xref: /freebsd/sys/x86/iommu/intel_idpgtbl.c (revision bc7512cc58af2e8bbe5bbf5ca0059b1daa1da897)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/malloc.h>
37 #include <sys/bus.h>
38 #include <sys/interrupt.h>
39 #include <sys/kernel.h>
40 #include <sys/ktr.h>
41 #include <sys/lock.h>
42 #include <sys/memdesc.h>
43 #include <sys/mutex.h>
44 #include <sys/proc.h>
45 #include <sys/rwlock.h>
46 #include <sys/rman.h>
47 #include <sys/sf_buf.h>
48 #include <sys/sysctl.h>
49 #include <sys/taskqueue.h>
50 #include <sys/tree.h>
51 #include <sys/uio.h>
52 #include <sys/vmem.h>
53 #include <vm/vm.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_kern.h>
56 #include <vm/vm_object.h>
57 #include <vm/vm_page.h>
58 #include <vm/vm_pager.h>
59 #include <vm/vm_map.h>
60 #include <dev/pci/pcireg.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 <dev/iommu/busdma_iommu.h>
68 #include <x86/iommu/intel_reg.h>
69 #include <x86/iommu/intel_dmar.h>
70 
71 static int domain_unmap_buf_locked(struct dmar_domain *domain,
72     iommu_gaddr_t base, iommu_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 	iommu_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     iommu_gaddr_t addr)
112 {
113 	vm_page_t m1;
114 	dmar_pte_t *pte;
115 	struct sf_buf *sf;
116 	iommu_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, IOMMU_PGF_OBJL |
124 	    IOMMU_PGF_WAITOK | IOMMU_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, IOMMU_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 			    IOMMU_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, iommu_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, iommu_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, iommu_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, iommu_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 & IOMMU_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 			    IOMMU_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->ref_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->ref_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)->ref_count += 1;
412 			m->ref_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, iommu_gaddr_t base,
424     iommu_gaddr_t size, vm_page_t *ma, uint64_t pflags, int flags)
425 {
426 	dmar_pte_t *pte;
427 	struct sf_buf *sf;
428 	iommu_gaddr_t pg_sz, base1;
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 	flags |= IOMMU_PGF_OBJL;
437 	TD_PREP_PINNED_ASSERT;
438 
439 	for (sf = NULL, pi = 0; size > 0; base += pg_sz, size -= pg_sz,
440 	    pi += run_sz) {
441 		for (lvl = 0, c = 0, superpage = false;; lvl++) {
442 			pg_sz = domain_page_size(domain, lvl);
443 			run_sz = pg_sz >> DMAR_PAGE_SHIFT;
444 			if (lvl == domain->pglvl - 1)
445 				break;
446 			/*
447 			 * Check if the current base suitable for the
448 			 * superpage mapping.  First, verify the level.
449 			 */
450 			if (!domain_is_sp_lvl(domain, lvl))
451 				continue;
452 			/*
453 			 * Next, look at the size of the mapping and
454 			 * alignment of both guest and host addresses.
455 			 */
456 			if (size < pg_sz || (base & (pg_sz - 1)) != 0 ||
457 			    (VM_PAGE_TO_PHYS(ma[pi]) & (pg_sz - 1)) != 0)
458 				continue;
459 			/* All passed, check host pages contiguouty. */
460 			if (c == 0) {
461 				for (c = 1; c < run_sz; c++) {
462 					if (VM_PAGE_TO_PHYS(ma[pi + c]) !=
463 					    VM_PAGE_TO_PHYS(ma[pi + c - 1]) +
464 					    PAGE_SIZE)
465 						break;
466 				}
467 			}
468 			if (c >= run_sz) {
469 				superpage = true;
470 				break;
471 			}
472 		}
473 		KASSERT(size >= pg_sz,
474 		    ("mapping loop overflow %p %jx %jx %jx", domain,
475 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
476 		KASSERT(pg_sz > 0, ("pg_sz 0 lvl %d", lvl));
477 		pte = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
478 		if (pte == NULL) {
479 			KASSERT((flags & IOMMU_PGF_WAITOK) == 0,
480 			    ("failed waitable pte alloc %p", domain));
481 			if (sf != NULL)
482 				dmar_unmap_pgtbl(sf);
483 			domain_unmap_buf_locked(domain, base1, base - base1,
484 			    flags);
485 			TD_PINNED_ASSERT;
486 			return (ENOMEM);
487 		}
488 		dmar_pte_store(&pte->pte, VM_PAGE_TO_PHYS(ma[pi]) | pflags |
489 		    (superpage ? DMAR_PTE_SP : 0));
490 		dmar_flush_pte_to_ram(domain->dmar, pte);
491 		sf_buf_page(sf)->ref_count += 1;
