xref: /titanic_50/usr/src/uts/i86pc/vm/htable.c (revision 24bb1048b1197ebc9afe761e4098573aeedfe8c9)
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 /*
23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2014 by Delphix. All rights reserved.
25  * Copyright 2015 Joyent, Inc.
26  */
27 
28 #include <sys/types.h>
29 #include <sys/sysmacros.h>
30 #include <sys/kmem.h>
31 #include <sys/atomic.h>
32 #include <sys/bitmap.h>
33 #include <sys/machparam.h>
34 #include <sys/machsystm.h>
35 #include <sys/mman.h>
36 #include <sys/systm.h>
37 #include <sys/cpuvar.h>
38 #include <sys/thread.h>
39 #include <sys/proc.h>
40 #include <sys/cpu.h>
41 #include <sys/kmem.h>
42 #include <sys/disp.h>
43 #include <sys/vmem.h>
44 #include <sys/vmsystm.h>
45 #include <sys/promif.h>
46 #include <sys/var.h>
47 #include <sys/x86_archext.h>
48 #include <sys/archsystm.h>
49 #include <sys/bootconf.h>
50 #include <sys/dumphdr.h>
51 #include <vm/seg_kmem.h>
52 #include <vm/seg_kpm.h>
53 #include <vm/hat.h>
54 #include <vm/hat_i86.h>
55 #include <sys/cmn_err.h>
56 #include <sys/panic.h>
57 
58 #ifdef __xpv
59 #include <sys/hypervisor.h>
60 #include <sys/xpv_panic.h>
61 #endif
62 
63 #include <sys/bootinfo.h>
64 #include <vm/kboot_mmu.h>
65 
66 static void x86pte_zero(htable_t *dest, uint_t entry, uint_t count);
67 
68 kmem_cache_t *htable_cache;
69 
70 /*
71  * The variable htable_reserve_amount, rather than HTABLE_RESERVE_AMOUNT,
72  * is used in order to facilitate testing of the htable_steal() code.
73  * By resetting htable_reserve_amount to a lower value, we can force
74  * stealing to occur.  The reserve amount is a guess to get us through boot.
75  */
76 #define	HTABLE_RESERVE_AMOUNT	(200)
77 uint_t htable_reserve_amount = HTABLE_RESERVE_AMOUNT;
78 kmutex_t htable_reserve_mutex;
79 uint_t htable_reserve_cnt;
80 htable_t *htable_reserve_pool;
81 
82 /*
83  * Used to hand test htable_steal().
84  */
85 #ifdef DEBUG
86 ulong_t force_steal = 0;
87 ulong_t ptable_cnt = 0;
88 #endif
89 
90 /*
91  * This variable is so that we can tune this via /etc/system
92  * Any value works, but a power of two <= mmu.ptes_per_table is best.
93  */
94 uint_t htable_steal_passes = 8;
95 
96 /*
97  * mutex stuff for access to htable hash
98  */
99 #define	NUM_HTABLE_MUTEX 128
100 kmutex_t htable_mutex[NUM_HTABLE_MUTEX];
101 #define	HTABLE_MUTEX_HASH(h) ((h) & (NUM_HTABLE_MUTEX - 1))
102 
103 #define	HTABLE_ENTER(h)	mutex_enter(&htable_mutex[HTABLE_MUTEX_HASH(h)]);
104 #define	HTABLE_EXIT(h)	mutex_exit(&htable_mutex[HTABLE_MUTEX_HASH(h)]);
105 
106 /*
107  * forward declarations
108  */
109 static void link_ptp(htable_t *higher, htable_t *new, uintptr_t vaddr);
110 static void unlink_ptp(htable_t *higher, htable_t *old, uintptr_t vaddr);
111 static void htable_free(htable_t *ht);
112 static x86pte_t *x86pte_access_pagetable(htable_t *ht, uint_t index);
113 static void x86pte_release_pagetable(htable_t *ht);
114 static x86pte_t x86pte_cas(htable_t *ht, uint_t entry, x86pte_t old,
115 	x86pte_t new);
116 
117 /*
118  * A counter to track if we are stealing or reaping htables. When non-zero
119  * htable_free() will directly free htables (either to the reserve or kmem)
120  * instead of putting them in a hat's htable cache.
121  */
122 uint32_t htable_dont_cache = 0;
123 
124 /*
125  * Track the number of active pagetables, so we can know how many to reap
126  */
127 static uint32_t active_ptables = 0;
128 
129 #ifdef __xpv
130 /*
131  * Deal with hypervisor complications.
132  */
133 void
134 xen_flush_va(caddr_t va)
135 {
136 	struct mmuext_op t;
137 	uint_t count;
138 
139 	if (IN_XPV_PANIC()) {
140 		mmu_tlbflush_entry((caddr_t)va);
141 	} else {
142 		t.cmd = MMUEXT_INVLPG_LOCAL;
143 		t.arg1.linear_addr = (uintptr_t)va;
144 		if (HYPERVISOR_mmuext_op(&t, 1, &count, DOMID_SELF) < 0)
145 			panic("HYPERVISOR_mmuext_op() failed");
146 		ASSERT(count == 1);
147 	}
148 }
149 
150 void
151 xen_gflush_va(caddr_t va, cpuset_t cpus)
152 {
153 	struct mmuext_op t;
154 	uint_t count;
155 
156 	if (IN_XPV_PANIC()) {
157 		mmu_tlbflush_entry((caddr_t)va);
158 		return;
159 	}
160 
161 	t.cmd = MMUEXT_INVLPG_MULTI;
162 	t.arg1.linear_addr = (uintptr_t)va;
163 	/*LINTED: constant in conditional context*/
164 	set_xen_guest_handle(t.arg2.vcpumask, &cpus);
165 	if (HYPERVISOR_mmuext_op(&t, 1, &count, DOMID_SELF) < 0)
166 		panic("HYPERVISOR_mmuext_op() failed");
167 	ASSERT(count == 1);
168 }
169 
170 void
171 xen_flush_tlb()
172 {
173 	struct mmuext_op t;
174 	uint_t count;
175 
176 	if (IN_XPV_PANIC()) {
177 		xpv_panic_reload_cr3();
178 	} else {
179 		t.cmd = MMUEXT_TLB_FLUSH_LOCAL;
180 		if (HYPERVISOR_mmuext_op(&t, 1, &count, DOMID_SELF) < 0)
181 			panic("HYPERVISOR_mmuext_op() failed");
182 		ASSERT(count == 1);
183 	}
184 }
185 
186 void
187 xen_gflush_tlb(cpuset_t cpus)
188 {
189 	struct mmuext_op t;
190 	uint_t count;
191 
192 	ASSERT(!IN_XPV_PANIC());
193 	t.cmd = MMUEXT_TLB_FLUSH_MULTI;
194 	/*LINTED: constant in conditional context*/
195 	set_xen_guest_handle(t.arg2.vcpumask, &cpus);
196 	if (HYPERVISOR_mmuext_op(&t, 1, &count, DOMID_SELF) < 0)
197 		panic("HYPERVISOR_mmuext_op() failed");
198 	ASSERT(count == 1);
199 }
200 
201 /*
202  * Install/Adjust a kpm mapping under the hypervisor.
203  * Value of "how" should be:
204  *	PT_WRITABLE | PT_VALID - regular kpm mapping
205  *	PT_VALID - make mapping read-only
206  *	0	- remove mapping
207  *
208  * returns 0 on success. non-zero for failure.
209  */
210 int
211 xen_kpm_page(pfn_t pfn, uint_t how)
212 {
213 	paddr_t pa = mmu_ptob((paddr_t)pfn);
214 	x86pte_t pte = PT_NOCONSIST | PT_REF | PT_MOD;
215 
216 	if (kpm_vbase == NULL)
217 		return (0);
218 
219 	if (how)
220 		pte |= pa_to_ma(pa) | how;
221 	else
222 		pte = 0;
223 	return (HYPERVISOR_update_va_mapping((uintptr_t)kpm_vbase + pa,
224 	    pte, UVMF_INVLPG | UVMF_ALL));
225 }
226 
227 void
228 xen_pin(pfn_t pfn, level_t lvl)
229 {
230 	struct mmuext_op t;
231 	uint_t count;
232 
233 	t.cmd = MMUEXT_PIN_L1_TABLE + lvl;
234 	t.arg1.mfn = pfn_to_mfn(pfn);
235 	if (HYPERVISOR_mmuext_op(&t, 1, &count, DOMID_SELF) < 0)
236 		panic("HYPERVISOR_mmuext_op() failed");
237 	ASSERT(count == 1);
238 }
239 
240 void
241 xen_unpin(pfn_t pfn)
242 {
243 	struct mmuext_op t;
244 	uint_t count;
245 
246 	t.cmd = MMUEXT_UNPIN_TABLE;
247 	t.arg1.mfn = pfn_to_mfn(pfn);
248 	if (HYPERVISOR_mmuext_op(&t, 1, &count, DOMID_SELF) < 0)
249 		panic("HYPERVISOR_mmuext_op() failed");
250 	ASSERT(count == 1);
251 }
252 
253 static void
254 xen_map(uint64_t pte, caddr_t va)
255 {
256 	if (HYPERVISOR_update_va_mapping((uintptr_t)va, pte,
257 	    UVMF_INVLPG | UVMF_LOCAL))
258 		panic("HYPERVISOR_update_va_mapping() failed");
259 }
260 #endif /* __xpv */
261 
262 /*
263  * Allocate a memory page for a hardware page table.
264  *
265  * A wrapper around page_get_physical(), with some extra checks.
266  */
267 static pfn_t
268 ptable_alloc(uintptr_t seed)
269 {
270 	pfn_t pfn;
271 	page_t *pp;
272 
273 	pfn = PFN_INVALID;
274 
275 	/*
276 	 * The first check is to see if there is memory in the system. If we
277 	 * drop to throttlefree, then fail the ptable_alloc() and let the
278 	 * stealing code kick in. Note that we have to do this test here,
279 	 * since the test in page_create_throttle() would let the NOSLEEP
280 	 * allocation go through and deplete the page reserves.
281 	 *
282 	 * The !NOMEMWAIT() lets pageout, fsflush, etc. skip this check.
283 	 */
284 	if (!NOMEMWAIT() && freemem <= throttlefree + 1)
285 		return (PFN_INVALID);
286 
287 #ifdef DEBUG
288 	/*
289 	 * This code makes htable_steal() easier to test. By setting
290 	 * force_steal we force pagetable allocations to fall
291 	 * into the stealing code. Roughly 1 in ever "force_steal"
292 	 * page table allocations will fail.
293 	 */
294 	if (proc_pageout != NULL && force_steal > 1 &&
295 	    ++ptable_cnt > force_steal) {
296 		ptable_cnt = 0;
297 		return (PFN_INVALID);
298 	}
299 #endif /* DEBUG */
300 
301 	pp = page_get_physical(seed);
302 	if (pp == NULL)
303 		return (PFN_INVALID);
304 	ASSERT(PAGE_SHARED(pp));
305 	pfn = pp->p_pagenum;
306 	if (pfn == PFN_INVALID)
307 		panic("ptable_alloc(): Invalid PFN!!");
308 	atomic_inc_32(&active_ptables);
309 	HATSTAT_INC(hs_ptable_allocs);
310 	return (pfn);
311 }
312 
313 /*
314  * Free an htable's associated page table page.  See the comments
315  * for ptable_alloc().
316  */
317 static void
318 ptable_free(pfn_t pfn)
319 {
320 	page_t *pp = page_numtopp_nolock(pfn);
321 
322 	/*
323 	 * need to destroy the page used for the pagetable
324 	 */
325 	ASSERT(pfn != PFN_INVALID);
326 	HATSTAT_INC(hs_ptable_frees);
327 	atomic_dec_32(&active_ptables);
328 	if (pp == NULL)
329 		panic("ptable_free(): no page for pfn!");
330 	ASSERT(PAGE_SHARED(pp));
331 	ASSERT(pfn == pp->p_pagenum);
332 	ASSERT(!IN_XPV_PANIC());
333 
334 	/*
335 	 * Get an exclusive lock, might have to wait for a kmem reader.
