xref: /titanic_41/usr/src/uts/i86pc/vm/hat_i86.c (revision cde2885fdf538266ee2a3b08dee2d5075ce8fa2b)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 
27 /*
28  * VM - Hardware Address Translation management for i386 and amd64
29  *
30  * Implementation of the interfaces described in <common/vm/hat.h>
31  *
32  * Nearly all the details of how the hardware is managed should not be
33  * visible outside this layer except for misc. machine specific functions
34  * that work in conjunction with this code.
35  *
36  * Routines used only inside of i86pc/vm start with hati_ for HAT Internal.
37  */
38 
39 #include <sys/machparam.h>
40 #include <sys/machsystm.h>
41 #include <sys/mman.h>
42 #include <sys/types.h>
43 #include <sys/systm.h>
44 #include <sys/cpuvar.h>
45 #include <sys/thread.h>
46 #include <sys/proc.h>
47 #include <sys/cpu.h>
48 #include <sys/kmem.h>
49 #include <sys/disp.h>
50 #include <sys/shm.h>
51 #include <sys/sysmacros.h>
52 #include <sys/machparam.h>
53 #include <sys/vmem.h>
54 #include <sys/vmsystm.h>
55 #include <sys/promif.h>
56 #include <sys/var.h>
57 #include <sys/x86_archext.h>
58 #include <sys/atomic.h>
59 #include <sys/bitmap.h>
60 #include <sys/controlregs.h>
61 #include <sys/bootconf.h>
62 #include <sys/bootsvcs.h>
63 #include <sys/bootinfo.h>
64 #include <sys/archsystm.h>
65 
66 #include <vm/seg_kmem.h>
67 #include <vm/hat_i86.h>
68 #include <vm/as.h>
69 #include <vm/seg.h>
70 #include <vm/page.h>
71 #include <vm/seg_kp.h>
72 #include <vm/seg_kpm.h>
73 #include <vm/vm_dep.h>
74 #ifdef __xpv
75 #include <sys/hypervisor.h>
76 #endif
77 #include <vm/kboot_mmu.h>
78 #include <vm/seg_spt.h>
79 
80 #include <sys/cmn_err.h>
81 
82 /*
83  * Basic parameters for hat operation.
84  */
85 struct hat_mmu_info mmu;
86 
87 /*
88  * The page that is the kernel's top level pagetable.
89  *
90  * For 32 bit PAE support on i86pc, the kernel hat will use the 1st 4 entries
91  * on this 4K page for its top level page table. The remaining groups of
92  * 4 entries are used for per processor copies of user VLP pagetables for
93  * running threads.  See hat_switch() and reload_pae32() for details.
94  *
95  * vlp_page[0..3] - level==2 PTEs for kernel HAT
96  * vlp_page[4..7] - level==2 PTEs for user thread on cpu 0
97  * vlp_page[8..11]  - level==2 PTE for user thread on cpu 1
98  * etc...
99  */
100 static x86pte_t *vlp_page;
101 
102 /*
103  * forward declaration of internal utility routines
104  */
105 static x86pte_t hati_update_pte(htable_t *ht, uint_t entry, x86pte_t expected,
106 	x86pte_t new);
107 
108 /*
109  * The kernel address space exists in all HATs. To implement this the
110  * kernel reserves a fixed number of entries in the topmost level(s) of page
111  * tables. The values are setup during startup and then copied to every user
112  * hat created by hat_alloc(). This means that kernelbase must be:
113  *
114  *	  4Meg aligned for 32 bit kernels
115  *	512Gig aligned for x86_64 64 bit kernel
116  *
117  * The hat_kernel_range_ts describe what needs to be copied from kernel hat
118  * to each user hat.
119  */
120 typedef struct hat_kernel_range {
121 	level_t		hkr_level;
122 	uintptr_t	hkr_start_va;
123 	uintptr_t	hkr_end_va;	/* zero means to end of memory */
124 } hat_kernel_range_t;
125 #define	NUM_KERNEL_RANGE 2
126 static hat_kernel_range_t kernel_ranges[NUM_KERNEL_RANGE];
127 static int num_kernel_ranges;
128 
129 uint_t use_boot_reserve = 1;	/* cleared after early boot process */
130 uint_t can_steal_post_boot = 0;	/* set late in boot to enable stealing */
131 
132 /*
133  * enable_1gpg: controls 1g page support for user applications.
134  * By default, 1g pages are exported to user applications. enable_1gpg can
135  * be set to 0 to not export.
136  */
137 int	enable_1gpg = 1;
138 
139 /*
140  * AMD shanghai processors provide better management of 1gb ptes in its tlb.
141  * By default, 1g page suppport will be disabled for pre-shanghai AMD
142  * processors that don't have optimal tlb support for the 1g page size.
143  * chk_optimal_1gtlb can be set to 0 to force 1g page support on sub-optimal
144  * processors.
145  */
146 int	chk_optimal_1gtlb = 1;
147 
148 
149 #ifdef DEBUG
150 uint_t	map1gcnt;
151 #endif
152 
153 
154 /*
155  * A cpuset for all cpus. This is used for kernel address cross calls, since
156  * the kernel addresses apply to all cpus.
157  */
158 cpuset_t khat_cpuset;
159 
160 /*
161  * management stuff for hat structures
162  */
163 kmutex_t	hat_list_lock;
164 kcondvar_t	hat_list_cv;
165 kmem_cache_t	*hat_cache;
166 kmem_cache_t	*hat_hash_cache;
167 kmem_cache_t	*vlp_hash_cache;
168 
169 /*
170  * Simple statistics
171  */
172 struct hatstats hatstat;
173 
174 /*
175  * Some earlier hypervisor versions do not emulate cmpxchg of PTEs
176  * correctly.  For such hypervisors we must set PT_USER for kernel
177  * entries ourselves (normally the emulation would set PT_USER for
178  * kernel entries and PT_USER|PT_GLOBAL for user entries).  pt_kern is
179  * thus set appropriately.  Note that dboot/kbm is OK, as only the full
180  * HAT uses cmpxchg() and the other paths (hypercall etc.) were never
181  * incorrect.
182  */
183 int pt_kern;
184 
185 /*
186  * useful stuff for atomic access/clearing/setting REF/MOD/RO bits in page_t's.
187  */
188 extern void atomic_orb(uchar_t *addr, uchar_t val);
189 extern void atomic_andb(uchar_t *addr, uchar_t val);
190 
191 #define	PP_GETRM(pp, rmmask)    (pp->p_nrm & rmmask)
192 #define	PP_ISMOD(pp)		PP_GETRM(pp, P_MOD)
193 #define	PP_ISREF(pp)		PP_GETRM(pp, P_REF)
194 #define	PP_ISRO(pp)		PP_GETRM(pp, P_RO)
195 
196 #define	PP_SETRM(pp, rm)	atomic_orb(&(pp->p_nrm), rm)
197 #define	PP_SETMOD(pp)		PP_SETRM(pp, P_MOD)
198 #define	PP_SETREF(pp)		PP_SETRM(pp, P_REF)
199 #define	PP_SETRO(pp)		PP_SETRM(pp, P_RO)
200 
201 #define	PP_CLRRM(pp, rm)	atomic_andb(&(pp->p_nrm), ~(rm))
202 #define	PP_CLRMOD(pp)   	PP_CLRRM(pp, P_MOD)
203 #define	PP_CLRREF(pp)   	PP_CLRRM(pp, P_REF)
204 #define	PP_CLRRO(pp)    	PP_CLRRM(pp, P_RO)
205 #define	PP_CLRALL(pp)		PP_CLRRM(pp, P_MOD | P_REF | P_RO)
206 
207 /*
208  * kmem cache constructor for struct hat
209  */
210 /*ARGSUSED*/
211 static int
212 hati_constructor(void *buf, void *handle, int kmflags)
213 {
214 	hat_t	*hat = buf;
215 
216 	mutex_init(&hat->hat_mutex, NULL, MUTEX_DEFAULT, NULL);
217 	bzero(hat->hat_pages_mapped,
218 	    sizeof (pgcnt_t) * (mmu.max_page_level + 1));
219 	hat->hat_ism_pgcnt = 0;
220 	hat->hat_stats = 0;
221 	hat->hat_flags = 0;
222 	CPUSET_ZERO(hat->hat_cpus);
223 	hat->hat_htable = NULL;
224 	hat->hat_ht_hash = NULL;
225 	return (0);
226 }
227 
228 /*
229  * Allocate a hat structure for as. We also create the top level
230  * htable and initialize it to contain the kernel hat entries.
231  */
232 hat_t *
233 hat_alloc(struct as *as)
234 {
235 	hat_t			*hat;
236 	htable_t		*ht;	/* top level htable */
237 	uint_t			use_vlp;
238 	uint_t			r;
239 	hat_kernel_range_t	*rp;
240 	uintptr_t		va;
241 	uintptr_t		eva;
242 	uint_t			start;
243 	uint_t			cnt;
244 	htable_t		*src;
245 
246 	/*
247 	 * Once we start creating user process HATs we can enable
248 	 * the htable_steal() code.
249 	 */
250 	if (can_steal_post_boot == 0)
251 		can_steal_post_boot = 1;
252 
253 	ASSERT(AS_WRITE_HELD(as, &as->a_lock));
254 	hat = kmem_cache_alloc(hat_cache, KM_SLEEP);
255 	hat->hat_as = as;
256 	mutex_init(&hat->hat_mutex, NULL, MUTEX_DEFAULT, NULL);
257 	ASSERT(hat->hat_flags == 0);
258 
259 #if defined(__xpv)
260 	/*
261 	 * No VLP stuff on the hypervisor due to the 64-bit split top level
262 	 * page tables.  On 32-bit it's not needed as the hypervisor takes
263 	 * care of copying the top level PTEs to a below 4Gig page.
264 	 */
265 	use_vlp = 0;
266 #else	/* __xpv */
267 	/* 32 bit processes uses a VLP style hat when running with PAE */
268 #if defined(__amd64)
269 	use_vlp = (ttoproc(curthread)->p_model == DATAMODEL_ILP32);
270 #elif defined(__i386)
271 	use_vlp = mmu.pae_hat;
272 #endif
273 #endif	/* __xpv */
274 	if (use_vlp) {
275 		hat->hat_flags = HAT_VLP;
276 		bzero(hat->hat_vlp_ptes, VLP_SIZE);
277 	}
278 
279 	/*
280 	 * Allocate the htable hash
281 	 */
282 	if ((hat->hat_flags & HAT_VLP)) {
283 		hat->hat_num_hash = mmu.vlp_hash_cnt;
284 		hat->hat_ht_hash = kmem_cache_alloc(vlp_hash_cache, KM_SLEEP);
285 	} else {
286 		hat->hat_num_hash = mmu.hash_cnt;
287 		hat->hat_ht_hash = kmem_cache_alloc(hat_hash_cache, KM_SLEEP);
288 	}
289 	bzero(hat->hat_ht_hash, hat->hat_num_hash * sizeof (htable_t *));
290 
291 	/*
292 	 * Initialize Kernel HAT entries at the top of the top level page
293 	 * tables for the new hat.
294 	 */
295 	hat->hat_htable = NULL;
296 	hat->hat_ht_cached = NULL;
297 	XPV_DISALLOW_MIGRATE();
298 	ht = htable_create(hat, (uintptr_t)0, TOP_LEVEL(hat), NULL);
299 	hat->hat_htable = ht;
300 
301 #if defined(__amd64)
302 	if (hat->hat_flags & HAT_VLP)
303 		goto init_done;
304 #endif
305 
306 	for (r = 0; r < num_kernel_ranges; ++r) {
307 		rp = &kernel_ranges[r];
308 		for (va = rp->hkr_start_va; va != rp->hkr_end_va;
309 		    va += cnt * LEVEL_SIZE(rp->hkr_level)) {
310 
311 			if (rp->hkr_level == TOP_LEVEL(hat))
312 				ht = hat->hat_htable;
313 			else
314 				ht = htable_create(hat, va, rp->hkr_level,
315 				    NULL);
316 
317 			start = htable_va2entry(va, ht);
318 			cnt = HTABLE_NUM_PTES(ht) - start;
319 			eva = va +
320 			    ((uintptr_t)cnt << LEVEL_SHIFT(rp->hkr_level));
321 			if (rp->hkr_end_va != 0 &&
322 			    (eva > rp->hkr_end_va || eva == 0))
323 				cnt = htable_va2entry(rp->hkr_end_va, ht) -
324 				    start;
325 
326 #if defined(__i386) && !defined(__xpv)
327 			if (ht->ht_flags & HTABLE_VLP) {
328 				bcopy(&vlp_page[start],
329 				    &hat->hat_vlp_ptes[start],
330 				    cnt * sizeof (x86pte_t));
331 				continue;
332 			}
333 #endif
334 			src = htable_lookup(kas.a_hat, va, rp->hkr_level);
335 			ASSERT(src != NULL);
336 			x86pte_copy(src, ht, start, cnt);
337 			htable_release(src);
338 		}
339 	}
340 
341 init_done:
342 
343 #if defined(__xpv)
344 	/*
345 	 * Pin top level page tables after initializing them
346 	 */
347 	xen_pin(hat->hat_htable->ht_pfn, mmu.max_level);
348 #if defined(__amd64)
349 	xen_pin(hat->hat_user_ptable, mmu.max_level);
350 #endif
351 #endif
352 	XPV_ALLOW_MIGRATE();
353 
354 	/*
355 	 * Put it at the start of the global list of all hats (used by stealing)
356 	 *
357 	 * kas.a_hat is not in the list but is instead used to find the
358 	 * first and last items in the list.
359 	 *
360 	 * - kas.a_hat->hat_next points to the start of the user hats.
361 	 *   The list ends where hat->hat_next == NULL
362 	 *
363 	 * - kas.a_hat->hat_prev points to the last of the user hats.
364 	 *   The list begins where hat->hat_prev == NULL
365 	 */
366 	mutex_enter(&hat_list_lock);
367 	hat->hat_prev = NULL;
368 	hat->hat_next = kas.a_hat->hat_next;
369 	if (hat->hat_next)
370 		hat->hat_next->hat_prev = hat;
371 	else
372 		kas.a_hat->hat_prev = hat;
373 	kas.a_hat->hat_next = hat;
374 	mutex_exit(&hat_list_lock);
375 
376 	return (hat);
377 }
378 
379 /*
380  * process has finished executing but as has not been cleaned up yet.
381  */
382 /*ARGSUSED*/
383 void
384 hat_free_start(hat_t *hat)
385 {
386 	ASSERT(AS_WRITE_HELD(hat->hat_as, &hat->hat_as->a_lock));
387 
388 	/*
389 	 * If the hat is currently a stealing victim, wait for the stealing
390 	 * to finish.  Once we mark it as HAT_FREEING, htable_steal()
391 	 * won't look at its pagetables anymore.
392 	 */
393 	mutex_enter(&hat_list_lock);
394 	while (hat->hat_flags & HAT_VICTIM)
395 		cv_wait(&hat_list_cv, &hat_list_lock);
396 	hat->hat_flags |= HAT_FREEING;
397 	mutex_exit(&hat_list_lock);
398 }
399 
400 /*
401  * An address space is being destroyed, so we destroy the associated hat.
402  */
403 void
404 hat_free_end(hat_t *hat)
405 {
406 	kmem_cache_t *cache;
407 
408 	ASSERT(hat->hat_flags & HAT_FREEING);
409 
410 	/*
411 	 * must not be running on the given hat
412 	 */
413 	ASSERT(CPU->cpu_current_hat != hat);
414 
415 	/*
416 	 * Remove it from the list of HATs
417 	 */
418 	mutex_enter(&hat_list_lock);
419 	if (hat->hat_prev)
420 		hat->hat_prev->hat_next = hat->hat_next;
421 	else
422 		kas.a_hat->hat_next = hat->hat_next;
423 	if (hat->hat_next)
424 		hat->hat_next->hat_prev = hat->hat_prev;
425 	else
426 		kas.a_hat->hat_prev = hat->hat_prev;
427 	mutex_exit(&hat_list_lock);
428 	hat->hat_next = hat->hat_prev = NULL;
429 
430 #if defined(__xpv)
431 	/*
432 	 * On the hypervisor, unpin top level page table(s)
433 	 */
434 	xen_unpin(hat->hat_htable->ht_pfn);
435 #if defined(__amd64)
436 	xen_unpin(hat->hat_user_ptable);
437 #endif
438 #endif
439 
440 	/*
441 	 * Make a pass through the htables freeing them all up.
442 	 */
443 	htable_purge_hat(hat);
444 
445 	/*
446 	 * Decide which kmem cache the hash table came from, then free it.
447 	 */
448 	if (hat->hat_flags & HAT_VLP)
449 		cache = vlp_hash_cache;
450 	else
451 		cache = hat_hash_cache;
452 	kmem_cache_free(cache, hat->hat_ht_hash);
453 	hat->hat_ht_hash = NULL;
454 
455 	hat->hat_flags = 0;
456 	kmem_cache_free(hat_cache, hat);
457 }
458 
459 /*
460  * round kernelbase down to a supported value to use for _userlimit
461  *
462  * userlimit must be aligned down to an entry in the top level htable.
463  * The one exception is for 32 bit HAT's running PAE.
464  */
465 uintptr_t
466 hat_kernelbase(uintptr_t va)
467 {
468 #if defined(__i386)
469 	va &= LEVEL_MASK(1);
470 #endif
471 	if (IN_VA_HOLE(va))
472 		panic("_userlimit %p will fall in VA hole\n", (void *)va);
473 	return (va);
474 }
475 
476 /*
477  *
478  */
479 static void
480 set_max_page_level()
481 {
482 	level_t lvl;
483 
484 	if (!kbm_largepage_support) {
485 		lvl = 0;
486 	} else {
487 		if (x86_feature & X86_1GPG) {
488 			lvl = 2;
489 			if (chk_optimal_1gtlb &&
490 			    cpuid_opteron_erratum(CPU, 6671130)) {
491 				lvl = 1;
492 			}
493 			if (plat_mnode_xcheck(LEVEL_SIZE(2) >>
494 			    LEVEL_SHIFT(0))) {
495 				lvl = 1;
496 			}
497 		} else {
498 			lvl = 1;
499 		}
500 	}
501 	mmu.max_page_level = lvl;
502 
503 	if ((lvl == 2) && (enable_1gpg == 0))
504 		mmu.umax_page_level = 1;
505 	else
506 		mmu.umax_page_level = lvl;
507 }
508 
509 /*
510  * Initialize hat data structures based on processor MMU information.
511  */
512 void
513 mmu_init(void)
514 {
515 	uint_t max_htables;
516 	uint_t pa_bits;
517 	uint_t va_bits;
518 	int i;
519 
520 	/*
521 	 * If CPU enabled the page table global bit, use it for the kernel
522 	 * This is bit 7 in CR4 (PGE - Page Global Enable).
523 	 */
524 	if ((x86_feature & X86_PGE) != 0 && (getcr4() & CR4_PGE) != 0)
525 		mmu.pt_global = PT_GLOBAL;
526 
527 	/*
528 	 * Detect NX and PAE usage.
529 	 */
530 	mmu.pae_hat = kbm_pae_support;
531 	if (kbm_nx_support)
532 		mmu.pt_nx = PT_NX;
533 	else
534 		mmu.pt_nx = 0;
535 
536 	/*
537 	 * Use CPU info to set various MMU parameters
538 	 */
539 	cpuid_get_addrsize(CPU, &pa_bits, &va_bits);
540 
541 	if (va_bits < sizeof (void *) * NBBY) {
542 		mmu.hole_start = (1ul << (va_bits - 1));
543 		mmu.hole_end = 0ul - mmu.hole_start - 1;
544 	} else {
545 		mmu.hole_end = 0;
546 		mmu.hole_start = mmu.hole_end - 1;
547 	}
548 #if defined(OPTERON_ERRATUM_121)
549 	/*
550 	 * If erratum 121 has already been detected at this time, hole_start
551 	 * contains the value to be subtracted from mmu.hole_start.
