xref: /freebsd/sys/x86/iommu/intel_idpgtbl.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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 <dev/pci/pcireg.h>
62 #include <machine/atomic.h>
63 #include <machine/bus.h>
64 #include <machine/cpu.h>
65 #include <machine/md_var.h>
66 #include <machine/specialreg.h>
67 #include <x86/include/busdma_impl.h>
68 #include <dev/iommu/busdma_iommu.h>
69 #include <x86/iommu/intel_reg.h>
70 #include <x86/iommu/intel_dmar.h>
71 
72 static int domain_unmap_buf_locked(struct dmar_domain *domain,
73     iommu_gaddr_t base, iommu_gaddr_t size, int flags);
74 
75 /*
76  * The cache of the identity mapping page tables for the DMARs.  Using
77  * the cache saves significant amount of memory for page tables by
78  * reusing the page tables, since usually DMARs are identical and have
79  * the same capabilities.  Still, cache records the information needed
80  * to match DMAR capabilities and page table format, to correctly
81  * handle different DMARs.
82  */
83 
84 struct idpgtbl {
85 	iommu_gaddr_t maxaddr;	/* Page table covers the guest address
86 				   range [0..maxaddr) */
87 	int pglvl;		/* Total page table levels ignoring
88 				   superpages */
89 	int leaf;		/* The last materialized page table
90 				   level, it is non-zero if superpages
91 				   are supported */
92 	vm_object_t pgtbl_obj;	/* The page table pages */
93 	LIST_ENTRY(idpgtbl) link;
94 };
95 
96 static struct sx idpgtbl_lock;
97 SX_SYSINIT(idpgtbl, &idpgtbl_lock, "idpgtbl");
98 static LIST_HEAD(, idpgtbl) idpgtbls = LIST_HEAD_INITIALIZER(idpgtbls);
99 static MALLOC_DEFINE(M_DMAR_IDPGTBL, "dmar_idpgtbl",
100     "Intel DMAR Identity mappings cache elements");
101 
102 /*
103  * Build the next level of the page tables for the identity mapping.
104  * - lvl is the level to build;
105  * - idx is the index of the page table page in the pgtbl_obj, which is
106  *   being allocated filled now;
107  * - addr is the starting address in the bus address space which is
108  *   mapped by the page table page.
109  */
110 static void
111 domain_idmap_nextlvl(struct idpgtbl *tbl, int lvl, vm_pindex_t idx,
112     iommu_gaddr_t addr)
113 {
114 	vm_page_t m1;
115 	dmar_pte_t *pte;
116 	struct sf_buf *sf;
117 	iommu_gaddr_t f, pg_sz;
118 	vm_pindex_t base;
119 	int i;
120 
121 	VM_OBJECT_ASSERT_LOCKED(tbl->pgtbl_obj);
122 	if (addr >= tbl->maxaddr)
123 		return;
124 	(void)dmar_pgalloc(tbl->pgtbl_obj, idx, IOMMU_PGF_OBJL |
125 	    IOMMU_PGF_WAITOK | IOMMU_PGF_ZERO);
126 	base = idx * DMAR_NPTEPG + 1; /* Index of the first child page of idx */
127 	pg_sz = pglvl_page_size(tbl->pglvl, lvl);
128 	if (lvl != tbl->leaf) {
129 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz)
130 			domain_idmap_nextlvl(tbl, lvl + 1, base + i, f);
131 	}
132 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
133 	pte = dmar_map_pgtbl(tbl->pgtbl_obj, idx, IOMMU_PGF_WAITOK, &sf);
134 	if (lvl == tbl->leaf) {
135 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) {
136 			if (f >= tbl->maxaddr)
137 				break;
138 			pte[i].pte = (DMAR_PTE_ADDR_MASK & f) |
139 			    DMAR_PTE_R | DMAR_PTE_W;
140 		}
141 	} else {
142 		for (i = 0, f = addr; i < DMAR_NPTEPG; i++, f += pg_sz) {
143 			if (f >= tbl->maxaddr)
144 				break;
145 			m1 = dmar_pgalloc(tbl->pgtbl_obj, base + i,
146 			    IOMMU_PGF_NOALLOC);
147 			KASSERT(m1 != NULL, ("lost page table page"));
148 			pte[i].pte = (DMAR_PTE_ADDR_MASK &
149 			    VM_PAGE_TO_PHYS(m1)) | DMAR_PTE_R | DMAR_PTE_W;
150 		}
151 	}
152 	/* domain_get_idmap_pgtbl flushes CPU cache if needed. */
153 	dmar_unmap_pgtbl(sf);
154 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
155 }
156 
157 /*
158  * Find a ready and compatible identity-mapping page table in the
159  * cache. If not found, populate the identity-mapping page table for
160  * the context, up to the maxaddr. The maxaddr byte is allowed to be
161  * not mapped, which is aligned with the definition of Maxmem as the
162  * highest usable physical address + 1.  If superpages are used, the
163  * maxaddr is typically mapped.
