xref: /freebsd/sys/x86/iommu/intel_idpgtbl.c (revision 389e4940069316fe667ffa263fa7d6390d0a960f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2013 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/malloc.h>
38 #include <sys/bus.h>
39 #include <sys/interrupt.h>
40 #include <sys/kernel.h>
41 #include <sys/ktr.h>
42 #include <sys/lock.h>
43 #include <sys/memdesc.h>
44 #include <sys/mutex.h>
45 #include <sys/proc.h>
46 #include <sys/rwlock.h>
47 #include <sys/rman.h>
48 #include <sys/sf_buf.h>
49 #include <sys/sysctl.h>
50 #include <sys/taskqueue.h>
51 #include <sys/tree.h>
52 #include <sys/uio.h>
53 #include <sys/vmem.h>
54 #include <vm/vm.h>
55 #include <vm/vm_extern.h>
56 #include <vm/vm_kern.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_pager.h>
60 #include <vm/vm_map.h>
61 #include <machine/atomic.h>
62 #include <machine/bus.h>
63 #include <machine/cpu.h>
64 #include <machine/md_var.h>
65 #include <machine/specialreg.h>
66 #include <x86/include/busdma_impl.h>
67 #include <x86/iommu/intel_reg.h>
68 #include <x86/iommu/busdma_dmar.h>
69 #include <x86/iommu/intel_dmar.h>
70 
71 static int domain_unmap_buf_locked(struct dmar_domain *domain,
72     dmar_gaddr_t base, dmar_gaddr_t size, int flags);
73 
74 /*
75  * The cache of the identity mapping page tables for the DMARs.  Using
76  * the cache saves significant amount of memory for page tables by
77  * reusing the page tables, since usually DMARs are identical and have
78  * the same capabilities.  Still, cache records the information needed
79  * to match DMAR capabilities and page table format, to correctly
80  * handle different DMARs.
81  */
82 
83 struct idpgtbl {
84 	dmar_gaddr_t maxaddr;	/* Page table covers the guest address
85 				   range [0..maxaddr) */
86 	int pglvl;		/* Total page table levels ignoring
87 				   superpages */
88 	int leaf;		/* The last materialized page table
89 				   level, it is non-zero if superpages
90 				   are supported */
91 	vm_object_t pgtbl_obj;	/* The page table pages */
92 	LIST_ENTRY(idpgtbl) link;
93 };
94 
95 static struct sx idpgtbl_lock;
96 SX_SYSINIT(idpgtbl, &idpgtbl_lock, "idpgtbl");
97 static LIST_HEAD(, idpgtbl) idpgtbls = LIST_HEAD_INITIALIZER(idpgtbls);
98 static MALLOC_DEFINE(M_DMAR_IDPGTBL, "dmar_idpgtbl",
99     "Intel DMAR Identity mappings cache elements");
100 
101 /*
102  * Build the next level of the page tables for the identity mapping.
103  * - lvl is the level to build;
104  * - idx is the index of the page table page in the pgtbl_obj, which is
105  *   being allocated filled now;
106  * - addr is the starting address in the bus address space which is
107  *   mapped by the page table page.
