1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Portions Copyright (c) 2010, Oracle and/or its affiliates. 23 * All rights reserved. 24 */ 25 /* 26 * Copyright (c) 2009, Intel Corporation. 27 * All rights reserved. 28 */ 29 30 /* 31 * DVMA code 32 * This file contains Intel IOMMU code that deals with DVMA 33 * i.e. DMA remapping. 34 */ 35 36 #include <sys/sysmacros.h> 37 #include <sys/pcie.h> 38 #include <sys/pci_cfgspace.h> 39 #include <vm/hat_i86.h> 40 #include <sys/memlist.h> 41 #include <sys/acpi/acpi.h> 42 #include <sys/acpica.h> 43 #include <sys/modhash.h> 44 #include <sys/immu.h> 45 46 #undef TEST 47 48 /* 49 * Macros based on PCI spec 50 */ 51 #define IMMU_PCI_REV2CLASS(r) ((r) >> 8) /* classcode from revid */ 52 #define IMMU_PCI_CLASS2BASE(c) ((c) >> 16) /* baseclass from classcode */ 53 #define IMMU_PCI_CLASS2SUB(c) (((c) >> 8) & 0xff); /* classcode */ 54 55 #define IMMU_CONTIG_PADDR(d, p) \ 56 ((d).dck_paddr && ((d).dck_paddr + IMMU_PAGESIZE) == (p)) 57 58 typedef struct dvma_arg { 59 immu_t *dva_immu; 60 dev_info_t *dva_rdip; 61 dev_info_t *dva_ddip; 62 domain_t *dva_domain; 63 int dva_level; 64 immu_flags_t dva_flags; 65 list_t *dva_list; 66 int dva_error; 67 } dvma_arg_t; 68 69 static domain_t *domain_create(immu_t *immu, dev_info_t *ddip, 70 dev_info_t *rdip, immu_flags_t immu_flags); 71 static immu_devi_t *create_immu_devi(dev_info_t *rdip, int bus, 72 int dev, int func, immu_flags_t immu_flags); 73 static void destroy_immu_devi(immu_devi_t *immu_devi); 74 static boolean_t dvma_map(immu_t *immu, domain_t *domain, uint64_t sdvma, 75 uint64_t nvpages, dcookie_t *dcookies, int dcount, dev_info_t *rdip, 76 immu_flags_t immu_flags); 77 78 /* Extern globals */ 79 extern struct memlist *phys_install; 80 81 82 /* static Globals */ 83 84 /* 85 * Used to setup DMA objects (memory regions) 86 * for DMA reads by IOMMU units 87 */ 88 static ddi_dma_attr_t immu_dma_attr = { 89 DMA_ATTR_V0, 90 0U, 91 0xffffffffU, 92 0xffffffffU, 93 MMU_PAGESIZE, /* MMU page aligned */ 94 0x1, 95 0x1, 96 0xffffffffU, 97 0xffffffffU, 98 1, 99 4, 100 0 101 }; 102 103 static ddi_device_acc_attr_t immu_acc_attr = { 104 DDI_DEVICE_ATTR_V0, 105 DDI_NEVERSWAP_ACC, 106 DDI_STRICTORDER_ACC 107 }; 108 109 110 /* globals private to this file */ 111 static kmutex_t immu_domain_lock; 112 static list_t immu_unity_domain_list; 113 static list_t immu_xlate_domain_list; 114 115 /* structure used to store idx into each level of the page tables */ 116 typedef struct xlate { 117 int xlt_level; 118 uint_t xlt_idx; 119 pgtable_t *xlt_pgtable; 120 } xlate_t; 121 122 /* 0 is reserved by Vt-d spec. Solaris reserves 1 */ 123 #define IMMU_UNITY_DID 1 124 125 static mod_hash_t *bdf_domain_hash; 126 127 static domain_t * 128 bdf_domain_lookup(immu_devi_t *immu_devi) 129 { 130 domain_t *domain; 131 int16_t seg = immu_devi->imd_seg; 132 int16_t bus = immu_devi->imd_bus; 133 int16_t devfunc = immu_devi->imd_devfunc; 134 uintptr_t bdf = (seg << 16 | bus << 8 | devfunc); 