xref: /freebsd/sys/x86/iommu/intel_ctx.c (revision 5ffd83dbcc34f10e07f6d3e968ae6365869615f4)
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/limits.h>
43 #include <sys/lock.h>
44 #include <sys/memdesc.h>
45 #include <sys/mutex.h>
46 #include <sys/proc.h>
47 #include <sys/rwlock.h>
48 #include <sys/rman.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 <contrib/dev/acpica/include/acpi.h>
62 #include <contrib/dev/acpica/include/accommon.h>
63 #include <dev/pci/pcireg.h>
64 #include <dev/pci/pcivar.h>
65 #include <machine/atomic.h>
66 #include <machine/bus.h>
67 #include <machine/md_var.h>
68 #include <machine/specialreg.h>
69 #include <x86/include/busdma_impl.h>
70 #include <dev/iommu/busdma_iommu.h>
71 #include <x86/iommu/intel_reg.h>
72 #include <x86/iommu/intel_dmar.h>
73 
74 static MALLOC_DEFINE(M_DMAR_CTX, "dmar_ctx", "Intel DMAR Context");
75 static MALLOC_DEFINE(M_DMAR_DOMAIN, "dmar_dom", "Intel DMAR Domain");
76 
77 static void dmar_domain_unload_task(void *arg, int pending);
78 static void dmar_unref_domain_locked(struct dmar_unit *dmar,
79     struct dmar_domain *domain);
80 static void dmar_domain_destroy(struct dmar_domain *domain);
81 
82 static void
83 dmar_ensure_ctx_page(struct dmar_unit *dmar, int bus)
84 {
85 	struct sf_buf *sf;
86 	dmar_root_entry_t *re;
87 	vm_page_t ctxm;
88 
89 	/*
90 	 * Allocated context page must be linked.
91 	 */
92 	ctxm = dmar_pgalloc(dmar->ctx_obj, 1 + bus, IOMMU_PGF_NOALLOC);
93 	if (ctxm != NULL)
94 		return;
95 
96 	/*
97 	 * Page not present, allocate and link.  Note that other
98 	 * thread might execute this sequence in parallel.  This
99 	 * should be safe, because the context entries written by both
100 	 * threads are equal.
101 	 */
102 	TD_PREP_PINNED_ASSERT;
103 	ctxm = dmar_pgalloc(dmar->ctx_obj, 1 + bus, IOMMU_PGF_ZERO |
104 	    IOMMU_PGF_WAITOK);
105 	re = dmar_map_pgtbl(dmar->ctx_obj, 0, IOMMU_PGF_NOALLOC, &sf);
106 	re += bus;
107 	dmar_pte_store(&re->r1, DMAR_ROOT_R1_P | (DMAR_ROOT_R1_CTP_MASK &
108 	    VM_PAGE_TO_PHYS(ctxm)));
109 	dmar_flush_root_to_ram(dmar, re);
110 	dmar_unmap_pgtbl(sf);
111 	TD_PINNED_ASSERT;
112 }
113 
114 static dmar_ctx_entry_t *
115 dmar_map_ctx_entry(struct dmar_ctx *ctx, struct sf_buf **sfp)
116 {
117 	struct dmar_unit *dmar;
118 	dmar_ctx_entry_t *ctxp;
119 
120 	dmar = (struct dmar_unit *)ctx->context.domain->iommu;
121 
122 	ctxp = dmar_map_pgtbl(dmar->ctx_obj, 1 +
123 	    PCI_RID2BUS(ctx->rid), IOMMU_PGF_NOALLOC | IOMMU_PGF_WAITOK, sfp);
124 	ctxp += ctx->rid & 0xff;
125 	return (ctxp);
126 }
127 
128 static void
129 device_tag_init(struct dmar_ctx *ctx, device_t dev)
130 {
131 	struct dmar_domain *domain;
132 	bus_addr_t maxaddr;
133 
134 	domain = (struct dmar_domain *)ctx->context.domain;
135 	maxaddr = MIN(domain->iodom.end, BUS_SPACE_MAXADDR);
136 	ctx->context.tag->common.ref_count = 1; /* Prevent free */
137 	ctx->context.tag->common.impl = &bus_dma_iommu_impl;
138 	ctx->context.tag->common.boundary = 0;
139 	ctx->context.tag->common.lowaddr = maxaddr;
140 	ctx->context.tag->common.highaddr = maxaddr;
141 	ctx->context.tag->common.maxsize = maxaddr;
142 	ctx->context.tag->common.nsegments = BUS_SPACE_UNRESTRICTED;
143 	ctx->context.tag->common.maxsegsz = maxaddr;
144 	ctx->context.tag->ctx = (struct iommu_ctx *)ctx;
145 	ctx->context.tag->owner = dev;
146 }
147 
148 static void
149 ctx_id_entry_init_one(dmar_ctx_entry_t *ctxp, struct dmar_domain *domain,
150     vm_page_t ctx_root)
151 {
152 	/*
153 	 * For update due to move, the store is not atomic.  It is
154 	 * possible that DMAR read upper doubleword, while low
155 	 * doubleword is not yet updated.  The domain id is stored in
156 	 * the upper doubleword, while the table pointer in the lower.