164  */
165 vm_object_t
166 domain_get_idmap_pgtbl(struct dmar_domain *domain, iommu_gaddr_t maxaddr)
167 {
168 	struct dmar_unit *unit;
169 	struct idpgtbl *tbl;
170 	vm_object_t res;
171 	vm_page_t m;
172 	int leaf, i;
173 
174 	leaf = 0; /* silence gcc */
175 
176 	/*
177 	 * First, determine where to stop the paging structures.
178 	 */
179 	for (i = 0; i < domain->pglvl; i++) {
180 		if (i == domain->pglvl - 1 || domain_is_sp_lvl(domain, i)) {
181 			leaf = i;
182 			break;
183 		}
184 	}
185 
186 	/*
187 	 * Search the cache for a compatible page table.  Qualified
188 	 * page table must map up to maxaddr, its level must be
189 	 * supported by the DMAR and leaf should be equal to the
190 	 * calculated value.  The later restriction could be lifted
191 	 * but I believe it is currently impossible to have any
192 	 * deviations for existing hardware.
193 	 */
194 	sx_slock(&idpgtbl_lock);
195 	LIST_FOREACH(tbl, &idpgtbls, link) {
196 		if (tbl->maxaddr >= maxaddr &&
197 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
198 		    tbl->leaf == leaf) {
199 			res = tbl->pgtbl_obj;
200 			vm_object_reference(res);
201 			sx_sunlock(&idpgtbl_lock);
202 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
203 			goto end;
204 		}
205 	}
206 
207 	/*
208 	 * Not found in cache, relock the cache into exclusive mode to
209 	 * be able to add element, and recheck cache again after the
210 	 * relock.
211 	 */
212 	sx_sunlock(&idpgtbl_lock);
213 	sx_xlock(&idpgtbl_lock);
214 	LIST_FOREACH(tbl, &idpgtbls, link) {
215 		if (tbl->maxaddr >= maxaddr &&
216 		    dmar_pglvl_supported(domain->dmar, tbl->pglvl) &&
217 		    tbl->leaf == leaf) {
218 			res = tbl->pgtbl_obj;
219 			vm_object_reference(res);
220 			sx_xunlock(&idpgtbl_lock);
221 			domain->pglvl = tbl->pglvl; /* XXXKIB ? */
222 			return (res);
223 		}
224 	}
225 
226 	/*
227 	 * Still not found, create new page table.
228 	 */
229 	tbl = malloc(sizeof(*tbl), M_DMAR_IDPGTBL, M_WAITOK);
230 	tbl->pglvl = domain->pglvl;
231 	tbl->leaf = leaf;
232 	tbl->maxaddr = maxaddr;
233 	tbl->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
234 	    IDX_TO_OFF(pglvl_max_pages(tbl->pglvl)), 0, 0, NULL);
235 	VM_OBJECT_WLOCK(tbl->pgtbl_obj);
236 	domain_idmap_nextlvl(tbl, 0, 0, 0);
237 	VM_OBJECT_WUNLOCK(tbl->pgtbl_obj);
238 	LIST_INSERT_HEAD(&idpgtbls, tbl, link);
239 	res = tbl->pgtbl_obj;
240 	vm_object_reference(res);
241 	sx_xunlock(&idpgtbl_lock);
242 
243 end:
244 	/*
245 	 * Table was found or created.
