1750  *	on every load.  This also implies merging.
1751  * HAT_STORECACHING_OK: the CPU may keep the data in the cache and push it to
1752  *	the device (perhaps with other data) at a later time.  The default is
1753  *	to push the data right away.  This also implies load caching.
1754  *
1755  * Equivalent of hat_memload(), but can be used for device memory where
1756  * there are no page_t's and we support additional flags (write merging, etc).
1757  * Note that we can have large page mappings with this interface.
1758  */
1759 int supported_devload_flags = HAT_LOAD | HAT_LOAD_LOCK |
1760 	HAT_LOAD_NOCONSIST | HAT_STRICTORDER | HAT_UNORDERED_OK |
1761 	HAT_MERGING_OK | HAT_LOADCACHING_OK | HAT_STORECACHING_OK;
1762 
1763 void
1764 hat_devload(
1765 	hat_t		*hat,
1766 	caddr_t		addr,
1767 	size_t		len,
1768 	pfn_t		pfn,
1769 	uint_t		attr,
1770 	int		flags)
1771 {
1772 	uintptr_t	va = ALIGN2PAGE(addr);
1773 	uintptr_t	eva = va + len;
1774 	level_t		level;
1775 	size_t		pgsize;
1776 	page_t		*pp;
1777 	int		f;	/* per PTE copy of flags  - maybe modified */
1778 	uint_t		a;	/* per PTE copy of attr */
1779 
1780 	XPV_DISALLOW_MIGRATE();
1781 	ASSERT(IS_PAGEALIGNED(va));
1782 	ASSERT(hat == kas.a_hat || eva <= _userlimit);
1783 	ASSERT(hat == kas.a_hat ||
1784 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1785 	ASSERT((flags & supported_devload_flags) == flags);
1786 
1787 	/*
1788 	 * handle all pages
1789 	 */
1790 	while (va < eva) {
1791 
1792 		/*
1793 		 * decide what level mapping to use (ie. pagesize)
1794 		 */
1795 		for (level = mmu.max_page_level; ; --level) {
1796 			pgsize = LEVEL_SIZE(level);
1797 			if (level == 0)
1798 				break;
1799 			if (IS_P2ALIGNED(va, pgsize) &&
1800 			    (eva - va) >= pgsize &&
1801 			    IS_P2ALIGNED(pfn, mmu_btop(pgsize))) {
1802 #ifdef DEBUG
1803 				if (level == 2)
1804 					map1gcnt++;
1805 #endif
1806 				break;
1807 			}
1808 		}
1809 
1810 		/*
1811 		 * If this is just memory then allow caching (this happens
1812 		 * for the nucleus pages) - though HAT_PLAT_NOCACHE can be used
1813 		 * to override that. If we don't have a page_t then make sure
1814 		 * NOCONSIST is set.
1815 		 */
1816 		a = attr;
1817 		f = flags;
1818 		if (!pf_is_memory(pfn))
1819 			f |= HAT_LOAD_NOCONSIST;
1820 		else if (!(a & HAT_PLAT_NOCACHE))
1821 			a |= HAT_STORECACHING_OK;
1822 
1823 		if (f & HAT_LOAD_NOCONSIST)
1824 			pp = NULL;
1825 		else
1826 			pp = page_numtopp_nolock(pfn);
1827 
1828 		/*
1829 		 * Check to make sure we are really trying to map a valid
1830 		 * memory page. The caller wishing to intentionally map
1831 		 * free memory pages will have passed the HAT_LOAD_NOCONSIST
1832 		 * flag, then pp will be NULL.
1833 		 */
1834 		if (pp != NULL) {
1835 			if (PP_ISFREE(pp)) {
1836 				panic("hat_devload: loading "
1837 				    "a mapping to free page %p", (void *)pp);
1838 			}
1839 
1840 			if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) {
1841 				panic("hat_devload: loading a mapping "
1842 				    "to an unlocked page %p",
1843 				    (void *)pp);
1844 			}
1845 		}
1846 
1847 		/*
1848 		 * load this page mapping
1849 		 */
1850 		ASSERT(!IN_VA_HOLE(va));
1851 		while (hati_load_common(hat, va, pp, a, f, level, pfn) != 0) {
1852 			if (level == 0)
1853 				panic("unexpected hati_load_common() failure");
1854 			--level;
1855 			pgsize = LEVEL_SIZE(level);
1856 		}
1857 
1858 		/*
1859 		 * move to next page
1860 		 */
1861 		va += pgsize;
1862 		pfn += mmu_btop(pgsize);
1863 	}
1864 	XPV_ALLOW_MIGRATE();
1865 }
1866 
1867 /*
1868  * void hat_unlock(hat, addr, len)
1869  *	unlock the mappings to a given range of addresses
1870  *
1871  * Locks are tracked by ht_lock_cnt in the htable.
1872  */
1873 void
1874 hat_unlock(hat_t *hat, caddr_t addr, size_t len)
1875 {
1876 	uintptr_t	vaddr = (uintptr_t)addr;
1877 	uintptr_t	eaddr = vaddr + len;
1878 	htable_t	*ht = NULL;
1879 
1880 	/*
1881 	 * kernel entries are always locked, we don't track lock counts
1882 	 */
1883 	ASSERT(hat == kas.a_hat || eaddr <= _userlimit);
1884 	ASSERT(IS_PAGEALIGNED(vaddr));
1885 	ASSERT(IS_PAGEALIGNED(eaddr));
1886 	if (hat == kas.a_hat)
1887 		return;
1888 	if (eaddr > _userlimit)
1889 		panic("hat_unlock() address out of range - above _userlimit");
1890 
1891 	XPV_DISALLOW_MIGRATE();
1892 	ASSERT(AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1893 	while (vaddr < eaddr) {
1894 		(void) htable_walk(hat, &ht, &vaddr, eaddr);
1895 		if (ht == NULL)
1896 			break;
1897 
1898 		ASSERT(!IN_VA_HOLE(vaddr));
1899 
1900 		if (ht->ht_lock_cnt < 1)
1901 			panic("hat_unlock(): lock_cnt < 1, "
1902 			    "htable=%p, vaddr=%p\n", (void *)ht, (void *)vaddr);
1903 		HTABLE_LOCK_DEC(ht);
1904 
1905 		vaddr += LEVEL_SIZE(ht->ht_level);
1906 	}
1907 	if (ht)
1908 		htable_release(ht);
1909 	XPV_ALLOW_MIGRATE();
1910 }
1911 
1912 /* ARGSUSED */
1913 void
1914 hat_unlock_region(struct hat *hat, caddr_t addr, size_t len,
1915     hat_region_cookie_t rcookie)
1916 {
1917 	panic("No shared region support on x86");
1918 }
1919 
1920 #if !defined(__xpv)
1921 /*
1922  * Cross call service routine to demap a virtual page on
1923  * the current CPU or flush all mappings in TLB.
1924  */
1925 /*ARGSUSED*/
1926 static int
1927 hati_demap_func(xc_arg_t a1, xc_arg_t a2, xc_arg_t a3)
1928 {
1929 	hat_t	*hat = (hat_t *)a1;
1930 	caddr_t	addr = (caddr_t)a2;
1931 
1932 	/*
1933 	 * If the target hat isn't the kernel and this CPU isn't operating
1934 	 * in the target hat, we can ignore the cross call.
1935 	 */
1936 	if (hat != kas.a_hat && hat != CPU->cpu_current_hat)
1937 		return (0);
1938 
1939 	/*
1940 	 * For a normal address, we just flush one page mapping
1941 	 */
1942 	if ((uintptr_t)addr != DEMAP_ALL_ADDR) {
1943 		mmu_tlbflush_entry(addr);
1944 		return (0);
1945 	}
1946 
1947 	/*
1948 	 * Otherwise we reload cr3 to effect a complete TLB flush.
1949 	 *
1950 	 * A reload of cr3 on a VLP process also means we must also recopy in
1951 	 * the pte values from the struct hat
1952 	 */
1953 	if (hat->hat_flags & HAT_VLP) {
1954 #if defined(__amd64)
1955 		x86pte_t *vlpptep = CPU->cpu_hat_info->hci_vlp_l2ptes;
1956 
1957 		VLP_COPY(hat->hat_vlp_ptes, vlpptep);
1958 #elif defined(__i386)
1959 		reload_pae32(hat, CPU);
1960 #endif
1961 	}
1962 	reload_cr3();
1963 	return (0);
1964 }
1965 
1966 /*
1967  * Flush all TLB entries, including global (ie. kernel) ones.
1968  */
1969 static void
1970 flush_all_tlb_entries(void)
1971 {
1972 	ulong_t cr4 = getcr4();
1973 
1974 	if (cr4 & CR4_PGE) {
1975 		setcr4(cr4 & ~(ulong_t)CR4_PGE);
1976 		setcr4(cr4);
1977 
1978 		/*
1979 		 * 32 bit PAE also needs to always reload_cr3()
1980 		 */
1981 		if (mmu.max_level == 2)
1982 			reload_cr3();
1983 	} else {
1984 		reload_cr3();
1985 	}
1986 }
1987 
1988 #define	TLB_CPU_HALTED	(01ul)
1989 #define	TLB_INVAL_ALL	(02ul)
1990 #define	CAS_TLB_INFO(cpu, old, new)	\
1991 	atomic_cas_ulong((ulong_t *)&(cpu)->cpu_m.mcpu_tlb_info, (old), (new))
1992 
1993 /*
1994  * Record that a CPU is going idle
1995  */
1996 void
1997 tlb_going_idle(void)
1998 {
1999 	atomic_or_ulong((ulong_t *)&CPU->cpu_m.mcpu_tlb_info, TLB_CPU_HALTED);
2000 }
2001 
2002 /*
2003  * Service a delayed TLB flush if coming out of being idle.
2004  * It will be called from cpu idle notification with interrupt disabled.
2005  */
2006 void
2007 tlb_service(void)
2008 {
2009 	ulong_t tlb_info;
2010 	ulong_t found;
2011 
2012 	/*
2013 	 * We only have to do something if coming out of being idle.
2014 	 */
2015 	tlb_info = CPU->cpu_m.mcpu_tlb_info;
2016 	if (tlb_info & TLB_CPU_HALTED) {
2017 		ASSERT(CPU->cpu_current_hat == kas.a_hat);
2018 
2019 		/*
2020 		 * Atomic clear and fetch of old state.
2021 		 */
2022 		while ((found = CAS_TLB_INFO(CPU, tlb_info, 0)) != tlb_info) {
2023 			ASSERT(found & TLB_CPU_HALTED);
2024 			tlb_info = found;
2025 			SMT_PAUSE();
2026 		}
2027 		if (tlb_info & TLB_INVAL_ALL)
2028 			flush_all_tlb_entries();
2029 	}
2030 }
2031 #endif /* !__xpv */
2032 
2033 /*
2034  * Internal routine to do cross calls to invalidate a range of pages on
2035  * all CPUs using a given hat.
2036  */
2037 void
2038 hat_tlb_inval(hat_t *hat, uintptr_t va)
2039 {
2040 	extern int	flushes_require_xcalls;	/* from mp_startup.c */
2041 	cpuset_t	justme;
2042 	cpuset_t	cpus_to_shootdown;
2043 #ifndef __xpv
2044 	cpuset_t	check_cpus;
2045 	cpu_t		*cpup;
2046 	int		c;
2047 #endif
2048 
2049 	/*
2050 	 * If the hat is being destroyed, there are no more users, so
2051 	 * demap need not do anything.
2052 	 */
2053 	if (hat->hat_flags & HAT_FREEING)
2054 		return;
2055 
2056 	/*
2057 	 * If demapping from a shared pagetable, we best demap the
2058 	 * entire set of user TLBs, since we don't know what addresses
2059 	 * these were shared at.
2060 	 */
2061 	if (hat->hat_flags & HAT_SHARED) {
2062 		hat = kas.a_hat;
2063 		va = DEMAP_ALL_ADDR;
2064 	}
2065 
2066 	/*
2067 	 * if not running with multiple CPUs, don't use cross calls
2068 	 */
2069 	if (panicstr || !flushes_require_xcalls) {
2070 #ifdef __xpv
2071 		if (va == DEMAP_ALL_ADDR)
2072 			xen_flush_tlb();
2073 		else
2074 			xen_flush_va((caddr_t)va);
2075 #else
2076 		(void) hati_demap_func((xc_arg_t)hat, (xc_arg_t)va, NULL);
2077 #endif
2078 		return;
2079 	}
2080 
2081 
2082 	/*
2083 	 * Determine CPUs to shootdown. Kernel changes always do all CPUs.
2084 	 * Otherwise it's just CPUs currently executing in this hat.
2085 	 */
2086 	kpreempt_disable();
2087 	CPUSET_ONLY(justme, CPU->cpu_id);
2088 	if (hat == kas.a_hat)
2089 		cpus_to_shootdown = khat_cpuset;
2090 	else
2091 		cpus_to_shootdown = hat->hat_cpus;
2092 
2093 #ifndef __xpv
2094 	/*
2095 	 * If any CPUs in the set are idle, just request a delayed flush
2096 	 * and avoid waking them up.