157 	 *
158 	 * There is no good solution, for the same reason it is wrong
159 	 * to clear P bit in the ctx entry for update.
160 	 */
161 	dmar_pte_store1(&ctxp->ctx2, DMAR_CTX2_DID(domain->domain) |
162 	    domain->awlvl);
163 	if (ctx_root == NULL) {
164 		dmar_pte_store1(&ctxp->ctx1, DMAR_CTX1_T_PASS | DMAR_CTX1_P);
165 	} else {
166 		dmar_pte_store1(&ctxp->ctx1, DMAR_CTX1_T_UNTR |
167 		    (DMAR_CTX1_ASR_MASK & VM_PAGE_TO_PHYS(ctx_root)) |
168 		    DMAR_CTX1_P);
169 	}
170 }
171 
172 static void
173 ctx_id_entry_init(struct dmar_ctx *ctx, dmar_ctx_entry_t *ctxp, bool move,
174     int busno)
175 {
176 	struct dmar_unit *unit;
177 	struct dmar_domain *domain;
178 	vm_page_t ctx_root;
179 	int i;
180 
181 	domain = (struct dmar_domain *)ctx->context.domain;
182 	unit = (struct dmar_unit *)domain->iodom.iommu;
183 	KASSERT(move || (ctxp->ctx1 == 0 && ctxp->ctx2 == 0),
184 	    ("dmar%d: initialized ctx entry %d:%d:%d 0x%jx 0x%jx",
185 	    unit->iommu.unit, busno, pci_get_slot(ctx->context.tag->owner),
186 	    pci_get_function(ctx->context.tag->owner),
187 	    ctxp->ctx1, ctxp->ctx2));
188 
189 	if ((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0 &&
190 	    (unit->hw_ecap & DMAR_ECAP_PT) != 0) {
191 		KASSERT(domain->pgtbl_obj == NULL,
192 		    ("ctx %p non-null pgtbl_obj", ctx));
193 		ctx_root = NULL;
194 	} else {
195 		ctx_root = dmar_pgalloc(domain->pgtbl_obj, 0,
196 		    IOMMU_PGF_NOALLOC);
197 	}
198 
199 	if (iommu_is_buswide_ctx((struct iommu_unit *)unit, busno)) {
200 		MPASS(!move);
201 		for (i = 0; i <= PCI_BUSMAX; i++) {
202 			ctx_id_entry_init_one(&ctxp[i], domain, ctx_root);
203 		}
204 	} else {
205 		ctx_id_entry_init_one(ctxp, domain, ctx_root);
206 	}
207 	dmar_flush_ctx_to_ram(unit, ctxp);
208 }
209 
210 static int
211 dmar_flush_for_ctx_entry(struct dmar_unit *dmar, bool force)
212 {
213 	int error;
214 
215 	/*
216 	 * If dmar declares Caching Mode as Set, follow 11.5 "Caching
217 	 * Mode Consideration" and do the (global) invalidation of the
218 	 * negative TLB entries.
219 	 */
220 	if ((dmar->hw_cap & DMAR_CAP_CM) == 0 && !force)
221 		return (0);
222 	if (dmar->qi_enabled) {
223 		dmar_qi_invalidate_ctx_glob_locked(dmar);
224 		if ((dmar->hw_ecap & DMAR_ECAP_DI) != 0 || force)
225 			dmar_qi_invalidate_iotlb_glob_locked(dmar);
226 		return (0);
227 	}
228 	error = dmar_inv_ctx_glob(dmar);
229 	if (error == 0 && ((dmar->hw_ecap & DMAR_ECAP_DI) != 0 || force))
230 		error = dmar_inv_iotlb_glob(dmar);
231 	return (error);
232 }
233 
234 static int
235 domain_init_rmrr(struct dmar_domain *domain, device_t dev, int bus,
236     int slot, int func, int dev_domain, int dev_busno,
237     const void *dev_path, int dev_path_len)
238 {
239 	struct iommu_map_entries_tailq rmrr_entries;
240 	struct iommu_map_entry *entry, *entry1;
241 	vm_page_t *ma;
242 	iommu_gaddr_t start, end;
243 	vm_pindex_t size, i;
244 	int error, error1;
245 
246 	error = 0;
247 	TAILQ_INIT(&rmrr_entries);
248 	dmar_dev_parse_rmrr(domain, dev_domain, dev_busno, dev_path,
249 	    dev_path_len, &rmrr_entries);
250 	TAILQ_FOREACH_SAFE(entry, &rmrr_entries, unroll_link, entry1) {
251 		/*
252 		 * VT-d specification requires that the start of an
253 		 * RMRR entry is 4k-aligned.  Buggy BIOSes put
254 		 * anything into the start and end fields.  Truncate
255 		 * and round as neccesary.