336 	 */
337 	if (!page_tryupgrade(pp)) {
338 		u_offset_t off = pp->p_offset;
339 		page_unlock(pp);
340 		pp = page_lookup(&kvp, off, SE_EXCL);
341 		if (pp == NULL)
342 			panic("page not found");
343 	}
344 #ifdef __xpv
345 	if (kpm_vbase && xen_kpm_page(pfn, PT_VALID | PT_WRITABLE) < 0)
346 		panic("failure making kpm r/w pfn=0x%lx", pfn);
347 #endif
348 	page_hashout(pp, NULL);
349 	page_free(pp, 1);
350 	page_unresv(1);
351 }
352 
353 /*
354  * Put one htable on the reserve list.
355  */
356 static void
357 htable_put_reserve(htable_t *ht)
358 {
359 	ht->ht_hat = NULL;		/* no longer tied to a hat */
360 	ASSERT(ht->ht_pfn == PFN_INVALID);
361 	HATSTAT_INC(hs_htable_rputs);
362 	mutex_enter(&htable_reserve_mutex);
363 	ht->ht_next = htable_reserve_pool;
364 	htable_reserve_pool = ht;
365 	++htable_reserve_cnt;
366 	mutex_exit(&htable_reserve_mutex);
367 }
368 
369 /*
370  * Take one htable from the reserve.
371  */
372 static htable_t *
373 htable_get_reserve(void)
374 {
375 	htable_t *ht = NULL;
376 
377 	mutex_enter(&htable_reserve_mutex);
378 	if (htable_reserve_cnt != 0) {
379 		ht = htable_reserve_pool;
380 		ASSERT(ht != NULL);
381 		ASSERT(ht->ht_pfn == PFN_INVALID);
382 		htable_reserve_pool = ht->ht_next;
383 		--htable_reserve_cnt;
384 		HATSTAT_INC(hs_htable_rgets);
385 	}
386 	mutex_exit(&htable_reserve_mutex);
387 	return (ht);
388 }
389 
390 /*
391  * Allocate initial htables and put them on the reserve list
392  */
393 void
394 htable_initial_reserve(uint_t count)
395 {
396 	htable_t *ht;
397 
398 	count += HTABLE_RESERVE_AMOUNT;
399 	while (count > 0) {
400 		ht = kmem_cache_alloc(htable_cache, KM_NOSLEEP);
401 		ASSERT(ht != NULL);
402 
403 		ASSERT(use_boot_reserve);
404 		ht->ht_pfn = PFN_INVALID;
405 		htable_put_reserve(ht);
406 		--count;
407 	}
408 }
409 
410 /*
411  * Readjust the reserves after a thread finishes using them.
412  */
413 void
414 htable_adjust_reserve()
415 {
416 	htable_t *ht;
417 
418 	/*
419 	 * Free any excess htables in the reserve list
420 	 */
421 	while (htable_reserve_cnt > htable_reserve_amount &&
422 	    !USE_HAT_RESERVES()) {
423 		ht = htable_get_reserve();
424 		if (ht == NULL)
425 			return;
426 		ASSERT(ht->ht_pfn == PFN_INVALID);
427 		kmem_cache_free(htable_cache, ht);
428 	}
429 }
430 
431 /*
432  * Search the active htables for one to steal. Start at a different hash
433  * bucket every time to help spread the pain of stealing
434  */
435 static void
436 htable_steal_active(hat_t *hat, uint_t cnt, uint_t threshold,
437     uint_t *stolen, htable_t **list)
438 {
439 	static uint_t	h_seed = 0;
440 	htable_t	*higher, *ht;
441 	uint_t		h, e, h_start;
442 	uintptr_t	va;
443 	x86pte_t	pte;
444 
445 	h = h_start = h_seed++ % hat->hat_num_hash;
446 	do {
447 		higher = NULL;
448 		HTABLE_ENTER(h);
449 		for (ht = hat->hat_ht_hash[h]; ht; ht = ht->ht_next) {
450 
451 			/*
452 			 * Can we rule out reaping?
453 			 */
454 			if (ht->ht_busy != 0 ||
455 			    (ht->ht_flags & HTABLE_SHARED_PFN) ||
456 			    ht->ht_level > 0 || ht->ht_valid_cnt > threshold ||
457 			    ht->ht_lock_cnt != 0)
458 				continue;
459 
460 			/*
461 			 * Increment busy so the htable can't disappear. We
462 			 * drop the htable mutex to avoid deadlocks with
463 			 * hat_pageunload() and the hment mutex while we
464 			 * call hat_pte_unmap()
465 			 */
466 			++ht->ht_busy;
467 			HTABLE_EXIT(h);
468 
469 			/*
470 			 * Try stealing.
471 			 * - unload and invalidate all PTEs
472 			 */
473 			for (e = 0, va = ht->ht_vaddr;
474 			    e < HTABLE_NUM_PTES(ht) && ht->ht_valid_cnt > 0 &&
475 			    ht->ht_busy == 1 && ht->ht_lock_cnt == 0;
476 			    ++e, va += MMU_PAGESIZE) {
477 				pte = x86pte_get(ht, e);
478 				if (!PTE_ISVALID(pte))
479 					continue;
480 				hat_pte_unmap(ht, e, HAT_UNLOAD, pte, NULL,
481 				    B_TRUE);
482 			}
483 
484 			/*
485 			 * Reacquire htable lock. If we didn't remove all
486 			 * mappings in the table, or another thread added a new
487 			 * mapping behind us, give up on this table.
488 			 */
489 			HTABLE_ENTER(h);
490 			if (ht->ht_busy != 1 || ht->ht_valid_cnt != 0 ||
491 			    ht->ht_lock_cnt != 0) {
492 				--ht->ht_busy;
493 				continue;
494 			}
495 
496 			/*
497 			 * Steal it and unlink the page table.
498 			 */
499 			higher = ht->ht_parent;
500 			unlink_ptp(higher, ht, ht->ht_vaddr);
501 
502 			/*
503 			 * remove from the hash list
504 			 */
505 			if (ht->ht_next)
506 				ht->ht_next->ht_prev = ht->ht_prev;
507 
508 			if (ht->ht_prev) {
509 				ht->ht_prev->ht_next = ht->ht_next;
510 			} else {
511 				ASSERT(hat->hat_ht_hash[h] == ht);
512 				hat->hat_ht_hash[h] = ht->ht_next;
513 			}
514 
515 			/*
516 			 * Break to outer loop to release the
517 			 * higher (ht_parent) pagetable. This
518 			 * spreads out the pain caused by
519 			 * pagefaults.
520 			 */
521 			ht->ht_next = *list;
522 			*list = ht;
523 			++*stolen;
524 			break;
525 		}
526 		HTABLE_EXIT(h);
527 		if (higher != NULL)
528 			htable_release(higher);
529 		if (++h == hat->hat_num_hash)
530 			h = 0;
531 	} while (*stolen < cnt && h != h_start);
532 }
533 
534 /*
535  * Move hat to the end of the kas list
536  */
537 static void
538 move_victim(hat_t *hat)
539 {
540 	ASSERT(MUTEX_HELD(&hat_list_lock));
541 
542 	/* unlink victim hat */
543 	if (hat->hat_prev)
544 		hat->hat_prev->hat_next = hat->hat_next;
545 	else
546 		kas.a_hat->hat_next = hat->hat_next;
547 
548 	if (hat->hat_next)
549 		hat->hat_next->hat_prev = hat->hat_prev;
550 	else
551 		kas.a_hat->hat_prev = hat->hat_prev;
552 	/* relink at end of hat list */
553 	hat->hat_next = NULL;
554 	hat->hat_prev = kas.a_hat->hat_prev;
555 	if (hat->hat_prev)
556 		hat->hat_prev->hat_next = hat;
557 	else
558 		kas.a_hat->hat_next = hat;
559 
560 	kas.a_hat->hat_prev = hat;
561 }
562 
563 /*
564  * This routine steals htables from user processes.  Called by htable_reap
565  * (reap=TRUE) or htable_alloc (reap=FALSE).
566  */
567 static htable_t *
568 htable_steal(uint_t cnt, boolean_t reap)
569 {
570 	hat_t		*hat = kas.a_hat;	/* list starts with khat */
571 	htable_t	*list = NULL;
572 	htable_t	*ht;
573 	uint_t		stolen = 0;
574 	uint_t		pass, passes;
575 	uint_t		threshold;
576 
577 	/*
578 	 * Limit htable_steal_passes to something reasonable
579 	 */
580 	if (htable_steal_passes == 0)
581 		htable_steal_passes = 1;
582 	if (htable_steal_passes > mmu.ptes_per_table)
583 		htable_steal_passes = mmu.ptes_per_table;
584 
585 	/*
586 	 * If we're stealing merely as part of kmem reaping (versus stealing
587 	 * to assure forward progress), we don't want to actually steal any
588 	 * active htables.  (Stealing active htables merely to give memory
589 	 * back to the system can inadvertently kick off an htable crime wave
590 	 * as active processes repeatedly steal htables from one another,
591 	 * plummeting the system into a kind of HAT lawlessness that can
592 	 * become so violent as to impede the one thing that can end it:  the
593 	 * freeing of memory via ARC reclaim and other means.)  So if we're
594 	 * reaping, we limit ourselves to the first pass that steals cached
595 	 * htables that aren't in use -- which gives memory back, but averts
596 	 * the entire breakdown of social order.
597 	 */
598 	passes = reap ? 0 : htable_steal_passes;
599 
600 	/*
601 	 * Loop through all user hats. The 1st pass takes cached htables that
602 	 * aren't in use. The later passes steal by removing mappings, too.
603 	 */
604 	atomic_inc_32(&htable_dont_cache);
605 	for (pass = 0; pass <= passes && stolen < cnt; ++pass) {
606 		threshold = pass * mmu.ptes_per_table / htable_steal_passes;
607 
608 		mutex_enter(&hat_list_lock);
609 
610 		/* skip the first hat (kernel) */
611 		hat = kas.a_hat->hat_next;
612 		for (;;) {
613 			/*
614 			 * Skip any hat that is already being stolen from.
615 			 *
616 			 * We skip SHARED hats, as these are dummy
617 			 * hats that host ISM shared page tables.
618 			 *
619 			 * We also skip if HAT_FREEING because hat_pte_unmap()
620 			 * won't zero out the PTE's. That would lead to hitting
621 			 * stale PTEs either here or under hat_unload() when we
622 			 * steal and unload the same page table in competing
623 			 * threads.
624 			 */
625 			while (hat != NULL &&
626 			    (hat->hat_flags &
627 			    (HAT_VICTIM | HAT_SHARED | HAT_FREEING)) != 0)
628 				hat = hat->hat_next;
629 
630 			if (hat == NULL)
631 				break;
632 
633 			/*
634 			 * Mark the HAT as a stealing victim so that it is
635 			 * not freed from under us, e.g. in as_free()
636 			 */
637 			hat->hat_flags |= HAT_VICTIM;
638 			mutex_exit(&hat_list_lock);
639 
640 			/*
641 			 * Take any htables from the hat's cached "free" list.
642 			 */
643 			hat_enter(hat);
644 			while ((ht = hat->hat_ht_cached) != NULL &&
645 			    stolen < cnt) {
646 				hat->hat_ht_cached = ht->ht_next;
647 				ht->ht_next = list;
648 				list = ht;
649 				++stolen;
650 			}
651 			hat_exit(hat);
652 
653 			/*
654 			 * Don't steal active htables on first pass.