246 	 *
247 	 * If DMAR does not snoop paging structures accesses, flush
248 	 * CPU cache to memory.  Note that dmar_unmap_pgtbl() coherent
249 	 * argument was possibly invalid at the time of the identity
250 	 * page table creation, since DMAR which was passed at the
251 	 * time of creation could be coherent, while current DMAR is
252 	 * not.
253 	 *
254 	 * If DMAR cannot look into the chipset write buffer, flush it
255 	 * as well.
256 	 */
257 	unit = domain->dmar;
258 	if (!DMAR_IS_COHERENT(unit)) {
259 		VM_OBJECT_WLOCK(res);
260 		for (m = vm_page_lookup(res, 0); m != NULL;
261 		     m = vm_page_next(m))
262 			pmap_invalidate_cache_pages(&m, 1);
263 		VM_OBJECT_WUNLOCK(res);
264 	}
265 	if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
266 		DMAR_LOCK(unit);
267 		dmar_flush_write_bufs(unit);
268 		DMAR_UNLOCK(unit);
269 	}
270 
271 	return (res);
272 }
273 
274 /*
275  * Return a reference to the identity mapping page table to the cache.
276  */
277 void
278 put_idmap_pgtbl(vm_object_t obj)
279 {
280 	struct idpgtbl *tbl, *tbl1;
281 	vm_object_t rmobj;
282 
283 	sx_slock(&idpgtbl_lock);
284 	KASSERT(obj->ref_count >= 2, ("lost cache reference"));
285 	vm_object_deallocate(obj);
286 
287 	/*
288 	 * Cache always owns one last reference on the page table object.
289 	 * If there is an additional reference, object must stay.
290 	 */
291 	if (obj->ref_count > 1) {
292 		sx_sunlock(&idpgtbl_lock);
293 		return;
294 	}
295 
296 	/*
297 	 * Cache reference is the last, remove cache element and free
298 	 * page table object, returning the page table pages to the
299 	 * system.
300 	 */
301 	sx_sunlock(&idpgtbl_lock);
302 	sx_xlock(&idpgtbl_lock);
303 	LIST_FOREACH_SAFE(tbl, &idpgtbls, link, tbl1) {
304 		rmobj = tbl->pgtbl_obj;
305 		if (rmobj->ref_count == 1) {
306 			LIST_REMOVE(tbl, link);
307 			atomic_subtract_int(&dmar_tbl_pagecnt,
308 			    rmobj->resident_page_count);
309 			vm_object_deallocate(rmobj);
310 			free(tbl, M_DMAR_IDPGTBL);
311 		}
312 	}
313 	sx_xunlock(&idpgtbl_lock);
314 }
315 
316 /*
317  * The core routines to map and unmap host pages at the given guest
318  * address.  Support superpages.
319  */
320 
321 /*
322  * Index of the pte for the guest address base in the page table at
323  * the level lvl.
324  */
325 static int
326 domain_pgtbl_pte_off(struct dmar_domain *domain, iommu_gaddr_t base, int lvl)
327 {
328 
329 	base >>= DMAR_PAGE_SHIFT + (domain->pglvl - lvl - 1) *
330 	    DMAR_NPTEPGSHIFT;
331 	return (base & DMAR_PTEMASK);
332 }
333 
334 /*
335  * Returns the page index of the page table page in the page table
336  * object, which maps the given address base at the page table level
337  * lvl.
