diff --git a/sys/dev/iommu/busdma_iommu.c b/sys/dev/iommu/busdma_iommu.c index 33549b4af011..4accafc64377 100644 --- a/sys/dev/iommu/busdma_iommu.c +++ b/sys/dev/iommu/busdma_iommu.c @@ -1,1097 +1,1084 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2013 The FreeBSD Foundation * * This software was developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * busdma_iommu.c, the implementation of the busdma(9) interface using * IOMMU units from Intel VT-d. */ static bool iommu_bus_dma_is_dev_disabled(int domain, int bus, int slot, int func) { char str[128], *env; int default_bounce; bool ret; static const char bounce_str[] = "bounce"; static const char iommu_str[] = "iommu"; static const char dmar_str[] = "dmar"; /* compatibility */ default_bounce = 0; env = kern_getenv("hw.busdma.default"); if (env != NULL) { if (strcmp(env, bounce_str) == 0) default_bounce = 1; else if (strcmp(env, iommu_str) == 0 || strcmp(env, dmar_str) == 0) default_bounce = 0; freeenv(env); } snprintf(str, sizeof(str), "hw.busdma.pci%d.%d.%d.%d", domain, bus, slot, func); env = kern_getenv(str); if (env == NULL) return (default_bounce != 0); if (strcmp(env, bounce_str) == 0) ret = true; else if (strcmp(env, iommu_str) == 0 || strcmp(env, dmar_str) == 0) ret = false; else ret = default_bounce != 0; freeenv(env); return (ret); } /* * Given original device, find the requester ID that will be seen by * the IOMMU unit and used for page table lookup. PCI bridges may take * ownership of transactions from downstream devices, so it may not be * the same as the BSF of the target device. In those cases, all * devices downstream of the bridge must share a single mapping * domain, and must collectively be assigned to use either IOMMU or * bounce mapping. */ device_t iommu_get_requester(device_t dev, uint16_t *rid) { devclass_t pci_class; device_t l, pci, pcib, pcip, pcibp, requester; int cap_offset; uint16_t pcie_flags; bool bridge_is_pcie; pci_class = devclass_find("pci"); l = requester = dev; *rid = pci_get_rid(dev); /* * Walk the bridge hierarchy from the target device to the * host port to find the translating bridge nearest the IOMMU * unit. */ for (;;) { pci = device_get_parent(l); KASSERT(pci != NULL, ("iommu_get_requester(%s): NULL parent " "for %s", device_get_name(dev), device_get_name(l))); KASSERT(device_get_devclass(pci) == pci_class, ("iommu_get_requester(%s): non-pci parent %s for %s", device_get_name(dev), device_get_name(pci), device_get_name(l))); pcib = device_get_parent(pci); KASSERT(pcib != NULL, ("iommu_get_requester(%s): NULL bridge " "for %s", device_get_name(dev), device_get_name(pci))); /* * The parent of our "bridge" isn't another PCI bus, * so pcib isn't a PCI->PCI bridge but rather a host * port, and the requester ID won't be translated * further. */ pcip = device_get_parent(pcib); if (device_get_devclass(pcip) != pci_class) break; pcibp = device_get_parent(pcip); if (pci_find_cap(l, PCIY_EXPRESS, &cap_offset) == 0) { /* * Do not stop the loop even if the target * device is PCIe, because it is possible (but * unlikely) to have a PCI->PCIe bridge * somewhere in the hierarchy. */ l = pcib; } else { /* * Device is not PCIe, it cannot be seen as a * requester by IOMMU unit. Check whether the * bridge is PCIe. */ bridge_is_pcie = pci_find_cap(pcib, PCIY_EXPRESS, &cap_offset) == 0; requester = pcib; /* * Check for a buggy PCIe/PCI bridge that * doesn't report the express capability. If * the bridge above it is express but isn't a * PCI bridge, then we know pcib is actually a * PCIe/PCI bridge. */ if (!bridge_is_pcie && pci_find_cap(pcibp, PCIY_EXPRESS, &cap_offset) == 0) { pcie_flags = pci_read_config(pcibp, cap_offset + PCIER_FLAGS, 2); if ((pcie_flags & PCIEM_FLAGS_TYPE) != PCIEM_TYPE_PCI_BRIDGE) bridge_is_pcie = true; } if (bridge_is_pcie) { /* * The current device is not PCIe, but * the bridge above it is. This is a * PCIe->PCI bridge. Assume that the * requester ID will be the secondary * bus number with slot and function * set to zero. * * XXX: Doesn't handle the case where * the bridge is PCIe->PCI-X, and the * bridge will only take ownership of * requests in some cases. We should * provide context entries with the * same page tables for taken and * non-taken transactions. */ *rid = PCI_RID(pci_get_bus(l), 0, 0); l = pcibp; } else { /* * Neither the device nor the bridge * above it are PCIe. This is a * conventional PCI->PCI bridge, which * will use the bridge's BSF as the * requester ID. */ *rid = pci_get_rid(pcib); l = pcib; } } } return (requester); } struct iommu_ctx * iommu_instantiate_ctx(struct iommu_unit *unit, device_t dev, bool rmrr) { device_t requester; struct iommu_ctx *ctx; bool disabled; uint16_t rid; requester = iommu_get_requester(dev, &rid); /* * If the user requested the IOMMU disabled for the device, we * cannot disable the IOMMU unit, due to possibility of other * devices on the same IOMMU unit still requiring translation. * Instead provide the identity mapping for the device * context. */ disabled = iommu_bus_dma_is_dev_disabled(pci_get_domain(requester), pci_get_bus(requester), pci_get_slot(requester), pci_get_function(requester)); ctx = iommu_get_ctx(unit, requester, rid, disabled, rmrr); if (ctx == NULL) return (NULL); if (disabled) { /* * Keep the first reference on context, release the * later refs. */ IOMMU_LOCK(unit); if ((ctx->flags & IOMMU_CTX_DISABLED) == 0) { ctx->flags |= IOMMU_CTX_DISABLED; IOMMU_UNLOCK(unit); } else { iommu_free_ctx_locked(unit, ctx); } ctx = NULL; } return (ctx); } struct iommu_ctx * iommu_get_dev_ctx(device_t dev) { struct iommu_unit *unit; unit = iommu_find(dev, bootverbose); /* Not in scope of any IOMMU ? */ if (unit == NULL) return (NULL); if (!unit->dma_enabled) return (NULL); #if defined(__amd64__) || defined(__i386__) dmar_quirks_pre_use(unit); dmar_instantiate_rmrr_ctxs(unit); #endif return (iommu_instantiate_ctx(unit, dev, false)); } bus_dma_tag_t iommu_get_dma_tag(device_t dev, device_t child) { struct iommu_ctx *ctx; bus_dma_tag_t res; ctx = iommu_get_dev_ctx(child); if (ctx == NULL) return (NULL); res = (bus_dma_tag_t)ctx->tag; return (res); } bool bus_dma_iommu_set_buswide(device_t dev) { struct iommu_unit *unit; device_t parent; u_int busno, slot, func; parent = device_get_parent(dev); if (device_get_devclass(parent) != devclass_find("pci")) return (false); unit = iommu_find(dev, bootverbose); if (unit == NULL) return (false); busno = pci_get_bus(dev); slot = pci_get_slot(dev); func = pci_get_function(dev); if (slot != 0 || func != 0) { if (bootverbose) { device_printf(dev, "iommu%d pci%d:%d:%d requested buswide busdma\n", unit->unit, busno, slot, func); } return (false); } iommu_set_buswide_ctx(unit, busno); return (true); } void iommu_set_buswide_ctx(struct iommu_unit *unit, u_int busno) { MPASS(busno <= PCI_BUSMAX); IOMMU_LOCK(unit); unit->buswide_ctxs[busno / NBBY / sizeof(uint32_t)] |= 1 << (busno % (NBBY * sizeof(uint32_t))); IOMMU_UNLOCK(unit); } bool iommu_is_buswide_ctx(struct iommu_unit *unit, u_int busno) { MPASS(busno <= PCI_BUSMAX); return ((unit->buswide_ctxs[busno / NBBY / sizeof(uint32_t)] & (1U << (busno % (NBBY * sizeof(uint32_t))))) != 0); } static MALLOC_DEFINE(M_IOMMU_DMAMAP, "iommu_dmamap", "IOMMU DMA Map"); static void iommu_bus_schedule_dmamap(struct iommu_unit *unit, struct bus_dmamap_iommu *map); static int iommu_bus_dma_tag_create(bus_dma_tag_t parent, bus_size_t alignment, bus_addr_t boundary, bus_addr_t lowaddr, bus_addr_t highaddr, bus_dma_filter_t *filter, void *filterarg, bus_size_t maxsize, int nsegments, bus_size_t maxsegsz, int flags, bus_dma_lock_t *lockfunc, void *lockfuncarg, bus_dma_tag_t *dmat) { struct bus_dma_tag_iommu *newtag, *oldtag; int error; *dmat = NULL; error = common_bus_dma_tag_create(parent != NULL ? &((struct bus_dma_tag_iommu *)parent)->common : NULL, alignment, boundary, lowaddr, highaddr, filter, filterarg, maxsize, nsegments, maxsegsz, flags, lockfunc, lockfuncarg, sizeof(struct bus_dma_tag_iommu), (void **)&newtag); if (error != 0) goto out; oldtag = (struct bus_dma_tag_iommu *)parent; newtag->common.impl = &bus_dma_iommu_impl; newtag->ctx = oldtag->ctx; newtag->owner = oldtag->owner; *dmat = (bus_dma_tag_t)newtag; out: CTR4(KTR_BUSDMA, "%s returned tag %p tag flags 0x%x error %d", __func__, newtag, (newtag != NULL ? newtag->common.flags : 0), error); return (error); } static int iommu_bus_dma_tag_set_domain(bus_dma_tag_t dmat) { return (0); } static int iommu_bus_dma_tag_destroy(bus_dma_tag_t dmat1) { struct bus_dma_tag_iommu *dmat, *parent; struct bus_dma_tag_iommu *dmat_copy __unused; int error; error = 0; dmat_copy = dmat = (struct bus_dma_tag_iommu *)dmat1; if (dmat != NULL) { if (dmat->map_count != 0) { error = EBUSY; goto out; } while (dmat != NULL) { parent = (struct bus_dma_tag_iommu *)dmat->common.parent; if (atomic_fetchadd_int(&dmat->common.ref_count, -1) == 1) { if (dmat == dmat->ctx->tag) iommu_free_ctx(dmat->ctx); free(dmat->segments, M_IOMMU_DMAMAP); free(dmat, M_DEVBUF); dmat = parent; } else dmat = NULL; } } out: CTR3(KTR_BUSDMA, "%s tag %p error %d", __func__, dmat_copy, error); return (error); } static bool iommu_bus_dma_id_mapped(bus_dma_tag_t dmat, vm_paddr_t buf, bus_size_t buflen) { return (false); } static int iommu_bus_dmamap_create(bus_dma_tag_t dmat, int flags, bus_dmamap_t *mapp) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; tag = (struct bus_dma_tag_iommu *)dmat; map = malloc_domainset(sizeof(*map), M_IOMMU_DMAMAP, DOMAINSET_PREF(tag->common.domain), M_NOWAIT | M_ZERO); if (map == NULL) { *mapp = NULL; return (ENOMEM); } if (tag->segments == NULL) { tag->segments = malloc_domainset(sizeof(bus_dma_segment_t) * tag->common.nsegments, M_IOMMU_DMAMAP, DOMAINSET_PREF(tag->common.domain), M_NOWAIT); if (tag->segments == NULL) { free(map, M_IOMMU_DMAMAP); *mapp = NULL; return (ENOMEM); } } IOMMU_DMAMAP_INIT(map); TAILQ_INIT(&map->map_entries); map->tag = tag; map->locked = true; map->cansleep = false; tag->map_count++; *mapp = (bus_dmamap_t)map; return (0); } static int iommu_bus_dmamap_destroy(bus_dma_tag_t dmat, bus_dmamap_t map1) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; tag = (struct bus_dma_tag_iommu *)dmat; map = (struct bus_dmamap_iommu *)map1; if (map != NULL) { IOMMU_DMAMAP_LOCK(map); if (!TAILQ_EMPTY(&map->map_entries)) { IOMMU_DMAMAP_UNLOCK(map); return (EBUSY); } IOMMU_DMAMAP_DESTROY(map); free(map, M_IOMMU_DMAMAP); } tag->map_count--; return (0); } static int iommu_bus_dmamem_alloc(bus_dma_tag_t dmat, void** vaddr, int flags, bus_dmamap_t *mapp) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; int error, mflags; vm_memattr_t attr; error = iommu_bus_dmamap_create(dmat, flags, mapp); if (error != 0) return (error); mflags = (flags & BUS_DMA_NOWAIT) != 0 ? M_NOWAIT : M_WAITOK; mflags |= (flags & BUS_DMA_ZERO) != 0 ? M_ZERO : 0; attr = (flags & BUS_DMA_NOCACHE) != 0 ? VM_MEMATTR_UNCACHEABLE : VM_MEMATTR_DEFAULT; tag = (struct bus_dma_tag_iommu *)dmat; map = (struct bus_dmamap_iommu *)*mapp; if (tag->common.maxsize < PAGE_SIZE && tag->common.alignment <= tag->common.maxsize && attr == VM_MEMATTR_DEFAULT) { *vaddr = malloc_domainset(tag->common.maxsize, M_DEVBUF, DOMAINSET_PREF(tag->common.domain), mflags); map->flags |= BUS_DMAMAP_IOMMU_MALLOC; } else { *vaddr = (void *)kmem_alloc_attr_domainset( DOMAINSET_PREF(tag->common.domain), tag->common.maxsize, mflags, 0ul, BUS_SPACE_MAXADDR, attr); map->flags |= BUS_DMAMAP_IOMMU_KMEM_ALLOC; } if (*vaddr == NULL) { iommu_bus_dmamap_destroy(dmat, *mapp); *mapp = NULL; return (ENOMEM); } return (0); } static void iommu_bus_dmamem_free(bus_dma_tag_t dmat, void *vaddr, bus_dmamap_t map1) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; tag = (struct bus_dma_tag_iommu *)dmat; map = (struct bus_dmamap_iommu *)map1; if ((map->flags & BUS_DMAMAP_IOMMU_MALLOC) != 0) { free(vaddr, M_DEVBUF); map->flags &= ~BUS_DMAMAP_IOMMU_MALLOC; } else { KASSERT((map->flags & BUS_DMAMAP_IOMMU_KMEM_ALLOC) != 0, ("iommu_bus_dmamem_free for non alloced map %p", map)); kmem_free((vm_offset_t)vaddr, tag->common.maxsize); map->flags &= ~BUS_DMAMAP_IOMMU_KMEM_ALLOC; } iommu_bus_dmamap_destroy(dmat, map1); } static int iommu_bus_dmamap_load_something1(struct bus_dma_tag_iommu *tag, struct bus_dmamap_iommu *map, vm_page_t *ma, int offset, bus_size_t buflen, int flags, bus_dma_segment_t *segs, int *segp, - struct iommu_map_entries_tailq *unroll_list) + struct iommu_map_entries_tailq *entries) { struct iommu_ctx *ctx; struct iommu_domain *domain; struct iommu_map_entry *entry; bus_size_t buflen1; int error, e_flags, idx, gas_flags, seg; KASSERT(offset < IOMMU_PAGE_SIZE, ("offset %d", offset)); if (segs == NULL) segs = tag->segments; ctx = tag->ctx; domain = ctx->domain; e_flags = IOMMU_MAP_ENTRY_READ | ((flags & BUS_DMA_NOWRITE) == 0 ? IOMMU_MAP_ENTRY_WRITE : 0); seg = *segp; error = 0; idx = 0; while (buflen > 0) { seg++; if (seg >= tag->common.nsegments) { error = EFBIG; break; } buflen1 = buflen > tag->common.maxsegsz ? tag->common.maxsegsz : buflen; /* * (Too) optimistically allow split if there are more * then one segments left. */ gas_flags = map->cansleep ? IOMMU_MF_CANWAIT : 0; if (seg + 1 < tag->common.nsegments) gas_flags |= IOMMU_MF_CANSPLIT; error = iommu_gas_map(domain, &tag->common, buflen1, offset, e_flags, gas_flags, ma + idx, &entry); if (error != 0) break; /* Update buflen1 in case buffer split. */ if (buflen1 > entry->end - entry->start - offset) buflen1 = entry->end - entry->start - offset; KASSERT(vm_addr_align_ok(entry->start + offset, tag->common.alignment), ("alignment failed: ctx %p start 0x%jx offset %x " "align 0x%jx", ctx, (uintmax_t)entry->start, offset, (uintmax_t)tag->common.alignment)); KASSERT(entry->end <= tag->common.lowaddr || entry->start >= tag->common.highaddr, ("entry placement failed: ctx %p start 0x%jx end 0x%jx " "lowaddr 0x%jx highaddr 0x%jx", ctx, (uintmax_t)entry->start, (uintmax_t)entry->end, (uintmax_t)tag->common.lowaddr, (uintmax_t)tag->common.highaddr)); KASSERT(vm_addr_bound_ok(entry->start + offset, buflen1, tag->common.boundary), ("boundary failed: ctx %p start 0x%jx end 0x%jx " "boundary 0x%jx", ctx, (uintmax_t)entry->start, (uintmax_t)entry->end, (uintmax_t)tag->common.boundary)); KASSERT(buflen1 <= tag->common.maxsegsz, ("segment too large: ctx %p start 0x%jx end 0x%jx " "buflen1 0x%jx maxsegsz 0x%jx", ctx, (uintmax_t)entry->start, (uintmax_t)entry->end, (uintmax_t)buflen1, (uintmax_t)tag->common.maxsegsz)); KASSERT((entry->flags & IOMMU_MAP_ENTRY_MAP) != 0, ("entry %p missing IOMMU_MAP_ENTRY_MAP", entry)); - IOMMU_DMAMAP_LOCK(map); - TAILQ_INSERT_TAIL(&map->map_entries, entry, dmamap_link); - IOMMU_DMAMAP_UNLOCK(map); - TAILQ_INSERT_TAIL(unroll_list, entry, unroll_link); + TAILQ_INSERT_TAIL(entries, entry, dmamap_link); segs[seg].ds_addr = entry->start + offset; segs[seg].ds_len = buflen1; idx += OFF_TO_IDX(offset + buflen1); offset += buflen1; offset &= IOMMU_PAGE_MASK; buflen -= buflen1; } if (error == 0) *segp = seg; return (error); } static int iommu_bus_dmamap_load_something(struct bus_dma_tag_iommu *tag, struct bus_dmamap_iommu *map, vm_page_t *ma, int offset, bus_size_t buflen, int flags, bus_dma_segment_t *segs, int *segp) { struct iommu_ctx *ctx; struct iommu_domain *domain; - struct iommu_map_entry *entry; - struct iommu_map_entries_tailq entries, unroll_list; + struct iommu_map_entries_tailq entries; int error; ctx = tag->ctx; domain = ctx->domain; atomic_add_long(&ctx->loads, 1); - TAILQ_INIT(&unroll_list); + TAILQ_INIT(&entries); error = iommu_bus_dmamap_load_something1(tag, map, ma, offset, - buflen, flags, segs, segp, &unroll_list); - if (error != 0 && !TAILQ_EMPTY(&unroll_list)) { + buflen, flags, segs, segp, &entries); + if (error == 0) { + IOMMU_DMAMAP_LOCK(map); + TAILQ_CONCAT(&map->map_entries, &entries, dmamap_link); + IOMMU_DMAMAP_UNLOCK(map); + } else if (!TAILQ_EMPTY(&entries)) { /* * The busdma interface does not allow us to report * partial buffer load, so unfortunately we have to * revert all work done. */ - TAILQ_INIT(&entries); - IOMMU_DMAMAP_LOCK(map); - TAILQ_FOREACH(entry, &unroll_list, unroll_link) { - /* - * No entries other than what we have created - * during the failed run might have been - * inserted there in between, since we own ctx - * pglock. - */ - TAILQ_REMOVE(&map->map_entries, entry, dmamap_link); - TAILQ_INSERT_TAIL(&entries, entry, dmamap_link); - } - IOMMU_DMAMAP_UNLOCK(map); IOMMU_DOMAIN_LOCK(domain); TAILQ_CONCAT(&domain->unload_entries, &entries, dmamap_link); IOMMU_DOMAIN_UNLOCK(domain); taskqueue_enqueue(domain->iommu->delayed_taskqueue, &domain->unload_task); } if (error == ENOMEM && (flags & BUS_DMA_NOWAIT) == 0 && !map->cansleep) error = EINPROGRESS; if (error == EINPROGRESS) iommu_bus_schedule_dmamap(domain->iommu, map); return (error); } static int iommu_bus_dmamap_load_ma(bus_dma_tag_t dmat, bus_dmamap_t map1, struct vm_page **ma, bus_size_t tlen, int ma_offs, int flags, bus_dma_segment_t *segs, int *segp) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; tag = (struct bus_dma_tag_iommu *)dmat; map = (struct bus_dmamap_iommu *)map1; return (iommu_bus_dmamap_load_something(tag, map, ma, ma_offs, tlen, flags, segs, segp)); } static int iommu_bus_dmamap_load_phys(bus_dma_tag_t dmat, bus_dmamap_t map1, vm_paddr_t buf, bus_size_t buflen, int flags, bus_dma_segment_t *segs, int *segp) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; vm_page_t *ma, fma; vm_paddr_t pstart, pend, paddr; int error, i, ma_cnt, mflags, offset; tag = (struct bus_dma_tag_iommu *)dmat; map = (struct bus_dmamap_iommu *)map1; pstart = trunc_page(buf); pend = round_page(buf + buflen); offset = buf & PAGE_MASK; ma_cnt = OFF_TO_IDX(pend - pstart); mflags = map->cansleep ? M_WAITOK : M_NOWAIT; ma = malloc(sizeof(vm_page_t) * ma_cnt, M_DEVBUF, mflags); if (ma == NULL) return (ENOMEM); fma = NULL; for (i = 0; i < ma_cnt; i++) { paddr = pstart + ptoa(i); ma[i] = PHYS_TO_VM_PAGE(paddr); if (ma[i] == NULL || VM_PAGE_TO_PHYS(ma[i]) != paddr) { /* * If PHYS_TO_VM_PAGE() returned NULL or the * vm_page was not initialized we'll use a * fake page. */ if (fma == NULL) { fma = malloc(sizeof(struct vm_page) * ma_cnt, M_DEVBUF, M_ZERO | mflags); if (fma == NULL) { free(ma, M_DEVBUF); return (ENOMEM); } } vm_page_initfake(&fma[i], pstart + ptoa(i), VM_MEMATTR_DEFAULT); ma[i] = &fma[i]; } } error = iommu_bus_dmamap_load_something(tag, map, ma, offset, buflen, flags, segs, segp); free(fma, M_DEVBUF); free(ma, M_DEVBUF); return (error); } static int iommu_bus_dmamap_load_buffer(bus_dma_tag_t dmat, bus_dmamap_t map1, void *buf, bus_size_t buflen, pmap_t pmap, int flags, bus_dma_segment_t *segs, int *segp) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; vm_page_t *ma, fma; vm_paddr_t pstart, pend, paddr; int error, i, ma_cnt, mflags, offset; tag = (struct bus_dma_tag_iommu *)dmat; map = (struct bus_dmamap_iommu *)map1; pstart = trunc_page((vm_offset_t)buf); pend = round_page((vm_offset_t)buf + buflen); offset = (vm_offset_t)buf & PAGE_MASK; ma_cnt = OFF_TO_IDX(pend - pstart); mflags = map->cansleep ? M_WAITOK : M_NOWAIT; ma = malloc(sizeof(vm_page_t) * ma_cnt, M_DEVBUF, mflags); if (ma == NULL) return (ENOMEM); fma = NULL; for (i = 0; i < ma_cnt; i++, pstart += PAGE_SIZE) { if (pmap == kernel_pmap) paddr = pmap_kextract(pstart); else paddr = pmap_extract(pmap, pstart); ma[i] = PHYS_TO_VM_PAGE(paddr); if (ma[i] == NULL || VM_PAGE_TO_PHYS(ma[i]) != paddr) { /* * If PHYS_TO_VM_PAGE() returned NULL or the * vm_page was not initialized we'll use a * fake page. */ if (fma == NULL) { fma = malloc(sizeof(struct vm_page) * ma_cnt, M_DEVBUF, M_ZERO | mflags); if (fma == NULL) { free(ma, M_DEVBUF); return (ENOMEM); } } vm_page_initfake(&fma[i], paddr, VM_MEMATTR_DEFAULT); ma[i] = &fma[i]; } } error = iommu_bus_dmamap_load_something(tag, map, ma, offset, buflen, flags, segs, segp); free(ma, M_DEVBUF); free(fma, M_DEVBUF); return (error); } static void iommu_bus_dmamap_waitok(bus_dma_tag_t dmat, bus_dmamap_t map1, struct memdesc *mem, bus_dmamap_callback_t *callback, void *callback_arg) { struct bus_dmamap_iommu *map; if (map1 == NULL) return; map = (struct bus_dmamap_iommu *)map1; map->mem = *mem; map->tag = (struct bus_dma_tag_iommu *)dmat; map->callback = callback; map->callback_arg = callback_arg; } static bus_dma_segment_t * iommu_bus_dmamap_complete(bus_dma_tag_t dmat, bus_dmamap_t map1, bus_dma_segment_t *segs, int nsegs, int error) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; tag = (struct bus_dma_tag_iommu *)dmat; map = (struct bus_dmamap_iommu *)map1; if (!map->locked) { KASSERT(map->cansleep, ("map not locked and not sleepable context %p", map)); /* * We are called from the delayed context. Relock the * driver. */ (tag->common.lockfunc)(tag->common.lockfuncarg, BUS_DMA_LOCK); map->locked = true; } if (segs == NULL) segs = tag->segments; return (segs); } /* * The limitations of busdma KPI forces the iommu to perform the actual * unload, consisting of the unmapping of the map entries page tables, * from the delayed context on i386, since page table page mapping * might require a sleep to be successfull. The unfortunate * consequence is that the DMA requests can be served some time after * the bus_dmamap_unload() call returned. * * On amd64, we assume that sf allocation cannot fail. */ static void iommu_bus_dmamap_unload(bus_dma_tag_t dmat, bus_dmamap_t map1) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; struct iommu_ctx *ctx; struct iommu_domain *domain; struct iommu_map_entries_tailq entries; tag = (struct bus_dma_tag_iommu *)dmat; map = (struct bus_dmamap_iommu *)map1; ctx = tag->ctx; domain = ctx->domain; atomic_add_long(&ctx->unloads, 1); TAILQ_INIT(&entries); IOMMU_DMAMAP_LOCK(map); TAILQ_CONCAT(&entries, &map->map_entries, dmamap_link); IOMMU_DMAMAP_UNLOCK(map); #if