xref: /freebsd/sys/x86/iommu/intel_idpgtbl.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2013 The FreeBSD Foundation
5  *
6  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
7  * under sponsorship from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/malloc.h>
34 #include <sys/bus.h>
35 #include <sys/interrupt.h>
36 #include <sys/kernel.h>
37 #include <sys/ktr.h>
38 #include <sys/lock.h>
39 #include <sys/memdesc.h>
40 #include <sys/mutex.h>
41 #include <sys/proc.h>
42 #include <sys/rwlock.h>
43 #include <sys/rman.h>
44 #include <sys/sf_buf.h>
45 #include <sys/sysctl.h>
46 #include <sys/taskqueue.h>
47 #include <sys/tree.h>
48 #include <sys/uio.h>
49 #include <sys/vmem.h>
50 #include <vm/vm.h>
51 #include <vm/vm_extern.h>
52 #include <vm/vm_kern.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vm_map.h>
57 #include <dev/pci/pcireg.h>
58 #include <machine/atomic.h>
59 #include <machine/bus.h>
60 #include <machine/cpu.h>
61 #include <machine/md_var.h>
62 #include <machine/specialreg.h>
63 #include <x86/include/busdma_impl.h>
64 #include <dev/iommu/busdma_iommu.h>
65 #include <x86/iommu/intel_reg.h>
66 #include <x86/iommu/intel_dmar.h>
67 
68 static int domain_unmap_buf_locked(struct dmar_domain *domain,
69     iommu_gaddr_t base, iommu_gaddr_t size, int flags);
70 
71 /*
72  * The cache of the identity mapping page tables for the DMARs.  Using
73  * the cache saves significant amount of memory for page tables by
74  * reusing the page tables, since usually DMARs are identical and have
75  * the same capabilities.  Still, cache records the information needed
76  * to match DMAR capabilities and page table format, to correctly
77  * handle different DMARs.
78  */
79 
80 struct idpgtbl {
81 	iommu_gaddr_t maxaddr;	/* Page table covers the guest address
82 				   range [0..maxaddr) */
83 	int pglvl;		/* Total page table levels ignoring
84 				   superpages */
85 	int leaf;		/* The last materialized page table
86 				   level, it is non-zero if superpages
87 				   are supported */
88 	vm_object_t pgtbl_obj;	/* The page table pages */
89 	LIST_ENTRY(idpgtbl) link;
90 };
91 
92 static struct sx idpgtbl_lock;
93 SX_SYSINIT(idpgtbl, &idpgtbl_lock, "idpgtbl");
94 static LIST_HEAD(, idpgtbl) idpgtbls = LIST_HEAD_INITIALIZER(idpgtbls);
95 static MALLOC_DEFINE(M_DMAR_IDPGTBL, "dmar_idpgtbl",
96     "Intel DMAR Identity mappings cache elements");
97 
98 /*
99  * Build the next level of the page tables for the identity mapping.
100  * - lvl is the level to build;
101  * - idx is the index of the page table page in the pgtbl_obj, which is
102  *   being allocated filled now;
103  * - addr is the starting address in the bus address space which is
104  *   mapped by the page table page.