338  */
339 static vm_pindex_t
340 domain_pgtbl_get_pindex(struct dmar_domain *domain, iommu_gaddr_t base, int lvl)
341 {
342 	vm_pindex_t idx, pidx;
343 	int i;
344 
345 	KASSERT(lvl >= 0 && lvl < domain->pglvl,
346 	    ("wrong lvl %p %d", domain, lvl));
347 
348 	for (pidx = idx = 0, i = 0; i < lvl; i++, pidx = idx) {
349 		idx = domain_pgtbl_pte_off(domain, base, i) +
350 		    pidx * DMAR_NPTEPG + 1;
351 	}
352 	return (idx);
353 }
354 
355 static dmar_pte_t *
356 domain_pgtbl_map_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl,
357     int flags, vm_pindex_t *idxp, struct sf_buf **sf)
358 {
359 	vm_page_t m;
360 	struct sf_buf *sfp;
361 	dmar_pte_t *pte, *ptep;
362 	vm_pindex_t idx, idx1;
363 
364 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
365 	KASSERT((flags & IOMMU_PGF_OBJL) != 0, ("lost PGF_OBJL"));
366 
367 	idx = domain_pgtbl_get_pindex(domain, base, lvl);
368 	if (*sf != NULL && idx == *idxp) {
369 		pte = (dmar_pte_t *)sf_buf_kva(*sf);
370 	} else {
371 		if (*sf != NULL)
372 			dmar_unmap_pgtbl(*sf);
373 		*idxp = idx;
374 retry:
375 		pte = dmar_map_pgtbl(domain->pgtbl_obj, idx, flags, sf);
376 		if (pte == NULL) {
377 			KASSERT(lvl > 0,
378 			    ("lost root page table page %p", domain));
379 			/*
380 			 * Page table page does not exist, allocate
381 			 * it and create a pte in the preceeding page level
382 			 * to reference the allocated page table page.
383 			 */
384 			m = dmar_pgalloc(domain->pgtbl_obj, idx, flags |
385 			    IOMMU_PGF_ZERO);
386 			if (m == NULL)
387 				return (NULL);
388 
389 			/*
390 			 * Prevent potential free while pgtbl_obj is
391 			 * unlocked in the recursive call to
392 			 * domain_pgtbl_map_pte(), if other thread did
393 			 * pte write and clean while the lock is
394 			 * dropped.
395 			 */
396 			m->ref_count++;
397 
398 			sfp = NULL;
399 			ptep = domain_pgtbl_map_pte(domain, base, lvl - 1,
400 			    flags, &idx1, &sfp);
401 			if (ptep == NULL) {
402 				KASSERT(m->pindex != 0,
403 				    ("loosing root page %p", domain));
404 				m->ref_count--;
405 				dmar_pgfree(domain->pgtbl_obj, m->pindex,
406 				    flags);
407 				return (NULL);
408 			}
409 			dmar_pte_store(&ptep->pte, DMAR_PTE_R | DMAR_PTE_W |
410 			    VM_PAGE_TO_PHYS(m));
411 			dmar_flush_pte_to_ram(domain->dmar, ptep);
412 			sf_buf_page(sfp)->ref_count += 1;
413 			m->ref_count--;
414 			dmar_unmap_pgtbl(sfp);
415 			/* Only executed once. */
416 			goto retry;
417 		}
418 	}
419 	pte += domain_pgtbl_pte_off(domain, base, lvl);
420 	return (pte);
421 }
422 
423 static int
424 domain_map_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base,
425     iommu_gaddr_t size, vm_page_t *ma, uint64_t pflags, int flags)
426 {
427 	dmar_pte_t *pte;
428 	struct sf_buf *sf;
429 	iommu_gaddr_t pg_sz, base1, size1;
430 	vm_pindex_t pi, c, idx, run_sz;
431 	int lvl;
432 	bool superpage;
433 
434 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
435 
436 	base1 = base;
437 	size1 = size;
438 	flags |= IOMMU_PGF_OBJL;
439 	TD_PREP_PINNED_ASSERT;
440 
441 	for (sf = NULL, pi = 0; size > 0; base += pg_sz, size -= pg_sz,
442 	    pi += run_sz) {
443 		for (lvl = 0, c = 0, superpage = false;; lvl++) {
444 			pg_sz = domain_page_size(domain, lvl);
445 			run_sz = pg_sz >> DMAR_PAGE_SHIFT;
446 			if (lvl == domain->pglvl - 1)
447 				break;
448 			/*
449 			 * Check if the current base suitable for the
450 			 * superpage mapping.  First, verify the level.