2097 	 */
2098 	check_cpus = cpus_to_shootdown;
2099 	for (c = 0; c < NCPU && !CPUSET_ISNULL(check_cpus); ++c) {
2100 		ulong_t tlb_info;
2101 
2102 		if (!CPU_IN_SET(check_cpus, c))
2103 			continue;
2104 		CPUSET_DEL(check_cpus, c);
2105 		cpup = cpu[c];
2106 		if (cpup == NULL)
2107 			continue;
2108 
2109 		tlb_info = cpup->cpu_m.mcpu_tlb_info;
2110 		while (tlb_info == TLB_CPU_HALTED) {
2111 			(void) CAS_TLB_INFO(cpup, TLB_CPU_HALTED,
2112 			    TLB_CPU_HALTED | TLB_INVAL_ALL);
2113 			SMT_PAUSE();
2114 			tlb_info = cpup->cpu_m.mcpu_tlb_info;
2115 		}
2116 		if (tlb_info == (TLB_CPU_HALTED | TLB_INVAL_ALL)) {
2117 			HATSTAT_INC(hs_tlb_inval_delayed);
2118 			CPUSET_DEL(cpus_to_shootdown, c);
2119 		}
2120 	}
2121 #endif
2122 
2123 	if (CPUSET_ISNULL(cpus_to_shootdown) ||
2124 	    CPUSET_ISEQUAL(cpus_to_shootdown, justme)) {
2125 
2126 #ifdef __xpv
2127 		if (va == DEMAP_ALL_ADDR)
2128 			xen_flush_tlb();
2129 		else
2130 			xen_flush_va((caddr_t)va);
2131 #else
2132 		(void) hati_demap_func((xc_arg_t)hat, (xc_arg_t)va, NULL);
2133 #endif
2134 
2135 	} else {
2136 
2137 		CPUSET_ADD(cpus_to_shootdown, CPU->cpu_id);
2138 #ifdef __xpv
2139 		if (va == DEMAP_ALL_ADDR)
2140 			xen_gflush_tlb(cpus_to_shootdown);
2141 		else
2142 			xen_gflush_va((caddr_t)va, cpus_to_shootdown);
2143 #else
2144 		xc_call((xc_arg_t)hat, (xc_arg_t)va, NULL,
2145 		    CPUSET2BV(cpus_to_shootdown), hati_demap_func);
2146 #endif
2147 
2148 	}
2149 	kpreempt_enable();
2150 }
2151 
2152 /*
2153  * Interior routine for HAT_UNLOADs from hat_unload_callback(),
2154  * hat_kmap_unload() OR from hat_steal() code.  This routine doesn't
2155  * handle releasing of the htables.
2156  */
2157 void
2158 hat_pte_unmap(
2159 	htable_t	*ht,
2160 	uint_t		entry,
2161 	uint_t		flags,
2162 	x86pte_t	old_pte,
2163 	void		*pte_ptr)
2164 {
2165 	hat_t		*hat = ht->ht_hat;
2166 	hment_t		*hm = NULL;
2167 	page_t		*pp = NULL;
2168 	level_t		l = ht->ht_level;
2169 	pfn_t		pfn;
2170 
2171 	/*
2172 	 * We always track the locking counts, even if nothing is unmapped
2173 	 */
2174 	if ((flags & HAT_UNLOAD_UNLOCK) != 0 && hat != kas.a_hat) {
2175 		ASSERT(ht->ht_lock_cnt > 0);
2176 		HTABLE_LOCK_DEC(ht);
2177 	}
2178 
2179 	/*
2180 	 * Figure out which page's mapping list lock to acquire using the PFN
2181 	 * passed in "old" PTE. We then attempt to invalidate the PTE.
2182 	 * If another thread, probably a hat_pageunload, has asynchronously
2183 	 * unmapped/remapped this address we'll loop here.
2184 	 */
2185 	ASSERT(ht->ht_busy > 0);
2186 	while (PTE_ISVALID(old_pte)) {
2187 		pfn = PTE2PFN(old_pte, l);
2188 		if (PTE_GET(old_pte, PT_SOFTWARE) >= PT_NOCONSIST) {
2189 			pp = NULL;
2190 		} else {
2191 #ifdef __xpv
2192 			if (pfn == PFN_INVALID)
2193 				panic("Invalid PFN, but not PT_NOCONSIST");
2194 #endif
2195 			pp = page_numtopp_nolock(pfn);
2196 			if (pp == NULL) {
2197 				panic("no page_t, not NOCONSIST: old_pte="
2198 				    FMT_PTE " ht=%lx entry=0x%x pte_ptr=%lx",
2199 				    old_pte, (uintptr_t)ht, entry,
2200 				    (uintptr_t)pte_ptr);
2201 			}
2202 			x86_hm_enter(pp);
2203 		}
2204 
2205 		old_pte = x86pte_inval(ht, entry, old_pte, pte_ptr);
2206 
2207 		/*
2208 		 * If the page hadn't changed we've unmapped it and can proceed
2209 		 */
2210 		if (PTE_ISVALID(old_pte) && PTE2PFN(old_pte, l) == pfn)
2211 			break;
2212 
2213 		/*
2214 		 * Otherwise, we'll have to retry with the current old_pte.
2215 		 * Drop the hment lock, since the pfn may have changed.
2216 		 */
2217 		if (pp != NULL) {
2218 			x86_hm_exit(pp);
2219 			pp = NULL;
2220 		} else {
2221 			ASSERT(PTE_GET(old_pte, PT_SOFTWARE) >= PT_NOCONSIST);
2222 		}
2223 	}
2224 
2225 	/*
2226 	 * If the old mapping wasn't valid, there's nothing more to do
2227 	 */
2228 	if (!PTE_ISVALID(old_pte)) {
2229 		if (pp != NULL)
2230 			x86_hm_exit(pp);
2231 		return;
2232 	}
2233 
2234 	/*
2235 	 * Take care of syncing any MOD/REF bits and removing the hment.
2236 	 */
2237 	if (pp != NULL) {
2238 		if (!(flags & HAT_UNLOAD_NOSYNC))
2239 			hati_sync_pte_to_page(pp, old_pte, l);
2240 		hm = hment_remove(pp, ht, entry);
2241 		x86_hm_exit(pp);
2242 		if (hm != NULL)
2243 			hment_free(hm);
2244 	}
2245 
2246 	/*
2247 	 * Handle book keeping in the htable and hat
2248 	 */
2249 	ASSERT(ht->ht_valid_cnt > 0);
2250 	HTABLE_DEC(ht->ht_valid_cnt);
2251 	PGCNT_DEC(hat, l);
2252 }
2253 
2254 /*
2255  * very cheap unload implementation to special case some kernel addresses
2256  */
2257 static void
2258 hat_kmap_unload(caddr_t addr, size_t len, uint_t flags)
2259 {
2260 	uintptr_t	va = (uintptr_t)addr;
2261 	uintptr_t	eva = va + len;
2262 	pgcnt_t		pg_index;
2263 	htable_t	*ht;
2264 	uint_t		entry;
2265 	x86pte_t	*pte_ptr;
2266 	x86pte_t	old_pte;
2267 
2268 	for (; va < eva; va += MMU_PAGESIZE) {
2269 		/*
2270 		 * Get the PTE
2271 		 */
2272 		pg_index = mmu_btop(va - mmu.kmap_addr);
2273 		pte_ptr = PT_INDEX_PTR(mmu.kmap_ptes, pg_index);
2274 		old_pte = GET_PTE(pte_ptr);
2275 
2276 		/*
2277 		 * get the htable / entry
2278 		 */
2279 		ht = mmu.kmap_htables[(va - mmu.kmap_htables[0]->ht_vaddr)
2280 		    >> LEVEL_SHIFT(1)];
2281 		entry = htable_va2entry(va, ht);
2282 
2283 		/*
2284 		 * use mostly common code to unmap it.
2285 		 */
2286 		hat_pte_unmap(ht, entry, flags, old_pte, pte_ptr);
2287 	}
2288 }
2289 
2290 
2291 /*
2292  * unload a range of virtual address space (no callback)
2293  */
2294 void
2295 hat_unload(hat_t *hat, caddr_t addr, size_t len, uint_t flags)
2296 {
2297 	uintptr_t va = (uintptr_t)addr;
2298 
2299 	XPV_DISALLOW_MIGRATE();
2300 	ASSERT(hat == kas.a_hat || va + len <= _userlimit);
2301 
2302 	/*
2303 	 * special case for performance.
2304 	 */
2305 	if (mmu.kmap_addr <= va && va < mmu.kmap_eaddr) {
2306 		ASSERT(hat == kas.a_hat);
2307 		hat_kmap_unload(addr, len, flags);
2308 	} else {
2309 		hat_unload_callback(hat, addr, len, flags, NULL);
2310 	}
2311 	XPV_ALLOW_MIGRATE();
2312 }
2313 
2314 /*
2315  * Do the callbacks for ranges being unloaded.
2316  */
2317 typedef struct range_info {
2318 	uintptr_t	rng_va;
2319 	ulong_t		rng_cnt;
2320 	level_t		rng_level;
2321 } range_info_t;
2322 
2323 static void
2324 handle_ranges(hat_callback_t *cb, uint_t cnt, range_info_t *range)
2325 {
2326 	/*
2327 	 * do callbacks to upper level VM system
2328 	 */
2329 	while (cb != NULL && cnt > 0) {
2330 		--cnt;
2331 		cb->hcb_start_addr = (caddr_t)range[cnt].rng_va;
2332 		cb->hcb_end_addr = cb->hcb_start_addr;
2333 		cb->hcb_end_addr +=
2334 		    range[cnt].rng_cnt << LEVEL_SIZE(range[cnt].rng_level);
2335 		cb->hcb_function(cb);
2336 	}
2337 }
2338 
2339 /*
2340  * Unload a given range of addresses (has optional callback)
2341  *
2342  * Flags:
2343  * define	HAT_UNLOAD		0x00
2344  * define	HAT_UNLOAD_NOSYNC	0x02
2345  * define	HAT_UNLOAD_UNLOCK	0x04
2346  * define	HAT_UNLOAD_OTHER	0x08 - not used
2347  * define	HAT_UNLOAD_UNMAP	0x10 - same as HAT_UNLOAD
2348  */
2349 #define	MAX_UNLOAD_CNT (8)
2350 void
2351 hat_unload_callback(
2352 	hat_t		*hat,
2353 	caddr_t		addr,
2354 	size_t		len,
2355 	uint_t		flags,
2356 	hat_callback_t	*cb)
2357 {
2358 	uintptr_t	vaddr = (uintptr_t)addr;
2359 	uintptr_t	eaddr = vaddr + len;
2360 	htable_t	*ht = NULL;
2361 	uint_t		entry;
2362 	uintptr_t	contig_va = (uintptr_t)-1L;
2363 	range_info_t	r[MAX_UNLOAD_CNT];
2364 	uint_t		r_cnt = 0;
2365 	x86pte_t	old_pte;
2366 
2367 	XPV_DISALLOW_MIGRATE();
2368 	ASSERT(hat == kas.a_hat || eaddr <= _userlimit);
2369 	ASSERT(IS_PAGEALIGNED(vaddr));
2370 	ASSERT(IS_PAGEALIGNED(eaddr));
2371 
2372 	/*
2373 	 * Special case a single page being unloaded for speed. This happens
2374 	 * quite frequently, COW faults after a fork() for example.
2375 	 */
2376 	if (cb == NULL && len == MMU_PAGESIZE) {
2377 		ht = htable_getpte(hat, vaddr, &entry, &old_pte, 0);
2378 		if (ht != NULL) {
2379 			if (PTE_ISVALID(old_pte))
2380 				hat_pte_unmap(ht, entry, flags, old_pte, NULL);
2381 			htable_release(ht);
2382 		}
2383 		XPV_ALLOW_MIGRATE();
2384 		return;
2385 	}
2386 
2387 	while (vaddr < eaddr) {
2388 		old_pte = htable_walk(hat, &ht, &vaddr, eaddr);
2389 		if (ht == NULL)
2390 			break;
2391 
2392 		ASSERT(!IN_VA_HOLE(vaddr));
2393 
2394 		if (vaddr < (uintptr_t)addr)
2395 			panic("hat_unload_callback(): unmap inside large page");
2396 
2397 		/*
2398 		 * We'll do the call backs for contiguous ranges
2399 		 */
2400 		if (vaddr != contig_va ||
2401 		    (r_cnt > 0 && r[r_cnt - 1].rng_level != ht->ht_level)) {
2402 			if (r_cnt == MAX_UNLOAD_CNT) {
2403 				handle_ranges(cb, r_cnt, r);
2404 				r_cnt = 0;
2405 			}
2406 			r[r_cnt].rng_va = vaddr;
2407 			r[r_cnt].rng_cnt = 0;
2408 			r[r_cnt].rng_level = ht->ht_level;
2409 			++r_cnt;
2410 		}
2411 
2412 		/*
2413 		 * Unload one mapping from the page tables.