105  */
106 static void
107 domain_idmap_nextlvl(struct idpgtbl *tbl, int lvl, vm_pindex_t idx,
108     iommu_gaddr_t addr)
109 {
110 	vm_page_t m1;
111 	dmar_pte_t *pte;
112 	struct sf_buf *sf;
113 	iommu_gaddr_t f, pg_sz;
114 	vm_pindex_t base;
115 	int i;
116 
117 	VM_OBJECT_ASSERT_LOCKED(tbl->pgtbl_obj);
118 	if (addr >= tbl->maxaddr)
119 		return;
120 	(void)dmar_pgalloc(tbl->pgtbl_obj, idx, IOMMU_PGF_OBJL |
121 	    IOMMU_PGF_WAITOK | IOMMU_PGF_ZERO);
122 	base = idx * DMAR_NPTEPG + 1; /* Index of the first child page of idx */
123 	pg_sz = pglvl_page_size(tbl->pglvl, lvl);
124 	if (lvl != tbl->leaf) {
125 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz)
126 			domain_idmap_nextlvl(tbl, lvl + 1, base + i, f);
127 	}
128 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
129 	pte = dmar_map_pgtbl(tbl->pgtbl_obj, idx, IOMMU_PGF_WAITOK, &sf);
130 	if (lvl == tbl->leaf) {
131 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) {
132 			if (f >= tbl->maxaddr)
133 				break;
134 			pte[i].pte = (DMAR_PTE_ADDR_MASK & f) |
135 			    DMAR_PTE_R | DMAR_PTE_W;
136 		}
137 	} else {
138 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) {
139 			if (f >= tbl->maxaddr)
140 				break;
141 			m1 = dmar_pgalloc(tbl->pgtbl_obj, base + i,
142 			    IOMMU_PGF_NOALLOC);
143 			KASSERT(m1 != NULL, ("lost page table page"));
144 			pte[i].pte = (DMAR_PTE_ADDR_MASK &
145 			    VM_PAGE_TO_PHYS(m1)) | DMAR_PTE_R | DMAR_PTE_W;
146 		}
147 	}
148 	/* domain_get_idmap_pgtbl flushes CPU cache if needed. */
149 	dmar_unmap_pgtbl(sf);
150 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
151 }
152 
153 /*
154  * Find a ready and compatible identity-mapping page table in the
155  * cache. If not found, populate the identity-mapping page table for
156  * the context, up to the maxaddr. The maxaddr byte is allowed to be
157  * not mapped, which is aligned with the definition of Maxmem as the
158  * highest usable physical address + 1.  If superpages are used, the
159  * maxaddr is typically mapped.
160  */
161 vm_object_t
162 domain_get_idmap_pgtbl(struct dmar_domain *domain, iommu_gaddr_t maxaddr)
163 {
164 	struct dmar_unit *unit;
165 	struct idpgtbl *tbl;
166 	vm_object_t res;
167 	vm_page_t m;
168 	int leaf, i;
169 
170 	leaf = 0; /* silence gcc */
171 
172 	/*
173 	 * First, determine where to stop the paging structures.
174 	 */
175 	for (i = 0; i < domain->pglvl; i++) {
176 		if (i == domain->pglvl - 1 || domain_is_sp_lvl(domain, i)) {
177 			leaf = i;
178 			break;
179 		}
180 	}
181 
182 	/*
183 	 * Search the cache for a compatible page table.  Qualified
184 	 * page table must map up to maxaddr, its level must be
185 	 * supported by the DMAR and leaf should be equal to the
186 	 * calculated value.  The later restriction could be lifted
187 	 * but I believe it is currently impossible to have any
188 	 * deviations for existing hardware.
189 	 */
190 	sx_slock(&idpgtbl_lock);
191 	LIST_FOREACH(tbl, &idpgtbls, link) {
192 		if (tbl->maxaddr >= maxaddr &&
193 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
194 		    tbl->leaf == leaf) {
195 			res = tbl->pgtbl_obj;
196 			vm_object_reference(res);
197 			sx_sunlock(&idpgtbl_lock);
198 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
199 			goto end;
200 		}
201 	}
202 
203 	/*
204 	 * Not found in cache, relock the cache into exclusive mode to
205 	 * be able to add element, and recheck cache again after the
206 	 * relock.
207 	 */
208 	sx_sunlock(&idpgtbl_lock);
209 	sx_xlock(&idpgtbl_lock);
210 	LIST_FOREACH(tbl, &idpgtbls, link) {
211 		if (tbl->maxaddr >= maxaddr &&
212 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
213 		    tbl->leaf == leaf) {
214 			res = tbl->pgtbl_obj;
215 			vm_object_reference(res);
216 			sx_xunlock(&idpgtbl_lock);
217 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
218 			return (res);
219 		}
220 	}
221 
222 	/*
223 	 * Still not found, create new page table.