108  */
109 static void
110 domain_idmap_nextlvl(struct idpgtbl *tbl, int lvl, vm_pindex_t idx,
111     dmar_gaddr_t addr)
112 {
113 	vm_page_t m1;
114 	dmar_pte_t *pte;
115 	struct sf_buf *sf;
116 	dmar_gaddr_t f, pg_sz;
117 	vm_pindex_t base;
118 	int i;
119 
120 	VM_OBJECT_ASSERT_LOCKED(tbl->pgtbl_obj);
121 	if (addr >= tbl->maxaddr)
122 		return;
123 	(void)dmar_pgalloc(tbl->pgtbl_obj, idx, DMAR_PGF_OBJL | DMAR_PGF_WAITOK |
124 	    DMAR_PGF_ZERO);
125 	base = idx * DMAR_NPTEPG + 1; /* Index of the first child page of idx */
126 	pg_sz = pglvl_page_size(tbl->pglvl, lvl);
127 	if (lvl != tbl->leaf) {
128 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz)
129 			domain_idmap_nextlvl(tbl, lvl + 1, base + i, f);
130 	}
131 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
132 	pte = dmar_map_pgtbl(tbl->pgtbl_obj, idx, DMAR_PGF_WAITOK, &sf);
133 	if (lvl == tbl->leaf) {
134 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) {
135 			if (f >= tbl->maxaddr)
136 				break;
137 			pte[i].pte = (DMAR_PTE_ADDR_MASK & f) |
138 			    DMAR_PTE_R | DMAR_PTE_W;
139 		}
140 	} else {
141 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) {
142 			if (f >= tbl->maxaddr)
143 				break;
144 			m1 = dmar_pgalloc(tbl->pgtbl_obj, base + i,
145 			    DMAR_PGF_NOALLOC);
146 			KASSERT(m1 != NULL, ("lost page table page"));
147 			pte[i].pte = (DMAR_PTE_ADDR_MASK &
148 			    VM_PAGE_TO_PHYS(m1)) | DMAR_PTE_R | DMAR_PTE_W;
149 		}
150 	}
151 	/* domain_get_idmap_pgtbl flushes CPU cache if needed. */
152 	dmar_unmap_pgtbl(sf);
153 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
154 }
155 
156 /*
157  * Find a ready and compatible identity-mapping page table in the
158  * cache. If not found, populate the identity-mapping page table for
159  * the context, up to the maxaddr. The maxaddr byte is allowed to be
160  * not mapped, which is aligned with the definition of Maxmem as the
161  * highest usable physical address + 1.  If superpages are used, the
162  * maxaddr is typically mapped.
163  */
164 vm_object_t
165 domain_get_idmap_pgtbl(struct dmar_domain *domain, dmar_gaddr_t maxaddr)
166 {
167 	struct dmar_unit *unit;
168 	struct idpgtbl *tbl;
169 	vm_object_t res;
170 	vm_page_t m;
171 	int leaf, i;
172 
173 	leaf = 0; /* silence gcc */
174 
175 	/*
176 	 * First, determine where to stop the paging structures.
177 	 */
178 	for (i = 0; i < domain->pglvl; i++) {
179 		if (i == domain->pglvl - 1 || domain_is_sp_lvl(domain, i)) {
180 			leaf = i;
181 			break;
182 		}
183 	}
184 
185 	/*
186 	 * Search the cache for a compatible page table.  Qualified
187 	 * page table must map up to maxaddr, its level must be
188 	 * supported by the DMAR and leaf should be equal to the
189 	 * calculated value.  The later restriction could be lifted
190 	 * but I believe it is currently impossible to have any
191 	 * deviations for existing hardware.
192 	 */
193 	sx_slock(&idpgtbl_lock);
194 	LIST_FOREACH(tbl, &idpgtbls, link) {
195 		if (tbl->maxaddr >= maxaddr &&
196 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
197 		    tbl->leaf == leaf) {
198 			res = tbl->pgtbl_obj;
199 			vm_object_reference(res);
200 			sx_sunlock(&idpgtbl_lock);
201 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
202 			goto end;
203 		}
204 	}
205 
206 	/*
207 	 * Not found in cache, relock the cache into exclusive mode to
208 	 * be able to add element, and recheck cache again after the
209 	 * relock.
210 	 */
211 	sx_sunlock(&idpgtbl_lock);
212 	sx_xlock(&idpgtbl_lock);
213 	LIST_FOREACH(tbl, &idpgtbls, link) {
214 		if (tbl->maxaddr >= maxaddr &&
215 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
216 		    tbl->leaf == leaf) {
217 			res = tbl->pgtbl_obj;
218 			vm_object_reference(res);
219 			sx_xunlock(&idpgtbl_lock);
220 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
221 			return (res);
222 		}
223 	}
224 
225 	/*
226 	 * Still not found, create new page table.
