xref: /freebsd/sys/x86/iommu/intel_ctx.c (revision d0b2dbfa0ecf2bbc9709efc5e20baf8e4b44bbbf)
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/cdefs.h>
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/malloc.h>
35 #include <sys/bus.h>
36 #include <sys/interrupt.h>
37 #include <sys/kernel.h>
38 #include <sys/ktr.h>
39 #include <sys/limits.h>
40 #include <sys/lock.h>
41 #include <sys/memdesc.h>
42 #include <sys/mutex.h>
43 #include <sys/proc.h>
44 #include <sys/rwlock.h>
45 #include <sys/rman.h>
46 #include <sys/sysctl.h>
47 #include <sys/taskqueue.h>
48 #include <sys/tree.h>
49 #include <sys/uio.h>
50 #include <sys/vmem.h>
51 #include <vm/vm.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_kern.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_pager.h>
57 #include <vm/vm_map.h>
58 #include <contrib/dev/acpica/include/acpi.h>
59 #include <contrib/dev/acpica/include/accommon.h>
60 #include <dev/pci/pcireg.h>
61 #include <dev/pci/pcivar.h>
62 #include <machine/atomic.h>
63 #include <machine/bus.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 MALLOC_DEFINE(M_DMAR_CTX, "dmar_ctx", "Intel DMAR Context");
72 static MALLOC_DEFINE(M_DMAR_DOMAIN, "dmar_dom", "Intel DMAR Domain");
73 
74 static void dmar_unref_domain_locked(struct dmar_unit *dmar,
75     struct dmar_domain *domain);
76 static void dmar_domain_destroy(struct dmar_domain *domain);
77 
78 static void
79 dmar_ensure_ctx_page(struct dmar_unit *dmar, int bus)
80 {
81 	struct sf_buf *sf;
82 	dmar_root_entry_t *re;
83 	vm_page_t ctxm;
84 
85 	/*
86 	 * Allocated context page must be linked.
87 	 */
88 	ctxm = dmar_pgalloc(dmar->ctx_obj, 1 + bus, IOMMU_PGF_NOALLOC);
89 	if (ctxm != NULL)
90 		return;
91 
92 	/*
93 	 * Page not present, allocate and link.  Note that other
94 	 * thread might execute this sequence in parallel.  This
95 	 * should be safe, because the context entries written by both
96 	 * threads are equal.
97 	 */
98 	TD_PREP_PINNED_ASSERT;
99 	ctxm = dmar_pgalloc(dmar->ctx_obj, 1 + bus, IOMMU_PGF_ZERO |
100 	    IOMMU_PGF_WAITOK);
101 	re = dmar_map_pgtbl(dmar->ctx_obj, 0, IOMMU_PGF_NOALLOC, &sf);
102 	re += bus;
103 	dmar_pte_store(&re->r1, DMAR_ROOT_R1_P | (DMAR_ROOT_R1_CTP_MASK &
104 	    VM_PAGE_TO_PHYS(ctxm)));
105 	dmar_flush_root_to_ram(dmar, re);
106 	dmar_unmap_pgtbl(sf);
107 	TD_PINNED_ASSERT;
108 }
109 
110 static dmar_ctx_entry_t *
111 dmar_map_ctx_entry(struct dmar_ctx *ctx, struct sf_buf **sfp)
112 {
113 	struct dmar_unit *dmar;
114 	dmar_ctx_entry_t *ctxp;
115 
116 	dmar = CTX2DMAR(ctx);
117 
118 	ctxp = dmar_map_pgtbl(dmar->ctx_obj, 1 + PCI_RID2BUS(ctx->context.rid),
119 	    IOMMU_PGF_NOALLOC | IOMMU_PGF_WAITOK, sfp);
120 	ctxp += ctx->context.rid & 0xff;
121 	return (ctxp);
122 }
123 
124 static void
125 device_tag_init(struct dmar_ctx *ctx, device_t dev)
126 {
127 	struct dmar_domain *domain;
128 	bus_addr_t maxaddr;
129 
130 	domain = CTX2DOM(ctx);
131 	maxaddr = MIN(domain->iodom.end, BUS_SPACE_MAXADDR);
132 	ctx->context.tag->common.ref_count = 1; /* Prevent free */
133 	ctx->context.tag->common.impl = &bus_dma_iommu_impl;
134 	ctx->context.tag->common.boundary = 0;
135 	ctx->context.tag->common.lowaddr = maxaddr;
136 	ctx->context.tag->common.highaddr = maxaddr;
137 	ctx->context.tag->common.maxsize = maxaddr;
138 	ctx->context.tag->common.nsegments = BUS_SPACE_UNRESTRICTED;
139 	ctx->context.tag->common.maxsegsz = maxaddr;
140 	ctx->context.tag->ctx = CTX2IOCTX(ctx);
141 	ctx->context.tag->owner = dev;
142 }
143 
144 static void
145 ctx_id_entry_init_one(dmar_ctx_entry_t *ctxp, struct dmar_domain *domain,
146     vm_page_t ctx_root)
147 {
148 	/*
149 	 * For update due to move, the store is not atomic.  It is
150 	 * possible that DMAR read upper doubleword, while low
151 	 * doubleword is not yet updated.  The domain id is stored in
152 	 * the upper doubleword, while the table pointer in the lower.
