2 * Copyright (c) 2013 The FreeBSD Foundation
5 * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
6 * under sponsorship from the FreeBSD Foundation.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/malloc.h>
38 #include <sys/interrupt.h>
39 #include <sys/kernel.h>
43 #include <sys/memdesc.h>
44 #include <sys/mutex.h>
45 #include <sys/sysctl.h>
47 #include <sys/taskqueue.h>
50 #include <dev/pci/pcireg.h>
51 #include <dev/pci/pcivar.h>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_kern.h>
55 #include <vm/vm_object.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_map.h>
58 #include <machine/atomic.h>
59 #include <machine/bus.h>
60 #include <machine/md_var.h>
61 #include <machine/specialreg.h>
62 #include <x86/include/busdma_impl.h>
63 #include <x86/iommu/intel_reg.h>
64 #include <x86/iommu/busdma_dmar.h>
65 #include <x86/iommu/intel_dmar.h>
68 * busdma_dmar.c, the implementation of the busdma(9) interface using
69 * DMAR units from Intel VT-d.
73 dmar_bus_dma_is_dev_disabled(int domain, int bus, int slot, int func)
77 snprintf(str, sizeof(str), "hw.busdma.pci%d.%d.%d.%d.bounce",
78 domain, bus, slot, func);
87 * Given original device, find the requester ID that will be seen by
88 * the DMAR unit and used for page table lookup. PCI bridges may take
89 * ownership of transactions from downstream devices, so it may not be
90 * the same as the BSF of the target device. In those cases, all
91 * devices downstream of the bridge must share a single mapping
92 * domain, and must collectively be assigned to use either DMAR or
96 dmar_get_requester(device_t dev, uint16_t *rid)
99 device_t l, pci, pcib, pcip, pcibp, requester;
104 pci_class = devclass_find("pci");
107 *rid = pci_get_rid(dev);
110 * Walk the bridge hierarchy from the target device to the
111 * host port to find the translating bridge nearest the DMAR
115 pci = device_get_parent(l);
116 KASSERT(pci != NULL, ("dmar_get_requester(%s): NULL parent "
117 "for %s", device_get_name(dev), device_get_name(l)));
118 KASSERT(device_get_devclass(pci) == pci_class,
119 ("dmar_get_requester(%s): non-pci parent %s for %s",
120 device_get_name(dev), device_get_name(pci),
121 device_get_name(l)));
123 pcib = device_get_parent(pci);
124 KASSERT(pcib != NULL, ("dmar_get_requester(%s): NULL bridge "
125 "for %s", device_get_name(dev), device_get_name(pci)));
128 * The parent of our "bridge" isn't another PCI bus,
129 * so pcib isn't a PCI->PCI bridge but rather a host
130 * port, and the requester ID won't be translated
133 pcip = device_get_parent(pcib);
134 if (device_get_devclass(pcip) != pci_class)
136 pcibp = device_get_parent(pcip);
138 if (pci_find_cap(l, PCIY_EXPRESS, &cap_offset) == 0) {
140 * Do not stop the loop even if the target
141 * device is PCIe, because it is possible (but
142 * unlikely) to have a PCI->PCIe bridge
143 * somewhere in the hierarchy.
148 * Device is not PCIe, it cannot be seen as a
149 * requester by DMAR unit. Check whether the
152 bridge_is_pcie = pci_find_cap(pcib, PCIY_EXPRESS,
157 * Check for a buggy PCIe/PCI bridge that
158 * doesn't report the express capability. If
159 * the bridge above it is express but isn't a
160 * PCI bridge, then we know pcib is actually a
163 if (!bridge_is_pcie && pci_find_cap(pcibp,
164 PCIY_EXPRESS, &cap_offset) == 0) {
165 pcie_flags = pci_read_config(pcibp,
166 cap_offset + PCIER_FLAGS, 2);
167 if ((pcie_flags & PCIEM_FLAGS_TYPE) !=
168 PCIEM_TYPE_PCI_BRIDGE)
169 bridge_is_pcie = true;
172 if (bridge_is_pcie) {
174 * The current device is not PCIe, but
175 * the bridge above it is. This is a
176 * PCIe->PCI bridge. Assume that the
177 * requester ID will be the secondary
178 * bus number with slot and function
181 * XXX: Doesn't handle the case where
182 * the bridge is PCIe->PCI-X, and the
183 * bridge will only take ownership of
184 * requests in some cases. We should
185 * provide context entries with the
186 * same page tables for taken and
187 * non-taken transactions.