227 	 */
228 	tbl = malloc(sizeof(*tbl), M_DMAR_IDPGTBL, M_WAITOK);
229 	tbl->pglvl = domain->pglvl;
230 	tbl->leaf = leaf;
231 	tbl->maxaddr = maxaddr;
232 	tbl->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
233 	    IDX_TO_OFF(pglvl_max_pages(tbl->pglvl)), 0, 0, NULL);
234 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
235 	domain_idmap_nextlvl(tbl, 0, 0, 0);
236 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
237 	LIST_INSERT_HEAD(&idpgtbls, tbl, link);
238 	res = tbl->pgtbl_obj;
239 	vm_object_reference(res);
240 	sx_xunlock(&idpgtbl_lock);
241 
242 end:
243 	/*
244 	 * Table was found or created.
245 	 *
246 	 * If DMAR does not snoop paging structures accesses, flush
247 	 * CPU cache to memory.  Note that dmar_unmap_pgtbl() coherent
248 	 * argument was possibly invalid at the time of the identity
249 	 * page table creation, since DMAR which was passed at the
250 	 * time of creation could be coherent, while current DMAR is
251 	 * not.
252 	 *
253 	 * If DMAR cannot look into the chipset write buffer, flush it
254 	 * as well.
255 	 */
256 	unit = domain->dmar;
257 	if (!DMAR_IS_COHERENT(unit)) {
258 		VM_OBJECT_WLOCK(res);
259 		for (m = vm_page_lookup(res, 0); m != NULL;
260 		     m = vm_page_next(m))
261 			pmap_invalidate_cache_pages(&m, 1);
262 		VM_OBJECT_WUNLOCK(res);
263 	}
264 	if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
265 		DMAR_LOCK(unit);
266 		dmar_flush_write_bufs(unit);
267 		DMAR_UNLOCK(unit);
268 	}
269 
270 	return (res);
271 }
272 
273 /*
274  * Return a reference to the identity mapping page table to the cache.
275  */
276 void
277 put_idmap_pgtbl(vm_object_t obj)
278 {
279 	struct idpgtbl *tbl, *tbl1;
280 	vm_object_t rmobj;
281 
282 	sx_slock(&idpgtbl_lock);
283 	KASSERT(obj->ref_count >= 2, ("lost cache reference"));
284 	vm_object_deallocate(obj);
285 
286 	/*
287 	 * Cache always owns one last reference on the page table object.
288 	 * If there is an additional reference, object must stay.
289 	 */
290 	if (obj->ref_count > 1) {
291 		sx_sunlock(&idpgtbl_lock);
292 		return;
293 	}
294 
295 	/*
296 	 * Cache reference is the last, remove cache element and free
297 	 * page table object, returning the page table pages to the
298 	 * system.
299 	 */
300 	sx_sunlock(&idpgtbl_lock);
301 	sx_xlock(&idpgtbl_lock);
302 	LIST_FOREACH_SAFE(tbl, &idpgtbls, link, tbl1) {
303 		rmobj = tbl->pgtbl_obj;
304 		if (rmobj->ref_count == 1) {
305 			LIST_REMOVE(tbl, link);
306 			atomic_subtract_int(&dmar_tbl_pagecnt,
307 			    rmobj->resident_page_count);
308 			vm_object_deallocate(rmobj);
309 			free(tbl, M_DMAR_IDPGTBL);
310 		}
311 	}
312 	sx_xunlock(&idpgtbl_lock);
313 }
314 
315 /*
316  * The core routines to map and unmap host pages at the given guest
317  * address.  Support superpages.
318  */
319 
320 /*
321  * Index of the pte for the guest address base in the page table at
322  * the level lvl.
323  */
324 static int
325 domain_pgtbl_pte_off(struct dmar_domain *domain, dmar_gaddr_t base, int lvl)
326 {
327 
328 	base >>= DMAR_PAGE_SHIFT + (domain->pglvl - lvl - 1) *
329 	    DMAR_NPTEPGSHIFT;
330 	return (base & DMAR_PTEMASK);
331 }
332 
333 /*
334  * Returns the page index of the page table page in the page table
335  * object, which maps the given address base at the page table level
336  * lvl.