153 	 *
154 	 * There is no good solution, for the same reason it is wrong
155 	 * to clear P bit in the ctx entry for update.
156 	 */
157 	dmar_pte_store1(&ctxp->ctx2, DMAR_CTX2_DID(domain->domain) |
158 	    domain->awlvl);
159 	if (ctx_root == NULL) {
160 		dmar_pte_store1(&ctxp->ctx1, DMAR_CTX1_T_PASS | DMAR_CTX1_P);
161 	} else {
162 		dmar_pte_store1(&ctxp->ctx1, DMAR_CTX1_T_UNTR |
163 		    (DMAR_CTX1_ASR_MASK & VM_PAGE_TO_PHYS(ctx_root)) |
164 		    DMAR_CTX1_P);
165 	}
166 }
167 
168 static void
169 ctx_id_entry_init(struct dmar_ctx *ctx, dmar_ctx_entry_t *ctxp, bool move,
170     int busno)
171 {
172 	struct dmar_unit *unit;
173 	struct dmar_domain *domain;
174 	vm_page_t ctx_root;
175 	int i;
176 
177 	domain = CTX2DOM(ctx);
178 	unit = DOM2DMAR(domain);
179 	KASSERT(move || (ctxp->ctx1 == 0 && ctxp->ctx2 == 0),
180 	    ("dmar%d: initialized ctx entry %d:%d:%d 0x%jx 0x%jx",
181 	    unit->iommu.unit, busno, pci_get_slot(ctx->context.tag->owner),
182 	    pci_get_function(ctx->context.tag->owner),
183 	    ctxp->ctx1, ctxp->ctx2));
184 
185 	if ((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0 &&
186 	    (unit->hw_ecap & DMAR_ECAP_PT) != 0) {
187 		KASSERT(domain->pgtbl_obj == NULL,
188 		    ("ctx %p non-null pgtbl_obj", ctx));
189 		ctx_root = NULL;
190 	} else {
191 		ctx_root = dmar_pgalloc(domain->pgtbl_obj, 0,
192 		    IOMMU_PGF_NOALLOC);
193 	}
194 
195 	if (iommu_is_buswide_ctx(DMAR2IOMMU(unit), busno)) {
196 		MPASS(!move);
197 		for (i = 0; i <= PCI_BUSMAX; i++) {
198 			ctx_id_entry_init_one(&ctxp[i], domain, ctx_root);
199 		}
200 	} else {
201 		ctx_id_entry_init_one(ctxp, domain, ctx_root);
202 	}
203 	dmar_flush_ctx_to_ram(unit, ctxp);
204 }
205 
206 static int
207 dmar_flush_for_ctx_entry(struct dmar_unit *dmar, bool force)
208 {
209 	int error;
210 
211 	/*
212 	 * If dmar declares Caching Mode as Set, follow 11.5 "Caching
213 	 * Mode Consideration" and do the (global) invalidation of the
214 	 * negative TLB entries.
215 	 */
216 	if ((dmar->hw_cap & DMAR_CAP_CM) == 0 && !force)
217 		return (0);
218 	if (dmar->qi_enabled) {
219 		dmar_qi_invalidate_ctx_glob_locked(dmar);
220 		if ((dmar->hw_ecap & DMAR_ECAP_DI) != 0 || force)
221 			dmar_qi_invalidate_iotlb_glob_locked(dmar);
222 		return (0);
223 	}
224 	error = dmar_inv_ctx_glob(dmar);
225 	if (error == 0 && ((dmar->hw_ecap & DMAR_ECAP_DI) != 0 || force))
226 		error = dmar_inv_iotlb_glob(dmar);
227 	return (error);
228 }
229 
230 static int
231 domain_init_rmrr(struct dmar_domain *domain, device_t dev, int bus,
232     int slot, int func, int dev_domain, int dev_busno,
233     const void *dev_path, int dev_path_len)
234 {
235 	struct iommu_map_entries_tailq rmrr_entries;
236 	struct iommu_map_entry *entry, *entry1;
237 	vm_page_t *ma;
238 	iommu_gaddr_t start, end;
239 	vm_pindex_t size, i;
240 	int error, error1;
241 
242 	error = 0;
243 	TAILQ_INIT(&rmrr_entries);
244 	dmar_dev_parse_rmrr(domain, dev_domain, dev_busno, dev_path,
245 	    dev_path_len, &rmrr_entries);
246 	TAILQ_FOREACH_SAFE(entry, &rmrr_entries, dmamap_link, entry1) {
247 		/*
248 		 * VT-d specification requires that the start of an
249 		 * RMRR entry is 4k-aligned.  Buggy BIOSes put
250 		 * anything into the start and end fields.  Truncate
251 		 * and round as neccesary.
252 		 *
253 		 * We also allow the overlapping RMRR entries, see
254 		 * iommu_gas_alloc_region().