135 136 if (seg < 0 || bus < 0 || devfunc < 0) { 137 return (NULL); 138 } 139 140 domain = NULL; 141 if (mod_hash_find(bdf_domain_hash, 142 (void *)bdf, (void *)&domain) == 0) { 143 ASSERT(domain); 144 ASSERT(domain->dom_did > 0); 145 return (domain); 146 } else { 147 return (NULL); 148 } 149 } 150 151 static void 152 bdf_domain_insert(immu_devi_t *immu_devi, domain_t *domain) 153 { 154 int16_t seg = immu_devi->imd_seg; 155 int16_t bus = immu_devi->imd_bus; 156 int16_t devfunc = immu_devi->imd_devfunc; 157 uintptr_t bdf = (seg << 16 | bus << 8 | devfunc); 158 int r; 159 160 if (seg < 0 || bus < 0 || devfunc < 0) { 161 return; 162 } 163 164 r = mod_hash_insert(bdf_domain_hash, (void *)bdf, (void *)domain); 165 ASSERT(r != MH_ERR_DUPLICATE); 166 ASSERT(r == 0); 167 } 168 169 static int 170 match_lpc(dev_info_t *pdip, void *arg) 171 { 172 immu_devi_t *immu_devi; 173 dvma_arg_t *dvap = (dvma_arg_t *)arg; 174 175 ASSERT(dvap->dva_error == DDI_FAILURE); 176 ASSERT(dvap->dva_ddip == NULL); 177 ASSERT(dvap->dva_list); 178 179 if (list_is_empty(dvap->dva_list)) { 180 return (DDI_WALK_TERMINATE); 181 } 182 183 immu_devi = list_head(dvap->dva_list); 184 for (; immu_devi; immu_devi = list_next(dvap->dva_list, 185 immu_devi)) { 186 ASSERT(immu_devi->imd_dip); 187 if (immu_devi->imd_dip == pdip) { 188 dvap->dva_ddip = pdip; 189 dvap->dva_error = DDI_SUCCESS; 190 return (DDI_WALK_TERMINATE); 191 } 192 } 193 194 return (DDI_WALK_CONTINUE); 195 } 196 197 static void 198 immu_devi_set_spclist(dev_info_t *dip, immu_t *immu) 199 { 200 list_t *spclist = NULL; 201 immu_devi_t *immu_devi; 202 203 ASSERT(MUTEX_HELD(&(DEVI(dip)->devi_lock))); 204 205 immu_devi = IMMU_DEVI(dip); 206 if (immu_devi->imd_display == B_TRUE) { 207 spclist = &(immu->immu_dvma_gfx_list); 208 } else if (immu_devi->imd_lpc == B_TRUE) { 209 spclist = &(immu->immu_dvma_lpc_list); 210 } 211 212 if (spclist) { 213 mutex_enter(&(immu->immu_lock)); 214 list_insert_head(spclist, immu_devi); 215 mutex_exit(&(immu->immu_lock)); 216 } 217 } 218 219 /* 220 * Set the immu_devi struct in the immu_devi field of a devinfo node 221 */ 222 int 223 immu_devi_set(dev_info_t *dip, immu_flags_t immu_flags) 224 { 225 int bus, dev, func; 226 immu_devi_t *new_imd; 227 immu_devi_t *immu_devi; 228 229 ASSERT(root_devinfo); 230 ASSERT(dip); 231 ASSERT(dip != root_devinfo); 232 233 immu_devi = immu_devi_get(dip); 234 if (immu_devi != NULL) { 235 return (DDI_SUCCESS); 236 } 237 238 bus = dev = func = -1; 239 240 /* 241 * Assume a new immu_devi struct is needed 242 */ 243 if (!DEVI_IS_PCI(dip) || acpica_get_bdf(dip, &bus, &dev, &func) != 0) { 244 /* 245 * No BDF. Set bus = -1 to indicate this. 246 * We still need to create a immu_devi struct 247 * though 248 */ 249 bus = -1; 250 dev = 0; 251 func = 0; 252 } 253 254 new_imd = create_immu_devi(dip, bus, dev, func, immu_flags); 255 if (new_imd == NULL) { 256 ddi_err(DER_WARN, dip, "Failed to create immu_devi " 257 "structure"); 258 return (DDI_FAILURE); 259 } 