1 /* 2 * QEMU emulation of an Intel IOMMU (VT-d) 3 * (DMA Remapping device) 4 * 5 * Copyright (C) 2013 Knut Omang, Oracle <knut.omang@oracle.com> 6 * Copyright (C) 2014 Le Tan, <tamlokveer@gmail.com> 7 * 8 * This program is free software; you can redistribute it and/or modify 9 * it under the terms of the GNU General Public License as published by 10 * the Free Software Foundation; either version 2 of the License, or 11 * (at your option) any later version. 12 13 * This program is distributed in the hope that it will be useful, 14 * but WITHOUT ANY WARRANTY; without even the implied warranty of 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 16 * GNU General Public License for more details. 17 18 * You should have received a copy of the GNU General Public License along 19 * with this program; if not, see <http://www.gnu.org/licenses/>. 20 */ 21 22 #include "qemu/osdep.h" 23 #include "qemu/error-report.h" 24 #include "qapi/error.h" 25 #include "hw/sysbus.h" 26 #include "exec/address-spaces.h" 27 #include "intel_iommu_internal.h" 28 #include "hw/pci/pci.h" 29 #include "hw/pci/pci_bus.h" 30 #include "hw/i386/pc.h" 31 #include "hw/i386/apic-msidef.h" 32 #include "hw/boards.h" 33 #include "hw/i386/x86-iommu.h" 34 #include "hw/pci-host/q35.h" 35 #include "sysemu/kvm.h" 36 #include "hw/i386/apic_internal.h" 37 #include "kvm_i386.h" 38 39 /*#define DEBUG_INTEL_IOMMU*/ 40 #ifdef DEBUG_INTEL_IOMMU 41 enum { 42 DEBUG_GENERAL, DEBUG_CSR, DEBUG_INV, DEBUG_MMU, DEBUG_FLOG, 43 DEBUG_CACHE, DEBUG_IR, 44 }; 45 #define VTD_DBGBIT(x) (1 << DEBUG_##x) 46 static int vtd_dbgflags = VTD_DBGBIT(GENERAL) | VTD_DBGBIT(CSR); 47 48 #define VTD_DPRINTF(what, fmt, ...) do { \ 49 if (vtd_dbgflags & VTD_DBGBIT(what)) { \ 50 fprintf(stderr, "(vtd)%s: " fmt "\n", __func__, \ 51 ## __VA_ARGS__); } \ 52 } while (0) 53 #else 54 #define VTD_DPRINTF(what, fmt, ...) do {} while (0) 55 #endif 56 57 static void vtd_define_quad(IntelIOMMUState *s, hwaddr addr, uint64_t val, 58 uint64_t wmask, uint64_t w1cmask) 59 { 60 stq_le_p(&s->csr[addr], val); 61 stq_le_p(&s->wmask[addr], wmask); 62 stq_le_p(&s->w1cmask[addr], w1cmask); 63 } 64 65 static void vtd_define_quad_wo(IntelIOMMUState *s, hwaddr addr, uint64_t mask) 66 { 67 stq_le_p(&s->womask[addr], mask); 68 } 69 70 static void vtd_define_long(IntelIOMMUState *s, hwaddr addr, uint32_t val, 71 uint32_t wmask, uint32_t w1cmask) 72 { 73 stl_le_p(&s->csr[addr], val); 74 stl_le_p(&s->wmask[addr], wmask); 75 stl_le_p(&s->w1cmask[addr], w1cmask); 76 } 77 78 static void vtd_define_long_wo(IntelIOMMUState *s, hwaddr addr, uint32_t mask) 79 { 80 stl_le_p(&s->womask[addr], mask); 81 } 82 83 /* "External" get/set operations */ 84 static void vtd_set_quad(IntelIOMMUState *s, hwaddr addr, uint64_t val) 85 { 86 uint64_t oldval = ldq_le_p(&s->csr[addr]); 87 uint64_t wmask = ldq_le_p(&s->wmask[addr]); 88 uint64_t w1cmask = ldq_le_p(&s->w1cmask[addr]); 89 stq_le_p(&s->csr[addr], 90 ((oldval & ~wmask) | (val & wmask)) & ~(w1cmask & val)); 91 } 92 93 static void vtd_set_long(IntelIOMMUState *s, hwaddr addr, uint32_t val) 94 { 95 uint32_t oldval = ldl_le_p(&s->csr[addr]); 96 uint32_t wmask = ldl_le_p(&s->wmask[addr]); 97 uint32_t w1cmask = ldl_le_p(&s->w1cmask[addr]); 98 stl_le_p(&s->csr[addr], 99 ((oldval & ~wmask) | (val & wmask)) & ~(w1cmask & val)); 100 } 101 102 static uint64_t vtd_get_quad(IntelIOMMUState *s, hwaddr addr) 103 { 104 uint64_t val = ldq_le_p(&s->csr[addr]); 105 uint64_t womask = ldq_le_p(&s->womask[addr]); 106 return val & ~womask; 107 } 108 109 static uint32_t vtd_get_long(IntelIOMMUState *s, hwaddr addr) 110 { 111 uint32_t val = ldl_le_p(&s->csr[addr]); 112 uint32_t womask = ldl_le_p(&s->womask[addr]); 113 return val & ~womask; 114 } 115 116 /* "Internal" get/set operations */ 117 static uint64_t vtd_get_quad_raw(IntelIOMMUState *s, hwaddr addr) 118 { 119 return ldq_le_p(&s->csr[addr]); 120 } 121 122 static uint32_t vtd_get_long_raw(IntelIOMMUState *s, hwaddr addr) 123 { 124 return ldl_le_p(&s->csr[addr]); 125 } 126 127 static void vtd_set_quad_raw(IntelIOMMUState *s, hwaddr addr, uint64_t val) 128 { 129 stq_le_p(&s->csr[addr], val); 130 } 131 132 static uint32_t vtd_set_clear_mask_long(IntelIOMMUState *s, hwaddr addr, 133 uint32_t clear, uint32_t mask) 134 { 135 uint32_t new_val = (ldl_le_p(&s->csr[addr]) & ~clear) | mask; 136 stl_le_p(&s->csr[addr], new_val); 137 return new_val; 138 } 139 140 static uint64_t vtd_set_clear_mask_quad(IntelIOMMUState *s, hwaddr addr, 141 uint64_t clear, uint64_t mask) 142 { 143 uint64_t new_val = (ldq_le_p(&s->csr[addr]) & ~clear) | mask; 144 stq_le_p(&s->csr[addr], new_val); 145 return new_val; 146 } 147 148 /* GHashTable functions */ 149 static gboolean vtd_uint64_equal(gconstpointer v1, gconstpointer v2) 150 { 151 return *((const uint64_t *)v1) == *((const uint64_t *)v2); 152 } 153 154 static guint vtd_uint64_hash(gconstpointer v) 155 { 156 return (guint)*(const uint64_t *)v; 157 } 158 159 static gboolean vtd_hash_remove_by_domain(gpointer key, gpointer value, 160 gpointer user_data) 161 { 162 VTDIOTLBEntry *entry = (VTDIOTLBEntry *)value; 163 uint16_t domain_id = *(uint16_t *)user_data; 164 return entry->domain_id == domain_id; 165 } 166 167 /* The shift of an addr for a certain level of paging structure */ 168 static inline uint32_t vtd_slpt_level_shift(uint32_t level) 169 { 170 return VTD_PAGE_SHIFT_4K + (level - 1) * VTD_SL_LEVEL_BITS; 171 } 172 173 static inline uint64_t vtd_slpt_level_page_mask(uint32_t level) 174 { 175 return ~((1ULL << vtd_slpt_level_shift(level)) - 1); 176 } 177 178 static gboolean vtd_hash_remove_by_page(gpointer key, gpointer value, 179 gpointer user_data) 180 { 181 VTDIOTLBEntry *entry = (VTDIOTLBEntry *)value; 182 VTDIOTLBPageInvInfo *info = (VTDIOTLBPageInvInfo *)user_data; 183 uint64_t gfn = (info->addr >> VTD_PAGE_SHIFT_4K) & info->mask; 184 uint64_t gfn_tlb = (info->addr & entry->mask) >> VTD_PAGE_SHIFT_4K; 185 return (entry->domain_id == info->domain_id) && 186 (((entry->gfn & info->mask) == gfn) || 187 (entry->gfn == gfn_tlb)); 188 } 189 190 /* Reset all the gen of VTDAddressSpace to zero and set the gen of 191 * IntelIOMMUState to 1. 192 */ 193 static void vtd_reset_context_cache(IntelIOMMUState *s) 194 { 195 VTDAddressSpace *vtd_as; 196 VTDBus *vtd_bus; 197 GHashTableIter bus_it; 198 uint32_t devfn_it; 199 200 g_hash_table_iter_init(&bus_it, s->vtd_as_by_busptr); 201 202 VTD_DPRINTF(CACHE, "global context_cache_gen=1"); 203 while (g_hash_table_iter_next (&bus_it, NULL, (void**)&vtd_bus)) { 204 for (devfn_it = 0; devfn_it < X86_IOMMU_PCI_DEVFN_MAX; ++devfn_it) { 205 vtd_as = vtd_bus->dev_as[devfn_it]; 206 if (!vtd_as) { 207 continue; 208 } 209 vtd_as->context_cache_entry.context_cache_gen = 0; 210 } 211 } 212 s->context_cache_gen = 1; 213 } 214 215 static void vtd_reset_iotlb(IntelIOMMUState *s) 216 { 217 assert(s->iotlb); 218 g_hash_table_remove_all(s->iotlb); 219 } 220 221 static uint64_t vtd_get_iotlb_key(uint64_t gfn, uint16_t source_id, 222 uint32_t level) 223 { 224 return gfn | ((uint64_t)(source_id) << VTD_IOTLB_SID_SHIFT) | 225 ((uint64_t)(level) << VTD_IOTLB_LVL_SHIFT); 226 } 227 228 static uint64_t vtd_get_iotlb_gfn(hwaddr addr, uint32_t level) 229 { 230 return (addr & vtd_slpt_level_page_mask(level)) >> VTD_PAGE_SHIFT_4K; 231 } 232 233 static VTDIOTLBEntry *vtd_lookup_iotlb(IntelIOMMUState *s, uint16_t source_id, 234 hwaddr addr) 235 { 236 VTDIOTLBEntry *entry; 237 uint64_t key; 238 int level; 239 240 for (level = VTD_SL_PT_LEVEL; level < VTD_SL_PML4_LEVEL; level++) { 241 key = vtd_get_iotlb_key(vtd_get_iotlb_gfn(addr, level), 242 source_id, level); 243 entry = g_hash_table_lookup(s->iotlb, &key); 244 if (entry) { 245 goto out; 246 } 247 } 248 249 out: 250 return entry; 251 } 252 253 static void vtd_update_iotlb(IntelIOMMUState *s, uint16_t source_id, 254 uint16_t domain_id, hwaddr addr, uint64_t slpte, 255 bool read_flags, bool write_flags, 256 uint32_t level) 257 { 258 VTDIOTLBEntry *entry = g_malloc(sizeof(*entry)); 259 uint64_t *key = g_malloc(sizeof(*key)); 260 uint64_t gfn = vtd_get_iotlb_gfn(addr, level); 261 262 VTD_DPRINTF(CACHE, "update iotlb sid 0x%"PRIx16 " gpa 0x%"PRIx64 263 " slpte 0x%"PRIx64 " did 0x%"PRIx16, source_id, addr, slpte, 264 domain_id); 265 if (g_hash_table_size(s->iotlb) >= VTD_IOTLB_MAX_SIZE) { 266 VTD_DPRINTF(CACHE, "iotlb exceeds size limit, forced to reset"); 267 vtd_reset_iotlb(s); 268 } 269 270 entry->gfn = gfn; 271 entry->domain_id = domain_id; 272 entry->slpte = slpte; 273 entry->read_flags = read_flags; 274 entry->write_flags = write_flags; 275 entry->mask = vtd_slpt_level_page_mask(level); 276 *key = vtd_get_iotlb_key(gfn, source_id, level); 277 g_hash_table_replace(s->iotlb, key, entry); 278 } 279 280 /* Given the reg addr of both the message data and address, generate an 281 * interrupt via MSI. 282 */ 283 static void vtd_generate_interrupt(IntelIOMMUState *s, hwaddr mesg_addr_reg, 284 hwaddr mesg_data_reg) 285 { 286 MSIMessage msi; 287 288 assert(mesg_data_reg < DMAR_REG_SIZE); 289 assert(mesg_addr_reg < DMAR_REG_SIZE); 290 291 msi.address = vtd_get_long_raw(s, mesg_addr_reg); 292 msi.data = vtd_get_long_raw(s, mesg_data_reg); 293 294 VTD_DPRINTF(FLOG, "msi: addr 0x%"PRIx64 " data 0x%"PRIx32, 295 msi.address, msi.data); 296 apic_get_class()->send_msi(&msi); 297 } 298 299 /* Generate a fault event to software via MSI if conditions are met. 300 * Notice that the value of FSTS_REG being passed to it should be the one 301 * before any update. 302 */ 303 static void vtd_generate_fault_event(IntelIOMMUState *s, uint32_t pre_fsts) 304 { 305 if (pre_fsts & VTD_FSTS_PPF || pre_fsts & VTD_FSTS_PFO || 306 pre_fsts & VTD_FSTS_IQE) { 307 VTD_DPRINTF(FLOG, "there are previous interrupt conditions " 308 "to be serviced by software, fault event is not generated " 309 "(FSTS_REG 0x%"PRIx32 ")", pre_fsts); 310 return; 311 } 312 vtd_set_clear_mask_long(s, DMAR_FECTL_REG, 0, VTD_FECTL_IP); 313 if (vtd_get_long_raw(s, DMAR_FECTL_REG) & VTD_FECTL_IM) { 314 VTD_DPRINTF(FLOG, "Interrupt Mask set, fault event is not generated"); 315 } else { 316 vtd_generate_interrupt(s, DMAR_FEADDR_REG, DMAR_FEDATA_REG); 317 vtd_set_clear_mask_long(s, DMAR_FECTL_REG, VTD_FECTL_IP, 0); 318 } 319 } 320 321 /* Check if the Fault (F) field of the Fault Recording Register referenced by 322 * @index is Set. 323 */ 324 static bool vtd_is_frcd_set(IntelIOMMUState *s, uint16_t index) 325 { 326 /* Each reg is 128-bit */ 327 hwaddr addr = DMAR_FRCD_REG_OFFSET + (((uint64_t)index) << 4); 328 addr += 8; /* Access the high 64-bit half */ 329 330 assert(index < DMAR_FRCD_REG_NR); 331 332 return vtd_get_quad_raw(s, addr) & VTD_FRCD_F; 333 } 334 335 /* Update the PPF field of Fault Status Register. 