451 			 */
452 			if (!domain_is_sp_lvl(domain, lvl))
453 				continue;
454 			/*
455 			 * Next, look at the size of the mapping and
456 			 * alignment of both guest and host addresses.
457 			 */
458 			if (size < pg_sz || (base & (pg_sz - 1)) != 0 ||
459 			    (VM_PAGE_TO_PHYS(ma[pi]) & (pg_sz - 1)) != 0)
460 				continue;
461 			/* All passed, check host pages contiguouty. */
462 			if (c == 0) {
463 				for (c = 1; c < run_sz; c++) {
464 					if (VM_PAGE_TO_PHYS(ma[pi + c]) !=
465 					    VM_PAGE_TO_PHYS(ma[pi + c - 1]) +
466 					    PAGE_SIZE)
467 						break;
468 				}
469 			}
470 			if (c >= run_sz) {
471 				superpage = true;
472 				break;
473 			}
474 		}
475 		KASSERT(size >= pg_sz,
476 		    ("mapping loop overflow %p %jx %jx %jx", domain,
477 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
478 		KASSERT(pg_sz > 0, ("pg_sz 0 lvl %d", lvl));
479 		pte = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
480 		if (pte == NULL) {
481 			KASSERT((flags & IOMMU_PGF_WAITOK) == 0,
482 			    ("failed waitable pte alloc %p", domain));
483 			if (sf != NULL)
484 				dmar_unmap_pgtbl(sf);
485 			domain_unmap_buf_locked(domain, base1, base - base1,
486 			    flags);
487 			TD_PINNED_ASSERT;
488 			return (ENOMEM);
489 		}
490 		dmar_pte_store(&pte->pte, VM_PAGE_TO_PHYS(ma[pi]) | pflags |
491 		    (superpage ? DMAR_PTE_SP : 0));
492 		dmar_flush_pte_to_ram(domain->dmar, pte);
493 		sf_buf_page(sf)->ref_count += 1;
494 	}
495 	if (sf != NULL)
496 		dmar_unmap_pgtbl(sf);
497 	TD_PINNED_ASSERT;
498 	return (0);
499 }
500 
501 static int
502 domain_map_buf(struct iommu_domain *iodom, iommu_gaddr_t base,
503     iommu_gaddr_t size, vm_page_t *ma, uint64_t eflags, int flags)
504 {
505 	struct dmar_domain *domain;
506 	struct dmar_unit *unit;
507 	uint64_t pflags;
508 	int error;
509 
510 	pflags = ((eflags & IOMMU_MAP_ENTRY_READ) != 0 ? DMAR_PTE_R : 0) |
511 	    ((eflags & IOMMU_MAP_ENTRY_WRITE) != 0 ? DMAR_PTE_W : 0) |
512 	    ((eflags & IOMMU_MAP_ENTRY_SNOOP) != 0 ? DMAR_PTE_SNP : 0) |
513 	    ((eflags & IOMMU_MAP_ENTRY_TM) != 0 ? DMAR_PTE_TM : 0);
514 
515 	domain = IODOM2DOM(iodom);
516 	unit = domain->dmar;
517 
518 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) == 0,
519 	    ("modifying idmap pagetable domain %p", domain));
520 	KASSERT((base & DMAR_PAGE_MASK) == 0,
521 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
522 	    (uintmax_t)size));
523 	KASSERT((size & DMAR_PAGE_MASK) == 0,
524 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
525 	    (uintmax_t)size));
526 	KASSERT(size > 0, ("zero size %p %jx %jx", domain, (uintmax_t)base,
527 	    (uintmax_t)size));
528 	KASSERT(base < (1ULL << domain->agaw),
529 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
530 	    (uintmax_t)size, domain->agaw));
531 	KASSERT(base + size < (1ULL << domain->agaw),