655 			 */
656 			if (pass != 0 && (stolen < cnt))
657 				htable_steal_active(hat, cnt, threshold,
658 				    &stolen, &list);
659 
660 			/*
661 			 * do synchronous teardown for the reap case so that
662 			 * we can forget hat; at this time, hat is
663 			 * guaranteed to be around because HAT_VICTIM is set
664 			 * (see htable_free() for similar code)
665 			 */
666 			for (ht = list; (ht) && (reap); ht = ht->ht_next) {
667 				if (ht->ht_hat == NULL)
668 					continue;
669 				ASSERT(ht->ht_hat == hat);
670 #if defined(__xpv) && defined(__amd64)
671 				if (!(ht->ht_flags & HTABLE_VLP) &&
672 				    ht->ht_level == mmu.max_level) {
673 					ptable_free(hat->hat_user_ptable);
674 					hat->hat_user_ptable = PFN_INVALID;
675 				}
676 #endif
677 				/*
678 				 * forget the hat
679 				 */
680 				ht->ht_hat = NULL;
681 			}
682 
683 			mutex_enter(&hat_list_lock);
684 
685 			/*
686 			 * Are we finished?
687 			 */
688 			if (stolen == cnt) {
689 				/*
690 				 * Try to spread the pain of stealing,
691 				 * move victim HAT to the end of the HAT list.
692 				 */
693 				if (pass >= 1 && cnt == 1 &&
694 				    kas.a_hat->hat_prev != hat)
695 					move_victim(hat);
696 				/*
697 				 * We are finished
698 				 */
699 			}
700 
701 			/*
702 			 * Clear the victim flag, hat can go away now (once
703 			 * the lock is dropped)
704 			 */
705 			if (hat->hat_flags & HAT_VICTIM) {
706 				ASSERT(hat != kas.a_hat);
707 				hat->hat_flags &= ~HAT_VICTIM;
708 				cv_broadcast(&hat_list_cv);
709 			}
710 
711 			/* move on to the next hat */
712 			hat = hat->hat_next;
713 		}
714 
715 		mutex_exit(&hat_list_lock);
716 
717 	}
718 	ASSERT(!MUTEX_HELD(&hat_list_lock));
719 
720 	atomic_dec_32(&htable_dont_cache);
721 	return (list);
722 }
723 
724 /*
725  * This is invoked from kmem when the system is low on memory.  We try
726  * to free hments, htables, and ptables to improve the memory situation.
727  */
728 /*ARGSUSED*/
729 static void
730 htable_reap(void *handle)
731 {
732 	uint_t		reap_cnt;
733 	htable_t	*list;
734 	htable_t	*ht;
735 
736 	HATSTAT_INC(hs_reap_attempts);
737 	if (!can_steal_post_boot)
738 		return;
739 
740 	/*
741 	 * Try to reap 5% of the page tables bounded by a maximum of
742 	 * 5% of physmem and a minimum of 10.
743 	 */
744 	reap_cnt = MAX(MIN(physmem / 20, active_ptables / 20), 10);
745 
746 	/*
747 	 * Note: htable_dont_cache should be set at the time of
748 	 * invoking htable_free()
749 	 */
750 	atomic_inc_32(&htable_dont_cache);
751 	/*
752 	 * Let htable_steal() do the work, we just call htable_free()
753 	 */
754 	XPV_DISALLOW_MIGRATE();
755 	list = htable_steal(reap_cnt, B_TRUE);
756 	XPV_ALLOW_MIGRATE();
757 	while ((ht = list) != NULL) {
758 		list = ht->ht_next;
759 		HATSTAT_INC(hs_reaped);
760 		htable_free(ht);
761 	}
762 	atomic_dec_32(&htable_dont_cache);
763 
764 	/*
765 	 * Free up excess reserves
766 	 */
767 	htable_adjust_reserve();
768 	hment_adjust_reserve();
769 }
770 
771 /*
772  * Allocate an htable, stealing one or using the reserve if necessary
773  */
774 static htable_t *
775 htable_alloc(
776 	hat_t		*hat,
777 	uintptr_t	vaddr,
778 	level_t		level,
779 	htable_t	*shared)
780 {
781 	htable_t	*ht = NULL;
782 	uint_t		is_vlp;
783 	uint_t		is_bare = 0;
784 	uint_t		need_to_zero = 1;
785 	int		kmflags = (can_steal_post_boot ? KM_NOSLEEP : KM_SLEEP);
786 
787 	if (level < 0 || level > TOP_LEVEL(hat))
788 		panic("htable_alloc(): level %d out of range\n", level);
789 
790 	is_vlp = (hat->hat_flags & HAT_VLP) && level == VLP_LEVEL;
791 	if (is_vlp || shared != NULL)
792 		is_bare = 1;
793 
794 	/*
795 	 * First reuse a cached htable from the hat_ht_cached field, this
796 	 * avoids unnecessary trips through kmem/page allocators.
797 	 */
798 	if (hat->hat_ht_cached != NULL && !is_bare) {
799 		hat_enter(hat);
800 		ht = hat->hat_ht_cached;
801 		if (ht != NULL) {
802 			hat->hat_ht_cached = ht->ht_next;
803 			need_to_zero = 0;
804 			/* XX64 ASSERT() they're all zero somehow */
805 			ASSERT(ht->ht_pfn != PFN_INVALID);
806 		}
807 		hat_exit(hat);
808 	}
809 
810 	if (ht == NULL) {
811 		/*
812 		 * Allocate an htable, possibly refilling the reserves.
813 		 */
814 		if (USE_HAT_RESERVES()) {
815 			ht = htable_get_reserve();
816 		} else {
817 			/*
818 			 * Donate successful htable allocations to the reserve.
819 			 */
820 			for (;;) {
821 				ht = kmem_cache_alloc(htable_cache, kmflags);
822 				if (ht == NULL)
823 					break;
824 				ht->ht_pfn = PFN_INVALID;
825 				if (USE_HAT_RESERVES() ||
826 				    htable_reserve_cnt >= htable_reserve_amount)
827 					break;
828 				htable_put_reserve(ht);
829 			}
830 		}
831 
832 		/*
833 		 * allocate a page for the hardware page table if needed
834 		 */
835 		if (ht != NULL && !is_bare) {
836 			ht->ht_hat = hat;
837 			ht->ht_pfn = ptable_alloc((uintptr_t)ht);
838 			if (ht->ht_pfn == PFN_INVALID) {
839 				if (USE_HAT_RESERVES())
840 					htable_put_reserve(ht);
841 				else
842 					kmem_cache_free(htable_cache, ht);
843 				ht = NULL;
844 			}
845 		}
846 	}
847 
848 	/*
849 	 * If allocations failed, kick off a kmem_reap() and resort to
850 	 * htable steal(). We may spin here if the system is very low on
851 	 * memory. If the kernel itself has consumed all memory and kmem_reap()
852 	 * can't free up anything, then we'll really get stuck here.
853 	 * That should only happen in a system where the administrator has
854 	 * misconfigured VM parameters via /etc/system.
855 	 */
856 	while (ht == NULL && can_steal_post_boot) {
857 		kmem_reap();
858 		ht = htable_steal(1, B_FALSE);
859 		HATSTAT_INC(hs_steals);
860 
861 		/*
862 		 * If we stole for a bare htable, release the pagetable page.
863 		 */
864 		if (ht != NULL) {
865 			if (is_bare) {
866 				ptable_free(ht->ht_pfn);
867 				ht->ht_pfn = PFN_INVALID;
868 #if defined(__xpv) && defined(__amd64)
869 			/*
870 			 * make stolen page table writable again in kpm
871 			 */
872 			} else if (kpm_vbase && xen_kpm_page(ht->ht_pfn,
873 			    PT_VALID | PT_WRITABLE) < 0) {
874 				panic("failure making kpm r/w pfn=0x%lx",
875 				    ht->ht_pfn);
876 #endif
877 			}
878 		}
879 	}
880 
881 	/*
882 	 * All attempts to allocate or steal failed. This should only happen
883 	 * if we run out of memory during boot, due perhaps to a huge
884 	 * boot_archive. At this point there's no way to continue.
885 	 */
886 	if (ht == NULL)
887 		panic("htable_alloc(): couldn't steal\n");
888 
889 #if defined(__amd64) && defined(__xpv)
890 	/*
891 	 * Under the 64-bit hypervisor, we have 2 top level page tables.
892 	 * If this allocation fails, we'll resort to stealing.
893 	 * We use the stolen page indirectly, by freeing the
894 	 * stolen htable first.
895 	 */
896 	if (level == mmu.max_level) {
897 		for (;;) {
898 			htable_t *stolen;
899 
900 			hat->hat_user_ptable = ptable_alloc((uintptr_t)ht + 1);
901 			if (hat->hat_user_ptable != PFN_INVALID)
902 				break;
903 			stolen = htable_steal(1, B_FALSE);
904 			if (stolen == NULL)
905 				panic("2nd steal ptable failed\n");
906 			htable_free(stolen);
907 		}
908 		block_zero_no_xmm(kpm_vbase + pfn_to_pa(hat->hat_user_ptable),
909 		    MMU_PAGESIZE);
910 	}
911 #endif
912 
913 	/*
914 	 * Shared page tables have all entries locked and entries may not
915 	 * be added or deleted.
916 	 */
917 	ht->ht_flags = 0;
918 	if (shared != NULL) {
919 		ASSERT(shared->ht_valid_cnt > 0);
920 		ht->ht_flags |= HTABLE_SHARED_PFN;
921 		ht->ht_pfn = shared->ht_pfn;
922 		ht->ht_lock_cnt = 0;
923 		ht->ht_valid_cnt = 0;		/* updated in hat_share() */
924 		ht->ht_shares = shared;
925 		need_to_zero = 0;
926 	} else {
927 		ht->ht_shares = NULL;
928 		ht->ht_lock_cnt = 0;
929 		ht->ht_valid_cnt = 0;
930 	}
931 
932 	/*
933 	 * setup flags, etc. for VLP htables
934 	 */
935 	if (is_vlp) {
936 		ht->ht_flags |= HTABLE_VLP;
937 		ASSERT(ht->ht_pfn == PFN_INVALID);
938 		need_to_zero = 0;
939 	}
940 
941 	/*
942 	 * fill in the htable
943 	 */
944 	ht->ht_hat = hat;
945 	ht->ht_parent = NULL;
946 	ht->ht_vaddr = vaddr;
947 	ht->ht_level = level;
948 	ht->ht_busy = 1;
949 	ht->ht_next = NULL;
950 	ht->ht_prev = NULL;
951 
952 	/*
953 	 * Zero out any freshly allocated page table
954 	 */
955 	if (need_to_zero)
956 		x86pte_zero(ht, 0, mmu.ptes_per_table);
957 
958 #if defined(__amd64) && defined(__xpv)
959 	if (!is_bare && kpm_vbase) {
960 		(void) xen_kpm_page(ht->ht_pfn, PT_VALID);
961 		if (level == mmu.max_level)
962 			(void) xen_kpm_page(hat->hat_user_ptable, PT_VALID);
963 	}
964 #endif
965 
966 	return (ht);
967 }
968 
969 /*
970  * Free up an htable, either to a hat's cached list, the reserves or
971  * back to kmem.
972  */
973 static void
974 htable_free(htable_t *ht)
975 {
976 	hat_t *hat = ht->ht_hat;
977 
978 	/*
979 	 * If the process isn't exiting, cache the free htable in the hat
980 	 * structure. We always do this for the boot time reserve. We don't
981 	 * do this if the hat is exiting or we are stealing/reaping htables.