2414 		 */
2415 		entry = htable_va2entry(vaddr, ht);
2416 		hat_pte_unmap(ht, entry, flags, old_pte, NULL);
2417 		ASSERT(ht->ht_level <= mmu.max_page_level);
2418 		vaddr += LEVEL_SIZE(ht->ht_level);
2419 		contig_va = vaddr;
2420 		++r[r_cnt - 1].rng_cnt;
2421 	}
2422 	if (ht)
2423 		htable_release(ht);
2424 
2425 	/*
2426 	 * handle last range for callbacks
2427 	 */
2428 	if (r_cnt > 0)
2429 		handle_ranges(cb, r_cnt, r);
2430 	XPV_ALLOW_MIGRATE();
2431 }
2432 
2433 /*
2434  * Invalidate a virtual address translation on a slave CPU during
2435  * panic() dumps.
2436  */
2437 void
2438 hat_flush_range(hat_t *hat, caddr_t va, size_t size)
2439 {
2440 	ssize_t sz;
2441 	caddr_t endva = va + size;
2442 
2443 	while (va < endva) {
2444 		sz = hat_getpagesize(hat, va);
2445 		if (sz < 0) {
2446 #ifdef __xpv
2447 			xen_flush_tlb();
2448 #else
2449 			flush_all_tlb_entries();
2450 #endif
2451 			break;
2452 		}
2453 #ifdef __xpv
2454 		xen_flush_va(va);
2455 #else
2456 		mmu_tlbflush_entry(va);
2457 #endif
2458 		va += sz;
2459 	}
2460 }
2461 
2462 /*
2463  * synchronize mapping with software data structures
2464  *
2465  * This interface is currently only used by the working set monitor
2466  * driver.
2467  */
2468 /*ARGSUSED*/
2469 void
2470 hat_sync(hat_t *hat, caddr_t addr, size_t len, uint_t flags)
2471 {
2472 	uintptr_t	vaddr = (uintptr_t)addr;
2473 	uintptr_t	eaddr = vaddr + len;
2474 	htable_t	*ht = NULL;
2475 	uint_t		entry;
2476 	x86pte_t	pte;
2477 	x86pte_t	save_pte;
2478 	x86pte_t	new;
2479 	page_t		*pp;
2480 
2481 	ASSERT(!IN_VA_HOLE(vaddr));
2482 	ASSERT(IS_PAGEALIGNED(vaddr));
2483 	ASSERT(IS_PAGEALIGNED(eaddr));
2484 	ASSERT(hat == kas.a_hat || eaddr <= _userlimit);
2485 
2486 	XPV_DISALLOW_MIGRATE();
2487 	for (; vaddr < eaddr; vaddr += LEVEL_SIZE(ht->ht_level)) {
2488 try_again:
2489 		pte = htable_walk(hat, &ht, &vaddr, eaddr);
2490 		if (ht == NULL)
2491 			break;
2492 		entry = htable_va2entry(vaddr, ht);
2493 
2494 		if (PTE_GET(pte, PT_SOFTWARE) >= PT_NOSYNC ||
2495 		    PTE_GET(pte, PT_REF | PT_MOD) == 0)
2496 			continue;
2497 
2498 		/*
2499 		 * We need to acquire the mapping list lock to protect
2500 		 * against hat_pageunload(), hat_unload(), etc.
2501 		 */
2502 		pp = page_numtopp_nolock(PTE2PFN(pte, ht->ht_level));
2503 		if (pp == NULL)
2504 			break;
2505 		x86_hm_enter(pp);
2506 		save_pte = pte;
2507 		pte = x86pte_get(ht, entry);
2508 		if (pte != save_pte) {
2509 			x86_hm_exit(pp);
2510 			goto try_again;
2511 		}
2512 		if (PTE_GET(pte, PT_SOFTWARE) >= PT_NOSYNC ||
2513 		    PTE_GET(pte, PT_REF | PT_MOD) == 0) {
2514 			x86_hm_exit(pp);
2515 			continue;
2516 		}
2517 
2518 		/*
2519 		 * Need to clear ref or mod bits. We may compete with
2520 		 * hardware updating the R/M bits and have to try again.
2521 		 */
2522 		if (flags == HAT_SYNC_ZERORM) {
2523 			new = pte;
2524 			PTE_CLR(new, PT_REF | PT_MOD);
2525 			pte = hati_update_pte(ht, entry, pte, new);
2526 			if (pte != 0) {
2527 				x86_hm_exit(pp);
2528 				goto try_again;
2529 			}
2530 		} else {
2531 			/*
2532 			 * sync the PTE to the page_t
2533 			 */
2534 			hati_sync_pte_to_page(pp, save_pte, ht->ht_level);
2535 		}
2536 		x86_hm_exit(pp);
2537 	}
2538 	if (ht)
2539 		htable_release(ht);
2540 	XPV_ALLOW_MIGRATE();
2541 }
2542 
2543 /*
2544  * void	hat_map(hat, addr, len, flags)
2545  */
2546 /*ARGSUSED*/
2547 void
2548 hat_map(hat_t *hat, caddr_t addr, size_t len, uint_t flags)
2549 {
2550 	/* does nothing */
2551 }
2552 
2553 /*
2554  * uint_t hat_getattr(hat, addr, *attr)
2555  *	returns attr for <hat,addr> in *attr.  returns 0 if there was a
2556  *	mapping and *attr is valid, nonzero if there was no mapping and
2557  *	*attr is not valid.
2558  */
2559 uint_t
2560 hat_getattr(hat_t *hat, caddr_t addr, uint_t *attr)
2561 {
2562 	uintptr_t	vaddr = ALIGN2PAGE(addr);
2563 	htable_t	*ht = NULL;
2564 	x86pte_t	pte;
2565 
2566 	ASSERT(hat == kas.a_hat || vaddr <= _userlimit);
2567 
2568 	if (IN_VA_HOLE(vaddr))
2569 		return ((uint_t)-1);
2570 
2571 	ht = htable_getpte(hat, vaddr, NULL, &pte, mmu.max_page_level);
2572 	if (ht == NULL)
2573 		return ((uint_t)-1);
2574 
2575 	if (!PTE_ISVALID(pte) || !PTE_ISPAGE(pte, ht->ht_level)) {
2576 		htable_release(ht);
2577 		return ((uint_t)-1);
2578 	}
2579 
2580 	*attr = PROT_READ;
2581 	if (PTE_GET(pte, PT_WRITABLE))
2582 		*attr |= PROT_WRITE;
2583 	if (PTE_GET(pte, PT_USER))
2584 		*attr |= PROT_USER;
2585 	if (!PTE_GET(pte, mmu.pt_nx))
2586 		*attr |= PROT_EXEC;
2587 	if (PTE_GET(pte, PT_SOFTWARE) >= PT_NOSYNC)
2588 		*attr |= HAT_NOSYNC;
2589 	htable_release(ht);
2590 	return (0);
2591 }
2592 
2593 /*
2594  * hat_updateattr() applies the given attribute change to an existing mapping
2595  */
2596 #define	HAT_LOAD_ATTR		1
2597 #define	HAT_SET_ATTR		2
2598 #define	HAT_CLR_ATTR		3
2599 
2600 static void
2601 hat_updateattr(hat_t *hat, caddr_t addr, size_t len, uint_t attr, int what)
2602 {
2603 	uintptr_t	vaddr = (uintptr_t)addr;
2604 	uintptr_t	eaddr = (uintptr_t)addr + len;
2605 	htable_t	*ht = NULL;
2606 	uint_t		entry;
2607 	x86pte_t	oldpte, newpte;
2608 	page_t		*pp;
2609 
2610 	XPV_DISALLOW_MIGRATE();
2611 	ASSERT(IS_PAGEALIGNED(vaddr));
2612 	ASSERT(IS_PAGEALIGNED(eaddr));
2613 	ASSERT(hat == kas.a_hat ||
2614 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
2615 	for (; vaddr < eaddr; vaddr += LEVEL_SIZE(ht->ht_level)) {
2616 try_again:
2617 		oldpte = htable_walk(hat, &ht, &vaddr, eaddr);
2618 		if (ht == NULL)
2619 			break;
2620 		if (PTE_GET(oldpte, PT_SOFTWARE) >= PT_NOCONSIST)
2621 			continue;
2622 
2623 		pp = page_numtopp_nolock(PTE2PFN(oldpte, ht->ht_level));
2624 		if (pp == NULL)
2625 			continue;
2626 		x86_hm_enter(pp);
2627 
2628 		newpte = oldpte;
2629 		/*
2630 		 * We found a page table entry in the desired range,
2631 		 * figure out the new attributes.
2632 		 */
2633 		if (what == HAT_SET_ATTR || what == HAT_LOAD_ATTR) {
2634 			if ((attr & PROT_WRITE) &&
2635 			    !PTE_GET(oldpte, PT_WRITABLE))
2636 				newpte |= PT_WRITABLE;
2637 
2638 			if ((attr & HAT_NOSYNC) &&
2639 			    PTE_GET(oldpte, PT_SOFTWARE) < PT_NOSYNC)
2640 				newpte |= PT_NOSYNC;
2641 
2642 			if ((attr & PROT_EXEC) && PTE_GET(oldpte, mmu.pt_nx))
2643 				newpte &= ~mmu.pt_nx;
2644 		}
2645 
2646 		if (what == HAT_LOAD_ATTR) {
2647 			if (!(attr & PROT_WRITE) &&
2648 			    PTE_GET(oldpte, PT_WRITABLE))
2649 				newpte &= ~PT_WRITABLE;
2650 
2651 			if (!(attr & HAT_NOSYNC) &&
2652 			    PTE_GET(oldpte, PT_SOFTWARE) >= PT_NOSYNC)
2653 				newpte &= ~PT_SOFTWARE;
2654 
2655 			if (!(attr & PROT_EXEC) && !PTE_GET(oldpte, mmu.pt_nx))
2656 				newpte |= mmu.pt_nx;
2657 		}
2658 
2659 		if (what == HAT_CLR_ATTR) {
2660 			if ((attr & PROT_WRITE) && PTE_GET(oldpte, PT_WRITABLE))
2661 				newpte &= ~PT_WRITABLE;
2662 
2663 			if ((attr & HAT_NOSYNC) &&
2664 			    PTE_GET(oldpte, PT_SOFTWARE) >= PT_NOSYNC)
2665 				newpte &= ~PT_SOFTWARE;
2666 
2667 			if ((attr & PROT_EXEC) && !PTE_GET(oldpte, mmu.pt_nx))
2668 				newpte |= mmu.pt_nx;
2669 		}
2670 
2671 		/*
2672 		 * Ensure NOSYNC/NOCONSIST mappings have REF and MOD set.
2673 		 * x86pte_set() depends on this.
2674 		 */
2675 		if (PTE_GET(newpte, PT_SOFTWARE) >= PT_NOSYNC)
2676 			newpte |= PT_REF | PT_MOD;
2677 
2678 		/*
2679 		 * what about PROT_READ or others? this code only handles:
2680 		 * EXEC, WRITE, NOSYNC
2681 		 */
2682 
2683 		/*
2684 		 * If new PTE really changed, update the table.
2685 		 */
2686 		if (newpte != oldpte) {
2687 			entry = htable_va2entry(vaddr, ht);
2688 			oldpte = hati_update_pte(ht, entry, oldpte, newpte);
2689 			if (oldpte != 0) {
2690 				x86_hm_exit(pp);
2691 				goto try_again;
2692 			}
2693 		}
2694 		x86_hm_exit(pp);
2695 	}
2696 	if (ht)
2697 		htable_release(ht);
2698 	XPV_ALLOW_MIGRATE();
2699 }
2700 
2701 /*
2702  * Various wrappers for hat_updateattr()
2703  */
2704 void
2705 hat_setattr(hat_t *hat, caddr_t addr, size_t len, uint_t attr)
2706 {
2707 	ASSERT(hat == kas.a_hat || (uintptr_t)addr + len <= _userlimit);
2708 	hat_updateattr(hat, addr, len, attr, HAT_SET_ATTR);
2709 }
2710 
2711 void
2712 hat_clrattr(hat_t *hat, caddr_t addr, size_t len, uint_t attr)
2713 {
2714 	ASSERT(hat == kas.a_hat || (uintptr_t)addr + len <= _userlimit);
2715 	hat_updateattr(hat, addr, len, attr, HAT_CLR_ATTR);
2716 }
2717 
2718 void
2719 hat_chgattr(hat_t *hat, caddr_t addr, size_t len, uint_t attr)
2720 {
2721 	ASSERT(hat == kas.a_hat || (uintptr_t)addr + len <= _userlimit);
2722 	hat_updateattr(hat, addr, len, attr, HAT_LOAD_ATTR);
2723 }
2724 
2725 void
2726 hat_chgprot(hat_t *hat, caddr_t addr, size_t len, uint_t vprot)
2727 {
2728 	ASSERT(hat == kas.a_hat || (uintptr_t)addr + len <= _userlimit);
2729 	hat_updateattr(hat, addr, len, vprot & HAT_PROT_MASK, HAT_LOAD_ATTR);
2730 }
2731 
2732 /*
2733  * size_t hat_getpagesize(hat, addr)
2734  *	returns pagesize in bytes for <hat, addr>. returns -1 of there is
2735  *	no mapping. This is an advisory call.