336 * Should be called whenever change the F field of any fault recording 337 * registers. 338 */ 339 static void vtd_update_fsts_ppf(IntelIOMMUState *s) 340 { 341 uint32_t i; 342 uint32_t ppf_mask = 0; 343 344 for (i = 0; i < DMAR_FRCD_REG_NR; i++) { 345 if (vtd_is_frcd_set(s, i)) { 346 ppf_mask = VTD_FSTS_PPF; 347 break; 348 } 349 } 350 vtd_set_clear_mask_long(s, DMAR_FSTS_REG, VTD_FSTS_PPF, ppf_mask); 351 VTD_DPRINTF(FLOG, "set PPF of FSTS_REG to %d", ppf_mask ? 1 : 0); 352 } 353 354 static void vtd_set_frcd_and_update_ppf(IntelIOMMUState *s, uint16_t index) 355 { 356 /* Each reg is 128-bit */ 357 hwaddr addr = DMAR_FRCD_REG_OFFSET + (((uint64_t)index) << 4); 358 addr += 8; /* Access the high 64-bit half */ 359 360 assert(index < DMAR_FRCD_REG_NR); 361 362 vtd_set_clear_mask_quad(s, addr, 0, VTD_FRCD_F); 363 vtd_update_fsts_ppf(s); 364 } 365 366 /* Must not update F field now, should be done later */ 367 static void vtd_record_frcd(IntelIOMMUState *s, uint16_t index, 368 uint16_t source_id, hwaddr addr, 369 VTDFaultReason fault, bool is_write) 370 { 371 uint64_t hi = 0, lo; 372 hwaddr frcd_reg_addr = DMAR_FRCD_REG_OFFSET + (((uint64_t)index) << 4); 373 374 assert(index < DMAR_FRCD_REG_NR); 375 376 lo = VTD_FRCD_FI(addr); 377 hi = VTD_FRCD_SID(source_id) | VTD_FRCD_FR(fault); 378 if (!is_write) { 379 hi |= VTD_FRCD_T; 380 } 381 vtd_set_quad_raw(s, frcd_reg_addr, lo); 382 vtd_set_quad_raw(s, frcd_reg_addr + 8, hi); 383 VTD_DPRINTF(FLOG, "record to FRCD_REG #%"PRIu16 ": hi 0x%"PRIx64 384 ", lo 0x%"PRIx64, index, hi, lo); 385 } 386 387 /* Try to collapse multiple pending faults from the same requester */ 388 static bool vtd_try_collapse_fault(IntelIOMMUState *s, uint16_t source_id) 389 { 390 uint32_t i; 391 uint64_t frcd_reg; 392 hwaddr addr = DMAR_FRCD_REG_OFFSET + 8; /* The high 64-bit half */ 393 394 for (i = 0; i < DMAR_FRCD_REG_NR; i++) { 395 frcd_reg = vtd_get_quad_raw(s, addr); 396 VTD_DPRINTF(FLOG, "frcd_reg #%d 0x%"PRIx64, i, frcd_reg); 397 if ((frcd_reg & VTD_FRCD_F) && 398 ((frcd_reg & VTD_FRCD_SID_MASK) == source_id)) { 399 return true; 400 } 401 addr += 16; /* 128-bit for each */ 402 } 403 return false; 404 } 405 406 /* Log and report an DMAR (address translation) fault to software */ 407 static void vtd_report_dmar_fault(IntelIOMMUState *s, uint16_t source_id, 408 hwaddr addr, VTDFaultReason fault, 409 bool is_write) 410 { 411 uint32_t fsts_reg = vtd_get_long_raw(s, DMAR_FSTS_REG); 412 413 assert(fault < VTD_FR_MAX); 414 415 if (fault == VTD_FR_RESERVED_ERR) { 416 /* This is not a normal fault reason case. Drop it. */ 417 return; 418 } 419 VTD_DPRINTF(FLOG, "sid 0x%"PRIx16 ", fault %d, addr 0x%"PRIx64 420 ", is_write %d", source_id, fault, addr, is_write); 421 if (fsts_reg & VTD_FSTS_PFO) { 422 VTD_DPRINTF(FLOG, "new fault is not recorded due to " 423 "Primary Fault Overflow"); 424 return; 425 } 426 if (vtd_try_collapse_fault(s, source_id)) { 427 VTD_DPRINTF(FLOG, "new fault is not recorded due to " 428 "compression of faults"); 429 return; 430 } 431 if (vtd_is_frcd_set(s, s->next_frcd_reg)) { 432 VTD_DPRINTF(FLOG, "Primary Fault Overflow and " 433 "new fault is not recorded, set PFO field"); 434 vtd_set_clear_mask_long(s, DMAR_FSTS_REG, 0, VTD_FSTS_PFO); 435 return; 436 } 437 438 vtd_record_frcd(s, s->next_frcd_reg, source_id, addr, fault, is_write); 439 440 if (fsts_reg & VTD_FSTS_PPF) { 441 VTD_DPRINTF(FLOG, "there are pending faults already, " 442 "fault event is not generated"); 443 vtd_set_frcd_and_update_ppf(s, s->next_frcd_reg); 444 s->next_frcd_reg++; 445 if (s->next_frcd_reg == DMAR_FRCD_REG_NR) { 446 s->next_frcd_reg = 0; 447 } 448 } else { 449 vtd_set_clear_mask_long(s, DMAR_FSTS_REG, VTD_FSTS_FRI_MASK, 450 VTD_FSTS_FRI(s->next_frcd_reg)); 451 vtd_set_frcd_and_update_ppf(s, s->next_frcd_reg); /* Will set PPF */ 452 s->next_frcd_reg++; 453 if (s->next_frcd_reg == DMAR_FRCD_REG_NR) { 454 s->next_frcd_reg = 0; 455 } 456 /* This case actually cause the PPF to be Set. 457 * So generate fault event (interrupt). 458 */ 459 vtd_generate_fault_event(s, fsts_reg); 460 } 461 } 462 463 /* Handle Invalidation Queue Errors of queued invalidation interface error 464 * conditions. 465 */ 466 static void vtd_handle_inv_queue_error(IntelIOMMUState *s) 467 { 468 uint32_t fsts_reg = vtd_get_long_raw(s, DMAR_FSTS_REG); 469 470 vtd_set_clear_mask_long(s, DMAR_FSTS_REG, 0, VTD_FSTS_IQE); 471 vtd_generate_fault_event(s, fsts_reg); 472 } 473 474 /* Set the IWC field and try to generate an invalidation completion interrupt */ 475 static void vtd_generate_completion_event(IntelIOMMUState *s) 476 { 477 VTD_DPRINTF(INV, "completes an invalidation wait command with " 478 "Interrupt Flag"); 479 if (vtd_get_long_raw(s, DMAR_ICS_REG) & VTD_ICS_IWC) { 480 VTD_DPRINTF(INV, "there is a previous interrupt condition to be " 481 "serviced by software, " 482 "new invalidation event is not generated"); 483 return; 484 } 485 vtd_set_clear_mask_long(s, DMAR_ICS_REG, 0, VTD_ICS_IWC); 486 vtd_set_clear_mask_long(s, DMAR_IECTL_REG, 0, VTD_IECTL_IP); 487 if (vtd_get_long_raw(s, DMAR_IECTL_REG) & VTD_IECTL_IM) { 488 VTD_DPRINTF(INV, "IM filed in IECTL_REG is set, new invalidation " 489 "event is not generated"); 490 return; 491 } else { 492 /* Generate the interrupt event */ 493 vtd_generate_interrupt(s, DMAR_IEADDR_REG, DMAR_IEDATA_REG); 494 vtd_set_clear_mask_long(s, DMAR_IECTL_REG, VTD_IECTL_IP, 0); 495 } 496 } 497 498 static inline bool vtd_root_entry_present(VTDRootEntry *root) 499 { 500 return root->val & VTD_ROOT_ENTRY_P; 501 } 502 503 static int vtd_get_root_entry(IntelIOMMUState *s, uint8_t index, 504 VTDRootEntry *re) 505 { 506 dma_addr_t addr; 507 508 addr = s->root + index * sizeof(*re); 509 if (dma_memory_read(&address_space_memory, addr, re, sizeof(*re))) { 510 VTD_DPRINTF(GENERAL, "error: fail to access root-entry at 0x%"PRIx64 511 " + %"PRIu8, s->root, index); 512 re->val = 0; 513 return -VTD_FR_ROOT_TABLE_INV; 514 } 515 re->val = le64_to_cpu(re->val); 516 return 0; 517 } 518 519 static inline bool vtd_context_entry_present(VTDContextEntry *context) 520 { 521 return context->lo & VTD_CONTEXT_ENTRY_P; 522 } 523 524 static int vtd_get_context_entry_from_root(VTDRootEntry *root, uint8_t index, 525 VTDContextEntry *ce) 526 { 527 dma_addr_t addr; 528 529 if (!vtd_root_entry_present(root)) { 530 VTD_DPRINTF(GENERAL, "error: root-entry is not present"); 531 return -VTD_FR_ROOT_ENTRY_P; 532 } 533 addr = (root->val & VTD_ROOT_ENTRY_CTP) + index * sizeof(*ce); 534 if (dma_memory_read(&address_space_memory, addr, ce, sizeof(*ce))) { 535 VTD_DPRINTF(GENERAL, "error: fail to access context-entry at 0x%"PRIx64 536 " + %"PRIu8, 537 (uint64_t)(root->val & VTD_ROOT_ENTRY_CTP), index); 538 return -VTD_FR_CONTEXT_TABLE_INV; 539 } 540 ce->lo = le64_to_cpu(ce->lo); 541 ce->hi = le64_to_cpu(ce->hi); 542 return 0; 543 } 544 545 static inline dma_addr_t vtd_get_slpt_base_from_context(VTDContextEntry *ce) 546 { 547 return ce->lo & VTD_CONTEXT_ENTRY_SLPTPTR; 548 } 549 550 static inline uint64_t vtd_get_slpte_addr(uint64_t slpte) 551 { 552 return slpte & VTD_SL_PT_BASE_ADDR_MASK; 553 } 554 555 /* Whether the pte indicates the address of the page frame */ 556 static inline bool vtd_is_last_slpte(uint64_t slpte, uint32_t level) 557 { 558 return level == VTD_SL_PT_LEVEL || (slpte & VTD_SL_PT_PAGE_SIZE_MASK); 559 } 560 561 /* Get the content of a spte located in @base_addr[@index] */ 562 static uint64_t vtd_get_slpte(dma_addr_t base_addr, uint32_t index) 563 { 564 uint64_t slpte; 565 566 assert(index < VTD_SL_PT_ENTRY_NR); 567 568 if (dma_memory_read(&address_space_memory, 569 base_addr + index * sizeof(slpte), &slpte, 570 sizeof(slpte))) { 571 slpte = (uint64_t)-1; 572 return slpte; 573 } 574 slpte = le64_to_cpu(slpte); 575 return slpte; 576 } 577 578 /* Given a gpa and the level of paging structure, return the offset of current 579 * level. 580 */ 581 static inline uint32_t vtd_gpa_level_offset(uint64_t gpa, uint32_t level) 582 { 583 return (gpa >> vtd_slpt_level_shift(level)) & 584 ((1ULL << VTD_SL_LEVEL_BITS) - 1); 585 } 586 587 /* Check Capability Register to see if the @level of page-table is supported */ 588 static inline bool vtd_is_level_supported(IntelIOMMUState *s, uint32_t level) 589 { 590 return VTD_CAP_SAGAW_MASK & s->cap & 591 (1ULL << (level - 2 + VTD_CAP_SAGAW_SHIFT)); 592 } 593 594 /* Get the page-table level that hardware should use for the second-level 595 * page-table walk from the Address Width field of context-entry. 596 */ 597 static inline uint32_t vtd_get_level_from_context_entry(VTDContextEntry *ce) 598 { 599 return 2 + (ce->hi & VTD_CONTEXT_ENTRY_AW); 600 } 601 602 static inline uint32_t vtd_get_agaw_from_context_entry(VTDContextEntry *ce) 603 { 604 return 30 + (ce->hi & VTD_CONTEXT_ENTRY_AW) * 9; 605 } 606 607 static const uint64_t vtd_paging_entry_rsvd_field[] = { 608 [0] = ~0ULL, 609 /* For not large page */ 610 [1] = 0x800ULL | ~(VTD_HAW_MASK | VTD_SL_IGN_COM), 611 [2] = 0x800ULL | ~(VTD_HAW_MASK | VTD_SL_IGN_COM), 612 [3] = 0x800ULL | ~(VTD_HAW_MASK | VTD_SL_IGN_COM), 613 [4] = 0x880ULL | ~(VTD_HAW_MASK | VTD_SL_IGN_COM), 614 /* For large page */ 615 [5] = 0x800ULL | ~(VTD_HAW_MASK | VTD_SL_IGN_COM), 616 [6] = 0x1ff800ULL | ~(VTD_HAW_MASK | VTD_SL_IGN_COM), 617 [7] = 0x3ffff800ULL | ~(VTD_HAW_MASK | VTD_SL_IGN_COM), 618 [8] = 0x880ULL | ~(VTD_HAW_MASK | VTD_SL_IGN_COM), 619 }; 620 621 static bool vtd_slpte_nonzero_rsvd(uint64_t slpte, uint32_t level) 622 { 623 if (slpte & VTD_SL_PT_PAGE_SIZE_MASK) { 624 /* Maybe large page */ 625 return slpte & vtd_paging_entry_rsvd_field[level + 4]; 626 } else { 627 return slpte & vtd_paging_entry_rsvd_field[level]; 628 } 629 } 630 631 /* Given the @gpa, get relevant @slptep. @slpte_level will be the last level 632 * of the translation, can be used for deciding the size of large page. 633 */ 634 static int vtd_gpa_to_slpte(VTDContextEntry *ce, uint64_t gpa, bool is_write, 635 uint64_t *slptep, uint32_t *slpte_level, 636 bool *reads, bool *writes) 637 { 638 dma_addr_t addr = vtd_get_slpt_base_from_context(ce); 639 uint32_t level = vtd_get_level_from_context_entry(ce); 640 uint32_t offset; 641 uint64_t slpte; 642 uint32_t ce_agaw = vtd_get_agaw_from_context_entry(ce); 643 uint64_t access_right_check; 644 645 /* Check if @gpa is above 2^X-1, where X is the minimum of MGAW in CAP_REG 646 * and AW in context-entry. 