337  */
338 static vm_pindex_t
339 domain_pgtbl_get_pindex(struct dmar_domain *domain, dmar_gaddr_t base, int lvl)
340 {
341 	vm_pindex_t idx, pidx;
342 	int i;
343 
344 	KASSERT(lvl >= 0 && lvl < domain->pglvl,
345 	    ("wrong lvl %p %d", domain, lvl));
346 
347 	for (pidx = idx = 0, i = 0; i < lvl; i++, pidx = idx) {
348 		idx = domain_pgtbl_pte_off(domain, base, i) +
349 		    pidx * DMAR_NPTEPG + 1;
350 	}
351 	return (idx);
352 }
353 
354 static dmar_pte_t *
355 domain_pgtbl_map_pte(struct dmar_domain *domain, dmar_gaddr_t base, int lvl,
356     int flags, vm_pindex_t *idxp, struct sf_buf **sf)
357 {
358 	vm_page_t m;
359 	struct sf_buf *sfp;
360 	dmar_pte_t *pte, *ptep;
361 	vm_pindex_t idx, idx1;
362 
363 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
364 	KASSERT((flags & DMAR_PGF_OBJL) != 0, ("lost PGF_OBJL"));
365 
366 	idx = domain_pgtbl_get_pindex(domain, base, lvl);
367 	if (*sf != NULL && idx == *idxp) {
368 		pte = (dmar_pte_t *)sf_buf_kva(*sf);
369 	} else {
370 		if (*sf != NULL)
371 			dmar_unmap_pgtbl(*sf);
372 		*idxp = idx;
373 retry:
374 		pte = dmar_map_pgtbl(domain->pgtbl_obj, idx, flags, sf);
375 		if (pte == NULL) {
376 			KASSERT(lvl > 0,
377 			    ("lost root page table page %p", domain));
378 			/*
379 			 * Page table page does not exist, allocate
380 			 * it and create a pte in the preceeding page level
381 			 * to reference the allocated page table page.
382 			 */
383 			m = dmar_pgalloc(domain->pgtbl_obj, idx, flags |
384 			    DMAR_PGF_ZERO);
385 			if (m == NULL)
386 				return (NULL);
387 
388 			/*
389 			 * Prevent potential free while pgtbl_obj is
390 			 * unlocked in the recursive call to
391 			 * domain_pgtbl_map_pte(), if other thread did
392 			 * pte write and clean while the lock is
393 			 * dropped.
394 			 */
395 			m->wire_count++;
396 
397 			sfp = NULL;
398 			ptep = domain_pgtbl_map_pte(domain, base, lvl - 1,
399 			    flags, &idx1, &sfp);
400 			if (ptep == NULL) {
401 				KASSERT(m->pindex != 0,
402 				    ("loosing root page %p", domain));
403 				m->wire_count--;
404 				dmar_pgfree(domain->pgtbl_obj, m->pindex,
405 				    flags);
406 				return (NULL);
407 			}
408 			dmar_pte_store(&ptep->pte, DMAR_PTE_R | DMAR_PTE_W |
409 			    VM_PAGE_TO_PHYS(m));
410 			dmar_flush_pte_to_ram(domain->dmar, ptep);
411 			sf_buf_page(sfp)->wire_count += 1;
412 			m->wire_count--;
413 			dmar_unmap_pgtbl(sfp);
414 			/* Only executed once. */
415 			goto retry;
416 		}
417 	}
418 	pte += domain_pgtbl_pte_off(domain, base, lvl);
419 	return (pte);
420 }
421 
422 static int
423 domain_map_buf_locked(struct dmar_domain *domain, dmar_gaddr_t base,
424     dmar_gaddr_t size, vm_page_t *ma, uint64_t pflags, int flags)
425 {
426 	dmar_pte_t *pte;
427 	struct sf_buf *sf;
428 	dmar_gaddr_t pg_sz, base1, size1;
429 	vm_pindex_t pi, c, idx, run_sz;
430 	int lvl;
431 	bool superpage;
432 
433 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
434 
435 	base1 = base;
436 	size1 = size;
437 	flags |= DMAR_PGF_OBJL;
438 	TD_PREP_PINNED_ASSERT;
439 
440 	for (sf = NULL, pi = 0; size > 0; base += pg_sz, size -= pg_sz,
441 	    pi += run_sz) {
442 		for (lvl = 0, c = 0, superpage = false;; lvl++) {
443 			pg_sz = domain_page_size(domain, lvl);
444 			run_sz = pg_sz >> DMAR_PAGE_SHIFT;
445 			if (lvl == domain->pglvl - 1)
446 				break;
447 			/*
448 			 * Check if the current base suitable for the
449 			 * superpage mapping.  First, verify the level.