189 *rid = PCI_RID(pci_get_bus(l), 0, 0);
193 * Neither the device nor the bridge
194 * above it are PCIe. This is a
195 * conventional PCI->PCI bridge, which
196 * will use the bridge's BSF as the
199 *rid = pci_get_rid(pcib);
208 dmar_instantiate_ctx(struct dmar_unit *dmar, device_t dev, bool rmrr)
211 struct dmar_ctx *ctx;
215 requester = dmar_get_requester(dev, &rid);
218 * If the user requested the IOMMU disabled for the device, we
219 * cannot disable the DMAR, due to possibility of other
220 * devices on the same DMAR still requiring translation.
221 * Instead provide the identity mapping for the device
224 disabled = dmar_bus_dma_is_dev_disabled(pci_get_domain(requester),
225 pci_get_bus(requester), pci_get_slot(requester),
226 pci_get_function(requester));
227 ctx = dmar_get_ctx(dmar, requester, rid, disabled, rmrr);
232 * Keep the first reference on context, release the
236 if ((ctx->flags & DMAR_CTX_DISABLED) == 0) {
237 ctx->flags |= DMAR_CTX_DISABLED;
240 dmar_free_ctx_locked(dmar, ctx);
248 dmar_get_dma_tag(device_t dev, device_t child)
250 struct dmar_unit *dmar;
251 struct dmar_ctx *ctx;
254 dmar = dmar_find(child);
255 /* Not in scope of any DMAR ? */
258 dmar_quirks_pre_use(dmar);
259 dmar_instantiate_rmrr_ctxs(dmar);
261 ctx = dmar_instantiate_ctx(dmar, child, false);
262 res = ctx == NULL ? NULL : (bus_dma_tag_t)&ctx->ctx_tag;
266 static MALLOC_DEFINE(M_DMAR_DMAMAP, "dmar_dmamap", "Intel DMAR DMA Map");
268 static void dmar_bus_schedule_dmamap(struct dmar_unit *unit,
269 struct bus_dmamap_dmar *map);
272 dmar_bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment,
273 bus_addr_t boundary, bus_addr_t lowaddr, bus_addr_t highaddr,
274 bus_dma_filter_t *filter, void *filterarg, bus_size_t maxsize,
275 int nsegments, bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc,
276 void *lockfuncarg, bus_dma_tag_t *dmat)
278 struct bus_dma_tag_dmar *newtag, *oldtag;
282 error = common_bus_dma_tag_create(parent != NULL ?