2736  */
2737 ssize_t
2738 hat_getpagesize(hat_t *hat, caddr_t addr)
2739 {
2740 	uintptr_t	vaddr = ALIGN2PAGE(addr);
2741 	htable_t	*ht;
2742 	size_t		pagesize;
2743 
2744 	ASSERT(hat == kas.a_hat || vaddr <= _userlimit);
2745 	if (IN_VA_HOLE(vaddr))
2746 		return (-1);
2747 	ht = htable_getpage(hat, vaddr, NULL);
2748 	if (ht == NULL)
2749 		return (-1);
2750 	pagesize = LEVEL_SIZE(ht->ht_level);
2751 	htable_release(ht);
2752 	return (pagesize);
2753 }
2754 
2755 
2756 
2757 /*
2758  * pfn_t hat_getpfnum(hat, addr)
2759  *	returns pfn for <hat, addr> or PFN_INVALID if mapping is invalid.
2760  */
2761 pfn_t
2762 hat_getpfnum(hat_t *hat, caddr_t addr)
2763 {
2764 	uintptr_t	vaddr = ALIGN2PAGE(addr);
2765 	htable_t	*ht;
2766 	uint_t		entry;
2767 	pfn_t		pfn = PFN_INVALID;
2768 
2769 	ASSERT(hat == kas.a_hat || vaddr <= _userlimit);
2770 	if (khat_running == 0)
2771 		return (PFN_INVALID);
2772 
2773 	if (IN_VA_HOLE(vaddr))
2774 		return (PFN_INVALID);
2775 
2776 	XPV_DISALLOW_MIGRATE();
2777 	/*
2778 	 * A very common use of hat_getpfnum() is from the DDI for kernel pages.
2779 	 * Use the kmap_ptes (which also covers the 32 bit heap) to speed
2780 	 * this up.
2781 	 */
2782 	if (mmu.kmap_addr <= vaddr && vaddr < mmu.kmap_eaddr) {
2783 		x86pte_t pte;
2784 		pgcnt_t pg_index;
2785 
2786 		pg_index = mmu_btop(vaddr - mmu.kmap_addr);
2787 		pte = GET_PTE(PT_INDEX_PTR(mmu.kmap_ptes, pg_index));
2788 		if (PTE_ISVALID(pte))
2789 			/*LINTED [use of constant 0 causes a lint warning] */
2790 			pfn = PTE2PFN(pte, 0);
2791 		XPV_ALLOW_MIGRATE();
2792 		return (pfn);
2793 	}
2794 
2795 	ht = htable_getpage(hat, vaddr, &entry);
2796 	if (ht == NULL) {
2797 		XPV_ALLOW_MIGRATE();
2798 		return (PFN_INVALID);
2799 	}
2800 	ASSERT(vaddr >= ht->ht_vaddr);
2801 	ASSERT(vaddr <= HTABLE_LAST_PAGE(ht));
2802 	pfn = PTE2PFN(x86pte_get(ht, entry), ht->ht_level);
2803 	if (ht->ht_level > 0)
2804 		pfn += mmu_btop(vaddr & LEVEL_OFFSET(ht->ht_level));
2805 	htable_release(ht);
2806 	XPV_ALLOW_MIGRATE();
2807 	return (pfn);
2808 }
2809 
2810 /*
2811  * int hat_probe(hat, addr)
2812  *	return 0 if no valid mapping is present.  Faster version
2813  *	of hat_getattr in certain architectures.
2814  */
2815 int
2816 hat_probe(hat_t *hat, caddr_t addr)
2817 {
2818 	uintptr_t	vaddr = ALIGN2PAGE(addr);
2819 	uint_t		entry;
2820 	htable_t	*ht;
2821 	pgcnt_t		pg_off;
2822 
2823 	ASSERT(hat == kas.a_hat || vaddr <= _userlimit);
2824 	ASSERT(hat == kas.a_hat ||
2825 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
2826 	if (IN_VA_HOLE(vaddr))
2827 		return (0);
2828 
2829 	/*
2830 	 * Most common use of hat_probe is from segmap. We special case it
2831 	 * for performance.
2832 	 */
2833 	if (mmu.kmap_addr <= vaddr && vaddr < mmu.kmap_eaddr) {
2834 		pg_off = mmu_btop(vaddr - mmu.kmap_addr);
2835 		if (mmu.pae_hat)
2836 			return (PTE_ISVALID(mmu.kmap_ptes[pg_off]));
2837 		else
2838 			return (PTE_ISVALID(
2839 			    ((x86pte32_t *)mmu.kmap_ptes)[pg_off]));
2840 	}
2841 
2842 	ht = htable_getpage(hat, vaddr, &entry);
2843 	htable_release(ht);
2844 	return (ht != NULL);
2845 }
2846 
2847 /*
2848  * Find out if the segment for hat_share()/hat_unshare() is DISM or locked ISM.
2849  */
2850 static int
2851 is_it_dism(hat_t *hat, caddr_t va)
2852 {
2853 	struct seg *seg;
2854 	struct shm_data *shmd;
2855 	struct spt_data *sptd;
2856 
2857 	seg = as_findseg(hat->hat_as, va, 0);
2858 	ASSERT(seg != NULL);
2859 	ASSERT(seg->s_base <= va);
2860 	shmd = (struct shm_data *)seg->s_data;
2861 	ASSERT(shmd != NULL);
2862 	sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2863 	ASSERT(sptd != NULL);
2864 	if (sptd->spt_flags & SHM_PAGEABLE)
2865 		return (1);
2866 	return (0);
2867 }
2868 
2869 /*
2870  * Simple implementation of ISM. hat_share() is similar to hat_memload_array(),
2871  * except that we use the ism_hat's existing mappings to determine the pages
2872  * and protections to use for this hat. If we find a full properly aligned
2873  * and sized pagetable, we will attempt to share the pagetable itself.
2874  */
2875 /*ARGSUSED*/
2876 int
2877 hat_share(
2878 	hat_t		*hat,
2879 	caddr_t		addr,
2880 	hat_t		*ism_hat,
2881 	caddr_t		src_addr,
2882 	size_t		len,	/* almost useless value, see below.. */
2883 	uint_t		ismszc)
2884 {
2885 	uintptr_t	vaddr_start = (uintptr_t)addr;
2886 	uintptr_t	vaddr;
2887 	uintptr_t	eaddr = vaddr_start + len;
2888 	uintptr_t	ism_addr_start = (uintptr_t)src_addr;
2889 	uintptr_t	ism_addr = ism_addr_start;
2890 	uintptr_t	e_ism_addr = ism_addr + len;
2891 	htable_t	*ism_ht = NULL;
2892 	htable_t	*ht;
2893 	x86pte_t	pte;
2894 	page_t		*pp;
2895 	pfn_t		pfn;
2896 	level_t		l;
2897 	pgcnt_t		pgcnt;
2898 	uint_t		prot;
2899 	int		is_dism;
2900 	int		flags;
2901 
2902 	/*
2903 	 * We might be asked to share an empty DISM hat by as_dup()
2904 	 */
2905 	ASSERT(hat != kas.a_hat);
2906 	ASSERT(eaddr <= _userlimit);
2907 	if (!(ism_hat->hat_flags & HAT_SHARED)) {
2908 		ASSERT(hat_get_mapped_size(ism_hat) == 0);
2909 		return (0);
2910 	}
2911 	XPV_DISALLOW_MIGRATE();
2912 
2913 	/*
2914 	 * The SPT segment driver often passes us a size larger than there are
2915 	 * valid mappings. That's because it rounds the segment size up to a
2916 	 * large pagesize, even if the actual memory mapped by ism_hat is less.
2917 	 */
2918 	ASSERT(IS_PAGEALIGNED(vaddr_start));
2919 	ASSERT(IS_PAGEALIGNED(ism_addr_start));
2920 	ASSERT(ism_hat->hat_flags & HAT_SHARED);
2921 	is_dism = is_it_dism(hat, addr);
2922 	while (ism_addr < e_ism_addr) {
2923 		/*
2924 		 * use htable_walk to get the next valid ISM mapping
2925 		 */
2926 		pte = htable_walk(ism_hat, &ism_ht, &ism_addr, e_ism_addr);
2927 		if (ism_ht == NULL)
2928 			break;
2929 
2930 		/*
2931 		 * First check to see if we already share the page table.
2932 		 */
2933 		l = ism_ht->ht_level;
2934 		vaddr = vaddr_start + (ism_addr - ism_addr_start);
2935 		ht = htable_lookup(hat, vaddr, l);
2936 		if (ht != NULL) {
2937 			if (ht->ht_flags & HTABLE_SHARED_PFN)
2938 				goto shared;
2939 			htable_release(ht);
2940 			goto not_shared;
2941 		}
2942 
2943 		/*
2944 		 * Can't ever share top table.
2945 		 */
2946 		if (l == mmu.max_level)
2947 			goto not_shared;
2948 
2949 		/*
2950 		 * Avoid level mismatches later due to DISM faults.
2951 		 */
2952 		if (is_dism && l > 0)
2953 			goto not_shared;
2954 
2955 		/*
2956 		 * addresses and lengths must align
2957 		 * table must be fully populated
2958 		 * no lower level page tables
2959 		 */
2960 		if (ism_addr != ism_ht->ht_vaddr ||
2961 		    (vaddr & LEVEL_OFFSET(l + 1)) != 0)
2962 			goto not_shared;
2963 
2964 		/*
2965 		 * The range of address space must cover a full table.
2966 		 */
2967 		if (e_ism_addr - ism_addr < LEVEL_SIZE(l + 1))
2968 			goto not_shared;
2969 
2970 		/*
2971 		 * All entries in the ISM page table must be leaf PTEs.
2972 		 */
2973 		if (l > 0) {
2974 			int e;
2975 
2976 			/*
2977 			 * We know the 0th is from htable_walk() above.
2978 			 */
2979 			for (e = 1; e < HTABLE_NUM_PTES(ism_ht); ++e) {
2980 				x86pte_t pte;
2981 				pte = x86pte_get(ism_ht, e);
2982 				if (!PTE_ISPAGE(pte, l))
2983 					goto not_shared;
2984 			}
2985 		}
2986 
2987 		/*
2988 		 * share the page table
2989 		 */
2990 		ht = htable_create(hat, vaddr, l, ism_ht);
2991 shared:
2992 		ASSERT(ht->ht_flags & HTABLE_SHARED_PFN);
2993 		ASSERT(ht->ht_shares == ism_ht);
2994 		hat->hat_ism_pgcnt +=
2995 		    (ism_ht->ht_valid_cnt - ht->ht_valid_cnt) <<
2996 		    (LEVEL_SHIFT(ht->ht_level) - MMU_PAGESHIFT);
2997 		ht->ht_valid_cnt = ism_ht->ht_valid_cnt;
2998 		htable_release(ht);
2999 		ism_addr = ism_ht->ht_vaddr + LEVEL_SIZE(l + 1);
3000 		htable_release(ism_ht);
3001 		ism_ht = NULL;
3002 		continue;
3003 
3004 not_shared:
3005 		/*
3006 		 * Unable to share the page table. Instead we will
3007 		 * create new mappings from the values in the ISM mappings.
3008 		 * Figure out what level size mappings to use;
3009 		 */
3010 		for (l = ism_ht->ht_level; l > 0; --l) {
3011 			if (LEVEL_SIZE(l) <= eaddr - vaddr &&
3012 			    (vaddr & LEVEL_OFFSET(l)) == 0)
3013 				break;
3014 		}
3015 
3016 		/*
3017 		 * The ISM mapping might be larger than the share area,
3018 		 * be careful to truncate it if needed.
3019 		 */
3020 		if (eaddr - vaddr >= LEVEL_SIZE(ism_ht->ht_level)) {
3021 			pgcnt = mmu_btop(LEVEL_SIZE(ism_ht->ht_level));
3022 		} else {
3023 			pgcnt = mmu_btop(eaddr - vaddr);
3024 			l = 0;
3025 		}
3026 
3027 		pfn = PTE2PFN(pte, ism_ht->ht_level);
3028 		ASSERT(pfn != PFN_INVALID);
3029 		while (pgcnt > 0) {
3030 			/*
3031 			 * Make a new pte for the PFN for this level.
3032 			 * Copy protections for the pte from the ISM pte.
3033 			 */
3034 			pp = page_numtopp_nolock(pfn);
3035 			ASSERT(pp != NULL);
3036 
3037 			prot = PROT_USER | PROT_READ | HAT_UNORDERED_OK;
3038 			if (PTE_GET(pte, PT_WRITABLE))
3039 				prot |= PROT_WRITE;
3040 			if (!PTE_GET(pte, PT_NX))
3041 				prot |= PROT_EXEC;
3042 
3043 			flags = HAT_LOAD;
3044 			if (!is_dism)
3045 				flags |= HAT_LOAD_LOCK | HAT_LOAD_NOCONSIST;
3046 			while (hati_load_common(hat, vaddr, pp, prot, flags,
3047 			    l, pfn) != 0) {
3048 				if (l == 0)
3049 					panic("hati_load_common() failure");
3050 				--l;
3051 			}
3052 
3053 			vaddr += LEVEL_SIZE(l);
3054 			ism_addr += LEVEL_SIZE(l);
3055 			pfn += mmu_btop(LEVEL_SIZE(l));
3056 			pgcnt -= mmu_btop(LEVEL_SIZE(l));
3057 		}
3058 	}
3059 	if (ism_ht != NULL)
3060 		htable_release(ism_ht);
3061 	XPV_ALLOW_MIGRATE();
3062 	return (0);
3063 }
3064 
3065 
3066 /*
3067  * hat_unshare() is similar to hat_unload_callback(), but
3068  * we have to look for empty shared pagetables. Note that
3069  * hat_unshare() is always invoked against an entire segment.