450 			 */
451 			if (!domain_is_sp_lvl(domain, lvl))
452 				continue;
453 			/*
454 			 * Next, look at the size of the mapping and
455 			 * alignment of both guest and host addresses.
456 			 */
457 			if (size < pg_sz || (base & (pg_sz - 1)) != 0 ||
458 			    (VM_PAGE_TO_PHYS(ma[pi]) & (pg_sz - 1)) != 0)
459 				continue;
460 			/* All passed, check host pages contiguouty. */
461 			if (c == 0) {
462 				for (c = 1; c < run_sz; c++) {
463 					if (VM_PAGE_TO_PHYS(ma[pi + c]) !=
464 					    VM_PAGE_TO_PHYS(ma[pi + c - 1]) +
465 					    PAGE_SIZE)
466 						break;
467 				}
468 			}
469 			if (c >= run_sz) {
470 				superpage = true;
471 				break;
472 			}
473 		}
474 		KASSERT(size >= pg_sz,
475 		    ("mapping loop overflow %p %jx %jx %jx", domain,
476 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
477 		KASSERT(pg_sz > 0, ("pg_sz 0 lvl %d", lvl));
478 		pte = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
479 		if (pte == NULL) {
480 			KASSERT((flags & DMAR_PGF_WAITOK) == 0,
481 			    ("failed waitable pte alloc %p", domain));
482 			if (sf != NULL)
483 				dmar_unmap_pgtbl(sf);
484 			domain_unmap_buf_locked(domain, base1, base - base1,
485 			    flags);
486 			TD_PINNED_ASSERT;
487 			return (ENOMEM);
488 		}
489 		dmar_pte_store(&pte->pte, VM_PAGE_TO_PHYS(ma[pi]) | pflags |
490 		    (superpage ? DMAR_PTE_SP : 0));
491 		dmar_flush_pte_to_ram(domain->dmar, pte);
492 		sf_buf_page(sf)->wire_count += 1;
493 	}
494 	if (sf != NULL)
495 		dmar_unmap_pgtbl(sf);
496 	TD_PINNED_ASSERT;
497 	return (0);
498 }
499 
500 int
501 domain_map_buf(struct dmar_domain *domain, dmar_gaddr_t base, dmar_gaddr_t size,
502     vm_page_t *ma, uint64_t pflags, int flags)
503 {
504 	struct dmar_unit *unit;
505 	int error;
506 
507 	unit = domain->dmar;
508 
509 	KASSERT((domain->flags & DMAR_DOMAIN_IDMAP) == 0,
510 	    ("modifying idmap pagetable domain %p", domain));
511 	KASSERT((base & DMAR_PAGE_MASK) == 0,
512 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
513 	    (uintmax_t)size));
514 	KASSERT((size & DMAR_PAGE_MASK) == 0,
515 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
516 	    (uintmax_t)size));
517 	KASSERT(size > 0, ("zero size %p %jx %jx", domain, (uintmax_t)base,
518 	    (uintmax_t)size));
519 	KASSERT(base < (1ULL << domain->agaw),
520 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
521 	    (uintmax_t)size, domain->agaw));
522 	KASSERT(base + size < (1ULL << domain->agaw),
523 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
524 	    (uintmax_t)size, domain->agaw));
525 	KASSERT(base + size > base,
526 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
527 	    (uintmax_t)size));