552 	 */
553 	ASSERT(hole_start == 0 || opteron_erratum_121 != 0);
554 	hole_start = mmu.hole_start - hole_start;
555 #else
556 	hole_start = mmu.hole_start;
557 #endif
558 	hole_end = mmu.hole_end;
559 
560 	mmu.highest_pfn = mmu_btop((1ull << pa_bits) - 1);
561 	if (mmu.pae_hat == 0 && pa_bits > 32)
562 		mmu.highest_pfn = PFN_4G - 1;
563 
564 	if (mmu.pae_hat) {
565 		mmu.pte_size = 8;	/* 8 byte PTEs */
566 		mmu.pte_size_shift = 3;
567 	} else {
568 		mmu.pte_size = 4;	/* 4 byte PTEs */
569 		mmu.pte_size_shift = 2;
570 	}
571 
572 	if (mmu.pae_hat && (x86_feature & X86_PAE) == 0)
573 		panic("Processor does not support PAE");
574 
575 	if ((x86_feature & X86_CX8) == 0)
576 		panic("Processor does not support cmpxchg8b instruction");
577 
578 #if defined(__amd64)
579 
580 	mmu.num_level = 4;
581 	mmu.max_level = 3;
582 	mmu.ptes_per_table = 512;
583 	mmu.top_level_count = 512;
584 
585 	mmu.level_shift[0] = 12;
586 	mmu.level_shift[1] = 21;
587 	mmu.level_shift[2] = 30;
588 	mmu.level_shift[3] = 39;
589 
590 #elif defined(__i386)
591 
592 	if (mmu.pae_hat) {
593 		mmu.num_level = 3;
594 		mmu.max_level = 2;
595 		mmu.ptes_per_table = 512;
596 		mmu.top_level_count = 4;
597 
598 		mmu.level_shift[0] = 12;
599 		mmu.level_shift[1] = 21;
600 		mmu.level_shift[2] = 30;
601 
602 	} else {
603 		mmu.num_level = 2;
604 		mmu.max_level = 1;
605 		mmu.ptes_per_table = 1024;
606 		mmu.top_level_count = 1024;
607 
608 		mmu.level_shift[0] = 12;
609 		mmu.level_shift[1] = 22;
610 	}
611 
612 #endif	/* __i386 */
613 
614 	for (i = 0; i < mmu.num_level; ++i) {
615 		mmu.level_size[i] = 1UL << mmu.level_shift[i];
616 		mmu.level_offset[i] = mmu.level_size[i] - 1;
617 		mmu.level_mask[i] = ~mmu.level_offset[i];
618 	}
619 
620 	set_max_page_level();
621 
622 	mmu_page_sizes = mmu.max_page_level + 1;
623 	mmu_exported_page_sizes = mmu.umax_page_level + 1;
624 
625 	/* restrict legacy applications from using pagesizes 1g and above */
626 	mmu_legacy_page_sizes =
627 	    (mmu_exported_page_sizes > 2) ? 2 : mmu_exported_page_sizes;
628 
629 
630 	for (i = 0; i <= mmu.max_page_level; ++i) {
631 		mmu.pte_bits[i] = PT_VALID | pt_kern;
632 		if (i > 0)
633 			mmu.pte_bits[i] |= PT_PAGESIZE;
634 	}
635 
636 	/*
637 	 * NOTE Legacy 32 bit PAE mode only has the P_VALID bit at top level.
638 	 */
639 	for (i = 1; i < mmu.num_level; ++i)
640 		mmu.ptp_bits[i] = PT_PTPBITS;
641 
642 #if defined(__i386)
643 	mmu.ptp_bits[2] = PT_VALID;
644 #endif
645 
646 	/*
647 	 * Compute how many hash table entries to have per process for htables.
648 	 * We start with 1 page's worth of entries.
649 	 *
650 	 * If physical memory is small, reduce the amount need to cover it.
651 	 */
652 	max_htables = physmax / mmu.ptes_per_table;
653 	mmu.hash_cnt = MMU_PAGESIZE / sizeof (htable_t *);
654 	while (mmu.hash_cnt > 16 && mmu.hash_cnt >= max_htables)
655 		mmu.hash_cnt >>= 1;
656 	mmu.vlp_hash_cnt = mmu.hash_cnt;
657 
658 #if defined(__amd64)
659 	/*
660 	 * If running in 64 bits and physical memory is large,
661 	 * increase the size of the cache to cover all of memory for
662 	 * a 64 bit process.
663 	 */
664 #define	HASH_MAX_LENGTH 4
665 	while (mmu.hash_cnt * HASH_MAX_LENGTH < max_htables)
666 		mmu.hash_cnt <<= 1;
667 #endif
668 }
669 
670 
671 /*
672  * initialize hat data structures
673  */
674 void
675 hat_init()
676 {
677 #if defined(__i386)
678 	/*
679 	 * _userlimit must be aligned correctly
680 	 */
681 	if ((_userlimit & LEVEL_MASK(1)) != _userlimit) {
682 		prom_printf("hat_init(): _userlimit=%p, not aligned at %p\n",
683 		    (void *)_userlimit, (void *)LEVEL_SIZE(1));
684 		halt("hat_init(): Unable to continue");
685 	}
686 #endif
687 
688 	cv_init(&hat_list_cv, NULL, CV_DEFAULT, NULL);
689 
690 	/*
691 	 * initialize kmem caches
692 	 */
693 	htable_init();
694 	hment_init();
695 
696 	hat_cache = kmem_cache_create("hat_t",
697 	    sizeof (hat_t), 0, hati_constructor, NULL, NULL,
698 	    NULL, 0, 0);
699 
700 	hat_hash_cache = kmem_cache_create("HatHash",
701 	    mmu.hash_cnt * sizeof (htable_t *), 0, NULL, NULL, NULL,
702 	    NULL, 0, 0);
703 
704 	/*
705 	 * VLP hats can use a smaller hash table size on large memroy machines
706 	 */
707 	if (mmu.hash_cnt == mmu.vlp_hash_cnt) {
708 		vlp_hash_cache = hat_hash_cache;
709 	} else {
710 		vlp_hash_cache = kmem_cache_create("HatVlpHash",
711 		    mmu.vlp_hash_cnt * sizeof (htable_t *), 0, NULL, NULL, NULL,
712 		    NULL, 0, 0);
713 	}
714 
715 	/*
716 	 * Set up the kernel's hat
717 	 */
718 	AS_LOCK_ENTER(&kas, &kas.a_lock, RW_WRITER);
719 	kas.a_hat = kmem_cache_alloc(hat_cache, KM_NOSLEEP);
720 	mutex_init(&kas.a_hat->hat_mutex, NULL, MUTEX_DEFAULT, NULL);
721 	kas.a_hat->hat_as = &kas;
722 	kas.a_hat->hat_flags = 0;
723 	AS_LOCK_EXIT(&kas, &kas.a_lock);
724 
725 	CPUSET_ZERO(khat_cpuset);
726 	CPUSET_ADD(khat_cpuset, CPU->cpu_id);
727 
728 	/*
729 	 * The kernel hat's next pointer serves as the head of the hat list .
730 	 * The kernel hat's prev pointer tracks the last hat on the list for
731 	 * htable_steal() to use.
732 	 */
733 	kas.a_hat->hat_next = NULL;
734 	kas.a_hat->hat_prev = NULL;
735 
736 	/*
737 	 * Allocate an htable hash bucket for the kernel
738 	 * XX64 - tune for 64 bit procs
739 	 */
740 	kas.a_hat->hat_num_hash = mmu.hash_cnt;
741 	kas.a_hat->hat_ht_hash = kmem_cache_alloc(hat_hash_cache, KM_NOSLEEP);
742 	bzero(kas.a_hat->hat_ht_hash, mmu.hash_cnt * sizeof (htable_t *));
743 
744 	/*
745 	 * zero out the top level and cached htable pointers
746 	 */
747 	kas.a_hat->hat_ht_cached = NULL;
748 	kas.a_hat->hat_htable = NULL;
749 
750 	/*
751 	 * Pre-allocate hrm_hashtab before enabling the collection of
752 	 * refmod statistics.  Allocating on the fly would mean us
753 	 * running the risk of suffering recursive mutex enters or
754 	 * deadlocks.
755 	 */
756 	hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *),
757 	    KM_SLEEP);
758 }
759 
760 /*
761  * Prepare CPU specific pagetables for VLP processes on 64 bit kernels.
762  *
763  * Each CPU has a set of 2 pagetables that are reused for any 32 bit
764  * process it runs. They are the top level pagetable, hci_vlp_l3ptes, and
765  * the next to top level table for the bottom 512 Gig, hci_vlp_l2ptes.
766  */
767 /*ARGSUSED*/
768 static void
769 hat_vlp_setup(struct cpu *cpu)
770 {
771 #if defined(__amd64) && !defined(__xpv)
772 	struct hat_cpu_info *hci = cpu->cpu_hat_info;
773 	pfn_t pfn;
774 
775 	/*
776 	 * allocate the level==2 page table for the bottom most
777 	 * 512Gig of address space (this is where 32 bit apps live)
778 	 */
779 	ASSERT(hci != NULL);
780 	hci->hci_vlp_l2ptes = kmem_zalloc(MMU_PAGESIZE, KM_SLEEP);
781 
782 	/*
783 	 * Allocate a top level pagetable and copy the kernel's
784 	 * entries into it. Then link in hci_vlp_l2ptes in the 1st entry.
785 	 */
786 	hci->hci_vlp_l3ptes = kmem_zalloc(MMU_PAGESIZE, KM_SLEEP);
787 	hci->hci_vlp_pfn =
788 	    hat_getpfnum(kas.a_hat, (caddr_t)hci->hci_vlp_l3ptes);
789 	ASSERT(hci->hci_vlp_pfn != PFN_INVALID);
790 	bcopy(vlp_page, hci->hci_vlp_l3ptes, MMU_PAGESIZE);
791 
792 	pfn = hat_getpfnum(kas.a_hat, (caddr_t)hci->hci_vlp_l2ptes);
793 	ASSERT(pfn != PFN_INVALID);
794 	hci->hci_vlp_l3ptes[0] = MAKEPTP(pfn, 2);
795 #endif /* __amd64 && !__xpv */
796 }
797 
798 /*ARGSUSED*/
799 static void
800 hat_vlp_teardown(cpu_t *cpu)
801 {
802 #if defined(__amd64) && !defined(__xpv)
803 	struct hat_cpu_info *hci;
804 
805 	if ((hci = cpu->cpu_hat_info) == NULL)
806 		return;
807 	if (hci->hci_vlp_l2ptes)
808 		kmem_free(hci->hci_vlp_l2ptes, MMU_PAGESIZE);
809 	if (hci->hci_vlp_l3ptes)
810 		kmem_free(hci->hci_vlp_l3ptes, MMU_PAGESIZE);
811 #endif
812 }
813 
814 #define	NEXT_HKR(r, l, s, e) {			\
815 	kernel_ranges[r].hkr_level = l;		\
816 	kernel_ranges[r].hkr_start_va = s;	\
817 	kernel_ranges[r].hkr_end_va = e;	\
818 	++r;					\
819 }
820 
821 /*
822  * Finish filling in the kernel hat.
823  * Pre fill in all top level kernel page table entries for the kernel's
824  * part of the address range.  From this point on we can't use any new
825  * kernel large pages if they need PTE's at max_level
826  *
827  * create the kmap mappings.
828  */
829 void
830 hat_init_finish(void)
831 {
832 	size_t		size;
833 	uint_t		r = 0;
834 	uintptr_t	va;
835 	hat_kernel_range_t *rp;
836 
837 
838 	/*
839 	 * We are now effectively running on the kernel hat.
840 	 * Clearing use_boot_reserve shuts off using the pre-allocated boot
841 	 * reserve for all HAT allocations.  From here on, the reserves are
842 	 * only used when avoiding recursion in kmem_alloc().
843 	 */
844 	use_boot_reserve = 0;
845 	htable_adjust_reserve();
846 
847 	/*
848 	 * User HATs are initialized with copies of all kernel mappings in
849 	 * higher level page tables. Ensure that those entries exist.
850 	 */
851 #if defined(__amd64)
852 
853 	NEXT_HKR(r, 3, kernelbase, 0);
854 #if defined(__xpv)
855 	NEXT_HKR(r, 3, HYPERVISOR_VIRT_START, HYPERVISOR_VIRT_END);
856 #endif
857 
858 #elif defined(__i386)
859 
860 #if !defined(__xpv)
861 	if (mmu.pae_hat) {
862 		va = kernelbase;
863 		if ((va & LEVEL_MASK(2)) != va) {
864 			va = P2ROUNDUP(va, LEVEL_SIZE(2));
865 			NEXT_HKR(r, 1, kernelbase, va);
866 		}
867 		if (va != 0)
868 			NEXT_HKR(r, 2, va, 0);
869 	} else
870 #endif /* __xpv */
871 		NEXT_HKR(r, 1, kernelbase, 0);
872 
873 #endif /* __i386 */
874 
875 	num_kernel_ranges = r;
876 
877 	/*
878 	 * Create all the kernel pagetables that will have entries
879 	 * shared to user HATs.
880 	 */
881 	for (r = 0; r < num_kernel_ranges; ++r) {
882 		rp = &kernel_ranges[r];
883 		for (va = rp->hkr_start_va; va != rp->hkr_end_va;
884 		    va += LEVEL_SIZE(rp->hkr_level)) {
885 			htable_t *ht;
886 
887 			if (IN_HYPERVISOR_VA(va))
888 				continue;
889 
890 			/* can/must skip if a page mapping already exists */
891 			if (rp->hkr_level <= mmu.max_page_level &&
892 			    (ht = htable_getpage(kas.a_hat, va, NULL)) !=
893 			    NULL) {
894 				htable_release(ht);
895 				continue;
896 			}
897 
898 			(void) htable_create(kas.a_hat, va, rp->hkr_level - 1,
899 			    NULL);
900 		}
901 	}
902 
903 	/*
904 	 * 32 bit PAE metal kernels use only 4 of the 512 entries in the
905 	 * page holding the top level pagetable. We use the remainder for
906 	 * the "per CPU" page tables for VLP processes.
907 	 * Map the top level kernel pagetable into the kernel to make
908 	 * it easy to use bcopy access these tables.
909 	 */
910 	if (mmu.pae_hat) {
911 		vlp_page = vmem_alloc(heap_arena, MMU_PAGESIZE, VM_SLEEP);
912 		hat_devload(kas.a_hat, (caddr_t)vlp_page, MMU_PAGESIZE,
913 		    kas.a_hat->hat_htable->ht_pfn,
914 #if !defined(__xpv)
915 		    PROT_WRITE |
916 #endif
917 		    PROT_READ | HAT_NOSYNC | HAT_UNORDERED_OK,
918 		    HAT_LOAD | HAT_LOAD_NOCONSIST);
919 	}
920 	hat_vlp_setup(CPU);
921 
922 	/*
923 	 * Create kmap (cached mappings of kernel PTEs)
924 	 * for 32 bit we map from segmap_start .. ekernelheap
925 	 * for 64 bit we map from segmap_start .. segmap_start + segmapsize;
926 	 */
927 #if defined(__i386)
928 	size = (uintptr_t)ekernelheap - segmap_start;
929 #elif defined(__amd64)
930 	size = segmapsize;
931 #endif
932 	hat_kmap_init((uintptr_t)segmap_start, size);
933 }
934 
935 /*
936  * On 32 bit PAE mode, PTE's are 64 bits, but ordinary atomic memory references
937  * are 32 bit, so for safety we must use cas64() to install these.
938  */
939 #ifdef __i386
940 static void
941 reload_pae32(hat_t *hat, cpu_t *cpu)
942 {
943 	x86pte_t *src;
944 	x86pte_t *dest;
945 	x86pte_t pte;
946 	int i;
947 
948 	/*
949 	 * Load the 4 entries of the level 2 page table into this
950 	 * cpu's range of the vlp_page and point cr3 at them.
951 	 */
952 	ASSERT(mmu.pae_hat);
953 	src = hat->hat_vlp_ptes;
954 	dest = vlp_page + (cpu->cpu_id + 1) * VLP_NUM_PTES;
955 	for (i = 0; i < VLP_NUM_PTES; ++i) {
956 		for (;;) {
957 			pte = dest[i];
958 			if (pte == src[i])
959 				break;
960 			if (cas64(dest + i, pte, src[i]) != src[i])
961 				break;
962 		}
963 	}
964 }
965 #endif
966 
967 /*
968  * Switch to a new active hat, maintaining bit masks to track active CPUs.
969  *
970  * On the 32-bit PAE hypervisor, %cr3 is a 64-bit value, on metal it
971  * remains a 32-bit value.
972  */
973 void
974 hat_switch(hat_t *hat)
975 {
976 	uint64_t	newcr3;
977 	cpu_t		*cpu = CPU;
978 	hat_t		*old = cpu->cpu_current_hat;
979 
980 	/*
981 	 * set up this information first, so we don't miss any cross calls
982 	 */
983 	if (old != NULL) {
984 		if (old == hat)
985 			return;
986 		if (old != kas.a_hat)
987 			CPUSET_ATOMIC_DEL(old->hat_cpus, cpu->cpu_id);
988 	}
989 
990 	/*
991 	 * Add this CPU to the active set for this HAT.
992 	 */
993 	if (hat != kas.a_hat) {
994 		CPUSET_ATOMIC_ADD(hat->hat_cpus, cpu->cpu_id);
995 	}
996 	cpu->cpu_current_hat = hat;
997 
998 	/*
999 	 * now go ahead and load cr3
1000 	 */
1001 	if (hat->hat_flags & HAT_VLP) {
1002 #if defined(__amd64)
1003 		x86pte_t *vlpptep = cpu->cpu_hat_info->hci_vlp_l2ptes;
1004 
1005 		VLP_COPY(hat->hat_vlp_ptes, vlpptep);
1006 		newcr3 = MAKECR3(cpu->cpu_hat_info->hci_vlp_pfn);
1007 #elif defined(__i386)
1008 		reload_pae32(hat, cpu);
1009 		newcr3 = MAKECR3(kas.a_hat->hat_htable->ht_pfn) +
1010 		    (cpu->cpu_id + 1) * VLP_SIZE;
1011 #endif
1012 	} else {
1013 		newcr3 = MAKECR3((uint64_t)hat->hat_htable->ht_pfn);
1014 	}
1015 #ifdef __xpv
1016 	{
1017 		struct mmuext_op t[2];
1018 		uint_t retcnt;
1019 		uint_t opcnt = 1;
1020 
1021 		t[0].cmd = MMUEXT_NEW_BASEPTR;
1022 		t[0].arg1.mfn = mmu_btop(pa_to_ma(newcr3));
1023 #if defined(__amd64)
1024 		/*
1025 		 * There's an interesting problem here, as to what to
1026 		 * actually specify when switching to the kernel hat.
1027 		 * For now we'll reuse the kernel hat again.
1028 		 */
1029 		t[1].cmd = MMUEXT_NEW_USER_BASEPTR;
1030 		if (hat == kas.a_hat)
1031 			t[1].arg1.mfn = mmu_btop(pa_to_ma(newcr3));
1032 		else
1033 			t[1].arg1.mfn = pfn_to_mfn(hat->hat_user_ptable);
1034 		++opcnt;
1035 #endif	/* __amd64 */
1036 		if (HYPERVISOR_mmuext_op(t, opcnt, &retcnt, DOMID_SELF) < 0)
1037 			panic("HYPERVISOR_mmu_update() failed");
1038 		ASSERT(retcnt == opcnt);
1039 
1040 	}
1041 #else
1042 	setcr3(newcr3);
1043 #endif
1044 	ASSERT(cpu == CPU);
1045 }
1046 
1047 /*
1048  * Utility to return a valid x86pte_t from protections, pfn, and level number
1049  */
1050 static x86pte_t
1051 hati_mkpte(pfn_t pfn, uint_t attr, level_t level, uint_t flags)
1052 {
1053 	x86pte_t	pte;
1054 	uint_t		cache_attr = attr & HAT_ORDER_MASK;
1055 
1056 	pte = MAKEPTE(pfn, level);
1057 
1058 	if (attr & PROT_WRITE)
1059 		PTE_SET(pte, PT_WRITABLE);
1060 
1061 	if (attr & PROT_USER)
1062 		PTE_SET(pte, PT_USER);
1063 
1064 	if (!(attr & PROT_EXEC))
1065 		PTE_SET(pte, mmu.pt_nx);
1066 
1067 	/*
1068 	 * Set the software bits used track ref/mod sync's and hments.
1069 	 * If not using REF/MOD, set them to avoid h/w rewriting PTEs.
1070 	 */
1071 	if (flags & HAT_LOAD_NOCONSIST)
1072 		PTE_SET(pte, PT_NOCONSIST | PT_REF | PT_MOD);
1073 	else if (attr & HAT_NOSYNC)
1074 		PTE_SET(pte, PT_NOSYNC | PT_REF | PT_MOD);
1075 
1076 	/*
1077 	 * Set the caching attributes in the PTE. The combination
1078 	 * of attributes are poorly defined, so we pay attention
1079 	 * to them in the given order.
1080 	 *
1081 	 * The test for HAT_STRICTORDER is different because it's defined
1082 	 * as "0" - which was a stupid thing to do, but is too late to change!
1083 	 */
1084 	if (cache_attr == HAT_STRICTORDER) {
1085 		PTE_SET(pte, PT_NOCACHE);
1086 	/*LINTED [Lint hates empty ifs, but it's the obvious way to do this] */
1087 	} else if (cache_attr & (HAT_UNORDERED_OK | HAT_STORECACHING_OK)) {
1088 		/* nothing to set */;
1089 	} else if (cache_attr & (HAT_MERGING_OK | HAT_LOADCACHING_OK)) {
1090 		PTE_SET(pte, PT_NOCACHE);
1091 		if (x86_feature & X86_PAT)
1092 			PTE_SET(pte, (level == 0) ? PT_PAT_4K : PT_PAT_LARGE);
1093 		else
1094 			PTE_SET(pte, PT_WRITETHRU);
1095 	} else {
1096 		panic("hati_mkpte(): bad caching attributes: %x\n", cache_attr);
1097 	}
1098 
1099 	return (pte);
1100 }
1101 
1102 /*
1103  * Duplicate address translations of the parent to the child.