256 		 *
257 		 * We also allow the overlapping RMRR entries, see
258 		 * iommu_gas_alloc_region().
259 		 */
260 		start = entry->start;
261 		end = entry->end;
262 		if (bootverbose)
263 			printf("dmar%d ctx pci%d:%d:%d RMRR [%#jx, %#jx]\n",
264 			    domain->iodom.iommu->unit, bus, slot, func,
265 			    (uintmax_t)start, (uintmax_t)end);
266 		entry->start = trunc_page(start);
267 		entry->end = round_page(end);
268 		if (entry->start == entry->end) {
269 			/* Workaround for some AMI (?) BIOSes */
270 			if (bootverbose) {
271 				if (dev != NULL)
272 					device_printf(dev, "");
273 				printf("pci%d:%d:%d ", bus, slot, func);
274 				printf("BIOS bug: dmar%d RMRR "
275 				    "region (%jx, %jx) corrected\n",
276 				    domain->iodom.iommu->unit, start, end);
277 			}
278 			entry->end += DMAR_PAGE_SIZE * 0x20;
279 		}
280 		size = OFF_TO_IDX(entry->end - entry->start);
281 		ma = malloc(sizeof(vm_page_t) * size, M_TEMP, M_WAITOK);
282 		for (i = 0; i < size; i++) {
283 			ma[i] = vm_page_getfake(entry->start + PAGE_SIZE * i,
284 			    VM_MEMATTR_DEFAULT);
285 		}
286 		error1 = iommu_gas_map_region((struct iommu_domain *)domain,
287 		    entry,
288 		    IOMMU_MAP_ENTRY_READ | IOMMU_MAP_ENTRY_WRITE,
289 		    IOMMU_MF_CANWAIT | IOMMU_MF_RMRR, ma);
290 		/*
291 		 * Non-failed RMRR entries are owned by context rb
292 		 * tree.  Get rid of the failed entry, but do not stop
293 		 * the loop.  Rest of the parsed RMRR entries are
294 		 * loaded and removed on the context destruction.
295 		 */
296 		if (error1 == 0 && entry->end != entry->start) {
297 			IOMMU_LOCK(domain->iodom.iommu);
298 			domain->refs++; /* XXXKIB prevent free */
299 			domain->iodom.flags |= IOMMU_DOMAIN_RMRR;
300 			IOMMU_UNLOCK(domain->iodom.iommu);
301 		} else {
302 			if (error1 != 0) {
303 				if (dev != NULL)
304 					device_printf(dev, "");
305 				printf("pci%d:%d:%d ", bus, slot, func);
306 				printf(
307 			    "dmar%d failed to map RMRR region (%jx, %jx) %d\n",
308 				    domain->iodom.iommu->unit, start, end,
309 				    error1);
310 				error = error1;
311 			}
312 			TAILQ_REMOVE(&rmrr_entries, entry, unroll_link);
313 			iommu_gas_free_entry((struct iommu_domain *)domain,
314 			    entry);
315 		}
316 		for (i = 0; i < size; i++)
317 			vm_page_putfake(ma[i]);
318 		free(ma, M_TEMP);
319 	}
320 	return (error);
321 }
322 
323 static struct dmar_domain *
324 dmar_domain_alloc(struct dmar_unit *dmar, bool id_mapped)
325 {
326 	struct iommu_domain *iodom;
327 	struct dmar_domain *domain;
328 	int error, id, mgaw;
329 
330 	id = alloc_unr(dmar->domids);
331 	if (id == -1)
332 		return (NULL);
333 	domain = malloc(sizeof(*domain), M_DMAR_DOMAIN, M_WAITOK | M_ZERO);
334 	iodom = (struct iommu_domain *)domain;
335 	domain->domain = id;
336 	LIST_INIT(&domain->contexts);
337 	RB_INIT(&domain->iodom.rb_root);
338 	TAILQ_INIT(&domain->iodom.unload_entries);
339 	TASK_INIT(&domain->iodom.unload_task, 0, dmar_domain_unload_task,
340 	    domain);
341 	mtx_init(&domain->iodom.lock, "dmardom", NULL, MTX_DEF);
342 	domain->dmar = dmar;
343 	domain->iodom.iommu = &dmar->iommu;
344 
345 	/*
346 	 * For now, use the maximal usable physical address of the
347 	 * installed memory to calculate the mgaw on id_mapped domain.
348 	 * It is useful for the identity mapping, and less so for the
349 	 * virtualized bus address space.