260 261 /* 262 * Check if some other thread allocated a immu_devi while we 263 * didn't own the lock. 264 */ 265 mutex_enter(&(DEVI(dip)->devi_lock)); 266 if (IMMU_DEVI(dip) == NULL) { 267 IMMU_DEVI_SET(dip, new_imd); 268 } else { 269 destroy_immu_devi(new_imd); 270 } 271 mutex_exit(&(DEVI(dip)->devi_lock)); 272 273 return (DDI_SUCCESS); 274 } 275 276 static dev_info_t * 277 get_lpc_devinfo(immu_t *immu, dev_info_t *rdip, immu_flags_t immu_flags) 278 { 279 dvma_arg_t dvarg = {0}; 280 dvarg.dva_list = &(immu->immu_dvma_lpc_list); 281 dvarg.dva_rdip = rdip; 282 dvarg.dva_error = DDI_FAILURE; 283 284 if (immu_walk_ancestor(rdip, NULL, match_lpc, 285 &dvarg, NULL, immu_flags) != DDI_SUCCESS) { 286 ddi_err(DER_MODE, rdip, "Could not walk ancestors to " 287 "find lpc_devinfo for ISA device"); 288 return (NULL); 289 } 290 291 if (dvarg.dva_error != DDI_SUCCESS || dvarg.dva_ddip == NULL) { 292 ddi_err(DER_MODE, rdip, "Could not find lpc_devinfo for " 293 "ISA device"); 294 return (NULL); 295 } 296 297 return (dvarg.dva_ddip); 298 } 299 300 static dev_info_t * 301 get_gfx_devinfo(dev_info_t *rdip) 302 { 303 immu_t *immu; 304 immu_devi_t *immu_devi; 305 list_t *list_gfx; 306 307 /* 308 * The GFX device may not be on the same IMMU unit as "agpgart" 309 * so search globally 310 */ 311 immu_devi = NULL; 312 immu = list_head(&immu_list); 313 for (; immu; immu = list_next(&immu_list, immu)) { 314 list_gfx = &(immu->immu_dvma_gfx_list); 315 if (!list_is_empty(list_gfx)) { 316 immu_devi = list_head(list_gfx); 317 break; 318 } 319 } 320 321 if (immu_devi == NULL) { 322 ddi_err(DER_WARN, rdip, "IMMU: No GFX device. " 323 "Cannot redirect agpgart"); 324 return (NULL); 325 } 326 327 /* list is not empty we checked above */ 328 ASSERT(immu_devi); 329 ASSERT(immu_devi->imd_dip); 330 331 ddi_err(DER_LOG, rdip, "IMMU: GFX redirect to %s", 332 ddi_node_name(immu_devi->imd_dip)); 333 334 return (immu_devi->imd_dip); 335 } 336 337 static immu_flags_t 338 dma_to_immu_flags(struct ddi_dma_req *dmareq) 339 { 340 immu_flags_t flags = 0; 341 342 if (dmareq->dmar_fp == DDI_DMA_SLEEP) { 343 flags |= IMMU_FLAGS_SLEEP; 344 } else { 345 flags |= IMMU_FLAGS_NOSLEEP; 346 } 347 348 #ifdef BUGGY_DRIVERS 349 350 flags |= (IMMU_FLAGS_READ | IMMU_FLAGS_WRITE); 351 352 #else 353 /* 354 * Read and write flags need to be reversed. 355 * DMA_READ means read from device and write 356 * to memory. So DMA read means DVMA write. 357 */ 358 if (dmareq->dmar_flags & DDI_DMA_READ) 359 flags |= IMMU_FLAGS_WRITE; 360 361 if (dmareq->dmar_flags & DDI_DMA_WRITE) 362 flags |= IMMU_FLAGS_READ; 363 364 /* 365 * Some buggy drivers specify neither READ or WRITE 366 * For such drivers set both read and write permissions 367 */ 368 if ((dmareq->dmar_flags & (DDI_DMA_READ | DDI_DMA_WRITE)) == 0) { 369 flags |= (IMMU_FLAGS_READ | IMMU_FLAGS_WRITE); 370 } 371 #endif 372 373 return (flags); 374 } 375 376 int 377 pgtable_ctor(void *buf, void *arg, int kmflag) 378 { 379 size_t actual_size = 0; 380 