diff --git a/sys/dev/iommu/iommu.h b/sys/dev/iommu/iommu.h index 65fefe3ada7b..fefd0f615be5 100644 --- a/sys/dev/iommu/iommu.h +++ b/sys/dev/iommu/iommu.h @@ -1,201 +1,204 @@ /*- * 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 */ + union { + TAILQ_ENTRY(iommu_map_entry) dmamap_link; /* DMA map entries */ + struct iommu_map_entry *tlb_flush_next; + }; RB_ENTRY(iommu_map_entry) rb_entry; /* Links for domain entries */ 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, bool cansleep); 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/dev/iommu/iommu_gas.c b/sys/dev/iommu/iommu_gas.c index ec456e2ec48b..bac15edcf849 100644 --- a/sys/dev/iommu/iommu_gas.c +++ b/sys/dev/iommu/iommu_gas.c @@ -1,870 +1,871 @@ /*- * 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("$FreeBSD$"); #define RB_AUGMENT(entry) iommu_gas_augment_entry(entry) #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 /* * Guest Address Space management. */ static uma_zone_t iommu_map_entry_zone; #ifdef INVARIANTS static int iommu_check_free; #endif static void intel_gas_init(void) { iommu_map_entry_zone = uma_zcreate("IOMMU_MAP_ENTRY", sizeof(struct iommu_map_entry), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NODUMP); } SYSINIT(intel_gas, SI_SUB_DRIVERS, SI_ORDER_FIRST, intel_gas_init, NULL); struct iommu_map_entry * iommu_gas_alloc_entry(struct iommu_domain *domain, u_int flags) { struct iommu_map_entry *res; KASSERT((flags & ~(IOMMU_PGF_WAITOK)) == 0, ("unsupported flags %x", flags)); res = uma_zalloc(iommu_map_entry_zone, ((flags & IOMMU_PGF_WAITOK) != 0 ? M_WAITOK : M_NOWAIT) | M_ZERO); - if (res != NULL) { + if (res != NULL && domain != NULL) { res->domain = domain; atomic_add_int(&domain->entries_cnt, 1); } return (res); } void iommu_gas_free_entry(struct iommu_domain *domain, struct iommu_map_entry *entry) { KASSERT(domain == entry->domain, ("mismatched free domain %p entry %p entry->domain %p", domain, entry, entry->domain)); - atomic_subtract_int(&domain->entries_cnt, 1); + if (domain != NULL) + atomic_subtract_int(&domain->entries_cnt, 1); uma_zfree(iommu_map_entry_zone, entry); } static int iommu_gas_cmp_entries(struct iommu_map_entry *a, struct iommu_map_entry *b) { /* Last entry have zero size, so <= */ KASSERT(a->start <= a->end, ("inverted entry %p (%jx, %jx)", a, (uintmax_t)a->start, (uintmax_t)a->end)); KASSERT(b->start <= b->end, ("inverted entry %p (%jx, %jx)", b, (uintmax_t)b->start, (uintmax_t)b->end)); KASSERT(a->end <= b->start || b->end <= a->start || a->end == a->start || b->end == b->start, ("overlapping entries %p (%jx, %jx) %p (%jx, %jx)", a, (uintmax_t)a->start, (uintmax_t)a->end, b, (uintmax_t)b->start, (uintmax_t)b->end)); if (a->end < b->end) return (-1); else if (b->end < a->end) return (1); return (0); } static void iommu_gas_augment_entry(struct iommu_map_entry *entry) { struct iommu_map_entry *child; iommu_gaddr_t free_down; free_down = 0; if ((child = RB_LEFT(entry, rb_entry)) != NULL) { free_down = MAX(free_down, child->free_down); free_down = MAX(free_down, entry->start - child->last); entry->first = child->first; } else entry->first = entry->start; if ((child = RB_RIGHT(entry, rb_entry)) != NULL) { free_down = MAX(free_down, child->free_down); free_down = MAX(free_down, child->first - entry->end); entry->last = child->last; } else entry->last = entry->end; entry->free_down = free_down; } RB_GENERATE(iommu_gas_entries_tree, iommu_map_entry, rb_entry, iommu_gas_cmp_entries); #ifdef INVARIANTS static void iommu_gas_check_free(struct iommu_domain *domain) { struct iommu_map_entry *entry, *l, *r; iommu_gaddr_t v; RB_FOREACH(entry, iommu_gas_entries_tree, &domain->rb_root) { KASSERT(domain == entry->domain, ("mismatched free domain %p entry %p entry->domain %p", domain, entry, entry->domain)); l = RB_LEFT(entry, rb_entry); r = RB_RIGHT(entry, rb_entry); v = 0; if (l != NULL) { v = MAX(v, l->free_down); v = MAX(v, entry->start - l->last); } if (r != NULL) { v = MAX(v, r->free_down); v = MAX(v, r->first - entry->end); } MPASS(entry->free_down == v); } } #endif static bool iommu_gas_rb_insert(struct iommu_domain *domain, struct iommu_map_entry *entry) { struct iommu_map_entry *found; found = RB_INSERT(iommu_gas_entries_tree, &domain->rb_root, entry); return (found == NULL); } static void iommu_gas_rb_remove(struct iommu_domain *domain, struct iommu_map_entry *entry) { RB_REMOVE(iommu_gas_entries_tree, &domain->rb_root, entry); } struct iommu_domain * iommu_get_ctx_domain(struct iommu_ctx *ctx) { return (ctx->domain); } void iommu_gas_init_domain(struct iommu_domain *domain) { struct iommu_map_entry *begin, *end; begin = iommu_gas_alloc_entry(domain, IOMMU_PGF_WAITOK); end = iommu_gas_alloc_entry(domain, IOMMU_PGF_WAITOK); IOMMU_DOMAIN_LOCK(domain); KASSERT(domain->entries_cnt == 2, ("dirty domain %p", domain)); KASSERT(RB_EMPTY(&domain->rb_root), ("non-empty entries %p", domain)); begin->start = 0; begin->end = IOMMU_PAGE_SIZE; begin->flags = IOMMU_MAP_ENTRY_PLACE | IOMMU_MAP_ENTRY_UNMAPPED; iommu_gas_rb_insert(domain, begin); end->start = domain->end; end->end = domain->end; end->flags = IOMMU_MAP_ENTRY_PLACE | IOMMU_MAP_ENTRY_UNMAPPED; iommu_gas_rb_insert(domain, end); domain->first_place = begin; domain->last_place = end; domain->flags |= IOMMU_DOMAIN_GAS_INITED; IOMMU_DOMAIN_UNLOCK(domain); } void iommu_gas_fini_domain(struct iommu_domain *domain) { struct iommu_map_entry *entry, *entry1; IOMMU_DOMAIN_ASSERT_LOCKED(domain); KASSERT(domain->entries_cnt == 2, ("domain still in use %p", domain)); entry = RB_MIN(iommu_gas_entries_tree, &domain->rb_root); KASSERT(entry->start == 0, ("start entry start %p", domain)); KASSERT(entry->end == IOMMU_PAGE_SIZE, ("start entry end %p", domain)); KASSERT(entry->flags == (IOMMU_MAP_ENTRY_PLACE | IOMMU_MAP_ENTRY_UNMAPPED), ("start entry flags %p", domain)); RB_REMOVE(iommu_gas_entries_tree, &domain->rb_root, entry); iommu_gas_free_entry(domain, entry); entry = RB_MAX(iommu_gas_entries_tree, &domain->rb_root); KASSERT(entry->start == domain->end, ("end entry start %p", domain)); KASSERT(entry->end == domain->end, ("end entry end %p", domain)); KASSERT(entry->flags == (IOMMU_MAP_ENTRY_PLACE | IOMMU_MAP_ENTRY_UNMAPPED), ("end entry flags %p", domain)); RB_REMOVE(iommu_gas_entries_tree, &domain->rb_root, entry); iommu_gas_free_entry(domain, entry); RB_FOREACH_SAFE(entry, iommu_gas_entries_tree, &domain->rb_root, entry1) { KASSERT((entry->flags & IOMMU_MAP_ENTRY_RMRR) != 0, ("non-RMRR entry left %p", domain)); RB_REMOVE(iommu_gas_entries_tree, &domain->rb_root, entry); iommu_gas_free_entry(domain, entry); } } struct