diff --git a/sys/x86/iommu/intel_dmar.h b/sys/x86/iommu/intel_dmar.h index c3163abf6f92..1a9b5041975c 100644 --- a/sys/x86/iommu/intel_dmar.h +++ b/sys/x86/iommu/intel_dmar.h @@ -1,426 +1,427 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * 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. */ #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 */ 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 */ }; #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) #define DMAR_INTR_FAULT 0 #define DMAR_INTR_QI 1 #define DMAR_INTR_TOTAL 2 struct dmar_unit { struct iommu_unit iommu; struct x86_unit_common x86c; uint16_t segment; uint64_t base; + int memdomain; /* Resources */ int reg_rid; struct resource *regs; /* 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; /* IR */ int ir_enabled; vm_paddr_t irt_phys; dmar_irte_t *irt; u_int irte_cnt; vmem_t *irtids; }; #define DMAR_LOCK(dmar) mtx_lock(&DMAR2IOMMU(dmar)->lock) #define DMAR_UNLOCK(dmar) mtx_unlock(&DMAR2IOMMU(dmar)->lock) #define DMAR_ASSERT_LOCKED(dmar) mtx_assert(&DMAR2IOMMU(dmar)->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 SYSCTL_DECL(_hw_iommu_dmar); 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); int domain_is_sp_lvl(struct dmar_domain *domain, 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); 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, iommu_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_disable_protected_regions(struct dmar_unit *unit); 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 iommu_unit *unit); void dmar_disable_fault_intr(struct iommu_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 iommu_unit *unit); void dmar_disable_qi_intr(struct iommu_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, 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 dmar_get_idmap_pgtbl(struct dmar_domain *domain, iommu_gaddr_t maxaddr); void dmar_put_idmap_pgtbl(vm_object_t obj); void dmar_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size); int dmar_domain_alloc_pgtbl(struct dmar_domain *domain); void dmar_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_method(struct iommu_unit *dmar, struct iommu_ctx *ctx); void dmar_free_ctx_method(struct iommu_ctx *ctx); struct dmar_ctx *dmar_find_ctx_locked(struct dmar_unit *dmar, uint16_t rid); struct iommu_ctx *dmar_get_ctx(struct iommu_unit *iommu, device_t dev, uint16_t rid, bool id_mapped, bool rmrr_init); void dmar_domain_unload_entry(struct iommu_map_entry *entry, bool free, bool cansleep); void dmar_domain_unload(struct iommu_domain *iodom, struct iommu_map_entries_tailq *entries, bool cansleep); 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); int dmar_alloc_msi_intr(device_t src, u_int *cookies, u_int count); int dmar_map_msi_intr(device_t src, u_int cpu, u_int vector, u_int cookie, uint64_t *addr, uint32_t *data); int dmar_unmap_msi_intr(device_t src, u_int cookie); int dmar_map_ioapic_intr(u_int ioapic_id, u_int cpu, u_int vector, bool edge, bool activehi, int irq, u_int *cookie, uint32_t *hi, uint32_t *lo); int dmar_unmap_ioapic_intr(u_int ioapic_id, u_int *cookie); extern int haw; extern int dmar_rmrr_enable; 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_drv.c b/sys/x86/iommu/intel_drv.c index f4a1ec06b721..22d04029f9ae 100644 --- a/sys/x86/iommu/intel_drv.c +++ b/sys/x86/iommu/intel_drv.c @@ -1,1303 +1,1336 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * 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 "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 #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 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); } int dmar_rmrr_enable = 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; TUNABLE_INT_FETCH("hw.dmar.rmrr_enable", &dmar_rmrr_enable); 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) iommu_high = BUS_SPACE_MAXADDR; else iommu_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_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++) iommu_release_intr(DMAR2IOMMU(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; } sysctl_ctx_free(&unit->iommu.sysctl_ctx); } #ifdef DEV_APIC static int dmar_remap_intr(device_t dev, device_t child, u_int irq) { struct dmar_unit *unit; struct iommu_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->x86c.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->msi_data = msi_data; dmd->msi_addr = msi_addr; (dmd->disable_intr)(DMAR2IOMMU(unit)); dmar_write4(unit, dmd->msi_data_reg, dmd->msi_data); dmar_write4(unit, dmd->msi_addr_reg, dmd->msi_addr); dmar_write4(unit, dmd->msi_uaddr_reg, dmd->msi_addr >> 32); (dmd->enable_intr)(DMAR2IOMMU(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)); } +/* Remapping Hardware Static Affinity Structure lookup */ +struct rhsa_iter_arg { + uint64_t base; + u_int proxim_dom; +}; + +static int +dmar_rhsa_iter(ACPI_DMAR_HEADER *dmarh, void *arg) +{ + struct rhsa_iter_arg *ria; + ACPI_DMAR_RHSA *adr; + + if (dmarh->Type == ACPI_DMAR_TYPE_HARDWARE_AFFINITY) { + ria = arg; + adr = (ACPI_DMAR_RHSA *)dmarh; + if (adr->BaseAddress == ria->base) + ria->proxim_dom = adr->ProximityDomain; + } + return (1); +} + static int