647 */ 648 if (gpa & ~((1ULL << MIN(ce_agaw, VTD_MGAW)) - 1)) { 649 VTD_DPRINTF(GENERAL, "error: gpa 0x%"PRIx64 " exceeds limits", gpa); 650 return -VTD_FR_ADDR_BEYOND_MGAW; 651 } 652 653 /* FIXME: what is the Atomics request here? */ 654 access_right_check = is_write ? VTD_SL_W : VTD_SL_R; 655 656 while (true) { 657 offset = vtd_gpa_level_offset(gpa, level); 658 slpte = vtd_get_slpte(addr, offset); 659 660 if (slpte == (uint64_t)-1) { 661 VTD_DPRINTF(GENERAL, "error: fail to access second-level paging " 662 "entry at level %"PRIu32 " for gpa 0x%"PRIx64, 663 level, gpa); 664 if (level == vtd_get_level_from_context_entry(ce)) { 665 /* Invalid programming of context-entry */ 666 return -VTD_FR_CONTEXT_ENTRY_INV; 667 } else { 668 return -VTD_FR_PAGING_ENTRY_INV; 669 } 670 } 671 *reads = (*reads) && (slpte & VTD_SL_R); 672 *writes = (*writes) && (slpte & VTD_SL_W); 673 if (!(slpte & access_right_check)) { 674 VTD_DPRINTF(GENERAL, "error: lack of %s permission for " 675 "gpa 0x%"PRIx64 " slpte 0x%"PRIx64, 676 (is_write ? "write" : "read"), gpa, slpte); 677 return is_write ? -VTD_FR_WRITE : -VTD_FR_READ; 678 } 679 if (vtd_slpte_nonzero_rsvd(slpte, level)) { 680 VTD_DPRINTF(GENERAL, "error: non-zero reserved field in second " 681 "level paging entry level %"PRIu32 " slpte 0x%"PRIx64, 682 level, slpte); 683 return -VTD_FR_PAGING_ENTRY_RSVD; 684 } 685 686 if (vtd_is_last_slpte(slpte, level)) { 687 *slptep = slpte; 688 *slpte_level = level; 689 return 0; 690 } 691 addr = vtd_get_slpte_addr(slpte); 692 level--; 693 } 694 } 695 696 /* Map a device to its corresponding domain (context-entry) */ 697 static int vtd_dev_to_context_entry(IntelIOMMUState *s, uint8_t bus_num, 698 uint8_t devfn, VTDContextEntry *ce) 699 { 700 VTDRootEntry re; 701 int ret_fr; 702 703 ret_fr = vtd_get_root_entry(s, bus_num, &re); 704 if (ret_fr) { 705 return ret_fr; 706 } 707 708 if (!vtd_root_entry_present(&re)) { 709 VTD_DPRINTF(GENERAL, "error: root-entry #%"PRIu8 " is not present", 710 bus_num); 711 return -VTD_FR_ROOT_ENTRY_P; 712 } else if (re.rsvd || (re.val & VTD_ROOT_ENTRY_RSVD)) { 713 VTD_DPRINTF(GENERAL, "error: non-zero reserved field in root-entry " 714 "hi 0x%"PRIx64 " lo 0x%"PRIx64, re.rsvd, re.val); 715 return -VTD_FR_ROOT_ENTRY_RSVD; 716 } 717 718 ret_fr = vtd_get_context_entry_from_root(&re, devfn, ce); 719 if (ret_fr) { 720 return ret_fr; 721 } 722 723 if (!vtd_context_entry_present(ce)) { 724 VTD_DPRINTF(GENERAL, 725 "error: context-entry #%"PRIu8 "(bus #%"PRIu8 ") " 726 "is not present", devfn, bus_num); 727 return -VTD_FR_CONTEXT_ENTRY_P; 728 } else if ((ce->hi & VTD_CONTEXT_ENTRY_RSVD_HI) || 729 (ce->lo & VTD_CONTEXT_ENTRY_RSVD_LO)) { 730 VTD_DPRINTF(GENERAL, 731 "error: non-zero reserved field in context-entry " 732 "hi 0x%"PRIx64 " lo 0x%"PRIx64, ce->hi, ce->lo); 733 return -VTD_FR_CONTEXT_ENTRY_RSVD; 734 } 735 /* Check if the programming of context-entry is valid */ 736 if (!vtd_is_level_supported(s, vtd_get_level_from_context_entry(ce))) { 737 VTD_DPRINTF(GENERAL, "error: unsupported Address Width value in " 738 "context-entry hi 0x%"PRIx64 " lo 0x%"PRIx64, 739 ce->hi, ce->lo); 740 return -VTD_FR_CONTEXT_ENTRY_INV; 741 } else if (ce->lo & VTD_CONTEXT_ENTRY_TT) { 742 VTD_DPRINTF(GENERAL, "error: unsupported Translation Type in " 743 "context-entry hi 0x%"PRIx64 " lo 0x%"PRIx64, 744 ce->hi, ce->lo); 745 return -VTD_FR_CONTEXT_ENTRY_INV; 746 } 747 return 0; 748 } 749 750 static inline uint16_t vtd_make_source_id(uint8_t bus_num, uint8_t devfn) 751 { 752 return ((bus_num & 0xffUL) << 8) | (devfn & 0xffUL); 753 } 754 755 static const bool vtd_qualified_faults[] = { 756 [VTD_FR_RESERVED] = false, 757 [VTD_FR_ROOT_ENTRY_P] = false, 758 [VTD_FR_CONTEXT_ENTRY_P] = true, 759 [VTD_FR_CONTEXT_ENTRY_INV] = true, 760 [VTD_FR_ADDR_BEYOND_MGAW] = true, 761 [VTD_FR_WRITE] = true, 762 [VTD_FR_READ] = true, 763 [VTD_FR_PAGING_ENTRY_INV] = true, 764 [VTD_FR_ROOT_TABLE_INV] = false, 765 [VTD_FR_CONTEXT_TABLE_INV] = false, 766 [VTD_FR_ROOT_ENTRY_RSVD] = false, 767 [VTD_FR_PAGING_ENTRY_RSVD] = true, 768 [VTD_FR_CONTEXT_ENTRY_TT] = true, 769 [VTD_FR_RESERVED_ERR] = false, 770 [VTD_FR_MAX] = false, 771 }; 772 773 /* To see if a fault condition is "qualified", which is reported to software 774 * only if the FPD field in the context-entry used to process the faulting 775 * request is 0. 776 */ 777 static inline bool vtd_is_qualified_fault(VTDFaultReason fault) 778 { 779 return vtd_qualified_faults[fault]; 780 } 781 782 static inline bool vtd_is_interrupt_addr(hwaddr addr) 783 { 784 return VTD_INTERRUPT_ADDR_FIRST <= addr && addr <= VTD_INTERRUPT_ADDR_LAST; 785 } 786 787 /* Map dev to context-entry then do a paging-structures walk to do a iommu 788 * translation. 789 * 790 * Called from RCU critical section. 791 * 792 * @bus_num: The bus number 793 * @devfn: The devfn, which is the combined of device and function number 794 * @is_write: The access is a write operation 795 * @entry: IOMMUTLBEntry that contain the addr to be translated and result 796 */ 797 static void vtd_do_iommu_translate(VTDAddressSpace *vtd_as, PCIBus *bus, 798 uint8_t devfn, hwaddr addr, bool is_write, 799 IOMMUTLBEntry *entry) 800 { 801 IntelIOMMUState *s = vtd_as->iommu_state; 802 VTDContextEntry ce; 803 uint8_t bus_num = pci_bus_num(bus); 804 VTDContextCacheEntry *cc_entry = &vtd_as->context_cache_entry; 805 uint64_t slpte, page_mask; 806 uint32_t level; 807 uint16_t source_id = vtd_make_source_id(bus_num, devfn); 808 int ret_fr; 809 bool is_fpd_set = false; 810 bool reads = true; 811 bool writes = true; 812 VTDIOTLBEntry *iotlb_entry; 813 814 /* Check if the request is in interrupt address range */ 815 if (vtd_is_interrupt_addr(addr)) { 816 if (is_write) { 817 /* FIXME: since we don't know the length of the access here, we 818 * treat Non-DWORD length write requests without PASID as 819 * interrupt requests, too. Withoud interrupt remapping support, 820 * we just use 1:1 mapping. 821 */ 822 VTD_DPRINTF(MMU, "write request to interrupt address " 823 "gpa 0x%"PRIx64, addr); 824 entry->iova = addr & VTD_PAGE_MASK_4K; 825 entry->translated_addr = addr & VTD_PAGE_MASK_4K; 826 entry->addr_mask = ~VTD_PAGE_MASK_4K; 827 entry->perm = IOMMU_WO; 828 return; 829 } else { 830 VTD_DPRINTF(GENERAL, "error: read request from interrupt address " 831 "gpa 0x%"PRIx64, addr); 832 vtd_report_dmar_fault(s, source_id, addr, VTD_FR_READ, is_write); 833 return; 834 } 835 } 836 /* Try to fetch slpte form IOTLB */ 837 iotlb_entry = vtd_lookup_iotlb(s, source_id, addr); 838 if (iotlb_entry) { 839 VTD_DPRINTF(CACHE, "hit iotlb sid 0x%"PRIx16 " gpa 0x%"PRIx64 840 " slpte 0x%"PRIx64 " did 0x%"PRIx16, source_id, addr, 841 iotlb_entry->slpte, iotlb_entry->domain_id); 842 slpte = iotlb_entry->slpte; 843 reads = iotlb_entry->read_flags; 844 writes = iotlb_entry->write_flags; 845 page_mask = iotlb_entry->mask; 846 goto out; 847 } 848 /* Try to fetch context-entry from cache first */ 849 if (cc_entry->context_cache_gen == s->context_cache_gen) { 850 VTD_DPRINTF(CACHE, "hit context-cache bus %d devfn %d " 851 "(hi %"PRIx64 " lo %"PRIx64 " gen %"PRIu32 ")", 852 bus_num, devfn, cc_entry->context_entry.hi, 853 cc_entry->context_entry.lo, cc_entry->context_cache_gen); 854 ce = cc_entry->context_entry; 855 is_fpd_set = ce.lo & VTD_CONTEXT_ENTRY_FPD; 856 } else { 857 ret_fr = vtd_dev_to_context_entry(s, bus_num, devfn, &ce); 858 is_fpd_set = ce.lo & VTD_CONTEXT_ENTRY_FPD; 859 if (ret_fr) { 860 ret_fr = -ret_fr; 861 if (is_fpd_set && vtd_is_qualified_fault(ret_fr)) { 862 VTD_DPRINTF(FLOG, "fault processing is disabled for DMA " 863 "requests through this context-entry " 864 "(with FPD Set)"); 865 } else { 866 vtd_report_dmar_fault(s, source_id, addr, ret_fr, is_write); 867 } 868 return; 869 } 870 /* Update context-cache */ 871 VTD_DPRINTF(CACHE, "update context-cache bus %d devfn %d " 872 "(hi %"PRIx64 " lo %"PRIx64 " gen %"PRIu32 "->%"PRIu32 ")", 873 bus_num, devfn, ce.hi, ce.lo, 874 cc_entry->context_cache_gen, s->context_cache_gen); 875 cc_entry->context_entry = ce; 876 cc_entry->context_cache_gen = s->context_cache_gen; 877 } 878 879 ret_fr = vtd_gpa_to_slpte(&ce, addr, is_write, &slpte, &level, 880 &reads, &writes); 881 if (ret_fr) { 882 ret_fr = -ret_fr; 883 if (is_fpd_set && vtd_is_qualified_fault(ret_fr)) { 884 VTD_DPRINTF(FLOG, "fault processing is disabled for DMA requests " 885 "through this context-entry (with FPD Set)"); 886 } else { 887 vtd_report_dmar_fault(s, source_id, addr, ret_fr, is_write); 888 } 889 return; 890 } 891 892 page_mask = vtd_slpt_level_page_mask(level); 893 vtd_update_iotlb(s, source_id, VTD_CONTEXT_ENTRY_DID(ce.hi), addr, slpte, 894 reads, writes, level); 895 out: 896 entry->iova = addr & page_mask; 897 entry->translated_addr = vtd_get_slpte_addr(slpte) & page_mask; 898 entry->addr_mask = ~page_mask; 899 entry->perm = (writes ? 2 : 0) + (reads ? 1 : 0); 900 } 901 902 static void vtd_root_table_setup(IntelIOMMUState *s) 903 { 904 s->root = vtd_get_quad_raw(s, DMAR_RTADDR_REG); 905 s->root_extended = s->root & VTD_RTADDR_RTT; 906 s->root &= VTD_RTADDR_ADDR_MASK; 907 908 VTD_DPRINTF(CSR, "root_table addr 0x%"PRIx64 " %s", s->root, 909 (s->root_extended ? "(extended)" : "")); 910 } 911 912 static void vtd_iec_notify_all(IntelIOMMUState *s, bool global, 913 uint32_t index, uint32_t mask) 914 { 915 x86_iommu_iec_notify_all(X86_IOMMU_DEVICE(s), global, index, mask); 916 } 917 918 static void vtd_interrupt_remap_table_setup(IntelIOMMUState *s) 919 { 920 uint64_t value = 0; 921 value = vtd_get_quad_raw(s, DMAR_IRTA_REG); 922 s->intr_size = 1UL << ((value & VTD_IRTA_SIZE_MASK) + 1); 923 s->intr_root = value & VTD_IRTA_ADDR_MASK; 924 s->intr_eime = value & VTD_IRTA_EIME; 925 926 /* Notify global invalidation */ 927 vtd_iec_notify_all(s, true, 0, 0); 928 929 VTD_DPRINTF(CSR, "int remap table addr 0x%"PRIx64 " size %"PRIu32, 930 s->intr_root, s->intr_size); 931 } 932 933 static void vtd_context_global_invalidate(IntelIOMMUState *s) 934 { 935 s->context_cache_gen++; 936 if (s->context_cache_gen == VTD_CONTEXT_CACHE_GEN_MAX) { 937 vtd_reset_context_cache(s); 938 } 939 } 940 941 942 /* Find the VTD address space currently associated with a given bus number, 943 */ 944 static VTDBus *vtd_find_as_from_bus_num(IntelIOMMUState *s, uint8_t bus_num) 945 { 946 VTDBus *vtd_bus = s->vtd_as_by_bus_num[bus_num]; 947 if (!vtd_bus) { 948 /* Iterate over the registered buses to find the one 949 * which currently hold this bus number, and update the bus_num lookup table: 950 */ 951 GHashTableIter iter; 952 953 g_hash_table_iter_init(&iter, s->vtd_as_by_busptr); 954 while (g_hash_table_iter_next (&iter, NULL, (void**)&vtd_bus)) { 955 if (pci_bus_num(vtd_bus->bus) == bus_num) { 956 s->vtd_as_by_bus_num[bus_num] = vtd_bus; 957 return vtd_bus; 958 } 959 } 960 } 961 return vtd_bus; 962 } 963 964 /* Do a context-cache device-selective invalidation. 