xref: /freebsd/sys/x86/iommu/intel_idpgtbl.c (revision 86c9d9918f1db7cdd968b60f8902466887bcd9e9)
1 /*-
2  * Copyright (c) 2013 The FreeBSD Foundation
3  * All rights reserved.
4  *
5  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
6  * under sponsorship from the FreeBSD Foundation.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/malloc.h>
36 #include <sys/bus.h>
37 #include <sys/interrupt.h>
38 #include <sys/kernel.h>
39 #include <sys/ktr.h>
40 #include <sys/lock.h>
41 #include <sys/memdesc.h>
42 #include <sys/mutex.h>
43 #include <sys/proc.h>
44 #include <sys/rwlock.h>
45 #include <sys/rman.h>
46 #include <sys/sf_buf.h>
47 #include <sys/sysctl.h>
48 #include <sys/taskqueue.h>
49 #include <sys/tree.h>
50 #include <sys/uio.h>
51 #include <sys/vmem.h>
52 #include <vm/vm.h>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_kern.h>
55 #include <vm/vm_object.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_pager.h>
58 #include <vm/vm_map.h>
59 #include <machine/atomic.h>
60 #include <machine/bus.h>
61 #include <machine/cpu.h>
62 #include <machine/md_var.h>
63 #include <machine/specialreg.h>
64 #include <x86/include/busdma_impl.h>
65 #include <x86/iommu/intel_reg.h>
66 #include <x86/iommu/busdma_dmar.h>
67 #include <x86/iommu/intel_dmar.h>
68 
69 static int domain_unmap_buf_locked(struct dmar_domain *domain,
70     dmar_gaddr_t base, dmar_gaddr_t size, int flags);
71 
72 /*
73  * The cache of the identity mapping page tables for the DMARs.  Using
74  * the cache saves significant amount of memory for page tables by
75  * reusing the page tables, since usually DMARs are identical and have
76  * the same capabilities.  Still, cache records the information needed
77  * to match DMAR capabilities and page table format, to correctly
78  * handle different DMARs.
79  */
80 
81 struct idpgtbl {
82 	dmar_gaddr_t maxaddr;	/* Page table covers the guest address
83 				   range [0..maxaddr) */
84 	int pglvl;		/* Total page table levels ignoring
85 				   superpages */
86 	int leaf;		/* The last materialized page table
87 				   level, it is non-zero if superpages
88 				   are supported */
89 	vm_object_t pgtbl_obj;	/* The page table pages */
90 	LIST_ENTRY(idpgtbl) link;
91 };
92 
93 static struct sx idpgtbl_lock;
94 SX_SYSINIT(idpgtbl, &idpgtbl_lock, "idpgtbl");
95 static LIST_HEAD(, idpgtbl) idpgtbls = LIST_HEAD_INITIALIZER(idpgtbls);
96 static MALLOC_DEFINE(M_DMAR_IDPGTBL, "dmar_idpgtbl",
97     "Intel DMAR Identity mappings cache elements");
98 
99 /*
100  * Build the next level of the page tables for the identity mapping.
101  * - lvl is the level to build;
102  * - idx is the index of the page table page in the pgtbl_obj, which is
103  *   being allocated filled now;
104  * - addr is the starting address in the bus address space which is
105  *   mapped by the page table page.