pgtable_t *pgtable; 381 int (*dmafp)(caddr_t); 382 caddr_t vaddr; 383 void *next; 384 385 ASSERT(buf); 386 ASSERT(arg == NULL); 387 388 pgtable = (pgtable_t *)buf; 389 390 dmafp = (kmflag & KM_NOSLEEP) ? DDI_DMA_DONTWAIT : DDI_DMA_SLEEP; 391 392 next = kmem_zalloc(IMMU_PAGESIZE, kmflag); 393 if (next == NULL) { 394 return (-1); 395 } 396 397 ASSERT(root_devinfo); 398 if (ddi_dma_alloc_handle(root_devinfo, &immu_dma_attr, 399 dmafp, NULL, &pgtable->hwpg_dmahdl) != DDI_SUCCESS) { 400 kmem_free(next, IMMU_PAGESIZE); 401 return (-1); 402 } 403 404 if (ddi_dma_mem_alloc(pgtable->hwpg_dmahdl, IMMU_PAGESIZE, 405 &immu_acc_attr, DDI_DMA_CONSISTENT | IOMEM_DATA_UNCACHED, 406 dmafp, NULL, &vaddr, &actual_size, 407 &pgtable->hwpg_memhdl) != DDI_SUCCESS) { 408 ddi_dma_free_handle(&pgtable->hwpg_dmahdl); 409 kmem_free(next, IMMU_PAGESIZE); 410 return (-1); 411 } 412 413 /* 414 * Memory allocation failure. Maybe a temporary condition 415 * so return error rather than panic, so we can try again 416 */ 417 if (actual_size < IMMU_PAGESIZE) { 418 ddi_dma_mem_free(&pgtable->hwpg_memhdl); 419 ddi_dma_free_handle(&pgtable->hwpg_dmahdl); 420 kmem_free(next, IMMU_PAGESIZE); 421 return (-1); 422 } 423 424 pgtable->hwpg_paddr = pfn_to_pa(hat_getpfnum(kas.a_hat, vaddr)); 425 pgtable->hwpg_vaddr = vaddr; 426 pgtable->swpg_next_array = next; 427 428 rw_init(&(pgtable->swpg_rwlock), NULL, RW_DEFAULT, NULL); 429 430 return (0); 431 } 432 433 void 434 pgtable_dtor(void *buf, void *arg) 435 { 436 pgtable_t *pgtable; 437 438 ASSERT(buf); 439 ASSERT(arg == NULL); 440 441 pgtable = (pgtable_t *)buf; 442 ASSERT(pgtable->swpg_next_array); 443 444 /* destroy will panic if lock is held. */ 445 rw_destroy(&(pgtable->swpg_rwlock)); 446 447 ddi_dma_mem_free(&pgtable->hwpg_memhdl); 448 ddi_dma_free_handle(&pgtable->hwpg_dmahdl); 449 kmem_free(pgtable->swpg_next_array, IMMU_PAGESIZE); 450 451 /* don't zero out hwpg_vaddr and swpg_next_array for debugging */ 452 } 453 454 /* 455 * pgtable_alloc() 456 * alloc a IOMMU pgtable structure. 457 * This same struct is used for root and context tables as well. 458 * This routine allocs the f/ollowing: 459 * - a pgtable_t struct 460 * - a HW page which holds PTEs/entries which is accesssed by HW 461 * so we set up DMA for this page 462 * - a SW page which is only for our bookeeping 463 * (for example to hold pointers to the next level pgtable). 464 * So a simple kmem_alloc suffices 465 */ 466 static pgtable_t * 467 pgtable_alloc(immu_t *immu, immu_flags_t immu_flags) 468 { 469 pgtable_t *pgtable; 470 int kmflags; 471 472 ASSERT(immu); 473 474 kmflags = (immu_flags & IMMU_FLAGS_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP; 475 476 pgtable = kmem_cache_alloc(immu_pgtable_cache, kmflags); 477 if (pgtable == NULL) { 478 return (NULL); 479 } 480 return (pgtable); 481 } 482 483 static void 484 pgtable_zero(immu_t *immu, pgtable_t *pgtable) 485 { 486 bzero(pgtable->hwpg_vaddr, IMMU_PAGESIZE); 487 bzero(pgtable->swpg_next_array, IMMU_PAGESIZE); 