283 &((struct bus_dma_tag_dmar *)parent)->common : NULL, alignment,
284 boundary, lowaddr, highaddr, filter, filterarg, maxsize,
285 nsegments, maxsegsz, flags, lockfunc, lockfuncarg,
286 sizeof(struct bus_dma_tag_dmar), (void **)&newtag);
290 oldtag = (struct bus_dma_tag_dmar *)parent;
291 newtag->common.impl = &bus_dma_dmar_impl;
292 newtag->ctx = oldtag->ctx;
293 newtag->owner = oldtag->owner;
295 *dmat = (bus_dma_tag_t)newtag;
297 CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d",
298 __func__, newtag, (newtag != NULL ? newtag->common.flags : 0),
304 dmar_bus_dma_tag_destroy(bus_dma_tag_t dmat1)
306 struct bus_dma_tag_dmar *dmat, *dmat_copy, *parent;
310 dmat_copy = dmat = (struct bus_dma_tag_dmar *)dmat1;
313 if (dmat->map_count != 0) {
317 while (dmat != NULL) {
318 parent = (struct bus_dma_tag_dmar *)dmat->common.parent;
319 if (atomic_fetchadd_int(&dmat->common.ref_count, -1) ==
321 if (dmat == &dmat->ctx->ctx_tag)
322 dmar_free_ctx(dmat->ctx);
323 free(dmat->segments, M_DMAR_DMAMAP);
324 free(dmat, M_DEVBUF);
331 CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error);
336 dmar_bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp)
338 struct bus_dma_tag_dmar *tag;
339 struct bus_dmamap_dmar *map;
341 tag = (struct bus_dma_tag_dmar *)dmat;
342 map = malloc(sizeof(*map), M_DMAR_DMAMAP, M_NOWAIT | M_ZERO);
347 if (tag->segments == NULL) {
348 tag->segments = malloc(sizeof(bus_dma_segment_t) *
349 tag->common.nsegments, M_DMAR_DMAMAP, M_NOWAIT);
350 if (tag->segments == NULL) {
351 free(map, M_DMAR_DMAMAP);
356 TAILQ_INIT(&map->map_entries);
359 map->cansleep = false;
361 *mapp = (bus_dmamap_t)map;
367 dmar_bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map1)
369 struct bus_dma_tag_dmar *tag;
370 struct bus_dmamap_dmar *map;
372 tag = (struct bus_dma_tag_dmar *)dmat;
373 map = (struct bus_dmamap_dmar *)map1;
375 DMAR_CTX_LOCK(tag->ctx);
376 if (!TAILQ_EMPTY(&map->map_entries)) {
377 DMAR_CTX_UNLOCK(tag->ctx);
380 DMAR_CTX_UNLOCK(tag->ctx);
381 free(map, M_DMAR_DMAMAP);
389 dmar_bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags,
392 struct bus_dma_tag_dmar *tag;
393 struct bus_dmamap_dmar *map;
397 error = dmar_bus_dmamap_create(dmat, flags, mapp);
401 mflags = (flags & BUS_DMA_NOWAIT) != 0 ? M_NOWAIT : M_WAITOK;
402 mflags |= (flags & BUS_DMA_ZERO) != 0 ? M_ZERO : 0;
403 attr = (flags & BUS_DMA_NOCACHE) != 0 ? VM_MEMATTR_UNCACHEABLE :
406 tag = (struct bus_dma_tag_dmar *)dmat;
407 map = (struct bus_dmamap_dmar *)*mapp;
409 if (tag->common.maxsize < PAGE_SIZE &&
410 tag->common.alignment <= tag->common.maxsize &&
411 attr == VM_MEMATTR_DEFAULT) {
412 *vaddr = malloc(tag->common.maxsize, M_DEVBUF, mflags);
413 map->flags |= BUS_DMAMAP_DMAR_MALLOC;
415 *vaddr = (void *)kmem_alloc_attr(kernel_arena,
416 tag->common.maxsize, mflags, 0ul, BUS_SPACE_MAXADDR,
418 map->flags |= BUS_DMAMAP_DMAR_KMEM_ALLOC;
420 if (*vaddr == NULL) {
421 dmar_bus_dmamap_destroy(dmat, *mapp);
429 dmar_bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map1)
431 struct bus_dma_tag_dmar *tag;
432 struct bus_dmamap_dmar *map;
434 tag = (struct bus_dma_tag_dmar *)dmat;
435 map = (struct bus_dmamap_dmar *)map1;
437 if ((map->flags & BUS_DMAMAP_DMAR_MALLOC) != 0) {
438 free(vaddr, M_DEVBUF);
439 map->flags &= ~BUS_DMAMAP_DMAR_MALLOC;
441 KASSERT((map->flags & BUS_DMAMAP_DMAR_KMEM_ALLOC) != 0,
442 ("dmar_bus_dmamem_free for non alloced map %p", map));
443 kmem_free(kernel_arena, (vm_offset_t)vaddr, tag->common.maxsize);
444 map->flags &= ~BUS_DMAMAP_DMAR_KMEM_ALLOC;
447 dmar_bus_dmamap_destroy(dmat, map1);
451 dmar_bus_dmamap_load_something1(struct bus_dma_tag_dmar *tag,
452 struct bus_dmamap_dmar *map, vm_page_t *ma, int offset, bus_size_t buflen,
453 int flags, bus_dma_segment_t *segs, int *segp,
454 struct dmar_map_entries_tailq *unroll_list)
456 struct dmar_ctx *ctx;
457 struct dmar_map_entry *entry;
460 int error, idx, gas_flags, seg;
462 KASSERT(offset < DMAR_PAGE_SIZE, ("offset %d", offset));
464 segs = tag->segments;
471 if (seg >= tag->common.nsegments) {
475 buflen1 = buflen > tag->common.maxsegsz ?