106  */
107 static void
108 domain_idmap_nextlvl(struct idpgtbl *tbl, int lvl, vm_pindex_t idx,
109     dmar_gaddr_t addr)
110 {
111 	vm_page_t m1;
112 	dmar_pte_t *pte;
113 	struct sf_buf *sf;
114 	dmar_gaddr_t f, pg_sz;
115 	vm_pindex_t base;
116 	int i;
117 
118 	VM_OBJECT_ASSERT_LOCKED(tbl->pgtbl_obj);
119 	if (addr >= tbl->maxaddr)
120 		return;
121 	(void)dmar_pgalloc(tbl->pgtbl_obj, idx, DMAR_PGF_OBJL | DMAR_PGF_WAITOK |
122 	    DMAR_PGF_ZERO);
123 	base = idx * DMAR_NPTEPG + 1; /* Index of the first child page of idx */
124 	pg_sz = pglvl_page_size(tbl->pglvl, lvl);
125 	if (lvl != tbl->leaf) {
126 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz)
127 			domain_idmap_nextlvl(tbl, lvl + 1, base + i, f);
128 	}
129 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
130 	pte = dmar_map_pgtbl(tbl->pgtbl_obj, idx, DMAR_PGF_WAITOK, &sf);
131 	if (lvl == tbl->leaf) {
132 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) {
133 			if (f >= tbl->maxaddr)
134 				break;
135 			pte[i].pte = (DMAR_PTE_ADDR_MASK & f) |
136 			    DMAR_PTE_R | DMAR_PTE_W;
137 		}
138 	} else {
139 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) {
140 			if (f >= tbl->maxaddr)
141 				break;
142 			m1 = dmar_pgalloc(tbl->pgtbl_obj, base + i,
143 			    DMAR_PGF_NOALLOC);
144 			KASSERT(m1 != NULL, ("lost page table page"));
145 			pte[i].pte = (DMAR_PTE_ADDR_MASK &
146 			    VM_PAGE_TO_PHYS(m1)) | DMAR_PTE_R | DMAR_PTE_W;
147 		}
148 	}
149 	/* domain_get_idmap_pgtbl flushes CPU cache if needed. */
150 	dmar_unmap_pgtbl(sf);
151 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
152 }
153 
154 /*
155  * Find a ready and compatible identity-mapping page table in the
156  * cache. If not found, populate the identity-mapping page table for
157  * the context, up to the maxaddr. The maxaddr byte is allowed to be
158  * not mapped, which is aligned with the definition of Maxmem as the
159  * highest usable physical address + 1.  If superpages are used, the
160  * maxaddr is typically mapped.
161  */
162 vm_object_t
163 domain_get_idmap_pgtbl(struct dmar_domain *domain, dmar_gaddr_t maxaddr)
164 {
165 	struct dmar_unit *unit;
166 	struct idpgtbl *tbl;
167 	vm_object_t res;
168 	vm_page_t m;
169 	int leaf, i;
170 
171 	leaf = 0; /* silence gcc */
172 
173 	/*
174 	 * First, determine where to stop the paging structures.
175 	 */
176 	for (i = 0; i < domain->pglvl; i++) {
177 		if (i == domain->pglvl - 1 || domain_is_sp_lvl(domain, i)) {
178 			leaf = i;
179 			break;
180 		}
181 	}
182 
183 	/*
184 	 * Search the cache for a compatible page table.  Qualified
185 	 * page table must map up to maxaddr, its level must be
186 	 * supported by the DMAR and leaf should be equal to the
187 	 * calculated value.  The later restriction could be lifted
188 	 * but I believe it is currently impossible to have any
189 	 * deviations for existing hardware.
190 	 */
191 	sx_slock(&idpgtbl_lock);
192 	LIST_FOREACH(tbl, &idpgtbls, link) {
193 		if (tbl->maxaddr >= maxaddr &&
194 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
195 		    tbl->leaf == leaf) {
196 			res = tbl->pgtbl_obj;
197 			vm_object_reference(res);
198 			sx_sunlock(&idpgtbl_lock);
199 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
200 			goto end;
201 		}
202 	}
203 
204 	/*
205 	 * Not found in cache, relock the cache into exclusive mode to
206 	 * be able to add element, and recheck cache again after the
207 	 * relock.
208 	 */
209 	sx_sunlock(&idpgtbl_lock);
210 	sx_xlock(&idpgtbl_lock);
211 	LIST_FOREACH(tbl, &idpgtbls, link) {
212 		if (tbl->maxaddr >= maxaddr &&
213 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
214 		    tbl->leaf == leaf) {
215 			res = tbl->pgtbl_obj;
216 			vm_object_reference(res);
217 			sx_xunlock(&idpgtbl_lock);
218 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
219 			return (res);
220 		}
221 	}
222 
223 	/*
224 	 * Still not found, create new page table.