532 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
533 	    (uintmax_t)size, domain->agaw));
534 	KASSERT(base + size > base,
535 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
536 	    (uintmax_t)size));
537 	KASSERT((pflags & (DMAR_PTE_R | DMAR_PTE_W)) != 0,
538 	    ("neither read nor write %jx", (uintmax_t)pflags));
539 	KASSERT((pflags & ~(DMAR_PTE_R | DMAR_PTE_W | DMAR_PTE_SNP |
540 	    DMAR_PTE_TM)) == 0,
541 	    ("invalid pte flags %jx", (uintmax_t)pflags));
542 	KASSERT((pflags & DMAR_PTE_SNP) == 0 ||
543 	    (unit->hw_ecap & DMAR_ECAP_SC) != 0,
544 	    ("PTE_SNP for dmar without snoop control %p %jx",
545 	    domain, (uintmax_t)pflags));
546 	KASSERT((pflags & DMAR_PTE_TM) == 0 ||
547 	    (unit->hw_ecap & DMAR_ECAP_DI) != 0,
548 	    ("PTE_TM for dmar without DIOTLB %p %jx",
549 	    domain, (uintmax_t)pflags));
550 	KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags));
551 
552 	DMAR_DOMAIN_PGLOCK(domain);
553 	error = domain_map_buf_locked(domain, base, size, ma, pflags, flags);
554 	DMAR_DOMAIN_PGUNLOCK(domain);
555 	if (error != 0)
556 		return (error);
557 
558 	if ((unit->hw_cap & DMAR_CAP_CM) != 0)
559 		domain_flush_iotlb_sync(domain, base, size);
560 	else if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
561 		/* See 11.1 Write Buffer Flushing. */
562 		DMAR_LOCK(unit);
563 		dmar_flush_write_bufs(unit);
564 		DMAR_UNLOCK(unit);
565 	}
566 	return (0);
567 }
568 
569 static void domain_unmap_clear_pte(struct dmar_domain *domain,
570     iommu_gaddr_t base, int lvl, int flags, dmar_pte_t *pte,
571     struct sf_buf **sf, bool free_fs);
572 
573 static void
574 domain_free_pgtbl_pde(struct dmar_domain *domain, iommu_gaddr_t base,
575     int lvl, int flags)
576 {
577 	struct sf_buf *sf;
578 	dmar_pte_t *pde;
579 	vm_pindex_t idx;
580 
581 	sf = NULL;
582 	pde = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
583 	domain_unmap_clear_pte(domain, base, lvl, flags, pde, &sf, true);
584 }
585 
586 static void
587 domain_unmap_clear_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl,
588     int flags, dmar_pte_t *pte, struct sf_buf **sf, bool free_sf)
589 {
590 	vm_page_t m;
591 
592 	dmar_pte_clear(&pte->pte);
593 	dmar_flush_pte_to_ram(domain->dmar, pte);
594 	m = sf_buf_page(*sf);
595 	if (free_sf) {
596 		dmar_unmap_pgtbl(*sf);
597 		*sf = NULL;
598 	}
599 	m->ref_count--;
600 	if (m->ref_count != 0)
601 		return;
602 	KASSERT(lvl != 0,
603 	    ("lost reference (lvl) on root pg domain %p base %jx lvl %d",
604 	    domain, (uintmax_t)base, lvl));
605 	KASSERT(m->pindex != 0,
606 	    ("lost reference (idx) on root pg domain %p base %jx lvl %d",
607 	    domain, (uintmax_t)base, lvl));
608 	dmar_pgfree(domain->pgtbl_obj, m->pindex, flags);
609 	domain_free_pgtbl_pde(domain, base, lvl - 1, flags);
610 }
611 
612 /*
613  * Assumes that the unmap is never partial.