1104  * This function really isn't used anymore.
1105  */
1106 /*ARGSUSED*/
1107 int
1108 hat_dup(hat_t *old, hat_t *new, caddr_t addr, size_t len, uint_t flag)
1109 {
1110 	ASSERT((uintptr_t)addr < kernelbase);
1111 	ASSERT(new != kas.a_hat);
1112 	ASSERT(old != kas.a_hat);
1113 	return (0);
1114 }
1115 
1116 /*
1117  * Allocate any hat resources required for a process being swapped in.
1118  */
1119 /*ARGSUSED*/
1120 void
1121 hat_swapin(hat_t *hat)
1122 {
1123 	/* do nothing - we let everything fault back in */
1124 }
1125 
1126 /*
1127  * Unload all translations associated with an address space of a process
1128  * that is being swapped out.
1129  */
1130 void
1131 hat_swapout(hat_t *hat)
1132 {
1133 	uintptr_t	vaddr = (uintptr_t)0;
1134 	uintptr_t	eaddr = _userlimit;
1135 	htable_t	*ht = NULL;
1136 	level_t		l;
1137 
1138 	XPV_DISALLOW_MIGRATE();
1139 	/*
1140 	 * We can't just call hat_unload(hat, 0, _userlimit...)  here, because
1141 	 * seg_spt and shared pagetables can't be swapped out.
1142 	 * Take a look at segspt_shmswapout() - it's a big no-op.
1143 	 *
1144 	 * Instead we'll walk through all the address space and unload
1145 	 * any mappings which we are sure are not shared, not locked.
1146 	 */
1147 	ASSERT(IS_PAGEALIGNED(vaddr));
1148 	ASSERT(IS_PAGEALIGNED(eaddr));
1149 	ASSERT(AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1150 	if ((uintptr_t)hat->hat_as->a_userlimit < eaddr)
1151 		eaddr = (uintptr_t)hat->hat_as->a_userlimit;
1152 
1153 	while (vaddr < eaddr) {
1154 		(void) htable_walk(hat, &ht, &vaddr, eaddr);
1155 		if (ht == NULL)
1156 			break;
1157 
1158 		ASSERT(!IN_VA_HOLE(vaddr));
1159 
1160 		/*
1161 		 * If the page table is shared skip its entire range.
1162 		 */
1163 		l = ht->ht_level;
1164 		if (ht->ht_flags & HTABLE_SHARED_PFN) {
1165 			vaddr = ht->ht_vaddr + LEVEL_SIZE(l + 1);
1166 			htable_release(ht);
1167 			ht = NULL;
1168 			continue;
1169 		}
1170 
1171 		/*
1172 		 * If the page table has no locked entries, unload this one.
1173 		 */
1174 		if (ht->ht_lock_cnt == 0)
1175 			hat_unload(hat, (caddr_t)vaddr, LEVEL_SIZE(l),
1176 			    HAT_UNLOAD_UNMAP);
1177 
1178 		/*
1179 		 * If we have a level 0 page table with locked entries,
1180 		 * skip the entire page table, otherwise skip just one entry.
1181 		 */
1182 		if (ht->ht_lock_cnt > 0 && l == 0)
1183 			vaddr = ht->ht_vaddr + LEVEL_SIZE(1);
1184 		else
1185 			vaddr += LEVEL_SIZE(l);
1186 	}
1187 	if (ht)
1188 		htable_release(ht);
1189 
1190 	/*
1191 	 * We're in swapout because the system is low on memory, so
1192 	 * go back and flush all the htables off the cached list.
1193 	 */
1194 	htable_purge_hat(hat);
1195 	XPV_ALLOW_MIGRATE();
1196 }
1197 
1198 /*
1199  * returns number of bytes that have valid mappings in hat.
1200  */
1201 size_t
1202 hat_get_mapped_size(hat_t *hat)
1203 {
1204 	size_t total = 0;
1205 	int l;
1206 
1207 	for (l = 0; l <= mmu.max_page_level; l++)
1208 		total += (hat->hat_pages_mapped[l] << LEVEL_SHIFT(l));
1209 	total += hat->hat_ism_pgcnt;
1210 
1211 	return (total);
1212 }
1213 
1214 /*
1215  * enable/disable collection of stats for hat.
1216  */
1217 int
1218 hat_stats_enable(hat_t *hat)
1219 {
1220 	atomic_add_32(&hat->hat_stats, 1);
1221 	return (1);
1222 }
1223 
1224 void
1225 hat_stats_disable(hat_t *hat)
1226 {
1227 	atomic_add_32(&hat->hat_stats, -1);
1228 }
1229 
1230 /*
1231  * Utility to sync the ref/mod bits from a page table entry to the page_t
1232  * We must be holding the mapping list lock when this is called.
1233  */
1234 static void
1235 hati_sync_pte_to_page(page_t *pp, x86pte_t pte, level_t level)
1236 {
1237 	uint_t	rm = 0;
1238 	pgcnt_t	pgcnt;
1239 
1240 	if (PTE_GET(pte, PT_SOFTWARE) >= PT_NOSYNC)
1241 		return;
1242 
1243 	if (PTE_GET(pte, PT_REF))
1244 		rm |= P_REF;
1245 
1246 	if (PTE_GET(pte, PT_MOD))
1247 		rm |= P_MOD;
1248 
1249 	if (rm == 0)
1250 		return;
1251 
1252 	/*
1253 	 * sync to all constituent pages of a large page
1254 	 */
1255 	ASSERT(x86_hm_held(pp));
1256 	pgcnt = page_get_pagecnt(level);
1257 	ASSERT(IS_P2ALIGNED(pp->p_pagenum, pgcnt));
1258 	for (; pgcnt > 0; --pgcnt) {
1259 		/*
1260 		 * hat_page_demote() can't decrease
1261 		 * pszc below this mapping size
1262 		 * since this large mapping existed after we
1263 		 * took mlist lock.
1264 		 */
1265 		ASSERT(pp->p_szc >= level);
1266 		hat_page_setattr(pp, rm);
1267 		++pp;
1268 	}
1269 }
1270 
1271 /*
1272  * This the set of PTE bits for PFN, permissions and caching
1273  * that are allowed to change on a HAT_LOAD_REMAP
1274  */
1275 #define	PT_REMAP_BITS							\
1276 	(PT_PADDR | PT_NX | PT_WRITABLE | PT_WRITETHRU |		\
1277 	PT_NOCACHE | PT_PAT_4K | PT_PAT_LARGE | PT_IGNORE | PT_REF | PT_MOD)
1278 
1279 #define	REMAPASSERT(EX)	if (!(EX)) panic("hati_pte_map: " #EX)
1280 /*
1281  * Do the low-level work to get a mapping entered into a HAT's pagetables
1282  * and in the mapping list of the associated page_t.
1283  */
1284 static int
1285 hati_pte_map(
1286 	htable_t	*ht,
1287 	uint_t		entry,
1288 	page_t		*pp,
1289 	x86pte_t	pte,
1290 	int		flags,
1291 	void		*pte_ptr)
1292 {
1293 	hat_t		*hat = ht->ht_hat;
1294 	x86pte_t	old_pte;
1295 	level_t		l = ht->ht_level;
1296 	hment_t		*hm;
1297 	uint_t		is_consist;
1298 	uint_t		is_locked;
1299 	int		rv = 0;
1300 
1301 	/*
1302 	 * Is this a consistant (ie. need mapping list lock) mapping?
1303 	 */
1304 	is_consist = (pp != NULL && (flags & HAT_LOAD_NOCONSIST) == 0);
1305 
1306 	/*
1307 	 * Track locked mapping count in the htable.  Do this first,
1308 	 * as we track locking even if there already is a mapping present.
1309 	 */
1310 	is_locked = (flags & HAT_LOAD_LOCK) != 0 && hat != kas.a_hat;
1311 	if (is_locked)
1312 		HTABLE_LOCK_INC(ht);
1313 
1314 	/*
1315 	 * Acquire the page's mapping list lock and get an hment to use.
1316 	 * Note that hment_prepare() might return NULL.
1317 	 */
1318 	if (is_consist) {
1319 		x86_hm_enter(pp);
1320 		hm = hment_prepare(ht, entry, pp);
1321 	}
1322 
1323 	/*
1324 	 * Set the new pte, retrieving the old one at the same time.
1325 	 */
1326 	old_pte = x86pte_set(ht, entry, pte, pte_ptr);
1327 
1328 	/*
1329 	 * Did we get a large page / page table collision?
1330 	 */
1331 	if (old_pte == LPAGE_ERROR) {
1332 		if (is_locked)
1333 			HTABLE_LOCK_DEC(ht);
1334 		rv = -1;
1335 		goto done;
1336 	}
1337 
1338 	/*
1339 	 * If the mapping didn't change there is nothing more to do.
1340 	 */
1341 	if (PTE_EQUIV(pte, old_pte))
1342 		goto done;
1343 
1344 	/*
1345 	 * Install a new mapping in the page's mapping list
1346 	 */
1347 	if (!PTE_ISVALID(old_pte)) {
1348 		if (is_consist) {
1349 			hment_assign(ht, entry, pp, hm);
1350 			x86_hm_exit(pp);
1351 		} else {
1352 			ASSERT(flags & HAT_LOAD_NOCONSIST);
1353 		}
1354 #if defined(__amd64)
1355 		if (ht->ht_flags & HTABLE_VLP) {
1356 			cpu_t *cpu = CPU;
1357 			x86pte_t *vlpptep = cpu->cpu_hat_info->hci_vlp_l2ptes;
1358 			VLP_COPY(hat->hat_vlp_ptes, vlpptep);
1359 		}
1360 #endif
1361 		HTABLE_INC(ht->ht_valid_cnt);
1362 		PGCNT_INC(hat, l);
1363 		return (rv);
1364 	}
1365 
1366 	/*
1367 	 * Remap's are more complicated:
1368 	 *  - HAT_LOAD_REMAP must be specified if changing the pfn.
1369 	 *    We also require that NOCONSIST be specified.
1370 	 *  - Otherwise only permission or caching bits may change.
1371 	 */
1372 	if (!PTE_ISPAGE(old_pte, l))
1373 		panic("non-null/page mapping pte=" FMT_PTE, old_pte);
1374 
1375 	if (PTE2PFN(old_pte, l) != PTE2PFN(pte, l)) {
1376 		REMAPASSERT(flags & HAT_LOAD_REMAP);
1377 		REMAPASSERT(flags & HAT_LOAD_NOCONSIST);
1378 		REMAPASSERT(PTE_GET(old_pte, PT_SOFTWARE) >= PT_NOCONSIST);
1379 		REMAPASSERT(pf_is_memory(PTE2PFN(old_pte, l)) ==
1380 		    pf_is_memory(PTE2PFN(pte, l)));
1381 		REMAPASSERT(!is_consist);
1382 	}
1383 
1384 	/*
1385 	 * We only let remaps change the certain bits in the PTE.
1386 	 */
1387 	if (PTE_GET(old_pte, ~PT_REMAP_BITS) != PTE_GET(pte, ~PT_REMAP_BITS))
1388 		panic("remap bits changed: old_pte="FMT_PTE", pte="FMT_PTE"\n",
1389 		    old_pte, pte);
1390 
1391 	/*
1392 	 * We don't create any mapping list entries on a remap, so release
1393 	 * any allocated hment after we drop the mapping list lock.
1394 	 */
1395 done:
1396 	if (is_consist) {
1397 		x86_hm_exit(pp);
1398 		if (hm != NULL)
1399 			hment_free(hm);
1400 	}
1401 	return (rv);
1402 }
1403 
1404 /*
1405  * Internal routine to load a single page table entry. This only fails if
1406  * we attempt to overwrite a page table link with a large page.
1407  */
1408 static int
1409 hati_load_common(
1410 	hat_t		*hat,
1411 	uintptr_t	va,
1412 	page_t		*pp,
1413 	uint_t		attr,
1414 	uint_t		flags,
1415 	level_t		level,
1416 	pfn_t		pfn)
1417 {
1418 	htable_t	*ht;
1419 	uint_t		entry;
1420 	x86pte_t	pte;
1421 	int		rv = 0;
1422 
1423 	/*
1424 	 * The number 16 is arbitrary and here to catch a recursion problem
1425 	 * early before we blow out the kernel stack.
1426 	 */
1427 	++curthread->t_hatdepth;
1428 	ASSERT(curthread->t_hatdepth < 16);
1429 
1430 	ASSERT(hat == kas.a_hat ||
1431 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1432 
1433 	if (flags & HAT_LOAD_SHARE)
1434 		hat->hat_flags |= HAT_SHARED;
1435 
1436 	/*
1437 	 * Find the page table that maps this page if it already exists.
1438 	 */
1439 	ht = htable_lookup(hat, va, level);
1440 
1441 	/*
1442 	 * We must have HAT_LOAD_NOCONSIST if page_t is NULL.
1443 	 */
1444 	if (pp == NULL)
1445 		flags |= HAT_LOAD_NOCONSIST;
1446 
1447 	if (ht == NULL) {
1448 		ht = htable_create(hat, va, level, NULL);
1449 		ASSERT(ht != NULL);
1450 	}
1451 	entry = htable_va2entry(va, ht);
1452 
1453 	/*
1454 	 * a bunch of paranoid error checking
1455 	 */
1456 	ASSERT(ht->ht_busy > 0);
1457 	if (ht->ht_vaddr > va || va > HTABLE_LAST_PAGE(ht))
1458 		panic("hati_load_common: bad htable %p, va %p",
1459 		    (void *)ht, (void *)va);
1460 	ASSERT(ht->ht_level == level);
1461 
1462 	/*
1463 	 * construct the new PTE
1464 	 */
1465 	if (hat == kas.a_hat)
1466 		attr &= ~PROT_USER;
1467 	pte = hati_mkpte(pfn, attr, level, flags);
1468 	if (hat == kas.a_hat && va >= kernelbase)
1469 		PTE_SET(pte, mmu.pt_global);
1470 
1471 	/*
1472 	 * establish the mapping
1473 	 */
1474 	rv = hati_pte_map(ht, entry, pp, pte, flags, NULL);
1475 
1476 	/*
1477 	 * release the htable and any reserves
1478 	 */
1479 	htable_release(ht);
1480 	--curthread->t_hatdepth;
1481 	return (rv);
1482 }
1483 
1484 /*
1485  * special case of hat_memload to deal with some kernel addrs for performance
1486  */
1487 static void
1488 hat_kmap_load(
1489 	caddr_t		addr,
1490 	page_t		*pp,
1491 	uint_t		attr,
1492 	uint_t		flags)
1493 {
1494 	uintptr_t	va = (uintptr_t)addr;
1495 	x86pte_t	pte;
1496 	pfn_t		pfn = page_pptonum(pp);
1497 	pgcnt_t		pg_off = mmu_btop(va - mmu.kmap_addr);
1498 	htable_t	*ht;
1499 	uint_t		entry;
1500 	void		*pte_ptr;
1501 
1502 	/*
1503 	 * construct the requested PTE
1504 	 */
1505 	attr &= ~PROT_USER;
1506 	attr |= HAT_STORECACHING_OK;
1507 	pte = hati_mkpte(pfn, attr, 0, flags);
1508 	PTE_SET(pte, mmu.pt_global);
1509 
1510 	/*
1511 	 * Figure out the pte_ptr and htable and use common code to finish up
1512 	 */
1513 	if (mmu.pae_hat)
1514 		pte_ptr = mmu.kmap_ptes + pg_off;
1515 	else
1516 		pte_ptr = (x86pte32_t *)mmu.kmap_ptes + pg_off;
1517 	ht = mmu.kmap_htables[(va - mmu.kmap_htables[0]->ht_vaddr) >>
1518 	    LEVEL_SHIFT(1)];
1519 	entry = htable_va2entry(va, ht);
1520 	++curthread->t_hatdepth;
1521 	ASSERT(curthread->t_hatdepth < 16);
1522 	(void) hati_pte_map(ht, entry, pp, pte, flags, pte_ptr);
1523 	--curthread->t_hatdepth;
1524 }
1525 
1526 /*
1527  * hat_memload() - load a translation to the given page struct
1528  *
1529  * Flags for hat_memload/hat_devload/hat_*attr.
1530  *
1531  * 	HAT_LOAD	Default flags to load a translation to the page.
1532  *
1533  * 	HAT_LOAD_LOCK	Lock down mapping resources; hat_map(), hat_memload(),
1534  *			and hat_devload().
1535  *
1536  *	HAT_LOAD_NOCONSIST Do not add mapping to page_t mapping list.
1537  *			sets PT_NOCONSIST
1538  *
1539  *	HAT_LOAD_SHARE	A flag to hat_memload() to indicate h/w page tables
1540  *			that map some user pages (not kas) is shared by more
1541  *			than one process (eg. ISM).
1542  *
1543  *	HAT_LOAD_REMAP	Reload a valid pte with a different page frame.
1544  *
1545  *	HAT_NO_KALLOC	Do not kmem_alloc while creating the mapping; at this
1546  *			point, it's setting up mapping to allocate internal
1547  *			hat layer data structures.  This flag forces hat layer
1548  *			to tap its reserves in order to prevent infinite
1549  *			recursion.
1550  *
1551  * The following is a protection attribute (like PROT_READ, etc.)
1552  *
1553  *	HAT_NOSYNC	set PT_NOSYNC - this mapping's ref/mod bits
1554  *			are never cleared.
1555  *
1556  * Installing new valid PTE's and creation of the mapping list
1557  * entry are controlled under the same lock. It's derived from the
1558  * page_t being mapped.
1559  */
1560 static uint_t supported_memload_flags =
1561 	HAT_LOAD | HAT_LOAD_LOCK | HAT_LOAD_ADV | HAT_LOAD_NOCONSIST |
1562 	HAT_LOAD_SHARE | HAT_NO_KALLOC | HAT_LOAD_REMAP | HAT_LOAD_TEXT;
1563 
1564 void
1565 hat_memload(
1566 	hat_t		*hat,
1567 	caddr_t		addr,
1568 	page_t		*pp,
1569 	uint_t		attr,
1570 	uint_t		flags)
1571 {
1572 	uintptr_t	va = (uintptr_t)addr;
1573 	level_t		level = 0;
1574 	pfn_t		pfn = page_pptonum(pp);
1575 
1576 	XPV_DISALLOW_MIGRATE();
1577 	ASSERT(IS_PAGEALIGNED(va));
1578 	ASSERT(hat == kas.a_hat || va < _userlimit);
1579 	ASSERT(hat == kas.a_hat ||
1580 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1581 	ASSERT((flags & supported_memload_flags) == flags);
1582 
1583 	ASSERT(!IN_VA_HOLE(va));
1584 	ASSERT(!PP_ISFREE(pp));
1585 
1586 	/*
1587 	 * kernel address special case for performance.
1588 	 */
1589 	if (mmu.kmap_addr <= va && va < mmu.kmap_eaddr) {
1590 		ASSERT(hat == kas.a_hat);
1591 		hat_kmap_load(addr, pp, attr, flags);
1592 		XPV_ALLOW_MIGRATE();
1593 		return;
1594 	}
1595 
1596 	/*
1597 	 * This is used for memory with normal caching enabled, so
1598 	 * always set HAT_STORECACHING_OK.
1599 	 */
1600 	attr |= HAT_STORECACHING_OK;
1601 	if (hati_load_common(hat, va, pp, attr, flags, level, pfn) != 0)
1602 		panic("unexpected hati_load_common() failure");
1603 	XPV_ALLOW_MIGRATE();
1604 }
1605 
1606 /* ARGSUSED */
1607 void
1608 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp,
1609     uint_t attr, uint_t flags, hat_region_cookie_t rcookie)
1610 {
1611 	hat_memload(hat, addr, pp, attr, flags);
1612 }
1613 
1614 /*
1615  * Load the given array of page structs using large pages when possible
1616  */
1617 void
1618 hat_memload_array(
1619 	hat_t		*hat,
1620 	caddr_t		addr,
1621 	size_t		len,
1622 	page_t		**pages,
1623 	uint_t		attr,
1624 	uint_t		flags)
1625 {
1626 	uintptr_t	va = (uintptr_t)addr;
1627 	uintptr_t	eaddr = va + len;
1628 	level_t		level;
1629 	size_t		pgsize;
1630 	pgcnt_t		pgindx = 0;
1631 	pfn_t		pfn;
1632 	pgcnt_t		i;
1633 
1634 	XPV_DISALLOW_MIGRATE();
1635 	ASSERT(IS_PAGEALIGNED(va));
1636 	ASSERT(hat == kas.a_hat || va + len <= _userlimit);
1637 	ASSERT(hat == kas.a_hat ||
1638 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1639 	ASSERT((flags & supported_memload_flags) == flags);
1640 
1641 	/*
1642 	 * memload is used for memory with full caching enabled, so
1643 	 * set HAT_STORECACHING_OK.