3070  */
3071 /*ARGSUSED*/
3072 void
3073 hat_unshare(hat_t *hat, caddr_t addr, size_t len, uint_t ismszc)
3074 {
3075 	uint64_t	vaddr = (uintptr_t)addr;
3076 	uintptr_t	eaddr = vaddr + len;
3077 	htable_t	*ht = NULL;
3078 	uint_t		need_demaps = 0;
3079 	int		flags = HAT_UNLOAD_UNMAP;
3080 	level_t		l;
3081 
3082 	ASSERT(hat != kas.a_hat);
3083 	ASSERT(eaddr <= _userlimit);
3084 	ASSERT(IS_PAGEALIGNED(vaddr));
3085 	ASSERT(IS_PAGEALIGNED(eaddr));
3086 	XPV_DISALLOW_MIGRATE();
3087 
3088 	/*
3089 	 * First go through and remove any shared pagetables.
3090 	 *
3091 	 * Note that it's ok to delay the TLB shootdown till the entire range is
3092 	 * finished, because if hat_pageunload() were to unload a shared
3093 	 * pagetable page, its hat_tlb_inval() will do a global TLB invalidate.
3094 	 */
3095 	l = mmu.max_page_level;
3096 	if (l == mmu.max_level)
3097 		--l;
3098 	for (; l >= 0; --l) {
3099 		for (vaddr = (uintptr_t)addr; vaddr < eaddr;
3100 		    vaddr = (vaddr & LEVEL_MASK(l + 1)) + LEVEL_SIZE(l + 1)) {
3101 			ASSERT(!IN_VA_HOLE(vaddr));
3102 			/*
3103 			 * find a pagetable that maps the current address
3104 			 */
3105 			ht = htable_lookup(hat, vaddr, l);
3106 			if (ht == NULL)
3107 				continue;
3108 			if (ht->ht_flags & HTABLE_SHARED_PFN) {
3109 				/*
3110 				 * clear page count, set valid_cnt to 0,
3111 				 * let htable_release() finish the job
3112 				 */
3113 				hat->hat_ism_pgcnt -= ht->ht_valid_cnt <<
3114 				    (LEVEL_SHIFT(ht->ht_level) - MMU_PAGESHIFT);
3115 				ht->ht_valid_cnt = 0;
3116 				need_demaps = 1;
3117 			}
3118 			htable_release(ht);
3119 		}
3120 	}
3121 
3122 	/*
3123 	 * flush the TLBs - since we're probably dealing with MANY mappings
3124 	 * we do just one CR3 reload.
3125 	 */
3126 	if (!(hat->hat_flags & HAT_FREEING) && need_demaps)
3127 		hat_tlb_inval(hat, DEMAP_ALL_ADDR);
3128 
3129 	/*
3130 	 * Now go back and clean up any unaligned mappings that
3131 	 * couldn't share pagetables.
3132 	 */
3133 	if (!is_it_dism(hat, addr))
3134 		flags |= HAT_UNLOAD_UNLOCK;
3135 	hat_unload(hat, addr, len, flags);
3136 	XPV_ALLOW_MIGRATE();
3137 }
3138 
3139 
3140 /*
3141  * hat_reserve() does nothing
3142  */
3143 /*ARGSUSED*/
3144 void
3145 hat_reserve(struct as *as, caddr_t addr, size_t len)
3146 {
3147 }
3148 
3149 
3150 /*
3151  * Called when all mappings to a page should have write permission removed.
3152  * Mostly stolen from hat_pagesync()
3153  */
3154 static void
3155 hati_page_clrwrt(struct page *pp)
3156 {
3157 	hment_t		*hm = NULL;
3158 	htable_t	*ht;
3159 	uint_t		entry;
3160 	x86pte_t	old;
3161 	x86pte_t	new;
3162 	uint_t		pszc = 0;
3163 
3164 	XPV_DISALLOW_MIGRATE();
3165 next_size:
3166 	/*
3167 	 * walk thru the mapping list clearing write permission
3168 	 */
3169 	x86_hm_enter(pp);
3170 	while ((hm = hment_walk(pp, &ht, &entry, hm)) != NULL) {
3171 		if (ht->ht_level < pszc)
3172 			continue;
3173 		old = x86pte_get(ht, entry);
3174 
3175 		for (;;) {
3176 			/*
3177 			 * Is this mapping of interest?
3178 			 */
3179 			if (PTE2PFN(old, ht->ht_level) != pp->p_pagenum ||
3180 			    PTE_GET(old, PT_WRITABLE) == 0)
3181 				break;
3182 
3183 			/*
3184 			 * Clear ref/mod writable bits. This requires cross
3185 			 * calls to ensure any executing TLBs see cleared bits.
3186 			 */
3187 			new = old;
3188 			PTE_CLR(new, PT_REF | PT_MOD | PT_WRITABLE);
3189 			old = hati_update_pte(ht, entry, old, new);
3190 			if (old != 0)
3191 				continue;
3192 
3193 			break;
3194 		}
3195 	}
3196 	x86_hm_exit(pp);
3197 	while (pszc < pp->p_szc) {
3198 		page_t *tpp;
3199 		pszc++;
3200 		tpp = PP_GROUPLEADER(pp, pszc);
3201 		if (pp != tpp) {
3202 			pp = tpp;
3203 			goto next_size;
3204 		}
3205 	}
3206 	XPV_ALLOW_MIGRATE();
3207 }
3208 
3209 /*
3210  * void hat_page_setattr(pp, flag)
3211  * void hat_page_clrattr(pp, flag)
3212  *	used to set/clr ref/mod bits.
3213  */
3214 void
3215 hat_page_setattr(struct page *pp, uint_t flag)
3216 {
3217 	vnode_t		*vp = pp->p_vnode;
3218 	kmutex_t	*vphm = NULL;
3219 	page_t		**listp;
3220 	int		noshuffle;
3221 
3222 	noshuffle = flag & P_NSH;
3223 	flag &= ~P_NSH;
3224 
3225 	if (PP_GETRM(pp, flag) == flag)
3226 		return;
3227 
3228 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) &&
3229 	    !noshuffle) {
3230 		vphm = page_vnode_mutex(vp);
3231 		mutex_enter(vphm);
3232 	}
3233 
3234 	PP_SETRM(pp, flag);
3235 
3236 	if (vphm != NULL) {
3237 
3238 		/*
3239 		 * Some File Systems examine v_pages for NULL w/o
3240 		 * grabbing the vphm mutex. Must not let it become NULL when
3241 		 * pp is the only page on the list.
3242 		 */
3243 		if (pp->p_vpnext != pp) {
3244 			page_vpsub(&vp->v_pages, pp);
3245 			if (vp->v_pages != NULL)
3246 				listp = &vp->v_pages->p_vpprev->p_vpnext;
3247 			else
3248 				listp = &vp->v_pages;
3249 			page_vpadd(listp, pp);
3250 		}
3251 		mutex_exit(vphm);
3252 	}
3253 }
3254 
3255 void
3256 hat_page_clrattr(struct page *pp, uint_t flag)
3257 {
3258 	vnode_t		*vp = pp->p_vnode;
3259 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
3260 
3261 	/*
3262 	 * Caller is expected to hold page's io lock for VMODSORT to work
3263 	 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod
3264 	 * bit is cleared.
3265 	 * We don't have assert to avoid tripping some existing third party
3266 	 * code. The dirty page is moved back to top of the v_page list
3267 	 * after IO is done in pvn_write_done().
3268 	 */
3269 	PP_CLRRM(pp, flag);
3270 
3271 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
3272 
3273 		/*
3274 		 * VMODSORT works by removing write permissions and getting
3275 		 * a fault when a page is made dirty. At this point
3276 		 * we need to remove write permission from all mappings
3277 		 * to this page.
3278 		 */
3279 		hati_page_clrwrt(pp);
3280 	}
3281 }
3282 
3283 /*
3284  *	If flag is specified, returns 0 if attribute is disabled
3285  *	and non zero if enabled.  If flag specifes multiple attributes
3286  *	then returns 0 if ALL attributes are disabled.  This is an advisory
3287  *	call.
3288  */
3289 uint_t
3290 hat_page_getattr(struct page *pp, uint_t flag)
3291 {
3292 	return (PP_GETRM(pp, flag));
3293 }
3294 
3295 
3296 /*
3297  * common code used by hat_pageunload() and hment_steal()
3298  */
3299 hment_t *
3300 hati_page_unmap(page_t *pp, htable_t *ht, uint_t entry)
3301 {
3302 	x86pte_t old_pte;
3303 	pfn_t pfn = pp->p_pagenum;
3304 	hment_t *hm;
3305 
3306 	/*
3307 	 * We need to acquire a hold on the htable in order to
3308 	 * do the invalidate. We know the htable must exist, since
3309 	 * unmap's don't release the htable until after removing any
3310 	 * hment. Having x86_hm_enter() keeps that from proceeding.
3311 	 */
3312 	htable_acquire(ht);
3313 
3314 	/*
3315 	 * Invalidate the PTE and remove the hment.
3316 	 */
3317 	old_pte = x86pte_inval(ht, entry, 0, NULL);
3318 	if (PTE2PFN(old_pte, ht->ht_level) != pfn) {
3319 		panic("x86pte_inval() failure found PTE = " FMT_PTE
3320 		    " pfn being unmapped is %lx ht=0x%lx entry=0x%x",
3321 		    old_pte, pfn, (uintptr_t)ht, entry);
3322 	}
3323 
3324 	/*
3325 	 * Clean up all the htable information for this mapping
3326 	 */
3327 	ASSERT(ht->ht_valid_cnt > 0);
3328 	HTABLE_DEC(ht->ht_valid_cnt);
3329 	PGCNT_DEC(ht->ht_hat, ht->ht_level);
3330 
3331 	/*
3332 	 * sync ref/mod bits to the page_t
3333 	 */
3334 	if (PTE_GET(old_pte, PT_SOFTWARE) < PT_NOSYNC)
3335 		hati_sync_pte_to_page(pp, old_pte, ht->ht_level);
3336 
3337 	/*
3338 	 * Remove the mapping list entry for this page.
3339 	 */
3340 	hm = hment_remove(pp, ht, entry);
3341 
3342 	/*
3343 	 * drop the mapping list lock so that we might free the
3344 	 * hment and htable.
3345 	 */
3346 	x86_hm_exit(pp);
3347 	htable_release(ht);
3348 	return (hm);
3349 }
3350 
3351 extern int	vpm_enable;
3352 /*
3353  * Unload all translations to a page. If the page is a subpage of a large
3354  * page, the large page mappings are also removed.
3355  *
3356  * The forceflags are unused.
3357  */
3358 
3359 /*ARGSUSED*/
3360 static int
3361 hati_pageunload(struct page *pp, uint_t pg_szcd, uint_t forceflag)
3362 {
3363 	page_t		*cur_pp = pp;
3364 	hment_t		*hm;
3365 	hment_t		*prev;
3366 	htable_t	*ht;
3367 	uint_t		entry;
3368 	level_t		level;
3369 
3370 	XPV_DISALLOW_MIGRATE();
3371 
3372 	/*
3373 	 * prevent recursion due to kmem_free()
3374 	 */
3375 	++curthread->t_hatdepth;
3376 	ASSERT(curthread->t_hatdepth < 16);
3377 
3378 #if defined(__amd64)
3379 	/*
3380 	 * clear the vpm ref.
3381 	 */
3382 	if (vpm_enable) {
3383 		pp->p_vpmref = 0;
3384 	}
3385 #endif
3386 	/*
3387 	 * The loop with next_size handles pages with multiple pagesize mappings
3388 	 */
3389 next_size:
3390 	for (;;) {
3391 
3392 		/*
3393 		 * Get a mapping list entry
3394 		 */
3395 		x86_hm_enter(cur_pp);
3396 		for (prev = NULL; ; prev = hm) {
3397 			hm = hment_walk(cur_pp, &ht, &entry, prev);
3398 			if (hm == NULL) {
3399 				x86_hm_exit(cur_pp);
3400 
3401 				/*
3402 				 * If not part of a larger page, we're done.
3403 				 */
3404 				if (cur_pp->p_szc <= pg_szcd) {
3405 					ASSERT(curthread->t_hatdepth > 0);
3406 					--curthread->t_hatdepth;
3407 					XPV_ALLOW_MIGRATE();
3408 					return (0);
3409 				}
3410 
3411 				/*
3412 				 * Else check the next larger page size.
3413 				 * hat_page_demote() may decrease p_szc
3414 				 * but that's ok we'll just take an extra
3415 				 * trip discover there're no larger mappings
3416 				 * and return.
3417 				 */
3418 				++pg_szcd;
3419 				cur_pp = PP_GROUPLEADER(cur_pp, pg_szcd);
3420 				goto next_size;
3421 			}
3422 
3423 			/*
3424 			 * If this mapping size matches, remove it.