255 		 */
256 		start = entry->start;
257 		end = entry->end;
258 		if (bootverbose)
259 			printf("dmar%d ctx pci%d:%d:%d RMRR [%#jx, %#jx]\n",
260 			    domain->iodom.iommu->unit, bus, slot, func,
261 			    (uintmax_t)start, (uintmax_t)end);
262 		entry->start = trunc_page(start);
263 		entry->end = round_page(end);
264 		if (entry->start == entry->end) {
265 			/* Workaround for some AMI (?) BIOSes */
266 			if (bootverbose) {
267 				if (dev != NULL)
268 					device_printf(dev, "");
269 				printf("pci%d:%d:%d ", bus, slot, func);
270 				printf("BIOS bug: dmar%d RMRR "
271 				    "region (%jx, %jx) corrected\n",
272 				    domain->iodom.iommu->unit, start, end);
273 			}
274 			entry->end += DMAR_PAGE_SIZE * 0x20;
275 		}
276 		size = OFF_TO_IDX(entry->end - entry->start);
277 		ma = malloc(sizeof(vm_page_t) * size, M_TEMP, M_WAITOK);
278 		for (i = 0; i < size; i++) {
279 			ma[i] = vm_page_getfake(entry->start + PAGE_SIZE * i,
280 			    VM_MEMATTR_DEFAULT);
281 		}
282 		error1 = iommu_gas_map_region(DOM2IODOM(domain), entry,
283 		    IOMMU_MAP_ENTRY_READ | IOMMU_MAP_ENTRY_WRITE,
284 		    IOMMU_MF_CANWAIT | IOMMU_MF_RMRR, ma);
285 		/*
286 		 * Non-failed RMRR entries are owned by context rb
287 		 * tree.  Get rid of the failed entry, but do not stop
288 		 * the loop.  Rest of the parsed RMRR entries are
289 		 * loaded and removed on the context destruction.
290 		 */
291 		if (error1 == 0 && entry->end != entry->start) {
292 			IOMMU_LOCK(domain->iodom.iommu);
293 			domain->refs++; /* XXXKIB prevent free */
294 			domain->iodom.flags |= IOMMU_DOMAIN_RMRR;
295 			IOMMU_UNLOCK(domain->iodom.iommu);
296 		} else {
297 			if (error1 != 0) {
298 				if (dev != NULL)
299 					device_printf(dev, "");
300 				printf("pci%d:%d:%d ", bus, slot, func);
301 				printf(
302 			    "dmar%d failed to map RMRR region (%jx, %jx) %d\n",
303 				    domain->iodom.iommu->unit, start, end,
304 				    error1);
305 				error = error1;
306 			}
307 			TAILQ_REMOVE(&rmrr_entries, entry, dmamap_link);
308 			iommu_gas_free_entry(entry);
309 		}
310 		for (i = 0; i < size; i++)
311 			vm_page_putfake(ma[i]);
312 		free(ma, M_TEMP);
313 	}
314 	return (error);
315 }
316 
317 /*
318  * PCI memory address space is shared between memory-mapped devices (MMIO) and
319  * host memory (which may be remapped by an IOMMU).  Device accesses to an
320  * address within a memory aperture in a PCIe root port will be treated as
321  * peer-to-peer and not forwarded to an IOMMU.  To avoid this, reserve the
322  * address space of the root port's memory apertures in the address space used
323  * by the IOMMU for remapping.
324  */
325 static int
326 dmar_reserve_pci_regions(struct dmar_domain *domain, device_t dev)
327 {
328 	struct iommu_domain *iodom;
329 	device_t root;
330 	uint32_t val;
331 	uint64_t base, limit;
332 	int error;
333 
334 	iodom = DOM2IODOM(domain);
335 
336 	root = pci_find_pcie_root_port(dev);
337 	if (root == NULL)
338 		return (0);
339 
340 	/* Disable downstream memory */
341 	base = PCI_PPBMEMBASE(0, pci_read_config(root, PCIR_MEMBASE_1, 2));
342 	limit = PCI_PPBMEMLIMIT(0, pci_read_config(root, PCIR_MEMLIMIT_1, 2));
343 	error = iommu_gas_reserve_region_extend(iodom, base, limit + 1);
344 	if (bootverbose || error != 0)
345 		device_printf(dev, "DMAR reserve [%#jx-%#jx] (error %d)\n",
346 		    base, limit + 1, error);
347 	if (error != 0)
348 		return (error);
349 
350 	/* Disable downstream prefetchable memory */
351 	val = pci_read_config(root, PCIR_PMBASEL_1, 2);
352 	if (val != 0 || pci_read_config(root, PCIR_PMLIMITL_1, 2) != 0) {
353 		if ((val & PCIM_BRPM_MASK) == PCIM_BRPM_64) {
354 			base = PCI_PPBMEMBASE(
355 			    pci_read_config(root, PCIR_PMBASEH_1, 4),
356 			    val);
357 			limit = PCI_PPBMEMLIMIT(
358 			    pci_read_config(root, PCIR_PMLIMITH_1, 4),
359 			    pci_read_config(root, PCIR_PMLIMITL_1, 2));
360 		} else {
361 			base = PCI_PPBMEMBASE(0, val);
362 			limit = PCI_PPBMEMLIMIT(0,
363 			    pci_read_config(root, PCIR_PMLIMITL_1, 2));
364 		}
365 		error = iommu_gas_reserve_region_extend(iodom, base,
366 		    limit + 1);
367 		if (bootverbose || error != 0)
368 			device_printf(dev, "DMAR reserve [%#jx-%#jx] "
369 			    "(error %d)\n", base, limit + 1, error);
370 		if (error != 0)
371 			return (error);
372 	}
373 
374 	return (error);
375 }
376 
377 static struct dmar_domain *
378 dmar_domain_alloc(struct dmar_unit *dmar, bool id_mapped)
379 {
380 	struct iommu_domain *iodom;
381 	struct iommu_unit *unit;
382 	struct dmar_domain *domain;
383 	int error, id, mgaw;
384 
385 	id = alloc_unr(dmar->domids);
386 	if (id == -1)
387 		return (NULL);
388 	domain = malloc(sizeof(*domain), M_DMAR_DOMAIN, M_WAITOK | M_ZERO);
389 	iodom = DOM2IODOM(domain);
390 	unit = DMAR2IOMMU(dmar);
391 	domain->domain = id;
392 	LIST_INIT(&domain->contexts);
393 	iommu_domain_init(unit, iodom, &dmar_domain_map_ops);
394 
395 	domain->dmar = dmar;
396 
397 	/*
398 	 * For now, use the maximal usable physical address of the
399 	 * installed memory to calculate the mgaw on id_mapped domain.