1644 	 */
1645 	attr |= HAT_STORECACHING_OK;
1646 
1647 	/*
1648 	 * handle all pages using largest possible pagesize
1649 	 */
1650 	while (va < eaddr) {
1651 		/*
1652 		 * decide what level mapping to use (ie. pagesize)
1653 		 */
1654 		pfn = page_pptonum(pages[pgindx]);
1655 		for (level = mmu.max_page_level; ; --level) {
1656 			pgsize = LEVEL_SIZE(level);
1657 			if (level == 0)
1658 				break;
1659 
1660 			if (!IS_P2ALIGNED(va, pgsize) ||
1661 			    (eaddr - va) < pgsize ||
1662 			    !IS_P2ALIGNED(pfn_to_pa(pfn), pgsize))
1663 				continue;
1664 
1665 			/*
1666 			 * To use a large mapping of this size, all the
1667 			 * pages we are passed must be sequential subpages
1668 			 * of the large page.
1669 			 * hat_page_demote() can't change p_szc because
1670 			 * all pages are locked.
1671 			 */
1672 			if (pages[pgindx]->p_szc >= level) {
1673 				for (i = 0; i < mmu_btop(pgsize); ++i) {
1674 					if (pfn + i !=
1675 					    page_pptonum(pages[pgindx + i]))
1676 						break;
1677 					ASSERT(pages[pgindx + i]->p_szc >=
1678 					    level);
1679 					ASSERT(pages[pgindx] + i ==
1680 					    pages[pgindx + i]);
1681 				}
1682 				if (i == mmu_btop(pgsize)) {
1683 #ifdef DEBUG
1684 					if (level == 2)
1685 						map1gcnt++;
1686 #endif
1687 					break;
1688 				}
1689 			}
1690 		}
1691 
1692 		/*
1693 		 * Load this page mapping. If the load fails, try a smaller
1694 		 * pagesize.
1695 		 */
1696 		ASSERT(!IN_VA_HOLE(va));
1697 		while (hati_load_common(hat, va, pages[pgindx], attr,
1698 		    flags, level, pfn) != 0) {
1699 			if (level == 0)
1700 				panic("unexpected hati_load_common() failure");
1701 			--level;
1702 			pgsize = LEVEL_SIZE(level);
1703 		}
1704 
1705 		/*
1706 		 * move to next page
1707 		 */
1708 		va += pgsize;
1709 		pgindx += mmu_btop(pgsize);
1710 	}
1711 	XPV_ALLOW_MIGRATE();
1712 }
1713 
1714 /* ARGSUSED */
1715 void
1716 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len,
1717     struct page **pps, uint_t attr, uint_t flags,
1718     hat_region_cookie_t rcookie)
1719 {
1720 	hat_memload_array(hat, addr, len, pps, attr, flags);
1721 }
1722 
1723 /*
1724  * void hat_devload(hat, addr, len, pf, attr, flags)
1725  *	load/lock the given page frame number
1726  *
1727  * Advisory ordering attributes. Apply only to device mappings.
1728  *
1729  * HAT_STRICTORDER: the CPU must issue the references in order, as the
1730  *	programmer specified.  This is the default.
1731  * HAT_UNORDERED_OK: the CPU may reorder the references (this is all kinds
1732  *	of reordering; store or load with store or load).
1733  * HAT_MERGING_OK: merging and batching: the CPU may merge individual stores
1734  *	to consecutive locations (for example, turn two consecutive byte
1735  *	stores into one halfword store), and it may batch individual loads
1736  *	(for example, turn two consecutive byte loads into one halfword load).
1737  *	This also implies re-ordering.
1738  * HAT_LOADCACHING_OK: the CPU may cache the data it fetches and reuse it
1739  *	until another store occurs.  The default is to fetch new data
1740  *	on every load.  This also implies merging.
1741  * HAT_STORECACHING_OK: the CPU may keep the data in the cache and push it to
1742  *	the device (perhaps with other data) at a later time.  The default is
1743  *	to push the data right away.  This also implies load caching.
1744  *
1745  * Equivalent of hat_memload(), but can be used for device memory where
1746  * there are no page_t's and we support additional flags (write merging, etc).
1747  * Note that we can have large page mappings with this interface.
1748  */
1749 int supported_devload_flags = HAT_LOAD | HAT_LOAD_LOCK |
1750 	HAT_LOAD_NOCONSIST | HAT_STRICTORDER | HAT_UNORDERED_OK |
1751 	HAT_MERGING_OK | HAT_LOADCACHING_OK | HAT_STORECACHING_OK;
1752 
1753 void
1754 hat_devload(
1755 	hat_t		*hat,
1756 	caddr_t		addr,
1757 	size_t		len,
1758 	pfn_t		pfn,
1759 	uint_t		attr,
1760 	int		flags)
1761 {
1762 	uintptr_t	va = ALIGN2PAGE(addr);
1763 	uintptr_t	eva = va + len;
1764 	level_t		level;
1765 	size_t		pgsize;
1766 	page_t		*pp;
1767 	int		f;	/* per PTE copy of flags  - maybe modified */
1768 	uint_t		a;	/* per PTE copy of attr */
1769 
1770 	XPV_DISALLOW_MIGRATE();
1771 	ASSERT(IS_PAGEALIGNED(va));
1772 	ASSERT(hat == kas.a_hat || eva <= _userlimit);
1773 	ASSERT(hat == kas.a_hat ||
1774 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1775 	ASSERT((flags & supported_devload_flags) == flags);
1776 
1777 	/*
1778 	 * handle all pages
1779 	 */
1780 	while (va < eva) {
1781 
1782 		/*
1783 		 * decide what level mapping to use (ie. pagesize)
1784 		 */
1785 		for (level = mmu.max_page_level; ; --level) {
1786 			pgsize = LEVEL_SIZE(level);
1787 			if (level == 0)
1788 				break;
1789 			if (IS_P2ALIGNED(va, pgsize) &&
1790 			    (eva - va) >= pgsize &&
1791 			    IS_P2ALIGNED(pfn, mmu_btop(pgsize))) {
1792 #ifdef DEBUG
1793 				if (level == 2)
1794 					map1gcnt++;
1795 #endif
1796 				break;
1797 			}
1798 		}
1799 
1800 		/*
1801 		 * If this is just memory then allow caching (this happens
1802 		 * for the nucleus pages) - though HAT_PLAT_NOCACHE can be used
1803 		 * to override that. If we don't have a page_t then make sure
1804 		 * NOCONSIST is set.
1805 		 */
1806 		a = attr;
1807 		f = flags;
1808 		if (!pf_is_memory(pfn))
1809 			f |= HAT_LOAD_NOCONSIST;
1810 		else if (!(a & HAT_PLAT_NOCACHE))
1811 			a |= HAT_STORECACHING_OK;
1812 
1813 		if (f & HAT_LOAD_NOCONSIST)
1814 			pp = NULL;
1815 		else
1816 			pp = page_numtopp_nolock(pfn);
1817 
1818 		/*
1819 		 * load this page mapping
1820 		 */
1821 		ASSERT(!IN_VA_HOLE(va));
1822 		while (hati_load_common(hat, va, pp, a, f, level, pfn) != 0) {
1823 			if (level == 0)
1824 				panic("unexpected hati_load_common() failure");
1825 			--level;
1826 			pgsize = LEVEL_SIZE(level);
1827 		}
1828 
1829 		/*
1830 		 * move to next page
1831 		 */
1832 		va += pgsize;
1833 		pfn += mmu_btop(pgsize);
1834 	}
1835 	XPV_ALLOW_MIGRATE();
1836 }
1837 
1838 /*
1839  * void hat_unlock(hat, addr, len)
1840  *	unlock the mappings to a given range of addresses
1841  *
1842  * Locks are tracked by ht_lock_cnt in the htable.
1843  */
1844 void
1845 hat_unlock(hat_t *hat, caddr_t addr, size_t len)
1846 {
1847 	uintptr_t	vaddr = (uintptr_t)addr;
1848 	uintptr_t	eaddr = vaddr + len;
1849 	htable_t	*ht = NULL;
1850 
1851 	/*
1852 	 * kernel entries are always locked, we don't track lock counts
1853 	 */
1854 	ASSERT(hat == kas.a_hat || eaddr <= _userlimit);
1855 	ASSERT(IS_PAGEALIGNED(vaddr));
1856 	ASSERT(IS_PAGEALIGNED(eaddr));
1857 	if (hat == kas.a_hat)
1858 		return;
1859 	if (eaddr > _userlimit)
1860 		panic("hat_unlock() address out of range - above _userlimit");
1861 
1862 	XPV_DISALLOW_MIGRATE();
1863 	ASSERT(AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
1864 	while (vaddr < eaddr) {
1865 		(void) htable_walk(hat, &ht, &vaddr, eaddr);
1866 		if (ht == NULL)
1867 			break;
1868 
1869 		ASSERT(!IN_VA_HOLE(vaddr));
1870 
1871 		if (ht->ht_lock_cnt < 1)
1872 			panic("hat_unlock(): lock_cnt < 1, "
1873 			    "htable=%p, vaddr=%p\n", (void *)ht, (void *)vaddr);
1874 		HTABLE_LOCK_DEC(ht);
1875 
1876 		vaddr += LEVEL_SIZE(ht->ht_level);
1877 	}
1878 	if (ht)
1879 		htable_release(ht);
1880 	XPV_ALLOW_MIGRATE();
1881 }
1882 
1883 /* ARGSUSED */
1884 void
1885 hat_unlock_region(struct hat *hat, caddr_t addr, size_t len,
1886     hat_region_cookie_t rcookie)
1887 {
1888 	panic("No shared region support on x86");
1889 }
1890 
1891 #if !defined(__xpv)
1892 /*
1893  * Cross call service routine to demap a virtual page on
1894  * the current CPU or flush all mappings in TLB.
1895  */
1896 /*ARGSUSED*/
1897 static int
1898 hati_demap_func(xc_arg_t a1, xc_arg_t a2, xc_arg_t a3)
1899 {
1900 	hat_t	*hat = (hat_t *)a1;
1901 	caddr_t	addr = (caddr_t)a2;
1902 
1903 	/*
1904 	 * If the target hat isn't the kernel and this CPU isn't operating
1905 	 * in the target hat, we can ignore the cross call.
1906 	 */
1907 	if (hat != kas.a_hat && hat != CPU->cpu_current_hat)
1908 		return (0);
1909 
1910 	/*
1911 	 * For a normal address, we just flush one page mapping
1912 	 */
1913 	if ((uintptr_t)addr != DEMAP_ALL_ADDR) {
1914 		mmu_tlbflush_entry(addr);
1915 		return (0);
1916 	}
1917 
1918 	/*
1919 	 * Otherwise we reload cr3 to effect a complete TLB flush.
1920 	 *
1921 	 * A reload of cr3 on a VLP process also means we must also recopy in
1922 	 * the pte values from the struct hat
1923 	 */
1924 	if (hat->hat_flags & HAT_VLP) {
1925 #if defined(__amd64)
1926 		x86pte_t *vlpptep = CPU->cpu_hat_info->hci_vlp_l2ptes;
1927 
1928 		VLP_COPY(hat->hat_vlp_ptes, vlpptep);
1929 #elif defined(__i386)
1930 		reload_pae32(hat, CPU);
1931 #endif
1932 	}
1933 	reload_cr3();
1934 	return (0);
1935 }
1936 
1937 /*
1938  * Flush all TLB entries, including global (ie. kernel) ones.
1939  */
1940 static void
1941 flush_all_tlb_entries(void)
1942 {
1943 	ulong_t cr4 = getcr4();
1944 
1945 	if (cr4 & CR4_PGE) {
1946 		setcr4(cr4 & ~(ulong_t)CR4_PGE);
1947 		setcr4(cr4);
1948 
1949 		/*
1950 		 * 32 bit PAE also needs to always reload_cr3()
1951 		 */
1952 		if (mmu.max_level == 2)
1953 			reload_cr3();
1954 	} else {
1955 		reload_cr3();
1956 	}
1957 }
1958 
1959 #define	TLB_CPU_HALTED	(01ul)
1960 #define	TLB_INVAL_ALL	(02ul)
1961 #define	CAS_TLB_INFO(cpu, old, new)	\
1962 	caslong((ulong_t *)&(cpu)->cpu_m.mcpu_tlb_info, (old), (new))
1963 
1964 /*
1965  * Record that a CPU is going idle
1966  */
1967 void
1968 tlb_going_idle(void)
1969 {
1970 	atomic_or_long((ulong_t *)&CPU->cpu_m.mcpu_tlb_info, TLB_CPU_HALTED);
1971 }
1972 
1973 /*
1974  * Service a delayed TLB flush if coming out of being idle.
1975  */
1976 void
1977 tlb_service(void)
1978 {
1979 	ulong_t flags = getflags();
1980 	ulong_t tlb_info;
1981 	ulong_t found;
1982 
1983 	/*
1984 	 * Be sure interrupts are off while doing this so that
1985 	 * higher level interrupts correctly wait for flushes to finish.
1986 	 */
1987 	if (flags & PS_IE)
1988 		flags = intr_clear();
1989 
1990 	/*
1991 	 * We only have to do something if coming out of being idle.
1992 	 */
1993 	tlb_info = CPU->cpu_m.mcpu_tlb_info;
1994 	if (tlb_info & TLB_CPU_HALTED) {
1995 		ASSERT(CPU->cpu_current_hat == kas.a_hat);
1996 
1997 		/*
1998 		 * Atomic clear and fetch of old state.
1999 		 */
2000 		while ((found = CAS_TLB_INFO(CPU, tlb_info, 0)) != tlb_info) {
2001 			ASSERT(found & TLB_CPU_HALTED);
2002 			tlb_info = found;
2003 			SMT_PAUSE();
2004 		}
2005 		if (tlb_info & TLB_INVAL_ALL)
2006 			flush_all_tlb_entries();
2007 	}
2008 
2009 	/*
2010 	 * Restore interrupt enable control bit.
2011 	 */
2012 	if (flags & PS_IE)
2013 		sti();
2014 }
2015 #endif /* !__xpv */
2016 
2017 /*
2018  * Internal routine to do cross calls to invalidate a range of pages on
2019  * all CPUs using a given hat.
2020  */
2021 void
2022 hat_tlb_inval(hat_t *hat, uintptr_t va)
2023 {
2024 	extern int	flushes_require_xcalls;	/* from mp_startup.c */
2025 	cpuset_t	justme;
2026 	cpuset_t	cpus_to_shootdown;
2027 #ifndef __xpv
2028 	cpuset_t	check_cpus;
2029 	cpu_t		*cpup;
2030 	int		c;
2031 #endif
2032 
2033 	/*
2034 	 * If the hat is being destroyed, there are no more users, so
2035 	 * demap need not do anything.
2036 	 */
2037 	if (hat->hat_flags & HAT_FREEING)
2038 		return;
2039 
2040 	/*
2041 	 * If demapping from a shared pagetable, we best demap the
2042 	 * entire set of user TLBs, since we don't know what addresses
2043 	 * these were shared at.
2044 	 */
2045 	if (hat->hat_flags & HAT_SHARED) {
2046 		hat = kas.a_hat;
2047 		va = DEMAP_ALL_ADDR;
2048 	}
2049 
2050 	/*
2051 	 * if not running with multiple CPUs, don't use cross calls
2052 	 */
2053 	if (panicstr || !flushes_require_xcalls) {
2054 #ifdef __xpv
2055 		if (va == DEMAP_ALL_ADDR)
2056 			xen_flush_tlb();
2057 		else
2058 			xen_flush_va((caddr_t)va);
2059 #else
2060 		(void) hati_demap_func((xc_arg_t)hat, (xc_arg_t)va, NULL);
2061 #endif
2062 		return;
2063 	}
2064 
2065 
2066 	/*
2067 	 * Determine CPUs to shootdown. Kernel changes always do all CPUs.
2068 	 * Otherwise it's just CPUs currently executing in this hat.
2069 	 */
2070 	kpreempt_disable();
2071 	CPUSET_ONLY(justme, CPU->cpu_id);
2072 	if (hat == kas.a_hat)
2073 		cpus_to_shootdown = khat_cpuset;
2074 	else
2075 		cpus_to_shootdown = hat->hat_cpus;
2076 
2077 #ifndef __xpv
2078 	/*
2079 	 * If any CPUs in the set are idle, just request a delayed flush
2080 	 * and avoid waking them up.
2081 	 */
2082 	check_cpus = cpus_to_shootdown;
2083 	for (c = 0; c < NCPU && !CPUSET_ISNULL(check_cpus); ++c) {
2084 		ulong_t tlb_info;
2085 
2086 		if (!CPU_IN_SET(check_cpus, c))
2087 			continue;
2088 		CPUSET_DEL(check_cpus, c);
2089 		cpup = cpu[c];
2090 		if (cpup == NULL)
2091 			continue;
2092 
2093 		tlb_info = cpup->cpu_m.mcpu_tlb_info;
2094 		while (tlb_info == TLB_CPU_HALTED) {
2095 			(void) CAS_TLB_INFO(cpup, TLB_CPU_HALTED,
2096 			    TLB_CPU_HALTED | TLB_INVAL_ALL);
2097 			SMT_PAUSE();
2098 			tlb_info = cpup->cpu_m.mcpu_tlb_info;
2099 		}
2100 		if (tlb_info == (TLB_CPU_HALTED | TLB_INVAL_ALL)) {
2101 			HATSTAT_INC(hs_tlb_inval_delayed);
2102 			CPUSET_DEL(cpus_to_shootdown, c);
2103 		}
2104 	}
2105 #endif
2106 
2107 	if (CPUSET_ISNULL(cpus_to_shootdown) ||
2108 	    CPUSET_ISEQUAL(cpus_to_shootdown, justme)) {
2109 
2110 #ifdef __xpv
2111 		if (va == DEMAP_ALL_ADDR)
2112 			xen_flush_tlb();
2113 		else
2114 			xen_flush_va((caddr_t)va);
2115 #else
2116 		(void) hati_demap_func((xc_arg_t)hat, (xc_arg_t)va, NULL);
2117 #endif
2118 
2119 	} else {
2120 
2121 		CPUSET_ADD(cpus_to_shootdown, CPU->cpu_id);
2122 #ifdef __xpv
2123 		if (va == DEMAP_ALL_ADDR)
2124 			xen_gflush_tlb(cpus_to_shootdown);
2125 		else
2126 			xen_gflush_va((caddr_t)va, cpus_to_shootdown);
2127 #else
2128 		xc_call((xc_arg_t)hat, (xc_arg_t)va, NULL, X_CALL_HIPRI,
2129 		    cpus_to_shootdown, hati_demap_func);
2130 #endif
2131 
2132 	}
2133 	kpreempt_enable();
2134 }
2135 
2136 /*
2137  * Interior routine for HAT_UNLOADs from hat_unload_callback(),
2138  * hat_kmap_unload() OR from hat_steal() code.  This routine doesn't
2139  * handle releasing of the htables.
2140  */
2141 void
2142 hat_pte_unmap(
2143 	htable_t	*ht,
2144 	uint_t		entry,
2145 	uint_t		flags,
2146 	x86pte_t	old_pte,
2147 	void		*pte_ptr)
2148 {
2149 	hat_t		*hat = ht->ht_hat;
2150 	hment_t		*hm = NULL;
2151 	page_t		*pp = NULL;
2152 	level_t		l = ht->ht_level;
2153 	pfn_t		pfn;
2154 
2155 	/*
2156 	 * We always track the locking counts, even if nothing is unmapped
2157 	 */
2158 	if ((flags & HAT_UNLOAD_UNLOCK) != 0 && hat != kas.a_hat) {
2159 		ASSERT(ht->ht_lock_cnt > 0);
2160 		HTABLE_LOCK_DEC(ht);
2161 	}
2162 
2163 	/*
2164 	 * Figure out which page's mapping list lock to acquire using the PFN
2165 	 * passed in "old" PTE. We then attempt to invalidate the PTE.
2166 	 * If another thread, probably a hat_pageunload, has asynchronously
2167 	 * unmapped/remapped this address we'll loop here.
2168 	 */
2169 	ASSERT(ht->ht_busy > 0);
2170 	while (PTE_ISVALID(old_pte)) {
2171 		pfn = PTE2PFN(old_pte, l);
2172 		if (PTE_GET(old_pte, PT_SOFTWARE) >= PT_NOCONSIST) {
2173 			pp = NULL;
2174 		} else {
2175 #ifdef __xpv
2176 			if (pfn == PFN_INVALID)
2177 				panic("Invalid PFN, but not PT_NOCONSIST");
2178 #endif
2179 			pp = page_numtopp_nolock(pfn);
2180 			if (pp == NULL) {
2181 				panic("no page_t, not NOCONSIST: old_pte="
2182 				    FMT_PTE " ht=%lx entry=0x%x pte_ptr=%lx",
2183 				    old_pte, (uintptr_t)ht, entry,
2184 				    (uintptr_t)pte_ptr);
2185 			}
2186 			x86_hm_enter(pp);
2187 		}
2188 
2189 		/*
2190 		 * If freeing the address space, check that the PTE
2191 		 * hasn't changed, as the mappings are no longer in use by
2192 		 * any thread, invalidation is unnecessary.