528 	KASSERT((pflags & (DMAR_PTE_R | DMAR_PTE_W)) != 0,
529 	    ("neither read nor write %jx", (uintmax_t)pflags));
530 	KASSERT((pflags & ~(DMAR_PTE_R | DMAR_PTE_W | DMAR_PTE_SNP |
531 	    DMAR_PTE_TM)) == 0,
532 	    ("invalid pte flags %jx", (uintmax_t)pflags));
533 	KASSERT((pflags & DMAR_PTE_SNP) == 0 ||
534 	    (unit->hw_ecap & DMAR_ECAP_SC) != 0,
535 	    ("PTE_SNP for dmar without snoop control %p %jx",
536 	    domain, (uintmax_t)pflags));
537 	KASSERT((pflags & DMAR_PTE_TM) == 0 ||
538 	    (unit->hw_ecap & DMAR_ECAP_DI) != 0,
539 	    ("PTE_TM for dmar without DIOTLB %p %jx",
540 	    domain, (uintmax_t)pflags));
541 	KASSERT((flags & ~DMAR_PGF_WAITOK) == 0, ("invalid flags %x", flags));
542 
543 	DMAR_DOMAIN_PGLOCK(domain);
544 	error = domain_map_buf_locked(domain, base, size, ma, pflags, flags);
545 	DMAR_DOMAIN_PGUNLOCK(domain);
546 	if (error != 0)
547 		return (error);
548 
549 	if ((unit->hw_cap & DMAR_CAP_CM) != 0)
550 		domain_flush_iotlb_sync(domain, base, size);
551 	else if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
552 		/* See 11.1 Write Buffer Flushing. */
553 		DMAR_LOCK(unit);
554 		dmar_flush_write_bufs(unit);
555 		DMAR_UNLOCK(unit);
556 	}
557 	return (0);
558 }
559 
560 static void domain_unmap_clear_pte(struct dmar_domain *domain,
561     dmar_gaddr_t base, int lvl, int flags, dmar_pte_t *pte,
562     struct sf_buf **sf, bool free_fs);
563 
564 static void
565 domain_free_pgtbl_pde(struct dmar_domain *domain, dmar_gaddr_t base,
566     int lvl, int flags)
567 {
568 	struct sf_buf *sf;
569 	dmar_pte_t *pde;
570 	vm_pindex_t idx;
571 
572 	sf = NULL;
573 	pde = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
574 	domain_unmap_clear_pte(domain, base, lvl, flags, pde, &sf, true);
575 }
576 
577 static void
578 domain_unmap_clear_pte(struct dmar_domain *domain, dmar_gaddr_t base, int lvl,
579     int flags, dmar_pte_t *pte, struct sf_buf **sf, bool free_sf)
580 {
581 	vm_page_t m;
582 
583 	dmar_pte_clear(&pte->pte);
584 	dmar_flush_pte_to_ram(domain->dmar, pte);
585 	m = sf_buf_page(*sf);
586 	if (free_sf) {
587 		dmar_unmap_pgtbl(*sf);
588 		*sf = NULL;
589 	}
590 	m->wire_count--;
591 	if (m->wire_count != 0)
592 		return;
593 	KASSERT(lvl != 0,
594 	    ("lost reference (lvl) on root pg domain %p base %jx lvl %d",
595 	    domain, (uintmax_t)base, lvl));
596 	KASSERT(m->pindex != 0,
597 	    ("lost reference (idx) on root pg domain %p base %jx lvl %d",
598 	    domain, (uintmax_t)base, lvl));
599 	dmar_pgfree(domain->pgtbl_obj, m->pindex, flags);
600 	domain_free_pgtbl_pde(domain, base, lvl - 1, flags);
601 }
602 
603 /*
604  * Assumes that the unmap is never partial.