iommu_gas_match_args { struct iommu_domain *domain; iommu_gaddr_t size; int offset; const struct bus_dma_tag_common *common; u_int gas_flags; struct iommu_map_entry *entry; }; /* * The interval [beg, end) is a free interval between two iommu_map_entries. * Addresses can be allocated only in the range [lbound, ubound). Try to * allocate space in the free interval, subject to the conditions expressed by * a, and return 'true' if and only if the allocation attempt succeeds. */ static bool iommu_gas_match_one(struct iommu_gas_match_args *a, iommu_gaddr_t beg, iommu_gaddr_t end, iommu_gaddr_t lbound, iommu_gaddr_t ubound) { struct iommu_map_entry *entry; iommu_gaddr_t first, size, start; bool found __diagused; int offset; /* * The prev->end is always aligned on the page size, which * causes page alignment for the entry->start too. * * Create IOMMU_PAGE_SIZE gaps before, after new entry * to ensure that out-of-bounds accesses fault. */ beg = MAX(beg + IOMMU_PAGE_SIZE, lbound); start = roundup2(beg, a->common->alignment); if (start < beg) return (false); end = MIN(end - IOMMU_PAGE_SIZE, ubound); offset = a->offset; size = a->size; if (start + offset + size > end) return (false); /* Check for and try to skip past boundary crossing. */ if (!vm_addr_bound_ok(start + offset, size, a->common->boundary)) { /* * The start + offset to start + offset + size region crosses * the boundary. Check if there is enough space after the next * boundary after the beg. */ first = start; beg = roundup2(start + offset + 1, a->common->boundary); start = roundup2(beg, a->common->alignment); if (start + offset + size > end || !vm_addr_bound_ok(start + offset, size, a->common->boundary)) { /* * Not enough space to align at the requested boundary, * or boundary is smaller than the size, but allowed to * split. We already checked that start + size does not * overlap ubound. * * XXXKIB. It is possible that beg is exactly at the * start of the next entry, then we do not have gap. * Ignore for now. */ if ((a->gas_flags & IOMMU_MF_CANSPLIT) == 0) return (false); size = beg - first - offset; start = first; } } entry = a->entry; entry->start = start; entry->end = start + roundup2(size + offset, IOMMU_PAGE_SIZE); entry->flags = IOMMU_MAP_ENTRY_MAP; found = iommu_gas_rb_insert(a->domain, entry); KASSERT(found, ("found dup %p start %jx size %jx", a->domain, (uintmax_t)start, (uintmax_t)size)); return (true); } /* Find the next entry that might abut a big-enough range. */ static struct iommu_map_entry * iommu_gas_next(struct iommu_map_entry *curr, iommu_gaddr_t min_free) { struct iommu_map_entry *next; if ((next = RB_RIGHT(curr, rb_entry)) != NULL && next->free_down >= min_free) { /* Find next entry in right subtree. */ do curr = next; while ((next = RB_LEFT(curr, rb_entry)) != NULL && next->free_down >= min_free); } else { /* Find next entry in a left-parent ancestor. */ while ((next = RB_PARENT(curr, rb_entry)) != NULL && curr == RB_RIGHT(next, rb_entry)) curr = next; curr = next; } return (curr); } static int iommu_gas_find_space(struct iommu_gas_match_args *a) { struct iommu_domain *domain; struct iommu_map_entry *curr, *first; iommu_gaddr_t addr, min_free; IOMMU_DOMAIN_ASSERT_LOCKED(a->domain); KASSERT(a->entry->flags == 0, ("dirty entry %p %p", a->domain, a->entry)); /* * If the subtree doesn't have free space for the requested allocation * plus two guard pages, skip it. */ min_free = 2 * IOMMU_PAGE_SIZE + roundup2(a->size + a->offset, IOMMU_PAGE_SIZE); /* * Find the first entry in the lower region that could abut a big-enough * range. */ curr = RB_ROOT(&a->domain->rb_root); first = NULL; while (curr != NULL && curr->free_down >= min_free) { first = curr; curr = RB_LEFT(curr, rb_entry); } /* * Walk the big-enough ranges until one satisfies alignment * requirements, or violates lowaddr address requirement. */ addr = a->common->lowaddr + 1; for (curr = first; curr != NULL; curr = iommu_gas_next(curr, min_free)) { if ((first = RB_LEFT(curr, rb_entry)) != NULL && iommu_gas_match_one(a, first->last, curr->start, 0, addr)) return (0); if (curr->end >= addr) { /* All remaining ranges >= addr */ break; } if ((first = RB_RIGHT(curr, rb_entry)) != NULL && iommu_gas_match_one(a, curr->end, first->first, 0, addr)) return (0); } /* * To resume the search at the start of the upper region, first climb to * the nearest ancestor that spans highaddr. Then find the last entry * before highaddr that could abut a big-enough range. */ addr = a->common->highaddr; while (curr != NULL && curr->last < addr) curr = RB_PARENT(curr, rb_entry); first = NULL; while (curr != NULL && curr->free_down >= min_free) { if (addr < curr->end) curr = RB_LEFT(curr, rb_entry); else { first = curr; curr = RB_RIGHT(curr, rb_entry); } } /* * Walk the remaining big-enough ranges until one satisfies alignment * requirements. */ domain = a->domain; for (curr = first; curr != NULL; curr = iommu_gas_next(curr, min_free)) { if ((first = RB_LEFT(curr, rb_entry)) != NULL && iommu_gas_match_one(a, first->last, curr->start, addr + 1, domain->end)) return (0); if ((first = RB_RIGHT(curr, rb_entry)) != NULL && iommu_gas_match_one(a, curr->end, first->first, addr + 1, domain->end)) return (0); } return (ENOMEM); } static int iommu_gas_alloc_region(struct iommu_domain *domain, struct iommu_map_entry *entry, u_int flags) { struct iommu_map_entry *next, *prev; bool found __diagused; IOMMU_DOMAIN_ASSERT_LOCKED(domain); if ((entry->start & IOMMU_PAGE_MASK) != 0 || (entry->end & IOMMU_PAGE_MASK) != 0) return (EINVAL); if (entry->start >= entry->end) return (EINVAL); if (entry->end >= domain->end) return (EINVAL); next = RB_NFIND(iommu_gas_entries_tree, &domain->rb_root, entry); KASSERT(next != NULL, ("next must be non-null %p %jx", domain, (uintmax_t)entry->start)); prev = RB_PREV(iommu_gas_entries_tree, &domain->rb_root, next); /* prev could be NULL */ /* * Adapt to broken BIOSes which specify overlapping RMRR * entries. * * XXXKIB: this does not handle a case when prev or next * entries are completely covered by the current one, which * extends both ways. */ if (prev != NULL && prev->end > entry->start && (prev->flags & IOMMU_MAP_ENTRY_PLACE) == 0) { if ((flags & IOMMU_MF_RMRR) == 0 || (prev->flags & IOMMU_MAP_ENTRY_RMRR) == 0) return (EBUSY); entry->start = prev->end; } if (next->start < entry->end && (next->flags & IOMMU_MAP_ENTRY_PLACE) == 0) { if ((flags & IOMMU_MF_RMRR) == 0 || (next->flags & IOMMU_MAP_ENTRY_RMRR) == 0) return (EBUSY); entry->end = next->start; } if (entry->end == entry->start) return (0); if (prev != NULL && prev->end > entry->start) { /* This assumes that prev is the placeholder entry. */ iommu_gas_rb_remove(domain, prev); prev = NULL; } if (next->start < entry->end) { iommu_gas_rb_remove(domain, next); next = NULL; } found = iommu_gas_rb_insert(domain, entry); KASSERT(found, ("found RMRR dup %p start %jx end %jx", domain, (uintmax_t)entry->start, (uintmax_t)entry->end)); if ((flags & IOMMU_MF_RMRR) != 0) entry->flags = IOMMU_MAP_ENTRY_RMRR; #ifdef INVARIANTS struct iommu_map_entry *ip, *in; ip = RB_PREV(iommu_gas_entries_tree, &domain->rb_root, entry); in = RB_NEXT(iommu_gas_entries_tree, &domain->rb_root, entry); KASSERT(prev == NULL || ip == prev, ("RMRR %p (%jx %jx) prev %p (%jx %jx) ins prev %p (%jx %jx)", entry, entry->start, entry->end, prev, prev == NULL ? 