225 	 */
226 	tbl = malloc(sizeof(*tbl), M_DMAR_IDPGTBL, M_WAITOK);
227 	tbl->pglvl = domain->pglvl;
228 	tbl->leaf = leaf;
229 	tbl->maxaddr = maxaddr;
230 	tbl->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
231 	    IDX_TO_OFF(pglvl_max_pages(tbl->pglvl)), 0, 0, NULL);
232 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
233 	domain_idmap_nextlvl(tbl, 0, 0, 0);
234 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
235 	LIST_INSERT_HEAD(&idpgtbls, tbl, link);
236 	res = tbl->pgtbl_obj;
237 	vm_object_reference(res);
238 	sx_xunlock(&idpgtbl_lock);
239 
240 end:
241 	/*
242 	 * Table was found or created.
243 	 *
244 	 * If DMAR does not snoop paging structures accesses, flush
245 	 * CPU cache to memory.  Note that dmar_unmap_pgtbl() coherent
246 	 * argument was possibly invalid at the time of the identity
247 	 * page table creation, since DMAR which was passed at the
248 	 * time of creation could be coherent, while current DMAR is
249 	 * not.
250 	 *
251 	 * If DMAR cannot look into the chipset write buffer, flush it
252 	 * as well.
253 	 */
254 	unit = domain->dmar;
255 	if (!DMAR_IS_COHERENT(unit)) {
256 		VM_OBJECT_WLOCK(res);
257 		for (m = vm_page_lookup(res, 0); m != NULL;
258 		     m = vm_page_next(m))
259 			pmap_invalidate_cache_pages(&m, 1);
260 		VM_OBJECT_WUNLOCK(res);
261 	}
262 	if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
263 		DMAR_LOCK(unit);
264 		dmar_flush_write_bufs(unit);
265 		DMAR_UNLOCK(unit);
266 	}
267 
268 	return (res);
269 }
270 
271 /*
272  * Return a reference to the identity mapping page table to the cache.
273  */
274 void
275 put_idmap_pgtbl(vm_object_t obj)
276 {
277 	struct idpgtbl *tbl, *tbl1;
278 	vm_object_t rmobj;
279 
280 	sx_slock(&idpgtbl_lock);
281 	KASSERT(obj->ref_count >= 2, ("lost cache reference"));
282 	vm_object_deallocate(obj);
283 
284 	/*
285 	 * Cache always owns one last reference on the page table object.
286 	 * If there is an additional reference, object must stay.
287 	 */
288 	if (obj->ref_count > 1) {
289 		sx_sunlock(&idpgtbl_lock);
290 		return;
291 	}
292 
293 	/*
294 	 * Cache reference is the last, remove cache element and free
295 	 * page table object, returning the page table pages to the
296 	 * system.
297 	 */
298 	sx_sunlock(&idpgtbl_lock);
299 	sx_xlock(&idpgtbl_lock);
300 	LIST_FOREACH_SAFE(tbl, &idpgtbls, link, tbl1) {
301 		rmobj = tbl->pgtbl_obj;
302 		if (rmobj->ref_count == 1) {
303 			LIST_REMOVE(tbl, link);
304 			atomic_subtract_int(&dmar_tbl_pagecnt,
305 			    rmobj->resident_page_count);
306 			vm_object_deallocate(rmobj);
307 			free(tbl, M_DMAR_IDPGTBL);
308 		}
309 	}
310 	sx_xunlock(&idpgtbl_lock);
311 }
312 
313 /*
314  * The core routines to map and unmap host pages at the given guest
315  * address.  Support superpages.
316  */
317 
318 /*
319  * Index of the pte for the guest address base in the page table at
320  * the level lvl.
321  */
322 static int
323 domain_pgtbl_pte_off(struct dmar_domain *domain, dmar_gaddr_t base, int lvl)
324 {
325 
326 	base >>= DMAR_PAGE_SHIFT + (domain->pglvl - lvl - 1) *
327 	    DMAR_NPTEPGSHIFT;
328 	return (base & DMAR_PTEMASK);
329 }
330 
331 /*
332  * Returns the page index of the page table page in the page table
333  * object, which maps the given address base at the page table level
334  * lvl.