982 	 */
983 	if (hat != NULL &&
984 	    !(ht->ht_flags & HTABLE_SHARED_PFN) &&
985 	    (use_boot_reserve ||
986 	    (!(hat->hat_flags & HAT_FREEING) && !htable_dont_cache))) {
987 		ASSERT((ht->ht_flags & HTABLE_VLP) == 0);
988 		ASSERT(ht->ht_pfn != PFN_INVALID);
989 		hat_enter(hat);
990 		ht->ht_next = hat->hat_ht_cached;
991 		hat->hat_ht_cached = ht;
992 		hat_exit(hat);
993 		return;
994 	}
995 
996 	/*
997 	 * If we have a hardware page table, free it.
998 	 * We don't free page tables that are accessed by sharing.
999 	 */
1000 	if (ht->ht_flags & HTABLE_SHARED_PFN) {
1001 		ASSERT(ht->ht_pfn != PFN_INVALID);
1002 	} else if (!(ht->ht_flags & HTABLE_VLP)) {
1003 		ptable_free(ht->ht_pfn);
1004 #if defined(__amd64) && defined(__xpv)
1005 		if (ht->ht_level == mmu.max_level && hat != NULL) {
1006 			ptable_free(hat->hat_user_ptable);
1007 			hat->hat_user_ptable = PFN_INVALID;
1008 		}
1009 #endif
1010 	}
1011 	ht->ht_pfn = PFN_INVALID;
1012 
1013 	/*
1014 	 * Free it or put into reserves.
1015 	 */
1016 	if (USE_HAT_RESERVES() || htable_reserve_cnt < htable_reserve_amount) {
1017 		htable_put_reserve(ht);
1018 	} else {
1019 		kmem_cache_free(htable_cache, ht);
1020 		htable_adjust_reserve();
1021 	}
1022 }
1023 
1024 
1025 /*
1026  * This is called when a hat is being destroyed or swapped out. We reap all
1027  * the remaining htables in the hat cache. If destroying all left over
1028  * htables are also destroyed.
1029  *
1030  * We also don't need to invalidate any of the PTPs nor do any demapping.
1031  */
1032 void
1033 htable_purge_hat(hat_t *hat)
1034 {
1035 	htable_t *ht;
1036 	int h;
1037 
1038 	/*
1039 	 * Purge the htable cache if just reaping.
1040 	 */
1041 	if (!(hat->hat_flags & HAT_FREEING)) {
1042 		atomic_inc_32(&htable_dont_cache);
1043 		for (;;) {
1044 			hat_enter(hat);
1045 			ht = hat->hat_ht_cached;
1046 			if (ht == NULL) {
1047 				hat_exit(hat);
1048 				break;
1049 			}
1050 			hat->hat_ht_cached = ht->ht_next;
1051 			hat_exit(hat);
1052 			htable_free(ht);
1053 		}
1054 		atomic_dec_32(&htable_dont_cache);
1055 		return;
1056 	}
1057 
1058 	/*
1059 	 * if freeing, no locking is needed
1060 	 */
1061 	while ((ht = hat->hat_ht_cached) != NULL) {
1062 		hat->hat_ht_cached = ht->ht_next;
1063 		htable_free(ht);
1064 	}
1065 
1066 	/*
1067 	 * walk thru the htable hash table and free all the htables in it.
1068 	 */
1069 	for (h = 0; h < hat->hat_num_hash; ++h) {
1070 		while ((ht = hat->hat_ht_hash[h]) != NULL) {
1071 			if (ht->ht_next)
1072 				ht->ht_next->ht_prev = ht->ht_prev;
1073 
1074 			if (ht->ht_prev) {
1075 				ht->ht_prev->ht_next = ht->ht_next;
1076 			} else {
1077 				ASSERT(hat->hat_ht_hash[h] == ht);
1078 				hat->hat_ht_hash[h] = ht->ht_next;
1079 			}
1080 			htable_free(ht);
1081 		}
1082 	}
1083 }
1084 
1085 /*
1086  * Unlink an entry for a table at vaddr and level out of the existing table
1087  * one level higher. We are always holding the HASH_ENTER() when doing this.
1088  */
1089 static void
1090 unlink_ptp(htable_t *higher, htable_t *old, uintptr_t vaddr)
1091 {
1092 	uint_t		entry = htable_va2entry(vaddr, higher);
1093 	x86pte_t	expect = MAKEPTP(old->ht_pfn, old->ht_level);
1094 	x86pte_t	found;
1095 	hat_t		*hat = old->ht_hat;
1096 
1097 	ASSERT(higher->ht_busy > 0);
1098 	ASSERT(higher->ht_valid_cnt > 0);
1099 	ASSERT(old->ht_valid_cnt == 0);
1100 	found = x86pte_cas(higher, entry, expect, 0);
1101 #ifdef __xpv
1102 	/*
1103 	 * This is weird, but Xen apparently automatically unlinks empty
1104 	 * pagetables from the upper page table. So allow PTP to be 0 already.
1105 	 */
1106 	if (found != expect && found != 0)
1107 #else
1108 	if (found != expect)
1109 #endif
1110 		panic("Bad PTP found=" FMT_PTE ", expected=" FMT_PTE,
1111 		    found, expect);
1112 
1113 	/*
1114 	 * When a top level VLP page table entry changes, we must issue
1115 	 * a reload of cr3 on all processors.
1116 	 *
1117 	 * If we don't need do do that, then we still have to INVLPG against
1118 	 * an address covered by the inner page table, as the latest processors
1119 	 * have TLB-like caches for non-leaf page table entries.
1120 	 */
1121 	if (!(hat->hat_flags & HAT_FREEING)) {
1122 		hat_tlb_inval(hat, (higher->ht_flags & HTABLE_VLP) ?
1123 		    DEMAP_ALL_ADDR : old->ht_vaddr);
1124 	}
1125 
1126 	HTABLE_DEC(higher->ht_valid_cnt);
1127 }
1128 
1129 /*
1130  * Link an entry for a new table at vaddr and level into the existing table
1131  * one level higher. We are always holding the HASH_ENTER() when doing this.
1132  */
1133 static void
1134 link_ptp(htable_t *higher, htable_t *new, uintptr_t vaddr)
1135 {
1136 	uint_t		entry = htable_va2entry(vaddr, higher);
1137 	x86pte_t	newptp = MAKEPTP(new->ht_pfn, new->ht_level);
1138 	x86pte_t	found;
1139 
1140 	ASSERT(higher->ht_busy > 0);
1141 
1142 	ASSERT(new->ht_level != mmu.max_level);
1143 
1144 	HTABLE_INC(higher->ht_valid_cnt);
1145 
1146 	found = x86pte_cas(higher, entry, 0, newptp);
1147 	if ((found & ~PT_REF) != 0)
1148 		panic("HAT: ptp not 0, found=" FMT_PTE, found);
1149 
1150 	/*
1151 	 * When any top level VLP page table entry changes, we must issue
1152 	 * a reload of cr3 on all processors using it.
1153 	 * We also need to do this for the kernel hat on PAE 32 bit kernel.
1154 	 */
1155 	if (
1156 #ifdef __i386
1157 	    (higher->ht_hat == kas.a_hat && higher->ht_level == VLP_LEVEL) ||
1158 #endif
1159 	    (higher->ht_flags & HTABLE_VLP))
1160 		hat_tlb_inval(higher->ht_hat, DEMAP_ALL_ADDR);
1161 }
1162 
1163 /*
1164  * Release of hold on an htable. If this is the last use and the pagetable
1165  * is empty we may want to free it, then recursively look at the pagetable
1166  * above it. The recursion is handled by the outer while() loop.
1167  *
1168  * On the metal, during process exit, we don't bother unlinking the tables from
1169  * upper level pagetables. They are instead handled in bulk by hat_free_end().
1170  * We can't do this on the hypervisor as we need the page table to be
1171  * implicitly unpinnned before it goes to the free page lists. This can't
1172  * happen unless we fully unlink it from the page table hierarchy.
1173  */
1174 void
1175 htable_release(htable_t *ht)
1176 {
1177 	uint_t		hashval;
1178 	htable_t	*shared;
1179 	htable_t	*higher;
1180 	hat_t		*hat;
1181 	uintptr_t	va;
1182 	level_t		level;
1183 
1184 	while (ht != NULL) {
1185 		shared = NULL;
1186 		for (;;) {
1187 			hat = ht->ht_hat;
1188 			va = ht->ht_vaddr;
1189 			level = ht->ht_level;
1190 			hashval = HTABLE_HASH(hat, va, level);
1191 
1192 			/*
1193 			 * The common case is that this isn't the last use of
1194 			 * an htable so we don't want to free the htable.
1195 			 */
1196 			HTABLE_ENTER(hashval);
1197 			ASSERT(ht->ht_valid_cnt >= 0);
1198 			ASSERT(ht->ht_busy > 0);
1199 			if (ht->ht_valid_cnt > 0)
1200 				break;
1201 			if (ht->ht_busy > 1)
1202 				break;
1203 			ASSERT(ht->ht_lock_cnt == 0);
1204 
1205 #if !defined(__xpv)
1206 			/*
1207 			 * we always release empty shared htables
1208 			 */
1209 			if (!(ht->ht_flags & HTABLE_SHARED_PFN)) {
1210 
1211 				/*
1212 				 * don't release if in address space tear down
1213 				 */
1214 				if (hat->hat_flags & HAT_FREEING)
1215 					break;
1216 
1217 				/*
1218 				 * At and above max_page_level, free if it's for
1219 				 * a boot-time kernel mapping below kernelbase.
1220 				 */
1221 				if (level >= mmu.max_page_level &&
1222 				    (hat != kas.a_hat || va >= kernelbase))
1223 					break;
1224 			}
1225 #endif /* __xpv */
1226 
1227 			/*
1228 			 * Remember if we destroy an htable that shares its PFN
1229 			 * from elsewhere.
1230 			 */
1231 			if (ht->ht_flags & HTABLE_SHARED_PFN) {
1232 				ASSERT(shared == NULL);
1233 				shared = ht->ht_shares;
1234 				HATSTAT_INC(hs_htable_unshared);
1235 			}
1236 
1237 			/*
1238 			 * Handle release of a table and freeing the htable_t.
1239 			 * Unlink it from the table higher (ie. ht_parent).
1240 			 */
1241 			higher = ht->ht_parent;
1242 			ASSERT(higher != NULL);
1243 
1244 			/*
1245 			 * Unlink the pagetable.
1246 			 */
1247 			unlink_ptp(higher, ht, va);
1248 
1249 			/*
1250 			 * remove this htable from its hash list
1251 			 */
1252 			if (ht->ht_next)
1253 				ht->ht_next->ht_prev = ht->ht_prev;
1254 
1255 			if (ht->ht_prev) {
1256 				ht->ht_prev->ht_next = ht->ht_next;
1257 			} else {
1258 				ASSERT(hat->hat_ht_hash[hashval] == ht);
1259 				hat->hat_ht_hash[hashval] = ht->ht_next;
1260 			}
1261 			HTABLE_EXIT(hashval);
1262 			htable_free(ht);
1263 			ht = higher;
1264 		}
1265 
1266 		ASSERT(ht->ht_busy >= 1);
1267 		--ht->ht_busy;
1268 		HTABLE_EXIT(hashval);
1269 
1270 		/*
1271 		 * If we released a shared htable, do a release on the htable
1272 		 * from which it shared
1273 		 */
1274 		ht = shared;
1275 	}
1276 }
1277 
1278 /*
1279  * Find the htable for the pagetable at the given level for the given address.