0 : prev->start, prev == NULL ? 0 : prev->end, ip, ip == NULL ? 0 : ip->start, ip == NULL ? 0 : ip->end)); KASSERT(next == NULL || in == next, ("RMRR %p (%jx %jx) next %p (%jx %jx) ins next %p (%jx %jx)", entry, entry->start, entry->end, next, next == NULL ? 0 : next->start, next == NULL ? 0 : next->end, in, in == NULL ? 0 : in->start, in == NULL ? 0 : in->end)); #endif return (0); } void iommu_gas_free_space(struct iommu_domain *domain, struct iommu_map_entry *entry) { IOMMU_DOMAIN_ASSERT_LOCKED(domain); KASSERT((entry->flags & (IOMMU_MAP_ENTRY_PLACE | IOMMU_MAP_ENTRY_RMRR | IOMMU_MAP_ENTRY_MAP)) == IOMMU_MAP_ENTRY_MAP, ("permanent entry %p %p", domain, entry)); iommu_gas_rb_remove(domain, entry); entry->flags &= ~IOMMU_MAP_ENTRY_MAP; #ifdef INVARIANTS if (iommu_check_free) iommu_gas_check_free(domain); #endif } void iommu_gas_free_region(struct iommu_domain *domain, struct iommu_map_entry *entry) { struct iommu_map_entry *next, *prev; IOMMU_DOMAIN_ASSERT_LOCKED(domain); KASSERT((entry->flags & (IOMMU_MAP_ENTRY_PLACE | IOMMU_MAP_ENTRY_RMRR | IOMMU_MAP_ENTRY_MAP)) == IOMMU_MAP_ENTRY_RMRR, ("non-RMRR entry %p %p", domain, entry)); prev = RB_PREV(iommu_gas_entries_tree, &domain->rb_root, entry); next = RB_NEXT(iommu_gas_entries_tree, &domain->rb_root, entry); iommu_gas_rb_remove(domain, entry); entry->flags &= ~IOMMU_MAP_ENTRY_RMRR; if (prev == NULL) iommu_gas_rb_insert(domain, domain->first_place); if (next == NULL) iommu_gas_rb_insert(domain, domain->last_place); } 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) { struct iommu_gas_match_args a; struct iommu_map_entry *entry; int error; KASSERT((flags & ~(IOMMU_MF_CANWAIT | IOMMU_MF_CANSPLIT)) == 0, ("invalid flags 0x%x", flags)); a.domain = domain; a.size = size; a.offset = offset; a.common = common; a.gas_flags = flags; entry = iommu_gas_alloc_entry(domain, (flags & IOMMU_MF_CANWAIT) != 0 ? IOMMU_PGF_WAITOK : 0); if (entry == NULL) return (ENOMEM); a.entry = entry; IOMMU_DOMAIN_LOCK(domain); error = iommu_gas_find_space(&a); if (error == ENOMEM) { IOMMU_DOMAIN_UNLOCK(domain); iommu_gas_free_entry(domain, entry); return (error); } #ifdef INVARIANTS if (iommu_check_free) iommu_gas_check_free(domain); #endif KASSERT(error == 0, ("unexpected error %d from iommu_gas_find_entry", error)); KASSERT(entry->end < domain->end, ("allocated GPA %jx, max GPA %jx", (uintmax_t)entry->end, (uintmax_t)domain->end)); entry->flags |= eflags; IOMMU_DOMAIN_UNLOCK(domain); error = domain->ops->map(domain, entry->start, entry->end - entry->start, ma, eflags, ((flags & IOMMU_MF_CANWAIT) != 0 ? IOMMU_PGF_WAITOK : 0)); if (error == ENOMEM) { iommu_domain_unload_entry(entry, true, (flags & IOMMU_MF_CANWAIT) != 0); return (error); } KASSERT(error == 0, ("unexpected error %d from domain_map_buf", error)); *res = entry; return (0); } int iommu_gas_map_region(struct iommu_domain *domain, struct iommu_map_entry *entry, u_int eflags, u_int flags, vm_page_t *ma) { iommu_gaddr_t start; int error; KASSERT(entry->flags == 0, ("used RMRR entry %p %p %x", domain, entry, entry->flags)); KASSERT((flags & ~(IOMMU_MF_CANWAIT | IOMMU_MF_RMRR)) == 0, ("invalid flags 0x%x", flags)); start = entry->start; IOMMU_DOMAIN_LOCK(domain); error = iommu_gas_alloc_region(domain, entry, flags); if (error != 0) { IOMMU_DOMAIN_UNLOCK(domain); return (error); } entry->flags |= eflags; IOMMU_DOMAIN_UNLOCK(domain); if (entry->end == entry->start) return (0); error = domain->ops->map(domain, entry->start, entry->end - entry->start, ma + OFF_TO_IDX(start - entry->start), eflags, ((flags & IOMMU_MF_CANWAIT) != 0 ? IOMMU_PGF_WAITOK : 0)); if (error == ENOMEM) { iommu_domain_unload_entry(entry, false, (flags & IOMMU_MF_CANWAIT) != 0); return (error); } KASSERT(error == 0, ("unexpected error %d from domain_map_buf", error)); return (0); } static int iommu_gas_reserve_region_locked(struct iommu_domain *domain, iommu_gaddr_t start, iommu_gaddr_t end, struct iommu_map_entry *entry) { int error; IOMMU_DOMAIN_ASSERT_LOCKED(domain); entry->start = start; entry->end = end; error = iommu_gas_alloc_region(domain, entry, IOMMU_MF_CANWAIT); if (error == 0) entry->flags |= IOMMU_MAP_ENTRY_UNMAPPED; return (error); } int iommu_gas_reserve_region(struct iommu_domain *domain, iommu_gaddr_t start, iommu_gaddr_t end, struct iommu_map_entry **entry0) { struct iommu_map_entry *entry; int error; entry = iommu_gas_alloc_entry(domain, IOMMU_PGF_WAITOK); IOMMU_DOMAIN_LOCK(domain); error = iommu_gas_reserve_region_locked(domain, start, end, entry); IOMMU_DOMAIN_UNLOCK(domain); if (error != 0) iommu_gas_free_entry(domain, entry); else if (entry0 != NULL) *entry0 = entry; return (error); } /* * As in iommu_gas_reserve_region, reserve [start, end), but allow for existing * entries. */ int iommu_gas_reserve_region_extend(struct iommu_domain *domain, iommu_gaddr_t start, iommu_gaddr_t end) { struct iommu_map_entry *entry, *next, *prev, key = {}; iommu_gaddr_t entry_start, entry_end; int error; error = 0; entry = NULL; end = ummin(end, domain->end); while (start < end) { /* Preallocate an entry. */ if (entry == NULL) entry = iommu_gas_alloc_entry(domain, IOMMU_PGF_WAITOK); /* Calculate the free region from here to the next entry. */ key.start = key.end = start; IOMMU_DOMAIN_LOCK(domain); next = RB_NFIND(iommu_gas_entries_tree, &domain->rb_root, &key); KASSERT(next != NULL, ("domain %p with end %#jx has no entry " "after %#jx", domain, (uintmax_t)domain->end, (uintmax_t)start)); entry_end = ummin(end, next->start); prev = RB_PREV(iommu_gas_entries_tree, &domain->rb_root, next); if (prev != NULL) entry_start = ummax(start, prev->end); else entry_start = start; start = next->end; /* Reserve the region if non-empty. */ if (entry_start != entry_end) { error = iommu_gas_reserve_region_locked(domain, entry_start, entry_end, entry); if (error != 0) { IOMMU_DOMAIN_UNLOCK(domain); break; } entry = NULL; } IOMMU_DOMAIN_UNLOCK(domain); } /* Release a preallocated entry if it was not used. */ if (entry != NULL) iommu_gas_free_entry(domain, entry); return (error); } void iommu_unmap_msi(struct iommu_ctx *ctx) { struct iommu_map_entry *entry; struct iommu_domain *domain; domain = ctx->domain; entry = domain->msi_entry; if (entry == NULL) return; domain->ops->unmap(domain, entry->start, entry->end - entry->start, IOMMU_PGF_WAITOK); IOMMU_DOMAIN_LOCK(domain); iommu_gas_free_space(domain, entry); IOMMU_DOMAIN_UNLOCK(domain); iommu_gas_free_entry(domain, entry); domain->msi_entry = NULL; domain->msi_base = 0; domain->msi_phys = 0; } int iommu_map_msi(struct iommu_ctx *ctx, iommu_gaddr_t size, int offset, u_int eflags, u_int flags, vm_page_t *ma) { struct iommu_domain *domain; struct iommu_map_entry *entry; int error; error = 0; domain = ctx->domain; /* Check if there is already an MSI page allocated */ IOMMU_DOMAIN_LOCK(domain); entry = domain->msi_entry; IOMMU_DOMAIN_UNLOCK(domain); if (entry == NULL) { error = iommu_gas_map(domain, &ctx->tag->common, size, offset, eflags, flags, ma, &entry); IOMMU_DOMAIN_LOCK(domain); if (error == 0) { if (domain->msi_entry == NULL) { MPASS(domain->msi_base == 0); MPASS(domain->msi_phys == 0); domain->msi_entry = entry; domain->msi_base = entry->start; domain->msi_phys = VM_PAGE_TO_PHYS(ma[0]); } else { /* * We lost the race and already have an * MSI page allocated. Free the unneeded entry. */ iommu_gas_free_entry(domain, entry); } } else if (domain->msi_entry != NULL) { /* * The allocation failed, but another succeeded. * Return success as there is a valid MSI page. */ error = 0; } IOMMU_DOMAIN_UNLOCK(domain); } return (error); } void iommu_translate_msi(struct iommu_domain *domain, uint64_t *addr) { *addr = (*addr - domain->msi_phys) + domain->msi_base; KASSERT(*addr >= domain->msi_entry->start, ("%s: Address is below the MSI entry start address (%jx < %jx)", __func__, (uintmax_t)*addr, (uintmax_t)domain->msi_entry->start)); KASSERT(*addr + sizeof(*addr) <= domain->msi_entry->end, ("%s: Address is above the MSI entry end address (%jx < %jx)", __func__, (uintmax_t)*addr, (uintmax_t)domain->msi_entry->end)); } SYSCTL_NODE(_hw, OID_AUTO, iommu, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, ""); #ifdef INVARIANTS SYSCTL_INT(_hw_iommu, OID_AUTO, check_free, CTLFLAG_RWTUN, &iommu_check_free, 0, "Check the GPA RBtree for free_down and free_after validity"); #endif diff --git a/sys/x86/iommu/intel_ctx.c b/sys/x86/iommu/intel_ctx.c index 98dfed7aaa17..b36531ec6d17 100644 --- a/sys/x86/iommu/intel_ctx.c +++ b/sys/x86/iommu/intel_ctx.c @@ -1,984 +1,984 @@ /*- * 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, 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, 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; } +/* + * If the given value for "free" is true, then the caller must not be using + * the entry's dmamap_link field. + */ void iommu_domain_unload_entry(struct iommu_map_entry *entry, bool free, bool cansleep) { struct dmar_domain *domain; struct dmar_unit *unit; domain = IODOM2DOM(entry->domain); unit = DOM2DMAR(domain); /* * If "free" is false, then the IOTLB invalidation must be performed * synchronously. Otherwise, the caller might free the entry before * dmar_qi_task() is finished processing it. */ if (unit->qi_enabled) { if (free) { DMAR_LOCK(unit); - dmar_qi_invalidate_locked(domain, entry->start, - entry->end - entry->start, &entry->gseq, true); - TAILQ_INSERT_TAIL(&unit->tlb_flush_entries, entry, - dmamap_link); + dmar_qi_invalidate_locked(domain, entry, true); DMAR_UNLOCK(unit); } else { dmar_qi_invalidate_sync(domain, entry->start, entry->end - entry->start, cansleep); dmar_domain_free_entry(entry, false); } } else { domain_flush_iotlb_sync(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, + while ((entry = TAILQ_FIRST(entries)) != NULL) { + TAILQ_REMOVE(entries, entry, dmamap_link); + dmar_qi_invalidate_locked(domain, entry, 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_dmar.h b/sys/x86/iommu/intel_dmar.h index 2840646f89a6..9fc6ac436677 100644 --- a/sys/x86/iommu/intel_dmar.h +++ b/sys/x86/iommu/intel_dmar.h @@ -1,456 +1,481 @@ /*- * 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. * * $FreeBSD$ */ #ifndef __X86_IOMMU_INTEL_DMAR_H #define __X86_IOMMU_INTEL_DMAR_H #include struct dmar_unit; /* * Locking annotations: * (u) - Protected by iommu unit lock * (d) - Protected by domain lock * (c) - Immutable after initialization */ /* * The domain abstraction. Most non-constant members of the domain * are protected by owning dmar unit lock, not by the domain lock. * Most important, the dmar lock protects the contexts list. * * The domain lock protects the address map for the domain, and list * of unload entries delayed. * * Page tables pages and pages content is protected by the vm object * lock pgtbl_obj, which contains the page tables pages. */ struct dmar_domain { struct iommu_domain iodom; int domain; /* (c) DID, written in context entry */ int mgaw; /* (c) Real max address width */ int agaw; /* (c) Adjusted guest address width */ int pglvl; /* (c) The pagelevel */ int awlvl; /* (c) The pagelevel as the bitmask, to set in context entry */ u_int ctx_cnt; /* (u) Number of contexts owned */ u_int refs; /* (u) Refs, including ctx */ struct dmar_unit *dmar; /* (c) */ LIST_ENTRY(dmar_domain) link; /* (u) Member in the dmar list */ LIST_HEAD(, dmar_ctx) contexts; /* (u) */ vm_object_t pgtbl_obj; /* (c) Page table pages */ u_int batch_no; }; struct dmar_ctx { struct iommu_ctx context; uint64_t last_fault_rec[2]; /* Last fault reported */ LIST_ENTRY(dmar_ctx) link; /* (u) Member in the domain list */ u_int refs; /* (u) References from tags */ }; #define DMAR_DOMAIN_PGLOCK(dom) VM_OBJECT_WLOCK((dom)->pgtbl_obj) #define DMAR_DOMAIN_PGTRYLOCK(dom) VM_OBJECT_TRYWLOCK((dom)->pgtbl_obj) #define DMAR_DOMAIN_PGUNLOCK(dom) VM_OBJECT_WUNLOCK((dom)->pgtbl_obj) #define DMAR_DOMAIN_ASSERT_PGLOCKED(dom) \ VM_OBJECT_ASSERT_WLOCKED((dom)->pgtbl_obj) #define DMAR_DOMAIN_LOCK(dom) mtx_lock(&(dom)->iodom.lock) #define DMAR_DOMAIN_UNLOCK(dom) mtx_unlock(&(dom)->iodom.lock) #define DMAR_DOMAIN_ASSERT_LOCKED(dom) mtx_assert(&(dom)->iodom.lock, MA_OWNED) #define DMAR2IOMMU(dmar) &((dmar)->iommu) #define IOMMU2DMAR(dmar) \ __containerof((dmar), struct dmar_unit, iommu) #define DOM2IODOM(domain) &((domain)->iodom) #define