335  */
336 static vm_pindex_t
337 domain_pgtbl_get_pindex(struct dmar_domain *domain, dmar_gaddr_t base, int lvl)
338 {
339 	vm_pindex_t idx, pidx;
340 	int i;
341 
342 	KASSERT(lvl >= 0 && lvl < domain->pglvl,
343 	    ("wrong lvl %p %d", domain, lvl));
344 
345 	for (pidx = idx = 0, i = 0; i < lvl; i++, pidx = idx) {
346 		idx = domain_pgtbl_pte_off(domain, base, i) +
347 		    pidx * DMAR_NPTEPG + 1;
348 	}
349 	return (idx);
350 }
351 
352 static dmar_pte_t *
353 domain_pgtbl_map_pte(struct dmar_domain *domain, dmar_gaddr_t base, int lvl,
354     int flags, vm_pindex_t *idxp, struct sf_buf **sf)
355 {
356 	vm_page_t m;
357 	struct sf_buf *sfp;
358 	dmar_pte_t *pte, *ptep;
359 	vm_pindex_t idx, idx1;
360 
361 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
362 	KASSERT((flags & DMAR_PGF_OBJL) != 0, ("lost PGF_OBJL"));
363 
364 	idx = domain_pgtbl_get_pindex(domain, base, lvl);
365 	if (*sf != NULL && idx == *idxp) {
366 		pte = (dmar_pte_t *)sf_buf_kva(*sf);
367 	} else {
368 		if (*sf != NULL)
369 			dmar_unmap_pgtbl(*sf);
370 		*idxp = idx;
371 retry:
372 		pte = dmar_map_pgtbl(domain->pgtbl_obj, idx, flags, sf);
373 		if (pte == NULL) {
374 			KASSERT(lvl > 0,
375 			    ("lost root page table page %p", domain));
376 			/*
377 			 * Page table page does not exist, allocate
378 			 * it and create a pte in the preceeding page level
379 			 * to reference the allocated page table page.
380 			 */
381 			m = dmar_pgalloc(domain->pgtbl_obj, idx, flags |
382 			    DMAR_PGF_ZERO);
383 			if (m == NULL)
384 				return (NULL);
385 
386 			/*
387 			 * Prevent potential free while pgtbl_obj is
388 			 * unlocked in the recursive call to
389 			 * domain_pgtbl_map_pte(), if other thread did
390 			 * pte write and clean while the lock is
391 			 * dropped.
392 			 */
393 			m->wire_count++;
394 
395 			sfp = NULL;
396 			ptep = domain_pgtbl_map_pte(domain, base, lvl - 1,
397 			    flags, &idx1, &sfp);
398 			if (ptep == NULL) {
399 				KASSERT(m->pindex != 0,
400 				    ("loosing root page %p", domain));
401 				m->wire_count--;
402 				dmar_pgfree(domain->pgtbl_obj, m->pindex,
403 				    flags);
404 				return (NULL);
405 			}
406 			dmar_pte_store(&ptep->pte, DMAR_PTE_R | DMAR_PTE_W |
407 			    VM_PAGE_TO_PHYS(m));
408 			dmar_flush_pte_to_ram(domain->dmar, ptep);
409 			sf_buf_page(sfp)->wire_count += 1;
410 			m->wire_count--;
411 			dmar_unmap_pgtbl(sfp);
412 			/* Only executed once. */
413 			goto retry;
414 		}
415 	}
416 	pte += domain_pgtbl_pte_off(domain, base, lvl);
417 	return (pte);
418 }
419 
420 static int
421 domain_map_buf_locked(struct dmar_domain *domain, dmar_gaddr_t base,
422     dmar_gaddr_t size, vm_page_t *ma, uint64_t pflags, int flags)
423 {
424 	dmar_pte_t *pte;
425 	struct sf_buf *sf;
426 	dmar_gaddr_t pg_sz, base1, size1;
427 	vm_pindex_t pi, c, idx, run_sz;
428 	int lvl;
429 	bool superpage;
430 
431 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
432 
433 	base1 = base;
434 	size1 = size;
435 	flags |= DMAR_PGF_OBJL;
436 	TD_PREP_PINNED_ASSERT;
437 
438 	for (sf = NULL, pi = 0; size > 0; base += pg_sz, size -= pg_sz,
439 	    pi += run_sz) {
440 		for (lvl = 0, c = 0, superpage = false;; lvl++) {
441 			pg_sz = domain_page_size(domain, lvl);
442 			run_sz = pg_sz >> DMAR_PAGE_SHIFT;
443 			if (lvl == domain->pglvl - 1)
444 				break;
445 			/*
446 			 * Check if the current base suitable for the
447 			 * superpage mapping.  First, verify the level.