965 * @func_mask: FM field after shifting 966 */ 967 static void vtd_context_device_invalidate(IntelIOMMUState *s, 968 uint16_t source_id, 969 uint16_t func_mask) 970 { 971 uint16_t mask; 972 VTDBus *vtd_bus; 973 VTDAddressSpace *vtd_as; 974 uint16_t devfn; 975 uint16_t devfn_it; 976 977 switch (func_mask & 3) { 978 case 0: 979 mask = 0; /* No bits in the SID field masked */ 980 break; 981 case 1: 982 mask = 4; /* Mask bit 2 in the SID field */ 983 break; 984 case 2: 985 mask = 6; /* Mask bit 2:1 in the SID field */ 986 break; 987 case 3: 988 mask = 7; /* Mask bit 2:0 in the SID field */ 989 break; 990 } 991 mask = ~mask; 992 VTD_DPRINTF(INV, "device-selective invalidation source 0x%"PRIx16 993 " mask %"PRIu16, source_id, mask); 994 vtd_bus = vtd_find_as_from_bus_num(s, VTD_SID_TO_BUS(source_id)); 995 if (vtd_bus) { 996 devfn = VTD_SID_TO_DEVFN(source_id); 997 for (devfn_it = 0; devfn_it < X86_IOMMU_PCI_DEVFN_MAX; ++devfn_it) { 998 vtd_as = vtd_bus->dev_as[devfn_it]; 999 if (vtd_as && ((devfn_it & mask) == (devfn & mask))) { 1000 VTD_DPRINTF(INV, "invalidate context-cahce of devfn 0x%"PRIx16, 1001 devfn_it); 1002 vtd_as->context_cache_entry.context_cache_gen = 0; 1003 } 1004 } 1005 } 1006 } 1007 1008 /* Context-cache invalidation 1009 * Returns the Context Actual Invalidation Granularity. 1010 * @val: the content of the CCMD_REG 1011 */ 1012 static uint64_t vtd_context_cache_invalidate(IntelIOMMUState *s, uint64_t val) 1013 { 1014 uint64_t caig; 1015 uint64_t type = val & VTD_CCMD_CIRG_MASK; 1016 1017 switch (type) { 1018 case VTD_CCMD_DOMAIN_INVL: 1019 VTD_DPRINTF(INV, "domain-selective invalidation domain 0x%"PRIx16, 1020 (uint16_t)VTD_CCMD_DID(val)); 1021 /* Fall through */ 1022 case VTD_CCMD_GLOBAL_INVL: 1023 VTD_DPRINTF(INV, "global invalidation"); 1024 caig = VTD_CCMD_GLOBAL_INVL_A; 1025 vtd_context_global_invalidate(s); 1026 break; 1027 1028 case VTD_CCMD_DEVICE_INVL: 1029 caig = VTD_CCMD_DEVICE_INVL_A; 1030 vtd_context_device_invalidate(s, VTD_CCMD_SID(val), VTD_CCMD_FM(val)); 1031 break; 1032 1033 default: 1034 VTD_DPRINTF(GENERAL, "error: invalid granularity"); 1035 caig = 0; 1036 } 1037 return caig; 1038 } 1039 1040 static void vtd_iotlb_global_invalidate(IntelIOMMUState *s) 1041 { 1042 vtd_reset_iotlb(s); 1043 } 1044 1045 static void vtd_iotlb_domain_invalidate(IntelIOMMUState *s, uint16_t domain_id) 1046 { 1047 g_hash_table_foreach_remove(s->iotlb, vtd_hash_remove_by_domain, 1048 &domain_id); 1049 } 1050 1051 static void vtd_iotlb_page_invalidate(IntelIOMMUState *s, uint16_t domain_id, 1052 hwaddr addr, uint8_t am) 1053 { 1054 VTDIOTLBPageInvInfo info; 1055 1056 assert(am <= VTD_MAMV); 1057 info.domain_id = domain_id; 1058 info.addr = addr; 1059 info.mask = ~((1 << am) - 1); 1060 g_hash_table_foreach_remove(s->iotlb, vtd_hash_remove_by_page, &info); 1061 } 1062 1063 /* Flush IOTLB 1064 * Returns the IOTLB Actual Invalidation Granularity. 1065 * @val: the content of the IOTLB_REG 1066 */ 1067 static uint64_t vtd_iotlb_flush(IntelIOMMUState *s, uint64_t val) 1068 { 1069 uint64_t iaig; 1070 uint64_t type = val & VTD_TLB_FLUSH_GRANU_MASK; 1071 uint16_t domain_id; 1072 hwaddr addr; 1073 uint8_t am; 1074 1075 switch (type) { 1076 case VTD_TLB_GLOBAL_FLUSH: 1077 VTD_DPRINTF(INV, "global invalidation"); 1078 iaig = VTD_TLB_GLOBAL_FLUSH_A; 1079 vtd_iotlb_global_invalidate(s); 1080 break; 1081 1082 case VTD_TLB_DSI_FLUSH: 1083 domain_id = VTD_TLB_DID(val); 1084 VTD_DPRINTF(INV, "domain-selective invalidation domain 0x%"PRIx16, 1085 domain_id); 1086 iaig = VTD_TLB_DSI_FLUSH_A; 1087 vtd_iotlb_domain_invalidate(s, domain_id); 1088 break; 1089 1090 case VTD_TLB_PSI_FLUSH: 1091 domain_id = VTD_TLB_DID(val); 1092 addr = vtd_get_quad_raw(s, DMAR_IVA_REG); 1093 am = VTD_IVA_AM(addr); 1094 addr = VTD_IVA_ADDR(addr); 1095 VTD_DPRINTF(INV, "page-selective invalidation domain 0x%"PRIx16 1096 " addr 0x%"PRIx64 " mask %"PRIu8, domain_id, addr, am); 1097 if (am > VTD_MAMV) { 1098 VTD_DPRINTF(GENERAL, "error: supported max address mask value is " 1099 "%"PRIu8, (uint8_t)VTD_MAMV); 1100 iaig = 0; 1101 break; 1102 } 1103 iaig = VTD_TLB_PSI_FLUSH_A; 1104 vtd_iotlb_page_invalidate(s, domain_id, addr, am); 1105 break; 1106 1107 default: 1108 VTD_DPRINTF(GENERAL, "error: invalid granularity"); 1109 iaig = 0; 1110 } 1111 return iaig; 1112 } 1113 1114 static inline bool vtd_queued_inv_enable_check(IntelIOMMUState *s) 1115 { 1116 return s->iq_tail == 0; 1117 } 1118 1119 static inline bool vtd_queued_inv_disable_check(IntelIOMMUState *s) 1120 { 1121 return s->qi_enabled && (s->iq_tail == s->iq_head) && 1122 (s->iq_last_desc_type == VTD_INV_DESC_WAIT); 1123 } 1124 1125 static void vtd_handle_gcmd_qie(IntelIOMMUState *s, bool en) 1126 { 1127 uint64_t iqa_val = vtd_get_quad_raw(s, DMAR_IQA_REG); 1128 1129 VTD_DPRINTF(INV, "Queued Invalidation Enable %s", (en ? "on" : "off")); 1130 if (en) { 1131 if (vtd_queued_inv_enable_check(s)) { 1132 s->iq = iqa_val & VTD_IQA_IQA_MASK; 1133 /* 2^(x+8) entries */ 1134 s->iq_size = 1UL << ((iqa_val & VTD_IQA_QS) + 8); 1135 s->qi_enabled = true; 1136 VTD_DPRINTF(INV, "DMAR_IQA_REG 0x%"PRIx64, iqa_val); 1137 VTD_DPRINTF(INV, "Invalidation Queue addr 0x%"PRIx64 " size %d", 1138 s->iq, s->iq_size); 1139 /* Ok - report back to driver */ 1140 vtd_set_clear_mask_long(s, DMAR_GSTS_REG, 0, VTD_GSTS_QIES); 1141 } else { 1142 VTD_DPRINTF(GENERAL, "error: can't enable Queued Invalidation: " 1143 "tail %"PRIu16, s->iq_tail); 1144 } 1145 } else { 1146 if (vtd_queued_inv_disable_check(s)) { 1147 /* disable Queued Invalidation */ 1148 vtd_set_quad_raw(s, DMAR_IQH_REG, 0); 1149 s->iq_head = 0; 1150 s->qi_enabled = false; 1151 /* Ok - report back to driver */ 1152 vtd_set_clear_mask_long(s, DMAR_GSTS_REG, VTD_GSTS_QIES, 0); 1153 } else { 1154 VTD_DPRINTF(GENERAL, "error: can't disable Queued Invalidation: " 1155 "head %"PRIu16 ", tail %"PRIu16 1156 ", last_descriptor %"PRIu8, 1157 s->iq_head, s->iq_tail, s->iq_last_desc_type); 1158 } 1159 } 1160 } 1161 1162 /* Set Root Table Pointer */ 1163 static void vtd_handle_gcmd_srtp(IntelIOMMUState *s) 1164 { 1165 VTD_DPRINTF(CSR, "set Root Table Pointer"); 1166 1167 vtd_root_table_setup(s); 1168 /* Ok - report back to driver */ 1169 vtd_set_clear_mask_long(s, DMAR_GSTS_REG, 0, VTD_GSTS_RTPS); 1170 } 1171 1172 /* Set Interrupt Remap Table Pointer */ 1173 static void vtd_handle_gcmd_sirtp(IntelIOMMUState *s) 1174 { 1175 VTD_DPRINTF(CSR, "set Interrupt Remap Table Pointer"); 1176 1177 vtd_interrupt_remap_table_setup(s); 1178 /* Ok - report back to driver */ 1179 vtd_set_clear_mask_long(s, DMAR_GSTS_REG, 0, VTD_GSTS_IRTPS); 1180 } 1181 1182 /* Handle Translation Enable/Disable */ 1183 static void vtd_handle_gcmd_te(IntelIOMMUState *s, bool en) 1184 { 1185 VTD_DPRINTF(CSR, "Translation Enable %s", (en ? "on" : "off")); 1186 1187 if (en) { 1188 s->dmar_enabled = true; 1189 /* Ok - report back to driver */ 1190 vtd_set_clear_mask_long(s, DMAR_GSTS_REG, 0, VTD_GSTS_TES); 1191 } else { 1192 s->dmar_enabled = false; 1193 1194 /* Clear the index of Fault Recording Register */ 1195 s->next_frcd_reg = 0; 1196 /* Ok - report back to driver */ 1197 vtd_set_clear_mask_long(s, DMAR_GSTS_REG, VTD_GSTS_TES, 0); 1198 } 1199 } 1200 1201 /* Handle Interrupt Remap Enable/Disable */ 1202 static void vtd_handle_gcmd_ire(IntelIOMMUState *s, bool en) 1203 { 1204 VTD_DPRINTF(CSR, "Interrupt Remap Enable %s", (en ? "on" : "off")); 1205 1206 if (en) { 1207 s->intr_enabled = true; 1208 /* Ok - report back to driver */ 1209 vtd_set_clear_mask_long(s, DMAR_GSTS_REG, 0, VTD_GSTS_IRES); 1210 } else { 1211 s->intr_enabled = false; 1212 /* Ok - report back to driver */ 1213 vtd_set_clear_mask_long(s, DMAR_GSTS_REG, VTD_GSTS_IRES, 0); 1214 } 1215 } 1216 1217 /* Handle write to Global Command Register */ 1218 static void vtd_handle_gcmd_write(IntelIOMMUState *s) 1219 { 1220 uint32_t status = vtd_get_long_raw(s, DMAR_GSTS_REG); 1221 uint32_t val = vtd_get_long_raw(s, DMAR_GCMD_REG); 1222 uint32_t changed = status ^ val; 1223 1224 VTD_DPRINTF(CSR, "value 0x%"PRIx32 " status 0x%"PRIx32, val, status); 1225 if (changed & VTD_GCMD_TE) { 1226 /* Translation enable/disable */ 1227 vtd_handle_gcmd_te(s, val & VTD_GCMD_TE); 1228 } 1229 if (val & VTD_GCMD_SRTP) { 1230 /* Set/update the root-table pointer */ 1231 vtd_handle_gcmd_srtp(s); 1232 } 1233 if (changed & VTD_GCMD_QIE) { 1234 /* Queued Invalidation Enable */ 1235 vtd_handle_gcmd_qie(s, val & VTD_GCMD_QIE); 1236 } 1237 if (val & VTD_GCMD_SIRTP) { 1238 /* Set/update the interrupt remapping root-table pointer */ 1239 vtd_handle_gcmd_sirtp(s); 1240 } 1241 if (changed & VTD_GCMD_IRE) { 1242 /* Interrupt remap enable/disable */ 1243 vtd_handle_gcmd_ire(s, val & VTD_GCMD_IRE); 1244 } 1245 } 1246 1247 /* Handle write to Context Command Register */ 1248 static void vtd_handle_ccmd_write(IntelIOMMUState *s) 1249 { 1250 uint64_t ret; 1251 uint64_t val = vtd_get_quad_raw(s, DMAR_CCMD_REG); 1252 1253 /* Context-cache invalidation request */ 1254 if (val & VTD_CCMD_ICC) { 1255 if (s->qi_enabled) { 1256 VTD_DPRINTF(GENERAL, "error: Queued Invalidation enabled, " 1257 "should not use register-based invalidation"); 1258 return; 1259 } 1260 ret = vtd_context_cache_invalidate(s, val); 1261 /* Invalidation completed. Change something to show */ 1262 vtd_set_clear_mask_quad(s, DMAR_CCMD_REG, VTD_CCMD_ICC, 0ULL); 1263 ret = vtd_set_clear_mask_quad(s, DMAR_CCMD_REG, VTD_CCMD_CAIG_MASK, 1264 ret); 1265 VTD_DPRINTF(INV, "CCMD_REG write-back val: 0x%"PRIx64, ret); 1266 } 1267 } 1268 1269 /* Handle write to IOTLB Invalidation Register */ 1270 static void vtd_handle_iotlb_write(IntelIOMMUState *s) 1271 { 1272 uint64_t ret; 1273 uint64_t val = vtd_get_quad_raw(s, DMAR_IOTLB_REG); 1274 1275 /* IOTLB invalidation request */ 1276 if (val & VTD_TLB_IVT) { 1277 if (s->qi_enabled) { 1278 VTD_DPRINTF(GENERAL, "error: Queued Invalidation enabled, " 1279 "should not use register-based invalidation"); 1280 return; 1281 } 1282 ret = vtd_iotlb_flush(s, val); 1283 /* Invalidation completed. Change something to show */ 1284 vtd_set_clear_mask_quad(s, DMAR_IOTLB_REG, VTD_TLB_IVT, 0ULL); 1285 ret = vtd_set_clear_mask_quad(s, DMAR_IOTLB_REG, 1286 VTD_TLB_FLUSH_GRANU_MASK_A, ret); 1287 VTD_DPRINTF(INV, "IOTLB_REG write-back val: 0x%"PRIx64, ret); 1288 } 1289 } 1290 1291 /* Fetch an Invalidation Descriptor from the Invalidation Queue */ 1292 static bool vtd_get_inv_desc(dma_addr_t base_addr, uint32_t offset, 1293 VTDInvDesc *inv_desc) 1294 { 1295 dma_addr_t addr = base_addr + offset * sizeof(*inv_desc); 1296 if (dma_memory_read(&address_space_memory, addr, inv_desc, 1297 sizeof(*inv_desc))) { 1298 VTD_DPRINTF(GENERAL, "error: fail to fetch Invalidation Descriptor " 1299 "base_addr 0x%"PRIx64 " offset %"PRIu32, base_addr, offset); 1300 inv_desc->lo = 0; 1301 inv_desc->hi = 0; 1302 1303 return false; 1304 } 1305 inv_desc->lo = le64_to_cpu(inv_desc->lo); 1306 inv_desc->hi = le64_to_cpu(inv_desc->hi); 1307 return true; 1308 } 1309 1310 static bool vtd_process_wait_desc(IntelIOMMUState *s, VTDInvDesc *inv_desc) 1311 { 1312 if ((inv_desc->hi & VTD_INV_DESC_WAIT_RSVD_HI) || 1313 (inv_desc->lo & VTD_INV_DESC_WAIT_RSVD_LO)) { 1314 VTD_DPRINTF(GENERAL, "error: non-zero reserved field in Invalidation " 1315 "Wait Descriptor hi 0x%"PRIx64 " lo 0x%"PRIx64, 1316 inv_desc->hi, inv_desc->lo); 1317 return false; 1318 } 1319 if (inv_desc->lo & VTD_INV_DESC_WAIT_SW) { 1320 /* Status Write */ 1321 uint32_t status_data = (uint32_t)(inv_desc->lo >> 1322 VTD_INV_DESC_WAIT_DATA_SHIFT); 1323 1324 assert(!(inv_desc->lo & VTD_INV_DESC_WAIT_IF)); 1325 1326 /* FIXME: need to be masked with HAW? */ 1327 dma_addr_t status_addr = inv_desc->hi; 1328 VTD_DPRINTF(INV, "status data 0x%x, status addr 0x%"PRIx64, 1329 status_data, status_addr); 1330 status_data = cpu_to_le32(status_data); 1331 if (dma_memory_write(&address_space_memory, status_addr, &status_data, 1332 sizeof(status_data))) { 1333 VTD_DPRINTF(GENERAL, "error: fail to perform a coherent write"); 1334 return false; 1335 } 1336 } else if (inv_desc->lo & VTD_INV_DESC_WAIT_IF) { 1337 /* Interrupt flag */ 1338 VTD_DPRINTF(INV, "Invalidation Wait Descriptor interrupt completion"); 1339 vtd_generate_completion_event(s); 1340 } else { 1341 VTD_DPRINTF(GENERAL, "error: invalid Invalidation Wait Descriptor: " 1342 "hi 0x%"PRIx64 " lo 0x%"PRIx64, inv_desc->hi, inv_desc->lo); 1343 return false; 1344 } 1345 return true; 1346 } 1347 1348 static bool vtd_process_context_cache_desc(IntelIOMMUState *s, 1349 VTDInvDesc *inv_desc) 1350 { 1351 if ((inv_desc->lo & VTD_INV_DESC_CC_RSVD) || inv_desc->hi) { 1352 VTD_DPRINTF(GENERAL, "error: non-zero reserved field in Context-cache " 1353 "Invalidate Descriptor"); 1354 return false; 1355 } 1356 switch (inv_desc->lo & VTD_INV_DESC_CC_G) { 1357 case VTD_INV_DESC_CC_DOMAIN: 1358 VTD_DPRINTF(INV, "domain-selective invalidation domain 0x%"PRIx16, 1359 (uint16_t)VTD_INV_DESC_CC_DID(inv_desc->lo)); 1360 /* Fall through */ 1361 case VTD_INV_DESC_CC_GLOBAL: 1362 VTD_DPRINTF(INV, "global invalidation"); 1363 vtd_context_global_invalidate(s); 1364 break; 1365 1366 case VTD_INV_DESC_CC_DEVICE: 1367 vtd_context_device_invalidate(s, VTD_INV_DESC_CC_SID(inv_desc->lo), 1368 VTD_INV_DESC_CC_FM(inv_desc->lo)); 1369 break; 1370 1371 default: 1372 VTD_DPRINTF(GENERAL, "error: invalid granularity in Context-cache " 1373 "Invalidate Descriptor hi 0x%"PRIx64 " lo 0x%"PRIx64, 1374 inv_desc->hi, inv_desc->lo); 1375 return false; 1376 } 1377 return true; 1378 } 1379 1380 static bool vtd_process_iotlb_desc(IntelIOMMUState *s, VTDInvDesc *inv_desc) 1381 { 1382 uint16_t domain_id; 1383 uint8_t am; 1384 hwaddr addr; 1385 1386 if ((inv_desc->lo & VTD_INV_DESC_IOTLB_RSVD_LO) || 1387 (inv_desc->hi & VTD_INV_DESC_IOTLB_RSVD_HI)) { 1388 VTD_DPRINTF(GENERAL, "error: non-zero reserved field in IOTLB " 1389 "Invalidate Descriptor hi 0x%"PRIx64 " lo 0x%"PRIx64, 1390 inv_desc->hi, inv_desc->lo); 1391 return false; 1392 } 1393 1394 switch (inv_desc->lo & VTD_INV_DESC_IOTLB_G) { 1395 case VTD_INV_DESC_IOTLB_GLOBAL: 1396 VTD_DPRINTF(INV, "global invalidation"); 1397 vtd_iotlb_global_invalidate(s); 1398 break; 1399 1400 case VTD_INV_DESC_IOTLB_DOMAIN: 1401 domain_id = VTD_INV_DESC_IOTLB_DID(inv_desc->lo); 1402 VTD_DPRINTF(INV, "domain-selective invalidation domain 0x%"PRIx16, 1403 domain_id); 1404 vtd_iotlb_domain_invalidate(s, domain_id); 1405 break; 1406 1407 case VTD_INV_DESC_IOTLB_PAGE: 1408 domain_id = VTD_INV_DESC_IOTLB_DID(inv_desc->lo); 1409 addr = VTD_INV_DESC_IOTLB_ADDR(inv_desc->hi); 1410 am = VTD_INV_DESC_IOTLB_AM(inv_desc->hi); 1411 VTD_DPRINTF(INV, "page-selective invalidation domain 0x%"PRIx16 1412 " addr 0x%"PRIx64 " mask %"PRIu8, domain_id, addr, am); 1413 if (am > VTD_MAMV) { 1414 VTD_DPRINTF(GENERAL, "error: supported max address mask value is " 1415 "%"PRIu8, (uint8_t)VTD_MAMV); 1416 return false; 1417 } 1418 vtd_iotlb_page_invalidate(s, domain_id, addr, am); 1419 break; 1420 1421 default: 1422 VTD_DPRINTF(GENERAL, "error: invalid granularity in IOTLB Invalidate " 1423 "Descriptor hi 0x%"PRIx64 " lo 0x%"PRIx64, 1424 inv_desc->hi, inv_desc->lo); 1425 return false; 1426 } 1427 return true; 1428 } 1429 1430 static bool vtd_process_inv_iec_desc(IntelIOMMUState *s, 1431 VTDInvDesc *inv_desc) 1432 { 1433 VTD_DPRINTF(INV, "inv ir glob %d index %d mask %d", 1434 inv_desc->iec.granularity, 1435 inv_desc->iec.index, 1436 inv_desc->iec.index_mask); 1437 1438 vtd_iec_notify_all(s, !inv_desc->iec.granularity, 1439 inv_desc->iec.index, 1440 inv_desc->iec.index_mask); 1441 1442 return true; 1443 } 1444 1445 static bool vtd_process_inv_desc(IntelIOMMUState *s) 1446 { 1447 VTDInvDesc inv_desc; 1448 uint8_t desc_type; 1449 1450 VTD_DPRINTF(INV, "iq head %"PRIu16, s->iq_head); 1451 if (!vtd_get_inv_desc(s->iq, s->iq_head, &inv_desc)) { 1452 s->iq_last_desc_type = VTD_INV_DESC_NONE; 1453 return false; 1454 } 1455 desc_type = inv_desc.lo & VTD_INV_DESC_TYPE; 1456 /* FIXME: should update at first or at last? */ 1457 s->iq_last_desc_type = desc_type; 1458 1459 switch (desc_type) { 1460 case VTD_INV_DESC_CC: 1461 VTD_DPRINTF(INV, "Context-cache Invalidate Descriptor hi 0x%"PRIx64 1462 " lo 0x%"PRIx64, inv_desc.hi, inv_desc.lo); 1463 if (!vtd_process_context_cache_desc(s, &inv_desc)) { 1464 return false; 1465 } 1466 break; 1467 1468 case VTD_INV_DESC_IOTLB: 1469 VTD_DPRINTF(INV, "IOTLB Invalidate Descriptor hi 0x%"PRIx64 1470 " lo 0x%"PRIx64, inv_desc.hi, inv_desc.lo); 1471 if (!vtd_process_iotlb_desc(s, &inv_desc)) { 1472 return false; 1473 } 1474 break; 1475 1476 case VTD_INV_DESC_WAIT: 1477 VTD_DPRINTF(INV, "Invalidation Wait Descriptor hi 0x%"PRIx64 1478 " lo 0x%"PRIx64, inv_desc.hi, inv_desc.lo); 1479 if (!vtd_process_wait_desc(s, &inv_desc)) { 1480 return false; 1481 } 1482 break; 1483 1484 case VTD_INV_DESC_IEC: 1485 VTD_DPRINTF(INV, "Invalidation Interrupt Entry Cache " 1486 "Descriptor hi 0x%"PRIx64 " lo 0x%"PRIx64, 1487 inv_desc.hi, inv_desc.lo); 1488 if (!vtd_process_inv_iec_desc(s, &inv_desc)) { 1489 return false; 1490 } 1491 break; 1492 1493 default: 1494 VTD_DPRINTF(GENERAL, "error: unkonw Invalidation Descriptor type " 1495 "hi 0x%"PRIx64 " lo 0x%"PRIx64 " type %"PRIu8, 1496 inv_desc.hi, inv_desc.lo, desc_type); 1497 return false; 1498 } 1499 s->iq_head++; 1500 if (s->iq_head == s->iq_size) { 1501 s->iq_head = 0; 1502 } 1503 return true; 1504 } 1505 1506 /* Try to fetch and process more Invalidation Descriptors */ 1507 static void vtd_fetch_inv_desc(IntelIOMMUState *s) 1508 { 1509 VTD_DPRINTF(INV, "fetch Invalidation Descriptors"); 1510 if (s->iq_tail >= s->iq_size) { 1511 /* Detects an invalid Tail pointer */ 1512 VTD_DPRINTF(GENERAL, "error: iq_tail is %"PRIu16 1513 " while iq_size is %"PRIu16, s->iq_tail, s->iq_size); 1514 vtd_handle_inv_queue_error(s); 1515 return; 1516 } 1517 while (s->iq_head != s->iq_tail) { 1518 if (!vtd_process_inv_desc(s)) { 1519 /* Invalidation Queue Errors */ 1520 vtd_handle_inv_queue_error(s); 1521 break; 1522 } 1523 /* Must update the IQH_REG in time */ 1524 vtd_set_quad_raw(s, DMAR_IQH_REG, 1525 (((uint64_t)(s->iq_head)) << VTD_IQH_QH_SHIFT) & 1526 VTD_IQH_QH_MASK); 1527 } 1528 } 1529 1530 /* Handle write to Invalidation Queue Tail Register */ 1531 static void vtd_handle_iqt_write(IntelIOMMUState *s) 1532 { 1533 uint64_t val = vtd_get_quad_raw(s, DMAR_IQT_REG); 1534 1535 s->iq_tail = VTD_IQT_QT(val); 1536 VTD_DPRINTF(INV, "set iq tail %"PRIu16, s->iq_tail); 1537 if (s->qi_enabled && !(vtd_get_long_raw(s, DMAR_FSTS_REG) & VTD_FSTS_IQE)) { 1538 /* Process Invalidation Queue here */ 1539 vtd_fetch_inv_desc(s); 1540 } 1541 } 1542 1543 static void vtd_handle_fsts_write(IntelIOMMUState *s) 1544 { 1545 uint32_t fsts_reg = vtd_get_long_raw(s, DMAR_FSTS_REG); 1546 uint32_t fectl_reg = vtd_get_long_raw(s, DMAR_FECTL_REG); 1547 uint32_t status_fields = VTD_FSTS_PFO | VTD_FSTS_PPF | VTD_FSTS_IQE; 1548 1549 if ((fectl_reg & VTD_FECTL_IP) && !(fsts_reg & status_fields)) { 1550 vtd_set_clear_mask_long(s, DMAR_FECTL_REG, VTD_FECTL_IP, 0); 1551 VTD_DPRINTF(FLOG, "all pending interrupt conditions serviced, clear " 1552 "IP field of FECTL_REG"); 1553 } 1554 /* FIXME: when IQE is Clear, should we try to fetch some Invalidation 1555 * Descriptors if there are any when Queued Invalidation is enabled? 1556 */ 1557 } 1558 1559 static void vtd_handle_fectl_write(IntelIOMMUState *s) 1560 { 1561 uint32_t fectl_reg; 1562 /* FIXME: when software clears the IM field, check the IP field. But do we 1563 * need to compare the old value and the new value to conclude that 1564 * software clears the IM field? Or just check if the IM field is zero? 1565 */ 1566 fectl_reg = vtd_get_long_raw(s, DMAR_FECTL_REG); 1567 if ((fectl_reg & VTD_FECTL_IP) && !(fectl_reg & VTD_FECTL_IM)) { 1568 vtd_generate_interrupt(s, DMAR_FEADDR_REG, DMAR_FEDATA_REG); 1569 vtd_set_clear_mask_long(s, DMAR_FECTL_REG, VTD_FECTL_IP, 0); 1570 VTD_DPRINTF(FLOG, "IM field is cleared, generate " 1571 "fault event interrupt"); 1572 } 1573 } 1574 1575 static void vtd_handle_ics_write(IntelIOMMUState *s) 1576 { 1577 uint32_t ics_reg = vtd_get_long_raw(s, DMAR_ICS_REG); 1578 uint32_t iectl_reg = vtd_get_long_raw(s, DMAR_IECTL_REG); 1579 1580 if ((iectl_reg & VTD_IECTL_IP) && !