224 	 */
225 	tbl = malloc(sizeof(*tbl), M_DMAR_IDPGTBL, M_WAITOK);
226 	tbl->pglvl = domain->pglvl;
227 	tbl->leaf = leaf;
228 	tbl->maxaddr = maxaddr;
229 	tbl->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
230 	    IDX_TO_OFF(pglvl_max_pages(tbl->pglvl)), 0, 0, NULL);
231 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
232 	domain_idmap_nextlvl(tbl, 0, 0, 0);
233 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
234 	LIST_INSERT_HEAD(&idpgtbls, tbl, link);
235 	res = tbl->pgtbl_obj;
236 	vm_object_reference(res);
237 	sx_xunlock(&idpgtbl_lock);
238 
239 end:
240 	/*
241 	 * Table was found or created.
242 	 *
243 	 * If DMAR does not snoop paging structures accesses, flush
244 	 * CPU cache to memory.  Note that dmar_unmap_pgtbl() coherent
245 	 * argument was possibly invalid at the time of the identity
246 	 * page table creation, since DMAR which was passed at the
247 	 * time of creation could be coherent, while current DMAR is
248 	 * not.
249 	 *
250 	 * If DMAR cannot look into the chipset write buffer, flush it
251 	 * as well.
252 	 */
253 	unit = domain->dmar;
254 	if (!DMAR_IS_COHERENT(unit)) {
255 		VM_OBJECT_WLOCK(res);
256 		for (m = vm_page_lookup(res, 0); m != NULL;
257 		     m = vm_page_next(m))
258 			pmap_invalidate_cache_pages(&m, 1);
259 		VM_OBJECT_WUNLOCK(res);
260 	}
261 	if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
262 		DMAR_LOCK(unit);
263 		dmar_flush_write_bufs(unit);
264 		DMAR_UNLOCK(unit);
265 	}
266 
267 	return (res);
268 }
269 
270 /*
271  * Return a reference to the identity mapping page table to the cache.
272  */
273 void
274 put_idmap_pgtbl(vm_object_t obj)
275 {
276 	struct idpgtbl *tbl, *tbl1;
277 	vm_object_t rmobj;
278 
279 	sx_slock(&idpgtbl_lock);
280 	KASSERT(obj->ref_count >= 2, ("lost cache reference"));
281 	vm_object_deallocate(obj);
282 
283 	/*
284 	 * Cache always owns one last reference on the page table object.
285 	 * If there is an additional reference, object must stay.
286 	 */
287 	if (obj->ref_count > 1) {
288 		sx_sunlock(&idpgtbl_lock);
289 		return;
290 	}
291 
292 	/*
293 	 * Cache reference is the last, remove cache element and free
294 	 * page table object, returning the page table pages to the
295 	 * system.
296 	 */
297 	sx_sunlock(&idpgtbl_lock);
298 	sx_xlock(&idpgtbl_lock);
299 	LIST_FOREACH_SAFE(tbl, &idpgtbls, link, tbl1) {
300 		rmobj = tbl->pgtbl_obj;
301 		if (rmobj->ref_count == 1) {
302 			LIST_REMOVE(tbl, link);
303 			atomic_subtract_int(&dmar_tbl_pagecnt,
304 			    rmobj->resident_page_count);
305 			vm_object_deallocate(rmobj);
306 			free(tbl, M_DMAR_IDPGTBL);
307 		}
308 	}
309 	sx_xunlock(&idpgtbl_lock);
310 }
311 
312 /*
313  * The core routines to map and unmap host pages at the given guest
314  * address.  Support superpages.
315  */
316 
317 /*
318  * Index of the pte for the guest address base in the page table at
319  * the level lvl.
320  */
321 static int
322 domain_pgtbl_pte_off(struct dmar_domain *domain, iommu_gaddr_t base, int lvl)
323 {
324 
325 	base >>= DMAR_PAGE_SHIFT + (domain->pglvl - lvl - 1) *
326 	    DMAR_NPTEPGSHIFT;
327 	return (base & DMAR_PTEMASK);
328 }
329 
330 /*
331  * Returns the page index of the page table page in the page table
332  * object, which maps the given address base at the page table level
333  * lvl.