350 	 */
351 	domain->iodom.end = id_mapped ? ptoa(Maxmem) : BUS_SPACE_MAXADDR;
352 	mgaw = dmar_maxaddr2mgaw(dmar, domain->iodom.end, !id_mapped);
353 	error = domain_set_agaw(domain, mgaw);
354 	if (error != 0)
355 		goto fail;
356 	if (!id_mapped)
357 		/* Use all supported address space for remapping. */
358 		domain->iodom.end = 1ULL << (domain->agaw - 1);
359 
360 	iommu_gas_init_domain((struct iommu_domain *)domain);
361 
362 	if (id_mapped) {
363 		if ((dmar->hw_ecap & DMAR_ECAP_PT) == 0) {
364 			domain->pgtbl_obj = domain_get_idmap_pgtbl(domain,
365 			    domain->iodom.end);
366 		}
367 		domain->iodom.flags |= IOMMU_DOMAIN_IDMAP;
368 	} else {
369 		error = domain_alloc_pgtbl(domain);
370 		if (error != 0)
371 			goto fail;
372 		/* Disable local apic region access */
373 		error = iommu_gas_reserve_region(iodom, 0xfee00000,
374 		    0xfeefffff + 1);
375 		if (error != 0)
376 			goto fail;
377 	}
378 	return (domain);
379 
380 fail:
381 	dmar_domain_destroy(domain);
382 	return (NULL);
383 }
384 
385 static struct dmar_ctx *
386 dmar_ctx_alloc(struct dmar_domain *domain, uint16_t rid)
387 {
388 	struct dmar_ctx *ctx;
389 
390 	ctx = malloc(sizeof(*ctx), M_DMAR_CTX, M_WAITOK | M_ZERO);
391 	ctx->context.domain = (struct iommu_domain *)domain;
392 	ctx->context.tag = malloc(sizeof(struct bus_dma_tag_iommu),
393 	    M_DMAR_CTX, M_WAITOK | M_ZERO);
394 	ctx->rid = rid;
395 	ctx->refs = 1;
396 	return (ctx);
397 }
398 
399 static void
400 dmar_ctx_link(struct dmar_ctx *ctx)
401 {
402 	struct dmar_domain *domain;
403 
404 	domain = (struct dmar_domain *)ctx->context.domain;
405 	IOMMU_ASSERT_LOCKED(domain->iodom.iommu);
406 	KASSERT(domain->refs >= domain->ctx_cnt,
407 	    ("dom %p ref underflow %d %d", domain, domain->refs,
408 	    domain->ctx_cnt));
409 	domain->refs++;
410 	domain->ctx_cnt++;
411 	LIST_INSERT_HEAD(&domain->contexts, ctx, link);
412 }
413 
414 static void
415 dmar_ctx_unlink(struct dmar_ctx *ctx)
416 {
417 	struct dmar_domain *domain;
418 
419 	domain = (struct dmar_domain *)ctx->context.domain;
420 	IOMMU_ASSERT_LOCKED(domain->iodom.iommu);
421 	KASSERT(domain->refs > 0,
422 	    ("domain %p ctx dtr refs %d", domain, domain->refs));
423 	KASSERT(domain->ctx_cnt >= domain->refs,
424 	    ("domain %p ctx dtr refs %d ctx_cnt %d", domain,
425 	    domain->refs, domain->ctx_cnt));
426 	domain->refs--;
427 	domain->ctx_cnt--;
428 	LIST_REMOVE(ctx, link);
429 }
430 
431 static void
432 dmar_domain_destroy(struct dmar_domain *domain)
433 {
434 	struct dmar_unit *dmar;
435 
436 	KASSERT(TAILQ_EMPTY(&domain->iodom.unload_entries),
437 	    ("unfinished unloads %p", domain));
438 	KASSERT(LIST_EMPTY(&domain->contexts),
439 	    ("destroying dom %p with contexts", domain));
440 	KASSERT(domain->ctx_cnt == 0,
441 	    ("destroying dom %p with ctx_cnt %d", domain, domain->ctx_cnt));
442 	KASSERT(domain->refs == 0,
443 	    ("destroying dom %p with refs %d", domain, domain->refs));
444 	if ((domain->iodom.flags & IOMMU_DOMAIN_GAS_INITED) != 0) {
445 		DMAR_DOMAIN_LOCK(domain);
446 		iommu_gas_fini_domain((struct iommu_domain *)domain);
447 		DMAR_DOMAIN_UNLOCK(domain);
448 	}
449 	if ((domain->iodom.flags & IOMMU_DOMAIN_PGTBL_INITED) != 0) {
450 		if (domain->pgtbl_obj != NULL)
451 			DMAR_DOMAIN_PGLOCK(domain);
452 		domain_free_pgtbl(domain);
453 	}
454 	mtx_destroy(&domain->iodom.lock);
455 	dmar = (struct dmar_unit *)domain->iodom.iommu;
456 	free_unr(dmar->domids, domain->domain);
457 	free(domain, M_DMAR_DOMAIN);
458 }
459 
460 static struct dmar_ctx *
461 dmar_get_ctx_for_dev1(struct dmar_unit *dmar, device_t dev, uint16_t rid,
462     int dev_domain, int dev_busno, const void *dev_path, int dev_path_len,
463     bool id_mapped, bool rmrr_init)
464 {
465 	struct dmar_domain *domain, *domain1;
466 	struct dmar_ctx *ctx, *ctx1;
467 	struct iommu_unit *unit;
468 	dmar_ctx_entry_t *ctxp;
469 	struct sf_buf *sf;
470 	int bus, slot, func, error;
471 	bool enable;
472 
473 	if (dev != NULL) {
474 		bus = pci_get_bus(dev);
475 		slot = pci_get_slot(dev);
476 		func = pci_get_function(dev);
477 	} else {
478 		bus = PCI_RID2BUS(rid);
479 		slot = PCI_RID2SLOT(rid);
480 		func = PCI_RID2FUNC(rid);
481 	}
482 	enable = false;
483 	TD_PREP_PINNED_ASSERT;
484 	unit = (struct iommu_unit *)dmar;
485 	DMAR_LOCK(dmar);
486 	KASSERT(!iommu_is_buswide_ctx(unit, bus) || (slot == 0 && func == 0),
487 	    ("iommu%d pci%d:%d:%d get_ctx for buswide", dmar->iommu.unit, bus,
488 	    slot, func));
489 	ctx = dmar_find_ctx_locked(dmar, rid);
490 	error = 0;
491 	if (ctx == NULL) {
492 		/*
493 		 * Perform the allocations which require sleep or have
494 		 * higher chance to succeed if the sleep is allowed.