2193 		 * If not freeing, do a full invalidate.
2194 		 *
2195 		 * On the hypervisor we must always remove mappings, as a
2196 		 * writable mapping left behind could cause a page table
2197 		 * allocation to fail.
2198 		 */
2199 #if !defined(__xpv)
2200 		if (hat->hat_flags & HAT_FREEING)
2201 			old_pte = x86pte_get(ht, entry);
2202 		else
2203 #endif
2204 			old_pte = x86pte_inval(ht, entry, old_pte, pte_ptr);
2205 
2206 		/*
2207 		 * If the page hadn't changed we've unmapped it and can proceed
2208 		 */
2209 		if (PTE_ISVALID(old_pte) && PTE2PFN(old_pte, l) == pfn)
2210 			break;
2211 
2212 		/*
2213 		 * Otherwise, we'll have to retry with the current old_pte.
2214 		 * Drop the hment lock, since the pfn may have changed.
2215 		 */
2216 		if (pp != NULL) {
2217 			x86_hm_exit(pp);
2218 			pp = NULL;
2219 		} else {
2220 			ASSERT(PTE_GET(old_pte, PT_SOFTWARE) >= PT_NOCONSIST);
2221 		}
2222 	}
2223 
2224 	/*
2225 	 * If the old mapping wasn't valid, there's nothing more to do
2226 	 */
2227 	if (!PTE_ISVALID(old_pte)) {
2228 		if (pp != NULL)
2229 			x86_hm_exit(pp);
2230 		return;
2231 	}
2232 
2233 	/*
2234 	 * Take care of syncing any MOD/REF bits and removing the hment.
2235 	 */
2236 	if (pp != NULL) {
2237 		if (!(flags & HAT_UNLOAD_NOSYNC))
2238 			hati_sync_pte_to_page(pp, old_pte, l);
2239 		hm = hment_remove(pp, ht, entry);
2240 		x86_hm_exit(pp);
2241 		if (hm != NULL)
2242 			hment_free(hm);
2243 	}
2244 
2245 	/*
2246 	 * Handle book keeping in the htable and hat
2247 	 */
2248 	ASSERT(ht->ht_valid_cnt > 0);
2249 	HTABLE_DEC(ht->ht_valid_cnt);
2250 	PGCNT_DEC(hat, l);
2251 }
2252 
2253 /*
2254  * very cheap unload implementation to special case some kernel addresses
2255  */
2256 static void
2257 hat_kmap_unload(caddr_t addr, size_t len, uint_t flags)
2258 {
2259 	uintptr_t	va = (uintptr_t)addr;
2260 	uintptr_t	eva = va + len;
2261 	pgcnt_t		pg_index;
2262 	htable_t	*ht;
2263 	uint_t		entry;
2264 	x86pte_t	*pte_ptr;
2265 	x86pte_t	old_pte;
2266 
2267 	for (; va < eva; va += MMU_PAGESIZE) {
2268 		/*
2269 		 * Get the PTE
2270 		 */
2271 		pg_index = mmu_btop(va - mmu.kmap_addr);
2272 		pte_ptr = PT_INDEX_PTR(mmu.kmap_ptes, pg_index);
2273 		old_pte = GET_PTE(pte_ptr);
2274 
2275 		/*
2276 		 * get the htable / entry
2277 		 */
2278 		ht = mmu.kmap_htables[(va - mmu.kmap_htables[0]->ht_vaddr)
2279 		    >> LEVEL_SHIFT(1)];
2280 		entry = htable_va2entry(va, ht);
2281 
2282 		/*
2283 		 * use mostly common code to unmap it.
2284 		 */
2285 		hat_pte_unmap(ht, entry, flags, old_pte, pte_ptr);
2286 	}
2287 }
2288 
2289 
2290 /*
2291  * unload a range of virtual address space (no callback)
2292  */
2293 void
2294 hat_unload(hat_t *hat, caddr_t addr, size_t len, uint_t flags)
2295 {
2296 	uintptr_t va = (uintptr_t)addr;
2297 
2298 	XPV_DISALLOW_MIGRATE();
2299 	ASSERT(hat == kas.a_hat || va + len <= _userlimit);
2300 
2301 	/*
2302 	 * special case for performance.
2303 	 */
2304 	if (mmu.kmap_addr <= va && va < mmu.kmap_eaddr) {
2305 		ASSERT(hat == kas.a_hat);
2306 		hat_kmap_unload(addr, len, flags);
2307 	} else {
2308 		hat_unload_callback(hat, addr, len, flags, NULL);
2309 	}
2310 	XPV_ALLOW_MIGRATE();
2311 }
2312 
2313 /*
2314  * Do the callbacks for ranges being unloaded.
2315  */
2316 typedef struct range_info {
2317 	uintptr_t	rng_va;
2318 	ulong_t		rng_cnt;
2319 	level_t		rng_level;
2320 } range_info_t;
2321 
2322 static void
2323 handle_ranges(hat_callback_t *cb, uint_t cnt, range_info_t *range)
2324 {
2325 	/*
2326 	 * do callbacks to upper level VM system
2327 	 */
2328 	while (cb != NULL && cnt > 0) {
2329 		--cnt;
2330 		cb->hcb_start_addr = (caddr_t)range[cnt].rng_va;
2331 		cb->hcb_end_addr = cb->hcb_start_addr;
2332 		cb->hcb_end_addr +=
2333 		    range[cnt].rng_cnt << LEVEL_SIZE(range[cnt].rng_level);
2334 		cb->hcb_function(cb);
2335 	}
2336 }
2337 
2338 /*
2339  * Unload a given range of addresses (has optional callback)
2340  *
2341  * Flags:
2342  * define	HAT_UNLOAD		0x00
2343  * define	HAT_UNLOAD_NOSYNC	0x02
2344  * define	HAT_UNLOAD_UNLOCK	0x04
2345  * define	HAT_UNLOAD_OTHER	0x08 - not used
2346  * define	HAT_UNLOAD_UNMAP	0x10 - same as HAT_UNLOAD
2347  */
2348 #define	MAX_UNLOAD_CNT (8)
2349 void
2350 hat_unload_callback(
2351 	hat_t		*hat,
2352 	caddr_t		addr,
2353 	size_t		len,
2354 	uint_t		flags,
2355 	hat_callback_t	*cb)
2356 {
2357 	uintptr_t	vaddr = (uintptr_t)addr;
2358 	uintptr_t	eaddr = vaddr + len;
2359 	htable_t	*ht = NULL;
2360 	uint_t		entry;
2361 	uintptr_t	contig_va = (uintptr_t)-1L;
2362 	range_info_t	r[MAX_UNLOAD_CNT];
2363 	uint_t		r_cnt = 0;
2364 	x86pte_t	old_pte;
2365 
2366 	XPV_DISALLOW_MIGRATE();
2367 	ASSERT(hat == kas.a_hat || eaddr <= _userlimit);
2368 	ASSERT(IS_PAGEALIGNED(vaddr));
2369 	ASSERT(IS_PAGEALIGNED(eaddr));
2370 
2371 	/*
2372 	 * Special case a single page being unloaded for speed. This happens
2373 	 * quite frequently, COW faults after a fork() for example.
2374 	 */
2375 	if (cb == NULL && len == MMU_PAGESIZE) {
2376 		ht = htable_getpte(hat, vaddr, &entry, &old_pte, 0);
2377 		if (ht != NULL) {
2378 			if (PTE_ISVALID(old_pte))
2379 				hat_pte_unmap(ht, entry, flags, old_pte, NULL);
2380 			htable_release(ht);
2381 		}
2382 		XPV_ALLOW_MIGRATE();
2383 		return;
2384 	}
2385 
2386 	while (vaddr < eaddr) {
2387 		old_pte = htable_walk(hat, &ht, &vaddr, eaddr);
2388 		if (ht == NULL)
2389 			break;
2390 
2391 		ASSERT(!IN_VA_HOLE(vaddr));
2392 
2393 		if (vaddr < (uintptr_t)addr)
2394 			panic("hat_unload_callback(): unmap inside large page");
2395 
2396 		/*
2397 		 * We'll do the call backs for contiguous ranges
2398 		 */
2399 		if (vaddr != contig_va ||
2400 		    (r_cnt > 0 && r[r_cnt - 1].rng_level != ht->ht_level)) {
2401 			if (r_cnt == MAX_UNLOAD_CNT) {
2402 				handle_ranges(cb, r_cnt, r);
2403 				r_cnt = 0;
2404 			}
2405 			r[r_cnt].rng_va = vaddr;
2406 			r[r_cnt].rng_cnt = 0;
2407 			r[r_cnt].rng_level = ht->ht_level;
2408 			++r_cnt;
2409 		}
2410 
2411 		/*
2412 		 * Unload one mapping from the page tables.
2413 		 */
2414 		entry = htable_va2entry(vaddr, ht);
2415 		hat_pte_unmap(ht, entry, flags, old_pte, NULL);
2416 		ASSERT(ht->ht_level <= mmu.max_page_level);
2417 		vaddr += LEVEL_SIZE(ht->ht_level);
2418 		contig_va = vaddr;
2419 		++r[r_cnt - 1].rng_cnt;
2420 	}
2421 	if (ht)
2422 		htable_release(ht);
2423 
2424 	/*
2425 	 * handle last range for callbacks
2426 	 */
2427 	if (r_cnt > 0)
2428 		handle_ranges(cb, r_cnt, r);
2429 	XPV_ALLOW_MIGRATE();
2430 }
2431 
2432 /*
2433  * synchronize mapping with software data structures
2434  *
2435  * This interface is currently only used by the working set monitor
2436  * driver.
2437  */
2438 /*ARGSUSED*/
2439 void
2440 hat_sync(hat_t *hat, caddr_t addr, size_t len, uint_t flags)
2441 {
2442 	uintptr_t	vaddr = (uintptr_t)addr;
2443 	uintptr_t	eaddr = vaddr + len;
2444 	htable_t	*ht = NULL;
2445 	uint_t		entry;
2446 	x86pte_t	pte;
2447 	x86pte_t	save_pte;
2448 	x86pte_t	new;
2449 	page_t		*pp;
2450 
2451 	ASSERT(!IN_VA_HOLE(vaddr));
2452 	ASSERT(IS_PAGEALIGNED(vaddr));
2453 	ASSERT(IS_PAGEALIGNED(eaddr));
2454 	ASSERT(hat == kas.a_hat || eaddr <= _userlimit);
2455 
2456 	XPV_DISALLOW_MIGRATE();
2457 	for (; vaddr < eaddr; vaddr += LEVEL_SIZE(ht->ht_level)) {
2458 try_again:
2459 		pte = htable_walk(hat, &ht, &vaddr, eaddr);
2460 		if (ht == NULL)
2461 			break;
2462 		entry = htable_va2entry(vaddr, ht);
2463 
2464 		if (PTE_GET(pte, PT_SOFTWARE) >= PT_NOSYNC ||
2465 		    PTE_GET(pte, PT_REF | PT_MOD) == 0)
2466 			continue;
2467 
2468 		/*
2469 		 * We need to acquire the mapping list lock to protect
2470 		 * against hat_pageunload(), hat_unload(), etc.
2471 		 */
2472 		pp = page_numtopp_nolock(PTE2PFN(pte, ht->ht_level));
2473 		if (pp == NULL)
2474 			break;
2475 		x86_hm_enter(pp);
2476 		save_pte = pte;
2477 		pte = x86pte_get(ht, entry);
2478 		if (pte != save_pte) {
2479 			x86_hm_exit(pp);
2480 			goto try_again;
2481 		}
2482 		if (PTE_GET(pte, PT_SOFTWARE) >= PT_NOSYNC ||
2483 		    PTE_GET(pte, PT_REF | PT_MOD) == 0) {
2484 			x86_hm_exit(pp);
2485 			continue;
2486 		}
2487 
2488 		/*
2489 		 * Need to clear ref or mod bits. We may compete with
2490 		 * hardware updating the R/M bits and have to try again.
2491 		 */
2492 		if (flags == HAT_SYNC_ZERORM) {
2493 			new = pte;
2494 			PTE_CLR(new, PT_REF | PT_MOD);
2495 			pte = hati_update_pte(ht, entry, pte, new);
2496 			if (pte != 0) {
2497 				x86_hm_exit(pp);
2498 				goto try_again;
2499 			}
2500 		} else {
2501 			/*
2502 			 * sync the PTE to the page_t
2503 			 */
2504 			hati_sync_pte_to_page(pp, save_pte, ht->ht_level);
2505 		}
2506 		x86_hm_exit(pp);
2507 	}
2508 	if (ht)
2509 		htable_release(ht);
2510 	XPV_ALLOW_MIGRATE();
2511 }
2512 
2513 /*
2514  * void	hat_map(hat, addr, len, flags)
2515  */
2516 /*ARGSUSED*/
2517 void
2518 hat_map(hat_t *hat, caddr_t addr, size_t len, uint_t flags)
2519 {
2520 	/* does nothing */
2521 }
2522 
2523 /*
2524  * uint_t hat_getattr(hat, addr, *attr)
2525  *	returns attr for <hat,addr> in *attr.  returns 0 if there was a
2526  *	mapping and *attr is valid, nonzero if there was no mapping and
2527  *	*attr is not valid.
2528  */
2529 uint_t
2530 hat_getattr(hat_t *hat, caddr_t addr, uint_t *attr)
2531 {
2532 	uintptr_t	vaddr = ALIGN2PAGE(addr);
2533 	htable_t	*ht = NULL;
2534 	x86pte_t	pte;
2535 
2536 	ASSERT(hat == kas.a_hat || vaddr <= _userlimit);
2537 
2538 	if (IN_VA_HOLE(vaddr))
2539 		return ((uint_t)-1);
2540 
2541 	ht = htable_getpte(hat, vaddr, NULL, &pte, mmu.max_page_level);
2542 	if (ht == NULL)
2543 		return ((uint_t)-1);
2544 
2545 	if (!PTE_ISVALID(pte) || !PTE_ISPAGE(pte, ht->ht_level)) {
2546 		htable_release(ht);
2547 		return ((uint_t)-1);
2548 	}
2549 
2550 	*attr = PROT_READ;
2551 	if (PTE_GET(pte, PT_WRITABLE))
2552 		*attr |= PROT_WRITE;
2553 	if (PTE_GET(pte, PT_USER))
2554 		*attr |= PROT_USER;
2555 	if (!PTE_GET(pte, mmu.pt_nx))
2556 		*attr |= PROT_EXEC;
2557 	if (PTE_GET(pte, PT_SOFTWARE) >= PT_NOSYNC)
2558 		*attr |= HAT_NOSYNC;
2559 	htable_release(ht);
2560 	return (0);
2561 }
2562 
2563 /*
2564  * hat_updateattr() applies the given attribute change to an existing mapping
2565  */
2566 #define	HAT_LOAD_ATTR		1
2567 #define	HAT_SET_ATTR		2
2568 #define	HAT_CLR_ATTR		3
2569 
2570 static void
2571 hat_updateattr(hat_t *hat, caddr_t addr, size_t len, uint_t attr, int what)
2572 {
2573 	uintptr_t	vaddr = (uintptr_t)addr;
2574 	uintptr_t	eaddr = (uintptr_t)addr + len;
2575 	htable_t	*ht = NULL;
2576 	uint_t		entry;
2577 	x86pte_t	oldpte, newpte;
2578 	page_t		*pp;
2579 
2580 	XPV_DISALLOW_MIGRATE();
2581 	ASSERT(IS_PAGEALIGNED(vaddr));
2582 	ASSERT(IS_PAGEALIGNED(eaddr));
2583 	ASSERT(hat == kas.a_hat ||
2584 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
2585 	for (; vaddr < eaddr; vaddr += LEVEL_SIZE(ht->ht_level)) {
2586 try_again:
2587 		oldpte = htable_walk(hat, &ht, &vaddr, eaddr);
2588 		if (ht == NULL)
2589 			break;
2590 		if (PTE_GET(oldpte, PT_SOFTWARE) >= PT_NOCONSIST)
2591 			continue;
2592 
2593 		pp = page_numtopp_nolock(PTE2PFN(oldpte, ht->ht_level));
2594 		if (pp == NULL)
2595 			continue;
2596 		x86_hm_enter(pp);
2597 
2598 		newpte = oldpte;
2599 		/*
2600 		 * We found a page table entry in the desired range,
2601 		 * figure out the new attributes.
2602 		 */
2603 		if (what == HAT_SET_ATTR || what == HAT_LOAD_ATTR) {
2604 			if ((attr & PROT_WRITE) &&
2605 			    !PTE_GET(oldpte, PT_WRITABLE))
2606 				newpte |= PT_WRITABLE;
2607 
2608 			if ((attr & HAT_NOSYNC) &&
2609 			    PTE_GET(oldpte, PT_SOFTWARE) < PT_NOSYNC)
2610 				newpte |= PT_NOSYNC;
2611 
2612 			if ((attr & PROT_EXEC) && PTE_GET(oldpte, mmu.pt_nx))
2613 				newpte &= ~mmu.pt_nx;
2614 		}
2615 
2616 		if (what == HAT_LOAD_ATTR) {
2617 			if (!(attr & PROT_WRITE) &&
2618 			    PTE_GET(oldpte, PT_WRITABLE))
2619 				newpte &= ~PT_WRITABLE;
2620 
2621 			if (!(attr & HAT_NOSYNC) &&
2622 			    PTE_GET(oldpte, PT_SOFTWARE) >= PT_NOSYNC)
2623 				newpte &= ~PT_SOFTWARE;
2624 
2625 			if (!(attr & PROT_EXEC) && !PTE_GET(oldpte, mmu.pt_nx))
2626 				newpte |= mmu.pt_nx;
2627 		}
2628 
2629 		if (what == HAT_CLR_ATTR) {
2630 			if ((attr & PROT_WRITE) && PTE_GET(oldpte, PT_WRITABLE))
2631 				newpte &= ~PT_WRITABLE;
2632 
2633 			if ((attr & HAT_NOSYNC) &&
2634 			    PTE_GET(oldpte, PT_SOFTWARE) >= PT_NOSYNC)
2635 				newpte &= ~PT_SOFTWARE;
2636 
2637 			if ((attr & PROT_EXEC) && !PTE_GET(oldpte, mmu.pt_nx))
2638 				newpte |= mmu.pt_nx;
2639 		}
2640 
2641 		/*
2642 		 * Ensure NOSYNC/NOCONSIST mappings have REF and MOD set.
2643 		 * x86pte_set() depends on this.
2644 		 */
2645 		if (PTE_GET(newpte, PT_SOFTWARE) >= PT_NOSYNC)
2646 			newpte |= PT_REF | PT_MOD;
2647 
2648 		/*
2649 		 * what about PROT_READ or others? this code only handles:
2650 		 * EXEC, WRITE, NOSYNC
2651 		 */
2652 
2653 		/*
2654 		 * If new PTE really changed, update the table.
2655 		 */
2656 		if (newpte != oldpte) {
2657 			entry = htable_va2entry(vaddr, ht);
2658 			oldpte = hati_update_pte(ht, entry, oldpte, newpte);
2659 			if (oldpte != 0) {
2660 				x86_hm_exit(pp);
2661 				goto try_again;
2662 			}
2663 		}
2664 		x86_hm_exit(pp);
2665 	}
2666 	if (ht)
2667 		htable_release(ht);
2668 	XPV_ALLOW_MIGRATE();
2669 }
2670 
2671 /*
2672  * Various wrappers for hat_updateattr()
2673  */
2674 void
2675 hat_setattr(hat_t *hat, caddr_t addr, size_t len, uint_t attr)
2676 {
2677 	ASSERT(hat == kas.a_hat || (uintptr_t)addr + len <= _userlimit);
2678 	hat_updateattr(hat, addr, len, attr, HAT_SET_ATTR);
2679 }
2680 
2681 void
2682 hat_clrattr(hat_t *hat, caddr_t addr, size_t len, uint_t attr)
2683 {
2684 	ASSERT(hat == kas.a_hat || (uintptr_t)addr + len <= _userlimit);
2685 	hat_updateattr(hat, addr, len, attr, HAT_CLR_ATTR);
2686 }
2687 
2688 void
2689 hat_chgattr(hat_t *hat, caddr_t addr, size_t len, uint_t attr)
2690 {
2691 	ASSERT(hat == kas.a_hat || (uintptr_t)addr + len <= _userlimit);
2692 	hat_updateattr(hat, addr, len, attr, HAT_LOAD_ATTR);
2693 }
2694 
2695 void
2696 hat_chgprot(hat_t *hat, caddr_t addr, size_t len, uint_t vprot)
2697 {
2698 	ASSERT(hat == kas.a_hat || (uintptr_t)addr + len <= _userlimit);
2699 	hat_updateattr(hat, addr, len, vprot & HAT_PROT_MASK, HAT_LOAD_ATTR);
2700 }
2701 
2702 /*
2703  * size_t hat_getpagesize(hat, addr)
2704  *	returns pagesize in bytes for <hat, addr>. returns -1 of there is
2705  *	no mapping. This is an advisory call.