dmar_attach(device_t dev) { struct dmar_unit *unit; ACPI_DMAR_HARDWARE_UNIT *dmaru; struct iommu_msi_data *dmd; + struct rhsa_iter_arg ria; uint64_t timeout; int disable_pmr; int i, error; unit = device_get_softc(dev); unit->iommu.unit = device_get_unit(dev); unit->iommu.dev = dev; sysctl_ctx_init(&unit->iommu.sysctl_ctx); 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"); dmar_devs[unit->iommu.unit] = NULL; 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); + unit->memdomain = -1; + ria.base = unit->base; + ria.proxim_dom = -1; + dmar_iterate_tbl(dmar_rhsa_iter, &ria); + if (ria.proxim_dom != -1) + unit->memdomain = acpi_map_pxm_to_vm_domainid(ria.proxim_dom); timeout = dmar_get_timeout(); TUNABLE_UINT64_FETCH("hw.iommu.dmar.timeout", &timeout); dmar_update_timeout(timeout); for (i = 0; i < DMAR_INTR_TOTAL; i++) unit->x86c.intrs[i].irq = -1; dmd = &unit->x86c.intrs[DMAR_INTR_FAULT]; dmd->name = "fault"; dmd->irq_rid = DMAR_FAULT_IRQ_RID; dmd->handler = dmar_fault_intr; dmd->msi_data_reg = DMAR_FEDATA_REG; dmd->msi_addr_reg = DMAR_FEADDR_REG; dmd->msi_uaddr_reg = DMAR_FEUADDR_REG; dmd->enable_intr = dmar_enable_fault_intr; dmd->disable_intr = dmar_disable_fault_intr; error = iommu_alloc_irq(DMAR2IOMMU(unit), DMAR_INTR_FAULT); if (error != 0) { dmar_release_resources(dev, unit); dmar_devs[unit->iommu.unit] = NULL; return (error); } dmar_write4(unit, dmd->msi_data_reg, dmd->msi_data); dmar_write4(unit, dmd->msi_addr_reg, dmd->msi_addr); dmar_write4(unit, dmd->msi_uaddr_reg, dmd->msi_addr >> 32); if (DMAR_HAS_QI(unit)) { dmd = &unit->x86c.intrs[DMAR_INTR_QI]; dmd->name = "qi"; dmd->irq_rid = DMAR_QI_IRQ_RID; dmd->handler = dmar_qi_intr; dmd->msi_data_reg = DMAR_IEDATA_REG; dmd->msi_addr_reg = DMAR_IEADDR_REG; dmd->msi_uaddr_reg = DMAR_IEUADDR_REG; dmd->enable_intr = dmar_enable_qi_intr; dmd->disable_intr = dmar_disable_qi_intr; error = iommu_alloc_irq(DMAR2IOMMU(unit), DMAR_INTR_QI); if (error != 0) { dmar_release_resources(dev, unit); dmar_devs[unit->iommu.unit] = NULL; return (error); } dmar_write4(unit, dmd->msi_data_reg, dmd->msi_data); dmar_write4(unit, dmd->msi_addr_reg, dmd->msi_addr); dmar_write4(unit, dmd->msi_uaddr_reg, dmd->msi_addr >> 32); } 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); + if (unit->memdomain != -1) { + unit->ctx_obj->domain.dr_policy = DOMAINSET_PREF( + unit->memdomain); + } /* * Allocate and load the root entry table pointer. Enable the * address translation after the required invalidations are * done. */ iommu_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); dmar_devs[unit->iommu.unit] = NULL; return (error); } error = dmar_inv_ctx_glob(unit); if (error != 0) { DMAR_UNLOCK(unit); dmar_release_resources(dev, unit); dmar_devs[unit->iommu.unit] = NULL; 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); dmar_devs[unit->iommu.unit] = NULL; return (error); } } DMAR_UNLOCK(unit); error = dmar_init_fault_log(unit); if (error != 0) { dmar_release_resources(dev, unit); dmar_devs[unit->iommu.unit] = NULL; return (error); } error = dmar_init_qi(unit); if (error != 0) { dmar_release_resources(dev, unit); dmar_devs[unit->iommu.unit] = NULL; return (error); } error = dmar_init_irt(unit); if (error != 0) { dmar_release_resources(dev, unit); dmar_devs[unit->iommu.unit] = NULL; return (error); } disable_pmr = 0; TUNABLE_INT_FETCH("hw.dmar.pmr.disable", &disable_pmr); if (disable_pmr) { error = dmar_disable_protected_regions(unit); if (error != 0) device_printf(dev, "Failed to disable protected regions\n"); } error = iommu_init_busdma(&unit->iommu); if (error != 0) { dmar_release_resources(dev, unit); dmar_devs[unit->iommu.unit] = NULL; return (error); } #ifdef NOTYET DMAR_LOCK(unit); error = dmar_enable_translation(unit); if (error != 0) { DMAR_UNLOCK(unit); dmar_release_resources(dev, unit); dmar_devs[unit->iommu.unit] = NULL; 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, 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"); } iommu_device_set_iommu_prop(dev, unit->iommu.dev); 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) { struct dmar_unit *unit; unit = dmar_find_nonpci(hpet_get_uid(dev), ACPI_DMAR_SCOPE_TYPE_HPET, rid); if (unit != NULL) iommu_device_set_iommu_prop(dev, unit->iommu.dev); return (unit); } struct dmar_unit * dmar_find_ioapic(u_int apic_id, uint16_t *rid) { struct dmar_unit *unit; device_t apic_dev; unit = dmar_find_nonpci(apic_id, ACPI_DMAR_SCOPE_TYPE_IOAPIC, rid); if (unit != NULL) { apic_dev = ioapic_get_dev(apic_id); if (apic_dev != NULL) iommu_device_set_iommu_prop(apic_dev, unit->iommu.dev); } return (unit); } 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 (!dmar_rmrr_enable) return (1); 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, 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(segment, 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 (!dmar_rmrr_enable) return (1); 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_disable_protected_regions(dmar); if (error != 0) printf("dmar%d: Failed to disable protected regions\n", 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(struct dmar_domain *domain, bool show_mappings) { struct iommu_domain *iodom; 