3425 			 */
3426 			level = ht->ht_level;
3427 			if (level == pg_szcd)
3428 				break;
3429 		}
3430 
3431 		/*
3432 		 * Remove the mapping list entry for this page.
3433 		 * Note this does the x86_hm_exit() for us.
3434 		 */
3435 		hm = hati_page_unmap(cur_pp, ht, entry);
3436 		if (hm != NULL)
3437 			hment_free(hm);
3438 	}
3439 }
3440 
3441 int
3442 hat_pageunload(struct page *pp, uint_t forceflag)
3443 {
3444 	ASSERT(PAGE_EXCL(pp));
3445 	return (hati_pageunload(pp, 0, forceflag));
3446 }
3447 
3448 /*
3449  * Unload all large mappings to pp and reduce by 1 p_szc field of every large
3450  * page level that included pp.
3451  *
3452  * pp must be locked EXCL. Even though no other constituent pages are locked
3453  * it's legal to unload large mappings to pp because all constituent pages of
3454  * large locked mappings have to be locked SHARED.  therefore if we have EXCL
3455  * lock on one of constituent pages none of the large mappings to pp are
3456  * locked.
3457  *
3458  * Change (always decrease) p_szc field starting from the last constituent
3459  * page and ending with root constituent page so that root's pszc always shows
3460  * the area where hat_page_demote() may be active.
3461  *
3462  * This mechanism is only used for file system pages where it's not always
3463  * possible to get EXCL locks on all constituent pages to demote the size code
3464  * (as is done for anonymous or kernel large pages).
3465  */
3466 void
3467 hat_page_demote(page_t *pp)
3468 {
3469 	uint_t		pszc;
3470 	uint_t		rszc;
3471 	uint_t		szc;
3472 	page_t		*rootpp;
3473 	page_t		*firstpp;
3474 	page_t		*lastpp;
3475 	pgcnt_t		pgcnt;
3476 
3477 	ASSERT(PAGE_EXCL(pp));
3478 	ASSERT(!PP_ISFREE(pp));
3479 	ASSERT(page_szc_lock_assert(pp));
3480 
3481 	if (pp->p_szc == 0)
3482 		return;
3483 
3484 	rootpp = PP_GROUPLEADER(pp, 1);
3485 	(void) hati_pageunload(rootpp, 1, HAT_FORCE_PGUNLOAD);
3486 
3487 	/*
3488 	 * all large mappings to pp are gone
3489 	 * and no new can be setup since pp is locked exclusively.
3490 	 *
3491 	 * Lock the root to make sure there's only one hat_page_demote()
3492 	 * outstanding within the area of this root's pszc.
3493 	 *
3494 	 * Second potential hat_page_demote() is already eliminated by upper
3495 	 * VM layer via page_szc_lock() but we don't rely on it and use our
3496 	 * own locking (so that upper layer locking can be changed without
3497 	 * assumptions that hat depends on upper layer VM to prevent multiple
3498 	 * hat_page_demote() to be issued simultaneously to the same large
3499 	 * page).
3500 	 */
3501 again:
3502 	pszc = pp->p_szc;
3503 	if (pszc == 0)
3504 		return;
3505 	rootpp = PP_GROUPLEADER(pp, pszc);
3506 	x86_hm_enter(rootpp);
3507 	/*
3508 	 * If root's p_szc is different from pszc we raced with another
3509 	 * hat_page_demote().  Drop the lock and try to find the root again.
3510 	 * If root's p_szc is greater than pszc previous hat_page_demote() is
3511 	 * not done yet.  Take and release mlist lock of root's root to wait
3512 	 * for previous hat_page_demote() to complete.
3513 	 */
3514 	if ((rszc = rootpp->p_szc) != pszc) {
3515 		x86_hm_exit(rootpp);
3516 		if (rszc > pszc) {
3517 			/* p_szc of a locked non free page can't increase */
3518 			ASSERT(pp != rootpp);
3519 
3520 			rootpp = PP_GROUPLEADER(rootpp, rszc);
3521 			x86_hm_enter(rootpp);
3522 			x86_hm_exit(rootpp);
3523 		}
3524 		goto again;
3525 	}
3526 	ASSERT(pp->p_szc == pszc);
3527 
3528 	/*
3529 	 * Decrement by 1 p_szc of every constituent page of a region that
3530 	 * covered pp. For example if original szc is 3 it gets changed to 2
3531 	 * everywhere except in region 2 that covered pp. Region 2 that
3532 	 * covered pp gets demoted to 1 everywhere except in region 1 that
3533 	 * covered pp. The region 1 that covered pp is demoted to region
3534 	 * 0. It's done this way because from region 3 we removed level 3
3535 	 * mappings, from region 2 that covered pp we removed level 2 mappings
3536 	 * and from region 1 that covered pp we removed level 1 mappings.  All
3537 	 * changes are done from from high pfn's to low pfn's so that roots
3538 	 * are changed last allowing one to know the largest region where
3539 	 * hat_page_demote() is stil active by only looking at the root page.
3540 	 *
3541 	 * This algorithm is implemented in 2 while loops. First loop changes
3542 	 * p_szc of pages to the right of pp's level 1 region and second
3543 	 * loop changes p_szc of pages of level 1 region that covers pp
3544 	 * and all pages to the left of level 1 region that covers pp.
3545 	 * In the first loop p_szc keeps dropping with every iteration
3546 	 * and in the second loop it keeps increasing with every iteration.
3547 	 *
3548 	 * First loop description: Demote pages to the right of pp outside of
3549 	 * level 1 region that covers pp.  In every iteration of the while
3550 	 * loop below find the last page of szc region and the first page of
3551 	 * (szc - 1) region that is immediately to the right of (szc - 1)
3552 	 * region that covers pp.  From last such page to first such page
3553 	 * change every page's szc to szc - 1. Decrement szc and continue
3554 	 * looping until szc is 1. If pp belongs to the last (szc - 1) region
3555 	 * of szc region skip to the next iteration.
3556 	 */
3557 	szc = pszc;
3558 	while (szc > 1) {
3559 		lastpp = PP_GROUPLEADER(pp, szc);
3560 		pgcnt = page_get_pagecnt(szc);
3561 		lastpp += pgcnt - 1;
3562 		firstpp = PP_GROUPLEADER(pp, (szc - 1));
3563 		pgcnt = page_get_pagecnt(szc - 1);
3564 		if (lastpp - firstpp < pgcnt) {
3565 			szc--;
3566 			continue;
3567 		}
3568 		firstpp += pgcnt;
3569 		while (lastpp != firstpp) {
3570 			ASSERT(lastpp->p_szc == pszc);
3571 			lastpp->p_szc = szc - 1;
3572 			lastpp--;
3573 		}
3574 		firstpp->p_szc = szc - 1;
3575 		szc--;
3576 	}
3577 
3578 	/*
3579 	 * Second loop description:
3580 	 * First iteration changes p_szc to 0 of every
3581 	 * page of level 1 region that covers pp.
3582 	 * Subsequent iterations find last page of szc region
3583 	 * immediately to the left of szc region that covered pp
3584 	 * and first page of (szc + 1) region that covers pp.
3585 	 * From last to first page change p_szc of every page to szc.
3586 	 * Increment szc and continue looping until szc is pszc.
3587 	 * If pp belongs to the fist szc region of (szc + 1) region
3588 	 * skip to the next iteration.
3589 	 *
3590 	 */
3591 	szc = 0;
3592 	while (szc < pszc) {
3593 		firstpp = PP_GROUPLEADER(pp, (szc + 1));
3594 		if (szc == 0) {
3595 			pgcnt = page_get_pagecnt(1);
3596 			lastpp = firstpp + (pgcnt - 1);
3597 		} else {
3598 			lastpp = PP_GROUPLEADER(pp, szc);
3599 			if (firstpp == lastpp) {
3600 				szc++;
3601 				continue;
3602 			}
3603 			lastpp--;
3604 			pgcnt = page_get_pagecnt(szc);
3605 		}
3606 		while (lastpp != firstpp) {
3607 			ASSERT(lastpp->p_szc == pszc);
3608 			lastpp->p_szc = szc;
3609 			lastpp--;
3610 		}
3611 		firstpp->p_szc = szc;
3612 		if (firstpp == rootpp)
3613 			break;
3614 		szc++;
3615 	}
3616 	x86_hm_exit(rootpp);
3617 }
3618 
3619 /*
3620  * get hw stats from hardware into page struct and reset hw stats
3621  * returns attributes of page
3622  * Flags for hat_pagesync, hat_getstat, hat_sync
3623  *
3624  * define	HAT_SYNC_ZERORM		0x01
3625  *
3626  * Additional flags for hat_pagesync
3627  *
3628  * define	HAT_SYNC_STOPON_REF	0x02
3629  * define	HAT_SYNC_STOPON_MOD	0x04
3630  * define	HAT_SYNC_STOPON_RM	0x06
3631  * define	HAT_SYNC_STOPON_SHARED	0x08
3632  */
3633 uint_t
3634 hat_pagesync(struct page *pp, uint_t flags)
3635 {
3636 	hment_t		*hm = NULL;
3637 	htable_t	*ht;
3638 	uint_t		entry;
3639 	x86pte_t	old, save_old;
3640 	x86pte_t	new;
3641 	uchar_t		nrmbits = P_REF|P_MOD|P_RO;
3642 	extern ulong_t	po_share;
3643 	page_t		*save_pp = pp;
3644 	uint_t		pszc = 0;
3645 
3646 	ASSERT(PAGE_LOCKED(pp) || panicstr);
3647 
3648 	if (PP_ISRO(pp) && (flags & HAT_SYNC_STOPON_MOD))
3649 		return (pp->p_nrm & nrmbits);
3650 
3651 	if ((flags & HAT_SYNC_ZERORM) == 0) {
3652 
3653 		if ((flags & HAT_SYNC_STOPON_REF) != 0 && PP_ISREF(pp))
3654 			return (pp->p_nrm & nrmbits);
3655 
3656 		if ((flags & HAT_SYNC_STOPON_MOD) != 0 && PP_ISMOD(pp))
3657 			return (pp->p_nrm & nrmbits);
3658 
3659 		if ((flags & HAT_SYNC_STOPON_SHARED) != 0 &&
3660 		    hat_page_getshare(pp) > po_share) {
3661 			if (PP_ISRO(pp))
3662 				PP_SETREF(pp);
3663 			return (pp->p_nrm & nrmbits);
3664 		}
3665 	}
3666 
3667 	XPV_DISALLOW_MIGRATE();
3668 next_size:
3669 	/*
3670 	 * walk thru the mapping list syncing (and clearing) ref/mod bits.
3671 	 */
3672 	x86_hm_enter(pp);
3673 	while ((hm = hment_walk(pp, &ht, &entry, hm)) != NULL) {
3674 		if (ht->ht_level < pszc)
3675 			continue;
3676 		old = x86pte_get(ht, entry);
3677 try_again:
3678 
3679 		ASSERT(PTE2PFN(old, ht->ht_level) == pp->p_pagenum);
3680 
3681 		if (PTE_GET(old, PT_REF | PT_MOD) == 0)
3682 			continue;
3683 
3684 		save_old = old;
3685 		if ((flags & HAT_SYNC_ZERORM) != 0) {
3686 
3687 			/*
3688 			 * Need to clear ref or mod bits. Need to demap
3689 			 * to make sure any executing TLBs see cleared bits.
3690 			 */
3691 			new = old;
3692 			PTE_CLR(new, PT_REF | PT_MOD);
3693 			old = hati_update_pte(ht, entry, old, new);
3694 			if (old != 0)
3695 				goto try_again;
3696 
3697 			old = save_old;
3698 		}
3699 
3700 		/*
3701 		 * Sync the PTE
3702 		 */
3703 		if (!(flags & HAT_SYNC_ZERORM) &&
3704 		    PTE_GET(old, PT_SOFTWARE) <= PT_NOSYNC)
3705 			hati_sync_pte_to_page(pp, old, ht->ht_level);
3706 
3707 		/*
3708 		 * can stop short if we found a ref'd or mod'd page
3709 		 */
3710 		if ((flags & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp) ||
3711 		    (flags & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)) {
3712 			x86_hm_exit(pp);
3713 			goto done;
3714 		}
3715 	}
3716 	x86_hm_exit(pp);
3717 	while (pszc < pp->p_szc) {
3718 		page_t *tpp;
3719 		pszc++;
3720 		tpp = PP_GROUPLEADER(pp, pszc);
3721 		if (pp != tpp) {
3722 			pp = tpp;
3723 			goto next_size;
3724 		}
3725 	}
3726 done:
3727 	XPV_ALLOW_MIGRATE();
3728 	return (save_pp->p_nrm & nrmbits);
3729 }
3730 
3731 /*
3732  * returns approx number of mappings to this pp.  A return of 0 implies
3733  * there are no mappings to the page.
3734  */
3735 ulong_t
3736 hat_page_getshare(page_t *pp)
3737 {
3738 	uint_t cnt;
3739 	cnt = hment_mapcnt(pp);
3740 #if defined(__amd64)
3741 	if (vpm_enable && pp->p_vpmref) {
3742 		cnt += 1;
3743 	}
3744 #endif
3745 	return (cnt);
3746 }
3747 
3748 /*
3749  * Return 1 the number of mappings exceeds sh_thresh. Return 0
3750  * otherwise.