2706  */
2707 ssize_t
2708 hat_getpagesize(hat_t *hat, caddr_t addr)
2709 {
2710 	uintptr_t	vaddr = ALIGN2PAGE(addr);
2711 	htable_t	*ht;
2712 	size_t		pagesize;
2713 
2714 	ASSERT(hat == kas.a_hat || vaddr <= _userlimit);
2715 	if (IN_VA_HOLE(vaddr))
2716 		return (-1);
2717 	ht = htable_getpage(hat, vaddr, NULL);
2718 	if (ht == NULL)
2719 		return (-1);
2720 	pagesize = LEVEL_SIZE(ht->ht_level);
2721 	htable_release(ht);
2722 	return (pagesize);
2723 }
2724 
2725 
2726 
2727 /*
2728  * pfn_t hat_getpfnum(hat, addr)
2729  *	returns pfn for <hat, addr> or PFN_INVALID if mapping is invalid.
2730  */
2731 pfn_t
2732 hat_getpfnum(hat_t *hat, caddr_t addr)
2733 {
2734 	uintptr_t	vaddr = ALIGN2PAGE(addr);
2735 	htable_t	*ht;
2736 	uint_t		entry;
2737 	pfn_t		pfn = PFN_INVALID;
2738 
2739 	ASSERT(hat == kas.a_hat || vaddr <= _userlimit);
2740 	if (khat_running == 0)
2741 		return (PFN_INVALID);
2742 
2743 	if (IN_VA_HOLE(vaddr))
2744 		return (PFN_INVALID);
2745 
2746 	XPV_DISALLOW_MIGRATE();
2747 	/*
2748 	 * A very common use of hat_getpfnum() is from the DDI for kernel pages.
2749 	 * Use the kmap_ptes (which also covers the 32 bit heap) to speed
2750 	 * this up.
2751 	 */
2752 	if (mmu.kmap_addr <= vaddr && vaddr < mmu.kmap_eaddr) {
2753 		x86pte_t pte;
2754 		pgcnt_t pg_index;
2755 
2756 		pg_index = mmu_btop(vaddr - mmu.kmap_addr);
2757 		pte = GET_PTE(PT_INDEX_PTR(mmu.kmap_ptes, pg_index));
2758 		if (PTE_ISVALID(pte))
2759 			/*LINTED [use of constant 0 causes a lint warning] */
2760 			pfn = PTE2PFN(pte, 0);
2761 		XPV_ALLOW_MIGRATE();
2762 		return (pfn);
2763 	}
2764 
2765 	ht = htable_getpage(hat, vaddr, &entry);
2766 	if (ht == NULL) {
2767 		XPV_ALLOW_MIGRATE();
2768 		return (PFN_INVALID);
2769 	}
2770 	ASSERT(vaddr >= ht->ht_vaddr);
2771 	ASSERT(vaddr <= HTABLE_LAST_PAGE(ht));
2772 	pfn = PTE2PFN(x86pte_get(ht, entry), ht->ht_level);
2773 	if (ht->ht_level > 0)
2774 		pfn += mmu_btop(vaddr & LEVEL_OFFSET(ht->ht_level));
2775 	htable_release(ht);
2776 	XPV_ALLOW_MIGRATE();
2777 	return (pfn);
2778 }
2779 
2780 /*
2781  * hat_getkpfnum() is an obsolete DDI routine, and its use is discouraged.
2782  * Use hat_getpfnum(kas.a_hat, ...) instead.
2783  *
2784  * We'd like to return PFN_INVALID if the mappings have underlying page_t's
2785  * but can't right now due to the fact that some software has grown to use
2786  * this interface incorrectly. So for now when the interface is misused,
2787  * return a warning to the user that in the future it won't work in the
2788  * way they're abusing it, and carry on.
2789  *
2790  * Note that hat_getkpfnum() is never supported on amd64.
2791  */
2792 #if !defined(__amd64)
2793 pfn_t
2794 hat_getkpfnum(caddr_t addr)
2795 {
2796 	pfn_t	pfn;
2797 	int badcaller = 0;
2798 
2799 	if (khat_running == 0)
2800 		panic("hat_getkpfnum(): called too early\n");
2801 	if ((uintptr_t)addr < kernelbase)
2802 		return (PFN_INVALID);
2803 
2804 	XPV_DISALLOW_MIGRATE();
2805 	if (segkpm && IS_KPM_ADDR(addr)) {
2806 		badcaller = 1;
2807 		pfn = hat_kpm_va2pfn(addr);
2808 	} else {
2809 		pfn = hat_getpfnum(kas.a_hat, addr);
2810 		badcaller = pf_is_memory(pfn);
2811 	}
2812 
2813 	if (badcaller)
2814 		hat_getkpfnum_badcall(caller());
2815 	XPV_ALLOW_MIGRATE();
2816 	return (pfn);
2817 }
2818 #endif /* __amd64 */
2819 
2820 /*
2821  * int hat_probe(hat, addr)
2822  *	return 0 if no valid mapping is present.  Faster version
2823  *	of hat_getattr in certain architectures.
2824  */
2825 int
2826 hat_probe(hat_t *hat, caddr_t addr)
2827 {
2828 	uintptr_t	vaddr = ALIGN2PAGE(addr);
2829 	uint_t		entry;
2830 	htable_t	*ht;
2831 	pgcnt_t		pg_off;
2832 
2833 	ASSERT(hat == kas.a_hat || vaddr <= _userlimit);
2834 	ASSERT(hat == kas.a_hat ||
2835 	    AS_LOCK_HELD(hat->hat_as, &hat->hat_as->a_lock));
2836 	if (IN_VA_HOLE(vaddr))
2837 		return (0);
2838 
2839 	/*
2840 	 * Most common use of hat_probe is from segmap. We special case it
2841 	 * for performance.
2842 	 */
2843 	if (mmu.kmap_addr <= vaddr && vaddr < mmu.kmap_eaddr) {
2844 		pg_off = mmu_btop(vaddr - mmu.kmap_addr);
2845 		if (mmu.pae_hat)
2846 			return (PTE_ISVALID(mmu.kmap_ptes[pg_off]));
2847 		else
2848 			return (PTE_ISVALID(
2849 			    ((x86pte32_t *)mmu.kmap_ptes)[pg_off]));
2850 	}
2851 
2852 	ht = htable_getpage(hat, vaddr, &entry);
2853 	htable_release(ht);
2854 	return (ht != NULL);
2855 }
2856 
2857 /*
2858  * Find out if the segment for hat_share()/hat_unshare() is DISM or locked ISM.
2859  */
2860 static int
2861 is_it_dism(hat_t *hat, caddr_t va)
2862 {
2863 	struct seg *seg;
2864 	struct shm_data *shmd;
2865 	struct spt_data *sptd;
2866 
2867 	seg = as_findseg(hat->hat_as, va, 0);
2868 	ASSERT(seg != NULL);
2869 	ASSERT(seg->s_base <= va);
2870 	shmd = (struct shm_data *)seg->s_data;
2871 	ASSERT(shmd != NULL);
2872 	sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2873 	ASSERT(sptd != NULL);
2874 	if (sptd->spt_flags & SHM_PAGEABLE)
2875 		return (1);
2876 	return (0);
2877 }
2878 
2879 /*
2880  * Simple implementation of ISM. hat_share() is similar to hat_memload_array(),
2881  * except that we use the ism_hat's existing mappings to determine the pages
2882  * and protections to use for this hat. If we find a full properly aligned
2883  * and sized pagetable, we will attempt to share the pagetable itself.
2884  */
2885 /*ARGSUSED*/
2886 int
2887 hat_share(
2888 	hat_t		*hat,
2889 	caddr_t		addr,
2890 	hat_t		*ism_hat,
2891 	caddr_t		src_addr,
2892 	size_t		len,	/* almost useless value, see below.. */
2893 	uint_t		ismszc)
2894 {
2895 	uintptr_t	vaddr_start = (uintptr_t)addr;
2896 	uintptr_t	vaddr;
2897 	uintptr_t	eaddr = vaddr_start + len;
2898 	uintptr_t	ism_addr_start = (uintptr_t)src_addr;
2899 	uintptr_t	ism_addr = ism_addr_start;
2900 	uintptr_t	e_ism_addr = ism_addr + len;
2901 	htable_t	*ism_ht = NULL;
2902 	htable_t	*ht;
2903 	x86pte_t	pte;
2904 	page_t		*pp;
2905 	pfn_t		pfn;
2906 	level_t		l;
2907 	pgcnt_t		pgcnt;
2908 	uint_t		prot;
2909 	int		is_dism;
2910 	int		flags;
2911 
2912 	/*
2913 	 * We might be asked to share an empty DISM hat by as_dup()
2914 	 */
2915 	ASSERT(hat != kas.a_hat);
2916 	ASSERT(eaddr <= _userlimit);
2917 	if (!(ism_hat->hat_flags & HAT_SHARED)) {
2918 		ASSERT(hat_get_mapped_size(ism_hat) == 0);
2919 		return (0);
2920 	}
2921 	XPV_DISALLOW_MIGRATE();
2922 
2923 	/*
2924 	 * The SPT segment driver often passes us a size larger than there are
2925 	 * valid mappings. That's because it rounds the segment size up to a
2926 	 * large pagesize, even if the actual memory mapped by ism_hat is less.
2927 	 */
2928 	ASSERT(IS_PAGEALIGNED(vaddr_start));
2929 	ASSERT(IS_PAGEALIGNED(ism_addr_start));
2930 	ASSERT(ism_hat->hat_flags & HAT_SHARED);
2931 	is_dism = is_it_dism(hat, addr);
2932 	while (ism_addr < e_ism_addr) {
2933 		/*
2934 		 * use htable_walk to get the next valid ISM mapping
2935 		 */
2936 		pte = htable_walk(ism_hat, &ism_ht, &ism_addr, e_ism_addr);
2937 		if (ism_ht == NULL)
2938 			break;
2939 
2940 		/*
2941 		 * First check to see if we already share the page table.
2942 		 */
2943 		l = ism_ht->ht_level;
2944 		vaddr = vaddr_start + (ism_addr - ism_addr_start);
2945 		ht = htable_lookup(hat, vaddr, l);
2946 		if (ht != NULL) {
2947 			if (ht->ht_flags & HTABLE_SHARED_PFN)
2948 				goto shared;
2949 			htable_release(ht);
2950 			goto not_shared;
2951 		}
2952 
2953 		/*
2954 		 * Can't ever share top table.
2955 		 */
2956 		if (l == mmu.max_level)
2957 			goto not_shared;
2958 
2959 		/*
2960 		 * Avoid level mismatches later due to DISM faults.
2961 		 */
2962 		if (is_dism && l > 0)
2963 			goto not_shared;
2964 
2965 		/*
2966 		 * addresses and lengths must align
2967 		 * table must be fully populated
2968 		 * no lower level page tables
2969 		 */
2970 		if (ism_addr != ism_ht->ht_vaddr ||
2971 		    (vaddr & LEVEL_OFFSET(l + 1)) != 0)
2972 			goto not_shared;
2973 
2974 		/*
2975 		 * The range of address space must cover a full table.
2976 		 */
2977 		if (e_ism_addr - ism_addr < LEVEL_SIZE(l + 1))
2978 			goto not_shared;
2979 
2980 		/*
2981 		 * All entries in the ISM page table must be leaf PTEs.
2982 		 */
2983 		if (l > 0) {
2984 			int e;
2985 
2986 			/*
2987 			 * We know the 0th is from htable_walk() above.
2988 			 */
2989 			for (e = 1; e < HTABLE_NUM_PTES(ism_ht); ++e) {
2990 				x86pte_t pte;
2991 				pte = x86pte_get(ism_ht, e);
2992 				if (!PTE_ISPAGE(pte, l))
2993 					goto not_shared;
2994 			}
2995 		}
2996 
2997 		/*
2998 		 * share the page table
2999 		 */
3000 		ht = htable_create(hat, vaddr, l, ism_ht);
3001 shared:
3002 		ASSERT(ht->ht_flags & HTABLE_SHARED_PFN);
3003 		ASSERT(ht->ht_shares == ism_ht);
3004 		hat->hat_ism_pgcnt +=
3005 		    (ism_ht->ht_valid_cnt - ht->ht_valid_cnt) <<
3006 		    (LEVEL_SHIFT(ht->ht_level) - MMU_PAGESHIFT);
3007 		ht->ht_valid_cnt = ism_ht->ht_valid_cnt;
3008 		htable_release(ht);
3009 		ism_addr = ism_ht->ht_vaddr + LEVEL_SIZE(l + 1);
3010 		htable_release(ism_ht);
3011 		ism_ht = NULL;
3012 		continue;
3013 
3014 not_shared:
3015 		/*
3016 		 * Unable to share the page table. Instead we will
3017 		 * create new mappings from the values in the ISM mappings.
3018 		 * Figure out what level size mappings to use;
3019 		 */
3020 		for (l = ism_ht->ht_level; l > 0; --l) {
3021 			if (LEVEL_SIZE(l) <= eaddr - vaddr &&
3022 			    (vaddr & LEVEL_OFFSET(l)) == 0)
3023 				break;
3024 		}
3025 
3026 		/*
3027 		 * The ISM mapping might be larger than the share area,
3028 		 * be careful to truncate it if needed.
3029 		 */
3030 		if (eaddr - vaddr >= LEVEL_SIZE(ism_ht->ht_level)) {
3031 			pgcnt = mmu_btop(LEVEL_SIZE(ism_ht->ht_level));
3032 		} else {
3033 			pgcnt = mmu_btop(eaddr - vaddr);
3034 			l = 0;
3035 		}
3036 
3037 		pfn = PTE2PFN(pte, ism_ht->ht_level);
3038 		ASSERT(pfn != PFN_INVALID);
3039 		while (pgcnt > 0) {
3040 			/*
3041 			 * Make a new pte for the PFN for this level.
3042 			 * Copy protections for the pte from the ISM pte.
3043 			 */
3044 			pp = page_numtopp_nolock(pfn);
3045 			ASSERT(pp != NULL);
3046 
3047 			prot = PROT_USER | PROT_READ | HAT_UNORDERED_OK;
3048 			if (PTE_GET(pte, PT_WRITABLE))
3049 				prot |= PROT_WRITE;
3050 			if (!PTE_GET(pte, PT_NX))
3051 				prot |= PROT_EXEC;
3052 
3053 			flags = HAT_LOAD;
3054 			if (!is_dism)
3055 				flags |= HAT_LOAD_LOCK | HAT_LOAD_NOCONSIST;
3056 			while (hati_load_common(hat, vaddr, pp, prot, flags,
3057 			    l, pfn) != 0) {
3058 				if (l == 0)
3059 					panic("hati_load_common() failure");
3060 				--l;
3061 			}
3062 
3063 			vaddr += LEVEL_SIZE(l);
3064 			ism_addr += LEVEL_SIZE(l);
3065 			pfn += mmu_btop(LEVEL_SIZE(l));
3066 			pgcnt -= mmu_btop(LEVEL_SIZE(l));
3067 		}
3068 	}
3069 	if (ism_ht != NULL)
3070 		htable_release(ism_ht);
3071 	XPV_ALLOW_MIGRATE();
3072 	return (0);
3073 }
3074 
3075 
3076 /*
3077  * hat_unshare() is similar to hat_unload_callback(), but
3078  * we have to look for empty shared pagetables. Note that
3079  * hat_unshare() is always invoked against an entire segment.
3080  */
3081 /*ARGSUSED*/
3082 void
3083 hat_unshare(hat_t *hat, caddr_t addr, size_t len, uint_t ismszc)
3084 {
3085 	uint64_t	vaddr = (uintptr_t)addr;
3086 	uintptr_t	eaddr = vaddr + len;
3087 	htable_t	*ht = NULL;
3088 	uint_t		need_demaps = 0;
3089 	int		flags = HAT_UNLOAD_UNMAP;
3090 	level_t		l;
3091 
3092 	ASSERT(hat != kas.a_hat);
3093 	ASSERT(eaddr <= _userlimit);
3094 	ASSERT(IS_PAGEALIGNED(vaddr));
3095 	ASSERT(IS_PAGEALIGNED(eaddr));
3096 	XPV_DISALLOW_MIGRATE();
3097 
3098 	/*
3099 	 * First go through and remove any shared pagetables.
3100 	 *
3101 	 * Note that it's ok to delay the TLB shootdown till the entire range is
3102 	 * finished, because if hat_pageunload() were to unload a shared
3103 	 * pagetable page, its hat_tlb_inval() will do a global TLB invalidate.
3104 	 */
3105 	l = mmu.max_page_level;
3106 	if (l == mmu.max_level)
3107 		--l;
3108 	for (; l >= 0; --l) {
3109 		for (vaddr = (uintptr_t)addr; vaddr < eaddr;
3110 		    vaddr = (vaddr & LEVEL_MASK(l + 1)) + LEVEL_SIZE(l + 1)) {
3111 			ASSERT(!IN_VA_HOLE(vaddr));
3112 			/*
3113 			 * find a pagetable that maps the current address
3114 			 */
3115 			ht = htable_lookup(hat, vaddr, l);
3116 			if (ht == NULL)
3117 				continue;
3118 			if (ht->ht_flags & HTABLE_SHARED_PFN) {
3119 				/*
3120 				 * clear page count, set valid_cnt to 0,
3121 				 * let htable_release() finish the job
3122 				 */
3123 				hat->hat_ism_pgcnt -= ht->ht_valid_cnt <<
3124 				    (LEVEL_SHIFT(ht->ht_level) - MMU_PAGESHIFT);
3125 				ht->ht_valid_cnt = 0;
3126 				need_demaps = 1;
3127 			}
3128 			htable_release(ht);
3129 		}
3130 	}
3131 
3132 	/*
3133 	 * flush the TLBs - since we're probably dealing with MANY mappings
3134 	 * we do just one CR3 reload.
3135 	 */
3136 	if (!(hat->hat_flags & HAT_FREEING) && need_demaps)
3137 		hat_tlb_inval(hat, DEMAP_ALL_ADDR);
3138 
3139 	/*
3140 	 * Now go back and clean up any unaligned mappings that
3141 	 * couldn't share pagetables.
3142 	 */
3143 	if (!is_it_dism(hat, addr))
3144 		flags |= HAT_UNLOAD_UNLOCK;
3145 	hat_unload(hat, addr, len, flags);
3146 	XPV_ALLOW_MIGRATE();
3147 }
3148 
3149 
3150 /*
3151  * hat_reserve() does nothing
3152  */
3153 /*ARGSUSED*/
3154 void
3155 hat_reserve(struct as *as, caddr_t addr, size_t len)
3156 {
3157 }
3158 
3159 
3160 /*
3161  * Called when all mappings to a page should have write permission removed.
3162  * Mostly stolem from hat_pagesync()
3163  */
3164 static void
3165 hati_page_clrwrt(struct page *pp)
3166 {
3167 	hment_t		*hm = NULL;
3168 	htable_t	*ht;
3169 	uint_t		entry;
3170 	x86pte_t	old;
3171 	x86pte_t	new;
3172 	uint_t		pszc = 0;
3173 
3174 	XPV_DISALLOW_MIGRATE();
3175 next_size:
3176 	/*
3177 	 * walk thru the mapping list clearing write permission
3178 	 */
3179 	x86_hm_enter(pp);
3180 	while ((hm = hment_walk(pp, &ht, &entry, hm)) != NULL) {
3181 		if (ht->ht_level < pszc)
3182 			continue;
3183 		old = x86pte_get(ht, entry);
3184 
3185 		for (;;) {
3186 			/*
3187 			 * Is this mapping of interest?
3188 			 */
3189 			if (PTE2PFN(old, ht->ht_level) != pp->p_pagenum ||
3190 			    PTE_GET(old, PT_WRITABLE) == 0)
3191 				break;
3192 
3193 			/*
3194 			 * Clear ref/mod writable bits. This requires cross
3195 			 * calls to ensure any executing TLBs see cleared bits.
3196 			 */
3197 			new = old;
3198 			PTE_CLR(new, PT_REF | PT_MOD | PT_WRITABLE);
3199 			old = hati_update_pte(ht, entry, old, new);
3200 			if (old != 0)
3201 				continue;
3202 
3203 			break;
3204 		}
3205 	}
3206 	x86_hm_exit(pp);
3207 	while (pszc < pp->p_szc) {
3208 		page_t *tpp;
3209 		pszc++;
3210 		tpp = PP_GROUPLEADER(pp, pszc);
3211 		if (pp != tpp) {
3212 			pp = tpp;
3213 			goto next_size;
3214 		}
3215 	}
3216 	XPV_ALLOW_MIGRATE();
3217 }
3218 
3219 /*
3220  * void hat_page_setattr(pp, flag)
3221  * void hat_page_clrattr(pp, flag)
3222  *	used to set/clr ref/mod bits.