492 	}
493 	if (sf != NULL)
494 		dmar_unmap_pgtbl(sf);
495 	TD_PINNED_ASSERT;
496 	return (0);
497 }
498 
499 static int
500 domain_map_buf(struct iommu_domain *iodom, iommu_gaddr_t base,
501     iommu_gaddr_t size, vm_page_t *ma, uint64_t eflags, int flags)
502 {
503 	struct dmar_domain *domain;
504 	struct dmar_unit *unit;
505 	uint64_t pflags;
506 	int error;
507 
508 	pflags = ((eflags & IOMMU_MAP_ENTRY_READ) != 0 ? DMAR_PTE_R : 0) |
509 	    ((eflags & IOMMU_MAP_ENTRY_WRITE) != 0 ? DMAR_PTE_W : 0) |
510 	    ((eflags & IOMMU_MAP_ENTRY_SNOOP) != 0 ? DMAR_PTE_SNP : 0) |
511 	    ((eflags & IOMMU_MAP_ENTRY_TM) != 0 ? DMAR_PTE_TM : 0);
512 
513 	domain = IODOM2DOM(iodom);
514 	unit = domain->dmar;
515 
516 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) == 0,
517 	    ("modifying idmap pagetable domain %p", domain));
518 	KASSERT((base & DMAR_PAGE_MASK) == 0,
519 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
520 	    (uintmax_t)size));
521 	KASSERT((size & DMAR_PAGE_MASK) == 0,
522 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
523 	    (uintmax_t)size));
524 	KASSERT(size > 0, ("zero size %p %jx %jx", domain, (uintmax_t)base,
525 	    (uintmax_t)size));
526 	KASSERT(base < (1ULL << domain->agaw),
527 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
528 	    (uintmax_t)size, domain->agaw));
529 	KASSERT(base + size < (1ULL << domain->agaw),
530 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
531 	    (uintmax_t)size, domain->agaw));
532 	KASSERT(base + size > base,
533 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
534 	    (uintmax_t)size));
535 	KASSERT((pflags & (DMAR_PTE_R | DMAR_PTE_W)) != 0,
536 	    ("neither read nor write %jx", (uintmax_t)pflags));
537 	KASSERT((pflags & ~(DMAR_PTE_R | DMAR_PTE_W | DMAR_PTE_SNP |
538 	    DMAR_PTE_TM)) == 0,
539 	    ("invalid pte flags %jx", (uintmax_t)pflags));
540 	KASSERT((pflags & DMAR_PTE_SNP) == 0 ||
541 	    (unit->hw_ecap & DMAR_ECAP_SC) != 0,
542 	    ("PTE_SNP for dmar without snoop control %p %jx",
543 	    domain, (uintmax_t)pflags));
544 	KASSERT((pflags & DMAR_PTE_TM) == 0 ||
545 	    (unit->hw_ecap & DMAR_ECAP_DI) != 0,
546 	    ("PTE_TM for dmar without DIOTLB %p %jx",
547 	    domain, (uintmax_t)pflags));
548 	KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags));
549 
550 	DMAR_DOMAIN_PGLOCK(domain);
551 	error = domain_map_buf_locked(domain, base, size, ma, pflags, flags);
552 	DMAR_DOMAIN_PGUNLOCK(domain);
553 	if (error != 0)
554 		return (error);
555 
556 	if ((unit->hw_cap & DMAR_CAP_CM) != 0)
557 		domain_flush_iotlb_sync(domain, base, size);
558 	else if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) {
559 		/* See 11.1 Write Buffer Flushing. */
560 		DMAR_LOCK(unit);
561 		dmar_flush_write_bufs(unit);
562 		DMAR_UNLOCK(unit);
563 	}
564 	return (0);
565 }
566 
567 static void domain_unmap_clear_pte(struct dmar_domain *domain,
568     iommu_gaddr_t base, int lvl, int flags, dmar_pte_t *pte,
569     struct sf_buf **sf, bool free_fs);
570 
571 static void
572 domain_free_pgtbl_pde(struct dmar_domain *domain, iommu_gaddr_t base,
573     int lvl, int flags)
574 {
575 	struct sf_buf *sf;
576 	dmar_pte_t *pde;
577 	vm_pindex_t idx;
578 
579 	sf = NULL;
580 	pde = domain_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf);
581 	domain_unmap_clear_pte(domain, base, lvl, flags, pde, &sf, true);
582 }
583 
584 static void
585 domain_unmap_clear_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl,
586     int flags, dmar_pte_t *pte, struct sf_buf **sf, bool free_sf)
587 {
588 	vm_page_t m;
589 
590 	dmar_pte_clear(&pte->pte);
591 	dmar_flush_pte_to_ram(domain->dmar, pte);
592 	m = sf_buf_page(*sf);
593 	if (free_sf) {
594 		dmar_unmap_pgtbl(*sf);
595 		*sf = NULL;
596 	}
597 	m->ref_count--;
598 	if (m->ref_count != 0)
599 		return;
600 	KASSERT(lvl != 0,
601 	    ("lost reference (lvl) on root pg domain %p base %jx lvl %d",
602 	    domain, (uintmax_t)base, lvl));
603 	KASSERT(m->pindex != 0,
604 	    ("lost reference (idx) on root pg domain %p base %jx lvl %d",
605 	    domain, (uintmax_t)base, lvl));
606 	dmar_pgfree(domain->pgtbl_obj, m->pindex, flags);
607 	domain_free_pgtbl_pde(domain, base, lvl - 1, flags);
608 }
609 
610 /*
611  * Assumes that the unmap is never partial.