(ics_reg & VTD_ICS_IWC)) { 1581 vtd_set_clear_mask_long(s, DMAR_IECTL_REG, VTD_IECTL_IP, 0); 1582 VTD_DPRINTF(INV, "pending completion interrupt condition serviced, " 1583 "clear IP field of IECTL_REG"); 1584 } 1585 } 1586 1587 static void vtd_handle_iectl_write(IntelIOMMUState *s) 1588 { 1589 uint32_t iectl_reg; 1590 /* FIXME: when software clears the IM field, check the IP field. But do we 1591 * need to compare the old value and the new value to conclude that 1592 * software clears the IM field? Or just check if the IM field is zero? 1593 */ 1594 iectl_reg = vtd_get_long_raw(s, DMAR_IECTL_REG); 1595 if ((iectl_reg & VTD_IECTL_IP) && !(iectl_reg & VTD_IECTL_IM)) { 1596 vtd_generate_interrupt(s, DMAR_IEADDR_REG, DMAR_IEDATA_REG); 1597 vtd_set_clear_mask_long(s, DMAR_IECTL_REG, VTD_IECTL_IP, 0); 1598 VTD_DPRINTF(INV, "IM field is cleared, generate " 1599 "invalidation event interrupt"); 1600 } 1601 } 1602 1603 static uint64_t vtd_mem_read(void *opaque, hwaddr addr, unsigned size) 1604 { 1605 IntelIOMMUState *s = opaque; 1606 uint64_t val; 1607 1608 if (addr + size > DMAR_REG_SIZE) { 1609 VTD_DPRINTF(GENERAL, "error: addr outside region: max 0x%"PRIx64 1610 ", got 0x%"PRIx64 " %d", 1611 (uint64_t)DMAR_REG_SIZE, addr, size); 1612 return (uint64_t)-1; 1613 } 1614 1615 switch (addr) { 1616 /* Root Table Address Register, 64-bit */ 1617 case DMAR_RTADDR_REG: 1618 if (size == 4) { 1619 val = s->root & ((1ULL << 32) - 1); 1620 } else { 1621 val = s->root; 1622 } 1623 break; 1624 1625 case DMAR_RTADDR_REG_HI: 1626 assert(size == 4); 1627 val = s->root >> 32; 1628 break; 1629 1630 /* Invalidation Queue Address Register, 64-bit */ 1631 case DMAR_IQA_REG: 1632 val = s->iq | (vtd_get_quad(s, DMAR_IQA_REG) & VTD_IQA_QS); 1633 if (size == 4) { 1634 val = val & ((1ULL << 32) - 1); 1635 } 1636 break; 1637 1638 case DMAR_IQA_REG_HI: 1639 assert(size == 4); 1640 val = s->iq >> 32; 1641 break; 1642 1643 default: 1644 if (size == 4) { 1645 val = vtd_get_long(s, addr); 1646 } else { 1647 val = vtd_get_quad(s, addr); 1648 } 1649 } 1650 VTD_DPRINTF(CSR, "addr 0x%"PRIx64 " size %d val 0x%"PRIx64, 1651 addr, size, val); 1652 return val; 1653 } 1654 1655 static void vtd_mem_write(void *opaque, hwaddr addr, 1656 uint64_t val, unsigned size) 1657 { 1658 IntelIOMMUState *s = opaque; 1659 1660 if (addr + size > DMAR_REG_SIZE) { 1661 VTD_DPRINTF(GENERAL, "error: addr outside region: max 0x%"PRIx64 1662 ", got 0x%"PRIx64 " %d", 1663 (uint64_t)DMAR_REG_SIZE, addr, size); 1664 return; 1665 } 1666 1667 switch (addr) { 1668 /* Global Command Register, 32-bit */ 1669 case DMAR_GCMD_REG: 1670 VTD_DPRINTF(CSR, "DMAR_GCMD_REG write addr 0x%"PRIx64 1671 ", size %d, val 0x%"PRIx64, addr, size, val); 1672 vtd_set_long(s, addr, val); 1673 vtd_handle_gcmd_write(s); 1674 break; 1675 1676 /* Context Command Register, 64-bit */ 1677 case DMAR_CCMD_REG: 1678 VTD_DPRINTF(CSR, "DMAR_CCMD_REG write addr 0x%"PRIx64 1679 ", size %d, val 0x%"PRIx64, addr, size, val); 1680 if (size == 4) { 1681 vtd_set_long(s, addr, val); 1682 } else { 1683 vtd_set_quad(s, addr, val); 1684 vtd_handle_ccmd_write(s); 1685 } 1686 break; 1687 1688 case DMAR_CCMD_REG_HI: 1689 VTD_DPRINTF(CSR, "DMAR_CCMD_REG_HI write addr 0x%"PRIx64 1690 ", size %d, val 0x%"PRIx64, addr, size, val); 1691 assert(size == 4); 1692 vtd_set_long(s, addr, val); 1693 vtd_handle_ccmd_write(s); 1694 break; 1695 1696 /* IOTLB Invalidation Register, 64-bit */ 1697 case DMAR_IOTLB_REG: 1698 VTD_DPRINTF(INV, "DMAR_IOTLB_REG write addr 0x%"PRIx64 1699 ", size %d, val 0x%"PRIx64, addr, size, val); 1700 if (size == 4) { 1701 vtd_set_long(s, addr, val); 1702 } else { 1703 vtd_set_quad(s, addr, val); 1704 vtd_handle_iotlb_write(s); 1705 } 1706 break; 1707 1708 case DMAR_IOTLB_REG_HI: 1709 VTD_DPRINTF(INV, "DMAR_IOTLB_REG_HI write addr 0x%"PRIx64 1710 ", size %d, val 0x%"PRIx64, addr, size, val); 1711 assert(size == 4); 1712 vtd_set_long(s, addr, val); 1713 vtd_handle_iotlb_write(s); 1714 break; 1715 1716 /* Invalidate Address Register, 64-bit */ 1717 case DMAR_IVA_REG: 1718 VTD_DPRINTF(INV, "DMAR_IVA_REG write addr 0x%"PRIx64 1719 ", size %d, val 0x%"PRIx64, addr, size, val); 1720 if (size == 4) { 1721 vtd_set_long(s, addr, val); 1722 } else { 1723 vtd_set_quad(s, addr, val); 1724 } 1725 break; 1726 1727 case DMAR_IVA_REG_HI: 1728 VTD_DPRINTF(INV, "DMAR_IVA_REG_HI write addr 0x%"PRIx64 1729 ", size %d, val 0x%"PRIx64, addr, size, val); 1730 assert(size == 4); 1731 vtd_set_long(s, addr, val); 1732 break; 1733 1734 /* Fault Status Register, 32-bit */ 1735 case DMAR_FSTS_REG: 1736 VTD_DPRINTF(FLOG, "DMAR_FSTS_REG write addr 0x%"PRIx64 1737 ", size %d, val 0x%"PRIx64, addr, size, val); 1738 assert(size == 4); 1739 vtd_set_long(s, addr, val); 1740 vtd_handle_fsts_write(s); 1741 break; 1742 1743 /* Fault Event Control Register, 32-bit */ 1744 case DMAR_FECTL_REG: 1745 VTD_DPRINTF(FLOG, "DMAR_FECTL_REG write addr 0x%"PRIx64 1746 ", size %d, val 0x%"PRIx64, addr, size, val); 1747 assert(size == 4); 1748 vtd_set_long(s, addr, val); 1749 vtd_handle_fectl_write(s); 1750 break; 1751 1752 /* Fault Event Data Register, 32-bit */ 1753 case DMAR_FEDATA_REG: 1754 VTD_DPRINTF(FLOG, "DMAR_FEDATA_REG write addr 0x%"PRIx64 1755 ", size %d, val 0x%"PRIx64, addr, size, val); 1756 assert(size == 4); 1757 vtd_set_long(s, addr, val); 1758 break; 1759 1760 /* Fault Event Address Register, 32-bit */ 1761 case DMAR_FEADDR_REG: 1762 VTD_DPRINTF(FLOG, "DMAR_FEADDR_REG write addr 0x%"PRIx64 1763 ", size %d, val 0x%"PRIx64, addr, size, val); 1764 assert(size == 4); 1765 vtd_set_long(s, addr, val); 1766 break; 1767 1768 /* Fault Event Upper Address Register, 32-bit */ 1769 case DMAR_FEUADDR_REG: 1770 VTD_DPRINTF(FLOG, "DMAR_FEUADDR_REG write addr 0x%"PRIx64 1771 ", size %d, val 0x%"PRIx64, addr, size, val); 1772 assert(size == 4); 1773 vtd_set_long(s, addr, val); 1774 break; 1775 1776 /* Protected Memory Enable Register, 32-bit */ 1777 case DMAR_PMEN_REG: 1778 VTD_DPRINTF(CSR, "DMAR_PMEN_REG write addr 0x%"PRIx64 1779 ", size %d, val 0x%"PRIx64, addr, size, val); 1780 assert(size == 4); 1781 vtd_set_long(s, addr, val); 1782 break; 1783 1784 /* Root Table Address Register, 64-bit */ 1785 case DMAR_RTADDR_REG: 1786 VTD_DPRINTF(CSR, "DMAR_RTADDR_REG write addr 0x%"PRIx64 1787 ", size %d, val 0x%"PRIx64, addr, size, val); 1788 if (size == 4) { 1789 vtd_set_long(s, addr, val); 1790 } else { 1791 vtd_set_quad(s, addr, val); 1792 } 1793 break; 1794 1795 case DMAR_RTADDR_REG_HI: 1796 VTD_DPRINTF(CSR, "DMAR_RTADDR_REG_HI write addr 0x%"PRIx64 1797 ", size %d, val 0x%"PRIx64, addr, size, val); 1798 assert(size == 4); 1799 vtd_set_long(s, addr, val); 1800 break; 1801 1802 /* Invalidation Queue Tail Register, 64-bit */ 1803 case DMAR_IQT_REG: 1804 VTD_DPRINTF(INV, "DMAR_IQT_REG write addr 0x%"PRIx64 1805 ", size %d, val 0x%"PRIx64, addr, size, val); 1806 if (size == 4) { 1807 vtd_set_long(s, addr, val); 1808 } else { 1809 vtd_set_quad(s, addr, val); 1810 } 1811 vtd_handle_iqt_write(s); 1812 break; 1813 1814 case DMAR_IQT_REG_HI: 1815 VTD_DPRINTF(INV, "DMAR_IQT_REG_HI write addr 0x%"PRIx64 1816 ", size %d, val 0x%"PRIx64, addr, size, val); 1817 assert(size == 4); 1818 vtd_set_long(s, addr, val); 1819 /* 19:63 of IQT_REG is RsvdZ, do nothing here */ 1820 break; 1821 1822 /* Invalidation Queue Address Register, 64-bit */ 1823 case DMAR_IQA_REG: 1824 VTD_DPRINTF(INV, "DMAR_IQA_REG write addr 0x%"PRIx64 1825 ", size %d, val 0x%"PRIx64, addr, size, val); 1826 if (size == 4) { 1827 vtd_set_long(s, addr, val); 1828 } else { 1829 vtd_set_quad(s, addr, val); 1830 } 1831 break; 1832 1833 case DMAR_IQA_REG_HI: 1834 VTD_DPRINTF(INV, "DMAR_IQA_REG_HI write addr 0x%"PRIx64 1835 ", size %d, val 0x%"PRIx64, addr, size, val); 1836 assert(size == 4); 1837 vtd_set_long(s, addr, val); 1838 break; 1839 1840 /* Invalidation Completion Status Register, 32-bit */ 1841 case DMAR_ICS_REG: 1842 VTD_DPRINTF(INV, "DMAR_ICS_REG write addr 0x%"PRIx64 1843 ", size %d, val 0x%"PRIx64, addr, size, val); 1844 assert(size == 4); 1845 vtd_set_long(s, addr, val); 1846 vtd_handle_ics_write(s); 1847 break; 1848 1849 /* Invalidation Event Control Register, 32-bit */ 1850 case DMAR_IECTL_REG: 1851 VTD_DPRINTF(INV, "DMAR_IECTL_REG write addr 0x%"PRIx64 1852 ", size %d, val 0x%"PRIx64, addr, size, val); 1853 assert(size == 4); 1854 vtd_set_long(s, addr, val); 1855 vtd_handle_iectl_write(s); 1856 break; 1857 1858 /* Invalidation Event Data Register, 32-bit */ 1859 case DMAR_IEDATA_REG: 1860 VTD_DPRINTF(INV, "DMAR_IEDATA_REG write addr 0x%"PRIx64 1861 ", size %d, val 0x%"PRIx64, addr, size, val); 1862 assert(size == 4); 1863 vtd_set_long(s, addr, val); 1864 break; 1865 1866 /* Invalidation Event Address Register, 32-bit */ 1867 case DMAR_IEADDR_REG: 1868 VTD_DPRINTF(INV, "DMAR_IEADDR_REG write addr 0x%"PRIx64 1869 ", size %d, val 0x%"PRIx64, addr, size, val); 1870 assert(size == 4); 1871 vtd_set_long(s, addr, val); 1872 break; 1873 1874 /* Invalidation Event Upper Address Register, 32-bit */ 1875 case DMAR_IEUADDR_REG: 1876 VTD_DPRINTF(INV, "DMAR_IEUADDR_REG write addr 0x%"PRIx64 1877 ", size %d, val 0x%"PRIx64, addr, size, val); 1878 assert(size == 4); 1879 vtd_set_long(s, addr, val); 1880 break; 1881 1882 /* Fault Recording Registers, 128-bit */ 1883 case DMAR_FRCD_REG_0_0: 1884 VTD_DPRINTF(FLOG, "DMAR_FRCD_REG_0_0 write addr 0x%"PRIx64 1885 ", size %d, val 0x%"PRIx64, addr, size, val); 1886 if (size == 4) { 1887 vtd_set_long(s, addr, val); 1888 } else { 1889 vtd_set_quad(s, addr, val); 1890 } 1891 break; 1892 1893 case DMAR_FRCD_REG_0_1: 1894 VTD_DPRINTF(FLOG, "DMAR_FRCD_REG_0_1 write addr 0x%"PRIx64 1895 ", size %d, val 0x%"PRIx64, addr, size, val); 1896 assert(size == 4); 1897 vtd_set_long(s, addr, val); 1898 break; 1899 1900 case DMAR_FRCD_REG_0_2: 1901 VTD_DPRINTF(FLOG, "DMAR_FRCD_REG_0_2 write addr 0x%"PRIx64 1902 ", size %d, val 0x%"PRIx64, addr, size, val); 1903 if (size == 4) { 1904 vtd_set_long(s, addr, val); 1905 } else { 1906 vtd_set_quad(s, addr, val); 1907 /* May clear bit 127 (Fault), update PPF */ 1908 vtd_update_fsts_ppf(s); 1909 } 1910 break; 1911 1912 case DMAR_FRCD_REG_0_3: 1913 VTD_DPRINTF(FLOG, "DMAR_FRCD_REG_0_3 write addr 0x%"PRIx64 1914 ", size %d, val 0x%"PRIx64, addr, size, val); 1915 assert(size == 4); 1916 vtd_set_long(s, addr, val); 1917 /* May clear bit 127 (Fault), update PPF */ 1918 vtd_update_fsts_ppf(s); 1919 break; 1920 1921 case DMAR_IRTA_REG: 1922 VTD_DPRINTF(IR, "DMAR_IRTA_REG write addr 0x%"PRIx64 1923 ", size %d, val 0x%"PRIx64, addr, size, val); 1924 