IODOM2DOM(domain) \ __containerof((domain), struct dmar_domain, iodom) #define CTX2IOCTX(ctx) &((ctx)->context) #define IOCTX2CTX(ctx) \ __containerof((ctx), struct dmar_ctx, context) #define CTX2DOM(ctx) IODOM2DOM((ctx)->context.domain) #define CTX2DMAR(ctx) (CTX2DOM(ctx)->dmar) #define DOM2DMAR(domain) ((domain)->dmar) struct dmar_msi_data { int irq; int irq_rid; struct resource *irq_res; void *intr_handle; int (*handler)(void *); int msi_data_reg; int msi_addr_reg; int msi_uaddr_reg; void (*enable_intr)(struct dmar_unit *); void (*disable_intr)(struct dmar_unit *); const char *name; }; #define DMAR_INTR_FAULT 0 #define DMAR_INTR_QI 1 #define DMAR_INTR_TOTAL 2 struct dmar_unit { struct iommu_unit iommu; device_t dev; uint16_t segment; uint64_t base; /* Resources */ int reg_rid; struct resource *regs; struct dmar_msi_data intrs[DMAR_INTR_TOTAL]; /* Hardware registers cache */ uint32_t hw_ver; uint64_t hw_cap; uint64_t hw_ecap; uint32_t hw_gcmd; /* Data for being a dmar */ LIST_HEAD(, dmar_domain) domains; struct unrhdr *domids; vm_object_t ctx_obj; u_int barrier_flags; /* Fault handler data */ struct mtx fault_lock; uint64_t *fault_log; int fault_log_head; int fault_log_tail; int fault_log_size; struct task fault_task; struct taskqueue *fault_taskqueue; /* QI */ int qi_enabled; vm_offset_t inv_queue; vm_size_t inv_queue_size; uint32_t inv_queue_avail; uint32_t inv_queue_tail; volatile uint32_t inv_waitd_seq_hw; /* hw writes there on wait descr completion */ uint64_t inv_waitd_seq_hw_phys; uint32_t inv_waitd_seq; /* next sequence number to use for wait descr */ u_int inv_waitd_gen; /* seq number generation AKA seq overflows */ u_int inv_seq_waiters; /* count of waiters for seq */ u_int inv_queue_full; /* informational counter */ /* IR */ int ir_enabled; vm_paddr_t irt_phys; dmar_irte_t *irt; u_int irte_cnt; vmem_t *irtids; - /* Delayed freeing of map entries queue processing */ - struct iommu_map_entries_tailq tlb_flush_entries; + /* + * Delayed freeing of map entries queue processing: + * + * tlb_flush_head and tlb_flush_tail are used to implement a FIFO + * queue that supports concurrent dequeues and enqueues. However, + * there can only be a single dequeuer (accessing tlb_flush_head) and + * a single enqueuer (accessing tlb_flush_tail) at a time. Since the + * unit's qi_task is the only dequeuer, it can access tlb_flush_head + * without any locking. In contrast, there may be multiple enqueuers, + * so the enqueuers acquire the iommu unit lock to serialize their + * accesses to tlb_flush_tail. + * + * In this FIFO queue implementation, the key to enabling concurrent + * dequeues and enqueues is that the dequeuer never needs to access + * tlb_flush_tail and the enqueuer never needs to access + * tlb_flush_head. In particular, tlb_flush_head and tlb_flush_tail + * are never NULL, so neither a dequeuer nor an enqueuer ever needs to + * update both. Instead, tlb_flush_head always points to a "zombie" + * struct, which previously held the last dequeued item. Thus, the + * zombie's next field actually points to the struct holding the first + * item in the queue. When an item is dequeued, the current zombie is + * finally freed, and the struct that held the just dequeued item + * becomes the new zombie. When the queue is empty, tlb_flush_tail + * also points to the zombie. + */ + struct iommu_map_entry *tlb_flush_head; + struct iommu_map_entry *tlb_flush_tail; struct task qi_task; struct taskqueue *qi_taskqueue; }; #define DMAR_LOCK(dmar) mtx_lock(&(dmar)->iommu.lock) #define DMAR_UNLOCK(dmar) mtx_unlock(&(dmar)->iommu.lock) #define DMAR_ASSERT_LOCKED(dmar) mtx_assert(&(dmar)->iommu.lock, MA_OWNED) #define DMAR_FAULT_LOCK(dmar) mtx_lock_spin(&(dmar)->fault_lock) #define DMAR_FAULT_UNLOCK(dmar) mtx_unlock_spin(&(dmar)->fault_lock) #define DMAR_FAULT_ASSERT_LOCKED(dmar) mtx_assert(&(dmar)->fault_lock, MA_OWNED) #define DMAR_IS_COHERENT(dmar) (((dmar)->hw_ecap & DMAR_ECAP_C) != 0) #define DMAR_HAS_QI(dmar) (((dmar)->hw_ecap & DMAR_ECAP_QI) != 0) #define DMAR_X2APIC(dmar) \ (x2apic_mode && ((dmar)->hw_ecap & DMAR_ECAP_EIM) != 0) /* Barrier ids */ #define DMAR_BARRIER_RMRR 0 #define DMAR_BARRIER_USEQ 1 struct dmar_unit *dmar_find(device_t dev, bool verbose); struct dmar_unit *dmar_find_hpet(device_t dev, uint16_t *rid); struct dmar_unit *dmar_find_ioapic(u_int apic_id, uint16_t *rid); u_int dmar_nd2mask(u_int nd); bool dmar_pglvl_supported(struct dmar_unit *unit, int pglvl); int domain_set_agaw(struct dmar_domain *domain, int mgaw); int dmar_maxaddr2mgaw(struct dmar_unit *unit, iommu_gaddr_t maxaddr, bool allow_less); vm_pindex_t pglvl_max_pages(int pglvl); int domain_is_sp_lvl(struct dmar_domain *domain, int lvl); iommu_gaddr_t pglvl_page_size(int total_pglvl, int lvl); iommu_gaddr_t domain_page_size(struct dmar_domain *domain, int lvl); int calc_am(struct dmar_unit *unit, iommu_gaddr_t base, iommu_gaddr_t size, iommu_gaddr_t *isizep); struct vm_page *dmar_pgalloc(vm_object_t obj, vm_pindex_t idx, int flags); void dmar_pgfree(vm_object_t obj, vm_pindex_t idx, int flags); void *dmar_map_pgtbl(vm_object_t obj, vm_pindex_t idx, int flags, struct sf_buf **sf); void dmar_unmap_pgtbl(struct sf_buf *sf); int dmar_load_root_entry_ptr(struct dmar_unit *unit); int dmar_inv_ctx_glob(struct dmar_unit *unit); int dmar_inv_iotlb_glob(struct dmar_unit *unit); int dmar_flush_write_bufs(struct dmar_unit *unit); void dmar_flush_pte_to_ram(struct dmar_unit *unit, dmar_pte_t *dst); void dmar_flush_ctx_to_ram(struct dmar_unit *unit, dmar_ctx_entry_t *dst); void dmar_flush_root_to_ram(struct dmar_unit *unit, dmar_root_entry_t *dst); int dmar_enable_translation(struct dmar_unit *unit); int dmar_disable_translation(struct dmar_unit *unit); int dmar_load_irt_ptr(struct dmar_unit *unit); int dmar_enable_ir(struct dmar_unit *unit); int dmar_disable_ir(struct dmar_unit *unit); bool dmar_barrier_enter(struct dmar_unit *dmar, u_int barrier_id); void dmar_barrier_exit(struct dmar_unit *dmar, u_int barrier_id); uint64_t dmar_get_timeout(void); void dmar_update_timeout(uint64_t newval); int dmar_fault_intr(void *arg); void dmar_enable_fault_intr(struct dmar_unit *unit); void dmar_disable_fault_intr(struct dmar_unit *unit); int dmar_init_fault_log(struct dmar_unit *unit); void dmar_fini_fault_log(struct dmar_unit *unit); int dmar_qi_intr(void *arg); void dmar_enable_qi_intr(struct dmar_unit *unit); void dmar_disable_qi_intr(struct dmar_unit *unit); int dmar_init_qi(struct dmar_unit *unit); void dmar_fini_qi(struct dmar_unit *unit); -void