448 			 */
449 			if (!domain_is_sp_lvl(domain, lvl))
450 				continue;
451 			/*
452 			 * Next, look at the size of the mapping and
453 			 * alignment of both guest and host addresses.
454 			 */
455 			if (size < pg_sz || (base & (pg_sz - 1)) != 0 ||
456 			    (VM_PAGE_TO_PHYS(ma[pi]) & (pg_sz - 1)) != 0)
457 				continue;
458 			/* All passed, check host pages contiguouty. */
459 			if (c == 0) {
460 				for (c = 1; c < run_sz; c++) {
461 					if (VM_PAGE_TO_PHYS(ma[pi + c]) !=
462 					    VM_PAGE_TO_PHYS(ma[pi + c - 1]) +
463 					    PAGE_SIZE)
464 						break;
465 				}
466 			}
467 			if (c >= run_sz) {
468 				superpage = true;
469 				break;
470 			}
471 		}
472 		KASSERT(size >= pg_sz,
473 		    ("mapping loop overflow %p %jx %jx %jx", domain,
474 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
475 		KASSERT(pg_sz > 0, ("pg_sz 0 lvl %d", lvl));
476 		pte = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
477 		if (pte == NULL) {
478 			KASSERT((flags & DMAR_PGF_WAITOK) == 0,
479 			    ("failed waitable pte alloc %p", domain));
480 			if (sf != NULL)
481 				dmar_unmap_pgtbl(sf);
482 			domain_unmap_buf_locked(domain, base1, base - base1,
483 			    flags);
484 			TD_PINNED_ASSERT;
485 			return (ENOMEM);
486 		}
487 		dmar_pte_store(&pte->pte, VM_PAGE_TO_PHYS(ma[pi]) | pflags |
488 		    (superpage ? DMAR_PTE_SP : 0));
489 		dmar_flush_pte_to_ram(domain->dmar, pte);
490 		sf_buf_page(sf)->wire_count += 1;
491 	}
492 	if (sf != NULL)
493 		dmar_unmap_pgtbl(sf);
494 	TD_PINNED_ASSERT;
495 	return (0);
496 }
497 
498 int
499 domain_map_buf(struct dmar_domain *domain, dmar_gaddr_t base, dmar_gaddr_t size,
500     vm_page_t *ma, uint64_t pflags, int flags)
501 {
502 	struct dmar_unit *unit;
503 	int error;
504 
505 	unit = domain->dmar;
506 
507 	KASSERT((domain->flags & DMAR_DOMAIN_IDMAP) == 0,
508 	    ("modifying idmap pagetable domain %p", domain));
509 	KASSERT((base & DMAR_PAGE_MASK) == 0,
510 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
511 	    (uintmax_t)size));
512 	KASSERT((size & DMAR_PAGE_MASK) == 0,
513 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
514 	    (uintmax_t)size));
515 	KASSERT(size > 0, ("zero size %p %jx %jx", domain, (uintmax_t)base,
516 	    (uintmax_t)size));
517 	KASSERT(base < (1ULL << domain->agaw),
518 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
519 	    (uintmax_t)size, domain->agaw));
520 	KASSERT(base + size < (1ULL << domain->agaw),
521 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
522 	    (uintmax_t)size, domain->agaw));
523 	KASSERT(base + size > base,
524 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
525 	    (uintmax_t)size));