334  */
335 static vm_pindex_t
336 domain_pgtbl_get_pindex(struct dmar_domain *domain, iommu_gaddr_t base, int lvl)
337 {
338 	vm_pindex_t idx, pidx;
339 	int i;
340 
341 	KASSERT(lvl >= 0 && lvl < domain->pglvl,
342 	    ("wrong lvl %p %d", domain, lvl));
343 
344 	for (pidx = idx = 0, i = 0; i < lvl; i++, pidx = idx) {
345 		idx = domain_pgtbl_pte_off(domain, base, i) +
346 		    pidx * DMAR_NPTEPG + 1;
347 	}
348 	return (idx);
349 }
350 
351 static dmar_pte_t *
352 domain_pgtbl_map_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl,
353     int flags, vm_pindex_t *idxp, struct sf_buf **sf)
354 {
355 	vm_page_t m;
356 	struct sf_buf *sfp;
357 	dmar_pte_t *pte, *ptep;
358 	vm_pindex_t idx, idx1;
359 
360 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
361 	KASSERT((flags & IOMMU_PGF_OBJL) != 0, ("lost PGF_OBJL"));
362 
363 	idx = domain_pgtbl_get_pindex(domain, base, lvl);
364 	if (*sf != NULL && idx == *idxp) {
365 		pte = (dmar_pte_t *)sf_buf_kva(*sf);
366 	} else {
367 		if (*sf != NULL)
368 			dmar_unmap_pgtbl(*sf);
369 		*idxp = idx;
370 retry:
371 		pte = dmar_map_pgtbl(domain->pgtbl_obj, idx, flags, sf);
372 		if (pte == NULL) {
373 			KASSERT(lvl > 0,
374 			    ("lost root page table page %p", domain));
375 			/*
376 			 * Page table page does not exist, allocate
377 			 * it and create a pte in the preceeding page level
378 			 * to reference the allocated page table page.
379 			 */
380 			m = dmar_pgalloc(domain->pgtbl_obj, idx, flags |
381 			    IOMMU_PGF_ZERO);
382 			if (m == NULL)
383 				return (NULL);
384 
385 			/*
386 			 * Prevent potential free while pgtbl_obj is
387 			 * unlocked in the recursive call to
388 			 * domain_pgtbl_map_pte(), if other thread did
389 			 * pte write and clean while the lock is
390 			 * dropped.
391 			 */
392 			m->ref_count++;
393 
394 			sfp = NULL;
395 			ptep = domain_pgtbl_map_pte(domain, base, lvl - 1,
396 			    flags, &idx1, &sfp);
397 			if (ptep == NULL) {
398 				KASSERT(m->pindex != 0,
399 				    ("loosing root page %p", domain));
400 				m->ref_count--;
401 				dmar_pgfree(domain->pgtbl_obj, m->pindex,
402 				    flags);
403 				return (NULL);
404 			}
405 			dmar_pte_store(&ptep->pte, DMAR_PTE_R | DMAR_PTE_W |
406 			    VM_PAGE_TO_PHYS(m));
407 			dmar_flush_pte_to_ram(domain->dmar, ptep);
408 			sf_buf_page(sfp)->ref_count += 1;
409 			m->ref_count--;
410 			dmar_unmap_pgtbl(sfp);
411 			/* Only executed once. */
412 			goto retry;
413 		}
414 	}
415 	pte += domain_pgtbl_pte_off(domain, base, lvl);
416 	return (pte);
417 }
418 
419 static int
420 domain_map_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base,
421     iommu_gaddr_t size, vm_page_t *ma, uint64_t pflags, int flags)
422 {
423 	dmar_pte_t *pte;
424 	struct sf_buf *sf;
425 	iommu_gaddr_t pg_sz, base1;
426 	vm_pindex_t pi, c, idx, run_sz;
427 	int lvl;
428 	bool superpage;
429 
430 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
431 
432 	base1 = base;
433 	flags |= IOMMU_PGF_OBJL;
434 	TD_PREP_PINNED_ASSERT;
435 
436 	for (sf = NULL, pi = 0; size > 0; base += pg_sz, size -= pg_sz,
437 	    pi += run_sz) {
438 		for (lvl = 0, c = 0, superpage = false;; lvl++) {
439 			pg_sz = domain_page_size(domain, lvl);
440 			run_sz = pg_sz >> DMAR_PAGE_SHIFT;
441 			if (lvl == domain->pglvl - 1)
442 				break;
443 			/*
444 			 * Check if the current base suitable for the
445 			 * superpage mapping.  First, verify the level.