dmar_qi_invalidate_locked(struct dmar_domain *domain, iommu_gaddr_t start, - iommu_gaddr_t size, struct iommu_qi_genseq *psec, bool emit_wait); +void dmar_qi_invalidate_locked(struct dmar_domain *domain, + struct iommu_map_entry *entry, bool emit_wait); void dmar_qi_invalidate_sync(struct dmar_domain *domain, iommu_gaddr_t start, iommu_gaddr_t size, bool cansleep); void dmar_qi_invalidate_ctx_glob_locked(struct dmar_unit *unit); void dmar_qi_invalidate_iotlb_glob_locked(struct dmar_unit *unit); void dmar_qi_invalidate_iec_glob(struct dmar_unit *unit); void dmar_qi_invalidate_iec(struct dmar_unit *unit, u_int start, u_int cnt); vm_object_t domain_get_idmap_pgtbl(struct dmar_domain *domain, iommu_gaddr_t maxaddr); void put_idmap_pgtbl(vm_object_t obj); void domain_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size); int domain_alloc_pgtbl(struct dmar_domain *domain); void domain_free_pgtbl(struct dmar_domain *domain); extern const struct iommu_domain_map_ops dmar_domain_map_ops; int dmar_dev_depth(device_t child); void dmar_dev_path(device_t child, int *busno, void *path1, int depth); struct dmar_ctx *dmar_get_ctx_for_dev(struct dmar_unit *dmar, device_t dev, uint16_t rid, bool id_mapped, bool 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); int dmar_move_ctx_to_domain(struct dmar_domain *domain, struct dmar_ctx *ctx); void dmar_free_ctx_locked(struct dmar_unit *dmar, struct dmar_ctx *ctx); void dmar_free_ctx(struct dmar_ctx *ctx); struct dmar_ctx *dmar_find_ctx_locked(struct dmar_unit *dmar, uint16_t rid); void dmar_domain_free_entry(struct iommu_map_entry *entry, bool free); 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); int dmar_instantiate_rmrr_ctxs(struct iommu_unit *dmar); void dmar_quirks_post_ident(struct dmar_unit *dmar); void dmar_quirks_pre_use(struct iommu_unit *dmar); int dmar_init_irt(struct dmar_unit *unit); void dmar_fini_irt(struct dmar_unit *unit); extern iommu_haddr_t dmar_high; extern int haw; extern int dmar_tbl_pagecnt; extern int dmar_batch_coalesce; static inline uint32_t dmar_read4(const struct dmar_unit *unit, int reg) { return (bus_read_4(unit->regs, reg)); } static inline uint64_t dmar_read8(const struct dmar_unit *unit, int reg) { #ifdef __i386__ uint32_t high, low; low = bus_read_4(unit->regs, reg); high = bus_read_4(unit->regs, reg + 4); return (low | ((uint64_t)high << 32)); #else return (bus_read_8(unit->regs, reg)); #endif } static inline void dmar_write4(const struct dmar_unit *unit, int reg, uint32_t val) { KASSERT(reg != DMAR_GCMD_REG || (val & DMAR_GCMD_TE) == (unit->hw_gcmd & DMAR_GCMD_TE), ("dmar%d clearing TE 0x%08x 0x%08x", unit->iommu.unit, unit->hw_gcmd, val)); bus_write_4(unit->regs, reg, val); } static inline void dmar_write8(const struct dmar_unit *unit, int reg, uint64_t val) { KASSERT(reg != DMAR_GCMD_REG, ("8byte GCMD write")); #ifdef __i386__ uint32_t high, low; low = val; high = val >> 32; bus_write_4(unit->regs, reg, low); bus_write_4(unit->regs, reg + 4, high); #else bus_write_8(unit->regs, reg, val); #endif } /* * dmar_pte_store and dmar_pte_clear ensure that on i386, 32bit writes * are issued in the correct order. For store, the lower word, * containing the P or R and W bits, is set only after the high word * is written. For clear, the P bit is cleared first, then the high * word is cleared. * * dmar_pte_update updates the pte. For amd64, the update is atomic. * For i386, it first disables the entry by clearing the word * containing the P bit, and then defer to dmar_pte_store. The locked * cmpxchg8b is probably available on any machine having DMAR support, * but interrupt translation table may be mapped uncached. */ static inline void dmar_pte_store1(volatile uint64_t *dst, uint64_t val) { #ifdef __i386__ volatile uint32_t *p; uint32_t hi, lo; hi = val >> 32; lo = val; p = (volatile uint32_t *)dst; *(p + 1) = hi; *p = lo; #else *dst = val; #endif } static inline void dmar_pte_store(volatile uint64_t *dst, uint64_t val) { KASSERT(*dst == 0, ("used pte %p oldval %jx newval %jx", dst, (uintmax_t)*dst, (uintmax_t)val)); dmar_pte_store1(dst, val); } static inline void dmar_pte_update(volatile uint64_t *dst, uint64_t val) { #ifdef __i386__ volatile uint32_t *p; p = (volatile uint32_t *)dst; *p = 0; #endif dmar_pte_store1(dst, val); } static inline void dmar_pte_clear(volatile uint64_t *dst) { #ifdef __i386__ volatile uint32_t *p; p = (volatile uint32_t *)dst; *p = 0; *(p + 1) = 0; #else *dst = 0; #endif } extern struct timespec dmar_hw_timeout; #define DMAR_WAIT_UNTIL(cond) \ { \ struct timespec last, curr; \ bool forever; \ \ if (dmar_hw_timeout.tv_sec == 0 && \ dmar_hw_timeout.tv_nsec == 0) { \ forever = true; \ } else { \ forever = false; \ nanouptime(&curr); \ timespecadd(&curr, &dmar_hw_timeout, &last); \ } \ for (;;) { \ if (cond) { \ error = 0; \ break; \ } \ nanouptime(&curr); \ if (!forever && timespeccmp(&last, &curr, <)) { \ error = ETIMEDOUT; \ break; \ } \ cpu_spinwait(); \ } \ } #ifdef INVARIANTS #define TD_PREP_PINNED_ASSERT \ int old_td_pinned; \ old_td_pinned = curthread->td_pinned #define TD_PINNED_ASSERT \ KASSERT(curthread->td_pinned == old_td_pinned, \ ("pin count leak: %d %d %s:%d", curthread->td_pinned, \ old_td_pinned, __FILE__, __LINE__)) #else #define TD_PREP_PINNED_ASSERT #define TD_PINNED_ASSERT #endif #endif diff --git a/sys/x86/iommu/intel_qi.c b/sys/x86/iommu/intel_qi.c index 32f01a2787b0..3a0012763000 100644 --- a/sys/x86/iommu/intel_qi.c +++ b/sys/x86/iommu/intel_qi.c @@ -1,495 +1,574 @@ /*- * 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 "opt_acpi.