if (size == 4) { 1925 vtd_set_long(s, addr, val); 1926 } else { 1927 vtd_set_quad(s, addr, val); 1928 } 1929 break; 1930 1931 case DMAR_IRTA_REG_HI: 1932 VTD_DPRINTF(IR, "DMAR_IRTA_REG_HI write addr 0x%"PRIx64 1933 ", size %d, val 0x%"PRIx64, addr, size, val); 1934 assert(size == 4); 1935 vtd_set_long(s, addr, val); 1936 break; 1937 1938 default: 1939 VTD_DPRINTF(GENERAL, "error: unhandled reg write addr 0x%"PRIx64 1940 ", size %d, val 0x%"PRIx64, addr, size, val); 1941 if (size == 4) { 1942 vtd_set_long(s, addr, val); 1943 } else { 1944 vtd_set_quad(s, addr, val); 1945 } 1946 } 1947 } 1948 1949 static IOMMUTLBEntry vtd_iommu_translate(MemoryRegion *iommu, hwaddr addr, 1950 bool is_write) 1951 { 1952 VTDAddressSpace *vtd_as = container_of(iommu, VTDAddressSpace, iommu); 1953 IntelIOMMUState *s = vtd_as->iommu_state; 1954 IOMMUTLBEntry ret = { 1955 .target_as = &address_space_memory, 1956 .iova = addr, 1957 .translated_addr = 0, 1958 .addr_mask = ~(hwaddr)0, 1959 .perm = IOMMU_NONE, 1960 }; 1961 1962 if (!s->dmar_enabled) { 1963 /* DMAR disabled, passthrough, use 4k-page*/ 1964 ret.iova = addr & VTD_PAGE_MASK_4K; 1965 ret.translated_addr = addr & VTD_PAGE_MASK_4K; 1966 ret.addr_mask = ~VTD_PAGE_MASK_4K; 1967 ret.perm = IOMMU_RW; 1968 return ret; 1969 } 1970 1971 vtd_do_iommu_translate(vtd_as, vtd_as->bus, vtd_as->devfn, addr, 1972 is_write, &ret); 1973 VTD_DPRINTF(MMU, 1974 "bus %"PRIu8 " slot %"PRIu8 " func %"PRIu8 " devfn %"PRIu8 1975 " gpa 0x%"PRIx64 " hpa 0x%"PRIx64, pci_bus_num(vtd_as->bus), 1976 VTD_PCI_SLOT(vtd_as->devfn), VTD_PCI_FUNC(vtd_as->devfn), 1977 vtd_as->devfn, addr, ret.translated_addr); 1978 return ret; 1979 } 1980 1981 static void vtd_iommu_notify_flag_changed(MemoryRegion *iommu, 1982 IOMMUNotifierFlag old, 1983 IOMMUNotifierFlag new) 1984 { 1985 VTDAddressSpace *vtd_as = container_of(iommu, VTDAddressSpace, iommu); 1986 1987 if (new & IOMMU_NOTIFIER_MAP) { 1988 error_report("Device at bus %s addr %02x.%d requires iommu " 1989 "notifier which is currently not supported by " 1990 "intel-iommu emulation", 1991 vtd_as->bus->qbus.name, PCI_SLOT(vtd_as->devfn), 1992 PCI_FUNC(vtd_as->devfn)); 1993 exit(1); 1994 } 1995 } 1996 1997 static const VMStateDescription vtd_vmstate = { 1998 .name = "iommu-intel", 1999 .unmigratable = 1, 2000 }; 2001 2002 static const MemoryRegionOps vtd_mem_ops = { 2003 .read = vtd_mem_read, 2004 .write = vtd_mem_write, 2005 .endianness = DEVICE_LITTLE_ENDIAN, 2006 .impl = { 2007 .min_access_size = 4, 2008 .max_access_size = 8, 2009 }, 2010 .valid = { 2011 .min_access_size = 4, 2012 .max_access_size = 8, 2013 }, 2014 }; 2015 2016 static Property vtd_properties[] = { 2017 DEFINE_PROP_UINT32("version", IntelIOMMUState, version, 0), 2018 DEFINE_PROP_ON_OFF_AUTO("eim", IntelIOMMUState, intr_eim, 2019 ON_OFF_AUTO_AUTO), 2020 DEFINE_PROP_BOOL("x-buggy-eim", IntelIOMMUState, buggy_eim, false), 2021 DEFINE_PROP_END_OF_LIST(), 2022 }; 2023 2024 /* Read IRTE entry with specific index */ 2025 static int vtd_irte_get(IntelIOMMUState *iommu, uint16_t index, 2026 VTD_IR_TableEntry *entry, uint16_t sid) 2027 { 2028 static const uint16_t vtd_svt_mask[VTD_SQ_MAX] = \ 2029 {0xffff, 0xfffb, 0xfff9, 0xfff8}; 2030 dma_addr_t addr = 0x00; 2031 uint16_t mask, source_id; 2032 uint8_t bus, bus_max, bus_min; 2033 2034 addr = iommu->intr_root + index * sizeof(*entry); 2035 if (dma_memory_read(&address_space_memory, addr, entry, 2036 sizeof(*entry))) { 2037 VTD_DPRINTF(GENERAL, "error: fail to access IR root at 0x%"PRIx64 2038 " + %"PRIu16, iommu->intr_root, index); 2039 return -VTD_FR_IR_ROOT_INVAL; 2040 } 2041 2042 if (!entry->irte.present) { 2043 VTD_DPRINTF(GENERAL, "error: present flag not set in IRTE" 2044 " entry index %u value 0x%"PRIx64 " 0x%"PRIx64, 2045 index, le64_to_cpu(entry->data[1]), 2046 le64_to_cpu(entry->data[0])); 2047 return -VTD_FR_IR_ENTRY_P; 2048 } 2049 2050 if (entry->irte.__reserved_0 || entry->irte.__reserved_1 || 2051 entry->irte.__reserved_2) { 2052 VTD_DPRINTF(GENERAL, "error: IRTE entry index %"PRIu16 2053 " reserved fields non-zero: 0x%"PRIx64 " 0x%"PRIx64, 2054 index, le64_to_cpu(entry->data[1]), 2055 le64_to_cpu(entry->data[0])); 2056 return -VTD_FR_IR_IRTE_RSVD; 2057 } 2058 2059 if (sid != X86_IOMMU_SID_INVALID) { 2060 /* Validate IRTE SID */ 2061 source_id = le32_to_cpu(entry->irte.source_id); 2062 switch (entry->irte.sid_vtype) { 2063 case VTD_SVT_NONE: 2064 VTD_DPRINTF(IR, "No SID validation for IRTE index %d", index); 2065 break; 2066 2067 case VTD_SVT_ALL: 2068 mask = vtd_svt_mask[entry->irte.sid_q]; 2069 if ((source_id & mask) != (sid & mask)) { 2070 VTD_DPRINTF(GENERAL, "SID validation for IRTE index " 2071 "%d failed (reqid 0x%04x sid 0x%04x)", index, 2072 sid, source_id); 2073 return -VTD_FR_IR_SID_ERR; 2074 } 2075 break; 2076 2077 case VTD_SVT_BUS: 2078 bus_max = source_id >> 8; 2079 bus_min = source_id & 0xff; 2080 bus = sid >> 8; 2081 if (bus > bus_max || bus < bus_min) { 2082 VTD_DPRINTF(GENERAL, "SID validation for IRTE index %d " 2083 "failed (bus %d outside %d-%d)", index, bus, 2084 bus_min, bus_max); 2085 return -VTD_FR_IR_SID_ERR; 2086 } 2087 break; 2088 2089 default: 2090 VTD_DPRINTF(GENERAL, "Invalid SVT bits (0x%x) in IRTE index " 2091 "%d", entry->irte.sid_vtype, index); 2092 /* Take this as verification failure. */ 2093 return -VTD_FR_IR_SID_ERR; 2094 break; 2095 } 2096 } 2097 2098 return 0; 2099 } 2100 2101 /* Fetch IRQ information of specific IR index */ 2102 static int vtd_remap_irq_get(IntelIOMMUState *iommu, uint16_t index, 2103 VTDIrq *irq, uint16_t sid) 2104 { 2105 VTD_IR_TableEntry irte = {}; 2106 int ret = 0; 2107 2108 ret = vtd_irte_get(iommu, index, &irte, sid); 2109 if (ret) { 2110 return ret; 2111 } 2112 2113 irq->trigger_mode = irte.irte.trigger_mode; 2114 irq->vector = irte.irte.vector; 2115 irq->delivery_mode = irte.irte.delivery_mode; 2116 irq->dest = le32_to_cpu(irte.irte.dest_id); 2117 if (!iommu->intr_eime) { 2118 #define VTD_IR_APIC_DEST_MASK (0xff00ULL) 2119 #define VTD_IR_APIC_DEST_SHIFT (8) 2120 irq->dest = (irq->dest & VTD_IR_APIC_DEST_MASK) >> 2121 VTD_IR_APIC_DEST_SHIFT; 2122 } 2123 irq->dest_mode = irte.irte.dest_mode; 2124 irq->redir_hint = irte.irte.redir_hint; 2125 2126 VTD_DPRINTF(IR, "remapping interrupt index %d: trig:%u,vec:%u," 2127 "deliver:%u,dest:%u,dest_mode:%u", index, 2128 irq->trigger_mode, irq->vector, irq->delivery_mode, 2129 irq->dest, irq->dest_mode); 2130 2131 return 0; 2132 } 2133 2134 /* Generate one MSI message from VTDIrq info */ 2135 static void vtd_generate_msi_message(VTDIrq *irq, MSIMessage *msg_out) 2136 { 2137 VTD_MSIMessage msg = {}; 2138 2139 /* Generate address bits */ 2140 msg.dest_mode = irq->dest_mode; 2141 msg.redir_hint = irq->redir_hint; 2142 msg.dest = irq->dest; 2143 msg.__addr_hi = irq->dest & 0xffffff00; 2144 msg.__addr_head = cpu_to_le32(0xfee); 2145 /* Keep this from original MSI address bits */ 2146 msg.__not_used = irq->msi_addr_last_bits; 2147 2148 /* Generate data bits */ 2149 msg.vector = irq->vector; 2150 msg.delivery_mode = irq->delivery_mode; 2151 msg.level = 1; 2152 msg.trigger_mode = irq->trigger_mode; 2153 2154 msg_out->address = msg.msi_addr; 2155 msg_out->data = msg.msi_data; 2156 } 2157 2158 /* Interrupt remapping for MSI/MSI-X entry */ 2159 static int vtd_interrupt_remap_msi(IntelIOMMUState *iommu, 2160 MSIMessage *origin, 2161 MSIMessage *translated, 2162 uint16_t sid) 2163 { 2164 int ret = 0; 2165 VTD_IR_MSIAddress addr; 2166 uint16_t index; 2167 VTDIrq irq = {}; 2168 2169 assert(origin && translated); 2170 2171 if (!iommu || !iommu->intr_enabled) { 2172 goto do_not_translate; 2173 } 2174 2175 if (origin->address & VTD_MSI_ADDR_HI_MASK) { 2176 VTD_DPRINTF(GENERAL, "error: MSI addr high 32 bits nonzero" 2177 " during interrupt remapping: 0x%"PRIx32, 2178 (uint32_t)((origin->address & VTD_MSI_ADDR_HI_MASK) >> \ 2179 VTD_MSI_ADDR_HI_SHIFT)); 2180 return -VTD_FR_IR_REQ_RSVD; 2181 } 2182 2183 addr.data = origin->address & VTD_MSI_ADDR_LO_MASK; 2184 if (addr.addr.__head != 0xfee) { 2185 VTD_DPRINTF(GENERAL, "error: MSI addr low 32 bits invalid: " 2186 "0x%"PRIx32, addr.data); 2187 return -VTD_FR_IR_REQ_RSVD; 2188 } 2189 2190 /* This is compatible mode. */ 2191 if (addr.addr.int_mode != VTD_IR_INT_FORMAT_REMAP) { 2192 goto do_not_translate; 2193 } 2194 2195 index = addr.addr.index_h << 15 | le16_to_cpu(addr.addr.index_l); 2196 2197 #define VTD_IR_MSI_DATA_SUBHANDLE (0x0000ffff) 2198 #define VTD_IR_MSI_DATA_RESERVED (0xffff0000) 2199 2200 if (addr.addr.sub_valid) { 2201 /* See VT-d spec 5.1.2.2 and 5.1.3 on subhandle */ 2202 index += origin->data & VTD_IR_MSI_DATA_SUBHANDLE; 2203 } 2204 2205 ret = vtd_remap_irq_get(iommu, index, &irq, sid); 2206 if (ret) { 2207 return ret; 2208 } 2209 2210 if (addr.addr.sub_valid) { 2211 VTD_DPRINTF(IR, "received MSI interrupt"); 2212 if (origin->data & VTD_IR_MSI_DATA_RESERVED) { 2213 VTD_DPRINTF(GENERAL, "error: MSI data bits non-zero for " 2214 "interrupt remappable entry: 0x%"PRIx32, 2215 origin->data); 2216 return -VTD_FR_IR_REQ_RSVD; 2217 } 2218 } else { 2219 uint8_t vector = origin->data & 0xff; 2220 uint8_t trigger_mode = (origin->data >> MSI_DATA_TRIGGER_SHIFT) & 0x1; 2221 2222 VTD_DPRINTF(IR, "received IOAPIC interrupt"); 2223 /* IOAPIC entry vector should be aligned with IRTE vector 2224 * (see vt-d spec 5.1.5.1). */ 2225 if (vector != irq.vector) { 2226 VTD_DPRINTF(GENERAL, "IOAPIC vector inconsistent: " 2227 "entry: %d, IRTE: %d, index: %d", 2228 vector, irq.vector, index); 2229 } 2230 2231 /* The Trigger Mode field must match the Trigger Mode in the IRTE. 2232 * (see vt-d spec 5.1.5.1). */ 2233 if (trigger_mode != irq.trigger_mode) { 2234 VTD_DPRINTF(GENERAL, "IOAPIC trigger mode inconsistent: " 2235 "entry: %u, IRTE: %u, index: %d", 2236 trigger_mode, irq.trigger_mode, index); 2237 } 2238 2239 } 2240 2241 /* 2242 * We'd better keep the last two bits, assuming that guest OS 2243 * might modify it. Keep it does not hurt after all. 2244 */ 2245 irq.msi_addr_last_bits = addr.addr.