446 			 */
447 			if (!domain_is_sp_lvl(domain, lvl))
448 				continue;
449 			/*
450 			 * Next, look at the size of the mapping and
451 			 * alignment of both guest and host addresses.
452 			 */
453 			if (size < pg_sz || (base & (pg_sz - 1)) != 0 ||
454 			    (VM_PAGE_TO_PHYS(ma[pi]) & (pg_sz - 1)) != 0)
455 				continue;
456 			/* All passed, check host pages contiguouty. */
457 			if (c == 0) {
458 				for (c = 1; c < run_sz; c++) {
459 					if (VM_PAGE_TO_PHYS(ma[pi + c]) !=
460 					    VM_PAGE_TO_PHYS(ma[pi + c - 1]) +
461 					    PAGE_SIZE)
462 						break;
463 				}
464 			}
465 			if (c >= run_sz) {
466 				superpage = true;
467 				break;
468 			}
469 		}
470 		KASSERT(size >= pg_sz,
471 		    ("mapping loop overflow %p %jx %jx %jx", domain,
472 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
473 		KASSERT(pg_sz > 0, ("pg_sz 0 lvl %d", lvl));
474 		pte = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
475 		if (pte == NULL) {
476 			KASSERT((flags & IOMMU_PGF_WAITOK) == 0,
477 			    ("failed waitable pte alloc %p", domain));
478 			if (sf != NULL)
479 				dmar_unmap_pgtbl(sf);
480 			domain_unmap_buf_locked(domain, base1, base - base1,
481 			    flags);
482 			TD_PINNED_ASSERT;
483 			return (ENOMEM);
484 		}
485 		dmar_pte_store(&pte->pte, VM_PAGE_TO_PHYS(ma[pi]) | pflags |
486 		    (superpage ? DMAR_PTE_SP : 0));
487 		dmar_flush_pte_to_ram(domain->dmar, pte);
488 		sf_buf_page(sf)->ref_count += 1;
489 	}
490 	if (sf != NULL)
491 		dmar_unmap_pgtbl(sf);
492 	TD_PINNED_ASSERT;
493 	return (0);
494 }
495 
496 static int
497 domain_map_buf(struct iommu_domain *iodom, iommu_gaddr_t base,
498     iommu_gaddr_t size, vm_page_t *ma, uint64_t eflags, int flags)
499 {
500 	struct dmar_domain *domain;
501 	struct dmar_unit *unit;
502 	uint64_t pflags;
503 	int error;
504 
505 	pflags = ((eflags & IOMMU_MAP_ENTRY_READ) != 0 ? DMAR_PTE_R : 0) |
506 	    ((eflags & IOMMU_MAP_ENTRY_WRITE) != 0 ? DMAR_PTE_W : 0) |
507 	    ((eflags & IOMMU_MAP_ENTRY_SNOOP) != 0 ? DMAR_PTE_SNP : 0) |
508 	    ((eflags & IOMMU_MAP_ENTRY_TM) != 0 ? DMAR_PTE_TM : 0);
509 
510 	domain = IODOM2DOM(iodom);
511 	unit = domain->dmar;
512 
513 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) == 0,
514 	    ("modifying idmap pagetable domain %p", domain));
515 	KASSERT((base & DMAR_PAGE_MASK) == 0,
516 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
517 	    (uintmax_t)size));
518 	KASSERT((size & DMAR_PAGE_MASK) == 0,
519 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
520 	    (uintmax_t)size));
521 	KASSERT(size > 0, ("zero size %p %jx %jx", domain, (uintmax_t)base,
522 	    (uintmax_t)size));
523 	KASSERT(base < (1ULL << domain->agaw),
524 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
525 	    (uintmax_t)size, domain->agaw));
526 	KASSERT(base + size < (1ULL << domain->agaw),