defined(IOMMU_DOMAIN_UNLOAD_SLEEP) IOMMU_DOMAIN_LOCK(domain); TAILQ_CONCAT(&domain->unload_entries, &entries, dmamap_link); IOMMU_DOMAIN_UNLOCK(domain); taskqueue_enqueue(domain->iommu->delayed_taskqueue, &domain->unload_task); #else THREAD_NO_SLEEPING(); iommu_domain_unload(domain, &entries, false); THREAD_SLEEPING_OK(); KASSERT(TAILQ_EMPTY(&entries), ("lazy iommu_ctx_unload %p", ctx)); #endif } static void iommu_bus_dmamap_sync(bus_dma_tag_t dmat, bus_dmamap_t map, bus_dmasync_op_t op) { } struct bus_dma_impl bus_dma_iommu_impl = { .tag_create = iommu_bus_dma_tag_create, .tag_destroy = iommu_bus_dma_tag_destroy, .tag_set_domain = iommu_bus_dma_tag_set_domain, .id_mapped = iommu_bus_dma_id_mapped, .map_create = iommu_bus_dmamap_create, .map_destroy = iommu_bus_dmamap_destroy, .mem_alloc = iommu_bus_dmamem_alloc, .mem_free = iommu_bus_dmamem_free, .load_phys = iommu_bus_dmamap_load_phys, .load_buffer = iommu_bus_dmamap_load_buffer, .load_ma = iommu_bus_dmamap_load_ma, .map_waitok = iommu_bus_dmamap_waitok, .map_complete = iommu_bus_dmamap_complete, .map_unload = iommu_bus_dmamap_unload, .map_sync = iommu_bus_dmamap_sync, }; static void iommu_bus_task_dmamap(void *arg, int pending) { struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; struct iommu_unit *unit; unit = arg; IOMMU_LOCK(unit); while ((map = TAILQ_FIRST(&unit->delayed_maps)) != NULL) { TAILQ_REMOVE(&unit->delayed_maps, map, delay_link); IOMMU_UNLOCK(unit); tag = map->tag; map->cansleep = true; map->locked = false; bus_dmamap_load_mem((bus_dma_tag_t)tag, (bus_dmamap_t)map, &map->mem, map->callback, map->callback_arg, BUS_DMA_WAITOK); map->cansleep = false; if (map->locked) { (tag->common.lockfunc)(tag->common.lockfuncarg, BUS_DMA_UNLOCK); } else map->locked = true; map->cansleep = false; IOMMU_LOCK(unit); } IOMMU_UNLOCK(unit); } static void iommu_bus_schedule_dmamap(struct iommu_unit *unit, struct bus_dmamap_iommu *map) { map->locked = false; IOMMU_LOCK(unit); TAILQ_INSERT_TAIL(&unit->delayed_maps, map, delay_link); IOMMU_UNLOCK(unit); taskqueue_enqueue(unit->delayed_taskqueue, &unit->dmamap_load_task); } int iommu_init_busdma(struct iommu_unit *unit) { int error; unit->dma_enabled = 1; error = TUNABLE_INT_FETCH("hw.iommu.dma", &unit->dma_enabled); if (error == 0) /* compatibility */ TUNABLE_INT_FETCH("hw.dmar.dma", &unit->dma_enabled); TAILQ_INIT(&unit->delayed_maps); TASK_INIT(&unit->dmamap_load_task, 0, iommu_bus_task_dmamap, unit); unit->delayed_taskqueue = taskqueue_create("iommu", M_WAITOK, taskqueue_thread_enqueue, &unit->delayed_taskqueue); taskqueue_start_threads(&unit->delayed_taskqueue, 1, PI_DISK, "iommu%d busdma taskq", unit->unit); return (0); } void iommu_fini_busdma(struct iommu_unit *unit) { if (unit->delayed_taskqueue == NULL) return; taskqueue_drain(unit->delayed_taskqueue, &unit->dmamap_load_task); taskqueue_free(unit->delayed_taskqueue); unit->delayed_taskqueue = NULL; } int bus_dma_iommu_load_ident(bus_dma_tag_t dmat, bus_dmamap_t map1, vm_paddr_t start, vm_size_t length, int flags) { struct bus_dma_tag_common *tc; struct bus_dma_tag_iommu *tag; struct bus_dmamap_iommu *map; struct iommu_ctx *ctx; struct iommu_domain *domain; struct iommu_map_entry *entry; vm_page_t *ma; vm_size_t i; int error; bool waitok; MPASS((start & PAGE_MASK) == 0); MPASS((length & PAGE_MASK) == 0); MPASS(length > 0); MPASS(start + length >= start); MPASS((flags & ~(BUS_DMA_NOWAIT | BUS_DMA_NOWRITE)) == 0); tc = (struct bus_dma_tag_common *)dmat; if (tc->impl != &bus_dma_iommu_impl) return (0); tag = (struct bus_dma_tag_iommu *)dmat; ctx = tag->ctx; domain = ctx->domain; map = (struct bus_dmamap_iommu *)map1; waitok = (flags & BUS_DMA_NOWAIT) != 0; entry = iommu_gas_alloc_entry(domain, waitok ? 0 : IOMMU_PGF_WAITOK); if (entry == NULL) return (ENOMEM); entry->start = start; entry->end = start + length; ma = malloc(sizeof(vm_page_t) * atop(length), M_TEMP, waitok ? M_WAITOK : M_NOWAIT); if (ma == NULL) { iommu_gas_free_entry(domain, entry); return (ENOMEM); } for (i = 0; i < atop(length); i++) { ma[i] = vm_page_getfake(entry->start + PAGE_SIZE * i, VM_MEMATTR_DEFAULT); } error = iommu_gas_map_region(domain, entry, IOMMU_MAP_ENTRY_READ | ((flags & BUS_DMA_NOWRITE) ? 0 : IOMMU_MAP_ENTRY_WRITE) | IOMMU_MAP_ENTRY_MAP, waitok ? IOMMU_MF_CANWAIT : 0, ma); if (error == 0) { IOMMU_DMAMAP_LOCK(map); TAILQ_INSERT_TAIL(&map->map_entries, entry, dmamap_link); IOMMU_DMAMAP_UNLOCK(map); } else { iommu_gas_free_entry(domain, entry); } for (i = 0; i < atop(length); i++) vm_page_putfake(ma[i]); free(ma, M_TEMP); return (error); } static void iommu_domain_unload_task(void *arg, int pending) { struct iommu_domain *domain; struct iommu_map_entries_tailq entries; domain = arg; TAILQ_INIT(&entries); for (;;) { IOMMU_DOMAIN_LOCK(domain); TAILQ_SWAP(&domain->unload_entries, &entries, iommu_map_entry, dmamap_link); IOMMU_DOMAIN_UNLOCK(domain); if (TAILQ_EMPTY(&entries)) break; iommu_domain_unload(domain, &entries, true); } } void iommu_domain_init(struct iommu_unit *unit, struct iommu_domain *domain, const struct iommu_domain_map_ops *ops) { domain->ops = ops; domain->iommu = unit; TASK_INIT(&domain->unload_task, 0, iommu_domain_unload_task, domain); RB_INIT(&domain->rb_root); TAILQ_INIT(&domain->unload_entries); mtx_init(&domain->lock, "iodom", NULL, MTX_DEF); } void iommu_domain_fini(struct iommu_domain *domain) { mtx_destroy(&domain->lock); } diff --git a/sys/dev/iommu/iommu.h b/sys/dev/iommu/iommu.h index 3800213a1d64..62b5659b6e83 100644 --- a/sys/dev/iommu/iommu.h +++ b/sys/dev/iommu/iommu.h @@ -1,202 +1,200 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2013 The FreeBSD Foundation * * This software was developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _DEV_IOMMU_IOMMU_H_ #define _DEV_IOMMU_IOMMU_H_ #include struct bus_dma_tag_common; struct iommu_map_entry; TAILQ_HEAD(iommu_map_entries_tailq, iommu_map_entry); RB_HEAD(iommu_gas_entries_tree, iommu_map_entry); RB_PROTOTYPE(iommu_gas_entries_tree, iommu_map_entry, rb_entry, iommu_gas_cmp_entries); struct iommu_qi_genseq { u_int gen; uint32_t seq; }; struct iommu_map_entry { iommu_gaddr_t start; iommu_gaddr_t end; iommu_gaddr_t first; /* Least start in subtree */ iommu_gaddr_t last; /* Greatest end in subtree */ iommu_gaddr_t free_down; /* Max free space below the current R/B tree node */ u_int flags; TAILQ_ENTRY(iommu_map_entry) dmamap_link; /* Link for dmamap entries */ RB_ENTRY(iommu_map_entry) rb_entry; /* Links for domain entries */ - TAILQ_ENTRY(iommu_map_entry) unroll_link; /* Link for unroll after - dmamap_load failure */ struct iommu_domain *domain; struct iommu_qi_genseq gseq; }; struct iommu_unit { struct mtx lock; device_t dev; int unit; int dma_enabled; /* Busdma delayed map load */ struct task dmamap_load_task; TAILQ_HEAD(, bus_dmamap_iommu) delayed_maps; struct taskqueue *delayed_taskqueue; /* * Bitmap of buses for which context must ignore slot:func, * duplicating the page table pointer into all context table * entries. This is a client-controlled quirk to support some * NTBs. */ uint32_t buswide_ctxs[(PCI_BUSMAX + 1) / NBBY / sizeof(uint32_t)]; }; struct iommu_domain_map_ops { int (*map)(struct iommu_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size, vm_page_t *ma, uint64_t pflags, int flags); int (*unmap)(struct iommu_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size, int flags); }; /* * Locking annotations: * (u) - Protected by iommu unit lock * (d) - Protected by domain lock * (c) - Immutable after initialization */ struct iommu_domain { struct iommu_unit *iommu; /* (c) */ const struct iommu_domain_map_ops *ops; struct mtx lock; /* (c) */ struct task unload_task; /* (c) */ u_int entries_cnt; /* (d) */ struct iommu_map_entries_tailq unload_entries; /* (d) Entries to unload */ struct iommu_gas_entries_tree rb_root; /* (d) */ iommu_gaddr_t end; /* (c) Highest address + 1 in the guest AS */ struct iommu_map_entry *first_place, *last_place; /* (d) */ struct iommu_map_entry *msi_entry; /* (d) Arch-specific */ iommu_gaddr_t msi_base; /* (d) Arch-specific */ vm_paddr_t msi_phys; /* (d) Arch-specific */ u_int flags; /* (u) */ }; struct iommu_ctx { struct iommu_domain *domain; /* (c) */ struct bus_dma_tag_iommu *tag; /* (c) Root tag */ u_long loads; /* atomic updates, for stat only */ u_long unloads; /* same */ u_int flags; /* (u) */ uint16_t rid; /* (c) pci RID */ }; /* struct iommu_ctx flags */ #define IOMMU_CTX_FAULTED 0x0001 /* Fault was reported, last_fault_rec is valid */ #define IOMMU_CTX_DISABLED 0x0002 /* Device is disabled, the ephemeral reference is kept to prevent context destruction */ #define IOMMU_DOMAIN_GAS_INITED 0x0001 #define IOMMU_DOMAIN_PGTBL_INITED 0x0002 #define IOMMU_DOMAIN_IDMAP 0x0010 /* Domain uses identity page table */ #define IOMMU_DOMAIN_RMRR 0x0020 /* Domain contains RMRR entry, cannot be turned off */ #define IOMMU_LOCK(unit) mtx_lock(&(unit)->lock) #define IOMMU_UNLOCK(unit) mtx_unlock(&(unit)->lock) #define IOMMU_ASSERT_LOCKED(unit) mtx_assert(&(unit)->lock, MA_OWNED) #define IOMMU_DOMAIN_LOCK(dom) mtx_lock(&(dom)->lock) #define IOMMU_DOMAIN_UNLOCK(dom) mtx_unlock(&(dom)->lock) #define IOMMU_DOMAIN_ASSERT_LOCKED(dom) mtx_assert(&(dom)->lock, MA_OWNED) void iommu_free_ctx(struct iommu_ctx *ctx); void iommu_free_ctx_locked(struct iommu_unit *iommu, struct iommu_ctx *ctx); struct iommu_ctx *iommu_get_ctx(struct iommu_unit *, device_t dev, uint16_t rid, bool id_mapped, bool rmrr_init); struct iommu_unit *iommu_find(device_t dev, bool verbose); void iommu_domain_unload_entry(struct iommu_map_entry *entry, bool free); void iommu_domain_unload(struct iommu_domain *domain, struct iommu_map_entries_tailq *entries, bool cansleep); struct iommu_ctx *iommu_instantiate_ctx(struct iommu_unit *iommu, device_t dev, bool rmrr); device_t iommu_get_requester(device_t dev, uint16_t *rid); int iommu_init_busdma(struct iommu_unit *unit); void iommu_fini_busdma(struct iommu_unit *unit); void iommu_gas_init_domain(struct iommu_domain *domain); void iommu_gas_fini_domain(struct iommu_domain *domain); struct iommu_map_entry *iommu_gas_alloc_entry(struct iommu_domain *domain, u_int flags); void iommu_gas_free_entry(struct iommu_domain *domain, struct iommu_map_entry *entry); void iommu_gas_free_space(struct iommu_domain *domain, struct iommu_map_entry *entry); int iommu_gas_map(struct iommu_domain *domain, const struct bus_dma_tag_common *common, iommu_gaddr_t size, int offset, u_int eflags, u_int flags, vm_page_t *ma, struct iommu_map_entry **res); void iommu_gas_free_region(struct iommu_domain *domain, struct iommu_map_entry *entry); int iommu_gas_map_region(struct iommu_domain *domain, struct iommu_map_entry *entry, u_int eflags, u_int flags, vm_page_t *ma); int iommu_gas_reserve_region(struct iommu_domain *domain, iommu_gaddr_t start, iommu_gaddr_t end, struct iommu_map_entry **entry0); int iommu_gas_reserve_region_extend(struct iommu_domain *domain, iommu_gaddr_t start, iommu_gaddr_t end); void iommu_set_buswide_ctx(struct iommu_unit *unit, u_int busno); bool iommu_is_buswide_ctx(struct iommu_unit *unit, u_int busno); void iommu_domain_init(struct iommu_unit *unit, struct iommu_domain *domain, const struct iommu_domain_map_ops *ops); void iommu_domain_fini(struct iommu_domain *domain); bool bus_dma_iommu_set_buswide(device_t dev); int bus_dma_iommu_load_ident(bus_dma_tag_t dmat, bus_dmamap_t map, vm_paddr_t start, vm_size_t length, int flags); bus_dma_tag_t iommu_get_dma_tag(device_t dev, device_t child); struct iommu_ctx *iommu_get_dev_ctx(device_t dev); struct iommu_domain *iommu_get_ctx_domain(struct iommu_ctx *ctx); SYSCTL_DECL(_hw_iommu); #endif /* !