400 	 * It is useful for the identity mapping, and less so for the
401 	 * virtualized bus address space.
402 	 */
403 	domain->iodom.end = id_mapped ? ptoa(Maxmem) : BUS_SPACE_MAXADDR;
404 	mgaw = dmar_maxaddr2mgaw(dmar, domain->iodom.end, !id_mapped);
405 	error = domain_set_agaw(domain, mgaw);
406 	if (error != 0)
407 		goto fail;
408 	if (!id_mapped)
409 		/* Use all supported address space for remapping. */
410 		domain->iodom.end = 1ULL << (domain->agaw - 1);
411 
412 	iommu_gas_init_domain(DOM2IODOM(domain));
413 
414 	if (id_mapped) {
415 		if ((dmar->hw_ecap & DMAR_ECAP_PT) == 0) {
416 			domain->pgtbl_obj = domain_get_idmap_pgtbl(domain,
417 			    domain->iodom.end);
418 		}
419 		domain->iodom.flags |= IOMMU_DOMAIN_IDMAP;
420 	} else {
421 		error = domain_alloc_pgtbl(domain);
422 		if (error != 0)
423 			goto fail;
424 		/* Disable local apic region access */
425 		error = iommu_gas_reserve_region(iodom, 0xfee00000,
426 		    0xfeefffff + 1, &iodom->msi_entry);
427 		if (error != 0)
428 			goto fail;
429 	}
430 	return (domain);
431 
432 fail:
433 	dmar_domain_destroy(domain);
434 	return (NULL);
435 }
436 
437 static struct dmar_ctx *
438 dmar_ctx_alloc(struct dmar_domain *domain, uint16_t rid)
439 {
440 	struct dmar_ctx *ctx;
441 
442 	ctx = malloc(sizeof(*ctx), M_DMAR_CTX, M_WAITOK | M_ZERO);
443 	ctx->context.domain = DOM2IODOM(domain);
444 	ctx->context.tag = malloc(sizeof(struct bus_dma_tag_iommu),
445 	    M_DMAR_CTX, M_WAITOK | M_ZERO);
446 	ctx->context.rid = rid;
447 	ctx->refs = 1;
448 	return (ctx);
449 }
450 
451 static void
452 dmar_ctx_link(struct dmar_ctx *ctx)
453 {
454 	struct dmar_domain *domain;
455 
456 	domain = CTX2DOM(ctx);
457 	IOMMU_ASSERT_LOCKED(domain->iodom.iommu);
458 	KASSERT(domain->refs >= domain->ctx_cnt,
459 	    ("dom %p ref underflow %d %d", domain, domain->refs,
460 	    domain->ctx_cnt));
461 	domain->refs++;
462 	domain->ctx_cnt++;
463 	LIST_INSERT_HEAD(&domain->contexts, ctx, link);
464 }
465 
466 static void
467 dmar_ctx_unlink(struct dmar_ctx *ctx)
468 {
469 	struct dmar_domain *domain;
470 
471 	domain = CTX2DOM(ctx);
472 	IOMMU_ASSERT_LOCKED(domain->iodom.iommu);
473 	KASSERT(domain->refs > 0,
474 	    ("domain %p ctx dtr refs %d", domain, domain->refs));
475 	KASSERT(domain->ctx_cnt >= domain->refs,
476 	    ("domain %p ctx dtr refs %d ctx_cnt %d", domain,
477 	    domain->refs, domain->ctx_cnt));
478 	domain->refs--;
479 	domain->ctx_cnt--;
480 	LIST_REMOVE(ctx, link);
481 }
482 
483 static void
484 dmar_domain_destroy(struct dmar_domain *domain)
485 {
486 	struct iommu_domain *iodom;
487 	struct dmar_unit *dmar;
488 
489 	iodom = DOM2IODOM(domain);
490 
491 	KASSERT(TAILQ_EMPTY(&domain->iodom.unload_entries),
492 	    ("unfinished unloads %p", domain));
493 	KASSERT(LIST_EMPTY(&domain->contexts),
494 	    ("destroying dom %p with contexts", domain));
495 	KASSERT(domain->ctx_cnt == 0,
496 	    ("destroying dom %p with ctx_cnt %d", domain, domain->ctx_cnt));