612  */
613 static int
614 domain_unmap_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base,
615     iommu_gaddr_t size, int flags)
616 {
617 	dmar_pte_t *pte;
618 	struct sf_buf *sf;
619 	vm_pindex_t idx;
620 	iommu_gaddr_t pg_sz;
621 	int lvl;
622 
623 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
624 	if (size == 0)
625 		return (0);
626 
627 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) == 0,
628 	    ("modifying idmap pagetable domain %p", domain));
629 	KASSERT((base & DMAR_PAGE_MASK) == 0,
630 	    ("non-aligned base %p %jx %jx", domain, (uintmax_t)base,
631 	    (uintmax_t)size));
632 	KASSERT((size & DMAR_PAGE_MASK) == 0,
633 	    ("non-aligned size %p %jx %jx", domain, (uintmax_t)base,
634 	    (uintmax_t)size));
635 	KASSERT(base < (1ULL << domain->agaw),
636 	    ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
637 	    (uintmax_t)size, domain->agaw));
638 	KASSERT(base + size < (1ULL << domain->agaw),
639 	    ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base,
640 	    (uintmax_t)size, domain->agaw));
641 	KASSERT(base + size > base,
642 	    ("size overflow %p %jx %jx", domain, (uintmax_t)base,
643 	    (uintmax_t)size));
644 	KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags));
645 
646 	pg_sz = 0; /* silence gcc */
647 	flags |= IOMMU_PGF_OBJL;
648 	TD_PREP_PINNED_ASSERT;
649 
650 	for (sf = NULL; size > 0; base += pg_sz, size -= pg_sz) {
651 		for (lvl = 0; lvl < domain->pglvl; lvl++) {
652 			if (lvl != domain->pglvl - 1 &&
653 			    !domain_is_sp_lvl(domain, lvl))
654 				continue;
655 			pg_sz = domain_page_size(domain, lvl);
656 			if (pg_sz > size)
657 				continue;
658 			pte = domain_pgtbl_map_pte(domain, base, lvl, flags,
659 			    &idx, &sf);
660 			KASSERT(pte != NULL,
661 			    ("sleeping or page missed %p %jx %d 0x%x",
662 			    domain, (uintmax_t)base, lvl, flags));
663 			if ((pte->pte & DMAR_PTE_SP) != 0 ||
664 			    lvl == domain->pglvl - 1) {
665 				domain_unmap_clear_pte(domain, base, lvl,
666 				    flags, pte, &sf, false);
667 				break;
668 			}
669 		}
670 		KASSERT(size >= pg_sz,
671 		    ("unmapping loop overflow %p %jx %jx %jx", domain,
672 		    (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz));
673 	}
674 	if (sf != NULL)
675 		dmar_unmap_pgtbl(sf);
676 	/*
677 	 * See 11.1 Write Buffer Flushing for an explanation why RWBF
678 	 * can be ignored there.