488 489 /* Dont need to flush the write we will flush when we use the entry */ 490 immu_regs_cpu_flush(immu, pgtable->hwpg_vaddr, IMMU_PAGESIZE); 491 } 492 493 static void 494 pgtable_free(immu_t *immu, pgtable_t *pgtable) 495 { 496 ASSERT(immu); 497 ASSERT(pgtable); 498 499 kmem_cache_free(immu_pgtable_cache, pgtable); 500 } 501 502 /* 503 * Function to identify a display device from the PCI class code 504 */ 505 static boolean_t 506 device_is_display(uint_t classcode) 507 { 508 static uint_t disp_classes[] = { 509 0x000100, 510 0x030000, 511 0x030001 512 }; 513 int i, nclasses = sizeof (disp_classes) / sizeof (uint_t); 514 515 for (i = 0; i < nclasses; i++) { 516 if (classcode == disp_classes[i]) 517 return (B_TRUE); 518 } 519 return (B_FALSE); 520 } 521 522 /* 523 * Function that determines if device is PCIEX and/or PCIEX bridge 524 */ 525 static boolean_t 526 device_is_pciex( 527 uchar_t bus, uchar_t dev, uchar_t func, boolean_t *is_pcib) 528 { 529 ushort_t cap; 530 ushort_t capsp; 531 ushort_t cap_count = PCI_CAP_MAX_PTR; 532 ushort_t status; 533 boolean_t is_pciex = B_FALSE; 534 535 *is_pcib = B_FALSE; 536 537 status = pci_getw_func(bus, dev, func, PCI_CONF_STAT); 538 if (!(status & PCI_STAT_CAP)) 539 return (B_FALSE); 540 541 capsp = pci_getb_func(bus, dev, func, PCI_CONF_CAP_PTR); 542 while (cap_count-- && capsp >= PCI_CAP_PTR_OFF) { 543 capsp &= PCI_CAP_PTR_MASK; 544 cap = pci_getb_func(bus, dev, func, capsp); 545 546 if (cap == PCI_CAP_ID_PCI_E) { 547 status = pci_getw_func(bus, dev, func, capsp + 2); 548 /* 549 * See section 7.8.2 of PCI-Express Base Spec v1.0a 550 * for Device/Port Type. 551 * PCIE_PCIECAP_DEV_TYPE_PCIE2PCI implies that the 552 * device is a PCIE2PCI bridge 553 */ 554 *is_pcib = 555 ((status & PCIE_PCIECAP_DEV_TYPE_MASK) == 556 PCIE_PCIECAP_DEV_TYPE_PCIE2PCI) ? B_TRUE : B_FALSE; 557 is_pciex = B_TRUE; 558 } 559 560 capsp = (*pci_getb_func)(bus, dev, func, 561 capsp + PCI_CAP_NEXT_PTR); 562 } 563 564 return (is_pciex); 565 } 566 567 568 /* 569 * immu_dvma_get_immu() 570 * get the immu unit structure for a dev_info node 571 */ 572 immu_t * 573 immu_dvma_get_immu(dev_info_t *dip, immu_flags_t immu_flags) 574 { 575 immu_devi_t *immu_devi; 576 immu_t *immu; 577 578 /* 579 * check if immu unit was already found earlier. 580 * If yes, then it will be stashed in immu_devi struct. 581 */ 582 immu_devi = immu_devi_get(dip); 583 if (immu_devi == NULL) { 584 if (immu_devi_set(dip, immu_flags) != DDI_SUCCESS) { 585 /* 586 * May fail because of low memory. Return error rather 587 * than panic as we want driver to rey again later 588 */ 589 ddi_err(DER_PANIC, dip, "immu_dvma_get_immu: " 590 "No immu_devi structure"); 591 /*NOTREACHED*/ 592 } 593 immu_devi = immu_devi_get(dip); 594 ASSERT(immu_devi); 595 } 596 597 mutex_enter(&(DEVI(dip)->devi_lock)); 598 if (immu_devi->imd_immu) { 599 immu = immu_devi->imd_immu; 600 mutex_exit(&(DEVI(dip)->devi_lock)); 601 return (immu); 602 } 603 