497 	KASSERT(domain->refs == 0,
498 	    ("destroying dom %p with refs %d", domain, domain->refs));
499 	if ((domain->iodom.flags & IOMMU_DOMAIN_GAS_INITED) != 0) {
500 		DMAR_DOMAIN_LOCK(domain);
501 		iommu_gas_fini_domain(iodom);
502 		DMAR_DOMAIN_UNLOCK(domain);
503 	}
504 	if ((domain->iodom.flags & IOMMU_DOMAIN_PGTBL_INITED) != 0) {
505 		if (domain->pgtbl_obj != NULL)
506 			DMAR_DOMAIN_PGLOCK(domain);
507 		domain_free_pgtbl(domain);
508 	}
509 	iommu_domain_fini(iodom);
510 	dmar = DOM2DMAR(domain);
511 	free_unr(dmar->domids, domain->domain);
512 	free(domain, M_DMAR_DOMAIN);
513 }
514 
515 static struct dmar_ctx *
516 dmar_get_ctx_for_dev1(struct dmar_unit *dmar, device_t dev, uint16_t rid,
517     int dev_domain, int dev_busno, const void *dev_path, int dev_path_len,
518     bool id_mapped, bool rmrr_init)
519 {
520 	struct dmar_domain *domain, *domain1;
521 	struct dmar_ctx *ctx, *ctx1;
522 	struct iommu_unit *unit __diagused;
523 	dmar_ctx_entry_t *ctxp;
524 	struct sf_buf *sf;
525 	int bus, slot, func, error;
526 	bool enable;
527 
528 	if (dev != NULL) {
529 		bus = pci_get_bus(dev);
530 		slot = pci_get_slot(dev);
531 		func = pci_get_function(dev);
532 	} else {
533 		bus = PCI_RID2BUS(rid);
534 		slot = PCI_RID2SLOT(rid);
535 		func = PCI_RID2FUNC(rid);
536 	}
537 	enable = false;
538 	TD_PREP_PINNED_ASSERT;
539 	unit = DMAR2IOMMU(dmar);
540 	DMAR_LOCK(dmar);
541 	KASSERT(!iommu_is_buswide_ctx(unit, bus) || (slot == 0 && func == 0),
542 	    ("iommu%d pci%d:%d:%d get_ctx for buswide", dmar->iommu.unit, bus,
543 	    slot, func));
544 	ctx = dmar_find_ctx_locked(dmar, rid);
545 	error = 0;
546 	if (ctx == NULL) {
547 		/*
548 		 * Perform the allocations which require sleep or have
549 		 * higher chance to succeed if the sleep is allowed.
550 		 */
551 		DMAR_UNLOCK(dmar);
552 		dmar_ensure_ctx_page(dmar, PCI_RID2BUS(rid));
553 		domain1 = dmar_domain_alloc(dmar, id_mapped);
554 		if (domain1 == NULL) {
555 			TD_PINNED_ASSERT;
556 			return (NULL);
557 		}
558 		if (!id_mapped) {
559 			error = domain_init_rmrr(domain1, dev, bus,
560 			    slot, func, dev_domain, dev_busno, dev_path,
561 			    dev_path_len);
562 			if (error == 0 && dev != NULL)
563 				error = dmar_reserve_pci_regions(domain1, dev);
564 			if (error != 0) {
565 				dmar_domain_destroy(domain1);
566 				TD_PINNED_ASSERT;
567 				return (NULL);
568 			}
569 		}
570 		ctx1 = dmar_ctx_alloc(domain1, rid);
571 		ctxp = dmar_map_ctx_entry(ctx1, &sf);
572 		DMAR_LOCK(dmar);
573 
574 		/*
575 		 * Recheck the contexts, other thread might have
576 		 * already allocated needed one.
577 		 */
578 		ctx = dmar_find_ctx_locked(dmar, rid);
579 		if (ctx == NULL) {
580 			domain = domain1;
581 			ctx = ctx1;
582 			dmar_ctx_link(ctx);
583 			ctx->context.tag->owner = dev;
584 			device_tag_init(ctx, dev);
585 
586 			/*
587 			 * This is the first activated context for the
588 			 * DMAR unit.  Enable the translation after
589 			 * everything is set up.