527 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
528 	    (uintmax_t)size, domain->agaw));
529 	KASSERT(base + size > base,
530 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
531 	    (uintmax_t)size));
532 	KASSERT((pflags & (DMAR_PTE_R | DMAR_PTE_W)) != 0,
533 	    ("neither read nor write %jx", (uintmax_t)pflags));
534 	KASSERT((pflags & ~(DMAR_PTE_R | DMAR_PTE_W | DMAR_PTE_SNP |
535 	    DMAR_PTE_TM)) == 0,
536 	    ("invalid pte flags %jx", (uintmax_t)pflags));
537 	KASSERT((pflags & DMAR_PTE_SNP) == 0 ||
538 	    (unit->hw_ecap & DMAR_ECAP_SC) != 0,
539 	    ("PTE_SNP for dmar without snoop control %p %jx",
540 	    domain, (uintmax_t)pflags));
541 	KASSERT((pflags & DMAR_PTE_TM) == 0 ||
542 	    (unit->hw_ecap & DMAR_ECAP_DI) != 0,
543 	    ("PTE_TM for dmar without DIOTLB %p %jx",
544 	    domain, (uintmax_t)pflags));
545 	KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags));
546 
547 	DMAR_DOMAIN_PGLOCK(domain);
548 	error = domain_map_buf_locked(domain, base, size, ma, pflags, flags);
549 	DMAR_DOMAIN_PGUNLOCK(domain);
550 	if (error != 0)
551 		return (error);
552 
553 	if ((unit->hw_cap & DMAR_CAP_CM) != 0)
554 		domain_flush_iotlb_sync(domain, base, size);
555 	else if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
556 		/* See 11.1 Write Buffer Flushing. */
557 		DMAR_LOCK(unit);
558 		dmar_flush_write_bufs(unit);
559 		DMAR_UNLOCK(unit);
560 	}
561 	return (0);
562 }
563 
564 static void domain_unmap_clear_pte(struct dmar_domain *domain,
565     iommu_gaddr_t base, int lvl, int flags, dmar_pte_t *pte,
566     struct sf_buf **sf, bool free_fs);
567 
568 static void
569 domain_free_pgtbl_pde(struct dmar_domain *domain, iommu_gaddr_t base,
570     int lvl, int flags)
571 {
572 	struct sf_buf *sf;
573 	dmar_pte_t *pde;
574 	vm_pindex_t idx;
575 
576 	sf = NULL;
577 	pde = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
578 	domain_unmap_clear_pte(domain, base, lvl, flags, pde, &sf, true);
579 }
580 
581 static void
582 domain_unmap_clear_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl,
583     int flags, dmar_pte_t *pte, struct sf_buf **sf, bool free_sf)
584 {
585 	vm_page_t m;
586 
587 	dmar_pte_clear(&pte->pte);
588 	dmar_flush_pte_to_ram(domain->dmar, pte);
589 	m = sf_buf_page(*sf);
590 	if (free_sf) {
591 		dmar_unmap_pgtbl(*sf);
592 		*sf = NULL;
593 	}
594 	m->ref_count--;
595 	if (m->ref_count != 0)
596 		return;
597 	KASSERT(lvl != 0,
598 	    ("lost reference (lvl) on root pg domain %p base %jx lvl %d",
599 	    domain, (uintmax_t)base, lvl));
600 	KASSERT(m->pindex != 0,
601 	    ("lost reference (idx) on root pg domain %p base %jx lvl %d",
602 	    domain, (uintmax_t)base, lvl));
603 	dmar_pgfree(domain->pgtbl_obj, m->pindex, flags);
604 	domain_free_pgtbl_pde(domain, base, lvl - 1, flags);
605 }
606 
607 /*
608  * Assumes that the unmap is never partial.