_DEV_IOMMU_IOMMU_H_ */ diff --git a/sys/x86/iommu/intel_ctx.c b/sys/x86/iommu/intel_ctx.c index b3573712de32..4e94936f27d4 100644 --- a/sys/x86/iommu/intel_ctx.c +++ b/sys/x86/iommu/intel_ctx.c @@ -1,973 +1,973 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2013 The FreeBSD Foundation * * This software was developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_DMAR_CTX, "dmar_ctx", "Intel DMAR Context"); static MALLOC_DEFINE(M_DMAR_DOMAIN, "dmar_dom", "Intel DMAR Domain"); static void dmar_unref_domain_locked(struct dmar_unit *dmar, struct dmar_domain *domain); static void dmar_domain_destroy(struct dmar_domain *domain); static void dmar_ensure_ctx_page(struct dmar_unit *dmar, int bus) { struct sf_buf *sf; dmar_root_entry_t *re; vm_page_t ctxm; /* * Allocated context page must be linked. */ ctxm = dmar_pgalloc(dmar->ctx_obj, 1 + bus, IOMMU_PGF_NOALLOC); if (ctxm != NULL) return; /* * Page not present, allocate and link. Note that other * thread might execute this sequence in parallel. This * should be safe, because the context entries written by both * threads are equal. */ TD_PREP_PINNED_ASSERT; ctxm = dmar_pgalloc(dmar->ctx_obj, 1 + bus, IOMMU_PGF_ZERO | IOMMU_PGF_WAITOK); re = dmar_map_pgtbl(dmar->ctx_obj, 0, IOMMU_PGF_NOALLOC, &sf); re += bus; dmar_pte_store(&re->r1, DMAR_ROOT_R1_P | (DMAR_ROOT_R1_CTP_MASK & VM_PAGE_TO_PHYS(ctxm))); dmar_flush_root_to_ram(dmar, re); dmar_unmap_pgtbl(sf); TD_PINNED_ASSERT; } static dmar_ctx_entry_t * dmar_map_ctx_entry(struct dmar_ctx *ctx, struct sf_buf **sfp) { struct dmar_unit *dmar; dmar_ctx_entry_t *ctxp; dmar = CTX2DMAR(ctx); ctxp = dmar_map_pgtbl(dmar->ctx_obj, 1 + PCI_RID2BUS(ctx->context.rid), IOMMU_PGF_NOALLOC | IOMMU_PGF_WAITOK, sfp); ctxp += ctx->context.rid & 0xff; return (ctxp); } static void device_tag_init(struct dmar_ctx *ctx, device_t dev) { struct dmar_domain *domain; bus_addr_t maxaddr; domain = CTX2DOM(ctx); maxaddr = MIN(domain->iodom.end, BUS_SPACE_MAXADDR); ctx->context.tag->common.ref_count = 1; /* Prevent free */ ctx->context.tag->common.impl = &bus_dma_iommu_impl; ctx->context.tag->common.boundary = 0; ctx->context.tag->common.lowaddr = maxaddr; ctx->context.tag->common.highaddr = maxaddr; ctx->context.tag->common.maxsize = maxaddr; ctx->context.tag->common.nsegments = BUS_SPACE_UNRESTRICTED; ctx->context.tag->common.maxsegsz = maxaddr; ctx->context.tag->ctx = CTX2IOCTX(ctx); ctx->context.tag->owner = dev; } static void ctx_id_entry_init_one(dmar_ctx_entry_t *ctxp, struct dmar_domain *domain, vm_page_t ctx_root) { /* * For update due to move, the store is not atomic. It is * possible that DMAR read upper doubleword, while low * doubleword is not yet updated. The domain id is stored in * the upper doubleword, while the table pointer in the lower. * * There is no good solution, for the same reason it is wrong * to clear P bit in the ctx entry for update. */ dmar_pte_store1(&ctxp->ctx2, DMAR_CTX2_DID(domain->domain) | domain->awlvl); if (ctx_root == NULL) { dmar_pte_store1(&ctxp->ctx1, DMAR_CTX1_T_PASS | DMAR_CTX1_P); } else { dmar_pte_store1(&ctxp->ctx1, DMAR_CTX1_T_UNTR | (DMAR_CTX1_ASR_MASK & VM_PAGE_TO_PHYS(ctx_root)) | DMAR_CTX1_P); } } static void ctx_id_entry_init(struct dmar_ctx *ctx, dmar_ctx_entry_t *ctxp, bool move, int busno) { struct dmar_unit *unit; struct dmar_domain *domain; vm_page_t ctx_root; int i; domain = CTX2DOM(ctx); unit = DOM2DMAR(domain); KASSERT(move || (ctxp->ctx1 == 0 && ctxp->ctx2 == 0), ("dmar%d: initialized ctx entry %d:%d:%d 0x%jx 0x%jx", unit->iommu.unit, busno, pci_get_slot(ctx->context.tag->owner), pci_get_function(ctx->context.tag->owner), ctxp->ctx1, ctxp->ctx2)); if ((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0 && (unit->hw_ecap & DMAR_ECAP_PT) != 0) { KASSERT(domain->pgtbl_obj == NULL, ("ctx %p non-null pgtbl_obj", ctx)); ctx_root = NULL; } else { ctx_root = dmar_pgalloc(domain->pgtbl_obj, 0, IOMMU_PGF_NOALLOC); } if (iommu_is_buswide_ctx(DMAR2IOMMU(unit), busno)) { MPASS(!move); for (i = 0; i <= PCI_BUSMAX; i++) { ctx_id_entry_init_one(&ctxp[i], domain, ctx_root); } } else { ctx_id_entry_init_one(ctxp, domain, ctx_root); } dmar_flush_ctx_to_ram(unit, ctxp); } static int dmar_flush_for_ctx_entry(struct dmar_unit *dmar, bool force) { int error; /* * If dmar declares Caching Mode as Set, follow 11.5 "Caching * Mode Consideration" and do the (global) invalidation of the * negative TLB entries. */ if ((dmar->hw_cap & DMAR_CAP_CM) == 0 && !force) return (0); if (dmar->qi_enabled) { dmar_qi_invalidate_ctx_glob_locked(dmar); if ((dmar->hw_ecap & DMAR_ECAP_DI) != 0 || force) dmar_qi_invalidate_iotlb_glob_locked(dmar); return (0); } error = dmar_inv_ctx_glob(dmar); if (error == 0 && ((dmar->hw_ecap & DMAR_ECAP_DI) != 0 || force)) error = dmar_inv_iotlb_glob(dmar); return (error); } static int domain_init_rmrr(struct dmar_domain *domain, device_t dev, int bus, int slot, int func, int dev_domain, int dev_busno, const void *dev_path, int dev_path_len) { struct iommu_map_entries_tailq rmrr_entries; struct iommu_map_entry *entry, *entry1; vm_page_t *ma; iommu_gaddr_t start, end; vm_pindex_t size, i; int error, error1; error = 0; TAILQ_INIT(&rmrr_entries); dmar_dev_parse_rmrr(domain, dev_domain, dev_busno, dev_path, dev_path_len, &rmrr_entries); - TAILQ_FOREACH_SAFE(entry, &rmrr_entries, unroll_link, entry1) { + TAILQ_FOREACH_SAFE(entry, &rmrr_entries, dmamap_link, entry1) { /* * VT-d specification requires that the start of an * RMRR entry is 4k-aligned. Buggy BIOSes put * anything into the start and end fields. Truncate * and round as neccesary. * * We also allow the overlapping RMRR entries, see * iommu_gas_alloc_region(). */ start = entry->start; end = entry->end; if (bootverbose) printf("dmar%d ctx pci%d:%d:%d RMRR [%#jx, %#jx]\n", domain->iodom.iommu->unit, bus, slot, func, (uintmax_t)start, (uintmax_t)end); entry->start = trunc_page(start); entry->end = round_page(end); if (entry->start == entry->end) { /* Workaround for some AMI (?) BIOSes */ if (bootverbose) { if (dev != NULL) device_printf(dev, ""); printf("pci%d:%d:%d ", bus, slot, func); printf("BIOS bug: dmar%d RMRR " "region (%jx, %jx) corrected\n", domain->iodom.iommu->unit, start, end); } entry->end += DMAR_PAGE_SIZE * 0x20; } size = OFF_TO_IDX(entry->end - entry->start); ma = malloc(sizeof(vm_page_t) * size, M_TEMP, M_WAITOK); for (i = 0; i < size; i++) { ma[i] = vm_page_getfake(entry->start + PAGE_SIZE * i, VM_MEMATTR_DEFAULT); } error1 = iommu_gas_map_region(DOM2IODOM(domain), entry, IOMMU_MAP_ENTRY_READ | IOMMU_MAP_ENTRY_WRITE, IOMMU_MF_CANWAIT | IOMMU_MF_RMRR, ma); /* * Non-failed RMRR entries are owned by context rb * tree. Get rid of the failed entry, but do not stop * the loop. Rest of the parsed RMRR entries are * loaded and removed on the context destruction. */ if (error1 == 0 && entry->end != entry->start) { IOMMU_LOCK(domain->iodom.iommu); domain->refs++; /* XXXKIB prevent free */ domain->iodom.flags |= IOMMU_DOMAIN_RMRR; IOMMU_UNLOCK(domain->iodom.iommu); } else { if (error1 != 0) { if (dev != NULL) device_printf(dev, ""); printf("pci%d:%d:%d ", bus, slot, func); printf( "dmar%d failed to map RMRR region (%jx, %jx) %d\n", domain->iodom.iommu->unit, start, end, error1); error = error1; } - TAILQ_REMOVE(&rmrr_entries, entry, unroll_link); + TAILQ_REMOVE(&rmrr_entries, entry, dmamap_link); iommu_gas_free_entry(DOM2IODOM(domain), entry); } for (i = 0; i < size; i++) vm_page_putfake(ma[i]); free(ma, M_TEMP); } return (error); } /* * PCI memory address space is shared between memory-mapped devices (MMIO) and * host memory (which may be remapped by an IOMMU). Device accesses to an * address within a memory aperture in a PCIe root port will be treated as * peer-to-peer and not forwarded to an IOMMU. To avoid this, reserve the * address space of the root port's memory apertures in the address space used * by the IOMMU for remapping. */ static int dmar_reserve_pci_regions(struct dmar_domain *domain, device_t dev) { struct iommu_domain *iodom; device_t root; uint32_t val; uint64_t base, limit; int error; iodom = DOM2IODOM(domain); root = pci_find_pcie_root_port(dev); if (root == NULL) return (0); /* Disable downstream memory */ base = PCI_PPBMEMBASE(0, pci_read_config(root, PCIR_MEMBASE_1, 2)); limit = PCI_PPBMEMLIMIT(0, pci_read_config(root, PCIR_MEMLIMIT_1, 2)); error = iommu_gas_reserve_region_extend(iodom, base, limit + 1); if (bootverbose || error != 0) device_printf(dev, "DMAR reserve [%#jx-%#jx] (error %d)\n", base, limit + 1, error); if (error != 0) return (error); /* Disable downstream prefetchable memory */ val = pci_read_config(root, PCIR_PMBASEL_1, 2); if (val != 0 || pci_read_config(root, PCIR_PMLIMITL_1, 2) != 0) { if ((val & PCIM_BRPM_MASK) == PCIM_BRPM_64) { base = PCI_PPBMEMBASE( pci_read_config(root, PCIR_PMBASEH_1, 4), val); limit = PCI_PPBMEMLIMIT( pci_read_config(root, PCIR_PMLIMITH_1, 4), pci_read_config(root, PCIR_PMLIMITL_1, 2)); } else { base = PCI_PPBMEMBASE(0, val); limit = PCI_PPBMEMLIMIT(0, pci_read_config(root, PCIR_PMLIMITL_1, 2)); } error = iommu_gas_reserve_region_extend(iodom, base, limit + 1); if (bootverbose || error != 0) device_printf(dev, "DMAR reserve [%#jx-%#jx] " "(error %d)\n", base, limit + 1, error); if (error != 0) return (error); } return (error); } static struct dmar_domain * dmar_domain_alloc(struct dmar_unit *dmar, bool id_mapped) { struct iommu_domain *iodom; struct iommu_unit *unit; struct dmar_domain *domain; int error, id, mgaw; id = alloc_unr(dmar->domids); if (id == -1) return (NULL); domain = malloc(sizeof(*domain), M_DMAR_DOMAIN, M_WAITOK | M_ZERO); iodom = DOM2IODOM(domain); unit = DMAR2IOMMU(dmar); domain->domain = id; LIST_INIT(&domain->contexts); iommu_domain_init(unit, iodom, &dmar_domain_map_ops); domain->dmar = dmar; /* * For now, use the maximal usable physical address of the * installed memory to calculate the mgaw on id_mapped domain. * It is useful for the identity mapping, and less so for the * virtualized bus address space. */ domain->iodom.end = id_mapped ? ptoa(Maxmem) : BUS_SPACE_MAXADDR; mgaw = dmar_maxaddr2mgaw(dmar, domain->iodom.end, !id_mapped); error = domain_set_agaw(domain, mgaw); if (error != 0) goto fail; if (!id_mapped) /* Use all supported address space for remapping. */ domain->iodom.end = 1ULL << (domain->agaw - 1); iommu_gas_init_domain(DOM2IODOM(domain)); if (id_mapped) { if ((dmar->hw_ecap & DMAR_ECAP_PT) == 0) { domain->pgtbl_obj = domain_get_idmap_pgtbl(domain, domain->iodom.end); } domain->iodom.flags |= IOMMU_DOMAIN_IDMAP; } else { error = domain_alloc_pgtbl(domain); if (error != 0) goto fail; /* Disable local apic region access */ error = iommu_gas_reserve_region(iodom, 0xfee00000, 0xfeefffff + 1, &iodom->msi_entry); if (error != 0) goto fail; } return (domain); fail: dmar_domain_destroy(domain); return (NULL); } static struct dmar_ctx * dmar_ctx_alloc(struct dmar_domain *domain, uint16_t rid) { struct dmar_ctx *ctx; ctx = malloc(sizeof(*ctx), M_DMAR_CTX, M_WAITOK | M_ZERO); ctx->context.domain = DOM2IODOM(domain); ctx->context.tag = malloc(sizeof(struct bus_dma_tag_iommu), M_DMAR_CTX, M_WAITOK | M_ZERO); ctx->context.rid = rid; ctx->refs = 1; return (ctx); } static void dmar_ctx_link(struct dmar_ctx *ctx) { struct dmar_domain *domain; domain = CTX2DOM(ctx); IOMMU_ASSERT_LOCKED(domain->iodom.iommu); KASSERT(domain->refs >= domain->ctx_cnt, ("dom %p ref underflow %d %d", domain, domain->refs, domain->ctx_cnt)); domain->refs++; domain->ctx_cnt++; LIST_INSERT_HEAD(&domain->contexts, ctx, link); } static void dmar_ctx_unlink(struct dmar_ctx *ctx) { struct dmar_domain *domain; domain = CTX2DOM(ctx); IOMMU_ASSERT_LOCKED(domain->iodom.iommu); KASSERT(domain->refs > 0, ("domain %p ctx dtr refs %d", domain, domain->refs)); KASSERT(domain->ctx_cnt >= domain->refs, ("domain %p ctx dtr refs %d ctx_cnt %d", domain, domain->refs, domain->ctx_cnt)); domain->refs--; domain->ctx_cnt--; LIST_REMOVE(ctx, link); } static void dmar_domain_destroy(struct dmar_domain *domain) { struct iommu_domain *iodom; struct dmar_unit *dmar; iodom = DOM2IODOM(domain); KASSERT(TAILQ_EMPTY(&domain->iodom.unload_entries), ("unfinished unloads %p", domain)); KASSERT(LIST_EMPTY(&domain->contexts), ("destroying dom %p with contexts", domain)); KASSERT(domain->ctx_cnt == 0, ("destroying dom %p with ctx_cnt %d", domain, domain->ctx_cnt)); KASSERT(domain->refs == 0, ("destroying dom %p with refs %d", domain, domain->refs)); if ((domain->iodom.flags & IOMMU_DOMAIN_GAS_INITED) != 0) { DMAR_DOMAIN_LOCK(domain); iommu_gas_fini_domain(iodom); DMAR_DOMAIN_UNLOCK(domain); } if ((domain->iodom.flags & IOMMU_DOMAIN_PGTBL_INITED) != 0) { if (domain->pgtbl_obj != NULL) DMAR_DOMAIN_PGLOCK(domain); domain_free_pgtbl(domain); } iommu_domain_fini(iodom); dmar = DOM2DMAR(domain); free_unr(dmar->domids, domain->domain); free(domain, M_DMAR_DOMAIN); } static struct dmar_ctx * dmar_get_ctx_for_dev1(struct dmar_unit *dmar, device_t dev, uint16_t rid, int dev_domain, int dev_busno, const void *dev_path, int dev_path_len, bool id_mapped, bool rmrr_init) { struct dmar_domain *domain, *domain1; struct dmar_ctx *ctx, *ctx1; struct iommu_unit *unit __diagused; dmar_ctx_entry_t *ctxp; struct sf_buf *sf; int bus, slot, func, error; bool enable; if (dev != NULL) { bus = pci_get_bus(dev); slot = pci_get_slot(dev); func = pci_get_function(dev); } else { bus = PCI_RID2BUS(rid); slot = PCI_RID2SLOT(rid); func = PCI_RID2FUNC(rid); } enable = false; TD_PREP_PINNED_ASSERT; unit = DMAR2IOMMU(dmar); DMAR_LOCK(dmar); KASSERT(!iommu_is_buswide_ctx(unit, bus) || (slot == 0 && func == 0), ("iommu%d pci%d:%d:%d get_ctx for buswide", dmar->iommu.unit, bus, slot, func)); ctx = dmar_find_ctx_locked(dmar, rid); error = 0; if (ctx == NULL) { /* * Perform the allocations which require sleep or have * higher chance to succeed if the sleep is allowed. */ DMAR_UNLOCK(dmar); dmar_ensure_ctx_page(dmar, PCI_RID2BUS(rid)); domain1 = dmar_domain_alloc(dmar, id_mapped); if (domain1 == NULL) { TD_PINNED_ASSERT; return (NULL); } if (!id_mapped) { error = domain_init_rmrr(domain1, dev, bus, slot, func, dev_domain, dev_busno, dev_path, dev_path_len); if (error == 0) error = dmar_reserve_pci_regions(domain1, dev); if (error != 0) { dmar_domain_destroy(domain1); TD_PINNED_ASSERT; return (NULL); } } ctx1 = dmar_ctx_alloc(domain1, rid); ctxp = dmar_map_ctx_entry(ctx1, &sf); DMAR_LOCK(dmar); /* * Recheck the contexts, other thread might have * already allocated needed one. */ ctx = dmar_find_ctx_locked(dmar, rid); if (ctx == NULL) { domain = domain1; ctx = ctx1; dmar_ctx_link(ctx); ctx->context.tag->owner = dev; device_tag_init(ctx, dev); /* * This is the first activated context for the * DMAR unit. Enable the translation after * everything is set up. */ if (LIST_EMPTY(&dmar->domains)) enable = true; LIST_INSERT_HEAD(&dmar->domains, domain, link); ctx_id_entry_init(ctx, ctxp, false, bus); if (dev != NULL) { device_printf(dev, "dmar%d pci%d:%d:%d:%d rid %x domain %d mgaw %d " "agaw %d %s-mapped\n", dmar->iommu.unit, dmar->segment, bus, slot, func, rid, domain->domain, domain->mgaw, domain->agaw, id_mapped ? "id" : "re"); } dmar_unmap_pgtbl(sf); } else { dmar_unmap_pgtbl(sf); dmar_domain_destroy(domain1); /* Nothing needs to be done to destroy ctx1. */ free(ctx1, M_DMAR_CTX); domain = CTX2DOM(ctx); ctx->refs++; /* tag referenced us */ } } else { domain = CTX2DOM(ctx); if (ctx->context.tag->owner == NULL) ctx->context.tag->owner = dev; ctx->refs++; /* tag referenced us */ } error = dmar_flush_for_ctx_entry(dmar, enable); if (error != 0) { dmar_free_ctx_locked(dmar, ctx); TD_PINNED_ASSERT; return (NULL); } /* * The dmar lock was potentially dropped between check for the * empty context list and now. Recheck the state of GCMD_TE * to avoid unneeded command. */ if (enable && !rmrr_init && (dmar->hw_gcmd & DMAR_GCMD_TE) == 0) { error = dmar_enable_translation(dmar); if (error == 0) { if (bootverbose) { printf("dmar%d: enabled translation\n", dmar->iommu.unit); } } else { printf("dmar%d: enabling translation failed, " "error %d\n", dmar->iommu.unit, error); dmar_free_ctx_locked(dmar, ctx); TD_PINNED_ASSERT; return (NULL); } } DMAR_UNLOCK(dmar); TD_PINNED_ASSERT; return (ctx); } struct dmar_ctx * dmar_get_ctx_for_dev(struct dmar_unit *dmar, device_t dev, uint16_t rid, bool id_mapped, bool rmrr_init) { int dev_domain, dev_path_len, dev_busno; dev_domain = pci_get_domain(dev); dev_path_len = dmar_dev_depth(dev); ACPI_DMAR_PCI_PATH dev_path[dev_path_len]; dmar_dev_path(dev, &dev_busno, dev_path, dev_path_len); return (dmar_get_ctx_for_dev1(dmar, dev, rid, dev_domain, dev_busno, dev_path, dev_path_len, id_mapped, rmrr_init)); } struct dmar_ctx * dmar_get_ctx_for_devpath(struct dmar_unit *dmar, uint16_t rid, int dev_domain, int dev_busno, const void *dev_path, int dev_path_len, bool id_mapped, bool rmrr_init) { return (dmar_get_ctx_for_dev1(dmar, NULL, rid, dev_domain, dev_busno, dev_path, dev_path_len, id_mapped, rmrr_init)); } int dmar_move_ctx_to_domain(struct dmar_domain *domain, struct dmar_ctx *ctx) { struct dmar_unit *dmar; struct dmar_domain *old_domain; dmar_ctx_entry_t *ctxp; struct sf_buf *sf; int error; dmar = domain->dmar; old_domain = CTX2DOM(ctx); if (domain == old_domain) return (0); KASSERT(old_domain->iodom.iommu == domain->iodom.iommu, ("domain %p %u moving between dmars %u %u", domain, domain->domain, old_domain->iodom.iommu->unit, domain->iodom.iommu->unit)); TD_PREP_PINNED_ASSERT; ctxp = dmar_map_ctx_entry(ctx, &sf); DMAR_LOCK(dmar); dmar_ctx_unlink(ctx); ctx->context.domain = &domain->iodom; dmar_ctx_link(ctx); ctx_id_entry_init(ctx, ctxp, true, PCI_BUSMAX + 100); dmar_unmap_pgtbl(sf); error = dmar_flush_for_ctx_entry(dmar, true); /* If flush failed, rolling back would not work as well. */ printf("dmar%d rid %x domain %d->%d %s-mapped\n", dmar->iommu.unit, ctx->context.rid, old_domain->domain, domain->domain, (domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0 ? "id" : "re"); dmar_unref_domain_locked(dmar, old_domain); TD_PINNED_ASSERT; return (error); } static void dmar_unref_domain_locked(struct dmar_unit *dmar, struct dmar_domain *domain) { DMAR_ASSERT_LOCKED(dmar); KASSERT(domain->refs >= 1, ("dmar %d domain %p refs %u", dmar->iommu.unit, domain, domain->refs)); KASSERT(domain->refs > domain->ctx_cnt, ("dmar %d domain %p refs %d ctx_cnt %d", dmar->iommu.unit, domain, domain->refs, domain->ctx_cnt)); if (domain->refs > 1) { domain->refs--; DMAR_UNLOCK(dmar); return; } KASSERT((domain->iodom.flags & IOMMU_DOMAIN_RMRR) == 0, ("lost ref on RMRR domain %p", domain)); LIST_REMOVE(domain, link); DMAR_UNLOCK(dmar); taskqueue_drain(dmar->iommu.delayed_taskqueue, &domain->iodom.unload_task); dmar_domain_destroy(domain); } void dmar_free_ctx_locked(struct dmar_unit *dmar, struct dmar_ctx *ctx) { struct sf_buf *sf; dmar_ctx_entry_t *ctxp; struct dmar_domain *domain; DMAR_ASSERT_LOCKED(dmar); KASSERT(ctx->refs >= 1, ("dmar %p ctx %p refs %u", dmar, ctx, ctx->refs)); /* * If our reference is not last, only the dereference should * be performed. */ if (ctx->refs > 1) { ctx->refs--; DMAR_UNLOCK(dmar); return; } KASSERT((ctx->context.flags & IOMMU_CTX_DISABLED) == 0, ("lost ref on disabled ctx %p", ctx)); /* * Otherwise, the context entry must be cleared before the * page table is destroyed. The mapping of the context * entries page could require sleep, unlock the dmar. */ DMAR_UNLOCK(dmar); TD_PREP_PINNED_ASSERT; ctxp = dmar_map_ctx_entry(ctx, &sf); DMAR_LOCK(dmar); KASSERT(ctx->refs >= 1, ("dmar %p ctx %p refs %u", dmar, ctx, ctx->refs)); /* * Other thread might have referenced the context, in which * case again only the dereference should be performed. */ if (ctx->refs > 1) { ctx->refs--; DMAR_UNLOCK(dmar); dmar_unmap_pgtbl(sf); TD_PINNED_ASSERT; return; } KASSERT((ctx->context.flags & IOMMU_CTX_DISABLED) == 0, ("lost ref on disabled ctx %p", ctx)); /* * Clear the context pointer and flush the caches. * XXXKIB: cannot do this if any RMRR entries are still present. */ dmar_pte_clear(&ctxp->ctx1); ctxp->ctx2 = 0; dmar_flush_ctx_to_ram(dmar, ctxp); dmar_inv_ctx_glob(dmar); if ((dmar->hw_ecap & DMAR_ECAP_DI) != 0) { if (dmar->qi_enabled) dmar_qi_invalidate_iotlb_glob_locked(dmar); else dmar_inv_iotlb_glob(dmar); } dmar_unmap_pgtbl(sf); domain = CTX2DOM(ctx); dmar_ctx_unlink(ctx); free(ctx->context.tag, M_DMAR_CTX); free(ctx, M_DMAR_CTX); dmar_unref_domain_locked(dmar, domain); TD_PINNED_ASSERT; } void dmar_free_ctx(struct dmar_ctx *ctx) { struct dmar_unit *dmar; dmar = CTX2DMAR(ctx); DMAR_LOCK(dmar); dmar_free_ctx_locked(dmar, ctx); } /* * Returns with the domain locked. */ struct dmar_ctx * dmar_find_ctx_locked(struct dmar_unit *dmar, uint16_t rid) { struct dmar_domain *domain; struct dmar_ctx *ctx; DMAR_ASSERT_LOCKED(dmar); LIST_FOREACH(domain, &dmar->domains, link) { LIST_FOREACH(ctx, &domain->contexts, link) { if (ctx->context.rid == rid) return (ctx); } } return (NULL); } void dmar_domain_free_entry(struct iommu_map_entry *entry, bool free) { struct iommu_domain *domain; domain = entry->domain; IOMMU_DOMAIN_LOCK(domain); if ((entry->flags & IOMMU_MAP_ENTRY_RMRR) != 0) iommu_gas_free_region(domain, entry); else iommu_gas_free_space(domain, entry); IOMMU_DOMAIN_UNLOCK(domain); if (free) iommu_gas_free_entry(domain, entry); else entry->flags = 0; } void iommu_domain_unload_entry(struct