590 			 */
591 			if (LIST_EMPTY(&dmar->domains))
592 				enable = true;
593 			LIST_INSERT_HEAD(&dmar->domains, domain, link);
594 			ctx_id_entry_init(ctx, ctxp, false, bus);
595 			if (dev != NULL) {
596 				device_printf(dev,
597 			    "dmar%d pci%d:%d:%d:%d rid %x domain %d mgaw %d "
598 				    "agaw %d %s-mapped\n",
599 				    dmar->iommu.unit, dmar->segment, bus, slot,
600 				    func, rid, domain->domain, domain->mgaw,
601 				    domain->agaw, id_mapped ? "id" : "re");
602 			}
603 			dmar_unmap_pgtbl(sf);
604 		} else {
605 			dmar_unmap_pgtbl(sf);
606 			dmar_domain_destroy(domain1);
607 			/* Nothing needs to be done to destroy ctx1. */
608 			free(ctx1, M_DMAR_CTX);
609 			domain = CTX2DOM(ctx);
610 			ctx->refs++; /* tag referenced us */
611 		}
612 	} else {
613 		domain = CTX2DOM(ctx);
614 		if (ctx->context.tag->owner == NULL)
615 			ctx->context.tag->owner = dev;
616 		ctx->refs++; /* tag referenced us */
617 	}
618 
619 	error = dmar_flush_for_ctx_entry(dmar, enable);
620 	if (error != 0) {
621 		dmar_free_ctx_locked(dmar, ctx);
622 		TD_PINNED_ASSERT;
623 		return (NULL);
624 	}
625 
626 	/*
627 	 * The dmar lock was potentially dropped between check for the
628 	 * empty context list and now.  Recheck the state of GCMD_TE
629 	 * to avoid unneeded command.
630 	 */
631 	if (enable && !rmrr_init && (dmar->hw_gcmd & DMAR_GCMD_TE) == 0) {
632 		error = dmar_disable_protected_regions(dmar);
633 		if (error != 0)
634 			printf("dmar%d: Failed to disable protected regions\n",
635 			    dmar->iommu.unit);
636 		error = dmar_enable_translation(dmar);
637 		if (error == 0) {
638 			if (bootverbose) {
639 				printf("dmar%d: enabled translation\n",
640 				    dmar->iommu.unit);
641 			}
642 		} else {
643 			printf("dmar%d: enabling translation failed, "
644 			    "error %d\n", dmar->iommu.unit, error);
645 			dmar_free_ctx_locked(dmar, ctx);
646 			TD_PINNED_ASSERT;
647 			return (NULL);
648 		}
649 	}
650 	DMAR_UNLOCK(dmar);
651 	TD_PINNED_ASSERT;
652 	return (ctx);
653 }
654 
655 struct dmar_ctx *
656 dmar_get_ctx_for_dev(struct dmar_unit *dmar, device_t dev, uint16_t rid,
657     bool id_mapped, bool rmrr_init)
658 {
659 	int dev_domain, dev_path_len, dev_busno;
660 
661 	dev_domain = pci_get_domain(dev);
662 	dev_path_len = dmar_dev_depth(dev);
663 	ACPI_DMAR_PCI_PATH dev_path[dev_path_len];
664 	dmar_dev_path(dev, &dev_busno, dev_path, dev_path_len);
665 	return (dmar_get_ctx_for_dev1(dmar, dev, rid, dev_domain, dev_busno,
666 	    dev_path, dev_path_len, id_mapped, rmrr_init));
667 }
668 
669 struct dmar_ctx *
670 dmar_get_ctx_for_devpath(struct dmar_unit *dmar, uint16_t rid,
671     int dev_domain, int dev_busno,
672     const void *dev_path, int dev_path_len,
673     bool id_mapped, bool rmrr_init)
674 {
675 
676 	return (dmar_get_ctx_for_dev1(dmar, NULL, rid, dev_domain, dev_busno,
677 	    dev_path, dev_path_len, id_mapped, rmrr_init));
678 }
679 
680 int
681 dmar_move_ctx_to_domain(struct dmar_domain *domain, struct dmar_ctx *ctx)
682 {
683 	struct dmar_unit *dmar;
684 	struct dmar_domain *old_domain;
685 	dmar_ctx_entry_t *ctxp;
686 	struct sf_buf *sf;
687 	int error;
688 
689 	dmar = domain->dmar;
690 	old_domain = CTX2DOM(ctx);
691 	if (domain == old_domain)
692 		return (0);
693 	KASSERT(old_domain->iodom.iommu == domain->iodom.iommu,
694 	    ("domain %p %u moving between dmars %u %u", domain,
695 	    domain->domain, old_domain->iodom.iommu->unit,
696 	    domain->iodom.iommu->unit));
697 	TD_PREP_PINNED_ASSERT;
698 
699 	ctxp = dmar_map_ctx_entry(ctx, &sf);
700 	DMAR_LOCK(dmar);
701 	dmar_ctx_unlink(ctx);
702 	ctx->context.domain = &domain->iodom;
703 	dmar_ctx_link(ctx);
704 	ctx_id_entry_init(ctx, ctxp, true, PCI_BUSMAX + 100);
705 	dmar_unmap_pgtbl(sf);
706 	error = dmar_flush_for_ctx_entry(dmar, true);
707 	/* If flush failed, rolling back would not work as well. */
708 	printf("dmar%d rid %x domain %d->%d %s-mapped\n",
709 	    dmar->iommu.unit, ctx->context.rid, old_domain->domain,
710 	    domain->domain, (domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0 ?