1280  * If found acquires a hold that eventually needs to be htable_release()d
1281  */
1282 htable_t *
1283 htable_lookup(hat_t *hat, uintptr_t vaddr, level_t level)
1284 {
1285 	uintptr_t	base;
1286 	uint_t		hashval;
1287 	htable_t	*ht = NULL;
1288 
1289 	ASSERT(level >= 0);
1290 	ASSERT(level <= TOP_LEVEL(hat));
1291 
1292 	if (level == TOP_LEVEL(hat)) {
1293 #if defined(__amd64)
1294 		/*
1295 		 * 32 bit address spaces on 64 bit kernels need to check
1296 		 * for overflow of the 32 bit address space
1297 		 */
1298 		if ((hat->hat_flags & HAT_VLP) && vaddr >= ((uint64_t)1 << 32))
1299 			return (NULL);
1300 #endif
1301 		base = 0;
1302 	} else {
1303 		base = vaddr & LEVEL_MASK(level + 1);
1304 	}
1305 
1306 	hashval = HTABLE_HASH(hat, base, level);
1307 	HTABLE_ENTER(hashval);
1308 	for (ht = hat->hat_ht_hash[hashval]; ht; ht = ht->ht_next) {
1309 		if (ht->ht_hat == hat &&
1310 		    ht->ht_vaddr == base &&
1311 		    ht->ht_level == level)
1312 			break;
1313 	}
1314 	if (ht)
1315 		++ht->ht_busy;
1316 
1317 	HTABLE_EXIT(hashval);
1318 	return (ht);
1319 }
1320 
1321 /*
1322  * Acquires a hold on a known htable (from a locked hment entry).
1323  */
1324 void
1325 htable_acquire(htable_t *ht)
1326 {
1327 	hat_t		*hat = ht->ht_hat;
1328 	level_t		level = ht->ht_level;
1329 	uintptr_t	base = ht->ht_vaddr;
1330 	uint_t		hashval = HTABLE_HASH(hat, base, level);
1331 
1332 	HTABLE_ENTER(hashval);
1333 #ifdef DEBUG
1334 	/*
1335 	 * make sure the htable is there
1336 	 */
1337 	{
1338 		htable_t	*h;
1339 
1340 		for (h = hat->hat_ht_hash[hashval];
1341 		    h && h != ht;
1342 		    h = h->ht_next)
1343 			;
1344 		ASSERT(h == ht);
1345 	}
1346 #endif /* DEBUG */
1347 	++ht->ht_busy;
1348 	HTABLE_EXIT(hashval);
1349 }
1350 
1351 /*
1352  * Find the htable for the pagetable at the given level for the given address.
1353  * If found acquires a hold that eventually needs to be htable_release()d
1354  * If not found the table is created.
1355  *
1356  * Since we can't hold a hash table mutex during allocation, we have to
1357  * drop it and redo the search on a create. Then we may have to free the newly
1358  * allocated htable if another thread raced in and created it ahead of us.
1359  */
1360 htable_t *
1361 htable_create(
1362 	hat_t		*hat,
1363 	uintptr_t	vaddr,
1364 	level_t		level,
1365 	htable_t	*shared)
1366 {
1367 	uint_t		h;
1368 	level_t		l;
1369 	uintptr_t	base;
1370 	htable_t	*ht;
1371 	htable_t	*higher = NULL;
1372 	htable_t	*new = NULL;
1373 
1374 	if (level < 0 || level > TOP_LEVEL(hat))
1375 		panic("htable_create(): level %d out of range\n", level);
1376 
1377 	/*
1378 	 * Create the page tables in top down order.
1379 	 */
1380 	for (l = TOP_LEVEL(hat); l >= level; --l) {
1381 		new = NULL;
1382 		if (l == TOP_LEVEL(hat))
1383 			base = 0;
1384 		else
1385 			base = vaddr & LEVEL_MASK(l + 1);
1386 
1387 		h = HTABLE_HASH(hat, base, l);
1388 try_again:
1389 		/*
1390 		 * look up the htable at this level
1391 		 */
1392 		HTABLE_ENTER(h);
1393 		if (l == TOP_LEVEL(hat)) {
1394 			ht = hat->hat_htable;
1395 		} else {
1396 			for (ht = hat->hat_ht_hash[h]; ht; ht = ht->ht_next) {
1397 				ASSERT(ht->ht_hat == hat);
1398 				if (ht->ht_vaddr == base &&
1399 				    ht->ht_level == l)
1400 					break;
1401 			}
1402 		}
1403 
1404 		/*
1405 		 * if we found the htable, increment its busy cnt
1406 		 * and if we had allocated a new htable, free it.
1407 		 */
1408 		if (ht != NULL) {
1409 			/*
1410 			 * If we find a pre-existing shared table, it must
1411 			 * share from the same place.
1412 			 */
1413 			if (l == level && shared && ht->ht_shares &&
1414 			    ht->ht_shares != shared) {
1415 				panic("htable shared from wrong place "
1416 				    "found htable=%p shared=%p",
1417 				    (void *)ht, (void *)shared);
1418 			}
1419 			++ht->ht_busy;
1420 			HTABLE_EXIT(h);
1421 			if (new)
1422 				htable_free(new);
1423 			if (higher != NULL)
1424 				htable_release(higher);
1425 			higher = ht;
1426 
1427 		/*
1428 		 * if we didn't find it on the first search
1429 		 * allocate a new one and search again
1430 		 */
1431 		} else if (new == NULL) {
1432 			HTABLE_EXIT(h);
1433 			new = htable_alloc(hat, base, l,
1434 			    l == level ? shared : NULL);
1435 			goto try_again;
1436 
1437 		/*
1438 		 * 2nd search and still not there, use "new" table
1439 		 * Link new table into higher, when not at top level.
1440 		 */
1441 		} else {
1442 			ht = new;
1443 			if (higher != NULL) {
1444 				link_ptp(higher, ht, base);
1445 				ht->ht_parent = higher;
1446 			}
1447 			ht->ht_next = hat->hat_ht_hash[h];
1448 			ASSERT(ht->ht_prev == NULL);
1449 			if (hat->hat_ht_hash[h])
1450 				hat->hat_ht_hash[h]->ht_prev = ht;
1451 			hat->hat_ht_hash[h] = ht;
1452 			HTABLE_EXIT(h);
1453 
1454 			/*
1455 			 * Note we don't do htable_release(higher).
1456 			 * That happens recursively when "new" is removed by
1457 			 * htable_release() or htable_steal().
1458 			 */
1459 			higher = ht;
1460 
1461 			/*
1462 			 * If we just created a new shared page table we
1463 			 * increment the shared htable's busy count, so that
1464 			 * it can't be the victim of a steal even if it's empty.
1465 			 */
1466 			if (l == level && shared) {
1467 				(void) htable_lookup(shared->ht_hat,
1468 				    shared->ht_vaddr, shared->ht_level);
1469 				HATSTAT_INC(hs_htable_shared);
1470 			}
1471 		}
1472 	}
1473 
1474 	return (ht);
1475 }
1476 
1477 /*
1478  * Inherit initial pagetables from the boot program. On the 64-bit
1479  * hypervisor we also temporarily mark the p_index field of page table
1480  * pages, so we know not to try making them writable in seg_kpm.
1481  */
1482 void
1483 htable_attach(
1484 	hat_t *hat,
1485 	uintptr_t base,
1486 	level_t level,
1487 	htable_t *parent,
1488 	pfn_t pfn)
1489 {
1490 	htable_t	*ht;
1491 	uint_t		h;
1492 	uint_t		i;
1493 	x86pte_t	pte;
1494 	x86pte_t	*ptep;
1495 	page_t		*pp;
1496 	extern page_t	*boot_claim_page(pfn_t);
1497 
1498 	ht = htable_get_reserve();
1499 	if (level == mmu.max_level)
1500 		kas.a_hat->hat_htable = ht;
1501 	ht->ht_hat = hat;
1502 	ht->ht_parent = parent;
1503 	ht->ht_vaddr = base;
1504 	ht->ht_level = level;
1505 	ht->ht_busy = 1;
1506 	ht->ht_next = NULL;
1507 	ht->ht_prev = NULL;
1508 	ht->ht_flags = 0;
1509 	ht->ht_pfn = pfn;
1510 	ht->ht_lock_cnt = 0;
1511 	ht->ht_valid_cnt = 0;
1512 	if (parent != NULL)
1513 		++parent->ht_busy;
1514 
1515 	h = HTABLE_HASH(hat, base, level);
1516 	HTABLE_ENTER(h);
1517 	ht->ht_next = hat->hat_ht_hash[h];
1518 	ASSERT(ht->ht_prev == NULL);
1519 	if (hat->hat_ht_hash[h])
1520 		hat->hat_ht_hash[h]->ht_prev = ht;
1521 	hat->hat_ht_hash[h] = ht;
1522 	HTABLE_EXIT(h);
1523 
1524 	/*
1525 	 * make sure the page table physical page is not FREE
1526 	 */
1527 	if (page_resv(1, KM_NOSLEEP) == 0)
1528 		panic("page_resv() failed in ptable alloc");
1529 
1530 	pp = boot_claim_page(pfn);
1531 	ASSERT(pp != NULL);
1532 
1533 	/*
1534 	 * Page table pages that were allocated by dboot or
1535 	 * in very early startup didn't go through boot_mapin()
1536 	 * and so won't have vnode/offsets. Fix that here.
1537 	 */
1538 	if (pp->p_vnode == NULL) {
1539 		/* match offset calculation in page_get_physical() */
1540 		u_offset_t offset = (uintptr_t)ht;
1541 		if (offset > kernelbase)
1542 			offset -= kernelbase;
1543 		offset <<= MMU_PAGESHIFT;
1544 #if defined(__amd64)
1545 		offset += mmu.hole_start;	/* something in VA hole */
1546 #else
1547 		offset += 1ULL << 40;		/* something > 4 Gig */
1548 #endif
1549 		ASSERT(page_exists(&kvp, offset) == NULL);
1550 		(void) page_hashin(pp, &kvp, offset, NULL);
1551 	}
1552 	page_downgrade(pp);
1553 #if defined(__xpv) && defined(__amd64)
1554 	/*
1555 	 * Record in the page_t that is a pagetable for segkpm setup.
1556 	 */
1557 	if (kpm_vbase)
1558 		pp->p_index = 1;
1559 #endif
1560 
1561 	/*
1562 	 * Count valid mappings and recursively attach lower level pagetables.
1563 	 */
1564 	ptep = kbm_remap_window(pfn_to_pa(pfn), 0);
1565 	for (i = 0; i < HTABLE_NUM_PTES(ht); ++i) {
1566 		if (mmu.pae_hat)
1567 			pte = ptep[i];
1568 		else
1569 			pte = ((x86pte32_t *)ptep)[i];
1570 		if (!IN_HYPERVISOR_VA(base) && PTE_ISVALID(pte)) {
1571 			++ht->ht_valid_cnt;
1572 			if (!PTE_ISPAGE(pte, level)) {
1573 				htable_attach(hat, base, level - 1,
1574 				    ht, PTE2PFN(pte, level));
1575 				ptep = kbm_remap_window(pfn_to_pa(pfn), 0);
1576 			}
1577 		}
1578 		base += LEVEL_SIZE(level);
1579 		if (base == mmu.hole_start)
1580 			base = (mmu.hole_end + MMU_PAGEOFFSET) & MMU_PAGEMASK;
1581 	}
1582 
1583 	/*
1584 	 * As long as all the mappings we had were below kernel base
1585 	 * we can release the htable.
1586 	 */
1587 	if (base < kernelbase)
1588 		htable_release(ht);
1589 }
1590 
1591 /*
1592  * Walk through a given htable looking for the first valid entry.  This
1593  * routine takes both a starting and ending address.  The starting address
1594  * is required to be within the htable provided by the caller, but there is
1595  * no such restriction on the ending address.
1596  *
1597  * If the routine finds a valid entry in the htable (at or beyond the
1598  * starting address), the PTE (and its address) will be returned.
1599  * This PTE may correspond to either a page or a pagetable - it is the
1600  * caller's responsibility to determine which.  If no valid entry is
1601  * found, 0 (and invalid PTE) and the next unexamined address will be
1602  * returned.
1603  *
1604  * The loop has been carefully coded for optimization.