614  */
615 static int
616 domain_unmap_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base,
617     iommu_gaddr_t size, int flags)
618 {
619 	dmar_pte_t *pte;
620 	struct sf_buf *sf;
621 	vm_pindex_t idx;
622 	iommu_gaddr_t pg_sz;
623 	int lvl;
624 
625 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
626 	if (size == 0)
627 		return (0);
628 
629 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) == 0,
630 	    ("modifying idmap pagetable domain %p", domain));
631 	KASSERT((base & DMAR_PAGE_MASK) == 0,
632 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
633 	    (uintmax_t)size));
634 	KASSERT((size & DMAR_PAGE_MASK) == 0,
635 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
636 	    (uintmax_t)size));
637 	KASSERT(base < (1ULL << domain->agaw),
638 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
639 	    (uintmax_t)size, domain->agaw));
640 	KASSERT(base + size < (1ULL << domain->agaw),
641 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
642 	    (uintmax_t)size, domain->agaw));
643 	KASSERT(base + size > base,
644 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
645 	    (uintmax_t)size));
646 	KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags));
647 
648 	pg_sz = 0; /* silence gcc */
649 	flags |= IOMMU_PGF_OBJL;
650 	TD_PREP_PINNED_ASSERT;
651 
652 	for (sf = NULL; size > 0; base += pg_sz, size -= pg_sz) {
653 		for (lvl = 0; lvl < domain->pglvl; lvl++) {
654 			if (lvl != domain->pglvl - 1 &&
655 			    !domain_is_sp_lvl(domain, lvl))
656 				continue;
657 			pg_sz = domain_page_size(domain, lvl);
658 			if (pg_sz > size)
659 				continue;
660 			pte = domain_pgtbl_map_pte(domain, base, lvl, flags,
661 			    &idx, &sf);
662 			KASSERT(pte != NULL,
663 			    ("sleeping or page missed %p %jx %d 0x%x",
664 			    domain, (uintmax_t)base, lvl, flags));
665 			if ((pte->pte & DMAR_PTE_SP) != 0 ||
666 			    lvl == domain->pglvl - 1) {
667 				domain_unmap_clear_pte(domain, base, lvl,
668 				    flags, pte, &sf, false);
669 				break;
670 			}
671 		}
672 		KASSERT(size >= pg_sz,
673 		    ("unmapping loop overflow %p %jx %jx %jx", domain,
674 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
675 	}
676 	if (sf != NULL)
677 		dmar_unmap_pgtbl(sf);
678 	/*
679 	 * See 11.1 Write Buffer Flushing for an explanation why RWBF
680 	 * can be ignored there.
681 	 */
682 
683 	TD_PINNED_ASSERT;
684 	return (0);
685 }
686 
687 static int
688 domain_unmap_buf(struct iommu_domain *iodom, iommu_gaddr_t base,
689     iommu_gaddr_t size, int flags)
690 {
691 	struct dmar_domain *domain;
692 	int error;
693 
694 	domain = IODOM2DOM(iodom);
695 
696 	DMAR_DOMAIN_PGLOCK(domain);
697 	error = domain_unmap_buf_locked(domain, base, size, flags);
698 	DMAR_DOMAIN_PGUNLOCK(domain);
699 	return (error);
700 }
701 
702 int
703 domain_alloc_pgtbl(struct dmar_domain *domain)
704 {
705 	vm_page_t m;
706 
707 	KASSERT(domain->pgtbl_obj == NULL,
708 	    ("already initialized %p", domain));
709 
710 	domain->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
711 	    IDX_TO_OFF(pglvl_max_pages(domain->pglvl)), 0, 0, NULL);
712 	DMAR_DOMAIN_PGLOCK(domain);
713 	m = dmar_pgalloc(domain->pgtbl_obj, 0, IOMMU_PGF_WAITOK |
714 	    IOMMU_PGF_ZERO | IOMMU_PGF_OBJL);
715 	/* No implicit free of the top level page table page. */