609  */
610 static int
611 domain_unmap_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base,
612     iommu_gaddr_t size, int flags)
613 {
614 	dmar_pte_t *pte;
615 	struct sf_buf *sf;
616 	vm_pindex_t idx;
617 	iommu_gaddr_t pg_sz;
618 	int lvl;
619 
620 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
621 	if (size == 0)
622 		return (0);
623 
624 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) == 0,
625 	    ("modifying idmap pagetable domain %p", domain));
626 	KASSERT((base & DMAR_PAGE_MASK) == 0,
627 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
628 	    (uintmax_t)size));
629 	KASSERT((size & DMAR_PAGE_MASK) == 0,
630 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
631 	    (uintmax_t)size));
632 	KASSERT(base < (1ULL << domain->agaw),
633 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
634 	    (uintmax_t)size, domain->agaw));
635 	KASSERT(base + size < (1ULL << domain->agaw),
636 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
637 	    (uintmax_t)size, domain->agaw));
638 	KASSERT(base + size > base,
639 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
640 	    (uintmax_t)size));
641 	KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags));
642 
643 	pg_sz = 0; /* silence gcc */
644 	flags |= IOMMU_PGF_OBJL;
645 	TD_PREP_PINNED_ASSERT;
646 
647 	for (sf = NULL; size > 0; base += pg_sz, size -= pg_sz) {
648 		for (lvl = 0; lvl < domain->pglvl; lvl++) {
649 			if (lvl != domain->pglvl - 1 &&
650 			    !domain_is_sp_lvl(domain, lvl))
651 				continue;
652 			pg_sz = domain_page_size(domain, lvl);
653 			if (pg_sz > size)
654 				continue;
655 			pte = domain_pgtbl_map_pte(domain, base, lvl, flags,
656 			    &idx, &sf);
657 			KASSERT(pte != NULL,
658 			    ("sleeping or page missed %p %jx %d 0x%x",
659 			    domain, (uintmax_t)base, lvl, flags));
660 			if ((pte->pte & DMAR_PTE_SP) != 0 ||
661 			    lvl == domain->pglvl - 1) {
662 				domain_unmap_clear_pte(domain, base, lvl,
663 				    flags, pte, &sf, false);
664 				break;
665 			}
666 		}
667 		KASSERT(size >= pg_sz,
668 		    ("unmapping loop overflow %p %jx %jx %jx", domain,
669 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
670 	}
671 	if (sf != NULL)
672 		dmar_unmap_pgtbl(sf);
673 	/*
674 	 * See 11.1 Write Buffer Flushing for an explanation why RWBF
675 	 * can be ignored there.
676 	 */
677 
678 	TD_PINNED_ASSERT;
679 	return (0);
680 }
681 
682 static int
683 domain_unmap_buf(struct iommu_domain *iodom, iommu_gaddr_t base,
684     iommu_gaddr_t size, int flags)
685 {
686 	struct dmar_domain *domain;
687 	int error;
688 
689 	domain = IODOM2DOM(iodom);
690 
691 	DMAR_DOMAIN_PGLOCK(domain);
692 	error = domain_unmap_buf_locked(domain, base, size, flags);
693 	DMAR_DOMAIN_PGUNLOCK(domain);
694 	return (error);
695 }
696 
697 int
698 domain_alloc_pgtbl(struct dmar_domain *domain)
699 {
700 	vm_page_t m;
701 
702 	KASSERT(domain->pgtbl_obj == NULL,
703 	    ("already initialized %p", domain));
704 
705 	domain->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
706 	    IDX_TO_OFF(pglvl_max_pages(domain->pglvl)), 0, 0, NULL);
707 	DMAR_DOMAIN_PGLOCK(domain);
708 	m = dmar_pgalloc(domain->pgtbl_obj, 0, IOMMU_PGF_WAITOK |
709 	    IOMMU_PGF_ZERO | IOMMU_PGF_OBJL);
710 	/* No implicit free of the top level page table page. */