495 		 */
496 		DMAR_UNLOCK(dmar);
497 		dmar_ensure_ctx_page(dmar, PCI_RID2BUS(rid));
498 		domain1 = dmar_domain_alloc(dmar, id_mapped);
499 		if (domain1 == NULL) {
500 			TD_PINNED_ASSERT;
501 			return (NULL);
502 		}
503 		if (!id_mapped) {
504 			error = domain_init_rmrr(domain1, dev, bus,
505 			    slot, func, dev_domain, dev_busno, dev_path,
506 			    dev_path_len);
507 			if (error != 0) {
508 				dmar_domain_destroy(domain1);
509 				TD_PINNED_ASSERT;
510 				return (NULL);
511 			}
512 		}
513 		ctx1 = dmar_ctx_alloc(domain1, rid);
514 		ctxp = dmar_map_ctx_entry(ctx1, &sf);
515 		DMAR_LOCK(dmar);
516 
517 		/*
518 		 * Recheck the contexts, other thread might have
519 		 * already allocated needed one.
520 		 */
521 		ctx = dmar_find_ctx_locked(dmar, rid);
522 		if (ctx == NULL) {
523 			domain = domain1;
524 			ctx = ctx1;
525 			dmar_ctx_link(ctx);
526 			ctx->context.tag->owner = dev;
527 			device_tag_init(ctx, dev);
528 
529 			/*
530 			 * This is the first activated context for the
531 			 * DMAR unit.  Enable the translation after
532 			 * everything is set up.
533 			 */
534 			if (LIST_EMPTY(&dmar->domains))
535 				enable = true;
536 			LIST_INSERT_HEAD(&dmar->domains, domain, link);
537 			ctx_id_entry_init(ctx, ctxp, false, bus);
538 			if (dev != NULL) {
539 				device_printf(dev,
540 			    "dmar%d pci%d:%d:%d:%d rid %x domain %d mgaw %d "
541 				    "agaw %d %s-mapped\n",
542 				    dmar->iommu.unit, dmar->segment, bus, slot,
543 				    func, rid, domain->domain, domain->mgaw,
544 				    domain->agaw, id_mapped ? "id" : "re");
545 			}
546 			dmar_unmap_pgtbl(sf);
547 		} else {
548 			dmar_unmap_pgtbl(sf);
549 			dmar_domain_destroy(domain1);
550 			/* Nothing needs to be done to destroy ctx1. */
551 			free(ctx1, M_DMAR_CTX);
552 			domain = (struct dmar_domain *)ctx->context.domain;
553 			ctx->refs++; /* tag referenced us */
554 		}
555 	} else {
556 		domain = (struct dmar_domain *)ctx->context.domain;
557 		if (ctx->context.tag->owner == NULL)
558 			ctx->context.tag->owner = dev;
559 		ctx->refs++; /* tag referenced us */
560 	}
561 
562 	error = dmar_flush_for_ctx_entry(dmar, enable);
563 	if (error != 0) {
564 		dmar_free_ctx_locked(dmar, ctx);
565 		TD_PINNED_ASSERT;
566 		return (NULL);
567 	}
568 
569 	/*
570 	 * The dmar lock was potentially dropped between check for the
571 	 * empty context list and now.  Recheck the state of GCMD_TE
572 	 * to avoid unneeded command.