526 	KASSERT((pflags & (DMAR_PTE_R | DMAR_PTE_W)) != 0,
527 	    ("neither read nor write %jx", (uintmax_t)pflags));
528 	KASSERT((pflags & ~(DMAR_PTE_R | DMAR_PTE_W | DMAR_PTE_SNP |
529 	    DMAR_PTE_TM)) == 0,
530 	    ("invalid pte flags %jx", (uintmax_t)pflags));
531 	KASSERT((pflags & DMAR_PTE_SNP) == 0 ||
532 	    (unit->hw_ecap & DMAR_ECAP_SC) != 0,
533 	    ("PTE_SNP for dmar without snoop control %p %jx",
534 	    domain, (uintmax_t)pflags));
535 	KASSERT((pflags & DMAR_PTE_TM) == 0 ||
536 	    (unit->hw_ecap & DMAR_ECAP_DI) != 0,
537 	    ("PTE_TM for dmar without DIOTLB %p %jx",
538 	    domain, (uintmax_t)pflags));
539 	KASSERT((flags & ~DMAR_PGF_WAITOK) == 0, ("invalid flags %x", flags));
540 
541 	DMAR_DOMAIN_PGLOCK(domain);
542 	error = domain_map_buf_locked(domain, base, size, ma, pflags, flags);
543 	DMAR_DOMAIN_PGUNLOCK(domain);
544 	if (error != 0)
545 		return (error);
546 
547 	if ((unit->hw_cap & DMAR_CAP_CM) != 0)
548 		domain_flush_iotlb_sync(domain, base, size);
549 	else if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
550 		/* See 11.1 Write Buffer Flushing. */
551 		DMAR_LOCK(unit);
552 		dmar_flush_write_bufs(unit);
553 		DMAR_UNLOCK(unit);
554 	}
555 	return (0);
556 }
557 
558 static void domain_unmap_clear_pte(struct dmar_domain *domain,
559     dmar_gaddr_t base, int lvl, int flags, dmar_pte_t *pte,
560     struct sf_buf **sf, bool free_fs);
561 
562 static void
563 domain_free_pgtbl_pde(struct dmar_domain *domain, dmar_gaddr_t base,
564     int lvl, int flags)
565 {
566 	struct sf_buf *sf;
567 	dmar_pte_t *pde;
568 	vm_pindex_t idx;
569 
570 	sf = NULL;
571 	pde = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
572 	domain_unmap_clear_pte(domain, base, lvl, flags, pde, &sf, true);
573 }
574 
575 static void
576 domain_unmap_clear_pte(struct dmar_domain *domain, dmar_gaddr_t base, int lvl,
577     int flags, dmar_pte_t *pte, struct sf_buf **sf, bool free_sf)
578 {
579 	vm_page_t m;
580 
581 	dmar_pte_clear(&pte->pte);
582 	dmar_flush_pte_to_ram(domain->dmar, pte);
583 	m = sf_buf_page(*sf);
584 	if (free_sf) {
585 		dmar_unmap_pgtbl(*sf);
586 		*sf = NULL;
587 	}
588 	m->wire_count--;
589 	if (m->wire_count != 0)
590 		return;
591 	KASSERT(lvl != 0,
592 	    ("lost reference (lvl) on root pg domain %p base %jx lvl %d",
593 	    domain, (uintmax_t)base, lvl));
594 	KASSERT(m->pindex != 0,
595 	    ("lost reference (idx) on root pg domain %p base %jx lvl %d",
596 	    domain, (uintmax_t)base, lvl));
597 	dmar_pgfree(domain->pgtbl_obj, m->pindex, flags);
598 	domain_free_pgtbl_pde(domain, base, lvl - 1, flags);
599 }
600 
601 /*
602  * Assumes that the unmap is never partial.