iommu_map_entry *entry, bool free) { struct dmar_domain *domain; struct dmar_unit *unit; domain = IODOM2DOM(entry->domain); unit = DOM2DMAR(domain); if (unit->qi_enabled) { DMAR_LOCK(unit); dmar_qi_invalidate_locked(IODOM2DOM(entry->domain), entry->start, entry->end - entry->start, &entry->gseq, true); if (!free) entry->flags |= IOMMU_MAP_ENTRY_QI_NF; TAILQ_INSERT_TAIL(&unit->tlb_flush_entries, entry, dmamap_link); DMAR_UNLOCK(unit); } else { domain_flush_iotlb_sync(IODOM2DOM(entry->domain), entry->start, entry->end - entry->start); dmar_domain_free_entry(entry, free); } } static bool dmar_domain_unload_emit_wait(struct dmar_domain *domain, struct iommu_map_entry *entry) { if (TAILQ_NEXT(entry, dmamap_link) == NULL) return (true); return (domain->batch_no++ % dmar_batch_coalesce == 0); } void iommu_domain_unload(struct iommu_domain *iodom, struct iommu_map_entries_tailq *entries, bool cansleep) { struct dmar_domain *domain; struct dmar_unit *unit; struct iommu_map_entry *entry, *entry1; int error __diagused; domain = IODOM2DOM(iodom); unit = DOM2DMAR(domain); TAILQ_FOREACH_SAFE(entry, entries, dmamap_link, entry1) { KASSERT((entry->flags & IOMMU_MAP_ENTRY_MAP) != 0, ("not mapped entry %p %p", domain, entry)); error = iodom->ops->unmap(iodom, entry->start, entry->end - entry->start, cansleep ? IOMMU_PGF_WAITOK : 0); KASSERT(error == 0, ("unmap %p error %d", domain, error)); if (!unit->qi_enabled) { domain_flush_iotlb_sync(domain, entry->start, entry->end - entry->start); TAILQ_REMOVE(entries, entry, dmamap_link); dmar_domain_free_entry(entry, true); } } if (TAILQ_EMPTY(entries)) return; KASSERT(unit->qi_enabled, ("loaded entry left")); DMAR_LOCK(unit); TAILQ_FOREACH(entry, entries, dmamap_link) { dmar_qi_invalidate_locked(domain, entry->start, entry->end - entry->start, &entry->gseq, dmar_domain_unload_emit_wait(domain, entry)); } TAILQ_CONCAT(&unit->tlb_flush_entries, entries, dmamap_link); DMAR_UNLOCK(unit); } struct iommu_ctx * iommu_get_ctx(struct iommu_unit *iommu, device_t dev, uint16_t rid, bool id_mapped, bool rmrr_init) { struct dmar_unit *dmar; struct dmar_ctx *ret; dmar = IOMMU2DMAR(iommu); ret = dmar_get_ctx_for_dev(dmar, dev, rid, id_mapped, rmrr_init); return (CTX2IOCTX(ret)); } void iommu_free_ctx_locked(struct iommu_unit *iommu, struct iommu_ctx *context) { struct dmar_unit *dmar; struct dmar_ctx *ctx; dmar = IOMMU2DMAR(iommu); ctx = IOCTX2CTX(context); dmar_free_ctx_locked(dmar, ctx); } void iommu_free_ctx(struct iommu_ctx *context) { struct dmar_ctx *ctx; ctx = IOCTX2CTX(context); dmar_free_ctx(ctx); } diff --git a/sys/x86/iommu/intel_drv.c b/sys/x86/iommu/intel_drv.c index 0b470d7bbf7a..c7647c1f7b1d 100644 --- a/sys/x86/iommu/intel_drv.c +++ b/sys/x86/iommu/intel_drv.c @@ -1,1336 +1,1336 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2013-2015 The FreeBSD Foundation * * This software was developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_acpi.h" #if defined(__amd64__) #define DEV_APIC #else #include "opt_apic.h" #endif #include "opt_ddb.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DEV_APIC #include "pcib_if.h" #include #include #include #endif #define DMAR_FAULT_IRQ_RID 0 #define DMAR_QI_IRQ_RID 1 #define DMAR_REG_RID 2 static devclass_t dmar_devclass; static device_t *dmar_devs; static int dmar_devcnt; typedef int (*dmar_iter_t)(ACPI_DMAR_HEADER *, void *); static void dmar_iterate_tbl(dmar_iter_t iter, void *arg) { ACPI_TABLE_DMAR *dmartbl; ACPI_DMAR_HEADER *dmarh; char *ptr, *ptrend; ACPI_STATUS status; status = AcpiGetTable(ACPI_SIG_DMAR, 1, (ACPI_TABLE_HEADER **)&dmartbl); if (ACPI_FAILURE(status)) return; ptr = (char *)dmartbl + sizeof(*dmartbl); ptrend = (char *)dmartbl + dmartbl->Header.Length; for (;;) { if (ptr >= ptrend) break; dmarh = (ACPI_DMAR_HEADER *)ptr; if (dmarh->Length <= 0) { printf("dmar_identify: corrupted DMAR table, l %d\n", dmarh->Length); break; } ptr += dmarh->Length; if (!iter(dmarh, arg)) break; } AcpiPutTable((ACPI_TABLE_HEADER *)dmartbl); } struct find_iter_args { int i; ACPI_DMAR_HARDWARE_UNIT *res; }; static int dmar_find_iter(ACPI_DMAR_HEADER *dmarh, void *arg) { struct find_iter_args *fia; if (dmarh->Type != ACPI_DMAR_TYPE_HARDWARE_UNIT) return (1); fia = arg; if (fia->i == 0) { fia->res = (ACPI_DMAR_HARDWARE_UNIT *)dmarh; return (0); } fia->i--; return (1); } static ACPI_DMAR_HARDWARE_UNIT * dmar_find_by_index(int idx) { struct find_iter_args fia; fia.i = idx; fia.res = NULL; dmar_iterate_tbl(dmar_find_iter, &fia); return (fia.res); } static int dmar_count_iter(ACPI_DMAR_HEADER *dmarh, void *arg) { if (dmarh->Type == ACPI_DMAR_TYPE_HARDWARE_UNIT) dmar_devcnt++; return (1); } static int dmar_enable = 0; static void dmar_identify(driver_t *driver, device_t parent) { ACPI_TABLE_DMAR *dmartbl; ACPI_DMAR_HARDWARE_UNIT *dmarh; ACPI_STATUS status; int i, error; if (acpi_disabled("dmar")) return; TUNABLE_INT_FETCH("hw.dmar.enable", &dmar_enable); if (!dmar_enable) return; status = AcpiGetTable(ACPI_SIG_DMAR, 1, (ACPI_TABLE_HEADER **)&dmartbl); if (ACPI_FAILURE(status)) return; haw = dmartbl->Width + 1; if ((1ULL << (haw + 1)) > BUS_SPACE_MAXADDR) dmar_high = BUS_SPACE_MAXADDR; else dmar_high = 1ULL << (haw + 1); if (bootverbose) { printf("DMAR HAW=%d flags=<%b>\n", dmartbl->Width, (unsigned)dmartbl->Flags, "\020\001INTR_REMAP\002X2APIC_OPT_OUT"); } AcpiPutTable((ACPI_TABLE_HEADER *)dmartbl); dmar_iterate_tbl(dmar_count_iter, NULL); if (dmar_devcnt == 0) return; dmar_devs = malloc(sizeof(device_t) * dmar_devcnt, M_DEVBUF, M_WAITOK | M_ZERO); for (i = 0; i < dmar_devcnt; i++) { dmarh = dmar_find_by_index(i); if (dmarh == NULL) { printf("dmar_identify: cannot find HWUNIT %d\n", i); continue; } dmar_devs[i] = BUS_ADD_CHILD(parent, 1, "dmar", i); if (dmar_devs[i] == NULL) { printf("dmar_identify: cannot create instance %d\n", i); continue; } error = bus_set_resource(dmar_devs[i], SYS_RES_MEMORY, DMAR_REG_RID, dmarh->Address, PAGE_SIZE); if (error != 0) { printf( "dmar%d: unable to alloc register window at 0x%08jx: error %d\n", i, (uintmax_t)dmarh->Address, error); device_delete_child(parent, dmar_devs[i]); dmar_devs[i] = NULL; } } } static int dmar_probe(device_t dev) { if (acpi_get_handle(dev) != NULL) return (ENXIO); device_set_desc(dev, "DMA remap"); return (BUS_PROBE_NOWILDCARD); } static void dmar_release_intr(device_t dev, struct dmar_unit *unit, int idx) { struct dmar_msi_data *dmd; dmd = &unit->intrs[idx]; if (dmd->irq == -1) return; bus_teardown_intr(dev, dmd->irq_res, dmd->intr_handle); bus_release_resource(dev, SYS_RES_IRQ, dmd->irq_rid, dmd->irq_res); bus_delete_resource(dev, SYS_RES_IRQ, dmd->irq_rid); PCIB_RELEASE_MSIX(device_get_parent(device_get_parent(dev)), dev, dmd->irq); dmd->irq = -1; } static void dmar_release_resources(device_t dev, struct dmar_unit *unit) { int i; iommu_fini_busdma(&unit->iommu); dmar_fini_irt(unit); dmar_fini_qi(unit); dmar_fini_fault_log(unit); for (i = 0; i < DMAR_INTR_TOTAL; i++) dmar_release_intr(dev, unit, i); if (unit->regs != NULL) { bus_deactivate_resource(dev, SYS_RES_MEMORY, unit->reg_rid, unit->regs); bus_release_resource(dev, SYS_RES_MEMORY, unit->reg_rid, unit->regs); unit->regs = NULL; } if (unit->domids != NULL) { delete_unrhdr(unit->domids); unit->domids = NULL; } if (unit->ctx_obj != NULL) { vm_object_deallocate(unit->ctx_obj); unit->ctx_obj = NULL; } } static int dmar_alloc_irq(device_t dev, struct dmar_unit *unit, int idx) { device_t pcib; struct dmar_msi_data *dmd; uint64_t msi_addr; uint32_t msi_data; int error; dmd = &unit->intrs[idx]; pcib = device_get_parent(device_get_parent(dev)); /* Really not pcib */ error = PCIB_ALLOC_MSIX(pcib, dev, &dmd->irq); if (error != 0) { device_printf(dev, "cannot allocate %s interrupt, %d\n", dmd->name, error); goto err1; } error = bus_set_resource(dev, SYS_RES_IRQ, dmd->irq_rid, dmd->irq, 1); if (error != 0) { device_printf(dev, "cannot set %s interrupt resource, %d\n", dmd->name, error); goto err2; } dmd->irq_res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &dmd->irq_rid, RF_ACTIVE); if (dmd->irq_res == NULL) { device_printf(dev, "cannot allocate resource for %s interrupt\n", dmd->name); error = ENXIO; goto err3; } error = bus_setup_intr(dev, dmd->irq_res, INTR_TYPE_MISC, dmd->handler, NULL, unit, &dmd->intr_handle); if (error != 0) { device_printf(dev, "cannot setup %s interrupt, %d\n", dmd->name, error); goto err4; } bus_describe_intr(dev, dmd->irq_res, dmd->intr_handle, "%s", dmd->name); error = PCIB_MAP_MSI(pcib, dev, dmd->irq, &msi_addr, &msi_data); if (error != 0) { device_printf(dev, "cannot map %s interrupt, %d\n", dmd->name, error); goto err5; } dmar_write4(unit, dmd->msi_data_reg, msi_data); dmar_write4(unit, dmd->msi_addr_reg, msi_addr); /* Only for xAPIC mode */ dmar_write4(unit, dmd->msi_uaddr_reg, msi_addr >> 32); return (0); err5: bus_teardown_intr(dev, dmd->irq_res, dmd->intr_handle); err4: bus_release_resource(dev, SYS_RES_IRQ, dmd->irq_rid, dmd->irq_res); err3: bus_delete_resource(dev, SYS_RES_IRQ, dmd->irq_rid); err2: PCIB_RELEASE_MSIX(pcib, dev, dmd->irq); dmd->irq = -1; err1: return (error); } #ifdef DEV_APIC static int dmar_remap_intr(device_t dev, device_t child, u_int irq) { struct dmar_unit *unit; struct dmar_msi_data *dmd; uint64_t msi_addr; uint32_t msi_data; int i, error; unit = device_get_softc(dev); for (i = 0; i < DMAR_INTR_TOTAL; i++) { dmd = &unit->intrs[i]; if (irq == dmd->irq) { error = PCIB_MAP_MSI(device_get_parent( device_get_parent(dev)), dev, irq, &msi_addr, &msi_data); if (error != 0) return (error); DMAR_LOCK(unit); (dmd->disable_intr)(unit); dmar_write4(unit, dmd->msi_data_reg, msi_data); dmar_write4(unit, dmd->msi_addr_reg, msi_addr); dmar_write4(unit, dmd->msi_uaddr_reg, msi_addr >> 32); (dmd->enable_intr)(unit); DMAR_UNLOCK(unit); return (0); } } return (ENOENT); } #endif static void dmar_print_caps(device_t dev, struct dmar_unit *unit, ACPI_DMAR_HARDWARE_UNIT *dmaru) { uint32_t caphi, ecaphi; device_printf(dev, "regs@0x%08jx, ver=%d.