573 	 */
574 	if (enable && !rmrr_init && (dmar->hw_gcmd & DMAR_GCMD_TE) == 0) {
575 		error = dmar_enable_translation(dmar);
576 		if (error == 0) {
577 			if (bootverbose) {
578 				printf("dmar%d: enabled translation\n",
579 				    dmar->iommu.unit);
580 			}
581 		} else {
582 			printf("dmar%d: enabling translation failed, "
583 			    "error %d\n", dmar->iommu.unit, error);
584 			dmar_free_ctx_locked(dmar, ctx);
585 			TD_PINNED_ASSERT;
586 			return (NULL);
587 		}
588 	}
589 	DMAR_UNLOCK(dmar);
590 	TD_PINNED_ASSERT;
591 	return (ctx);
592 }
593 
594 struct dmar_ctx *
595 dmar_get_ctx_for_dev(struct dmar_unit *dmar, device_t dev, uint16_t rid,
596     bool id_mapped, bool rmrr_init)
597 {
598 	int dev_domain, dev_path_len, dev_busno;
599 
600 	dev_domain = pci_get_domain(dev);
601 	dev_path_len = dmar_dev_depth(dev);
602 	ACPI_DMAR_PCI_PATH dev_path[dev_path_len];
603 	dmar_dev_path(dev, &dev_busno, dev_path, dev_path_len);
604 	return (dmar_get_ctx_for_dev1(dmar, dev, rid, dev_domain, dev_busno,
605 	    dev_path, dev_path_len, id_mapped, rmrr_init));
606 }
607 
608 struct dmar_ctx *
609 dmar_get_ctx_for_devpath(struct dmar_unit *dmar, uint16_t rid,
610     int dev_domain, int dev_busno,
611     const void *dev_path, int dev_path_len,
612     bool id_mapped, bool rmrr_init)
613 {
614 
615 	return (dmar_get_ctx_for_dev1(dmar, NULL, rid, dev_domain, dev_busno,
616 	    dev_path, dev_path_len, id_mapped, rmrr_init));
617 }
618 
619 int
620 dmar_move_ctx_to_domain(struct dmar_domain *domain, struct dmar_ctx *ctx)
621 {
622 	struct dmar_unit *dmar;
623 	struct dmar_domain *old_domain;
624 	dmar_ctx_entry_t *ctxp;
625 	struct sf_buf *sf;
626 	int error;
627 
628 	dmar = domain->dmar;
629 	old_domain = (struct dmar_domain *)ctx->context.domain;
630 	if (domain == old_domain)
631 		return (0);
632 	KASSERT(old_domain->iodom.iommu == domain->iodom.iommu,
633 	    ("domain %p %u moving between dmars %u %u", domain,
634 	    domain->domain, old_domain->iodom.iommu->unit,
635 	    domain->iodom.iommu->unit));
636 	TD_PREP_PINNED_ASSERT;
637 
638 	ctxp = dmar_map_ctx_entry(ctx, &sf);
639 	DMAR_LOCK(dmar);
640 	dmar_ctx_unlink(ctx);
641 	ctx->context.domain = &domain->iodom;
642 	dmar_ctx_link(ctx);
643 	ctx_id_entry_init(ctx, ctxp, true, PCI_BUSMAX + 100);
644 	dmar_unmap_pgtbl(sf);
645 	error = dmar_flush_for_ctx_entry(dmar, true);
646 	/* If flush failed, rolling back would not work as well. */
647 	printf("dmar%d rid %x domain %d->%d %s-mapped\n",
648 	    dmar->iommu.unit, ctx->rid, old_domain->domain, domain->domain,
649 	    (domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0 ? "id" : "re");
650 	dmar_unref_domain_locked(dmar, old_domain);
651 	TD_PINNED_ASSERT;
652 	return (error);
653 }
654 
655 static void
656 dmar_unref_domain_locked(struct dmar_unit *dmar, struct dmar_domain *domain)
657 {
658 
659 	DMAR_ASSERT_LOCKED(dmar);
660 	KASSERT(domain->refs >= 1,
661 	    ("dmar %d domain %p refs %u", dmar->iommu.unit, domain,
662 	    domain->refs));
663 	KASSERT(domain->refs > domain->ctx_cnt,
664 	    ("dmar %d domain %p refs %d ctx_cnt %d", dmar->iommu.unit, domain,
665 	    domain->refs, domain->ctx_cnt));
666 
667 	if (domain->refs > 1) {
668 		domain->refs--;
669 		DMAR_UNLOCK(dmar);
670 		return;
671 	}
672 
673 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_RMRR) == 0,
674 	    ("lost ref on RMRR domain %p", domain));
675 
676 	LIST_REMOVE(domain, link);
677 	DMAR_UNLOCK(dmar);
678 
679 	taskqueue_drain(dmar->iommu.delayed_taskqueue,
680 	    &domain->iodom.unload_task);
681 	dmar_domain_destroy(domain);
682 }
683 
684 void
685 dmar_free_ctx_locked(struct dmar_unit *dmar, struct dmar_ctx *ctx)
686 {
687 	struct sf_buf *sf;
688 	dmar_ctx_entry_t *ctxp;
689 	struct dmar_domain *domain;
690 
691 	DMAR_ASSERT_LOCKED(dmar);
692 	KASSERT(ctx->refs >= 1,
693 	    ("dmar %p ctx %p refs %u", dmar, ctx, ctx->refs));
694 
695 	/*
696 	 * If our reference is not last, only the dereference should
697 	 * be performed.