3223  */
3224 void
3225 hat_page_setattr(struct page *pp, uint_t flag)
3226 {
3227 	vnode_t		*vp = pp->p_vnode;
3228 	kmutex_t	*vphm = NULL;
3229 	page_t		**listp;
3230 	int		noshuffle;
3231 
3232 	noshuffle = flag & P_NSH;
3233 	flag &= ~P_NSH;
3234 
3235 	if (PP_GETRM(pp, flag) == flag)
3236 		return;
3237 
3238 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) &&
3239 	    !noshuffle) {
3240 		vphm = page_vnode_mutex(vp);
3241 		mutex_enter(vphm);
3242 	}
3243 
3244 	PP_SETRM(pp, flag);
3245 
3246 	if (vphm != NULL) {
3247 
3248 		/*
3249 		 * Some File Systems examine v_pages for NULL w/o
3250 		 * grabbing the vphm mutex. Must not let it become NULL when
3251 		 * pp is the only page on the list.
3252 		 */
3253 		if (pp->p_vpnext != pp) {
3254 			page_vpsub(&vp->v_pages, pp);
3255 			if (vp->v_pages != NULL)
3256 				listp = &vp->v_pages->p_vpprev->p_vpnext;
3257 			else
3258 				listp = &vp->v_pages;
3259 			page_vpadd(listp, pp);
3260 		}
3261 		mutex_exit(vphm);
3262 	}
3263 }
3264 
3265 void
3266 hat_page_clrattr(struct page *pp, uint_t flag)
3267 {
3268 	vnode_t		*vp = pp->p_vnode;
3269 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
3270 
3271 	/*
3272 	 * Caller is expected to hold page's io lock for VMODSORT to work
3273 	 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod
3274 	 * bit is cleared.
3275 	 * We don't have assert to avoid tripping some existing third party
3276 	 * code. The dirty page is moved back to top of the v_page list
3277 	 * after IO is done in pvn_write_done().
3278 	 */
3279 	PP_CLRRM(pp, flag);
3280 
3281 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
3282 
3283 		/*
3284 		 * VMODSORT works by removing write permissions and getting
3285 		 * a fault when a page is made dirty. At this point
3286 		 * we need to remove write permission from all mappings
3287 		 * to this page.
3288 		 */
3289 		hati_page_clrwrt(pp);
3290 	}
3291 }
3292 
3293 /*
3294  *	If flag is specified, returns 0 if attribute is disabled
3295  *	and non zero if enabled.  If flag specifes multiple attributs
3296  *	then returns 0 if ALL atriibutes are disabled.  This is an advisory
3297  *	call.
3298  */
3299 uint_t
3300 hat_page_getattr(struct page *pp, uint_t flag)
3301 {
3302 	return (PP_GETRM(pp, flag));
3303 }
3304 
3305 
3306 /*
3307  * common code used by hat_pageunload() and hment_steal()
3308  */
3309 hment_t *
3310 hati_page_unmap(page_t *pp, htable_t *ht, uint_t entry)
3311 {
3312 	x86pte_t old_pte;
3313 	pfn_t pfn = pp->p_pagenum;
3314 	hment_t *hm;
3315 
3316 	/*
3317 	 * We need to acquire a hold on the htable in order to
3318 	 * do the invalidate. We know the htable must exist, since
3319 	 * unmap's don't release the htable until after removing any
3320 	 * hment. Having x86_hm_enter() keeps that from proceeding.
3321 	 */
3322 	htable_acquire(ht);
3323 
3324 	/*
3325 	 * Invalidate the PTE and remove the hment.
3326 	 */
3327 	old_pte = x86pte_inval(ht, entry, 0, NULL);
3328 	if (PTE2PFN(old_pte, ht->ht_level) != pfn) {
3329 		panic("x86pte_inval() failure found PTE = " FMT_PTE
3330 		    " pfn being unmapped is %lx ht=0x%lx entry=0x%x",
3331 		    old_pte, pfn, (uintptr_t)ht, entry);
3332 	}
3333 
3334 	/*
3335 	 * Clean up all the htable information for this mapping
3336 	 */
3337 	ASSERT(ht->ht_valid_cnt > 0);
3338 	HTABLE_DEC(ht->ht_valid_cnt);
3339 	PGCNT_DEC(ht->ht_hat, ht->ht_level);
3340 
3341 	/*
3342 	 * sync ref/mod bits to the page_t
3343 	 */
3344 	if (PTE_GET(old_pte, PT_SOFTWARE) < PT_NOSYNC)
3345 		hati_sync_pte_to_page(pp, old_pte, ht->ht_level);
3346 
3347 	/*
3348 	 * Remove the mapping list entry for this page.
3349 	 */
3350 	hm = hment_remove(pp, ht, entry);
3351 
3352 	/*
3353 	 * drop the mapping list lock so that we might free the
3354 	 * hment and htable.
3355 	 */
3356 	x86_hm_exit(pp);
3357 	htable_release(ht);
3358 	return (hm);
3359 }
3360 
3361 extern int	vpm_enable;
3362 /*
3363  * Unload all translations to a page. If the page is a subpage of a large
3364  * page, the large page mappings are also removed.
3365  *
3366  * The forceflags are unused.
3367  */
3368 
3369 /*ARGSUSED*/
3370 static int
3371 hati_pageunload(struct page *pp, uint_t pg_szcd, uint_t forceflag)
3372 {
3373 	page_t		*cur_pp = pp;
3374 	hment_t		*hm;
3375 	hment_t		*prev;
3376 	htable_t	*ht;
3377 	uint_t		entry;
3378 	level_t		level;
3379 
3380 	XPV_DISALLOW_MIGRATE();
3381 #if defined(__amd64)
3382 	/*
3383 	 * clear the vpm ref.
3384 	 */
3385 	if (vpm_enable) {
3386 		pp->p_vpmref = 0;
3387 	}
3388 #endif
3389 	/*
3390 	 * The loop with next_size handles pages with multiple pagesize mappings
3391 	 */
3392 next_size:
3393 	for (;;) {
3394 
3395 		/*
3396 		 * Get a mapping list entry
3397 		 */
3398 		x86_hm_enter(cur_pp);
3399 		for (prev = NULL; ; prev = hm) {
3400 			hm = hment_walk(cur_pp, &ht, &entry, prev);
3401 			if (hm == NULL) {
3402 				x86_hm_exit(cur_pp);
3403 
3404 				/*
3405 				 * If not part of a larger page, we're done.
3406 				 */
3407 				if (cur_pp->p_szc <= pg_szcd) {
3408 					XPV_ALLOW_MIGRATE();
3409 					return (0);
3410 				}
3411 
3412 				/*
3413 				 * Else check the next larger page size.
3414 				 * hat_page_demote() may decrease p_szc
3415 				 * but that's ok we'll just take an extra
3416 				 * trip discover there're no larger mappings
3417 				 * and return.
3418 				 */
3419 				++pg_szcd;
3420 				cur_pp = PP_GROUPLEADER(cur_pp, pg_szcd);
3421 				goto next_size;
3422 			}
3423 
3424 			/*
3425 			 * If this mapping size matches, remove it.
3426 			 */
3427 			level = ht->ht_level;
3428 			if (level == pg_szcd)
3429 				break;
3430 		}
3431 
3432 		/*
3433 		 * Remove the mapping list entry for this page.
3434 		 * Note this does the x86_hm_exit() for us.
3435 		 */
3436 		hm = hati_page_unmap(cur_pp, ht, entry);
3437 		if (hm != NULL)
3438 			hment_free(hm);
3439 	}
3440 }
3441 
3442 int
3443 hat_pageunload(struct page *pp, uint_t forceflag)
3444 {
3445 	ASSERT(PAGE_EXCL(pp));
3446 	return (hati_pageunload(pp, 0, forceflag));
3447 }
3448 
3449 /*
3450  * Unload all large mappings to pp and reduce by 1 p_szc field of every large
3451  * page level that included pp.
3452  *
3453  * pp must be locked EXCL. Even though no other constituent pages are locked
3454  * it's legal to unload large mappings to pp because all constituent pages of
3455  * large locked mappings have to be locked SHARED.  therefore if we have EXCL
3456  * lock on one of constituent pages none of the large mappings to pp are
3457  * locked.
3458  *
3459  * Change (always decrease) p_szc field starting from the last constituent
3460  * page and ending with root constituent page so that root's pszc always shows
3461  * the area where hat_page_demote() may be active.
3462  *
3463  * This mechanism is only used for file system pages where it's not always
3464  * possible to get EXCL locks on all constituent pages to demote the size code
3465  * (as is done for anonymous or kernel large pages).
3466  */
3467 void
3468 hat_page_demote(page_t *pp)
3469 {
3470 	uint_t		pszc;
3471 	uint_t		rszc;
3472 	uint_t		szc;
3473 	page_t		*rootpp;
3474 	page_t		*firstpp;
3475 	page_t		*lastpp;
3476 	pgcnt_t		pgcnt;
3477 
3478 	ASSERT(PAGE_EXCL(pp));
3479 	ASSERT(!PP_ISFREE(pp));
3480 	ASSERT(page_szc_lock_assert(pp));
3481 
3482 	if (pp->p_szc == 0)
3483 		return;
3484 
3485 	rootpp = PP_GROUPLEADER(pp, 1);
3486 	(void) hati_pageunload(rootpp, 1, HAT_FORCE_PGUNLOAD);
3487 
3488 	/*
3489 	 * all large mappings to pp are gone
3490 	 * and no new can be setup since pp is locked exclusively.
3491 	 *
3492 	 * Lock the root to make sure there's only one hat_page_demote()
3493 	 * outstanding within the area of this root's pszc.
3494 	 *
3495 	 * Second potential hat_page_demote() is already eliminated by upper
3496 	 * VM layer via page_szc_lock() but we don't rely on it and use our
3497 	 * own locking (so that upper layer locking can be changed without
3498 	 * assumptions that hat depends on upper layer VM to prevent multiple
3499 	 * hat_page_demote() to be issued simultaneously to the same large
3500 	 * page).
3501 	 */
3502 again:
3503 	pszc = pp->p_szc;
3504 	if (pszc == 0)
3505 		return;
3506 	rootpp = PP_GROUPLEADER(pp, pszc);
3507 	x86_hm_enter(rootpp);
3508 	/*
3509 	 * If root's p_szc is different from pszc we raced with another
3510 	 * hat_page_demote().  Drop the lock and try to find the root again.
3511 	 * If root's p_szc is greater than pszc previous hat_page_demote() is
3512 	 * not done yet.  Take and release mlist lock of root's root to wait
3513 	 * for previous hat_page_demote() to complete.
3514 	 */
3515 	if ((rszc = rootpp->p_szc) != pszc) {
3516 		x86_hm_exit(rootpp);
3517 		if (rszc > pszc) {
3518 			/* p_szc of a locked non free page can't increase */
3519 			ASSERT(pp != rootpp);
3520 
3521 			rootpp = PP_GROUPLEADER(rootpp, rszc);
3522 			x86_hm_enter(rootpp);
3523 			x86_hm_exit(rootpp);
3524 		}
3525 		goto again;
3526 	}
3527 	ASSERT(pp->p_szc == pszc);
3528 
3529 	/*
3530 	 * Decrement by 1 p_szc of every constituent page of a region that
3531 	 * covered pp. For example if original szc is 3 it gets changed to 2
3532 	 * everywhere except in region 2 that covered pp. Region 2 that
3533 	 * covered pp gets demoted to 1 everywhere except in region 1 that
3534 	 * covered pp. The region 1 that covered pp is demoted to region
3535 	 * 0. It's done this way because from region 3 we removed level 3
3536 	 * mappings, from region 2 that covered pp we removed level 2 mappings
3537 	 * and from region 1 that covered pp we removed level 1 mappings.  All
3538 	 * changes are done from from high pfn's to low pfn's so that roots
3539 	 * are changed last allowing one to know the largest region where
3540 	 * hat_page_demote() is stil active by only looking at the root page.
3541 	 *
3542 	 * This algorithm is implemented in 2 while loops. First loop changes
3543 	 * p_szc of pages to the right of pp's level 1 region and second
3544 	 * loop changes p_szc of pages of level 1 region that covers pp
3545 	 * and all pages to the left of level 1 region that covers pp.
3546 	 * In the first loop p_szc keeps dropping with every iteration
3547 	 * and in the second loop it keeps increasing with every iteration.
3548 	 *
3549 	 * First loop description: Demote pages to the right of pp outside of
3550 	 * level 1 region that covers pp.  In every iteration of the while
3551 	 * loop below find the last page of szc region and the first page of
3552 	 * (szc - 1) region that is immediately to the right of (szc - 1)
3553 	 * region that covers pp.  From last such page to first such page
3554 	 * change every page's szc to szc - 1. Decrement szc and continue
3555 	 * looping until szc is 1. If pp belongs to the last (szc - 1) region
3556 	 * of szc region skip to the next iteration.
3557 	 */
3558 	szc = pszc;
3559 	while (szc > 1) {
3560 		lastpp = PP_GROUPLEADER(pp, szc);
3561 		pgcnt = page_get_pagecnt(szc);
3562 		lastpp += pgcnt - 1;
3563 		firstpp = PP_GROUPLEADER(pp, (szc - 1));
3564 		pgcnt = page_get_pagecnt(szc - 1);
3565 		if (lastpp - firstpp < pgcnt) {
3566 			szc--;
3567 			continue;
3568 		}
3569 		firstpp += pgcnt;
3570 		while (lastpp != firstpp) {
3571 			ASSERT(lastpp->p_szc == pszc);
3572 			lastpp->p_szc = szc - 1;
3573 			lastpp--;
3574 		}
3575 		firstpp->p_szc = szc - 1;
3576 		szc--;
3577 	}
3578 
3579 	/*
3580 	 * Second loop description:
3581 	 * First iteration changes p_szc to 0 of every
3582 	 * page of level 1 region that covers pp.
3583 	 * Subsequent iterations find last page of szc region
3584 	 * immediately to the left of szc region that covered pp
3585 	 * and first page of (szc + 1) region that covers pp.
3586 	 * From last to first page change p_szc of every page to szc.
3587 	 * Increment szc and continue looping until szc is pszc.
3588 	 * If pp belongs to the fist szc region of (szc + 1) region
3589 	 * skip to the next iteration.
3590 	 *
3591 	 */
3592 	szc = 0;
3593 	while (szc < pszc) {
3594 		firstpp = PP_GROUPLEADER(pp, (szc + 1));
3595 		if (szc == 0) {
3596 			pgcnt = page_get_pagecnt(1);
3597 			lastpp = firstpp + (pgcnt - 1);
3598 		} else {
3599 			lastpp = PP_GROUPLEADER(pp, szc);
3600 			if (firstpp == lastpp) {
3601 				szc++;
3602 				continue;
3603 			}
3604 			lastpp--;
3605 			pgcnt = page_get_pagecnt(szc);
3606 		}
3607 		while (lastpp != firstpp) {
3608 			ASSERT(lastpp->p_szc == pszc);
3609 			lastpp->p_szc = szc;
3610 			lastpp--;
3611 		}
3612 		firstpp->p_szc = szc;
3613 		if (firstpp == rootpp)
3614 			break;
3615 		szc++;
3616 	}
3617 	x86_hm_exit(rootpp);
3618 }
3619 
3620 /*
3621  * get hw stats from hardware into page struct and reset hw stats
3622  * returns attributes of page
3623  * Flags for hat_pagesync, hat_getstat, hat_sync
3624  *
3625  * define	HAT_SYNC_ZERORM		0x01
3626  *
3627  * Additional flags for hat_pagesync
3628  *
3629  * define	HAT_SYNC_STOPON_REF	0x02
3630  * define	HAT_SYNC_STOPON_MOD	0x04
3631  * define	HAT_SYNC_STOPON_RM	0x06
3632  * define	HAT_SYNC_STOPON_SHARED	0x08
3633  */
3634 uint_t
3635 hat_pagesync(struct page *pp, uint_t flags)
3636 {
3637 	hment_t		*hm = NULL;
3638 	htable_t	*ht;
3639 	uint_t		entry;
3640 	x86pte_t	old, save_old;
3641 	x86pte_t	new;
3642 	uchar_t		nrmbits = P_REF|P_MOD|P_RO;
3643 	extern ulong_t	po_share;
3644 	page_t		*save_pp = pp;
3645 	uint_t		pszc = 0;
3646 
3647 	ASSERT(PAGE_LOCKED(pp) || panicstr);
3648 
3649 	if (PP_ISRO(pp) && (flags & HAT_SYNC_STOPON_MOD))
3650 		return (pp->p_nrm & nrmbits);
3651 
3652 	if ((flags & HAT_SYNC_ZERORM) == 0) {
3653 
3654 		if ((flags & HAT_SYNC_STOPON_REF) != 0 && PP_ISREF(pp))
3655 			return (pp->p_nrm & nrmbits);
3656 
3657 		if ((flags & HAT_SYNC_STOPON_MOD) != 0 && PP_ISMOD(pp))
3658 			return (pp->p_nrm & nrmbits);
3659 
3660 		if ((flags & HAT_SYNC_STOPON_SHARED) != 0 &&
3661 		    hat_page_getshare(pp) > po_share) {
3662 			if (PP_ISRO(pp))
3663 				PP_SETREF(pp);
3664 			return (pp->p_nrm & nrmbits);
3665 		}
3666 	}
3667 
3668 	XPV_DISALLOW_MIGRATE();
3669 next_size:
3670 	/*
3671 	 * walk thru the mapping list syncing (and clearing) ref/mod bits.
3672 	 */
3673 	x86_hm_enter(pp);
3674 	while ((hm = hment_walk(pp, &ht, &entry, hm)) != NULL) {
3675 		if (ht->ht_level < pszc)
3676 			continue;
3677 		old = x86pte_get(ht, entry);
3678 try_again:
3679 
3680 		ASSERT(PTE2PFN(old, ht->ht_level) == pp->p_pagenum);
3681 
3682 		if (PTE_GET(old, PT_REF | PT_MOD) == 0)
3683 			continue;
3684 
3685 		save_old = old;
3686 		if ((flags & HAT_SYNC_ZERORM) != 0) {
3687 
3688 			/*
3689 			 * Need to clear ref or mod bits. Need to demap
3690 			 * to make sure any executing TLBs see cleared bits.
3691 			 */
3692 			new = old;
3693 			PTE_CLR(new, PT_REF | PT_MOD);
3694 			old = hati_update_pte(ht, entry, old, new);
3695 			if (old != 0)
3696 				goto try_again;
3697 
3698 			old = save_old;
3699 		}
3700 
3701 		/*
3702 		 * Sync the PTE
3703 		 */
3704 		if (!(flags & HAT_SYNC_ZERORM) &&
3705 		    PTE_GET(old, PT_SOFTWARE) <= PT_NOSYNC)
3706 			hati_sync_pte_to_page(pp, old, ht->ht_level);
3707 
3708 		/*
3709 		 * can stop short if we found a ref'd or mod'd page
3710 		 */
3711 		if ((flags & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp) ||
3712 		    (flags & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)) {
3713 			x86_hm_exit(pp);
3714 			goto done;
3715 		}
3716 	}
3717 	x86_hm_exit(pp);
3718 	while (pszc < pp->p_szc) {
3719 		page_t *tpp;
3720 		pszc++;
3721 		tpp = PP_GROUPLEADER(pp, pszc);
3722 		if (pp != tpp) {
3723 			pp = tpp;
3724 			goto next_size;
3725 		}
3726 	}
3727 done:
3728 	XPV_ALLOW_MIGRATE();
3729 	return (save_pp->p_nrm & nrmbits);
3730 }
3731 
3732 /*
3733  * returns approx number of mappings to this pp.  A return of 0 implies
3734  * there are no mappings to the page.
3735  */
3736 ulong_t
3737 hat_page_getshare(page_t *pp)
3738 {
3739 	uint_t cnt;
3740 	cnt = hment_mapcnt(pp);
3741 #if defined(__amd64)
3742 	if (vpm_enable && pp->p_vpmref) {
3743 		cnt += 1;
3744 	}
3745 #endif
3746 	return (cnt);
3747 }
3748 
3749 /*
3750  * Return 1 the number of mappings exceeds sh_thresh. Return 0
3751  * otherwise.
3752  */
3753 int
3754 hat_page_checkshare(page_t *pp, ulong_t sh_thresh)
3755 {
3756 	return (hat_page_getshare(pp) > sh_thresh);
3757 }
3758 
3759 /*
3760  * hat_softlock isn't supported anymore
3761  */
3762 /*ARGSUSED*/
3763 faultcode_t
3764 hat_softlock(
3765 	hat_t *hat,
3766 	caddr_t addr,
3767 	size_t *len,
3768 	struct page **page_array,
3769 	uint_t flags)
3770 {
3771 	return (FC_NOSUPPORT);
3772 }
3773 
3774 
3775 
3776 /*
3777  * Routine to expose supported HAT features to platform independent code.