716 	m->ref_count = 1;
717 	DMAR_DOMAIN_PGUNLOCK(domain);
718 	DMAR_LOCK(domain->dmar);
719 	domain->iodom.flags |= IOMMU_DOMAIN_PGTBL_INITED;
720 	DMAR_UNLOCK(domain->dmar);
721 	return (0);
722 }
723 
724 void
725 domain_free_pgtbl(struct dmar_domain *domain)
726 {
727 	vm_object_t obj;
728 	vm_page_t m;
729 
730 	obj = domain->pgtbl_obj;
731 	if (obj == NULL) {
732 		KASSERT((domain->dmar->hw_ecap & DMAR_ECAP_PT) != 0 &&
733 		    (domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0,
734 		    ("lost pagetable object domain %p", domain));
735 		return;
736 	}
737 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
738 	domain->pgtbl_obj = NULL;
739 
740 	if ((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0) {
741 		put_idmap_pgtbl(obj);
742 		domain->iodom.flags &= ~IOMMU_DOMAIN_IDMAP;
743 		return;
744 	}
745 
746 	/* Obliterate ref_counts */
747 	VM_OBJECT_ASSERT_WLOCKED(obj);
748 	for (m = vm_page_lookup(obj, 0); m != NULL; m = vm_page_next(m))
749 		m->ref_count = 0;
750 	VM_OBJECT_WUNLOCK(obj);
751 	vm_object_deallocate(obj);
752 }
753 
754 static inline uint64_t
755 domain_wait_iotlb_flush(struct dmar_unit *unit, uint64_t wt, int iro)
756 {
757 	uint64_t iotlbr;
758 
759 	dmar_write8(unit, iro + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
760 	    DMAR_IOTLB_DR | DMAR_IOTLB_DW | wt);
761 	for (;;) {
762 		iotlbr = dmar_read8(unit, iro + DMAR_IOTLB_REG_OFF);
763 		if ((iotlbr & DMAR_IOTLB_IVT) == 0)
764 			break;
765 		cpu_spinwait();
766 	}
767 	return (iotlbr);
768 }
769 
770 void
771 domain_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base,
772     iommu_gaddr_t size)
773 {
774 	struct dmar_unit *unit;
775 	iommu_gaddr_t isize;
776 	uint64_t iotlbr;
777 	int am, iro;
778 
779 	unit = domain->dmar;
780 	KASSERT(!unit->qi_enabled, ("dmar%d: sync iotlb flush call",
781 	    unit->iommu.unit));
782 	iro = DMAR_ECAP_IRO(unit->hw_ecap) * 16;
783 	DMAR_LOCK(unit);
784 	if ((unit->hw_cap & DMAR_CAP_PSI) == 0 || size > 2 * 1024 * 1024) {
785 		iotlbr = domain_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_DOM |
786 		    DMAR_IOTLB_DID(domain->domain), iro);
787 		KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
788 		    DMAR_IOTLB_IAIG_INVLD,
789 		    ("dmar%d: invalidation failed %jx", unit->iommu.unit,
790 		    (uintmax_t)iotlbr));
791 	} else {
792 		for (; size > 0; base += isize, size -= isize) {
793 			am = calc_am(unit, base, size, &isize);
794 			dmar_write8(unit, iro, base | am);
795 			iotlbr = domain_wait_iotlb_flush(unit,
796 			    DMAR_IOTLB_IIRG_PAGE |
797 			    DMAR_IOTLB_DID(domain->domain), iro);
798 			KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
799 			    DMAR_IOTLB_IAIG_INVLD,
800 			    ("dmar%d: PSI invalidation failed "
801 			    "iotlbr 0x%jx base 0x%jx size 0x%jx am %d",
802 			    unit->iommu.unit, (uintmax_t)iotlbr,
803 			    (uintmax_t)base, (uintmax_t)size, am));
804 			/*
805 			 * Any non-page granularity covers whole guest
806 			 * address space for the domain.
807 			 */
808 			if ((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
809 			    DMAR_IOTLB_IAIG_PAGE)
810 				break;
811 		}
812 	}
813 	DMAR_UNLOCK(unit);
814 }
815 
816 const struct iommu_domain_map_ops dmar_domain_map_ops = {
817 	.map = domain_map_buf,
818 	.unmap = domain_unmap_buf,
819 };
820