711 	    "id" : "re");
712 	dmar_unref_domain_locked(dmar, old_domain);
713 	TD_PINNED_ASSERT;
714 	return (error);
715 }
716 
717 static void
718 dmar_unref_domain_locked(struct dmar_unit *dmar, struct dmar_domain *domain)
719 {
720 
721 	DMAR_ASSERT_LOCKED(dmar);
722 	KASSERT(domain->refs >= 1,
723 	    ("dmar %d domain %p refs %u", dmar->iommu.unit, domain,
724 	    domain->refs));
725 	KASSERT(domain->refs > domain->ctx_cnt,
726 	    ("dmar %d domain %p refs %d ctx_cnt %d", dmar->iommu.unit, domain,
727 	    domain->refs, domain->ctx_cnt));
728 
729 	if (domain->refs > 1) {
730 		domain->refs--;
731 		DMAR_UNLOCK(dmar);
732 		return;
733 	}
734 
735 	KASSERT((domain->iodom.flags & IOMMU_DOMAIN_RMRR) == 0,
736 	    ("lost ref on RMRR domain %p", domain));
737 
738 	LIST_REMOVE(domain, link);
739 	DMAR_UNLOCK(dmar);
740 
741 	taskqueue_drain(dmar->iommu.delayed_taskqueue,
742 	    &domain->iodom.unload_task);
743 	dmar_domain_destroy(domain);
744 }
745 
746 void
747 dmar_free_ctx_locked(struct dmar_unit *dmar, struct dmar_ctx *ctx)
748 {
749 	struct sf_buf *sf;
750 	dmar_ctx_entry_t *ctxp;
751 	struct dmar_domain *domain;
752 
753 	DMAR_ASSERT_LOCKED(dmar);
754 	KASSERT(ctx->refs >= 1,
755 	    ("dmar %p ctx %p refs %u", dmar, ctx, ctx->refs));
756 
757 	/*
758 	 * If our reference is not last, only the dereference should
759 	 * be performed.
760 	 */
761 	if (ctx->refs > 1) {
762 		ctx->refs--;
763 		DMAR_UNLOCK(dmar);
764 		return;
765 	}
766 
767 	KASSERT((ctx->context.flags & IOMMU_CTX_DISABLED) == 0,
768 	    ("lost ref on disabled ctx %p", ctx));
769 
770 	/*
771 	 * Otherwise, the context entry must be cleared before the
772 	 * page table is destroyed.  The mapping of the context
773 	 * entries page could require sleep, unlock the dmar.
774 	 */
775 	DMAR_UNLOCK(dmar);
776 	TD_PREP_PINNED_ASSERT;
777 	ctxp = dmar_map_ctx_entry(ctx, &sf);
778 	DMAR_LOCK(dmar);
779 	KASSERT(ctx->refs >= 1,
780 	    ("dmar %p ctx %p refs %u", dmar, ctx, ctx->refs));
781 
782 	/*
783 	 * Other thread might have referenced the context, in which
784 	 * case again only the dereference should be performed.
785 	 */
786 	if (ctx->refs > 1) {
787 		ctx->refs--;
788 		DMAR_UNLOCK(dmar);
789 		dmar_unmap_pgtbl(sf);
790 		TD_PINNED_ASSERT;
791 		return;
792 	}
793 
794 	KASSERT((ctx->context.flags & IOMMU_CTX_DISABLED) == 0,
795 	    ("lost ref on disabled ctx %p", ctx));
796 
797 	/*
798 	 * Clear the context pointer and flush the caches.
799 	 * XXXKIB: cannot do this if any RMRR entries are still present.
800 	 */
801 	dmar_pte_clear(&ctxp->ctx1);
802 	ctxp->ctx2 = 0;
803 	dmar_flush_ctx_to_ram(dmar, ctxp);
804 	dmar_inv_ctx_glob(dmar);
805 	if ((dmar->hw_ecap & DMAR_ECAP_DI) != 0) {
806 		if (dmar->qi_enabled)
807 			dmar_qi_invalidate_iotlb_glob_locked(dmar);
808 		else
809 			dmar_inv_iotlb_glob(dmar);
810 	}
811 	dmar_unmap_pgtbl(sf);
812 	domain = CTX2DOM(ctx);
813 	dmar_ctx_unlink(ctx);
814 	free(ctx->context.tag, M_DMAR_CTX);
815 	free(ctx, M_DMAR_CTX);
816 	dmar_unref_domain_locked(dmar, domain);
817 	TD_PINNED_ASSERT;
818 }
819 
820 void
821 dmar_free_ctx(struct dmar_ctx *ctx)
822 {
823 	struct dmar_unit *dmar;
824 
825 	dmar = CTX2DMAR(ctx);
826 	DMAR_LOCK(dmar);
827 	dmar_free_ctx_locked(dmar, ctx);
828 }
829 
830 /*
831  * Returns with the domain locked.