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 static bool dmar_qi_seq_processed(const struct dmar_unit *unit, const struct iommu_qi_genseq *pseq) { + u_int gen; - return (pseq->gen < unit->inv_waitd_gen || - (pseq->gen == unit->inv_waitd_gen && - pseq->seq <= unit->inv_waitd_seq_hw)); + gen = unit->inv_waitd_gen; + return (pseq->gen < gen || + (pseq->gen == gen && pseq->seq <= unit->inv_waitd_seq_hw)); } static int dmar_enable_qi(struct dmar_unit *unit) { int error; DMAR_ASSERT_LOCKED(unit); unit->hw_gcmd |= DMAR_GCMD_QIE; dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd); DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_QIES) != 0)); return (error); } static int dmar_disable_qi(struct dmar_unit *unit) { int error; DMAR_ASSERT_LOCKED(unit); unit->hw_gcmd &= ~DMAR_GCMD_QIE; dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd); DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_QIES) == 0)); return (error); } static void dmar_qi_advance_tail(struct dmar_unit *unit) { DMAR_ASSERT_LOCKED(unit); dmar_write4(unit, DMAR_IQT_REG, unit->inv_queue_tail); } static void dmar_qi_ensure(struct dmar_unit *unit, int descr_count) { uint32_t head; int bytes; DMAR_ASSERT_LOCKED(unit); bytes = descr_count << DMAR_IQ_DESCR_SZ_SHIFT; for (;;) { if (bytes <= unit->inv_queue_avail) break; /* refill */ head = dmar_read4(unit, DMAR_IQH_REG); head &= DMAR_IQH_MASK; unit->inv_queue_avail = head - unit->inv_queue_tail - DMAR_IQ_DESCR_SZ; if (head <= unit->inv_queue_tail) unit->inv_queue_avail += unit->inv_queue_size; if (bytes <= unit->inv_queue_avail) break; /* * No space in the queue, do busy wait. Hardware must * make a progress. But first advance the tail to * inform the descriptor streamer about entries we * might have already filled, otherwise they could * clog the whole queue.. + * + * See dmar_qi_invalidate_locked() for a discussion + * about data race prevention. */ dmar_qi_advance_tail(unit); unit->inv_queue_full++; cpu_spinwait(); } unit->inv_queue_avail -= bytes; } static void dmar_qi_emit(struct dmar_unit *unit, uint64_t data1, uint64_t data2) { DMAR_ASSERT_LOCKED(unit); *(volatile uint64_t *)(unit->inv_queue + unit->inv_queue_tail) = data1; unit->inv_queue_tail += DMAR_IQ_DESCR_SZ / 2; KASSERT(unit->inv_queue_tail <= unit->inv_queue_size, ("tail overflow 0x%x 0x%jx", unit->inv_queue_tail, (uintmax_t)unit->inv_queue_size)); unit->inv_queue_tail &= unit->inv_queue_size - 1; *(volatile uint64_t *)(unit->inv_queue + unit->inv_queue_tail) = data2; unit->inv_queue_tail += DMAR_IQ_DESCR_SZ / 2; KASSERT(unit->inv_queue_tail <= unit->inv_queue_size, ("tail overflow 0x%x 0x%jx", unit->inv_queue_tail, (uintmax_t)unit->inv_queue_size)); unit->inv_queue_tail &= unit->inv_queue_size - 1; } static void dmar_qi_emit_wait_descr(struct dmar_unit *unit, uint32_t seq, bool intr, bool memw, bool fence) { DMAR_ASSERT_LOCKED(unit); dmar_qi_emit(unit, DMAR_IQ_DESCR_WAIT_ID | (intr ? DMAR_IQ_DESCR_WAIT_IF : 0) | (memw ? DMAR_IQ_DESCR_WAIT_SW : 0) | (fence ? DMAR_IQ_DESCR_WAIT_FN : 0) | (memw ? DMAR_IQ_DESCR_WAIT_SD(seq) : 0), memw ? unit->inv_waitd_seq_hw_phys : 0); } static void dmar_qi_emit_wait_seq(struct dmar_unit *unit, struct iommu_qi_genseq *pseq, bool emit_wait) { struct iommu_qi_genseq gsec; uint32_t seq; KASSERT(pseq != NULL, ("wait descriptor with no place for seq")); DMAR_ASSERT_LOCKED(unit); if (unit->inv_waitd_seq == 0xffffffff) { gsec.gen = unit->inv_waitd_gen; gsec.seq = unit->inv_waitd_seq; dmar_qi_ensure(unit, 1); dmar_qi_emit_wait_descr(unit, gsec.seq, false, true, false); dmar_qi_advance_tail(unit); while (!dmar_qi_seq_processed(unit, &gsec)) cpu_spinwait(); unit->inv_waitd_gen++; unit->inv_waitd_seq = 1; } seq = unit->inv_waitd_seq++; pseq->gen = unit->inv_waitd_gen; pseq->seq = seq; if (emit_wait) { dmar_qi_ensure(unit, 1); dmar_qi_emit_wait_descr(unit, seq, true, true, false); } } +/* + * To avoid missed wakeups, callers must increment the unit's waiters count + * before advancing the tail past the wait descriptor. + */ static void dmar_qi_wait_for_seq(struct dmar_unit *unit, const struct iommu_qi_genseq *gseq, bool nowait) { DMAR_ASSERT_LOCKED(unit); - unit->inv_seq_waiters++; + KASSERT(unit->inv_seq_waiters > 0, ("%s: no waiters", __func__)); while (!dmar_qi_seq_processed(unit, gseq)) { if (cold || nowait) { cpu_spinwait(); } else { msleep(&unit->inv_seq_waiters, &unit->iommu.lock, 0, "dmarse", hz); } } unit->inv_seq_waiters--; } -void -dmar_qi_invalidate_locked(struct dmar_domain *domain, iommu_gaddr_t base, +static void +dmar_qi_invalidate_emit(struct dmar_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size, struct iommu_qi_genseq *pseq, bool emit_wait) { struct dmar_unit *unit; iommu_gaddr_t isize; int am; unit = domain->dmar; DMAR_ASSERT_LOCKED(unit); for (; size > 0; base += isize, size -= isize) { am = calc_am(unit, base, size, &isize); dmar_qi_ensure(unit, 1); dmar_qi_emit(unit, DMAR_IQ_DESCR_IOTLB_INV | DMAR_IQ_DESCR_IOTLB_PAGE | DMAR_IQ_DESCR_IOTLB_DW | DMAR_IQ_DESCR_IOTLB_DR | DMAR_IQ_DESCR_IOTLB_DID(domain->domain), base | am); } dmar_qi_emit_wait_seq(unit, pseq, emit_wait); +} + +/* + * The caller must not be using the entry's dmamap_link field. + */ +void +dmar_qi_invalidate_locked(struct dmar_domain *domain, + struct iommu_map_entry *entry, bool emit_wait) +{ + struct dmar_unit *unit; + + unit = domain->dmar; + DMAR_ASSERT_LOCKED(unit); + dmar_qi_invalidate_emit(domain, entry->start, entry->end - + entry->start, &entry->gseq, emit_wait); + + /* + * To avoid a data race in dmar_qi_task(), the entry's gseq must be + * initialized before the entry is added to the TLB flush list, and the + * entry must be added to that list before the tail is advanced. More + * precisely, the tail must not be advanced past the wait descriptor + * that will generate the interrupt that schedules dmar_qi_task() for + * execution before the entry is added to the list. While an earlier + * call to dmar_qi_ensure() might have advanced the tail, it will not + * advance it past the wait descriptor. + * + * See the definition of struct dmar_unit for more information on + * synchronization. + */ + entry->tlb_flush_next = NULL; + atomic_store_rel_ptr((uintptr_t *)&unit->tlb_flush_tail->tlb_flush_next, + (uintptr_t)entry); + unit->tlb_flush_tail = entry; + dmar_qi_advance_tail(unit); } void dmar_qi_invalidate_sync(struct dmar_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size, bool cansleep) { struct dmar_unit *unit; struct iommu_qi_genseq gseq; unit = domain->dmar; DMAR_LOCK(unit); - dmar_qi_invalidate_locked(domain, base, size, &gseq, true); + dmar_qi_invalidate_emit(domain, base, size, &gseq, true); + + /* + * To avoid a missed wakeup in dmar_qi_task(), the unit's waiters count + * must be incremented before the tail is advanced. + */ + unit->inv_seq_waiters++; + + dmar_qi_advance_tail(unit); dmar_qi_wait_for_seq(unit, &gseq, !cansleep); DMAR_UNLOCK(unit); } void dmar_qi_invalidate_ctx_glob_locked(struct dmar_unit *unit) { struct iommu_qi_genseq gseq; DMAR_ASSERT_LOCKED(unit); dmar_qi_ensure(unit, 2); dmar_qi_emit(unit, DMAR_IQ_DESCR_CTX_INV | DMAR_IQ_DESCR_CTX_GLOB, 0); dmar_qi_emit_wait_seq(unit, &gseq, true); + /* See dmar_qi_invalidate_sync(). */ + unit->inv_seq_waiters++; dmar_qi_advance_tail(unit); dmar_qi_wait_for_seq(unit, &gseq, false); } void dmar_qi_invalidate_iotlb_glob_locked(struct dmar_unit *unit) { struct iommu_qi_genseq gseq; DMAR_ASSERT_LOCKED(unit); dmar_qi_ensure(unit, 2); dmar_qi_emit(unit, DMAR_IQ_DESCR_IOTLB_INV | DMAR_IQ_DESCR_IOTLB_GLOB | DMAR_IQ_DESCR_IOTLB_DW | DMAR_IQ_DESCR_IOTLB_DR, 0); dmar_qi_emit_wait_seq(unit, &gseq, true); + /* See dmar_qi_invalidate_sync(). */ + unit->inv_seq_waiters++; dmar_qi_advance_tail(unit); dmar_qi_wait_for_seq(unit, &gseq, false); } void dmar_qi_invalidate_iec_glob(struct dmar_unit *unit) { struct iommu_qi_genseq gseq; DMAR_ASSERT_LOCKED(unit); dmar_qi_ensure(unit, 2); dmar_qi_emit(unit, DMAR_IQ_DESCR_IEC_INV, 0); dmar_qi_emit_wait_seq(unit, &gseq, true); + /* See dmar_qi_invalidate_sync(). */ + unit->inv_seq_waiters++; dmar_qi_advance_tail(unit); dmar_qi_wait_for_seq(unit, &gseq, false); } void dmar_qi_invalidate_iec(struct dmar_unit *unit, u_int start, u_int cnt) { struct iommu_qi_genseq gseq; u_int c, l; DMAR_ASSERT_LOCKED(unit); KASSERT(start < unit->irte_cnt && start < start + cnt && start + cnt <= unit->irte_cnt, ("inv iec overflow %d %d %d", unit->irte_cnt, start, cnt)); for (; cnt > 0; cnt -= c, start += c) { l = ffs(start | cnt) - 1; c = 1 << l; dmar_qi_ensure(unit, 1); dmar_qi_emit(unit, DMAR_IQ_DESCR_IEC_INV | DMAR_IQ_DESCR_IEC_IDX | DMAR_IQ_DESCR_IEC_IIDX(start) | DMAR_IQ_DESCR_IEC_IM(l), 0); } dmar_qi_ensure(unit, 1); dmar_qi_emit_wait_seq(unit, &gseq, true); + + /* + * Since dmar_qi_wait_for_seq() will not sleep, this increment's + * placement relative to advancing the tail doesn't matter. + */ + unit->inv_seq_waiters++; + dmar_qi_advance_tail(unit); /* * The caller of the function, in particular, * dmar_ir_program_irte(), may be called from the context * where the sleeping is forbidden (in fact, the * intr_table_lock mutex may be held, locked from * intr_shuffle_irqs()). Wait for the invalidation completion * using the busy wait. * * The impact on the interrupt input setup code is small, the * expected overhead is comparable with the chipset register * read. It is more harmful for the parallel DMA operations, * since we own the dmar unit lock until whole invalidation * queue is processed, which includes requests possibly issued * before our request. */ dmar_qi_wait_for_seq(unit, &gseq, true); } int dmar_qi_intr(void *arg) { struct dmar_unit *unit; unit = arg; KASSERT(unit->qi_enabled, ("dmar%d: QI is not enabled", unit->iommu.unit)); taskqueue_enqueue(unit->qi_taskqueue, &unit->qi_task); return (FILTER_HANDLED); } static void dmar_qi_task(void *arg, int pending __unused) { struct dmar_unit *unit; - struct iommu_map_entry *entry; + struct iommu_domain *domain; + struct iommu_map_entry *entry, *head; uint32_t ics; unit = arg; /* * Request an interrupt on the completion of the next invalidation * wait descriptor with the IF field set. */ ics = dmar_read4(unit, DMAR_ICS_REG); if ((ics & DMAR_ICS_IWC) != 0) { ics = DMAR_ICS_IWC; dmar_write4(unit, DMAR_ICS_REG, ics); } - DMAR_LOCK(unit); for (;;) { - entry = TAILQ_FIRST(&unit->tlb_flush_entries); + head = unit->tlb_flush_head; + entry = (struct iommu_map_entry *) + atomic_load_acq_ptr((uintptr_t *)&head->tlb_flush_next); if (entry == NULL) break; if (!dmar_qi_seq_processed(unit, &entry->gseq)) break; - TAILQ_REMOVE(&unit->tlb_flush_entries, entry, dmamap_link); - DMAR_UNLOCK(unit); - dmar_domain_free_entry(entry, true); - DMAR_LOCK(unit); + unit->tlb_flush_head = entry; + iommu_gas_free_entry(head->domain, head); + 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 (unit->inv_seq_waiters > 0) + if (unit->inv_seq_waiters > 0) { + /* + * Acquire the DMAR lock so that wakeup() is called only after + * the waiter is sleeping. + */ + DMAR_LOCK(unit); wakeup(&unit->inv_seq_waiters); - DMAR_UNLOCK(unit); + DMAR_UNLOCK(unit); + } } int dmar_init_qi(struct dmar_unit *unit) { uint64_t iqa; uint32_t ics; int qi_sz; if (!DMAR_HAS_QI(unit) || (unit->hw_cap & DMAR_CAP_CM) != 0) return (0); unit->qi_enabled = 1; TUNABLE_INT_FETCH("hw.dmar.qi", &unit->qi_enabled); if (!unit->qi_enabled) return (0); - TAILQ_INIT(&unit->tlb_flush_entries); + unit->tlb_flush_head = unit->tlb_flush_tail = + iommu_gas_alloc_entry(NULL, 0); TASK_INIT(&unit->qi_task, 0, dmar_qi_task, unit); unit->qi_taskqueue = taskqueue_create_fast("dmarqf", M_WAITOK, taskqueue_thread_enqueue, &unit->qi_taskqueue); taskqueue_start_threads(&unit->qi_taskqueue, 1, PI_AV, "dmar%d qi taskq", unit->iommu.unit); unit->inv_waitd_gen = 0; unit->inv_waitd_seq = 1; qi_sz = DMAR_IQA_QS_DEF; TUNABLE_INT_FETCH("hw.dmar.qi_size", &qi_sz); if (qi_sz > DMAR_IQA_QS_MAX) qi_sz = DMAR_IQA_QS_MAX; unit->inv_queue_size = (1ULL << qi_sz) * PAGE_SIZE; /* Reserve one descriptor to prevent wraparound. */ unit->inv_queue_avail = unit->inv_queue_size - DMAR_IQ_DESCR_SZ; /* The invalidation queue reads by DMARs are always coherent. */ unit->inv_queue = kmem_alloc_contig(unit->inv_queue_size, M_WAITOK | M_ZERO, 0, dmar_high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); unit->inv_waitd_seq_hw_phys = pmap_kextract( (vm_offset_t)&unit->inv_waitd_seq_hw); DMAR_LOCK(unit); dmar_write8(unit, DMAR_IQT_REG, 0); iqa = pmap_kextract(unit->inv_queue); iqa |= qi_sz; dmar_write8(unit, DMAR_IQA_REG, iqa); dmar_enable_qi(unit); ics = dmar_read4(unit, DMAR_ICS_REG); if ((ics & DMAR_ICS_IWC) != 0) { ics = DMAR_ICS_IWC; dmar_write4(unit, DMAR_ICS_REG, ics); } dmar_enable_qi_intr(unit); DMAR_UNLOCK(unit); return (0); } void dmar_fini_qi(struct dmar_unit *unit) { struct iommu_qi_genseq gseq; if (!unit->qi_enabled) return; taskqueue_drain(unit->qi_taskqueue, &unit->qi_task); taskqueue_free(unit->qi_taskqueue); unit->qi_taskqueue = NULL; DMAR_LOCK(unit); /* quisce */ dmar_qi_ensure(unit, 1); dmar_qi_emit_wait_seq(unit, &gseq, true); + /* See dmar_qi_invalidate_sync_locked(). */ + unit->inv_seq_waiters++; dmar_qi_advance_tail(unit); dmar_qi_wait_for_seq(unit, &gseq, false); /* only after the quisce, disable queue */ dmar_disable_qi_intr(unit); dmar_disable_qi(unit); KASSERT(unit->inv_seq_waiters == 0, ("dmar%d: waiters on disabled queue", unit->iommu.unit)); DMAR_UNLOCK(unit); kmem_free(unit->inv_queue, unit->inv_queue_size); unit->inv_queue = 0; unit->inv_queue_size = 0; unit->qi_enabled = 0; } void dmar_enable_qi_intr(struct dmar_unit *unit) { uint32_t iectl; DMAR_ASSERT_LOCKED(unit); KASSERT(DMAR_HAS_QI(unit), ("dmar%d: QI is not supported", unit->iommu.unit)); iectl = dmar_read4(unit, DMAR_IECTL_REG); iectl &= ~DMAR_IECTL_IM; dmar_write4(unit, DMAR_IECTL_REG, iectl); } void dmar_disable_qi_intr(struct dmar_unit *unit) { uint32_t iectl; DMAR_ASSERT_LOCKED(unit); KASSERT(DMAR_HAS_QI(unit), ("dmar%d: QI is not supported", unit->iommu.unit)); iectl = dmar_read4(unit, DMAR_IECTL_REG); dmar_write4(unit, DMAR_IECTL_REG, iectl | DMAR_IECTL_IM); }