1605  */
1606 static x86pte_t
1607 htable_scan(htable_t *ht, uintptr_t *vap, uintptr_t eaddr)
1608 {
1609 	uint_t e;
1610 	x86pte_t found_pte = (x86pte_t)0;
1611 	caddr_t pte_ptr;
1612 	caddr_t end_pte_ptr;
1613 	int l = ht->ht_level;
1614 	uintptr_t va = *vap & LEVEL_MASK(l);
1615 	size_t pgsize = LEVEL_SIZE(l);
1616 
1617 	ASSERT(va >= ht->ht_vaddr);
1618 	ASSERT(va <= HTABLE_LAST_PAGE(ht));
1619 
1620 	/*
1621 	 * Compute the starting index and ending virtual address
1622 	 */
1623 	e = htable_va2entry(va, ht);
1624 
1625 	/*
1626 	 * The following page table scan code knows that the valid
1627 	 * bit of a PTE is in the lowest byte AND that x86 is little endian!!
1628 	 */
1629 	pte_ptr = (caddr_t)x86pte_access_pagetable(ht, 0);
1630 	end_pte_ptr = (caddr_t)PT_INDEX_PTR(pte_ptr, HTABLE_NUM_PTES(ht));
1631 	pte_ptr = (caddr_t)PT_INDEX_PTR((x86pte_t *)pte_ptr, e);
1632 	while (!PTE_ISVALID(*pte_ptr)) {
1633 		va += pgsize;
1634 		if (va >= eaddr)
1635 			break;
1636 		pte_ptr += mmu.pte_size;
1637 		ASSERT(pte_ptr <= end_pte_ptr);
1638 		if (pte_ptr == end_pte_ptr)
1639 			break;
1640 	}
1641 
1642 	/*
1643 	 * if we found a valid PTE, load the entire PTE
1644 	 */
1645 	if (va < eaddr && pte_ptr != end_pte_ptr)
1646 		found_pte = GET_PTE((x86pte_t *)pte_ptr);
1647 	x86pte_release_pagetable(ht);
1648 
1649 #if defined(__amd64)
1650 	/*
1651 	 * deal with VA hole on amd64
1652 	 */
1653 	if (l == mmu.max_level && va >= mmu.hole_start && va <= mmu.hole_end)
1654 		va = mmu.hole_end + va - mmu.hole_start;
1655 #endif /* __amd64 */
1656 
1657 	*vap = va;
1658 	return (found_pte);
1659 }
1660 
1661 /*
1662  * Find the address and htable for the first populated translation at or
1663  * above the given virtual address.  The caller may also specify an upper
1664  * limit to the address range to search.  Uses level information to quickly
1665  * skip unpopulated sections of virtual address spaces.
1666  *
1667  * If not found returns NULL. When found, returns the htable and virt addr
1668  * and has a hold on the htable.
1669  */
1670 x86pte_t
1671 htable_walk(
1672 	struct hat *hat,
1673 	htable_t **htp,
1674 	uintptr_t *vaddr,
1675 	uintptr_t eaddr)
1676 {
1677 	uintptr_t va = *vaddr;
1678 	htable_t *ht;
1679 	htable_t *prev = *htp;
1680 	level_t l;
1681 	level_t max_mapped_level;
1682 	x86pte_t pte;
1683 
1684 	ASSERT(eaddr > va);
1685 
1686 	/*
1687 	 * If this is a user address, then we know we need not look beyond
1688 	 * kernelbase.
1689 	 */
1690 	ASSERT(hat == kas.a_hat || eaddr <= kernelbase ||
1691 	    eaddr == HTABLE_WALK_TO_END);
1692 	if (hat != kas.a_hat && eaddr == HTABLE_WALK_TO_END)
1693 		eaddr = kernelbase;
1694 
1695 	/*
1696 	 * If we're coming in with a previous page table, search it first
1697 	 * without doing an htable_lookup(), this should be frequent.
1698 	 */
1699 	if (prev) {
1700 		ASSERT(prev->ht_busy > 0);
1701 		ASSERT(prev->ht_vaddr <= va);
1702 		l = prev->ht_level;
1703 		if (va <= HTABLE_LAST_PAGE(prev)) {
1704 			pte = htable_scan(prev, &va, eaddr);
1705 
1706 			if (PTE_ISPAGE(pte, l)) {
1707 				*vaddr = va;
1708 				*htp = prev;
1709 				return (pte);
1710 			}
1711 		}
1712 
1713 		/*
1714 		 * We found nothing in the htable provided by the caller,
1715 		 * so fall through and do the full search
1716 		 */
1717 		htable_release(prev);
1718 	}
1719 
1720 	/*
1721 	 * Find the level of the largest pagesize used by this HAT.
1722 	 */
1723 	if (hat->hat_ism_pgcnt > 0) {
1724 		max_mapped_level = mmu.umax_page_level;
1725 	} else {
1726 		max_mapped_level = 0;
1727 		for (l = 1; l <= mmu.max_page_level; ++l)
1728 			if (hat->hat_pages_mapped[l] != 0)
1729 				max_mapped_level = l;
1730 	}
1731 
1732 	while (va < eaddr && va >= *vaddr) {
1733 		/*
1734 		 *  Find lowest table with any entry for given address.
1735 		 */
1736 		for (l = 0; l <= TOP_LEVEL(hat); ++l) {
1737 			ht = htable_lookup(hat, va, l);
1738 			if (ht != NULL) {
1739 				pte = htable_scan(ht, &va, eaddr);
1740 				if (PTE_ISPAGE(pte, l)) {
1741 					VERIFY(!IN_VA_HOLE(va));
1742 					*vaddr = va;
1743 					*htp = ht;
1744 					return (pte);
1745 				}
1746 				htable_release(ht);
1747 				break;
1748 			}
1749 
1750 			/*
1751 			 * No htable at this level for the address. If there
1752 			 * is no larger page size that could cover it, we can
1753 			 * skip right to the start of the next page table.
1754 			 */
1755 			ASSERT(l < TOP_LEVEL(hat));
1756 			if (l >= max_mapped_level) {
1757 				va = NEXT_ENTRY_VA(va, l + 1);
1758 				if (va >= eaddr)
1759 					break;
1760 			}
1761 		}
1762 	}
1763 
1764 	*vaddr = 0;
1765 	*htp = NULL;
1766 	return (0);
1767 }
1768 
1769 /*
1770  * Find the htable and page table entry index of the given virtual address
1771  * with pagesize at or below given level.
1772  * If not found returns NULL. When found, returns the htable, sets
1773  * entry, and has a hold on the htable.
1774  */
1775 htable_t *
1776 htable_getpte(
1777 	struct hat *hat,
1778 	uintptr_t vaddr,
1779 	uint_t *entry,
1780 	x86pte_t *pte,
1781 	level_t level)
1782 {
1783 	htable_t	*ht;
1784 	level_t		l;
1785 	uint_t		e;
1786 
1787 	ASSERT(level <= mmu.max_page_level);
1788 
1789 	for (l = 0; l <= level; ++l) {
1790 		ht = htable_lookup(hat, vaddr, l);
1791 		if (ht == NULL)
1792 			continue;
1793 		e = htable_va2entry(vaddr, ht);
1794 		if (entry != NULL)
1795 			*entry = e;
1796 		if (pte != NULL)
1797 			*pte = x86pte_get(ht, e);
1798 		return (ht);
1799 	}
1800 	return (NULL);
1801 }
1802 
1803 /*
1804  * Find the htable and page table entry index of the given virtual address.
1805  * There must be a valid page mapped at the given address.
1806  * If not found returns NULL. When found, returns the htable, sets
1807  * entry, and has a hold on the htable.
1808  */
1809 htable_t *
1810 htable_getpage(struct hat *hat, uintptr_t vaddr, uint_t *entry)
1811 {
1812 	htable_t	*ht;
1813 	uint_t		e;
1814 	x86pte_t	pte;
1815 
1816 	ht = htable_getpte(hat, vaddr, &e, &pte, mmu.max_page_level);
1817 	if (ht == NULL)
1818 		return (NULL);
1819 
1820 	if (entry)
1821 		*entry = e;
1822 
1823 	if (PTE_ISPAGE(pte, ht->ht_level))
1824 		return (ht);
1825 	htable_release(ht);
1826 	return (NULL);
1827 }
1828 
1829 
1830 void
1831 htable_init()
1832 {
1833 	/*
1834 	 * To save on kernel VA usage, we avoid debug information in 32 bit
1835 	 * kernels.
1836 	 */
1837 #if defined(__amd64)
1838 	int	kmem_flags = KMC_NOHASH;
1839 #elif defined(__i386)
1840 	int	kmem_flags = KMC_NOHASH | KMC_NODEBUG;
1841 #endif
1842 
1843 	/*
1844 	 * initialize kmem caches
1845 	 */
1846 	htable_cache = kmem_cache_create("htable_t",
1847 	    sizeof (htable_t), 0, NULL, NULL,
1848 	    htable_reap, NULL, hat_memload_arena, kmem_flags);
1849 }
1850 
1851 /*
1852  * get the pte index for the virtual address in the given htable's pagetable
1853  */
1854 uint_t
1855 htable_va2entry(uintptr_t va, htable_t *ht)
1856 {
1857 	level_t	l = ht->ht_level;
1858 
1859 	ASSERT(va >= ht->ht_vaddr);
1860 	ASSERT(va <= HTABLE_LAST_PAGE(ht));
1861 	return ((va >> LEVEL_SHIFT(l)) & (HTABLE_NUM_PTES(ht) - 1));
1862 }
1863 
1864 /*
1865  * Given an htable and the index of a pte in it, return the virtual address
1866  * of the page.
1867  */
1868 uintptr_t
1869 htable_e2va(htable_t *ht, uint_t entry)
1870 {
1871 	level_t	l = ht->ht_level;
1872 	uintptr_t va;
1873 
1874 	ASSERT(entry < HTABLE_NUM_PTES(ht));
1875 	va = ht->ht_vaddr + ((uintptr_t)entry << LEVEL_SHIFT(l));
1876 
1877 	/*
1878 	 * Need to skip over any VA hole in top level table
1879 	 */
1880 #if defined(__amd64)
1881 	if (ht->ht_level == mmu.max_level && va >= mmu.hole_start)
1882 		va += ((mmu.hole_end - mmu.hole_start) + 1);
1883 #endif
1884 
1885 	return (va);
1886 }
1887 
1888 /*
1889  * The code uses compare and swap instructions to read/write PTE's to
1890  * avoid atomicity problems, since PTEs can be 8 bytes on 32 bit systems.
1891  * will naturally be atomic.
1892  *
1893  * The combination of using kpreempt_disable()/_enable() and the hci_mutex
1894  * are used to ensure that an interrupt won't overwrite a temporary mapping
1895  * while it's in use. If an interrupt thread tries to access a PTE, it will
1896  * yield briefly back to the pinned thread which holds the cpu's hci_mutex.
1897  */
1898 void
1899 x86pte_cpu_init(cpu_t *cpu)
1900 {
1901 	struct hat_cpu_info *hci;
1902 
1903 	hci = kmem_zalloc(sizeof (*hci), KM_SLEEP);
1904 	mutex_init(&hci->hci_mutex, NULL, MUTEX_DEFAULT, NULL);
1905 	cpu->cpu_hat_info = hci;
1906 }
1907 
1908 void
1909 x86pte_cpu_fini(cpu_t *cpu)
1910 {
1911 	struct hat_cpu_info *hci = cpu->cpu_hat_info;
1912 
1913 	kmem_free(hci, sizeof (*hci));
1914 	cpu->cpu_hat_info = NULL;
1915 }
1916 
1917 #ifdef __i386
1918 /*
1919  * On 32 bit kernels, loading a 64 bit PTE is a little tricky
1920  */
1921 x86pte_t
1922 get_pte64(x86pte_t *ptr)
1923 {
1924 	volatile uint32_t *p = (uint32_t *)ptr;
1925 	x86pte_t t;
1926 
1927 	ASSERT(mmu.pae_hat != 0);
1928 	for (;;) {
1929 		t = p[0];
1930 		t |= (uint64_t)p[1] << 32;
1931 		if ((t & 0xffffffff) == p[0])
1932 			return (t);
1933 	}
1934 }
1935 #endif /* __i386 */
1936 
1937 /*
1938  * Disable preemption and establish a mapping to the pagetable with the
1939  * given pfn. This is optimized for there case where it's the same
1940  * pfn as we last used referenced from this CPU.