832  */
833 struct dmar_ctx *
834 dmar_find_ctx_locked(struct dmar_unit *dmar, uint16_t rid)
835 {
836 	struct dmar_domain *domain;
837 	struct dmar_ctx *ctx;
838 
839 	DMAR_ASSERT_LOCKED(dmar);
840 
841 	LIST_FOREACH(domain, &dmar->domains, link) {
842 		LIST_FOREACH(ctx, &domain->contexts, link) {
843 			if (ctx->context.rid == rid)
844 				return (ctx);
845 		}
846 	}
847 	return (NULL);
848 }
849 
850 void
851 dmar_domain_free_entry(struct iommu_map_entry *entry, bool free)
852 {
853 	if ((entry->flags & IOMMU_MAP_ENTRY_RMRR) != 0)
854 		iommu_gas_free_region(entry);
855 	else
856 		iommu_gas_free_space(entry);
857 	if (free)
858 		iommu_gas_free_entry(entry);
859 	else
860 		entry->flags = 0;
861 }
862 
863 /*
864  * If the given value for "free" is true, then the caller must not be using
865  * the entry's dmamap_link field.
866  */
867 void
868 iommu_domain_unload_entry(struct iommu_map_entry *entry, bool free,
869     bool cansleep)
870 {
871 	struct dmar_domain *domain;
872 	struct dmar_unit *unit;
873 
874 	domain = IODOM2DOM(entry->domain);
875 	unit = DOM2DMAR(domain);
876 
877 	/*
878 	 * If "free" is false, then the IOTLB invalidation must be performed
879 	 * synchronously.  Otherwise, the caller might free the entry before
880 	 * dmar_qi_task() is finished processing it.
881 	 */
882 	if (unit->qi_enabled) {
883 		if (free) {
884 			DMAR_LOCK(unit);
885 			dmar_qi_invalidate_locked(domain, entry, true);
886 			DMAR_UNLOCK(unit);
887 		} else {
888 			dmar_qi_invalidate_sync(domain, entry->start,
889 			    entry->end - entry->start, cansleep);
890 			dmar_domain_free_entry(entry, false);
891 		}
892 	} else {
893 		domain_flush_iotlb_sync(domain, entry->start, entry->end -
894 		    entry->start);
895 		dmar_domain_free_entry(entry, free);
896 	}
897 }
898 
899 static bool
900 dmar_domain_unload_emit_wait(struct dmar_domain *domain,
901     struct iommu_map_entry *entry)
902 {
903 
904 	if (TAILQ_NEXT(entry, dmamap_link) == NULL)
905 		return (true);
906 	return (domain->batch_no++ % dmar_batch_coalesce == 0);
907 }
908 
909 void
910 iommu_domain_unload(struct iommu_domain *iodom,
911     struct iommu_map_entries_tailq *entries, bool cansleep)
912 {
913 	struct dmar_domain *domain;
914 	struct dmar_unit *unit;
915 	struct iommu_map_entry *entry, *entry1;
916 	int error __diagused;
917 
918 	domain = IODOM2DOM(iodom);
919 	unit = DOM2DMAR(domain);
920 
921 	TAILQ_FOREACH_SAFE(entry, entries, dmamap_link, entry1) {
922 		KASSERT((entry->flags & IOMMU_MAP_ENTRY_MAP) != 0,
923 		    ("not mapped entry %p %p", domain, entry));
924 		error = iodom->ops->unmap(iodom, entry->start, entry->end -
925 		    entry->start, cansleep ? IOMMU_PGF_WAITOK : 0);
926 		KASSERT(error == 0, ("unmap %p error %d", domain, error));
927 		if (!unit->qi_enabled) {
928 			domain_flush_iotlb_sync(domain, entry->start,
929 			    entry->end - entry->start);
930 			TAILQ_REMOVE(entries, entry, dmamap_link);
931 			dmar_domain_free_entry(entry, true);
932 		}
933 	}
934 	if (TAILQ_EMPTY(entries))
935 		return;
936 
937 	KASSERT(unit->qi_enabled, ("loaded entry left"));
938 	DMAR_LOCK(unit);
939 	while ((entry = TAILQ_FIRST(entries)) != NULL) {
940 		TAILQ_REMOVE(entries, entry, dmamap_link);
941 		dmar_qi_invalidate_locked(domain, entry,
942 		    dmar_domain_unload_emit_wait(domain, entry));
943 	}
944 	DMAR_UNLOCK(unit);
945 }
946 
947 struct iommu_ctx *
948 iommu_get_ctx(struct iommu_unit *iommu, device_t dev, uint16_t rid,
949     bool id_mapped, bool rmrr_init)
950 {
951 	struct dmar_unit *dmar;
952 	struct dmar_ctx *ret;
953 
954 	dmar = IOMMU2DMAR(iommu);
955 
956 	ret = dmar_get_ctx_for_dev(dmar, dev, rid, id_mapped, rmrr_init);
957 
958 	return (CTX2IOCTX(ret));
959 }
960 
961 void
962 iommu_free_ctx_locked(struct iommu_unit *iommu, struct iommu_ctx *context)
963 {
964 	struct dmar_unit *dmar;
965 	struct dmar_ctx *ctx;
966 
967 	dmar = IOMMU2DMAR(iommu);
968 	ctx = IOCTX2CTX(context);
969 
970 	dmar_free_ctx_locked(dmar, ctx);
971 }
972 
973 void
974 iommu_free_ctx(struct iommu_ctx *context)
975 {
976 	struct dmar_ctx *ctx;
977 
978 	ctx = IOCTX2CTX(context);
979 
980 	dmar_free_ctx(ctx);
981 }
982