603  */
604 static int
605 domain_unmap_buf_locked(struct dmar_domain *domain, dmar_gaddr_t base,
606     dmar_gaddr_t size, int flags)
607 {
608 	dmar_pte_t *pte;
609 	struct sf_buf *sf;
610 	vm_pindex_t idx;
611 	dmar_gaddr_t pg_sz;
612 	int lvl;
613 
614 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
615 	if (size == 0)
616 		return (0);
617 
618 	KASSERT((domain->flags & DMAR_DOMAIN_IDMAP) == 0,
619 	    ("modifying idmap pagetable domain %p", domain));
620 	KASSERT((base & DMAR_PAGE_MASK) == 0,
621 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
622 	    (uintmax_t)size));
623 	KASSERT((size & DMAR_PAGE_MASK) == 0,
624 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
625 	    (uintmax_t)size));
626 	KASSERT(base < (1ULL << domain->agaw),
627 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
628 	    (uintmax_t)size, domain->agaw));
629 	KASSERT(base + size < (1ULL << domain->agaw),
630 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
631 	    (uintmax_t)size, domain->agaw));
632 	KASSERT(base + size > base,
633 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
634 	    (uintmax_t)size));
635 	KASSERT((flags & ~DMAR_PGF_WAITOK) == 0, ("invalid flags %x", flags));
636 
637 	pg_sz = 0; /* silence gcc */
638 	flags |= DMAR_PGF_OBJL;
639 	TD_PREP_PINNED_ASSERT;
640 
641 	for (sf = NULL; size > 0; base += pg_sz, size -= pg_sz) {
642 		for (lvl = 0; lvl < domain->pglvl; lvl++) {
643 			if (lvl != domain->pglvl - 1 &&
644 			    !domain_is_sp_lvl(domain, lvl))
645 				continue;
646 			pg_sz = domain_page_size(domain, lvl);
647 			if (pg_sz > size)
648 				continue;
649 			pte = domain_pgtbl_map_pte(domain, base, lvl, flags,
650 			    &idx, &sf);
651 			KASSERT(pte != NULL,
652 			    ("sleeping or page missed %p %jx %d 0x%x",
653 			    domain, (uintmax_t)base, lvl, flags));
654 			if ((pte->pte & DMAR_PTE_SP) != 0 ||
655 			    lvl == domain->pglvl - 1) {
656 				domain_unmap_clear_pte(domain, base, lvl,
657 				    flags, pte, &sf, false);
658 				break;
659 			}
660 		}
661 		KASSERT(size >= pg_sz,
662 		    ("unmapping loop overflow %p %jx %jx %jx", domain,
663 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
664 	}
665 	if (sf != NULL)
666 		dmar_unmap_pgtbl(sf);
667 	/*
668 	 * See 11.1 Write Buffer Flushing for an explanation why RWBF
669 	 * can be ignored there.
670 	 */
671 
672 	TD_PINNED_ASSERT;
673 	return (0);
674 }
675 
676 int
677 domain_unmap_buf(struct dmar_domain *domain, dmar_gaddr_t base,
678     dmar_gaddr_t size, int flags)
679 {
680 	int error;
681 
682 	DMAR_DOMAIN_PGLOCK(domain);
683 	error = domain_unmap_buf_locked(domain, base, size, flags);
684 	DMAR_DOMAIN_PGUNLOCK(domain);
685 	return (error);
686 }
687 
688 int
689 domain_alloc_pgtbl(struct dmar_domain *domain)
690 {
691 	vm_page_t m;
692 
693 	KASSERT(domain->pgtbl_obj == NULL,
694 	    ("already initialized %p", domain));
695 
696 	domain->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
697 	    IDX_TO_OFF(pglvl_max_pages(domain->pglvl)), 0, 0, NULL);
698 	DMAR_DOMAIN_PGLOCK(domain);
699 	m = dmar_pgalloc(domain->pgtbl_obj, 0, DMAR_PGF_WAITOK |
700 	    DMAR_PGF_ZERO | DMAR_PGF_OBJL);