679 	 */
680 
681 	TD_PINNED_ASSERT;
682 	return (0);
683 }
684 
685 static int
686 domain_unmap_buf(struct iommu_domain *iodom, iommu_gaddr_t base,
687     iommu_gaddr_t size, int flags)
688 {
689 	struct dmar_domain *domain;
690 	int error;
691 
692 	domain = IODOM2DOM(iodom);
693 
694 	DMAR_DOMAIN_PGLOCK(domain);
695 	error = domain_unmap_buf_locked(domain, base, size, flags);
696 	DMAR_DOMAIN_PGUNLOCK(domain);
697 	return (error);
698 }
699 
700 int
701 domain_alloc_pgtbl(struct dmar_domain *domain)
702 {
703 	vm_page_t m;
704 
705 	KASSERT(domain->pgtbl_obj == NULL,
706 	    ("already initialized %p", domain));
707 
708 	domain->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL,
709 	    IDX_TO_OFF(pglvl_max_pages(domain->pglvl)), 0, 0, NULL);
710 	DMAR_DOMAIN_PGLOCK(domain);
711 	m = dmar_pgalloc(domain->pgtbl_obj, 0, IOMMU_PGF_WAITOK |
712 	    IOMMU_PGF_ZERO | IOMMU_PGF_OBJL);
713 	/* No implicit free of the top level page table page. */
714 	m->ref_count = 1;
715 	DMAR_DOMAIN_PGUNLOCK(domain);
716 	DMAR_LOCK(domain->dmar);
717 	domain->iodom.flags |= IOMMU_DOMAIN_PGTBL_INITED;
718 	DMAR_UNLOCK(domain->dmar);
719 	return (0);
720 }
721 
722 void
723 domain_free_pgtbl(struct dmar_domain *domain)
724 {
725 	vm_object_t obj;
726 	vm_page_t m;
727 
728 	obj = domain->pgtbl_obj;
729 	if (obj == NULL) {
730 		KASSERT((domain->dmar->hw_ecap & DMAR_ECAP_PT) != 0 &&
731 		    (domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0,
732 		    ("lost pagetable object domain %p", domain));
733 		return;
734 	}
735 	DMAR_DOMAIN_ASSERT_PGLOCKED(domain);
736 	domain->pgtbl_obj = NULL;
737 
738 	if ((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0) {
739 		put_idmap_pgtbl(obj);
740 		domain->iodom.flags &= ~IOMMU_DOMAIN_IDMAP;
741 		return;
742 	}
743 
744 	/* Obliterate ref_counts */
745 	VM_OBJECT_ASSERT_WLOCKED(obj);
746 	for (m = vm_page_lookup(obj, 0); m != NULL; m = vm_page_next(m))
747 		m->ref_count = 0;
748 	VM_OBJECT_WUNLOCK(obj);
749 	vm_object_deallocate(obj);
750 }
751 
752 static inline uint64_t
753 domain_wait_iotlb_flush(struct dmar_unit *unit, uint64_t wt, int iro)
754 {
755 	uint64_t iotlbr;
756 
757 	dmar_write8(unit, iro + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT |
758 	    DMAR_IOTLB_DR | DMAR_IOTLB_DW | wt);
759 	for (;;) {
760 		iotlbr = dmar_read8(unit, iro + DMAR_IOTLB_REG_OFF);
761 		if ((iotlbr & DMAR_IOTLB_IVT) == 0)
762 			break;
763 		cpu_spinwait();
764 	}
765 	return (iotlbr);
766 }
767 
768 void
769 domain_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base,
770     iommu_gaddr_t size)
771 {
772 	struct dmar_unit *unit;
773 	iommu_gaddr_t isize;
774 	uint64_t iotlbr;
775 	int am, iro;
776 
777 	unit = domain->dmar;
778 	KASSERT(!unit->qi_enabled, ("dmar%d: sync iotlb flush call",
779 	    unit->iommu.unit));
780 	iro = DMAR_ECAP_IRO(unit->hw_ecap) * 16;
781 	DMAR_LOCK(unit);
782 	if ((unit->hw_cap & DMAR_CAP_PSI) == 0 || size > 2 * 1024 * 1024) {
783 		iotlbr = domain_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_DOM |
784 		    DMAR_IOTLB_DID(domain->domain), iro);
785 		KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
786 		    DMAR_IOTLB_IAIG_INVLD,
787 		    ("dmar%d: invalidation failed %jx", unit->iommu.unit,
788 		    (uintmax_t)iotlbr));
789 	} else {
790 		for (; size > 0; base += isize, size -= isize) {
791 			am = calc_am(unit, base, size, &isize);
792 			dmar_write8(unit, iro, base | am);
793 			iotlbr = domain_wait_iotlb_flush(unit,
794 			    DMAR_IOTLB_IIRG_PAGE |
795 			    DMAR_IOTLB_DID(domain->domain), iro);
796 			KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
797 			    DMAR_IOTLB_IAIG_INVLD,
798 			    ("dmar%d: PSI invalidation failed "
799 			    "iotlbr 0x%jx base 0x%jx size 0x%jx am %d",
800 			    unit->iommu.unit, (uintmax_t)iotlbr,
801 			    (uintmax_t)base, (uintmax_t)size, am));
802 			/*
803 			 * Any non-page granularity covers whole guest
804 			 * address space for the domain.
805 			 */
806 			if ((iotlbr & DMAR_IOTLB_IAIG_MASK) !=
807 			    DMAR_IOTLB_IAIG_PAGE)
808 				break;
809 		}
810 	}
811 	DMAR_UNLOCK(unit);
812 }
813 
814 const struct iommu_domain_map_ops dmar_domain_map_ops = {
815 	.map = domain_map_buf,
816 	.unmap = domain_unmap_buf,
817 };
818