__not_care; 2246 2247 /* Translate VTDIrq to MSI message */ 2248 vtd_generate_msi_message(&irq, translated); 2249 2250 VTD_DPRINTF(IR, "mapping MSI 0x%"PRIx64":0x%"PRIx32 " -> " 2251 "0x%"PRIx64":0x%"PRIx32, origin->address, origin->data, 2252 translated->address, translated->data); 2253 return 0; 2254 2255 do_not_translate: 2256 memcpy(translated, origin, sizeof(*origin)); 2257 return 0; 2258 } 2259 2260 static int vtd_int_remap(X86IOMMUState *iommu, MSIMessage *src, 2261 MSIMessage *dst, uint16_t sid) 2262 { 2263 return vtd_interrupt_remap_msi(INTEL_IOMMU_DEVICE(iommu), 2264 src, dst, sid); 2265 } 2266 2267 static MemTxResult vtd_mem_ir_read(void *opaque, hwaddr addr, 2268 uint64_t *data, unsigned size, 2269 MemTxAttrs attrs) 2270 { 2271 return MEMTX_OK; 2272 } 2273 2274 static MemTxResult vtd_mem_ir_write(void *opaque, hwaddr addr, 2275 uint64_t value, unsigned size, 2276 MemTxAttrs attrs) 2277 { 2278 int ret = 0; 2279 MSIMessage from = {}, to = {}; 2280 uint16_t sid = X86_IOMMU_SID_INVALID; 2281 2282 from.address = (uint64_t) addr + VTD_INTERRUPT_ADDR_FIRST; 2283 from.data = (uint32_t) value; 2284 2285 if (!attrs.unspecified) { 2286 /* We have explicit Source ID */ 2287 sid = attrs.requester_id; 2288 } 2289 2290 ret = vtd_interrupt_remap_msi(opaque, &from, &to, sid); 2291 if (ret) { 2292 /* TODO: report error */ 2293 VTD_DPRINTF(GENERAL, "int remap fail for addr 0x%"PRIx64 2294 " data 0x%"PRIx32, from.address, from.data); 2295 /* Drop this interrupt */ 2296 return MEMTX_ERROR; 2297 } 2298 2299 VTD_DPRINTF(IR, "delivering MSI 0x%"PRIx64":0x%"PRIx32 2300 " for device sid 0x%04x", 2301 to.address, to.data, sid); 2302 2303 apic_get_class()->send_msi(&to); 2304 2305 return MEMTX_OK; 2306 } 2307 2308 static const MemoryRegionOps vtd_mem_ir_ops = { 2309 .read_with_attrs = vtd_mem_ir_read, 2310 .write_with_attrs = vtd_mem_ir_write, 2311 .endianness = DEVICE_LITTLE_ENDIAN, 2312 .impl = { 2313 .min_access_size = 4, 2314 .max_access_size = 4, 2315 }, 2316 .valid = { 2317 .min_access_size = 4, 2318 .max_access_size = 4, 2319 }, 2320 }; 2321 2322 VTDAddressSpace *vtd_find_add_as(IntelIOMMUState *s, PCIBus *bus, int devfn) 2323 { 2324 uintptr_t key = (uintptr_t)bus; 2325 VTDBus *vtd_bus = g_hash_table_lookup(s->vtd_as_by_busptr, &key); 2326 VTDAddressSpace *vtd_dev_as; 2327 2328 if (!vtd_bus) { 2329 /* No corresponding free() */ 2330 vtd_bus = g_malloc0(sizeof(VTDBus) + sizeof(VTDAddressSpace *) * \ 2331 X86_IOMMU_PCI_DEVFN_MAX); 2332 vtd_bus->bus = bus; 2333 key = (uintptr_t)bus; 2334 g_hash_table_insert(s->vtd_as_by_busptr, &key, vtd_bus); 2335 } 2336 2337 vtd_dev_as = vtd_bus->dev_as[devfn]; 2338 2339 if (!vtd_dev_as) { 2340 vtd_bus->dev_as[devfn] = vtd_dev_as = g_malloc0(sizeof(VTDAddressSpace)); 2341 2342 vtd_dev_as->bus = bus; 2343 vtd_dev_as->devfn = (uint8_t)devfn; 2344 vtd_dev_as->iommu_state = s; 2345 vtd_dev_as->context_cache_entry.context_cache_gen = 0; 2346 memory_region_init_iommu(&vtd_dev_as->iommu, OBJECT(s), 2347 &s->iommu_ops, "intel_iommu", UINT64_MAX); 2348 memory_region_init_io(&vtd_dev_as->iommu_ir, OBJECT(s), 2349 &vtd_mem_ir_ops, s, "intel_iommu_ir", 2350 VTD_INTERRUPT_ADDR_SIZE); 2351 memory_region_add_subregion(&vtd_dev_as->iommu, VTD_INTERRUPT_ADDR_FIRST, 2352 &vtd_dev_as->iommu_ir); 2353 address_space_init(&vtd_dev_as->as, 2354 &vtd_dev_as->iommu, "intel_iommu"); 2355 } 2356 return vtd_dev_as; 2357 } 2358 2359 /* Do the initialization. It will also be called when reset, so pay 2360 * attention when adding new initialization stuff. 2361 */ 2362 static void vtd_init(IntelIOMMUState *s) 2363 { 2364 X86IOMMUState *x86_iommu = X86_IOMMU_DEVICE(s); 2365 2366 memset(s->csr, 0, DMAR_REG_SIZE); 2367 memset(s->wmask, 0, DMAR_REG_SIZE); 2368 memset(s->w1cmask, 0, DMAR_REG_SIZE); 2369 memset(s->womask, 0, DMAR_REG_SIZE); 2370 2371 s->iommu_ops.translate = vtd_iommu_translate; 2372 s->iommu_ops.notify_flag_changed = vtd_iommu_notify_flag_changed; 2373 s->root = 0; 2374 s->root_extended = false; 2375 s->dmar_enabled = false; 2376 s->iq_head = 0; 2377 s->iq_tail = 0; 2378 s->iq = 0; 2379 s->iq_size = 0; 2380 s->qi_enabled = false; 2381 s->iq_last_desc_type = VTD_INV_DESC_NONE; 2382 s->next_frcd_reg = 0; 2383 s->cap = VTD_CAP_FRO | VTD_CAP_NFR | VTD_CAP_ND | VTD_CAP_MGAW | 2384 VTD_CAP_SAGAW | VTD_CAP_MAMV | VTD_CAP_PSI | VTD_CAP_SLLPS; 2385 s->ecap = VTD_ECAP_QI | VTD_ECAP_IRO; 2386 2387 if (x86_iommu->intr_supported) { 2388 s->ecap |= VTD_ECAP_IR | VTD_ECAP_MHMV; 2389 if (s->intr_eim == ON_OFF_AUTO_ON) { 2390 s->ecap |= VTD_ECAP_EIM; 2391 } 2392 assert(s->intr_eim != ON_OFF_AUTO_AUTO); 2393 } 2394 2395 vtd_reset_context_cache(s); 2396 vtd_reset_iotlb(s); 2397 2398 /* Define registers with default values and bit semantics */ 2399 vtd_define_long(s, DMAR_VER_REG, 0x10UL, 0, 0); 2400 vtd_define_quad(s, DMAR_CAP_REG, s->cap, 0, 0); 2401 vtd_define_quad(s, DMAR_ECAP_REG, s->ecap, 0, 0); 2402 vtd_define_long(s, DMAR_GCMD_REG, 0, 0xff800000UL, 0); 2403 vtd_define_long_wo(s, DMAR_GCMD_REG, 0xff800000UL); 2404 vtd_define_long(s, DMAR_GSTS_REG, 0, 0, 0); 2405 vtd_define_quad(s, DMAR_RTADDR_REG, 0, 0xfffffffffffff000ULL, 0); 2406 vtd_define_quad(s, DMAR_CCMD_REG, 0, 0xe0000003ffffffffULL, 0); 2407 vtd_define_quad_wo(s, DMAR_CCMD_REG, 0x3ffff0000ULL); 2408 2409 /* Advanced Fault Logging not supported */ 2410 vtd_define_long(s, DMAR_FSTS_REG, 0, 0, 0x11UL); 2411 vtd_define_long(s, DMAR_FECTL_REG, 0x80000000UL, 0x80000000UL, 0); 2412 vtd_define_long(s, DMAR_FEDATA_REG, 0, 0x0000ffffUL, 0); 2413 vtd_define_long(s, DMAR_FEADDR_REG, 0, 0xfffffffcUL, 0); 2414 2415 /* Treated as RsvdZ when EIM in ECAP_REG is not supported 2416 * vtd_define_long(s, DMAR_FEUADDR_REG, 0, 0xffffffffUL, 0); 2417 */ 2418 vtd_define_long(s, DMAR_FEUADDR_REG, 0, 0, 0); 2419 2420 /* Treated as RO for implementations that PLMR and PHMR fields reported 2421 * as Clear in the CAP_REG. 2422 * vtd_define_long(s, DMAR_PMEN_REG, 0, 0x80000000UL, 0); 2423 */ 2424 vtd_define_long(s, DMAR_PMEN_REG, 0, 0, 0); 2425 2426 vtd_define_quad(s, DMAR_IQH_REG, 0, 0, 0); 2427 vtd_define_quad(s, DMAR_IQT_REG, 0, 0x7fff0ULL, 0); 2428 vtd_define_quad(s, DMAR_IQA_REG, 0, 0xfffffffffffff007ULL, 0); 2429 vtd_define_long(s, DMAR_ICS_REG, 0, 0, 0x1UL); 2430 vtd_define_long(s, DMAR_IECTL_REG, 0x80000000UL, 0x80000000UL, 0); 2431 vtd_define_long(s, DMAR_IEDATA_REG, 0, 0xffffffffUL, 0); 2432 vtd_define_long(s, DMAR_IEADDR_REG, 0, 0xfffffffcUL, 0); 2433 /* Treadted as RsvdZ when EIM in ECAP_REG is not supported */ 2434 vtd_define_long(s, DMAR_IEUADDR_REG, 0, 0, 0); 2435 2436 /* IOTLB registers */ 2437 vtd_define_quad(s, DMAR_IOTLB_REG, 0, 0Xb003ffff00000000ULL, 0); 2438 vtd_define_quad(s, DMAR_IVA_REG, 0, 0xfffffffffffff07fULL, 0); 2439 vtd_define_quad_wo(s, DMAR_IVA_REG, 0xfffffffffffff07fULL); 2440 2441 /* Fault Recording Registers, 128-bit */ 2442 vtd_define_quad(s, DMAR_FRCD_REG_0_0, 0, 0, 0); 2443 vtd_define_quad(s, DMAR_FRCD_REG_0_2, 0, 0, 0x8000000000000000ULL); 2444 2445 /* 2446 * Interrupt remapping registers. 2447 */ 2448 vtd_define_quad(s, DMAR_IRTA_REG, 0, 0xfffffffffffff80fULL, 0); 2449 } 2450 2451 /* Should not reset address_spaces when reset because devices will still use 2452 * the address space they got at first (won't ask the bus again). 2453 */ 2454 static void vtd_reset(DeviceState *dev) 2455 { 2456 IntelIOMMUState *s = INTEL_IOMMU_DEVICE(dev); 2457 2458 VTD_DPRINTF(GENERAL, ""); 2459 vtd_init(s); 2460 } 2461 2462 static AddressSpace *vtd_host_dma_iommu(PCIBus *bus, void *opaque, int devfn) 2463 { 2464 IntelIOMMUState *s = opaque; 2465 VTDAddressSpace *vtd_as; 2466 2467 assert(0 <= devfn && devfn < X86_IOMMU_PCI_DEVFN_MAX); 2468 2469 vtd_as = vtd_find_add_as(s, bus, devfn); 2470 return &vtd_as->as; 2471 } 2472 2473 static bool vtd_decide_config(IntelIOMMUState *s, Error **errp) 2474 { 2475 X86IOMMUState *x86_iommu = X86_IOMMU_DEVICE(s); 2476 2477 /* Currently Intel IOMMU IR only support "kernel-irqchip={off|split}" */ 2478 if (x86_iommu->intr_supported && kvm_irqchip_in_kernel() && 2479 !kvm_irqchip_is_split()) { 2480 error_setg(errp, "Intel Interrupt Remapping cannot work with " 2481 "kernel-irqchip=on, please use 'split|off'."); 2482 return false; 2483 } 2484 if (s->intr_eim == ON_OFF_AUTO_ON && !x86_iommu->intr_supported) { 2485 error_setg(errp, "eim=on cannot be selected without intremap=on"); 2486 return false; 2487 } 2488 2489 if (s->intr_eim == ON_OFF_AUTO_AUTO) { 2490 s->intr_eim = (kvm_irqchip_in_kernel() || s->buggy_eim) 2491 && x86_iommu->intr_supported ? 2492 ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF; 2493 } 2494 if (s->intr_eim == ON_OFF_AUTO_ON && !s->buggy_eim) { 2495 if (!kvm_irqchip_in_kernel()) { 2496 error_setg(errp, "eim=on requires accel=kvm,kernel-irqchip=split"); 2497 return false; 2498 } 2499 if (!kvm_enable_x2apic()) { 2500 error_setg(errp, "eim=on requires support on the KVM side" 2501 "(X2APIC_API, first shipped in v4.7)"); 2502 return false; 2503 } 2504 } 2505 2506 return true; 2507 } 2508 2509 static void vtd_realize(DeviceState *dev, Error **errp) 2510 { 2511 PCMachineState *pcms = PC_MACHINE(qdev_get_machine()); 2512 PCIBus *bus = pcms->bus; 2513 IntelIOMMUState *s = INTEL_IOMMU_DEVICE(dev); 2514 X86IOMMUState *x86_iommu = X86_IOMMU_DEVICE(dev); 2515 2516 VTD_DPRINTF(GENERAL, ""); 2517 x86_iommu->type = TYPE_INTEL; 2518 2519 if (!vtd_decide_config(s, errp)) { 2520 return; 2521 } 2522 2523 memset(s->vtd_as_by_bus_num, 0, sizeof(s->vtd_as_by_bus_num)); 2524 memory_region_init_io(&s->csrmem, OBJECT(s), &vtd_mem_ops, s, 2525 "intel_iommu", DMAR_REG_SIZE); 2526 sysbus_init_mmio(SYS_BUS_DEVICE(s), &s->csrmem); 2527 /* No corresponding destroy */ 2528 s->iotlb = g_hash_table_new_full(vtd_uint64_hash, vtd_uint64_equal, 2529 g_free, g_free); 2530 s->vtd_as_by_busptr = g_hash_table_new_full(vtd_uint64_hash, vtd_uint64_equal, 2531 g_free, g_free); 2532 vtd_init(s); 2533 sysbus_mmio_map(SYS_BUS_DEVICE(s), 0, Q35_HOST_BRIDGE_IOMMU_ADDR); 2534 pci_setup_iommu(bus, vtd_host_dma_iommu, dev); 2535 /* Pseudo address space under root PCI bus. */ 2536 pcms->ioapic_as = vtd_host_dma_iommu(bus, s, Q35_PSEUDO_DEVFN_IOAPIC); 2537 } 2538 2539 static void vtd_class_init(ObjectClass *klass, void *data) 2540 { 2541 DeviceClass *dc = DEVICE_CLASS(klass); 2542 X86IOMMUClass *x86_class = X86_IOMMU_CLASS(klass); 2543 2544 dc->reset = vtd_reset; 2545 dc->vmsd = &vtd_vmstate; 2546 dc->props = vtd_properties; 2547 dc->hotpluggable = false; 2548 x86_class->realize = vtd_realize; 2549 x86_class->int_remap = vtd_int_remap; 2550 } 2551 2552 static const TypeInfo vtd_info = { 2553 .name = TYPE_INTEL_IOMMU_DEVICE, 2554 .parent = TYPE_X86_IOMMU_DEVICE, 2555 .instance_size = sizeof(IntelIOMMUState), 2556 .class_init = vtd_class_init, 2557 }; 2558 2559 static void vtd_register_types(void) 2560 { 2561 VTD_DPRINTF(GENERAL, ""); 2562 type_register_static(&vtd_info); 2563 } 2564 2565 type_init(vtd_register_types) 2566