476 tag->common.maxsegsz : buflen;
477 size = round_page(offset + buflen1);
480 * (Too) optimistically allow split if there are more
481 * then one segments left.
483 gas_flags = map->cansleep ? DMAR_GM_CANWAIT : 0;
484 if (seg + 1 < tag->common.nsegments)
485 gas_flags |= DMAR_GM_CANSPLIT;
487 error = dmar_gas_map(ctx, &tag->common, size, offset,
488 DMAR_MAP_ENTRY_READ | DMAR_MAP_ENTRY_WRITE,
489 gas_flags, ma + idx, &entry);
492 if ((gas_flags & DMAR_GM_CANSPLIT) != 0) {
493 KASSERT(size >= entry->end - entry->start,
494 ("split increased entry size %jx %jx %jx",
495 (uintmax_t)size, (uintmax_t)entry->start,
496 (uintmax_t)entry->end));
497 size = entry->end - entry->start;
501 KASSERT(entry->end - entry->start == size,
502 ("no split allowed %jx %jx %jx",
503 (uintmax_t)size, (uintmax_t)entry->start,
504 (uintmax_t)entry->end));
506 if (offset + buflen1 > size)
507 buflen1 = size - offset;
508 if (buflen1 > tag->common.maxsegsz)
509 buflen1 = tag->common.maxsegsz;
511 KASSERT(((entry->start + offset) & (tag->common.alignment - 1))
513 ("alignment failed: ctx %p start 0x%jx offset %x "
514 "align 0x%jx", ctx, (uintmax_t)entry->start, offset,
515 (uintmax_t)tag->common.alignment));
516 KASSERT(entry->end <= tag->common.lowaddr ||
517 entry->start >= tag->common.highaddr,
518 ("entry placement failed: ctx %p start 0x%jx end 0x%jx "
519 "lowaddr 0x%jx highaddr 0x%jx", ctx,
520 (uintmax_t)entry->start, (uintmax_t)entry->end,
521 (uintmax_t)tag->common.lowaddr,
522 (uintmax_t)tag->common.highaddr));
523 KASSERT(dmar_test_boundary(entry->start + offset, buflen1,
524 tag->common.boundary),
525 ("boundary failed: ctx %p start 0x%jx end 0x%jx "
526 "boundary 0x%jx", ctx, (uintmax_t)entry->start,
527 (uintmax_t)entry->end, (uintmax_t)tag->common.boundary));
528 KASSERT(buflen1 <= tag->common.maxsegsz,
529 ("segment too large: ctx %p start 0x%jx end 0x%jx "
530 "buflen1 0x%jx maxsegsz 0x%jx", ctx,
531 (uintmax_t)entry->start, (uintmax_t)entry->end,
532 (uintmax_t)buflen1, (uintmax_t)tag->common.maxsegsz));
535 TAILQ_INSERT_TAIL(&map->map_entries, entry, dmamap_link);
536 entry->flags |= DMAR_MAP_ENTRY_MAP;
537 DMAR_CTX_UNLOCK(ctx);
538 TAILQ_INSERT_TAIL(unroll_list, entry, unroll_link);
540 segs[seg].ds_addr = entry->start + offset;