3751  */
3752 int
3753 hat_page_checkshare(page_t *pp, ulong_t sh_thresh)
3754 {
3755 	return (hat_page_getshare(pp) > sh_thresh);
3756 }
3757 
3758 /*
3759  * hat_softlock isn't supported anymore
3760  */
3761 /*ARGSUSED*/
3762 faultcode_t
3763 hat_softlock(
3764 	hat_t *hat,
3765 	caddr_t addr,
3766 	size_t *len,
3767 	struct page **page_array,
3768 	uint_t flags)
3769 {
3770 	return (FC_NOSUPPORT);
3771 }
3772 
3773 
3774 
3775 /*
3776  * Routine to expose supported HAT features to platform independent code.
3777  */
3778 /*ARGSUSED*/
3779 int
3780 hat_supported(enum hat_features feature, void *arg)
3781 {
3782 	switch (feature) {
3783 
3784 	case HAT_SHARED_PT:	/* this is really ISM */
3785 		return (1);
3786 
3787 	case HAT_DYNAMIC_ISM_UNMAP:
3788 		return (0);
3789 
3790 	case HAT_VMODSORT:
3791 		return (1);
3792 
3793 	case HAT_SHARED_REGIONS:
3794 		return (0);
3795 
3796 	default:
3797 		panic("hat_supported() - unknown feature");
3798 	}
3799 	return (0);
3800 }
3801 
3802 /*
3803  * Called when a thread is exiting and has been switched to the kernel AS
3804  */
3805 void
3806 hat_thread_exit(kthread_t *thd)
3807 {
3808 	ASSERT(thd->t_procp->p_as == &kas);
3809 	XPV_DISALLOW_MIGRATE();
3810 	hat_switch(thd->t_procp->p_as->a_hat);
3811 	XPV_ALLOW_MIGRATE();
3812 }
3813 
3814 /*
3815  * Setup the given brand new hat structure as the new HAT on this cpu's mmu.
3816  */
3817 /*ARGSUSED*/
3818 void
3819 hat_setup(hat_t *hat, int flags)
3820 {
3821 	XPV_DISALLOW_MIGRATE();
3822 	kpreempt_disable();
3823 
3824 	hat_switch(hat);
3825 
3826 	kpreempt_enable();
3827 	XPV_ALLOW_MIGRATE();
3828 }
3829 
3830 /*
3831  * Prepare for a CPU private mapping for the given address.
3832  *
3833  * The address can only be used from a single CPU and can be remapped
3834  * using hat_mempte_remap().  Return the address of the PTE.
3835  *
3836  * We do the htable_create() if necessary and increment the valid count so
3837  * the htable can't disappear.  We also hat_devload() the page table into
3838  * kernel so that the PTE is quickly accessed.
3839  */
3840 hat_mempte_t
3841 hat_mempte_setup(caddr_t addr)
3842 {
3843 	uintptr_t	va = (uintptr_t)addr;
3844 	htable_t	*ht;
3845 	uint_t		entry;
3846 	x86pte_t	oldpte;
3847 	hat_mempte_t	p;
3848 
3849 	ASSERT(IS_PAGEALIGNED(va));
3850 	ASSERT(!IN_VA_HOLE(va));
3851 	++curthread->t_hatdepth;
3852 	XPV_DISALLOW_MIGRATE();
3853 	ht = htable_getpte(kas.a_hat, va, &entry, &oldpte, 0);
3854 	if (ht == NULL) {
3855 		ht = htable_create(kas.a_hat, va, 0, NULL);
3856 		entry = htable_va2entry(va, ht);
3857 		ASSERT(ht->ht_level == 0);
3858 		oldpte = x86pte_get(ht, entry);
3859 	}
3860 	if (PTE_ISVALID(oldpte))
3861 		panic("hat_mempte_setup(): address already mapped"
3862 		    "ht=%p, entry=%d, pte=" FMT_PTE, (void *)ht, entry, oldpte);
3863 
3864 	/*
3865 	 * increment ht_valid_cnt so that the pagetable can't disappear
3866 	 */
3867 	HTABLE_INC(ht->ht_valid_cnt);
3868 
3869 	/*
3870 	 * return the PTE physical address to the caller.
3871 	 */
3872 	htable_release(ht);
3873 	XPV_ALLOW_MIGRATE();
3874 	p = PT_INDEX_PHYSADDR(pfn_to_pa(ht->ht_pfn), entry);
3875 	--curthread->t_hatdepth;
3876 	return (p);
3877 }
3878 
3879 /*
3880  * Release a CPU private mapping for the given address.
3881  * We decrement the htable valid count so it might be destroyed.
3882  */
3883 /*ARGSUSED1*/
3884 void
3885 hat_mempte_release(caddr_t addr, hat_mempte_t pte_pa)
3886 {
3887 	htable_t	*ht;
3888 
3889 	XPV_DISALLOW_MIGRATE();
3890 	/*
3891 	 * invalidate any left over mapping and decrement the htable valid count
3892 	 */
3893 #ifdef __xpv
3894 	if (HYPERVISOR_update_va_mapping((uintptr_t)addr, 0,
3895 	    UVMF_INVLPG | UVMF_LOCAL))
3896 		panic("HYPERVISOR_update_va_mapping() failed");
3897 #else
3898 	{
3899 		x86pte_t *pteptr;
3900 
3901 		pteptr = x86pte_mapin(mmu_btop(pte_pa),
3902 		    (pte_pa & MMU_PAGEOFFSET) >> mmu.pte_size_shift, NULL);
3903 		if (mmu.pae_hat)
3904 			*pteptr = 0;
3905 		else
3906 			*(x86pte32_t *)pteptr = 0;
3907 		mmu_tlbflush_entry(addr);
3908 		x86pte_mapout();
3909 	}
3910 #endif
3911 
3912 	ht = htable_getpte(kas.a_hat, ALIGN2PAGE(addr), NULL, NULL, 0);
3913 	if (ht == NULL)
3914 		panic("hat_mempte_release(): invalid address");
3915 	ASSERT(ht->ht_level == 0);
3916 	HTABLE_DEC(ht->ht_valid_cnt);
3917 	htable_release(ht);
3918 	XPV_ALLOW_MIGRATE();
3919 }
3920 
3921 /*
3922  * Apply a temporary CPU private mapping to a page. We flush the TLB only
3923  * on this CPU, so this ought to have been called with preemption disabled.
3924  */
3925 void
3926 hat_mempte_remap(
3927 	pfn_t		pfn,
3928 	caddr_t		addr,
3929 	hat_mempte_t	pte_pa,
3930 	uint_t		attr,
3931 	uint_t		flags)
3932 {
3933 	uintptr_t	va = (uintptr_t)addr;
3934 	x86pte_t	pte;
3935 
3936 	/*
3937 	 * Remap the given PTE to the new page's PFN. Invalidate only
3938 	 * on this CPU.
3939 	 */
3940 #ifdef DEBUG
3941 	htable_t	*ht;
3942 	uint_t		entry;
3943 
3944 	ASSERT(IS_PAGEALIGNED(va));
3945 	ASSERT(!IN_VA_HOLE(va));
3946 	ht = htable_getpte(kas.a_hat, va, &entry, NULL, 0);
3947 	ASSERT(ht != NULL);
3948 	ASSERT(ht->ht_level == 0);
3949 	ASSERT(ht->ht_valid_cnt > 0);
3950 	ASSERT(ht->ht_pfn == mmu_btop(pte_pa));
3951 	htable_release(ht);
3952 #endif
3953 	XPV_DISALLOW_MIGRATE();
3954 	pte = hati_mkpte(pfn, attr, 0, flags);
3955 #ifdef __xpv
3956 	if (HYPERVISOR_update_va_mapping(va, pte, UVMF_INVLPG | UVMF_LOCAL))
3957 		panic("HYPERVISOR_update_va_mapping() failed");
3958 #else
3959 	{
3960 		x86pte_t *pteptr;
3961 
3962 		pteptr = x86pte_mapin(mmu_btop(pte_pa),
3963 		    (pte_pa & MMU_PAGEOFFSET) >> mmu.pte_size_shift, NULL);
3964 		if (mmu.pae_hat)
3965 			*(x86pte_t *)pteptr = pte;
3966 		else
3967 			*(x86pte32_t *)pteptr = (x86pte32_t)pte;
3968 		mmu_tlbflush_entry(addr);
3969 		x86pte_mapout();
3970 	}
3971 #endif
3972 	XPV_ALLOW_MIGRATE();
3973 }
3974 
3975 
3976 
3977 /*
3978  * Hat locking functions
3979  * XXX - these two functions are currently being used by hatstats
3980  * 	they can be removed by using a per-as mutex for hatstats.
3981  */
3982 void
3983 hat_enter(hat_t *hat)
3984 {
3985 	mutex_enter(&hat->hat_mutex);
3986 }
3987 
3988 void
3989 hat_exit(hat_t *hat)
3990 {
3991 	mutex_exit(&hat->hat_mutex);
3992 }
3993 
3994 /*
3995  * HAT part of cpu initialization.
3996  */
3997 void
3998 hat_cpu_online(struct cpu *cpup)
3999 {
4000 	if (cpup != CPU) {
4001 		x86pte_cpu_init(cpup);
4002 		hat_vlp_setup(cpup);
4003 	}
4004 	CPUSET_ATOMIC_ADD(khat_cpuset, cpup->cpu_id);
4005 }
4006 
4007 /*
4008  * HAT part of cpu deletion.
4009  * (currently, we only call this after the cpu is safely passivated.)
4010  */
4011 void
4012 hat_cpu_offline(struct cpu *cpup)
4013 {
4014 	ASSERT(cpup != CPU);
4015 
4016 	CPUSET_ATOMIC_DEL(khat_cpuset, cpup->cpu_id);
4017 	hat_vlp_teardown(cpup);
4018 	x86pte_cpu_fini(cpup);
4019 }
4020 
4021 /*
4022  * Function called after all CPUs are brought online.
4023  * Used to remove low address boot mappings.
4024  */
4025 void
4026 clear_boot_mappings(uintptr_t low, uintptr_t high)
4027 {
4028 	uintptr_t vaddr = low;
4029 	htable_t *ht = NULL;
4030 	level_t level;
4031 	uint_t entry;
4032 	x86pte_t pte;
4033 
4034 	/*
4035 	 * On 1st CPU we can unload the prom mappings, basically we blow away
4036 	 * all virtual mappings under _userlimit.
4037 	 */
4038 	while (vaddr < high) {
4039 		pte = htable_walk(kas.a_hat, &ht, &vaddr, high);
4040 		if (ht == NULL)
4041 			break;
4042 
4043 		level = ht->ht_level;
4044 		entry = htable_va2entry(vaddr, ht);
4045 		ASSERT(level <= mmu.max_page_level);
4046 		ASSERT(PTE_ISPAGE(pte, level));
4047 
4048 		/*
4049 		 * Unload the mapping from the page tables.
4050 		 */
4051 		(void) x86pte_inval(ht, entry, 0, NULL);
4052 		ASSERT(ht->ht_valid_cnt > 0);
4053 		HTABLE_DEC(ht->ht_valid_cnt);
4054 		PGCNT_DEC(ht->ht_hat, ht->ht_level);
4055 
4056 		vaddr += LEVEL_SIZE(ht->ht_level);
4057 	}
4058 	if (ht)
4059 		htable_release(ht);
4060 }
4061 
4062 /*
4063  * Atomically update a new translation for a single page.  If the
4064  * currently installed PTE doesn't match the value we expect to find,
4065  * it's not updated and we return the PTE we found.
4066  *
4067  * If activating nosync or NOWRITE and the page was modified we need to sync
4068  * with the page_t. Also sync with page_t if clearing ref/mod bits.
4069  */
4070 static x86pte_t
4071 hati_update_pte(htable_t *ht, uint_t entry, x86pte_t expected, x86pte_t new)
4072 {
4073 	page_t		*pp;
4074 	uint_t		rm = 0;
4075 	x86pte_t	replaced;
4076 
4077 	if (PTE_GET(expected, PT_SOFTWARE) < PT_NOSYNC &&
4078 	    PTE_GET(expected, PT_MOD | PT_REF) &&
4079 	    (PTE_GET(new, PT_NOSYNC) || !PTE_GET(new, PT_WRITABLE) ||
4080 	    !PTE_GET(new, PT_MOD | PT_REF))) {
4081 
4082 		ASSERT(!pfn_is_foreign(PTE2PFN(expected, ht->ht_level)));
4083 		pp = page_numtopp_nolock(PTE2PFN(expected, ht->ht_level));
4084 		ASSERT(pp != NULL);
4085 		if (PTE_GET(expected, PT_MOD))
4086 			rm |= P_MOD;
4087 		if (PTE_GET(expected, PT_REF))
4088 			rm |= P_REF;
4089 		PTE_CLR(new, PT_MOD | PT_REF);
4090 	}
4091 
4092 	replaced = x86pte_update(ht, entry, expected, new);
4093 	if (replaced != expected)
4094 		return (replaced);
4095 
4096 	if (rm) {
4097 		/*
4098 		 * sync to all constituent pages of a large page
4099 		 */
4100 		pgcnt_t pgcnt = page_get_pagecnt(ht->ht_level);
4101 		ASSERT(IS_P2ALIGNED(pp->p_pagenum, pgcnt));
4102 		while (pgcnt-- > 0) {
4103 			/*
4104 			 * hat_page_demote() can't decrease
4105 			 * pszc below this mapping size
4106 			 * since large mapping existed after we
4107 			 * took mlist lock.