711 	m->ref_count = 1;
712 	DMAR_DOMAIN_PGUNLOCK(domain);
713 	DMAR_LOCK(domain->dmar);
714 	domain->iodom.flags |= IOMMU_DOMAIN_PGTBL_INITED;
715 	DMAR_UNLOCK(domain->dmar);
716 	return (0);
717 }
718 
719 void
720 domain_free_pgtbl(struct dmar_domain *domain)
721 {
722 	vm_object_t obj;
723 	vm_page_t m;
724 
725 	obj = domain->pgtbl_obj;
726 	if (obj == NULL) {
727 		KASSERT((domain->dmar->hw_ecap & DMAR_ECAP_PT) != 0 &&
728 		    (domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0,
729 		    ("lost pagetable object domain %p", domain));
730 		return;
731 	}
732 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
733 	domain->pgtbl_obj = NULL;
734 
735 	if ((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0) {
736 		put_idmap_pgtbl(obj);
737 		domain->iodom.flags &= ~IOMMU_DOMAIN_IDMAP;
738 		return;
739 	}
740 
741 	/* Obliterate ref_counts */
742 	VM_OBJECT_ASSERT_WLOCKED(obj);
743 	for (m = vm_page_lookup(obj, 0); m != NULL; m = vm_page_next(m))
744 		m->ref_count = 0;
745 	VM_OBJECT_WUNLOCK(obj);
746 	vm_object_deallocate(obj);
747 }
748 
749 static inline uint64_t
750 domain_wait_iotlb_flush(struct dmar_unit *unit, uint64_t wt, int iro)
751 {
752 	uint64_t iotlbr;
753 
754 	dmar_write8(unit, iro + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
755 	    DMAR_IOTLB_DR | DMAR_IOTLB_DW | wt);
756 	for (;;) {
757 		iotlbr = dmar_read8(unit, iro + DMAR_IOTLB_REG_OFF);
758 		if ((iotlbr & DMAR_IOTLB_IVT) == 0)
759 			break;
760 		cpu_spinwait();
761 	}
762 	return (iotlbr);
763 }
764 
765 void
766 domain_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base,
767     iommu_gaddr_t size)
768 {
769 	struct dmar_unit *unit;
770 	iommu_gaddr_t isize;
771 	uint64_t iotlbr;
772 	int am, iro;
773 
774 	unit = domain->dmar;
775 	KASSERT(!unit->qi_enabled, ("dmar%d: sync iotlb flush call",
776 	    unit->iommu.unit));
777 	iro = DMAR_ECAP_IRO(unit->hw_ecap) * 16;
778 	DMAR_LOCK(unit);
779 	if ((unit->hw_cap & DMAR_CAP_PSI) == 0 || size > 2 * 1024 * 1024) {
780 		iotlbr = domain_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_DOM |
781 		    DMAR_IOTLB_DID(domain->domain), iro);
782 		KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
783 		    DMAR_IOTLB_IAIG_INVLD,
784 		    ("dmar%d: invalidation failed %jx", unit->iommu.unit,
785 		    (uintmax_t)iotlbr));
786 	} else {
787 		for (; size > 0; base += isize, size -= isize) {
788 			am = calc_am(unit, base, size, &isize);
789 			dmar_write8(unit, iro, base | am);
790 			iotlbr = domain_wait_iotlb_flush(unit,
791 			    DMAR_IOTLB_IIRG_PAGE |
792 			    DMAR_IOTLB_DID(domain->domain), iro);
793 			KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
794 			    DMAR_IOTLB_IAIG_INVLD,
795 			    ("dmar%d: PSI invalidation failed "
796 			    "iotlbr 0x%jx base 0x%jx size 0x%jx am %d",
797 			    unit->iommu.unit, (uintmax_t)iotlbr,
798 			    (uintmax_t)base, (uintmax_t)size, am));
799 			/*
800 			 * Any non-page granularity covers whole guest
801 			 * address space for the domain.
802 			 */
803 			if ((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
804 			    DMAR_IOTLB_IAIG_PAGE)
805 				break;
806 		}
807 	}
808 	DMAR_UNLOCK(unit);
809 }
810 
811 const struct iommu_domain_map_ops dmar_domain_map_ops = {
812 	.map = domain_map_buf,
813 	.unmap = domain_unmap_buf,
814 };
815