541 segs[seg].ds_len = buflen1;
543 idx += OFF_TO_IDX(trunc_page(offset + buflen1));
545 offset &= DMAR_PAGE_MASK;
554 dmar_bus_dmamap_load_something(struct bus_dma_tag_dmar *tag,
555 struct bus_dmamap_dmar *map, vm_page_t *ma, int offset, bus_size_t buflen,
556 int flags, bus_dma_segment_t *segs, int *segp)
558 struct dmar_ctx *ctx;
559 struct dmar_map_entry *entry, *entry1;
560 struct dmar_map_entries_tailq unroll_list;
564 atomic_add_long(&ctx->loads, 1);
566 TAILQ_INIT(&unroll_list);
567 error = dmar_bus_dmamap_load_something1(tag, map, ma, offset,
568 buflen, flags, segs, segp, &unroll_list);
571 * The busdma interface does not allow us to report
572 * partial buffer load, so unfortunately we have to
573 * revert all work done.
576 TAILQ_FOREACH_SAFE(entry, &unroll_list, unroll_link,
579 * No entries other than what we have created
580 * during the failed run might have been
581 * inserted there in between, since we own ctx
584 TAILQ_REMOVE(&map->map_entries, entry, dmamap_link);
585 TAILQ_REMOVE(&unroll_list, entry, unroll_link);
586 TAILQ_INSERT_TAIL(&ctx->unload_entries, entry,
589 DMAR_CTX_UNLOCK(ctx);
590 taskqueue_enqueue(ctx->dmar->delayed_taskqueue,
594 if (error == ENOMEM && (flags & BUS_DMA_NOWAIT) == 0 &&
597 if (error == EINPROGRESS)
598 dmar_bus_schedule_dmamap(ctx->dmar, map);
603 dmar_bus_dmamap_load_ma(bus_dma_tag_t dmat, bus_dmamap_t map1,
604 struct vm_page **ma, bus_size_t tlen, int ma_offs, int flags,
605 bus_dma_segment_t *segs, int *segp)
607 struct bus_dma_tag_dmar *tag;
608 struct bus_dmamap_dmar *map;
610 tag = (struct bus_dma_tag_dmar *)dmat;
611 map = (struct bus_dmamap_dmar *)map1;
612 return (dmar_bus_dmamap_load_something(tag, map, ma, ma_offs, tlen,
617 dmar_bus_dmamap_load_phys(bus_dma_tag_t dmat, bus_dmamap_t map1,
618 vm_paddr_t buf, bus_size_t buflen, int flags, bus_dma_segment_t *segs,
621 struct bus_dma_tag_dmar *tag;
622 struct bus_dmamap_dmar *map;
624 vm_paddr_t pstart, pend;
625 int error, i, ma_cnt, offset;
627 tag = (struct bus_dma_tag_dmar *)dmat;
628 map = (struct bus_dmamap_dmar *)map1;
629 pstart = trunc_page(buf);
630 pend = round_page(buf + buflen);
631 offset = buf & PAGE_MASK;
632 ma_cnt = OFF_TO_IDX(pend - pstart);
633 ma = malloc(sizeof(vm_page_t) * ma_cnt, M_DEVBUF, map->cansleep ?