4108 			 */
4109 			ASSERT(pp->p_szc >= ht->ht_level);
4110 			hat_page_setattr(pp, rm);
4111 			++pp;
4112 		}
4113 	}
4114 
4115 	return (0);
4116 }
4117 
4118 /* ARGSUSED */
4119 void
4120 hat_join_srd(struct hat *hat, vnode_t *evp)
4121 {
4122 }
4123 
4124 /* ARGSUSED */
4125 hat_region_cookie_t
4126 hat_join_region(struct hat *hat,
4127     caddr_t r_saddr,
4128     size_t r_size,
4129     void *r_obj,
4130     u_offset_t r_objoff,
4131     uchar_t r_perm,
4132     uchar_t r_pgszc,
4133     hat_rgn_cb_func_t r_cb_function,
4134     uint_t flags)
4135 {
4136 	panic("No shared region support on x86");
4137 	return (HAT_INVALID_REGION_COOKIE);
4138 }
4139 
4140 /* ARGSUSED */
4141 void
4142 hat_leave_region(struct hat *hat, hat_region_cookie_t rcookie, uint_t flags)
4143 {
4144 	panic("No shared region support on x86");
4145 }
4146 
4147 /* ARGSUSED */
4148 void
4149 hat_dup_region(struct hat *hat, hat_region_cookie_t rcookie)
4150 {
4151 	panic("No shared region support on x86");
4152 }
4153 
4154 
4155 /*
4156  * Kernel Physical Mapping (kpm) facility
4157  *
4158  * Most of the routines needed to support segkpm are almost no-ops on the
4159  * x86 platform.  We map in the entire segment when it is created and leave
4160  * it mapped in, so there is no additional work required to set up and tear
4161  * down individual mappings.  All of these routines were created to support
4162  * SPARC platforms that have to avoid aliasing in their virtually indexed
4163  * caches.
4164  *
4165  * Most of the routines have sanity checks in them (e.g. verifying that the
4166  * passed-in page is locked).  We don't actually care about most of these
4167  * checks on x86, but we leave them in place to identify problems in the
4168  * upper levels.
4169  */
4170 
4171 /*
4172  * Map in a locked page and return the vaddr.
4173  */
4174 /*ARGSUSED*/
4175 caddr_t
4176 hat_kpm_mapin(struct page *pp, struct kpme *kpme)
4177 {
4178 	caddr_t		vaddr;
4179 
4180 #ifdef DEBUG
4181 	if (kpm_enable == 0) {
4182 		cmn_err(CE_WARN, "hat_kpm_mapin: kpm_enable not set\n");
4183 		return ((caddr_t)NULL);
4184 	}
4185 
4186 	if (pp == NULL || PAGE_LOCKED(pp) == 0) {
4187 		cmn_err(CE_WARN, "hat_kpm_mapin: pp zero or not locked\n");
4188 		return ((caddr_t)NULL);
4189 	}
4190 #endif
4191 
4192 	vaddr = hat_kpm_page2va(pp, 1);
4193 
4194 	return (vaddr);
4195 }
4196 
4197 /*
4198  * Mapout a locked page.
4199  */
4200 /*ARGSUSED*/
4201 void
4202 hat_kpm_mapout(struct page *pp, struct kpme *kpme, caddr_t vaddr)
4203 {
4204 #ifdef DEBUG
4205 	if (kpm_enable == 0) {
4206 		cmn_err(CE_WARN, "hat_kpm_mapout: kpm_enable not set\n");
4207 		return;
4208 	}
4209 
4210 	if (IS_KPM_ADDR(vaddr) == 0) {
4211 		cmn_err(CE_WARN, "hat_kpm_mapout: no kpm address\n");
4212 		return;
4213 	}
4214 
4215 	if (pp == NULL || PAGE_LOCKED(pp) == 0) {
4216 		cmn_err(CE_WARN, "hat_kpm_mapout: page zero or not locked\n");
4217 		return;
4218 	}
4219 #endif
4220 }
4221 
4222 /*
4223  * hat_kpm_mapin_pfn is used to obtain a kpm mapping for physical
4224  * memory addresses that are not described by a page_t.  It can
4225  * also be used for normal pages that are not locked, but beware
4226  * this is dangerous - no locking is performed, so the identity of
4227  * the page could change.  hat_kpm_mapin_pfn is not supported when
4228  * vac_colors > 1, because the chosen va depends on the page identity,
4229  * which could change.
4230  * The caller must only pass pfn's for valid physical addresses; violation
4231  * of this rule will cause panic.
4232  */
4233 caddr_t
4234 hat_kpm_mapin_pfn(pfn_t pfn)
4235 {
4236 	caddr_t paddr, vaddr;
4237 
4238 	if (kpm_enable == 0)
4239 		return ((caddr_t)NULL);
4240 
4241 	paddr = (caddr_t)ptob(pfn);
4242 	vaddr = (uintptr_t)kpm_vbase + paddr;
4243 
4244 	return ((caddr_t)vaddr);
4245 }
4246 
4247 /*ARGSUSED*/
4248 void
4249 hat_kpm_mapout_pfn(pfn_t pfn)
4250 {
4251 	/* empty */
4252 }
4253 
4254 /*
4255  * Return the kpm virtual address for a specific pfn
4256  */
4257 caddr_t
4258 hat_kpm_pfn2va(pfn_t pfn)
4259 {
4260 	uintptr_t vaddr = (uintptr_t)kpm_vbase + mmu_ptob(pfn);
4261 
4262 	ASSERT(!pfn_is_foreign(pfn));
4263 	return ((caddr_t)vaddr);
4264 }
4265 
4266 /*
4267  * Return the kpm virtual address for the page at pp.
4268  */
4269 /*ARGSUSED*/
4270 caddr_t
4271 hat_kpm_page2va(struct page *pp, int checkswap)
4272 {
4273 	return (hat_kpm_pfn2va(pp->p_pagenum));
4274 }
4275 
4276 /*
4277  * Return the page frame number for the kpm virtual address vaddr.
4278  */
4279 pfn_t
4280 hat_kpm_va2pfn(caddr_t vaddr)
4281 {
4282 	pfn_t		pfn;
4283 
4284 	ASSERT(IS_KPM_ADDR(vaddr));
4285 
4286 	pfn = (pfn_t)btop(vaddr - kpm_vbase);
4287 
4288 	return (pfn);
4289 }
4290 
4291 
4292 /*
4293  * Return the page for the kpm virtual address vaddr.
4294  */
4295 page_t *
4296 hat_kpm_vaddr2page(caddr_t vaddr)
4297 {
4298 	pfn_t		pfn;
4299 
4300 	ASSERT(IS_KPM_ADDR(vaddr));
4301 
4302 	pfn = hat_kpm_va2pfn(vaddr);
4303 
4304 	return (page_numtopp_nolock(pfn));
4305 }
4306 
4307 /*
4308  * hat_kpm_fault is called from segkpm_fault when we take a page fault on a
4309  * KPM page.  This should never happen on x86
4310  */
4311 int
4312 hat_kpm_fault(hat_t *hat, caddr_t vaddr)
4313 {
4314 	panic("pagefault in seg_kpm.  hat: 0x%p  vaddr: 0x%p",
4315 	    (void *)hat, (void *)vaddr);
4316 
4317 	return (0);
4318 }
4319 
4320 /*ARGSUSED*/
4321 void
4322 hat_kpm_mseghash_clear(int nentries)
4323 {}
4324 
4325 /*ARGSUSED*/
4326 void
4327 hat_kpm_mseghash_update(pgcnt_t inx, struct memseg *msp)
4328 {}
4329 
4330 #ifndef	__xpv
4331 void
4332 hat_kpm_addmem_mseg_update(struct memseg *msp, pgcnt_t nkpmpgs,
4333 	offset_t kpm_pages_off)
4334 {
4335 	_NOTE(ARGUNUSED(nkpmpgs, kpm_pages_off));
4336 	pfn_t base, end;
4337 
4338 	/*
4339 	 * kphysm_add_memory_dynamic() does not set nkpmpgs
4340 	 * when page_t memory is externally allocated.  That
4341 	 * code must properly calculate nkpmpgs in all cases
4342 	 * if nkpmpgs needs to be used at some point.
4343 	 */
4344 
4345 	/*
4346 	 * The meta (page_t) pages for dynamically added memory are allocated
4347 	 * either from the incoming memory itself or from existing memory.
4348 	 * In the former case the base of the incoming pages will be different
4349 	 * than the base of the dynamic segment so call memseg_get_start() to
4350 	 * get the actual base of the incoming memory for each case.
4351 	 */
4352 
4353 	base = memseg_get_start(msp);
4354 	end = msp->pages_end;
4355 
4356 	hat_devload(kas.a_hat, kpm_vbase + mmu_ptob(base),
4357 	    mmu_ptob(end - base), base, PROT_READ | PROT_WRITE,
4358 	    HAT_LOAD | HAT_LOAD_LOCK | HAT_LOAD_NOCONSIST);
4359 }
4360 
4361 void
4362 hat_kpm_addmem_mseg_insert(struct memseg *msp)
4363 {
4364 	_NOTE(ARGUNUSED(msp));
4365 }
4366 
4367 void
4368 hat_kpm_addmem_memsegs_update(struct memseg *msp)
4369 {
4370 	_NOTE(ARGUNUSED(msp));
4371 }
4372 
4373 /*
4374  * Return end of metadata for an already setup memseg.
4375  * X86 platforms don't need per-page meta data to support kpm.
4376  */
4377 caddr_t
4378 hat_kpm_mseg_reuse(struct memseg *msp)
4379 {
4380 	return ((caddr_t)msp->epages);
4381 }
4382 
4383 void
4384 hat_kpm_delmem_mseg_update(struct memseg *msp, struct memseg **mspp)
4385 {
4386 	_NOTE(ARGUNUSED(msp, mspp));
4387 	ASSERT(0);
4388 }
4389 
4390 void
4391 hat_kpm_split_mseg_update(struct memseg *msp, struct memseg **mspp,
4392 	struct memseg *lo, struct memseg *mid, struct memseg *hi)
4393 {
4394 	_NOTE(ARGUNUSED(msp, mspp, lo, mid, hi));
4395 	ASSERT(0);
4396 }
4397 
4398 /*
4399  * Walk the memsegs chain, applying func to each memseg span.
4400  */
4401 void
4402 hat_kpm_walk(void (*func)(void *, void *, size_t), void *arg)
4403 {
4404 	pfn_t	pbase, pend;
4405 	void	*base;
4406 	size_t	size;
4407 	struct memseg *msp;
4408 
4409 	for (msp = memsegs; msp; msp = msp->next) {
4410 		pbase = msp->pages_base;
4411 		pend = msp->pages_end;
4412 		base = ptob(pbase) + kpm_vbase;
4413 		size = ptob(pend - pbase);
4414 		func(arg, base, size);
4415 	}
4416 }
4417 
4418 #else	/* __xpv */
4419 
4420 /*
4421  * There are specific Hypervisor calls to establish and remove mappings
4422  * to grant table references and the privcmd driver. We have to ensure
4423  * that a page table actually exists.
4424  */
4425 void
4426 hat_prepare_mapping(hat_t *hat, caddr_t addr, uint64_t *pte_ma)
4427 {
4428 	maddr_t base_ma;
4429 	htable_t *ht;
4430 	uint_t entry;
4431 
4432 	ASSERT(IS_P2ALIGNED((uintptr_t)addr, MMU_PAGESIZE));
4433 	XPV_DISALLOW_MIGRATE();
4434 	ht = htable_create(hat, (uintptr_t)addr, 0, NULL);
4435 
4436 	/*
4437 	 * if an address for pte_ma is passed in, return the MA of the pte
4438 	 * for this specific address.  This address is only valid as long
4439 	 * as the htable stays locked.
4440 	 */
4441 	if (pte_ma != NULL) {
4442 		entry = htable_va2entry((uintptr_t)addr, ht);
4443 		base_ma = pa_to_ma(ptob(ht->ht_pfn));
4444 		*pte_ma = base_ma + (entry << mmu.pte_size_shift);
4445 	}
4446 	XPV_ALLOW_MIGRATE();
4447 }
4448 
4449 void
4450 hat_release_mapping(hat_t *hat, caddr_t addr)
4451 {
4452 	htable_t *ht;
4453 
4454 	ASSERT(IS_P2ALIGNED((uintptr_t)addr, MMU_PAGESIZE));
4455 	XPV_DISALLOW_MIGRATE();
4456 	ht = htable_lookup(hat, (uintptr_t)addr, 0);
4457 	ASSERT(ht != NULL);
4458 	ASSERT(ht->ht_busy >= 2);
4459 	htable_release(ht);
4460 	htable_release(ht);
4461 	XPV_ALLOW_MIGRATE();
4462 }
4463 #endif	/* __xpv */
4464