701 	/* No implicit free of the top level page table page. */
702 	m->wire_count = 1;
703 	DMAR_DOMAIN_PGUNLOCK(domain);
704 	DMAR_LOCK(domain->dmar);
705 	domain->flags |= DMAR_DOMAIN_PGTBL_INITED;
706 	DMAR_UNLOCK(domain->dmar);
707 	return (0);
708 }
709 
710 void
711 domain_free_pgtbl(struct dmar_domain *domain)
712 {
713 	vm_object_t obj;
714 	vm_page_t m;
715 
716 	obj = domain->pgtbl_obj;
717 	if (obj == NULL) {
718 		KASSERT((domain->dmar->hw_ecap & DMAR_ECAP_PT) != 0 &&
719 		    (domain->flags & DMAR_DOMAIN_IDMAP) != 0,
720 		    ("lost pagetable object domain %p", domain));
721 		return;
722 	}
723 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
724 	domain->pgtbl_obj = NULL;
725 
726 	if ((domain->flags & DMAR_DOMAIN_IDMAP) != 0) {
727 		put_idmap_pgtbl(obj);
728 		domain->flags &= ~DMAR_DOMAIN_IDMAP;
729 		return;
730 	}
731 
732 	/* Obliterate wire_counts */
733 	VM_OBJECT_ASSERT_WLOCKED(obj);
734 	for (m = vm_page_lookup(obj, 0); m != NULL; m = vm_page_next(m))
735 		m->wire_count = 0;
736 	VM_OBJECT_WUNLOCK(obj);
737 	vm_object_deallocate(obj);
738 }
739 
740 static inline uint64_t
741 domain_wait_iotlb_flush(struct dmar_unit *unit, uint64_t wt, int iro)
742 {
743 	uint64_t iotlbr;
744 
745 	dmar_write8(unit, iro + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
746 	    DMAR_IOTLB_DR | DMAR_IOTLB_DW | wt);
747 	for (;;) {
748 		iotlbr = dmar_read8(unit, iro + DMAR_IOTLB_REG_OFF);
749 		if ((iotlbr & DMAR_IOTLB_IVT) == 0)
750 			break;
751 		cpu_spinwait();
752 	}
753 	return (iotlbr);
754 }
755 
756 void
757 domain_flush_iotlb_sync(struct dmar_domain *domain, dmar_gaddr_t base,
758     dmar_gaddr_t size)
759 {
760 	struct dmar_unit *unit;
761 	dmar_gaddr_t isize;
762 	uint64_t iotlbr;
763 	int am, iro;
764 
765 	unit = domain->dmar;
766 	KASSERT(!unit->qi_enabled, ("dmar%d: sync iotlb flush call",
767 	    unit->unit));
768 	iro = DMAR_ECAP_IRO(unit->hw_ecap) * 16;
769 	DMAR_LOCK(unit);
770 	if ((unit->hw_cap & DMAR_CAP_PSI) == 0 || size > 2 * 1024 * 1024) {
771 		iotlbr = domain_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_DOM |
772 		    DMAR_IOTLB_DID(domain->domain), iro);
773 		KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
774 		    DMAR_IOTLB_IAIG_INVLD,
775 		    ("dmar%d: invalidation failed %jx", unit->unit,
776 		    (uintmax_t)iotlbr));
777 	} else {
778 		for (; size > 0; base += isize, size -= isize) {
779 			am = calc_am(unit, base, size, &isize);
780 			dmar_write8(unit, iro, base | am);
781 			iotlbr = domain_wait_iotlb_flush(unit,
782 			    DMAR_IOTLB_IIRG_PAGE |
783 			    DMAR_IOTLB_DID(domain->domain), iro);
784 			KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
785 			    DMAR_IOTLB_IAIG_INVLD,
786 			    ("dmar%d: PSI invalidation failed "
787 			    "iotlbr 0x%jx base 0x%jx size 0x%jx am %d",
788 			    unit->unit, (uintmax_t)iotlbr,
789 			    (uintmax_t)base, (uintmax_t)size, am));
790 			/*
791 			 * Any non-page granularity covers whole guest
792 			 * address space for the domain.
793 			 */
794 			if ((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
795 			    DMAR_IOTLB_IAIG_PAGE)
796 				break;
797 		}
798 	}
799 	DMAR_UNLOCK(unit);
800 }
801