%d, seg=%d, flags=<%b>\n", (uintmax_t)dmaru->Address, DMAR_MAJOR_VER(unit->hw_ver), DMAR_MINOR_VER(unit->hw_ver), dmaru->Segment, dmaru->Flags, "\020\001INCLUDE_ALL_PCI"); caphi = unit->hw_cap >> 32; device_printf(dev, "cap=%b,", (u_int)unit->hw_cap, "\020\004AFL\005WBF\006PLMR\007PHMR\010CM\027ZLR\030ISOCH"); printf("%b, ", caphi, "\020\010PSI\027DWD\030DRD\031FL1GP\034PSI"); printf("ndoms=%d, sagaw=%d, mgaw=%d, fro=%d, nfr=%d, superp=%d", DMAR_CAP_ND(unit->hw_cap), DMAR_CAP_SAGAW(unit->hw_cap), DMAR_CAP_MGAW(unit->hw_cap), DMAR_CAP_FRO(unit->hw_cap), DMAR_CAP_NFR(unit->hw_cap), DMAR_CAP_SPS(unit->hw_cap)); if ((unit->hw_cap & DMAR_CAP_PSI) != 0) printf(", mamv=%d", DMAR_CAP_MAMV(unit->hw_cap)); printf("\n"); ecaphi = unit->hw_ecap >> 32; device_printf(dev, "ecap=%b,", (u_int)unit->hw_ecap, "\020\001C\002QI\003DI\004IR\005EIM\007PT\010SC\031ECS\032MTS" "\033NEST\034DIS\035PASID\036PRS\037ERS\040SRS"); printf("%b, ", ecaphi, "\020\002NWFS\003EAFS"); printf("mhmw=%d, iro=%d\n", DMAR_ECAP_MHMV(unit->hw_ecap), DMAR_ECAP_IRO(unit->hw_ecap)); } static int dmar_attach(device_t dev) { struct dmar_unit *unit; ACPI_DMAR_HARDWARE_UNIT *dmaru; uint64_t timeout; int i, error; unit = device_get_softc(dev); unit->dev = dev; unit->iommu.unit = device_get_unit(dev); unit->iommu.dev = dev; dmaru = dmar_find_by_index(unit->iommu.unit); if (dmaru == NULL) return (EINVAL); unit->segment = dmaru->Segment; unit->base = dmaru->Address; unit->reg_rid = DMAR_REG_RID; unit->regs = bus_alloc_resource_any(dev, SYS_RES_MEMORY, &unit->reg_rid, RF_ACTIVE); if (unit->regs == NULL) { device_printf(dev, "cannot allocate register window\n"); return (ENOMEM); } unit->hw_ver = dmar_read4(unit, DMAR_VER_REG); unit->hw_cap = dmar_read8(unit, DMAR_CAP_REG); unit->hw_ecap = dmar_read8(unit, DMAR_ECAP_REG); if (bootverbose) dmar_print_caps(dev, unit, dmaru); dmar_quirks_post_ident(unit); timeout = dmar_get_timeout(); TUNABLE_UINT64_FETCH("hw.dmar.timeout", &timeout); dmar_update_timeout(timeout); for (i = 0; i < DMAR_INTR_TOTAL; i++) unit->intrs[i].irq = -1; unit->intrs[DMAR_INTR_FAULT].name = "fault"; unit->intrs[DMAR_INTR_FAULT].irq_rid = DMAR_FAULT_IRQ_RID; unit->intrs[DMAR_INTR_FAULT].handler = dmar_fault_intr; unit->intrs[DMAR_INTR_FAULT].msi_data_reg = DMAR_FEDATA_REG; unit->intrs[DMAR_INTR_FAULT].msi_addr_reg = DMAR_FEADDR_REG; unit->intrs[DMAR_INTR_FAULT].msi_uaddr_reg = DMAR_FEUADDR_REG; unit->intrs[DMAR_INTR_FAULT].enable_intr = dmar_enable_fault_intr; unit->intrs[DMAR_INTR_FAULT].disable_intr = dmar_disable_fault_intr; error = dmar_alloc_irq(dev, unit, DMAR_INTR_FAULT); if (error != 0) { dmar_release_resources(dev, unit); return (error); } if (DMAR_HAS_QI(unit)) { unit->intrs[DMAR_INTR_QI].name = "qi"; unit->intrs[DMAR_INTR_QI].irq_rid = DMAR_QI_IRQ_RID; unit->intrs[DMAR_INTR_QI].handler = dmar_qi_intr; unit->intrs[DMAR_INTR_QI].msi_data_reg = DMAR_IEDATA_REG; unit->intrs[DMAR_INTR_QI].msi_addr_reg = DMAR_IEADDR_REG; unit->intrs[DMAR_INTR_QI].msi_uaddr_reg = DMAR_IEUADDR_REG; unit->intrs[DMAR_INTR_QI].enable_intr = dmar_enable_qi_intr; unit->intrs[DMAR_INTR_QI].disable_intr = dmar_disable_qi_intr; error = dmar_alloc_irq(dev, unit, DMAR_INTR_QI); if (error != 0) { dmar_release_resources(dev, unit); return (error); } } mtx_init(&unit->iommu.lock, "dmarhw", NULL, MTX_DEF); unit->domids = new_unrhdr(0, dmar_nd2mask(DMAR_CAP_ND(unit->hw_cap)), &unit->iommu.lock); LIST_INIT(&unit->domains); /* * 9.2 "Context Entry": * When Caching Mode (CM) field is reported as Set, the * domain-id value of zero is architecturally reserved. * Software must not use domain-id value of zero * when CM is Set. */ if ((unit->hw_cap & DMAR_CAP_CM) != 0) alloc_unr_specific(unit->domids, 0); unit->ctx_obj = vm_pager_allocate(OBJT_PHYS, NULL, IDX_TO_OFF(1 + DMAR_CTX_CNT), 0, 0, NULL); /* * Allocate and load the root entry table pointer. Enable the * address translation after the required invalidations are * done. */ dmar_pgalloc(unit->ctx_obj, 0, IOMMU_PGF_WAITOK | IOMMU_PGF_ZERO); DMAR_LOCK(unit); error = dmar_load_root_entry_ptr(unit); if (error != 0) { DMAR_UNLOCK(unit); dmar_release_resources(dev, unit); return (error); } error = dmar_inv_ctx_glob(unit); if (error != 0) { DMAR_UNLOCK(unit); dmar_release_resources(dev, unit); return (error); } if ((unit->hw_ecap & DMAR_ECAP_DI) != 0) { error = dmar_inv_iotlb_glob(unit); if (error != 0) { DMAR_UNLOCK(unit); dmar_release_resources(dev, unit); return (error); } } DMAR_UNLOCK(unit); error = dmar_init_fault_log(unit); if (error != 0) { dmar_release_resources(dev, unit); return (error); } error = dmar_init_qi(unit); if (error != 0) { dmar_release_resources(dev, unit); return (error); } error = dmar_init_irt(unit); if (error != 0) { dmar_release_resources(dev, unit); return (error); } error = iommu_init_busdma(&unit->iommu); if (error != 0) { dmar_release_resources(dev, unit); return (error); } #ifdef NOTYET DMAR_LOCK(unit); error = dmar_enable_translation(unit); if (error != 0) { DMAR_UNLOCK(unit); dmar_release_resources(dev, unit); return (error); } DMAR_UNLOCK(unit); #endif return (0); } static int dmar_detach(device_t dev) { return (EBUSY); } static int dmar_suspend(device_t dev) { return (0); } static int dmar_resume(device_t dev) { /* XXXKIB */ return (0); } static device_method_t dmar_methods[] = { DEVMETHOD(device_identify, dmar_identify), DEVMETHOD(device_probe, dmar_probe), DEVMETHOD(device_attach, dmar_attach), DEVMETHOD(device_detach, dmar_detach), DEVMETHOD(device_suspend, dmar_suspend), DEVMETHOD(device_resume, dmar_resume), #ifdef DEV_APIC DEVMETHOD(bus_remap_intr, dmar_remap_intr), #endif DEVMETHOD_END }; static driver_t dmar_driver = { "dmar", dmar_methods, sizeof(struct dmar_unit), }; DRIVER_MODULE(dmar, acpi, dmar_driver, dmar_devclass, 0, 0); MODULE_DEPEND(dmar, acpi, 1, 1, 1); static void dmar_print_path(int busno, int depth, const ACPI_DMAR_PCI_PATH *path) { int i; printf("[%d, ", busno); for (i = 0; i < depth; i++) { if (i != 0) printf(", "); printf("(%d, %d)", path[i].Device, path[i].Function); } printf("]"); } int dmar_dev_depth(device_t child) { devclass_t pci_class; device_t bus, pcib; int depth; pci_class = devclass_find("pci"); for (depth = 1; ; depth++) { bus = device_get_parent(child); pcib = device_get_parent(bus); if (device_get_devclass(device_get_parent(pcib)) != pci_class) return (depth); child = pcib; } } void dmar_dev_path(device_t child, int *busno, void *path1, int depth) { devclass_t pci_class; device_t bus, pcib; ACPI_DMAR_PCI_PATH *path; pci_class = devclass_find("pci"); path = path1; for (depth--; depth != -1; depth--) { path[depth].Device = pci_get_slot(child); path[depth].Function = pci_get_function(child); bus = device_get_parent(child); pcib = device_get_parent(bus); if (device_get_devclass(device_get_parent(pcib)) != pci_class) { /* reached a host bridge */ *busno = pcib_get_bus(bus); return; } child = pcib; } panic("wrong depth"); } static int dmar_match_pathes(int busno1, const ACPI_DMAR_PCI_PATH *path1, int depth1, int busno2, const ACPI_DMAR_PCI_PATH *path2, int depth2, enum AcpiDmarScopeType scope_type) { int i, depth; if (busno1 != busno2) return (0); if (scope_type == ACPI_DMAR_SCOPE_TYPE_ENDPOINT && depth1 != depth2) return (0); depth = depth1; if (depth2 < depth) depth = depth2; for (i = 0; i < depth; i++) { if (path1[i].Device != path2[i].Device || path1[i].Function != path2[i].Function) return (0); } return (1); } static int dmar_match_devscope(ACPI_DMAR_DEVICE_SCOPE *devscope, int dev_busno, const ACPI_DMAR_PCI_PATH *dev_path, int dev_path_len) { ACPI_DMAR_PCI_PATH *path; int path_len; if (devscope->Length < sizeof(*devscope)) { printf("dmar_match_devscope: corrupted DMAR table, dl %d\n", devscope->Length); return (-1); } if (devscope->EntryType != ACPI_DMAR_SCOPE_TYPE_ENDPOINT && devscope->EntryType != ACPI_DMAR_SCOPE_TYPE_BRIDGE) return (0); path_len = devscope->Length - sizeof(*devscope); if (path_len % 2 != 0) { printf("dmar_match_devscope: corrupted DMAR table, dl %d\n", devscope->Length); return (-1); } path_len /= 2; path = (ACPI_DMAR_PCI_PATH *)(devscope + 1); if (path_len == 0) { printf("dmar_match_devscope: corrupted DMAR table, dl %d\n", devscope->Length); return (-1); } return (dmar_match_pathes(devscope->Bus, path, path_len, dev_busno, dev_path, dev_path_len, devscope->EntryType)); } static bool dmar_match_by_path(struct dmar_unit *unit, int dev_domain, int dev_busno, const ACPI_DMAR_PCI_PATH *dev_path, int dev_path_len, const char **banner) { ACPI_DMAR_HARDWARE_UNIT *dmarh; ACPI_DMAR_DEVICE_SCOPE *devscope; char *ptr, *ptrend; int match; dmarh = dmar_find_by_index(unit->iommu.unit); if (dmarh == NULL) return (false); if (dmarh->Segment != dev_domain) return (false); if ((dmarh->Flags & ACPI_DMAR_INCLUDE_ALL) != 0) { if (banner != NULL) *banner = "INCLUDE_ALL"; return (true); } ptr = (char *)dmarh + sizeof(*dmarh); ptrend = (char *)dmarh + dmarh->Header.Length; while (ptr < ptrend) { devscope = (ACPI_DMAR_DEVICE_SCOPE *)ptr; ptr += devscope->Length; match = dmar_match_devscope(devscope, dev_busno, dev_path, dev_path_len); if (match == -1) return (false); if (match == 1) { if (banner != NULL) *banner = "specific match"; return (true); } } return (false); } static struct dmar_unit * dmar_find_by_scope(int dev_domain, int dev_busno, const ACPI_DMAR_PCI_PATH *dev_path, int dev_path_len) { struct dmar_unit *unit; int i; for (i = 0; i < dmar_devcnt; i++) { if (dmar_devs[i] == NULL) continue; unit = device_get_softc(dmar_devs[i]); if (dmar_match_by_path(unit, dev_domain, dev_busno, dev_path, dev_path_len, NULL)) return (unit); } return (NULL); } struct dmar_unit * dmar_find(device_t dev, bool verbose) { struct dmar_unit *unit; const char *banner; int i, dev_domain, dev_busno, dev_path_len; /* * This function can only handle PCI(e) devices. */ if (device_get_devclass(device_get_parent(dev)) != devclass_find("pci")) return (NULL); dev_domain = pci_get_domain(dev); dev_path_len = dmar_dev_depth(dev); ACPI_DMAR_PCI_PATH dev_path[dev_path_len]; dmar_dev_path(dev, &dev_busno, dev_path, dev_path_len); banner = ""; for (i = 0; i < dmar_devcnt; i++) { if (dmar_devs[i] == NULL) continue; unit = device_get_softc(dmar_devs[i]); if (dmar_match_by_path(unit, dev_domain, dev_busno, dev_path, dev_path_len, &banner)) break; } if (i == dmar_devcnt) return (NULL); if (verbose) { device_printf(dev, "pci%d:%d:%d:%d matched dmar%d by %s", dev_domain, pci_get_bus(dev), pci_get_slot(dev), pci_get_function(dev), unit->iommu.unit, banner); printf(" scope path "); dmar_print_path(dev_busno, dev_path_len, dev_path); printf("\n"); } return (unit); } static struct dmar_unit * dmar_find_nonpci(u_int id, u_int entry_type, uint16_t *rid) { device_t dmar_dev; struct dmar_unit *unit; ACPI_DMAR_HARDWARE_UNIT *dmarh; ACPI_DMAR_DEVICE_SCOPE *devscope; ACPI_DMAR_PCI_PATH *path; char *ptr, *ptrend; #ifdef DEV_APIC int error; #endif int i; for (i = 0; i < dmar_devcnt; i++) { dmar_dev = dmar_devs[i]; if (dmar_dev == NULL) continue; unit = (struct dmar_unit *)device_get_softc(dmar_dev); dmarh = dmar_find_by_index(i); if (dmarh == NULL) continue; ptr = (char *)dmarh + sizeof(*dmarh); ptrend = (char *)dmarh + dmarh->Header.Length; for (;;) { if (ptr >= ptrend) break; devscope = (ACPI_DMAR_DEVICE_SCOPE *)ptr; ptr += devscope->Length; if (devscope->EntryType != entry_type) continue; if (devscope->EnumerationId != id) continue; #ifdef DEV_APIC if (entry_type == ACPI_DMAR_SCOPE_TYPE_IOAPIC) { error = ioapic_get_rid(id, rid); /* * If our IOAPIC has PCI bindings then * use the PCI device rid. */ if (error == 0) return (unit); } #endif if (devscope->Length - sizeof(ACPI_DMAR_DEVICE_SCOPE) == 2) { if (rid != NULL) { path = (ACPI_DMAR_PCI_PATH *) (devscope + 1); *rid = PCI_RID(devscope->Bus, path->Device, path->Function); } return (unit); } printf( "dmar_find_nonpci: id %d type %d path length != 2\n", id, entry_type); break; } } return (NULL); } struct dmar_unit * dmar_find_hpet(device_t