3778  */
3779 /*ARGSUSED*/
3780 int
3781 hat_supported(enum hat_features feature, void *arg)
3782 {
3783 	switch (feature) {
3784 
3785 	case HAT_SHARED_PT:	/* this is really ISM */
3786 		return (1);
3787 
3788 	case HAT_DYNAMIC_ISM_UNMAP:
3789 		return (0);
3790 
3791 	case HAT_VMODSORT:
3792 		return (1);
3793 
3794 	case HAT_SHARED_REGIONS:
3795 		return (0);
3796 
3797 	default:
3798 		panic("hat_supported() - unknown feature");
3799 	}
3800 	return (0);
3801 }
3802 
3803 /*
3804  * Called when a thread is exiting and has been switched to the kernel AS
3805  */
3806 void
3807 hat_thread_exit(kthread_t *thd)
3808 {
3809 	ASSERT(thd->t_procp->p_as == &kas);
3810 	XPV_DISALLOW_MIGRATE();
3811 	hat_switch(thd->t_procp->p_as->a_hat);
3812 	XPV_ALLOW_MIGRATE();
3813 }
3814 
3815 /*
3816  * Setup the given brand new hat structure as the new HAT on this cpu's mmu.
3817  */
3818 /*ARGSUSED*/
3819 void
3820 hat_setup(hat_t *hat, int flags)
3821 {
3822 	XPV_DISALLOW_MIGRATE();
3823 	kpreempt_disable();
3824 
3825 	hat_switch(hat);
3826 
3827 	kpreempt_enable();
3828 	XPV_ALLOW_MIGRATE();
3829 }
3830 
3831 /*
3832  * Prepare for a CPU private mapping for the given address.
3833  *
3834  * The address can only be used from a single CPU and can be remapped
3835  * using hat_mempte_remap().  Return the address of the PTE.
3836  *
3837  * We do the htable_create() if necessary and increment the valid count so
3838  * the htable can't disappear.  We also hat_devload() the page table into
3839  * kernel so that the PTE is quickly accessed.
3840  */
3841 hat_mempte_t
3842 hat_mempte_setup(caddr_t addr)
3843 {
3844 	uintptr_t	va = (uintptr_t)addr;
3845 	htable_t	*ht;
3846 	uint_t		entry;
3847 	x86pte_t	oldpte;
3848 	hat_mempte_t	p;
3849 
3850 	ASSERT(IS_PAGEALIGNED(va));
3851 	ASSERT(!IN_VA_HOLE(va));
3852 	++curthread->t_hatdepth;
3853 	XPV_DISALLOW_MIGRATE();
3854 	ht = htable_getpte(kas.a_hat, va, &entry, &oldpte, 0);
3855 	if (ht == NULL) {
3856 		ht = htable_create(kas.a_hat, va, 0, NULL);
3857 		entry = htable_va2entry(va, ht);
3858 		ASSERT(ht->ht_level == 0);
3859 		oldpte = x86pte_get(ht, entry);
3860 	}
3861 	if (PTE_ISVALID(oldpte))
3862 		panic("hat_mempte_setup(): address already mapped"
3863 		    "ht=%p, entry=%d, pte=" FMT_PTE, (void *)ht, entry, oldpte);
3864 
3865 	/*
3866 	 * increment ht_valid_cnt so that the pagetable can't disappear
3867 	 */
3868 	HTABLE_INC(ht->ht_valid_cnt);
3869 
3870 	/*
3871 	 * return the PTE physical address to the caller.
3872 	 */
3873 	htable_release(ht);
3874 	XPV_ALLOW_MIGRATE();
3875 	p = PT_INDEX_PHYSADDR(pfn_to_pa(ht->ht_pfn), entry);
3876 	--curthread->t_hatdepth;
3877 	return (p);
3878 }
3879 
3880 /*
3881  * Release a CPU private mapping for the given address.
3882  * We decrement the htable valid count so it might be destroyed.
3883  */
3884 /*ARGSUSED1*/
3885 void
3886 hat_mempte_release(caddr_t addr, hat_mempte_t pte_pa)
3887 {
3888 	htable_t	*ht;
3889 
3890 	XPV_DISALLOW_MIGRATE();
3891 	/*
3892 	 * invalidate any left over mapping and decrement the htable valid count
3893 	 */
3894 #ifdef __xpv
3895 	if (HYPERVISOR_update_va_mapping((uintptr_t)addr, 0,
3896 	    UVMF_INVLPG | UVMF_LOCAL))
3897 		panic("HYPERVISOR_update_va_mapping() failed");
3898 #else
3899 	{
3900 		x86pte_t *pteptr;
3901 
3902 		pteptr = x86pte_mapin(mmu_btop(pte_pa),
3903 		    (pte_pa & MMU_PAGEOFFSET) >> mmu.pte_size_shift, NULL);
3904 		if (mmu.pae_hat)
3905 			*pteptr = 0;
3906 		else
3907 			*(x86pte32_t *)pteptr = 0;
3908 		mmu_tlbflush_entry(addr);
3909 		x86pte_mapout();
3910 	}
3911 #endif
3912 
3913 	ht = htable_getpte(kas.a_hat, ALIGN2PAGE(addr), NULL, NULL, 0);
3914 	if (ht == NULL)
3915 		panic("hat_mempte_release(): invalid address");
3916 	ASSERT(ht->ht_level == 0);
3917 	HTABLE_DEC(ht->ht_valid_cnt);
3918 	htable_release(ht);
3919 	XPV_ALLOW_MIGRATE();
3920 }
3921 
3922 /*
3923  * Apply a temporary CPU private mapping to a page. We flush the TLB only
3924  * on this CPU, so this ought to have been called with preemption disabled.
3925  */
3926 void
3927 hat_mempte_remap(
3928 	pfn_t		pfn,
3929 	caddr_t		addr,
3930 	hat_mempte_t	pte_pa,
3931 	uint_t		attr,
3932 	uint_t		flags)
3933 {
3934 	uintptr_t	va = (uintptr_t)addr;
3935 	x86pte_t	pte;
3936 
3937 	/*
3938 	 * Remap the given PTE to the new page's PFN. Invalidate only
3939 	 * on this CPU.
3940 	 */
3941 #ifdef DEBUG
3942 	htable_t	*ht;
3943 	uint_t		entry;
3944 
3945 	ASSERT(IS_PAGEALIGNED(va));
3946 	ASSERT(!IN_VA_HOLE(va));
3947 	ht = htable_getpte(kas.a_hat, va, &entry, NULL, 0);
3948 	ASSERT(ht != NULL);
3949 	ASSERT(ht->ht_level == 0);
3950 	ASSERT(ht->ht_valid_cnt > 0);
3951 	ASSERT(ht->ht_pfn == mmu_btop(pte_pa));
3952 	htable_release(ht);
3953 #endif
3954 	XPV_DISALLOW_MIGRATE();
3955 	pte = hati_mkpte(pfn, attr, 0, flags);
3956 #ifdef __xpv
3957 	if (HYPERVISOR_update_va_mapping(va, pte, UVMF_INVLPG | UVMF_LOCAL))
3958 		panic("HYPERVISOR_update_va_mapping() failed");
3959 #else
3960 	{
3961 		x86pte_t *pteptr;
3962 
3963 		pteptr = x86pte_mapin(mmu_btop(pte_pa),
3964 		    (pte_pa & MMU_PAGEOFFSET) >> mmu.pte_size_shift, NULL);
3965 		if (mmu.pae_hat)
3966 			*(x86pte_t *)pteptr = pte;
3967 		else
3968 			*(x86pte32_t *)pteptr = (x86pte32_t)pte;
3969 		mmu_tlbflush_entry(addr);
3970 		x86pte_mapout();
3971 	}
3972 #endif
3973 	XPV_ALLOW_MIGRATE();
3974 }
3975 
3976 
3977 
3978 /*
3979  * Hat locking functions
3980  * XXX - these two functions are currently being used by hatstats
3981  * 	they can be removed by using a per-as mutex for hatstats.
3982  */
3983 void
3984 hat_enter(hat_t *hat)
3985 {
3986 	mutex_enter(&hat->hat_mutex);
3987 }
3988 
3989 void
3990 hat_exit(hat_t *hat)
3991 {
3992 	mutex_exit(&hat->hat_mutex);
3993 }
3994 
3995 /*
3996  * HAT part of cpu initialization.
3997  */
3998 void
3999 hat_cpu_online(struct cpu *cpup)
4000 {
4001 	if (cpup != CPU) {
4002 		x86pte_cpu_init(cpup);
4003 		hat_vlp_setup(cpup);
4004 	}
4005 	CPUSET_ATOMIC_ADD(khat_cpuset, cpup->cpu_id);
4006 }
4007 
4008 /*
4009  * HAT part of cpu deletion.
4010  * (currently, we only call this after the cpu is safely passivated.)
4011  */
4012 void
4013 hat_cpu_offline(struct cpu *cpup)
4014 {
4015 	ASSERT(cpup != CPU);
4016 
4017 	CPUSET_ATOMIC_DEL(khat_cpuset, cpup->cpu_id);
4018 	x86pte_cpu_fini(cpup);
4019 	hat_vlp_teardown(cpup);
4020 }
4021 
4022 /*
4023  * Function called after all CPUs are brought online.
4024  * Used to remove low address boot mappings.
4025  */
4026 void
4027 clear_boot_mappings(uintptr_t low, uintptr_t high)
4028 {
4029 	uintptr_t vaddr = low;
4030 	htable_t *ht = NULL;
4031 	level_t level;
4032 	uint_t entry;
4033 	x86pte_t pte;
4034 
4035 	/*
4036 	 * On 1st CPU we can unload the prom mappings, basically we blow away
4037 	 * all virtual mappings under _userlimit.
4038 	 */
4039 	while (vaddr < high) {
4040 		pte = htable_walk(kas.a_hat, &ht, &vaddr, high);
4041 		if (ht == NULL)
4042 			break;
4043 
4044 		level = ht->ht_level;
4045 		entry = htable_va2entry(vaddr, ht);
4046 		ASSERT(level <= mmu.max_page_level);
4047 		ASSERT(PTE_ISPAGE(pte, level));
4048 
4049 		/*
4050 		 * Unload the mapping from the page tables.
4051 		 */
4052 		(void) x86pte_inval(ht, entry, 0, NULL);
4053 		ASSERT(ht->ht_valid_cnt > 0);
4054 		HTABLE_DEC(ht->ht_valid_cnt);
4055 		PGCNT_DEC(ht->ht_hat, ht->ht_level);
4056 
4057 		vaddr += LEVEL_SIZE(ht->ht_level);
4058 	}
4059 	if (ht)
4060 		htable_release(ht);
4061 }
4062 
4063 /*
4064  * Atomically update a new translation for a single page.  If the
4065  * currently installed PTE doesn't match the value we expect to find,
4066  * it's not updated and we return the PTE we found.
4067  *
4068  * If activating nosync or NOWRITE and the page was modified we need to sync
4069  * with the page_t. Also sync with page_t if clearing ref/mod bits.
4070  */
4071 static x86pte_t
4072 hati_update_pte(htable_t *ht, uint_t entry, x86pte_t expected, x86pte_t new)
4073 {
4074 	page_t		*pp;
4075 	uint_t		rm = 0;
4076 	x86pte_t	replaced;
4077 
4078 	if (PTE_GET(expected, PT_SOFTWARE) < PT_NOSYNC &&
4079 	    PTE_GET(expected, PT_MOD | PT_REF) &&
4080 	    (PTE_GET(new, PT_NOSYNC) || !PTE_GET(new, PT_WRITABLE) ||
4081 	    !PTE_GET(new, PT_MOD | PT_REF))) {
4082 
4083 		ASSERT(!pfn_is_foreign(PTE2PFN(expected, ht->ht_level)));
4084 		pp = page_numtopp_nolock(PTE2PFN(expected, ht->ht_level));
4085 		ASSERT(pp != NULL);
4086 		if (PTE_GET(expected, PT_MOD))
4087 			rm |= P_MOD;
4088 		if (PTE_GET(expected, PT_REF))
4089 			rm |= P_REF;
4090 		PTE_CLR(new, PT_MOD | PT_REF);
4091 	}
4092 
4093 	replaced = x86pte_update(ht, entry, expected, new);
4094 	if (replaced != expected)
4095 		return (replaced);
4096 
4097 	if (rm) {
4098 		/*
4099 		 * sync to all constituent pages of a large page
4100 		 */
4101 		pgcnt_t pgcnt = page_get_pagecnt(ht->ht_level);
4102 		ASSERT(IS_P2ALIGNED(pp->p_pagenum, pgcnt));
4103 		while (pgcnt-- > 0) {
4104 			/*
4105 			 * hat_page_demote() can't decrease
4106 			 * pszc below this mapping size
4107 			 * since large mapping existed after we
4108 			 * took mlist lock.
4109 			 */
4110 			ASSERT(pp->p_szc >= ht->ht_level);
4111 			hat_page_setattr(pp, rm);
4112 			++pp;
4113 		}
4114 	}
4115 
4116 	return (0);
4117 }
4118 
4119 /* ARGSUSED */
4120 void
4121 hat_join_srd(struct hat *hat, vnode_t *evp)
4122 {
4123 }
4124 
4125 /* ARGSUSED */
4126 hat_region_cookie_t
4127 hat_join_region(struct hat *hat,
4128     caddr_t r_saddr,
4129     size_t r_size,
4130     void *r_obj,
4131     u_offset_t r_objoff,
4132     uchar_t r_perm,
4133     uchar_t r_pgszc,
4134     hat_rgn_cb_func_t r_cb_function,
4135     uint_t flags)
4136 {
4137 	panic("No shared region support on x86");
4138 	return (HAT_INVALID_REGION_COOKIE);
4139 }
4140 
4141 /* ARGSUSED */
4142 void
4143 hat_leave_region(struct hat *hat, hat_region_cookie_t rcookie, uint_t flags)
4144 {
4145 	panic("No shared region support on x86");
4146 }
4147 
4148 /* ARGSUSED */
4149 void
4150 hat_dup_region(struct hat *hat, hat_region_cookie_t rcookie)
4151 {
4152 	panic("No shared region support on x86");
4153 }
4154 
4155 
4156 /*
4157  * Kernel Physical Mapping (kpm) facility
4158  *
4159  * Most of the routines needed to support segkpm are almost no-ops on the
4160  * x86 platform.  We map in the entire segment when it is created and leave
4161  * it mapped in, so there is no additional work required to set up and tear
4162  * down individual mappings.  All of these routines were created to support
4163  * SPARC platforms that have to avoid aliasing in their virtually indexed
4164  * caches.
4165  *
4166  * Most of the routines have sanity checks in them (e.g. verifying that the
4167  * passed-in page is locked).  We don't actually care about most of these
4168  * checks on x86, but we leave them in place to identify problems in the
4169  * upper levels.
4170  */
4171 
4172 /*
4173  * Map in a locked page and return the vaddr.
4174  */
4175 /*ARGSUSED*/
4176 caddr_t
4177 hat_kpm_mapin(struct page *pp, struct kpme *kpme)
4178 {
4179 	caddr_t		vaddr;
4180 
4181 #ifdef DEBUG
4182 	if (kpm_enable == 0) {
4183 		cmn_err(CE_WARN, "hat_kpm_mapin: kpm_enable not set\n");
4184 		return ((caddr_t)NULL);
4185 	}
4186 
4187 	if (pp == NULL || PAGE_LOCKED(pp) == 0) {
4188 		cmn_err(CE_WARN, "hat_kpm_mapin: pp zero or not locked\n");
4189 		return ((caddr_t)NULL);
4190 	}
4191 #endif
4192 
4193 	vaddr = hat_kpm_page2va(pp, 1);
4194 
4195 	return (vaddr);
4196 }
4197 
4198 /*
4199  * Mapout a locked page.
4200  */
4201 /*ARGSUSED*/
4202 void
4203 hat_kpm_mapout(struct page *pp, struct kpme *kpme, caddr_t vaddr)
4204 {
4205 #ifdef DEBUG
4206 	if (kpm_enable == 0) {
4207 		cmn_err(CE_WARN, "hat_kpm_mapout: kpm_enable not set\n");
4208 		return;
4209 	}
4210 
4211 	if (IS_KPM_ADDR(vaddr) == 0) {
4212 		cmn_err(CE_WARN, "hat_kpm_mapout: no kpm address\n");
4213 		return;
4214 	}
4215 
4216 	if (pp == NULL || PAGE_LOCKED(pp) == 0) {
4217 		cmn_err(CE_WARN, "hat_kpm_mapout: page zero or not locked\n");
4218 		return;
4219 	}
4220 #endif
4221 }
4222 
4223 /*
4224  * Return the kpm virtual address for a specific pfn
4225  */
4226 caddr_t
4227 hat_kpm_pfn2va(pfn_t pfn)
4228 {
4229 	uintptr_t vaddr = (uintptr_t)kpm_vbase + mmu_ptob(pfn);
4230 
4231 	ASSERT(!pfn_is_foreign(pfn));
4232 	return ((caddr_t)vaddr);
4233 }
4234 
4235 /*
4236  * Return the kpm virtual address for the page at pp.
4237  */
4238 /*ARGSUSED*/
4239 caddr_t
4240 hat_kpm_page2va(struct page *pp, int checkswap)
4241 {
4242 	return (hat_kpm_pfn2va(pp->p_pagenum));
4243 }
4244 
4245 /*
4246  * Return the page frame number for the kpm virtual address vaddr.
4247  */
4248 pfn_t
4249 hat_kpm_va2pfn(caddr_t vaddr)
4250 {
4251 	pfn_t		pfn;
4252 
4253 	ASSERT(IS_KPM_ADDR(vaddr));
4254 
4255 	pfn = (pfn_t)btop(vaddr - kpm_vbase);
4256 
4257 	return (pfn);
4258 }
4259 
4260 
4261 /*
4262  * Return the page for the kpm virtual address vaddr.
4263  */
4264 page_t *
4265 hat_kpm_vaddr2page(caddr_t vaddr)
4266 {
4267 	pfn_t		pfn;
4268 
4269 	ASSERT(IS_KPM_ADDR(vaddr));
4270 
4271 	pfn = hat_kpm_va2pfn(vaddr);
4272 
4273 	return (page_numtopp_nolock(pfn));
4274 }
4275 
4276 /*
4277  * hat_kpm_fault is called from segkpm_fault when we take a page fault on a
4278  * KPM page.  This should never happen on x86
4279  */
4280 int
4281 hat_kpm_fault(hat_t *hat, caddr_t vaddr)
4282 {
4283 	panic("pagefault in seg_kpm.  hat: 0x%p  vaddr: 0x%p",
4284 	    (void *)hat, (void *)vaddr);
4285 
4286 	return (0);
4287 }
4288 
4289 /*ARGSUSED*/
4290 void
4291 hat_kpm_mseghash_clear(int nentries)
4292 {}
4293 
4294 /*ARGSUSED*/
4295 void
4296 hat_kpm_mseghash_update(pgcnt_t inx, struct memseg *msp)
4297 {}
4298 
4299 #ifdef __xpv
4300 /*
4301  * There are specific Hypervisor calls to establish and remove mappings
4302  * to grant table references and the privcmd driver. We have to ensure
4303  * that a page table actually exists.
4304  */
4305 void
4306 hat_prepare_mapping(hat_t *hat, caddr_t addr, uint64_t *pte_ma)
4307 {
4308 	maddr_t base_ma;
4309 	htable_t *ht;
4310 	uint_t entry;
4311 
4312 	ASSERT(IS_P2ALIGNED((uintptr_t)addr, MMU_PAGESIZE));
4313 	XPV_DISALLOW_MIGRATE();
4314 	ht = htable_create(hat, (uintptr_t)addr, 0, NULL);
4315 
4316 	/*
4317 	 * if an address for pte_ma is passed in, return the MA of the pte
4318 	 * for this specific address.  This address is only valid as long
4319 	 * as the htable stays locked.
4320 	 */
4321 	if (pte_ma != NULL) {
4322 		entry = htable_va2entry((uintptr_t)addr, ht);
4323 		base_ma = pa_to_ma(ptob(ht->ht_pfn));
4324 		*pte_ma = base_ma + (entry << mmu.pte_size_shift);
4325 	}
4326 	XPV_ALLOW_MIGRATE();
4327 }
4328 
4329 void
4330 hat_release_mapping(hat_t *hat, caddr_t addr)
4331 {
4332 	htable_t *ht;
4333 
4334 	ASSERT(IS_P2ALIGNED((uintptr_t)addr, MMU_PAGESIZE));
4335 	XPV_DISALLOW_MIGRATE();
4336 	ht = htable_lookup(hat, (uintptr_t)addr, 0);
4337 	ASSERT(ht != NULL);
4338 	ASSERT(ht->ht_busy >= 2);
4339 	htable_release(ht);
4340 	htable_release(ht);
4341 	XPV_ALLOW_MIGRATE();
4342 									}
4343 #endif
4344