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); iommu_db_domain_print_contexts(iodom); if (show_mappings) iommu_db_domain_print_mappings(iodom); } DB_SHOW_COMMAND_FLAGS(dmar_domain, db_dmar_print_domain, CS_OWN) { struct dmar_unit *unit; struct dmar_domain *domain; struct iommu_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->iodom.contexts, link) { if (pci_domain == unit->segment && bus == pci_get_bus(ctx->tag->owner) && device == pci_get_slot(ctx->tag->owner) && function == pci_get_function(ctx->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%jx@%p/phys@%jx next seq 0x%x gen 0x%x\n", (uintmax_t)unit->x86c.inv_queue, (uintmax_t)dmar_read8(unit, DMAR_IQA_REG), (uintmax_t)unit->x86c.inv_queue_size, dmar_read4(unit, DMAR_IQH_REG), dmar_read4(unit, DMAR_IQT_REG), unit->x86c.inv_queue_avail, dmar_read4(unit, DMAR_ICS_REG), dmar_read4(unit, DMAR_IECTL_REG), (uintmax_t)unit->x86c.inv_waitd_seq_hw, &unit->x86c.inv_waitd_seq_hw, (uintmax_t)unit->x86c.inv_waitd_seq_hw_phys, unit->x86c.inv_waitd_seq, unit->x86c.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 static struct iommu_unit * dmar_find_method(device_t dev, bool verbose) { struct dmar_unit *dmar; dmar = dmar_find(dev, verbose); return (&dmar->iommu); } static struct x86_unit_common * dmar_get_x86_common(struct iommu_unit *unit) { struct dmar_unit *dmar; dmar = IOMMU2DMAR(unit); return (&dmar->x86c); } static void dmar_unit_pre_instantiate_ctx(struct iommu_unit *unit) { dmar_quirks_pre_use(unit); dmar_instantiate_rmrr_ctxs(unit); } static struct x86_iommu dmar_x86_iommu = { .get_x86_common = dmar_get_x86_common, .unit_pre_instantiate_ctx = dmar_unit_pre_instantiate_ctx, .domain_unload_entry = dmar_domain_unload_entry, .domain_unload = dmar_domain_unload, .get_ctx = dmar_get_ctx, .free_ctx_locked = dmar_free_ctx_locked_method, .free_ctx = dmar_free_ctx_method, .find = dmar_find_method, .alloc_msi_intr = dmar_alloc_msi_intr, .map_msi_intr = dmar_map_msi_intr, .unmap_msi_intr = dmar_unmap_msi_intr, .map_ioapic_intr = dmar_map_ioapic_intr, .unmap_ioapic_intr = dmar_unmap_ioapic_intr, }; static void x86_iommu_set_intel(void *arg __unused) { if (cpu_vendor_id == CPU_VENDOR_INTEL) set_x86_iommu(&dmar_x86_iommu); } SYSINIT(x86_iommu, SI_SUB_TUNABLES, SI_ORDER_ANY, x86_iommu_set_intel, NULL); diff --git a/sys/x86/iommu/intel_idpgtbl.c b/sys/x86/iommu/intel_idpgtbl.c index 6f66106822fe..b133dc875515 100644 --- a/sys/x86/iommu/intel_idpgtbl.c +++ b/sys/x86/iommu/intel_idpgtbl.c @@ -1,793 +1,804 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * 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 #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 #include static int dmar_unmap_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size, int flags, struct iommu_map_entry *entry); /* * The cache of the identity mapping page tables for the DMARs. Using * the cache saves significant amount of memory for page tables by * reusing the page tables, since usually DMARs are identical and have * the same capabilities. Still, cache records the information needed * to match DMAR capabilities and page table format, to correctly * handle different DMARs. */ struct idpgtbl { iommu_gaddr_t maxaddr; /* Page table covers the guest address range [0..maxaddr) */ int pglvl; /* Total page table levels ignoring superpages */ int leaf; /* The last materialized page table level, it is non-zero if superpages are supported */ vm_object_t pgtbl_obj; /* The page table pages */ LIST_ENTRY(idpgtbl) link; }; static struct sx idpgtbl_lock; SX_SYSINIT(idpgtbl, &idpgtbl_lock, "idpgtbl"); static LIST_HEAD(, idpgtbl) idpgtbls = LIST_HEAD_INITIALIZER(idpgtbls); static MALLOC_DEFINE(M_DMAR_IDPGTBL, "dmar_idpgtbl", "Intel DMAR Identity mappings cache elements"); /* * Build the next level of the page tables for the identity mapping. * - lvl is the level to build; * - idx is the index of the page table page in the pgtbl_obj, which is * being allocated filled now; * - addr is the starting address in the bus address space which is * mapped by the page table page. */ static void dmar_idmap_nextlvl(struct idpgtbl *tbl, int lvl, vm_pindex_t idx, iommu_gaddr_t addr) { vm_page_t m1; iommu_pte_t *pte; struct sf_buf *sf; iommu_gaddr_t f, pg_sz; vm_pindex_t base; int i; VM_OBJECT_ASSERT_LOCKED(tbl->pgtbl_obj); if (addr >= tbl->maxaddr) return; (void)iommu_pgalloc(tbl->pgtbl_obj, idx, IOMMU_PGF_OBJL | IOMMU_PGF_WAITOK | IOMMU_PGF_ZERO); base = idx * IOMMU_NPTEPG + 1; /* Index of the first child page of idx */ pg_sz = pglvl_page_size(tbl->pglvl, lvl); if (lvl != tbl->leaf) { for (i = 0, f = addr; i < IOMMU_NPTEPG; i++, f += pg_sz) dmar_idmap_nextlvl(tbl, lvl + 1, base + i, f); } VM_OBJECT_WUNLOCK(tbl->pgtbl_obj); pte = iommu_map_pgtbl(tbl->pgtbl_obj, idx, IOMMU_PGF_WAITOK, &sf); if (lvl == tbl->leaf) { for (i = 0, f = addr; i < IOMMU_NPTEPG; i++, f += pg_sz) { if (f >= tbl->maxaddr) break; pte[i].pte = (DMAR_PTE_ADDR_MASK & f) | DMAR_PTE_R | DMAR_PTE_W; } } else { for (i = 0, f = addr; i < IOMMU_NPTEPG; i++, f += pg_sz) { if (f >= tbl->maxaddr) break; m1 = iommu_pgalloc(tbl->pgtbl_obj, base + i, IOMMU_PGF_NOALLOC); KASSERT(m1 != NULL, ("lost page table page")); pte[i].pte = (DMAR_PTE_ADDR_MASK & VM_PAGE_TO_PHYS(m1)) | DMAR_PTE_R | DMAR_PTE_W; } } /* dmar_get_idmap_pgtbl flushes CPU cache if needed. */ iommu_unmap_pgtbl(sf); VM_OBJECT_WLOCK(tbl->pgtbl_obj); } /* * Find a ready and compatible identity-mapping page table in the * cache. If not found, populate the identity-mapping page table for * the context, up to the maxaddr. The maxaddr byte is allowed to be * not mapped, which is aligned with the definition of Maxmem as the * highest usable physical address + 1. If superpages are used, the * maxaddr is typically mapped. */ vm_object_t dmar_get_idmap_pgtbl(struct dmar_domain *domain, iommu_gaddr_t maxaddr) { struct dmar_unit *unit; struct idpgtbl *tbl; vm_object_t res; vm_page_t m; int leaf, i; leaf = 0; /* silence gcc */ /* * First, determine where to stop the paging structures. */ for (i = 0; i < domain->pglvl; i++) { if (i == domain->pglvl - 1 || domain_is_sp_lvl(domain, i)) { leaf = i; break; } } /* * Search the cache for a compatible page table. Qualified * page table must map up to maxaddr, its level must be * supported by the DMAR and leaf should be equal to the * calculated value. The later restriction could be lifted * but I believe it is currently impossible to have any * deviations for existing hardware. */ sx_slock(&idpgtbl_lock); LIST_FOREACH(tbl, &idpgtbls, link) { if (tbl->maxaddr >= maxaddr && dmar_pglvl_supported(domain->dmar, tbl->pglvl) && tbl->leaf == leaf) { res = tbl->pgtbl_obj; vm_object_reference(res); sx_sunlock(&idpgtbl_lock); domain->pglvl = tbl->pglvl; /* XXXKIB ? */ goto end; } } /* * Not found in cache, relock the cache into exclusive mode to * be able to add element, and recheck cache again after the * relock. */ sx_sunlock(&idpgtbl_lock); sx_xlock(&idpgtbl_lock); LIST_FOREACH(tbl, &idpgtbls, link) { if (tbl->maxaddr >= maxaddr && dmar_pglvl_supported(domain->dmar, tbl->pglvl) && tbl->leaf == leaf) { res = tbl->pgtbl_obj; vm_object_reference(res); sx_xunlock(&idpgtbl_lock); domain->pglvl = tbl->pglvl; /* XXXKIB ? */ return (res); } } /* * Still not found, create new page table. */ tbl = malloc(sizeof(*tbl), M_DMAR_IDPGTBL, M_WAITOK); tbl->pglvl = domain->pglvl; tbl->leaf = leaf; tbl->maxaddr = maxaddr; tbl->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL, IDX_TO_OFF(pglvl_max_pages(tbl->pglvl)), 0, 0, NULL); + /* + * Do not set NUMA policy, the identity table might be used + * by more than one unit. + */ VM_OBJECT_WLOCK(tbl->pgtbl_obj); dmar_idmap_nextlvl(tbl, 0, 0, 0); VM_OBJECT_WUNLOCK(tbl->pgtbl_obj); LIST_INSERT_HEAD(&idpgtbls, tbl, link); res = tbl->pgtbl_obj; vm_object_reference(res); sx_xunlock(&idpgtbl_lock); end: /* * Table was found or created. * * If DMAR does not snoop paging structures accesses, flush * CPU cache to memory. Note that dmar_unmap_pgtbl() coherent * argument was possibly invalid at the time of the identity * page table creation, since DMAR which was passed at the * time of creation could be coherent, while current DMAR is * not. * * If DMAR cannot look into the chipset write buffer, flush it * as well. */ unit = domain->dmar; if (!DMAR_IS_COHERENT(unit)) { VM_OBJECT_WLOCK(res); for (m = vm_page_lookup(res, 0); m != NULL; m = vm_page_next(m)) pmap_invalidate_cache_pages(&m, 1); VM_OBJECT_WUNLOCK(res); } if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) { DMAR_LOCK(unit); dmar_flush_write_bufs(unit); DMAR_UNLOCK(unit); } return (res); } /* * Return a reference to the identity mapping page table to the cache. */ void dmar_put_idmap_pgtbl(vm_object_t obj) { struct idpgtbl *tbl, *tbl1; vm_object_t rmobj; sx_slock(&idpgtbl_lock); KASSERT(obj->ref_count >= 2, ("lost cache reference")); vm_object_deallocate(obj); /* * Cache always owns one last reference on the page table object. * If there is an additional reference, object must stay. */ if (obj->ref_count > 1) { sx_sunlock(&idpgtbl_lock); return; } /* * Cache reference is the last, remove cache element and free * page table object, returning the page table pages to the * system. */ sx_sunlock(&idpgtbl_lock); sx_xlock(&idpgtbl_lock); LIST_FOREACH_SAFE(tbl, &idpgtbls, link, tbl1) { rmobj = tbl->pgtbl_obj; if (rmobj->ref_count == 1) { LIST_REMOVE(tbl, link); atomic_subtract_int(&iommu_tbl_pagecnt, rmobj->resident_page_count); vm_object_deallocate(rmobj); free(tbl, M_DMAR_IDPGTBL); } } sx_xunlock(&idpgtbl_lock); } /* * The core routines to map and unmap host pages at the given guest * address. Support superpages. */ static iommu_pte_t * dmar_pgtbl_map_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl, int flags, vm_pindex_t *idxp, struct sf_buf **sf) { vm_page_t m; struct sf_buf *sfp; iommu_pte_t *pte, *ptep; vm_pindex_t idx, idx1; DMAR_DOMAIN_ASSERT_PGLOCKED(domain); KASSERT((flags & IOMMU_PGF_OBJL) != 0, ("lost PGF_OBJL")); idx = pglvl_pgtbl_get_pindex(domain->pglvl, base, lvl); if (*sf != NULL && idx == *idxp) { pte = (iommu_pte_t *)sf_buf_kva(*sf); } else { if (*sf != NULL) iommu_unmap_pgtbl(*sf); *idxp = idx; retry: pte = iommu_map_pgtbl(domain->pgtbl_obj, idx, flags, sf); if (pte == NULL) { KASSERT(lvl > 0, ("lost root page table page %p", domain)); /* * Page table page does not exist, allocate * it and create a