634 M_WAITOK : M_NOWAIT);
637 for (i = 0; i < ma_cnt; i++)
638 ma[i] = PHYS_TO_VM_PAGE(pstart + i * PAGE_SIZE);
639 error = dmar_bus_dmamap_load_something(tag, map, ma, offset, buflen,
646 dmar_bus_dmamap_load_buffer(bus_dma_tag_t dmat, bus_dmamap_t map1, void *buf,
647 bus_size_t buflen, pmap_t pmap, int flags, bus_dma_segment_t *segs,
650 struct bus_dma_tag_dmar *tag;
651 struct bus_dmamap_dmar *map;
653 vm_paddr_t pstart, pend, paddr;
654 int error, i, ma_cnt, offset;
656 tag = (struct bus_dma_tag_dmar *)dmat;
657 map = (struct bus_dmamap_dmar *)map1;
658 pstart = trunc_page((vm_offset_t)buf);
659 pend = round_page((vm_offset_t)buf + buflen);
660 offset = (vm_offset_t)buf & PAGE_MASK;
661 ma_cnt = OFF_TO_IDX(pend - pstart);
662 ma = malloc(sizeof(vm_page_t) * ma_cnt, M_DEVBUF, map->cansleep ?
663 M_WAITOK : M_NOWAIT);
668 * If dumping, do not attempt to call
669 * PHYS_TO_VM_PAGE() at all. It may return non-NULL
670 * but the vm_page returned might be not initialized,
671 * e.g. for the kernel itself.
673 KASSERT(pmap == kernel_pmap, ("non-kernel address write"));
674 fma = malloc(sizeof(struct vm_page) * ma_cnt, M_DEVBUF,
675 M_ZERO | (map->cansleep ? M_WAITOK : M_NOWAIT));
680 for (i = 0; i < ma_cnt; i++, pstart += PAGE_SIZE) {
681 paddr = pmap_kextract(pstart);
682 vm_page_initfake(&fma[i], paddr, VM_MEMATTR_DEFAULT);
687 for (i = 0; i < ma_cnt; i++, pstart += PAGE_SIZE) {
688 if (pmap == kernel_pmap)
689 paddr = pmap_kextract(pstart);
691 paddr = pmap_extract(pmap, pstart);
692 ma[i] = PHYS_TO_VM_PAGE(paddr);
693 KASSERT(VM_PAGE_TO_PHYS(ma[i]) == paddr,
694 ("PHYS_TO_VM_PAGE failed %jx %jx m %p",
695 (uintmax_t)paddr, (uintmax_t)VM_PAGE_TO_PHYS(ma[i]),
699 error = dmar_bus_dmamap_load_something(tag, map, ma, offset, buflen,
707 dmar_bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map1,
708 struct memdesc *mem, bus_dmamap_callback_t *callback, void *callback_arg)
710 struct bus_dmamap_dmar *map;
714 map = (struct bus_dmamap_dmar *)map1;
716 map->tag = (struct bus_dma_tag_dmar *)dmat;
717 map->callback = callback;
718 map->callback_arg = callback_arg;
721 static bus_dma_segment_t *
722 dmar_bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map1,
723 bus_dma_segment_t *segs, int nsegs, int error)
725 struct bus_dma_tag_dmar *tag;
726 struct bus_dmamap_dmar *map;
728 tag = (struct bus_dma_tag_dmar *)dmat;
729 map = (struct bus_dmamap_dmar *)map1;
732 KASSERT(map->cansleep,
733 ("map not locked and not sleepable context %p", map));
736 * We are called from the delayed context. Relock the
739 (tag->common.lockfunc)(tag->common.lockfuncarg, BUS_DMA_LOCK);
744 segs = tag->segments;
749 * The limitations of busdma KPI forces the dmar to perform the actual
750 * unload, consisting of the unmapping of the map entries page tables,
751 * from the delayed context on i386, since page table page mapping
752 * might require a sleep to be successfull. The unfortunate
753 * consequence is that the DMA requests can be served some time after
754 * the bus_dmamap_unload() call returned.
756 * On amd64, we assume that sf allocation cannot fail.