mutex_exit(&(DEVI(dip)->devi_lock)); 604 605 immu = immu_dmar_get_immu(dip); 606 if (immu == NULL) { 607 ddi_err(DER_PANIC, dip, "immu_dvma_get_immu: " 608 "Cannot find immu_t for device"); 609 /*NOTREACHED*/ 610 } 611 612 /* 613 * Check if some other thread found immu 614 * while lock was not held 615 */ 616 immu_devi = immu_devi_get(dip); 617 /* immu_devi should be present as we found it earlier */ 618 if (immu_devi == NULL) { 619 ddi_err(DER_PANIC, dip, 620 "immu_dvma_get_immu: No immu_devi structure"); 621 /*NOTREACHED*/ 622 } 623 624 mutex_enter(&(DEVI(dip)->devi_lock)); 625 if (immu_devi->imd_immu == NULL) { 626 /* nobody else set it, so we should do it */ 627 immu_devi->imd_immu = immu; 628 immu_devi_set_spclist(dip, immu); 629 } else { 630 /* 631 * if some other thread got immu before 632 * us, it should get the same results 633 */ 634 if (immu_devi->imd_immu != immu) { 635 ddi_err(DER_PANIC, dip, "Multiple " 636 "immu units found for device. Expected (%p), " 637 "actual (%p)", (void *)immu, 638 (void *)immu_devi->imd_immu); 639 mutex_exit(&(DEVI(dip)->devi_lock)); 640 /*NOTREACHED*/ 641 } 642 } 643 mutex_exit(&(DEVI(dip)->devi_lock)); 644 645 return (immu); 646 } 647 648 649 /* ############################# IMMU_DEVI code ############################ */ 650 651 /* 652 * Allocate a immu_devi structure and initialize it 653 */ 654 static immu_devi_t * 655 create_immu_devi(dev_info_t *rdip, int bus, int dev, int func, 656 immu_flags_t immu_flags) 657 { 658 uchar_t baseclass, subclass; 659 uint_t classcode, revclass; 660 immu_devi_t *immu_devi; 661 boolean_t pciex = B_FALSE; 662 int kmflags; 663 boolean_t is_pcib = B_FALSE; 664 665 /* bus == -1 indicate non-PCI device (no BDF) */ 666 ASSERT(bus == -1 || bus >= 0); 667 ASSERT(dev >= 0); 668 ASSERT(func >= 0); 669 670 kmflags = (immu_flags & IMMU_FLAGS_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP; 671 immu_devi = kmem_zalloc(sizeof (immu_devi_t), kmflags); 672 if (immu_devi == NULL) { 673 ddi_err(DER_WARN, rdip, "Failed to allocate memory for " 674 "Intel IOMMU immu_devi structure"); 675 return (NULL); 676 } 677 immu_devi->imd_dip = rdip; 678 immu_devi->imd_seg = 0; /* Currently seg can only be 0 */ 679 immu_devi->imd_bus = bus; 680 immu_devi->imd_pcib_type = IMMU_PCIB_BAD; 681 682 if (bus == -1) { 683 immu_devi->imd_pcib_type = IMMU_PCIB_NOBDF; 684 return (immu_devi); 685 } 686 687 immu_devi->imd_devfunc = IMMU_PCI_DEVFUNC(dev, func); 688 immu_devi->imd_sec = 0; 689 immu_devi->imd_sub = 0; 690 691 revclass = pci_getl_func(bus, dev, func, PCI_CONF_REVID); 692 693 classcode = IMMU_PCI_REV2CLASS(revclass); 694 baseclass = IMMU_PCI_CLASS2BASE(classcode); 695 subclass = IMMU_PCI_CLASS2SUB(classcode); 696 697 if (baseclass == PCI_CLASS_BRIDGE && subclass == PCI_BRIDGE_PCI) { 698 699 immu_devi->imd_sec = pci_getb_func(bus, dev, func, 700 PCI_BCNF_SECBUS); 701 immu_devi->imd_sub = pci_getb_func(bus, dev, func, 702 PCI_BCNF_SUBBUS); 703 704 pciex = device_is_pciex(bus, dev, func, &is_pcib); 