pte in the preceeding page level * to reference the allocated page table page. */ m = iommu_pgalloc(domain->pgtbl_obj, idx, flags | IOMMU_PGF_ZERO); if (m == NULL) return (NULL); /* * Prevent potential free while pgtbl_obj is * unlocked in the recursive call to * domain_pgtbl_map_pte(), if other thread did * pte write and clean while the lock is * dropped. */ vm_page_wire(m); sfp = NULL; ptep = dmar_pgtbl_map_pte(domain, base, lvl - 1, flags, &idx1, &sfp); if (ptep == NULL) { KASSERT(m->pindex != 0, ("loosing root page %p", domain)); vm_page_unwire_noq(m); iommu_pgfree(domain->pgtbl_obj, m->pindex, flags, NULL); return (NULL); } dmar_pte_store(&ptep->pte, DMAR_PTE_R | DMAR_PTE_W | VM_PAGE_TO_PHYS(m)); dmar_flush_pte_to_ram(domain->dmar, ptep); vm_page_wire(sf_buf_page(sfp)); vm_page_unwire_noq(m); iommu_unmap_pgtbl(sfp); /* Only executed once. */ goto retry; } } pte += pglvl_pgtbl_pte_off(domain->pglvl, base, lvl); return (pte); } static int dmar_map_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size, vm_page_t *ma, uint64_t pflags, int flags, struct iommu_map_entry *entry) { iommu_pte_t *pte; struct sf_buf *sf; iommu_gaddr_t pg_sz, base1; vm_pindex_t pi, c, idx, run_sz; int lvl; bool superpage; DMAR_DOMAIN_ASSERT_PGLOCKED(domain); base1 = base; flags |= IOMMU_PGF_OBJL; TD_PREP_PINNED_ASSERT; for (sf = NULL, pi = 0; size > 0; base += pg_sz, size -= pg_sz, pi += run_sz) { for (lvl = 0, c = 0, superpage = false;; lvl++) { pg_sz = domain_page_size(domain, lvl); run_sz = pg_sz >> IOMMU_PAGE_SHIFT; if (lvl == domain->pglvl - 1) break; /* * Check if the current base suitable for the * superpage mapping. First, verify the level. */ if (!domain_is_sp_lvl(domain, lvl)) continue; /* * Next, look at the size of the mapping and * alignment of both guest and host addresses. */ if (size < pg_sz || (base & (pg_sz - 1)) != 0 || (VM_PAGE_TO_PHYS(ma[pi]) & (pg_sz - 1)) != 0) continue; /* All passed, check host pages contiguouty. */ if (c == 0) { for (c = 1; c < run_sz; c++) { if (VM_PAGE_TO_PHYS(ma[pi + c]) != VM_PAGE_TO_PHYS(ma[pi + c - 1]) + PAGE_SIZE) break; } } if (c >= run_sz) { superpage = true; break; } } KASSERT(size >= pg_sz, ("mapping loop overflow %p %jx %jx %jx", domain, (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz)); KASSERT(pg_sz > 0, ("pg_sz 0 lvl %d", lvl)); pte = dmar_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf); if (pte == NULL) { KASSERT((flags & IOMMU_PGF_WAITOK) == 0, ("failed waitable pte alloc %p", domain)); if (sf != NULL) iommu_unmap_pgtbl(sf); dmar_unmap_buf_locked(domain, base1, base - base1, flags, entry); TD_PINNED_ASSERT; return (ENOMEM); } dmar_pte_store(&pte->pte, VM_PAGE_TO_PHYS(ma[pi]) | pflags | (superpage ? DMAR_PTE_SP : 0)); dmar_flush_pte_to_ram(domain->dmar, pte); vm_page_wire(sf_buf_page(sf)); } if (sf != NULL) iommu_unmap_pgtbl(sf); TD_PINNED_ASSERT; return (0); } static int dmar_map_buf(struct iommu_domain *iodom, struct iommu_map_entry *entry, vm_page_t *ma, uint64_t eflags, int flags) { struct dmar_domain *domain; struct dmar_unit *unit; iommu_gaddr_t base, size; uint64_t pflags; int error; base = entry->start; size = entry->end - entry->start; pflags = ((eflags & IOMMU_MAP_ENTRY_READ) != 0 ? DMAR_PTE_R : 0) | ((eflags & IOMMU_MAP_ENTRY_WRITE) != 0 ? DMAR_PTE_W : 0) | ((eflags & IOMMU_MAP_ENTRY_SNOOP) != 0 ? DMAR_PTE_SNP : 0) | ((eflags & IOMMU_MAP_ENTRY_TM) != 0 ? DMAR_PTE_TM : 0); domain = IODOM2DOM(iodom); unit = domain->dmar; KASSERT((iodom->flags & IOMMU_DOMAIN_IDMAP) == 0, ("modifying idmap pagetable domain %p", domain)); KASSERT((base & IOMMU_PAGE_MASK) == 0, ("non-aligned base %p %jx %jx", domain, (uintmax_t)base, (uintmax_t)size)); KASSERT((size & IOMMU_PAGE_MASK) == 0, ("non-aligned size %p %jx %jx", domain, (uintmax_t)base, (uintmax_t)size)); KASSERT(size > 0, ("zero size %p %jx %jx", domain, (uintmax_t)base, (uintmax_t)size)); KASSERT(base < (1ULL << domain->agaw), ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base, (uintmax_t)size, domain->agaw)); KASSERT(base + size < (1ULL << domain->agaw), ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base, (uintmax_t)size, domain->agaw)); KASSERT(base + size > base, ("size overflow %p %jx %jx", domain, (uintmax_t)base, (uintmax_t)size)); KASSERT((pflags & (DMAR_PTE_R | DMAR_PTE_W)) != 0, ("neither read nor write %jx", (uintmax_t)pflags)); KASSERT((pflags & ~(DMAR_PTE_R | DMAR_PTE_W | DMAR_PTE_SNP | DMAR_PTE_TM)) == 0, ("invalid pte flags %jx", (uintmax_t)pflags)); KASSERT((pflags & DMAR_PTE_SNP) == 0 || (unit->hw_ecap & DMAR_ECAP_SC) != 0, ("PTE_SNP for dmar without snoop control %p %jx", domain, (uintmax_t)pflags)); KASSERT((pflags & DMAR_PTE_TM) == 0 || (unit->hw_ecap & DMAR_ECAP_DI) != 0, ("PTE_TM for dmar without DIOTLB %p %jx", domain, (uintmax_t)pflags)); KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags)); DMAR_DOMAIN_PGLOCK(domain); error = dmar_map_buf_locked(domain, base, size, ma, pflags, flags, entry); DMAR_DOMAIN_PGUNLOCK(domain); if (error != 0) return (error); if ((unit->hw_cap & DMAR_CAP_CM) != 0) dmar_flush_iotlb_sync(domain, base, size); else if ((unit->hw_cap & DMAR_CAP_RWBF) != 0) { /* See 11.1 Write Buffer Flushing. */ DMAR_LOCK(unit); dmar_flush_write_bufs(unit); DMAR_UNLOCK(unit); } return (0); } static void dmar_unmap_clear_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl, int flags, iommu_pte_t *pte, struct sf_buf **sf, struct iommu_map_entry *entry, bool free_fs); static void dmar_free_pgtbl_pde(struct dmar_domain *domain, iommu_gaddr_t base, int lvl, int flags, struct iommu_map_entry *entry) { struct sf_buf *sf; iommu_pte_t *pde; vm_pindex_t idx; sf = NULL; pde = dmar_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf); dmar_unmap_clear_pte(domain, base, lvl, flags, pde, &sf, entry, true); } static void dmar_unmap_clear_pte(struct dmar_domain *domain, iommu_gaddr_t base, int lvl, int flags, iommu_pte_t *pte, struct sf_buf **sf, struct iommu_map_entry *entry, bool free_sf) { vm_page_t m; dmar_pte_clear(&pte->pte); dmar_flush_pte_to_ram(domain->dmar, pte); m = sf_buf_page(*sf); if (free_sf) { iommu_unmap_pgtbl(*sf); *sf = NULL; } if (!vm_page_unwire_noq(m)) return; KASSERT(lvl != 0, ("lost reference (lvl) on root pg domain %p base %jx lvl %d", domain, (uintmax_t)base, lvl)); KASSERT(m->pindex != 0, ("lost reference (idx) on root pg domain %p base %jx lvl %d", domain, (uintmax_t)base, lvl)); iommu_pgfree(domain->pgtbl_obj, m->pindex, flags, entry); dmar_free_pgtbl_pde(domain, base, lvl - 1, flags, entry); } /* * Assumes that the unmap is never partial. */ static int dmar_unmap_buf_locked(struct dmar_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size, int flags, struct iommu_map_entry *entry) { iommu_pte_t *pte; struct sf_buf *sf; vm_pindex_t idx; iommu_gaddr_t pg_sz; int lvl; DMAR_DOMAIN_ASSERT_PGLOCKED(domain); if (size == 0) return (0); KASSERT((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) == 0, ("modifying idmap pagetable domain %p", domain)); KASSERT((base & IOMMU_PAGE_MASK) == 0, ("non-aligned base %p %jx %jx", domain, (uintmax_t)base, (uintmax_t)size)); KASSERT((size & IOMMU_PAGE_MASK) == 0, ("non-aligned size %p %jx %jx", domain, (uintmax_t)base, (uintmax_t)size)); KASSERT(base < (1ULL << domain->agaw), ("base too high %p %jx %jx agaw %d", domain, (uintmax_t)base, (uintmax_t)size, domain->agaw)); KASSERT(base + size < (1ULL << domain->agaw), ("end too high %p %jx %jx agaw %d", domain, (uintmax_t)base, (uintmax_t)size, domain->agaw)); KASSERT(base + size > base, ("size overflow %p %jx %jx", domain, (uintmax_t)base, (uintmax_t)size)); KASSERT((flags & ~IOMMU_PGF_WAITOK) == 0, ("invalid flags %x", flags)); pg_sz = 0; /* silence gcc */ flags |= IOMMU_PGF_OBJL; TD_PREP_PINNED_ASSERT; for (sf = NULL; size > 0; base += pg_sz, size -= pg_sz) { for (lvl = 0; lvl < domain->pglvl; lvl++) { if (lvl != domain->pglvl - 1 && !domain_is_sp_lvl(domain, lvl)) continue; pg_sz = domain_page_size(domain, lvl); if (pg_sz > size) continue; pte = dmar_pgtbl_map_pte(domain, base, lvl, flags, &idx, &sf); KASSERT(pte != NULL, ("sleeping or page missed %p %jx %d 0x%x", domain, (uintmax_t)base, lvl, flags)); if ((pte->pte & DMAR_PTE_SP) != 0 || lvl == domain->pglvl - 1) { dmar_unmap_clear_pte(domain, base, lvl, flags, pte, &sf, entry, false); break; } } KASSERT(size >= pg_sz, ("unmapping loop overflow %p %jx %jx %jx", domain, (uintmax_t)base, (uintmax_t)size, (uintmax_t)pg_sz)); } if (sf != NULL) iommu_unmap_pgtbl(sf); /* * See 11.1 Write Buffer Flushing for an explanation why RWBF * can be ignored there. */ TD_PINNED_ASSERT; return (0); } static int dmar_unmap_buf(struct iommu_domain *iodom, struct iommu_map_entry *entry, int flags) { struct dmar_domain *domain; int error; domain = IODOM2DOM(iodom); DMAR_DOMAIN_PGLOCK(domain); error = dmar_unmap_buf_locked(domain, entry->start, entry->end - entry->start, flags, entry); DMAR_DOMAIN_PGUNLOCK(domain); return (error); } int dmar_domain_alloc_pgtbl(struct dmar_domain *domain) { vm_page_t m; + struct dmar_unit *unit; KASSERT(domain->pgtbl_obj == NULL, ("already initialized %p", domain)); + unit = domain->dmar; domain->pgtbl_obj = vm_pager_allocate(OBJT_PHYS, NULL, IDX_TO_OFF(pglvl_max_pages(domain->pglvl)), 0, 0, NULL); + if (unit->memdomain != -1) { + domain->pgtbl_obj->domain.dr_policy = DOMAINSET_PREF( + unit->memdomain); + } DMAR_DOMAIN_PGLOCK(domain); m = iommu_pgalloc(domain->pgtbl_obj, 0, IOMMU_PGF_WAITOK | IOMMU_PGF_ZERO | IOMMU_PGF_OBJL); /* No implicit free of the top level page table page. */ vm_page_wire(m); DMAR_DOMAIN_PGUNLOCK(domain); - DMAR_LOCK(domain->dmar); + DMAR_LOCK(unit); domain->iodom.flags |= IOMMU_DOMAIN_PGTBL_INITED; - DMAR_UNLOCK(domain->dmar); + DMAR_UNLOCK(unit); return (0); } void dmar_domain_free_pgtbl(struct dmar_domain *domain) { vm_object_t obj; vm_page_t m; obj = domain->pgtbl_obj; if (obj == NULL) { KASSERT((domain->dmar->hw_ecap & DMAR_ECAP_PT) != 0 && (domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0, ("lost pagetable object domain %p", domain)); return; } DMAR_DOMAIN_ASSERT_PGLOCKED(domain); domain->pgtbl_obj = NULL; if ((domain->iodom.flags & IOMMU_DOMAIN_IDMAP) != 0) { dmar_put_idmap_pgtbl(obj); domain->iodom.flags &= ~IOMMU_DOMAIN_IDMAP; return; } /* Obliterate ref_counts */ VM_OBJECT_ASSERT_WLOCKED(obj); for (m = vm_page_lookup(obj, 0); m != NULL; m = vm_page_next(m)) { vm_page_clearref(m); vm_wire_sub(1); } VM_OBJECT_WUNLOCK(obj); vm_object_deallocate(obj); } static inline uint64_t dmar_wait_iotlb_flush(struct dmar_unit *unit, uint64_t wt, int iro) { uint64_t iotlbr; dmar_write8(unit, iro + DMAR_IOTLB_REG_OFF, DMAR_IOTLB_IVT | DMAR_IOTLB_DR | DMAR_IOTLB_DW | wt); for (;;) { iotlbr = dmar_read8(unit, iro + DMAR_IOTLB_REG_OFF); if ((iotlbr & DMAR_IOTLB_IVT) == 0) break; cpu_spinwait(); } return (iotlbr); } void dmar_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base, iommu_gaddr_t size) { struct dmar_unit *unit; iommu_gaddr_t isize; uint64_t iotlbr; int am, iro; unit = domain->dmar; KASSERT(!unit->qi_enabled, ("dmar%d: sync iotlb flush call", unit->iommu.unit)); iro = DMAR_ECAP_IRO(unit->hw_ecap) * 16; DMAR_LOCK(unit); if ((unit->hw_cap & DMAR_CAP_PSI) == 0 || size > 2 * 1024 * 1024) { iotlbr = dmar_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_DOM | DMAR_IOTLB_DID(domain->domain), iro); KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) != DMAR_IOTLB_IAIG_INVLD, ("dmar%d: invalidation failed %jx", unit->iommu.unit, (uintmax_t)iotlbr)); } else { for (; size > 0; base += isize, size -= isize) { am = calc_am(unit, base, size, &isize); dmar_write8(unit, iro, base | am); iotlbr = dmar_wait_iotlb_flush(unit, DMAR_IOTLB_IIRG_PAGE | DMAR_IOTLB_DID(domain->domain), iro); KASSERT((iotlbr & DMAR_IOTLB_IAIG_MASK) != DMAR_IOTLB_IAIG_INVLD, ("dmar%d: PSI invalidation failed " "iotlbr 0x%jx base 0x%jx size 0x%jx am %d", unit->iommu.unit, (uintmax_t)iotlbr, (uintmax_t)base, (uintmax_t)size, am)); /* * Any non-page granularity covers whole guest * address space for the domain. */ if ((iotlbr & DMAR_IOTLB_IAIG_MASK) != DMAR_IOTLB_IAIG_PAGE) break; } } DMAR_UNLOCK(unit); } const struct iommu_domain_map_ops dmar_domain_map_ops = { .map = dmar_map_buf, .unmap = dmar_unmap_buf, }; diff --git a/sys/x86/iommu/intel_intrmap.c b/sys/x86/iommu/intel_intrmap.c index 87bb7c791dd0..06e41523624b 100644 --- a/sys/x86/iommu/intel_intrmap.c +++ b/sys/x86/iommu/intel_intrmap.c @@ -1,378 +1,389 @@ /*- * Copyright (c) 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 #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 struct dmar_unit *dmar_ir_find(device_t src, uint16_t *rid, int *is_dmar); static void dmar_ir_program_irte(struct dmar_unit *unit, u_int idx, uint64_t low, uint16_t rid); static int dmar_ir_free_irte(struct dmar_unit *unit, u_int cookie); int dmar_alloc_msi_intr(device_t src, u_int *cookies, u_int count) { struct dmar_unit *unit; vmem_addr_t vmem_res; u_int idx, i; int error; unit = dmar_ir_find(src, NULL, NULL); if (unit == NULL || !unit->ir_enabled) { for (i = 0; i < count; i++) cookies[i] = -1; return (EOPNOTSUPP); } error = vmem_alloc(unit->irtids, count, M_FIRSTFIT | M_NOWAIT, &vmem_res); if (error != 0) { KASSERT(error != EOPNOTSUPP, ("impossible EOPNOTSUPP from vmem")); return (error); } idx = vmem_res; for (i = 0; i < count; i++) cookies[i] = idx + i; return (0); } int dmar_map_msi_intr(device_t src, u_int cpu, u_int vector, u_int cookie, uint64_t *addr, uint32_t *data) { struct dmar_unit *unit; uint64_t low; uint16_t rid; int is_dmar; unit = dmar_ir_find(src, &rid, &is_dmar); if (is_dmar) { KASSERT(unit == NULL, ("DMAR cannot translate itself")); /* * See VT-d specification, 5.1.6 Remapping Hardware - * Interrupt Programming. */ *data = vector; *addr = MSI_INTEL_ADDR_BASE | ((cpu & 0xff) << 12); if (x2apic_mode) *addr |= ((uint64_t)cpu & 0xffffff00) << 32; else KASSERT(cpu <= 0xff, ("cpu id too big %d", cpu)); return (0); } if (unit == NULL || !unit->ir_enabled || cookie == -1) return (EOPNOTSUPP); low = (DMAR_X2APIC(unit) ? DMAR_IRTE1_DST_x2APIC(cpu) : DMAR_IRTE1_DST_xAPIC(cpu)) | DMAR_IRTE1_V(vector) | DMAR_IRTE1_DLM_FM | DMAR_IRTE1_TM_EDGE | DMAR_IRTE1_RH_DIRECT | DMAR_IRTE1_DM_PHYSICAL | DMAR_IRTE1_P; dmar_ir_program_irte(unit, cookie, low, rid); if (addr != NULL) { /* * See VT-d specification, 5.1.5.2 MSI and MSI-X * Register Programming. */ *addr = MSI_INTEL_ADDR_BASE | ((cookie & 0x7fff) << 5) | ((cookie & 0x8000) << 2) | 0x18; *data = 0; } return (0); } int dmar_unmap_msi_intr(device_t src, u_int cookie) { struct dmar_unit *unit; if (cookie == -1) return (0); unit = dmar_ir_find(src, NULL, NULL); return (dmar_ir_free_irte(unit, cookie)); } int dmar_map_ioapic_intr(u_int ioapic_id, u_int cpu, u_int vector, bool edge, bool activehi, int irq, u_int *cookie, uint32_t *hi, uint32_t *lo) { struct dmar_unit *unit; vmem_addr_t vmem_res; uint64_t low, iorte; u_int idx; int error; uint16_t rid; unit = dmar_find_ioapic(ioapic_id, &rid); if (unit == NULL || !unit->ir_enabled) { *cookie = -1; return (EOPNOTSUPP); } error = vmem_alloc(unit->irtids, 1, M_FIRSTFIT | M_NOWAIT, &vmem_res); if (error != 0) { KASSERT(error != EOPNOTSUPP, ("impossible EOPNOTSUPP from vmem")); return (error); } idx = vmem_res; low = 0; switch (irq) { case IRQ_EXTINT: low |= DMAR_IRTE1_DLM_ExtINT; break; case IRQ_NMI: low |= DMAR_IRTE1_DLM_NMI; break; case IRQ_SMI: low |= DMAR_IRTE1_DLM_SMI; break; default: KASSERT(vector != 0, ("No vector for IRQ %u", irq)); low |= DMAR_IRTE1_DLM_FM | DMAR_IRTE1_V(vector); break; } low |= (DMAR_X2APIC(unit) ? DMAR_IRTE1_DST_x2APIC(cpu) : DMAR_IRTE1_DST_xAPIC(cpu)) | (edge ? DMAR_IRTE1_TM_EDGE : DMAR_IRTE1_TM_LEVEL) | DMAR_IRTE1_RH_DIRECT | DMAR_IRTE1_DM_PHYSICAL | DMAR_IRTE1_P; dmar_ir_program_irte(unit, idx, low, rid); if (hi != NULL) { /* * See VT-d specification, 5.1.5.1 I/OxAPIC * Programming. */ iorte = (1ULL << 48) | ((uint64_t)(idx & 0x7fff) << 49) | ((idx & 0x8000) != 0 ? (1 << 11) : 0) | (edge ? IOART_TRGREDG : IOART_TRGRLVL) | (activehi ? IOART_INTAHI : IOART_INTALO) | IOART_DELFIXED | vector; *hi = iorte >> 32; *lo = iorte; } *cookie = idx; return (0); } int dmar_unmap_ioapic_intr(u_int ioapic_id, u_int *cookie) { struct dmar_unit *unit; u_int idx; idx = *cookie; if (idx == -1) return (0); *cookie = -1; unit = dmar_find_ioapic(ioapic_id, NULL); KASSERT(unit != NULL && unit->ir_enabled, ("unmap: cookie %d unit %p", idx, unit)); return (dmar_ir_free_irte(unit, idx)); } static struct dmar_unit * dmar_ir_find(device_t src, uint16_t *rid, int *is_dmar) { devclass_t src_class; struct dmar_unit *unit; /* * We need to determine if the interrupt source generates FSB * interrupts. If yes, it is either DMAR, in which case * interrupts are not remapped. Or it is HPET, and interrupts * are remapped. For HPET, source id is reported by HPET * record in DMAR ACPI table. */ if (is_dmar != NULL) *is_dmar = FALSE; src_class = device_get_devclass(src); if (src_class == devclass_find("dmar")) { unit = NULL; if (is_dmar != NULL) *is_dmar = TRUE; } else if (src_class == devclass_find("hpet")) { unit = dmar_find_hpet(src, rid); } else { unit = dmar_find(src, bootverbose); if (unit != NULL && rid != NULL) iommu_get_requester(src, rid); } return (unit); } static void dmar_ir_program_irte(struct dmar_unit *unit, u_int idx, uint64_t low, uint16_t rid) { dmar_irte_t *irte; uint64_t high; KASSERT(idx < unit->irte_cnt, ("bad cookie %d %d", idx, unit->irte_cnt)); irte = &(unit->irt[idx]); high = DMAR_IRTE2_SVT_RID | DMAR_IRTE2_SQ_RID | DMAR_IRTE2_SID_RID(rid); if (bootverbose) { device_printf(unit->iommu.dev, "programming irte[%d] rid %#x high %#jx low %#jx\n", idx, rid, (uintmax_t)high, (uintmax_t)low); } DMAR_LOCK(unit); if ((irte->irte1 & DMAR_IRTE1_P) != 0) { /* * The rte is already valid. Assume that the request * is to remap the interrupt for balancing. Only low * word of rte needs to be changed. Assert that the * high word contains expected value. */ KASSERT(irte->irte2 == high, ("irte2 mismatch, %jx %jx", (uintmax_t)irte->irte2, (uintmax_t)high)); dmar_pte_update(&irte->irte1, low); } else { dmar_pte_store(&irte->irte2, high); dmar_pte_store(&irte->irte1, low); } dmar_qi_invalidate_iec(unit, idx, 1); DMAR_UNLOCK(unit); } static int dmar_ir_free_irte(struct dmar_unit *unit, u_int cookie) { dmar_irte_t *irte; KASSERT(unit != NULL && unit->ir_enabled, ("unmap: cookie %d unit %p", cookie, unit)); KASSERT(cookie < unit->irte_cnt, ("bad cookie %u %u", cookie, unit->irte_cnt)); irte = &(unit->irt[cookie]); dmar_pte_clear(&irte->irte1); dmar_pte_clear(&irte->irte2); DMAR_LOCK(unit); dmar_qi_invalidate_iec(unit, cookie, 1); DMAR_UNLOCK(unit); vmem_free(unit->irtids, cookie, 1); return (0); } int dmar_init_irt(struct dmar_unit *unit) { SYSCTL_ADD_INT(&unit->iommu.sysctl_ctx, SYSCTL_CHILDREN(device_get_sysctl_tree(unit->iommu.dev)), OID_AUTO, "ir", CTLFLAG_RD, &unit->ir_enabled, 0, "Interrupt remapping ops enabled"); if ((unit->hw_ecap & DMAR_ECAP_IR) == 0) return (0); unit->ir_enabled = 1; TUNABLE_INT_FETCH("hw.dmar.ir", &unit->ir_enabled); TUNABLE_INT_FETCH("hw.iommu.ir", &unit->ir_enabled); if (!unit->ir_enabled) return (0); if (!unit->qi_enabled) { unit->ir_enabled = 0; if (bootverbose) device_printf(unit->iommu.dev, "QI disabled, disabling interrupt remapping\n"); return (0); } unit->irte_cnt = roundup_pow_of_two(num_io_irqs); - unit->irt = kmem_alloc_contig(unit->irte_cnt * sizeof(dmar_irte_t), - M_ZERO | M_WAITOK, 0, iommu_high, PAGE_SIZE, 0, - DMAR_IS_COHERENT(unit) ? - VM_MEMATTR_DEFAULT : VM_MEMATTR_UNCACHEABLE); + if (unit->memdomain == -1) { + unit->irt = kmem_alloc_contig( + unit->irte_cnt * sizeof(dmar_irte_t), + M_ZERO | M_WAITOK, 0, iommu_high, PAGE_SIZE, 0, + DMAR_IS_COHERENT(unit) ? + VM_MEMATTR_DEFAULT : VM_MEMATTR_UNCACHEABLE); + } else { + unit->irt = kmem_alloc_contig_domainset( + DOMAINSET_PREF(unit->memdomain), + unit->irte_cnt * sizeof(dmar_irte_t), + M_ZERO | M_WAITOK, 0, iommu_high, PAGE_SIZE, 0, + DMAR_IS_COHERENT(unit) ? + VM_MEMATTR_DEFAULT : VM_MEMATTR_UNCACHEABLE); + } if (unit->irt == NULL) return (ENOMEM); unit->irt_phys = pmap_kextract((vm_offset_t)unit->irt); unit->irtids = vmem_create("dmarirt", 0, unit->irte_cnt, 1, 0, M_FIRSTFIT | M_NOWAIT); DMAR_LOCK(unit); dmar_load_irt_ptr(unit); dmar_qi_invalidate_iec_glob(unit); DMAR_UNLOCK(unit); /* * Initialize mappings for already configured interrupt pins. * Required, because otherwise the interrupts fault without * irtes. */ intr_reprogram(); DMAR_LOCK(unit); dmar_enable_ir(unit); DMAR_UNLOCK(unit); return (0); } void dmar_fini_irt(struct dmar_unit *unit) { unit->ir_enabled = 0; if (unit->irt != NULL) { dmar_disable_ir(unit); dmar_qi_invalidate_iec_glob(unit); vmem_destroy(unit->irtids); kmem_free(unit->irt, unit->irte_cnt * sizeof(dmar_irte_t)); } }