698 	 */
699 	if (ctx->refs > 1) {
700 		ctx->refs--;
701 		DMAR_UNLOCK(dmar);
702 		return;
703 	}
704 
705 	KASSERT((ctx->context.flags & IOMMU_CTX_DISABLED) == 0,
706 	    ("lost ref on disabled ctx %p", ctx));
707 
708 	/*
709 	 * Otherwise, the context entry must be cleared before the
710 	 * page table is destroyed.  The mapping of the context
711 	 * entries page could require sleep, unlock the dmar.
712 	 */
713 	DMAR_UNLOCK(dmar);
714 	TD_PREP_PINNED_ASSERT;
715 	ctxp = dmar_map_ctx_entry(ctx, &sf);
716 	DMAR_LOCK(dmar);
717 	KASSERT(ctx->refs >= 1,
718 	    ("dmar %p ctx %p refs %u", dmar, ctx, ctx->refs));
719 
720 	/*
721 	 * Other thread might have referenced the context, in which
722 	 * case again only the dereference should be performed.
723 	 */
724 	if (ctx->refs > 1) {
725 		ctx->refs--;
726 		DMAR_UNLOCK(dmar);
727 		dmar_unmap_pgtbl(sf);
728 		TD_PINNED_ASSERT;
729 		return;
730 	}
731 
732 	KASSERT((ctx->context.flags & IOMMU_CTX_DISABLED) == 0,
733 	    ("lost ref on disabled ctx %p", ctx));
734 
735 	/*
736 	 * Clear the context pointer and flush the caches.
737 	 * XXXKIB: cannot do this if any RMRR entries are still present.
738 	 */
739 	dmar_pte_clear(&ctxp->ctx1);
740 	ctxp->ctx2 = 0;
741 	dmar_flush_ctx_to_ram(dmar, ctxp);
742 	dmar_inv_ctx_glob(dmar);
743 	if ((dmar->hw_ecap & DMAR_ECAP_DI) != 0) {
744 		if (dmar->qi_enabled)
745 			dmar_qi_invalidate_iotlb_glob_locked(dmar);
746 		else
747 			dmar_inv_iotlb_glob(dmar);
748 	}
749 	dmar_unmap_pgtbl(sf);
750 	domain = (struct dmar_domain *)ctx->context.domain;
751 	dmar_ctx_unlink(ctx);
752 	free(ctx->context.tag, M_DMAR_CTX);
753 	free(ctx, M_DMAR_CTX);
754 	dmar_unref_domain_locked(dmar, domain);
755 	TD_PINNED_ASSERT;
756 }
757 
758 void
759 dmar_free_ctx(struct dmar_ctx *ctx)
760 {
761 	struct dmar_unit *dmar;
762 
763 	dmar = (struct dmar_unit *)ctx->context.domain->iommu;
764 	DMAR_LOCK(dmar);
765 	dmar_free_ctx_locked(dmar, ctx);
766 }
767 
768 /*
769  * Returns with the domain locked.
770  */
771 struct dmar_ctx *
772 dmar_find_ctx_locked(struct dmar_unit *dmar, uint16_t rid)
773 {
774 	struct dmar_domain *domain;
775 	struct dmar_ctx *ctx;
776 
777 	DMAR_ASSERT_LOCKED(dmar);
778 
779 	LIST_FOREACH(domain, &dmar->domains, link) {
780 		LIST_FOREACH(ctx, &domain->contexts, link) {
781 			if (ctx->rid == rid)
782 				return (ctx);
783 		}
784 	}
785 	return (NULL);
786 }
787 
788 void
789 dmar_domain_free_entry(struct iommu_map_entry *entry, bool free)
790 {
791 	struct iommu_domain *domain;
792 
793 	domain = entry->domain;
794 	IOMMU_DOMAIN_LOCK(domain);
795 	if ((entry->flags & IOMMU_MAP_ENTRY_RMRR) != 0)
796 		iommu_gas_free_region(domain, entry);
797 	else
798 		iommu_gas_free_space(domain, entry);
799 	IOMMU_DOMAIN_UNLOCK(domain);
800 	if (free)
801 		iommu_gas_free_entry(domain, entry);
802 	else
803 		entry->flags = 0;
804 }
805 
806 void
807 dmar_domain_unload_entry(struct iommu_map_entry *entry, bool free)
808 {
809 	struct dmar_domain *domain;
810 	struct dmar_unit *unit;
811 
812 	domain = (struct dmar_domain *)entry->domain;
813 	unit = (struct dmar_unit *)domain->iodom.iommu;
814 	if (unit->qi_enabled) {
815 		DMAR_LOCK(unit);
816 		dmar_qi_invalidate_locked((struct dmar_domain *)entry->domain,
817 		    entry->start, entry->end - entry->start, &entry->gseq,
818 		    true);
819 		if (!free)
820 			entry->flags |= IOMMU_MAP_ENTRY_QI_NF;
821 		TAILQ_INSERT_TAIL(&unit->tlb_flush_entries, entry, dmamap_link);
822 		DMAR_UNLOCK(unit);
823 	} else {
824 		domain_flush_iotlb_sync((struct dmar_domain *)entry->domain,