759 dmar_bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map1)
761 struct bus_dma_tag_dmar *tag;
762 struct bus_dmamap_dmar *map;
763 struct dmar_ctx *ctx;
764 #if defined(__amd64__)
765 struct dmar_map_entries_tailq entries;
768 tag = (struct bus_dma_tag_dmar *)dmat;
769 map = (struct bus_dmamap_dmar *)map1;
771 atomic_add_long(&ctx->unloads, 1);
773 #if defined(__i386__)
775 TAILQ_CONCAT(&ctx->unload_entries, &map->map_entries, dmamap_link);
776 DMAR_CTX_UNLOCK(ctx);
777 taskqueue_enqueue(ctx->dmar->delayed_taskqueue, &ctx->unload_task);
778 #else /* defined(__amd64__) */
779 TAILQ_INIT(&entries);
781 TAILQ_CONCAT(&entries, &map->map_entries, dmamap_link);
782 DMAR_CTX_UNLOCK(ctx);
783 THREAD_NO_SLEEPING();
784 dmar_ctx_unload(ctx, &entries, false);
785 THREAD_SLEEPING_OK();
786 KASSERT(TAILQ_EMPTY(&entries), ("lazy dmar_ctx_unload %p", ctx));
791 dmar_bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map,
796 struct bus_dma_impl bus_dma_dmar_impl = {
797 .tag_create = dmar_bus_dma_tag_create,
798 .tag_destroy = dmar_bus_dma_tag_destroy,
799 .map_create = dmar_bus_dmamap_create,
800 .map_destroy = dmar_bus_dmamap_destroy,
801 .mem_alloc = dmar_bus_dmamem_alloc,
802 .mem_free = dmar_bus_dmamem_free,
803 .load_phys = dmar_bus_dmamap_load_phys,
804 .load_buffer = dmar_bus_dmamap_load_buffer,
805 .load_ma = dmar_bus_dmamap_load_ma,
806 .map_waitok = dmar_bus_dmamap_waitok,
807 .map_complete = dmar_bus_dmamap_complete,
808 .map_unload = dmar_bus_dmamap_unload,
809 .map_sync = dmar_bus_dmamap_sync
813 dmar_bus_task_dmamap(void *arg, int pending)
815 struct bus_dma_tag_dmar *tag;
816 struct bus_dmamap_dmar *map;
817 struct dmar_unit *unit;
821 while ((map = TAILQ_FIRST(&unit->delayed_maps)) != NULL) {
822 TAILQ_REMOVE(&unit->delayed_maps, map, delay_link);
825 map->cansleep = true;
827 bus_dmamap_load_mem((bus_dma_tag_t)tag, (bus_dmamap_t)map,
828 &map->mem, map->callback, map->callback_arg,
830 map->cansleep = false;
832 (tag->common.lockfunc)(tag->common.lockfuncarg,
836 map->cansleep = false;
843 dmar_bus_schedule_dmamap(struct dmar_unit *unit, struct bus_dmamap_dmar *map)
848 TAILQ_INSERT_TAIL(&unit->delayed_maps, map, delay_link);
850 taskqueue_enqueue(unit->delayed_taskqueue, &unit->dmamap_load_task);
854 dmar_init_busdma(struct dmar_unit *unit)
857 TAILQ_INIT(&unit->delayed_maps);
858 TASK_INIT(&unit->dmamap_load_task, 0, dmar_bus_task_dmamap, unit);
859 unit->delayed_taskqueue = taskqueue_create("dmar", M_WAITOK,
860 taskqueue_thread_enqueue, &unit->delayed_taskqueue);
861 taskqueue_start_threads(&unit->delayed_taskqueue, 1, PI_DISK,
862 "dmar%d busdma taskq", unit->unit);
867 dmar_fini_busdma(struct dmar_unit *unit)
870 if (unit->delayed_taskqueue == NULL)
873 taskqueue_drain(unit->delayed_taskqueue, &unit->dmamap_load_task);
874 taskqueue_free(unit->delayed_taskqueue);
875 unit->delayed_taskqueue = NULL;