1941  */
1942 static x86pte_t *
1943 x86pte_access_pagetable(htable_t *ht, uint_t index)
1944 {
1945 	/*
1946 	 * VLP pagetables are contained in the hat_t
1947 	 */
1948 	if (ht->ht_flags & HTABLE_VLP)
1949 		return (PT_INDEX_PTR(ht->ht_hat->hat_vlp_ptes, index));
1950 	return (x86pte_mapin(ht->ht_pfn, index, ht));
1951 }
1952 
1953 /*
1954  * map the given pfn into the page table window.
1955  */
1956 /*ARGSUSED*/
1957 x86pte_t *
1958 x86pte_mapin(pfn_t pfn, uint_t index, htable_t *ht)
1959 {
1960 	x86pte_t *pteptr;
1961 	x86pte_t pte = 0;
1962 	x86pte_t newpte;
1963 	int x;
1964 
1965 	ASSERT(pfn != PFN_INVALID);
1966 
1967 	if (!khat_running) {
1968 		caddr_t va = kbm_remap_window(pfn_to_pa(pfn), 1);
1969 		return (PT_INDEX_PTR(va, index));
1970 	}
1971 
1972 	/*
1973 	 * If kpm is available, use it.
1974 	 */
1975 	if (kpm_vbase)
1976 		return (PT_INDEX_PTR(hat_kpm_pfn2va(pfn), index));
1977 
1978 	/*
1979 	 * Disable preemption and grab the CPU's hci_mutex
1980 	 */
1981 	kpreempt_disable();
1982 	ASSERT(CPU->cpu_hat_info != NULL);
1983 	mutex_enter(&CPU->cpu_hat_info->hci_mutex);
1984 	x = PWIN_TABLE(CPU->cpu_id);
1985 	pteptr = (x86pte_t *)PWIN_PTE_VA(x);
1986 #ifndef __xpv
1987 	if (mmu.pae_hat)
1988 		pte = *pteptr;
1989 	else
1990 		pte = *(x86pte32_t *)pteptr;
1991 #endif
1992 
1993 	newpte = MAKEPTE(pfn, 0) | mmu.pt_global | mmu.pt_nx;
1994 
1995 	/*
1996 	 * For hardware we can use a writable mapping.
1997 	 */
1998 #ifdef __xpv
1999 	if (IN_XPV_PANIC())
2000 #endif
2001 		newpte |= PT_WRITABLE;
2002 
2003 	if (!PTE_EQUIV(newpte, pte)) {
2004 
2005 #ifdef __xpv
2006 		if (!IN_XPV_PANIC()) {
2007 			xen_map(newpte, PWIN_VA(x));
2008 		} else
2009 #endif
2010 		{
2011 			XPV_ALLOW_PAGETABLE_UPDATES();
2012 			if (mmu.pae_hat)
2013 				*pteptr = newpte;
2014 			else
2015 				*(x86pte32_t *)pteptr = newpte;
2016 			XPV_DISALLOW_PAGETABLE_UPDATES();
2017 			mmu_tlbflush_entry((caddr_t)(PWIN_VA(x)));
2018 		}
2019 	}
2020 	return (PT_INDEX_PTR(PWIN_VA(x), index));
2021 }
2022 
2023 /*
2024  * Release access to a page table.
2025  */
2026 static void
2027 x86pte_release_pagetable(htable_t *ht)
2028 {
2029 	/*
2030 	 * nothing to do for VLP htables
2031 	 */
2032 	if (ht->ht_flags & HTABLE_VLP)
2033 		return;
2034 
2035 	x86pte_mapout();
2036 }
2037 
2038 void
2039 x86pte_mapout(void)
2040 {
2041 	if (kpm_vbase != NULL || !khat_running)
2042 		return;
2043 
2044 	/*
2045 	 * Drop the CPU's hci_mutex and restore preemption.
2046 	 */
2047 #ifdef __xpv
2048 	if (!IN_XPV_PANIC()) {
2049 		uintptr_t va;
2050 
2051 		/*
2052 		 * We need to always clear the mapping in case a page
2053 		 * that was once a page table page is ballooned out.
2054 		 */
2055 		va = (uintptr_t)PWIN_VA(PWIN_TABLE(CPU->cpu_id));
2056 		(void) HYPERVISOR_update_va_mapping(va, 0,
2057 		    UVMF_INVLPG | UVMF_LOCAL);
2058 	}
2059 #endif
2060 	mutex_exit(&CPU->cpu_hat_info->hci_mutex);
2061 	kpreempt_enable();
2062 }
2063 
2064 /*
2065  * Atomic retrieval of a pagetable entry
2066  */
2067 x86pte_t
2068 x86pte_get(htable_t *ht, uint_t entry)
2069 {
2070 	x86pte_t	pte;
2071 	x86pte_t	*ptep;
2072 
2073 	/*
2074 	 * Be careful that loading PAE entries in 32 bit kernel is atomic.
2075 	 */
2076 	ASSERT(entry < mmu.ptes_per_table);
2077 	ptep = x86pte_access_pagetable(ht, entry);
2078 	pte = GET_PTE(ptep);
2079 	x86pte_release_pagetable(ht);
2080 	return (pte);
2081 }
2082 
2083 /*
2084  * Atomic unconditional set of a page table entry, it returns the previous
2085  * value. For pre-existing mappings if the PFN changes, then we don't care
2086  * about the old pte's REF / MOD bits. If the PFN remains the same, we leave
2087  * the MOD/REF bits unchanged.
2088  *
2089  * If asked to overwrite a link to a lower page table with a large page
2090  * mapping, this routine returns the special value of LPAGE_ERROR. This
2091  * allows the upper HAT layers to retry with a smaller mapping size.
2092  */
2093 x86pte_t
2094 x86pte_set(htable_t *ht, uint_t entry, x86pte_t new, void *ptr)
2095 {
2096 	x86pte_t	old;
2097 	x86pte_t	prev;
2098 	x86pte_t	*ptep;
2099 	level_t		l = ht->ht_level;
2100 	x86pte_t	pfn_mask = (l != 0) ? PT_PADDR_LGPG : PT_PADDR;
2101 	x86pte_t	n;
2102 	uintptr_t	addr = htable_e2va(ht, entry);
2103 	hat_t		*hat = ht->ht_hat;
2104 
2105 	ASSERT(new != 0); /* don't use to invalidate a PTE, see x86pte_update */
2106 	ASSERT(!(ht->ht_flags & HTABLE_SHARED_PFN));
2107 	if (ptr == NULL)
2108 		ptep = x86pte_access_pagetable(ht, entry);
2109 	else
2110 		ptep = ptr;
2111 
2112 	/*
2113 	 * Install the new PTE. If remapping the same PFN, then
2114 	 * copy existing REF/MOD bits to new mapping.
2115 	 */
2116 	do {
2117 		prev = GET_PTE(ptep);
2118 		n = new;
2119 		if (PTE_ISVALID(n) && (prev & pfn_mask) == (new & pfn_mask))
2120 			n |= prev & (PT_REF | PT_MOD);
2121 
2122 		/*
2123 		 * Another thread may have installed this mapping already,
2124 		 * flush the local TLB and be done.
2125 		 */
2126 		if (prev == n) {
2127 			old = new;
2128 #ifdef __xpv
2129 			if (!IN_XPV_PANIC())
2130 				xen_flush_va((caddr_t)addr);
2131 			else
2132 #endif
2133 				mmu_tlbflush_entry((caddr_t)addr);
2134 			goto done;
2135 		}
2136 
2137 		/*
2138 		 * Detect if we have a collision of installing a large
2139 		 * page mapping where there already is a lower page table.
2140 		 */
2141 		if (l > 0 && (prev & PT_VALID) && !(prev & PT_PAGESIZE)) {
2142 			old = LPAGE_ERROR;
2143 			goto done;
2144 		}
2145 
2146 		XPV_ALLOW_PAGETABLE_UPDATES();
2147 		old = CAS_PTE(ptep, prev, n);
2148 		XPV_DISALLOW_PAGETABLE_UPDATES();
2149 	} while (old != prev);
2150 
2151 	/*
2152 	 * Do a TLB demap if needed, ie. the old pte was valid.
2153 	 *
2154 	 * Note that a stale TLB writeback to the PTE here either can't happen
2155 	 * or doesn't matter. The PFN can only change for NOSYNC|NOCONSIST
2156 	 * mappings, but they were created with REF and MOD already set, so
2157 	 * no stale writeback will happen.
2158 	 *
2159 	 * Segmap is the only place where remaps happen on the same pfn and for
2160 	 * that we want to preserve the stale REF/MOD bits.
2161 	 */
2162 	if (old & PT_REF)
2163 		hat_tlb_inval(hat, addr);
2164 
2165 done:
2166 	if (ptr == NULL)
2167 		x86pte_release_pagetable(ht);
2168 	return (old);
2169 }
2170 
2171 /*
2172  * Atomic compare and swap of a page table entry. No TLB invalidates are done.
2173  * This is used for links between pagetables of different levels.
2174  * Note we always create these links with dirty/access set, so they should
2175  * never change.
2176  */
2177 x86pte_t
2178 x86pte_cas(htable_t *ht, uint_t entry, x86pte_t old, x86pte_t new)
2179 {
2180 	x86pte_t	pte;
2181 	x86pte_t	*ptep;
2182 #ifdef __xpv
2183 	/*
2184 	 * We can't use writable pagetables for upper level tables, so fake it.
2185 	 */
2186 	mmu_update_t t[2];
2187 	int cnt = 1;
2188 	int count;
2189 	maddr_t ma;
2190 
2191 	if (!IN_XPV_PANIC()) {
2192 		ASSERT(!(ht->ht_flags & HTABLE_VLP));	/* no VLP yet */
2193 		ma = pa_to_ma(PT_INDEX_PHYSADDR(pfn_to_pa(ht->ht_pfn), entry));
2194 		t[0].ptr = ma | MMU_NORMAL_PT_UPDATE;
2195 		t[0].val = new;
2196 
2197 #if defined(__amd64)
2198 		/*
2199 		 * On the 64-bit hypervisor we need to maintain the user mode
2200 		 * top page table too.
2201 		 */
2202 		if (ht->ht_level == mmu.max_level && ht->ht_hat != kas.a_hat) {
2203 			ma = pa_to_ma(PT_INDEX_PHYSADDR(pfn_to_pa(
2204 			    ht->ht_hat->hat_user_ptable), entry));
2205 			t[1].ptr = ma | MMU_NORMAL_PT_UPDATE;
2206 			t[1].val = new;
2207 			++cnt;
2208 		}
2209 #endif	/* __amd64 */
2210 
2211 		if (HYPERVISOR_mmu_update(t, cnt, &count, DOMID_SELF))
2212 			panic("HYPERVISOR_mmu_update() failed");
2213 		ASSERT(count == cnt);
2214 		return (old);
2215 	}
2216 #endif
2217 	ptep = x86pte_access_pagetable(ht, entry);
2218 	XPV_ALLOW_PAGETABLE_UPDATES();
2219 	pte = CAS_PTE(ptep, old, new);
2220 	XPV_DISALLOW_PAGETABLE_UPDATES();
2221 	x86pte_release_pagetable(ht);
2222 	return (pte);
2223 }
2224 
2225 /*
2226  * Invalidate a page table entry as long as it currently maps something that
2227  * matches the value determined by expect.