825 		    entry->start, entry->end - entry->start);
826 		dmar_domain_free_entry(entry, free);
827 	}
828 }
829 
830 static bool
831 dmar_domain_unload_emit_wait(struct dmar_domain *domain,
832     struct iommu_map_entry *entry)
833 {
834 
835 	if (TAILQ_NEXT(entry, dmamap_link) == NULL)
836 		return (true);
837 	return (domain->batch_no++ % dmar_batch_coalesce == 0);
838 }
839 
840 void
841 dmar_domain_unload(struct dmar_domain *domain,
842     struct iommu_map_entries_tailq *entries, bool cansleep)
843 {
844 	struct dmar_unit *unit;
845 	struct iommu_map_entry *entry, *entry1;
846 	int error;
847 
848 	unit = (struct dmar_unit *)domain->iodom.iommu;
849 
850 	TAILQ_FOREACH_SAFE(entry, entries, dmamap_link, entry1) {
851 		KASSERT((entry->flags & IOMMU_MAP_ENTRY_MAP) != 0,
852 		    ("not mapped entry %p %p", domain, entry));
853 		error = domain_unmap_buf(domain, entry->start, entry->end -
854 		    entry->start, cansleep ? IOMMU_PGF_WAITOK : 0);
855 		KASSERT(error == 0, ("unmap %p error %d", domain, error));
856 		if (!unit->qi_enabled) {
857 			domain_flush_iotlb_sync(domain, entry->start,
858 			    entry->end - entry->start);
859 			TAILQ_REMOVE(entries, entry, dmamap_link);
860 			dmar_domain_free_entry(entry, true);
861 		}
862 	}
863 	if (TAILQ_EMPTY(entries))
864 		return;
865 
866 	KASSERT(unit->qi_enabled, ("loaded entry left"));
867 	DMAR_LOCK(unit);
868 	TAILQ_FOREACH(entry, entries, dmamap_link) {
869 		dmar_qi_invalidate_locked(domain, entry->start, entry->end -
870 		    entry->start, &entry->gseq,
871 		    dmar_domain_unload_emit_wait(domain, entry));
872 	}
873 	TAILQ_CONCAT(&unit->tlb_flush_entries, entries, dmamap_link);
874 	DMAR_UNLOCK(unit);
875 }
876 
877 static void
878 dmar_domain_unload_task(void *arg, int pending)
879 {
880 	struct dmar_domain *domain;
881 	struct iommu_map_entries_tailq entries;
882 
883 	domain = arg;
884 	TAILQ_INIT(&entries);
885 
886 	for (;;) {
887 		DMAR_DOMAIN_LOCK(domain);
888 		TAILQ_SWAP(&domain->iodom.unload_entries, &entries,
889 		    iommu_map_entry, dmamap_link);
890 		DMAR_DOMAIN_UNLOCK(domain);
891 		if (TAILQ_EMPTY(&entries))
892 			break;
893 		dmar_domain_unload(domain, &entries, true);
894 	}
895 }
896 
897 struct iommu_ctx *
898 iommu_get_ctx(struct iommu_unit *iommu, device_t dev, uint16_t rid,
899     bool id_mapped, bool rmrr_init)
900 {
901 	struct dmar_unit *dmar;
902 	struct dmar_ctx *ret;
903 
904 	dmar = (struct dmar_unit *)iommu;
905 
906 	ret = dmar_get_ctx_for_dev(dmar, dev, rid, id_mapped, rmrr_init);
907 
908 	return ((struct iommu_ctx *)ret);
909 }
910 
911 void
912 iommu_free_ctx_locked(struct iommu_unit *iommu, struct iommu_ctx *context)
913 {
914 	struct dmar_unit *dmar;
915 	struct dmar_ctx *ctx;
916 
917 	dmar = (struct dmar_unit *)iommu;
918 	ctx = (struct dmar_ctx *)context;
919 
920 	dmar_free_ctx_locked(dmar, ctx);
921 }
922 
923 void
924 iommu_free_ctx(struct iommu_ctx *context)
925 {
926 	struct dmar_unit *dmar;
927 	struct dmar_ctx *ctx;
928 
929 	ctx = (struct dmar_ctx *)context;
930 	dmar = (struct dmar_unit *)ctx->context.domain->iommu;
931 
932 	dmar_free_ctx(ctx);
933 }
934 
935 void
936 iommu_domain_unload_entry(struct iommu_map_entry *entry, bool free)
937 {
938 
939 	dmar_domain_unload_entry(entry, free);
940 }
941 
942 void
943 iommu_domain_unload(struct iommu_domain *iodom,
944     struct iommu_map_entries_tailq *entries, bool cansleep)
945 {
946 	struct dmar_domain *domain;
947 
948 	domain = (struct dmar_domain *)iodom;
949 
950 	dmar_domain_unload(domain, entries, cansleep);
951 }
952