dev, uint16_t *rid) { return (dmar_find_nonpci(hpet_get_uid(dev), ACPI_DMAR_SCOPE_TYPE_HPET, rid)); } struct dmar_unit * dmar_find_ioapic(u_int apic_id, uint16_t *rid) { return (dmar_find_nonpci(apic_id, ACPI_DMAR_SCOPE_TYPE_IOAPIC, rid)); } struct rmrr_iter_args { struct dmar_domain *domain; int dev_domain; int dev_busno; const ACPI_DMAR_PCI_PATH *dev_path; int dev_path_len; struct iommu_map_entries_tailq *rmrr_entries; }; static int dmar_rmrr_iter(ACPI_DMAR_HEADER *dmarh, void *arg) { struct rmrr_iter_args *ria; ACPI_DMAR_RESERVED_MEMORY *resmem; ACPI_DMAR_DEVICE_SCOPE *devscope; struct iommu_map_entry *entry; char *ptr, *ptrend; int match; if (dmarh->Type != ACPI_DMAR_TYPE_RESERVED_MEMORY) return (1); ria = arg; resmem = (ACPI_DMAR_RESERVED_MEMORY *)dmarh; if (resmem->Segment != ria->dev_domain) return (1); ptr = (char *)resmem + sizeof(*resmem); ptrend = (char *)resmem + resmem->Header.Length; for (;;) { if (ptr >= ptrend) break; devscope = (ACPI_DMAR_DEVICE_SCOPE *)ptr; ptr += devscope->Length; match = dmar_match_devscope(devscope, ria->dev_busno, ria->dev_path, ria->dev_path_len); if (match == 1) { entry = iommu_gas_alloc_entry(DOM2IODOM(ria->domain), IOMMU_PGF_WAITOK); entry->start = resmem->BaseAddress; /* The RMRR entry end address is inclusive. */ entry->end = resmem->EndAddress; TAILQ_INSERT_TAIL(ria->rmrr_entries, entry, - unroll_link); + dmamap_link); } } return (1); } void dmar_dev_parse_rmrr(struct dmar_domain *domain, int dev_domain, int dev_busno, const void *dev_path, int dev_path_len, struct iommu_map_entries_tailq *rmrr_entries) { struct rmrr_iter_args ria; ria.domain = domain; ria.dev_domain = dev_domain; ria.dev_busno = dev_busno; ria.dev_path = (const ACPI_DMAR_PCI_PATH *)dev_path; ria.dev_path_len = dev_path_len; ria.rmrr_entries = rmrr_entries; dmar_iterate_tbl(dmar_rmrr_iter, &ria); } struct inst_rmrr_iter_args { struct dmar_unit *dmar; }; static device_t dmar_path_dev(int segment, int path_len, int busno, const ACPI_DMAR_PCI_PATH *path, uint16_t *rid) { device_t dev; int i; dev = NULL; for (i = 0; i < path_len; i++) { dev = pci_find_dbsf(segment, busno, path->Device, path->Function); if (i != path_len - 1) { busno = pci_cfgregread(busno, path->Device, path->Function, PCIR_SECBUS_1, 1); path++; } } *rid = PCI_RID(busno, path->Device, path->Function); return (dev); } static int dmar_inst_rmrr_iter(ACPI_DMAR_HEADER *dmarh, void *arg) { const ACPI_DMAR_RESERVED_MEMORY *resmem; const ACPI_DMAR_DEVICE_SCOPE *devscope; struct inst_rmrr_iter_args *iria; const char *ptr, *ptrend; device_t dev; struct dmar_unit *unit; int dev_path_len; uint16_t rid; iria = arg; if (dmarh->Type != ACPI_DMAR_TYPE_RESERVED_MEMORY) return (1); resmem = (ACPI_DMAR_RESERVED_MEMORY *)dmarh; if (resmem->Segment != iria->dmar->segment) return (1); ptr = (const char *)resmem + sizeof(*resmem); ptrend = (const char *)resmem + resmem->Header.Length; for (;;) { if (ptr >= ptrend) break; devscope = (const ACPI_DMAR_DEVICE_SCOPE *)ptr; ptr += devscope->Length; /* XXXKIB bridge */ if (devscope->EntryType != ACPI_DMAR_SCOPE_TYPE_ENDPOINT) continue; rid = 0; dev_path_len = (devscope->Length - sizeof(ACPI_DMAR_DEVICE_SCOPE)) / 2; dev = dmar_path_dev(resmem->Segment, dev_path_len, devscope->Bus, (const ACPI_DMAR_PCI_PATH *)(devscope + 1), &rid); if (dev == NULL) { if (bootverbose) { printf("dmar%d no dev found for RMRR " "[%#jx, %#jx] rid %#x scope path ", iria->dmar->iommu.unit, (uintmax_t)resmem->BaseAddress, (uintmax_t)resmem->EndAddress, rid); dmar_print_path(devscope->Bus, dev_path_len, (const ACPI_DMAR_PCI_PATH *)(devscope + 1)); printf("\n"); } unit = dmar_find_by_scope(resmem->Segment, devscope->Bus, (const ACPI_DMAR_PCI_PATH *)(devscope + 1), dev_path_len); if (iria->dmar != unit) continue; dmar_get_ctx_for_devpath(iria->dmar, rid, resmem->Segment, devscope->Bus, (const ACPI_DMAR_PCI_PATH *)(devscope + 1), dev_path_len, false, true); } else { unit = dmar_find(dev, false); if (iria->dmar != unit) continue; iommu_instantiate_ctx(&(iria)->dmar->iommu, dev, true); } } return (1); } /* * Pre-create all contexts for the DMAR which have RMRR entries. */ int dmar_instantiate_rmrr_ctxs(struct iommu_unit *unit) { struct dmar_unit *dmar; struct inst_rmrr_iter_args iria; int error; dmar = IOMMU2DMAR(unit); if (!dmar_barrier_enter(dmar, DMAR_BARRIER_RMRR)) return (0); error = 0; iria.dmar = dmar; dmar_iterate_tbl(dmar_inst_rmrr_iter, &iria); DMAR_LOCK(dmar); if (!LIST_EMPTY(&dmar->domains)) { KASSERT((dmar->hw_gcmd & DMAR_GCMD_TE) == 0, ("dmar%d: RMRR not handled but translation is already enabled", dmar->iommu.unit)); error = dmar_enable_translation(dmar); if (bootverbose) { if (error == 0) { printf("dmar%d: enabled translation\n", dmar->iommu.unit); } else { printf("dmar%d: enabling translation failed, " "error %d\n", dmar->iommu.unit, error); } } } dmar_barrier_exit(dmar, DMAR_BARRIER_RMRR); return (error); } #ifdef DDB #include #include static void dmar_print_domain_entry(const struct iommu_map_entry *entry) { struct iommu_map_entry *l, *r; db_printf( " start %jx end %jx first %jx last %jx free_down %jx flags %x ", entry->start, entry->end, entry->first, entry->last, entry->free_down, entry->flags); db_printf("left "); l = RB_LEFT(entry, rb_entry); if (l == NULL) db_printf("NULL "); else db_printf("%jx ", l->start); db_printf("right "); r = RB_RIGHT(entry, rb_entry); if (r == NULL) db_printf("NULL"); else db_printf("%jx", r->start); db_printf("\n"); } static void dmar_print_ctx(struct dmar_ctx *ctx) { db_printf( " @%p pci%d:%d:%d refs %d flags %x loads %lu unloads %lu\n", ctx, pci_get_bus(ctx->context.tag->owner), pci_get_slot(ctx->context.tag->owner), pci_get_function(ctx->context.tag->owner), ctx->refs, ctx->context.flags, ctx->context.loads, ctx->context.unloads); } static void dmar_print_domain(struct dmar_domain *domain, bool show_mappings) { struct iommu_domain *iodom; struct iommu_map_entry *entry; struct dmar_ctx *ctx; iodom = DOM2IODOM(domain); db_printf( " @%p dom %d mgaw %d agaw %d pglvl %d end %jx refs %d\n" " ctx_cnt %d flags %x pgobj %p map_ents %u\n", domain, domain->domain, domain->mgaw, domain->agaw, domain->pglvl, (uintmax_t)domain->iodom.end, domain->refs, domain->ctx_cnt, domain->iodom.flags, domain->pgtbl_obj, domain->iodom.entries_cnt); if (!LIST_EMPTY(&domain->contexts)) { db_printf(" Contexts:\n"); LIST_FOREACH(ctx, &domain->contexts, link) dmar_print_ctx(ctx); } if (!show_mappings) return; db_printf(" mapped:\n"); RB_FOREACH(entry, iommu_gas_entries_tree, &iodom->rb_root) { dmar_print_domain_entry(entry); if (db_pager_quit) break; } if (db_pager_quit) return; db_printf(" unloading:\n"); TAILQ_FOREACH(entry, &domain->iodom.unload_entries, dmamap_link) { dmar_print_domain_entry(entry); if (db_pager_quit) break; } } DB_FUNC(dmar_domain, db_dmar_print_domain, db_show_table, CS_OWN, NULL) { struct dmar_unit *unit; struct dmar_domain *domain; struct dmar_ctx *ctx; bool show_mappings, valid; int pci_domain, bus, device, function, i, t; db_expr_t radix; valid = false; radix = db_radix; db_radix = 10; t = db_read_token(); if (t == tSLASH) { t = db_read_token(); if (t != tIDENT) { db_printf("Bad modifier\n"); db_radix = radix; db_skip_to_eol(); return; } show_mappings = strchr(db_tok_string, 'm') != NULL; t = db_read_token(); } else { show_mappings = false; } if (t == tNUMBER) { pci_domain = db_tok_number; t = db_read_token(); if (t == tNUMBER) { bus = db_tok_number; t = db_read_token(); if (t == tNUMBER) { device = db_tok_number; t = db_read_token(); if (t == tNUMBER) { function = db_tok_number; valid = true; } } } } db_radix = radix; db_skip_to_eol(); if (!valid) { db_printf("usage: show dmar_domain [/m] " " \n"); return; } for (i = 0; i < dmar_devcnt; i++) { unit = device_get_softc(dmar_devs[i]); LIST_FOREACH(domain, &unit->domains, link) { LIST_FOREACH(ctx, &domain->contexts, link) { if (pci_domain == unit->segment && bus == pci_get_bus(ctx->context.tag->owner) && device == pci_get_slot(ctx->context.tag->owner) && function == pci_get_function(ctx->context.tag->owner)) { dmar_print_domain(domain, show_mappings); goto out; } } } } out:; } static void dmar_print_one(int idx, bool show_domains, bool show_mappings) { struct dmar_unit *unit; struct dmar_domain *domain; int i, frir; unit = device_get_softc(dmar_devs[idx]); db_printf("dmar%d at %p, root at 0x%jx, ver 0x%x\n", unit->iommu.unit, unit, dmar_read8(unit, DMAR_RTADDR_REG), dmar_read4(unit, DMAR_VER_REG)); db_printf("cap 0x%jx ecap 0x%jx gsts 0x%x fsts 0x%x fectl 0x%x\n", (uintmax_t)dmar_read8(unit, DMAR_CAP_REG), (uintmax_t)dmar_read8(unit, DMAR_ECAP_REG), dmar_read4(unit, DMAR_GSTS_REG), dmar_read4(unit, DMAR_FSTS_REG), dmar_read4(unit, DMAR_FECTL_REG)); if (unit->ir_enabled) { db_printf("ir is enabled; IRT @%p phys 0x%jx maxcnt %d\n", unit->irt, (uintmax_t)unit->irt_phys, unit->irte_cnt); } db_printf("fed 0x%x fea 0x%x feua 0x%x\n", dmar_read4(unit, DMAR_FEDATA_REG), dmar_read4(unit, DMAR_FEADDR_REG), dmar_read4(unit, DMAR_FEUADDR_REG)); db_printf("primary fault log:\n"); for (i = 0; i < DMAR_CAP_NFR(unit->hw_cap); i++) { frir = (DMAR_CAP_FRO(unit->hw_cap) + i) * 16; db_printf(" %d at 0x%x: %jx %jx\n", i, frir, (uintmax_t)dmar_read8(unit, frir), (uintmax_t)dmar_read8(unit, frir + 8)); } if (DMAR_HAS_QI(unit)) { db_printf("ied 0x%x iea 0x%x ieua 0x%x\n", dmar_read4(unit, DMAR_IEDATA_REG), dmar_read4(unit, DMAR_IEADDR_REG), dmar_read4(unit, DMAR_IEUADDR_REG)); if (unit->qi_enabled) { db_printf("qi is enabled: queue @0x%jx (IQA 0x%jx) " "size 0x%jx\n" " head 0x%x tail 0x%x avail 0x%x status 0x%x ctrl 0x%x\n" " hw compl 0x%x@%p/phys@%jx next seq 0x%x gen 0x%x\n", (uintmax_t)unit->inv_queue, (uintmax_t)dmar_read8(unit, DMAR_IQA_REG), (uintmax_t)unit->inv_queue_size, dmar_read4(unit, DMAR_IQH_REG), dmar_read4(unit, DMAR_IQT_REG), unit->inv_queue_avail, dmar_read4(unit, DMAR_ICS_REG), dmar_read4(unit, DMAR_IECTL_REG), unit->inv_waitd_seq_hw, &unit->inv_waitd_seq_hw, (uintmax_t)unit->inv_waitd_seq_hw_phys, unit->inv_waitd_seq, unit->inv_waitd_gen); } else { db_printf("qi is disabled\n"); } } if (show_domains) { db_printf("domains:\n"); LIST_FOREACH(domain, &unit->domains, link) { dmar_print_domain(domain, show_mappings); if (db_pager_quit) break; } } } DB_SHOW_COMMAND(dmar, db_dmar_print) { bool show_domains, show_mappings; show_domains = strchr(modif, 'd') != NULL; show_mappings = strchr(modif, 'm') != NULL; if (!have_addr) { db_printf("usage: show dmar [/d] [/m] index\n"); return; } dmar_print_one((int)addr, show_domains, show_mappings); } DB_SHOW_ALL_COMMAND(dmars, db_show_all_dmars) { int i; bool show_domains, show_mappings; show_domains = strchr(modif, 'd') != NULL; show_mappings = strchr(modif, 'm') != NULL; for (i = 0; i < dmar_devcnt; i++) { dmar_print_one(i, show_domains, show_mappings); if (db_pager_quit) break; } } #endif struct iommu_unit * iommu_find(device_t dev, bool verbose) { struct dmar_unit *dmar; dmar = dmar_find(dev, verbose); return (&dmar->iommu); }