2228  *
2229  * If tlb is set, also invalidates any TLB entries.
2230  *
2231  * Returns the previous value of the PTE.
2232  */
2233 x86pte_t
2234 x86pte_inval(
2235 	htable_t *ht,
2236 	uint_t entry,
2237 	x86pte_t expect,
2238 	x86pte_t *pte_ptr,
2239 	boolean_t tlb)
2240 {
2241 	x86pte_t	*ptep;
2242 	x86pte_t	oldpte;
2243 	x86pte_t	found;
2244 
2245 	ASSERT(!(ht->ht_flags & HTABLE_SHARED_PFN));
2246 	ASSERT(ht->ht_level <= mmu.max_page_level);
2247 
2248 	if (pte_ptr != NULL)
2249 		ptep = pte_ptr;
2250 	else
2251 		ptep = x86pte_access_pagetable(ht, entry);
2252 
2253 #if defined(__xpv)
2254 	/*
2255 	 * If exit()ing just use HYPERVISOR_mmu_update(), as we can't be racing
2256 	 * with anything else.
2257 	 */
2258 	if ((ht->ht_hat->hat_flags & HAT_FREEING) && !IN_XPV_PANIC()) {
2259 		int count;
2260 		mmu_update_t t[1];
2261 		maddr_t ma;
2262 
2263 		oldpte = GET_PTE(ptep);
2264 		if (expect != 0 && (oldpte & PT_PADDR) != (expect & PT_PADDR))
2265 			goto done;
2266 		ma = pa_to_ma(PT_INDEX_PHYSADDR(pfn_to_pa(ht->ht_pfn), entry));
2267 		t[0].ptr = ma | MMU_NORMAL_PT_UPDATE;
2268 		t[0].val = 0;
2269 		if (HYPERVISOR_mmu_update(t, 1, &count, DOMID_SELF))
2270 			panic("HYPERVISOR_mmu_update() failed");
2271 		ASSERT(count == 1);
2272 		goto done;
2273 	}
2274 #endif /* __xpv */
2275 
2276 	/*
2277 	 * Note that the loop is needed to handle changes due to h/w updating
2278 	 * of PT_MOD/PT_REF.
2279 	 */
2280 	do {
2281 		oldpte = GET_PTE(ptep);
2282 		if (expect != 0 && (oldpte & PT_PADDR) != (expect & PT_PADDR))
2283 			goto done;
2284 		XPV_ALLOW_PAGETABLE_UPDATES();
2285 		found = CAS_PTE(ptep, oldpte, 0);
2286 		XPV_DISALLOW_PAGETABLE_UPDATES();
2287 	} while (found != oldpte);
2288 	if (tlb && (oldpte & (PT_REF | PT_MOD)))
2289 		hat_tlb_inval(ht->ht_hat, htable_e2va(ht, entry));
2290 
2291 done:
2292 	if (pte_ptr == NULL)
2293 		x86pte_release_pagetable(ht);
2294 	return (oldpte);
2295 }
2296 
2297 /*
2298  * Change a page table entry af it currently matches the value in expect.
2299  */
2300 x86pte_t
2301 x86pte_update(
2302 	htable_t *ht,
2303 	uint_t entry,
2304 	x86pte_t expect,
2305 	x86pte_t new)
2306 {
2307 	x86pte_t	*ptep;
2308 	x86pte_t	found;
2309 
2310 	ASSERT(new != 0);
2311 	ASSERT(!(ht->ht_flags & HTABLE_SHARED_PFN));
2312 	ASSERT(ht->ht_level <= mmu.max_page_level);
2313 
2314 	ptep = x86pte_access_pagetable(ht, entry);
2315 	XPV_ALLOW_PAGETABLE_UPDATES();
2316 	found = CAS_PTE(ptep, expect, new);
2317 	XPV_DISALLOW_PAGETABLE_UPDATES();
2318 	if (found == expect) {
2319 		hat_tlb_inval(ht->ht_hat, htable_e2va(ht, entry));
2320 
2321 		/*
2322 		 * When removing write permission *and* clearing the
2323 		 * MOD bit, check if a write happened via a stale
2324 		 * TLB entry before the TLB shootdown finished.
2325 		 *
2326 		 * If it did happen, simply re-enable write permission and
2327 		 * act like the original CAS failed.
2328 		 */
2329 		if ((expect & (PT_WRITABLE | PT_MOD)) == PT_WRITABLE &&
2330 		    (new & (PT_WRITABLE | PT_MOD)) == 0 &&
2331 		    (GET_PTE(ptep) & PT_MOD) != 0) {
2332 			do {
2333 				found = GET_PTE(ptep);
2334 				XPV_ALLOW_PAGETABLE_UPDATES();
2335 				found =
2336 				    CAS_PTE(ptep, found, found | PT_WRITABLE);
2337 				XPV_DISALLOW_PAGETABLE_UPDATES();
2338 			} while ((found & PT_WRITABLE) == 0);
2339 		}
2340 	}
2341 	x86pte_release_pagetable(ht);
2342 	return (found);
2343 }
2344 
2345 #ifndef __xpv
2346 /*
2347  * Copy page tables - this is just a little more complicated than the
2348  * previous routines. Note that it's also not atomic! It also is never
2349  * used for VLP pagetables.
2350  */
2351 void
2352 x86pte_copy(htable_t *src, htable_t *dest, uint_t entry, uint_t count)
2353 {
2354 	caddr_t	src_va;
2355 	caddr_t dst_va;
2356 	size_t size;
2357 	x86pte_t *pteptr;
2358 	x86pte_t pte;
2359 
2360 	ASSERT(khat_running);
2361 	ASSERT(!(dest->ht_flags & HTABLE_VLP));
2362 	ASSERT(!(src->ht_flags & HTABLE_VLP));
2363 	ASSERT(!(src->ht_flags & HTABLE_SHARED_PFN));
2364 	ASSERT(!(dest->ht_flags & HTABLE_SHARED_PFN));
2365 
2366 	/*
2367 	 * Acquire access to the CPU pagetable windows for the dest and source.
2368 	 */
2369 	dst_va = (caddr_t)x86pte_access_pagetable(dest, entry);
2370 	if (kpm_vbase) {
2371 		src_va = (caddr_t)
2372 		    PT_INDEX_PTR(hat_kpm_pfn2va(src->ht_pfn), entry);
2373 	} else {
2374 		uint_t x = PWIN_SRC(CPU->cpu_id);
2375 
2376 		/*
2377 		 * Finish defining the src pagetable mapping
2378 		 */
2379 		src_va = (caddr_t)PT_INDEX_PTR(PWIN_VA(x), entry);
2380 		pte = MAKEPTE(src->ht_pfn, 0) | mmu.pt_global | mmu.pt_nx;
2381 		pteptr = (x86pte_t *)PWIN_PTE_VA(x);
2382 		if (mmu.pae_hat)
2383 			*pteptr = pte;
2384 		else
2385 			*(x86pte32_t *)pteptr = pte;
2386 		mmu_tlbflush_entry((caddr_t)(PWIN_VA(x)));
2387 	}
2388 
2389 	/*
2390 	 * now do the copy
2391 	 */
2392 	size = count << mmu.pte_size_shift;
2393 	bcopy(src_va, dst_va, size);
2394 
2395 	x86pte_release_pagetable(dest);
2396 }
2397 
2398 #else /* __xpv */
2399 
2400 /*
2401  * The hypervisor only supports writable pagetables at level 0, so we have
2402  * to install these 1 by 1 the slow way.
2403  */
2404 void
2405 x86pte_copy(htable_t *src, htable_t *dest, uint_t entry, uint_t count)
2406 {
2407 	caddr_t	src_va;
2408 	x86pte_t pte;
2409 
2410 	ASSERT(!IN_XPV_PANIC());
2411 	src_va = (caddr_t)x86pte_access_pagetable(src, entry);
2412 	while (count) {
2413 		if (mmu.pae_hat)
2414 			pte = *(x86pte_t *)src_va;
2415 		else
2416 			pte = *(x86pte32_t *)src_va;
2417 		if (pte != 0) {
2418 			set_pteval(pfn_to_pa(dest->ht_pfn), entry,
2419 			    dest->ht_level, pte);
2420 #ifdef __amd64
2421 			if (dest->ht_level == mmu.max_level &&
2422 			    htable_e2va(dest, entry) < HYPERVISOR_VIRT_END)
2423 				set_pteval(
2424 				    pfn_to_pa(dest->ht_hat->hat_user_ptable),
2425 				    entry, dest->ht_level, pte);
2426 #endif
2427 		}
2428 		--count;
2429 		++entry;
2430 		src_va += mmu.pte_size;
2431 	}
2432 	x86pte_release_pagetable(src);
2433 }
2434 #endif /* __xpv */
2435 
2436 /*
2437  * Zero page table entries - Note this doesn't use atomic stores!
2438  */
2439 static void
2440 x86pte_zero(htable_t *dest, uint_t entry, uint_t count)
2441 {
2442 	caddr_t dst_va;
2443 	size_t size;
2444 #ifdef __xpv
2445 	int x;
2446 	x86pte_t newpte;
2447 #endif
2448 
2449 	/*
2450 	 * Map in the page table to be zeroed.
2451 	 */
2452 	ASSERT(!(dest->ht_flags & HTABLE_SHARED_PFN));
2453 	ASSERT(!(dest->ht_flags & HTABLE_VLP));
2454 
2455 	/*
2456 	 * On the hypervisor we don't use x86pte_access_pagetable() since
2457 	 * in this case the page is not pinned yet.
2458 	 */
2459 #ifdef __xpv
2460 	if (kpm_vbase == NULL) {
2461 		kpreempt_disable();
2462 		ASSERT(CPU->cpu_hat_info != NULL);
2463 		mutex_enter(&CPU->cpu_hat_info->hci_mutex);
2464 		x = PWIN_TABLE(CPU->cpu_id);
2465 		newpte = MAKEPTE(dest->ht_pfn, 0) | PT_WRITABLE;
2466 		xen_map(newpte, PWIN_VA(x));
2467 		dst_va = (caddr_t)PT_INDEX_PTR(PWIN_VA(x), entry);
2468 	} else
2469 #endif
2470 		dst_va = (caddr_t)x86pte_access_pagetable(dest, entry);
2471 
2472 	size = count << mmu.pte_size_shift;
2473 	ASSERT(size > BLOCKZEROALIGN);
2474 #ifdef __i386
2475 	if (!is_x86_feature(x86_featureset, X86FSET_SSE2))
2476 		bzero(dst_va, size);
2477 	else
2478 #endif
2479 		block_zero_no_xmm(dst_va, size);
2480 
2481 #ifdef __xpv
2482 	if (kpm_vbase == NULL) {
2483 		xen_map(0, PWIN_VA(x));
2484 		mutex_exit(&CPU->cpu_hat_info->hci_mutex);
2485 		kpreempt_enable();
2486 	} else
2487 #endif
2488 		x86pte_release_pagetable(dest);
2489 }
2490 
2491 /*
2492  * Called to ensure that all pagetables are in the system dump
2493  */
2494 void
2495 hat_dump(void)
2496 {
2497 	hat_t *hat;
2498 	uint_t h;
2499 	htable_t *ht;
2500 
2501 	/*
2502 	 * Dump all page tables
2503 	 */
2504 	for (hat = kas.a_hat; hat != NULL; hat = hat->hat_next) {
2505 		for (h = 0; h < hat->hat_num_hash; ++h) {
2506 			for (ht = hat->hat_ht_hash[h]; ht; ht = ht->ht_next) {
2507 				if ((ht->ht_flags & HTABLE_VLP) == 0)
2508 					dump_page(ht->ht_pfn);
2509 			}
2510 		}
2511 	}
2512 }
2513