605  */
606 static int
607 domain_unmap_buf_locked(struct dmar_domain *domain, dmar_gaddr_t base,
608     dmar_gaddr_t size, int flags)
609 {
610 	dmar_pte_t *pte;
611 	struct sf_buf *sf;
612 	vm_pindex_t idx;
613 	dmar_gaddr_t pg_sz;
614 	int lvl;
615 
616 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
617 	if (size == 0)
618 		return (0);
619 
620 	KASSERT((domain->flags & DMAR_DOMAIN_IDMAP) == 0,
621 	    ("modifying idmap pagetable domain %p", domain));
622 	KASSERT((base & DMAR_PAGE_MASK) == 0,
623 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
624 	    (uintmax_t)size));
625 	KASSERT((size & DMAR_PAGE_MASK) == 0,
626 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
627 	    (uintmax_t)size));
628 	KASSERT(base < (1ULL << domain->agaw),
629 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
630 	    (uintmax_t)size, domain->agaw));
631 	KASSERT(base + size < (1ULL << domain->agaw),
632 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
633 	    (uintmax_t)size, domain->agaw));
634 	KASSERT(base + size > base,
635 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
636 	    (uintmax_t)size));
637 	KASSERT((flags & ~DMAR_PGF_WAITOK) == 0, ("invalid flags %x", flags));
638 
639 	pg_sz = 0; /* silence gcc */
640 	flags |= DMAR_PGF_OBJL;
641 	TD_PREP_PINNED_ASSERT;
642 
643 	for (sf = NULL; size > 0; base += pg_sz, size -= pg_sz) {
644 		for (lvl = 0; lvl < domain->pglvl; lvl++) {
645 			if (lvl != domain->pglvl - 1 &&
646 			    !domain_is_sp_lvl(domain, lvl))
647 				continue;
648 			pg_sz = domain_page_size(domain, lvl);
649 			if (pg_sz > size)
650 				continue;
651 			pte = domain_pgtbl_map_pte(domain, base, lvl, flags,
652 			    &idx, &sf);
653 			KASSERT(pte != NULL,
654 			    ("sleeping or page missed %p %jx %d 0x%x",
655 			    domain, (uintmax_t)base, lvl, flags));
656 			if ((pte->pte & DMAR_PTE_SP) != 0 ||
657 			    lvl == domain->pglvl - 1) {
658 				domain_unmap_clear_pte(domain, base, lvl,
659 				    flags, pte, &sf, false);
660 				break;
661 			}
662 		}
663 		KASSERT(size >= pg_sz,
664 		    ("unmapping loop overflow %p %jx %jx %jx", domain,
665 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
666 	}
667 	if (sf != NULL)
668 		dmar_unmap_pgtbl(sf);
669 	/*
670 	 * See 11.1 Write Buffer Flushing for an explanation why RWBF
671 	 * can be ignored there.
672 	 */
673 
674 	TD_PINNED_ASSERT;
675 	return (0);
676 }
677 
678 int
679 domain_unmap_buf(struct dmar_domain *domain, dmar_gaddr_t base,
680     dmar_gaddr_t size, int flags)
681 {
682 	int error;
683 
684 	DMAR_DOMAIN_PGLOCK(domain);
685 	error = domain_unmap_buf_locked(domain, base, size, flags);
686 	DMAR_DOMAIN_PGUNLOCK(domain);
687 	return (error);
688 }
689 
690 int
691 domain_alloc_pgtbl(struct dmar_domain *domain)
692 {
693 	vm_page_t m;
694 
695 	KASSERT(domain->pgtbl_obj == NULL,
696 	    ("already initialized %p", domain));
697 
698 	domain->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
699 	    IDX_TO_OFF(pglvl_max_pages(domain->pglvl)), 0, 0, NULL);
700 	DMAR_DOMAIN_PGLOCK(domain);
701 	m = dmar_pgalloc(domain->pgtbl_obj, 0, DMAR_PGF_WAITOK |
702 	    DMAR_PGF_ZERO | DMAR_PGF_OBJL);
703 	/* No implicit free of the top level page table page. */
704 	m->wire_count = 1;
705 	DMAR_DOMAIN_PGUNLOCK(domain);
706 	DMAR_LOCK(domain->dmar);
707 	domain->flags |= DMAR_DOMAIN_PGTBL_INITED;
708 	DMAR_UNLOCK(domain->dmar);
709 	return (0);
710 }
711 
712 void
713 domain_free_pgtbl(struct dmar_domain *domain)
714 {
715 	vm_object_t obj;
716 	vm_page_t m;
717 
718 	obj = domain->pgtbl_obj;
719 	if (obj == NULL) {
720 		KASSERT((domain->dmar->hw_ecap & DMAR_ECAP_PT) != 0 &&
721 		    (domain->flags & DMAR_DOMAIN_IDMAP) != 0,
722 		    ("lost pagetable object domain %p", domain));
723 		return;
724 	}
725 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
726 	domain->pgtbl_obj = NULL;
727 
728 	if ((domain->flags & DMAR_DOMAIN_IDMAP) != 0) {
729 		put_idmap_pgtbl(obj);
730 		domain->flags &= ~DMAR_DOMAIN_IDMAP;
731 		return;
732 	}
733 
734 	/* Obliterate wire_counts */
735 	VM_OBJECT_ASSERT_WLOCKED(obj);
736 	for (m = vm_page_lookup(obj, 0); m != NULL; m = vm_page_next(m))
737 		m->wire_count = 0;
738 	VM_OBJECT_WUNLOCK(obj);
739 	vm_object_deallocate(obj);
740 }
741 
742 static inline uint64_t
743 domain_wait_iotlb_flush(struct dmar_unit *unit, uint64_t wt, int iro)
744 {
745 	uint64_t iotlbr;
746 
747 	dmar_write8(unit, iro + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
748 	    DMAR_IOTLB_DR | DMAR_IOTLB_DW | wt);
749 	for (;;) {
750 		iotlbr = dmar_read8(unit, iro + DMAR_IOTLB_REG_OFF);
751 		if ((iotlbr & DMAR_IOTLB_IVT) == 0)
752 			break;
753 		cpu_spinwait();
754 	}
755 	return (iotlbr);
756 }
757 
758 void
759 domain_flush_iotlb_sync(struct dmar_domain *domain, dmar_gaddr_t base,
760     dmar_gaddr_t size)
761 {
762 	struct dmar_unit *unit;
763 	dmar_gaddr_t isize;
764 	uint64_t iotlbr;
765 	int am, iro;
766 
767 	unit = domain->dmar;
768 	KASSERT(!unit->qi_enabled, ("dmar%d: sync iotlb flush call",
769 	    unit->unit));
770 	iro = DMAR_ECAP_IRO(unit->hw_ecap) * 16;
771 	DMAR_LOCK(unit);
772 	if ((unit->hw_cap & DMAR_CAP_PSI) == 0 || size > 2 * 1024 * 1024) {
773 		iotlbr = domain_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_DOM |
774 		    DMAR_IOTLB_DID(domain->domain), iro);
775 		KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
776 		    DMAR_IOTLB_IAIG_INVLD,
777 		    ("dmar%d: invalidation failed %jx", unit->unit,
778 		    (uintmax_t)iotlbr));
779 	} else {
780 		for (; size > 0; base += isize, size -= isize) {
781 			am = calc_am(unit, base, size, &isize);
782 			dmar_write8(unit, iro, base | am);
783 			iotlbr = domain_wait_iotlb_flush(unit,
784 			    DMAR_IOTLB_IIRG_PAGE |
785 			    DMAR_IOTLB_DID(domain->domain), iro);
786 			KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
787 			    DMAR_IOTLB_IAIG_INVLD,
788 			    ("dmar%d: PSI invalidation failed "
789 			    "iotlbr 0x%jx base 0x%jx size 0x%jx am %d",
790 			    unit->unit, (uintmax_t)iotlbr,
791 			    (uintmax_t)base, (uintmax_t)size, am));
792 			/*
793 			 * Any non-page granularity covers whole guest
794 			 * address space for the domain.
795 			 */
796 			if ((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
797 			    DMAR_IOTLB_IAIG_PAGE)
798 				break;
799 		}
800 	}
801 	DMAR_UNLOCK(unit);
802 }
803