705 if (pciex == B_TRUE && is_pcib == B_TRUE) { 706 immu_devi->imd_pcib_type = IMMU_PCIB_PCIE_PCI; 707 } else if (pciex == B_TRUE) { 708 immu_devi->imd_pcib_type = IMMU_PCIB_PCIE_PCIE; 709 } else { 710 immu_devi->imd_pcib_type = IMMU_PCIB_PCI_PCI; 711 } 712 } else { 713 immu_devi->imd_pcib_type = IMMU_PCIB_ENDPOINT; 714 } 715 716 /* check for certain special devices */ 717 immu_devi->imd_display = device_is_display(classcode); 718 719 immu_devi->imd_lpc = ((baseclass == PCI_CLASS_BRIDGE) && 720 (subclass == PCI_BRIDGE_ISA)) ? B_TRUE : B_FALSE; 721 722 immu_devi->imd_domain = NULL; 723 724 immu_devi->imd_dvma_flags = immu_global_dvma_flags; 725 726 return (immu_devi); 727 } 728 729 static void 730 destroy_immu_devi(immu_devi_t *immu_devi) 731 { 732 kmem_free(immu_devi, sizeof (immu_devi_t)); 733 } 734 735 static domain_t * 736 immu_devi_domain(dev_info_t *rdip, dev_info_t **ddipp) 737 { 738 immu_devi_t *immu_devi; 739 domain_t *domain; 740 dev_info_t *ddip; 741 742 ASSERT(rdip); 743 ASSERT(ddipp); 744 745 *ddipp = NULL; 746 747 immu_devi = immu_devi_get(rdip); 748 if (immu_devi == NULL) { 749 return (NULL); 750 } 751 752 mutex_enter(&(DEVI(rdip)->devi_lock)); 753 domain = immu_devi->imd_domain; 754 ddip = immu_devi->imd_ddip; 755 mutex_exit(&(DEVI(rdip)->devi_lock)); 756 757 if (domain) { 758 ASSERT(domain->dom_did > 0); 759 ASSERT(ddip); 760 *ddipp = ddip; 761 } 762 763 return (domain); 764 765 } 766 767 /* ############################# END IMMU_DEVI code ######################## */ 768 /* ############################# DOMAIN code ############################### */ 769 770 /* 771 * This routine always succeeds 772 */ 773 static int 774 did_alloc(immu_t *immu, dev_info_t *rdip, 775 dev_info_t *ddip, immu_flags_t immu_flags) 776 { 777 int did; 778 779 ASSERT(immu); 780 ASSERT(rdip); 781 ASSERT(rdip != root_devinfo); 782 783 did = (uintptr_t)vmem_alloc(immu->immu_did_arena, 1, 784 (immu_flags & IMMU_FLAGS_NOSLEEP) ? VM_NOSLEEP : VM_SLEEP); 785 786 if (did == 0) { 787 ASSERT(immu->immu_unity_domain); 788 ASSERT(immu->immu_unity_domain->dom_did > 0); 789 ddi_err(DER_WARN, rdip, "device domain-id alloc error" 790 " domain-device: %s%d. immu unit is %s. Using " 791 "unity domain with domain-id (%d)", 792 ddi_driver_name(ddip), ddi_get_instance(ddip), 793 immu->immu_name, immu->immu_unity_domain->dom_did); 794 did = immu->immu_unity_domain->dom_did; 795 } 796 797 return (did); 798 } 799 800 static int 801 get_branch_domain(dev_info_t *pdip, void *arg) 802 { 803 immu_devi_t *immu_devi; 804 domain_t *domain; 805 dev_info_t *ddip; 806 immu_t *immu; 807 dvma_arg_t *dvp = (dvma_arg_t *)arg; 808 809 ASSERT(pdip); 810 ASSERT(dvp); 811 ASSERT(dvp->dva_rdip); 812 813 /* 814 * The field dvp->dva_rdip is a work-in-progress 815 * and gets updated as we walk up the ancestor 816 * tree. The final ddip is set only when we reach 817 * the top of the tree. So the dvp->dva_ddip field cannot 818 * be relied on until we reach the top of the field. 819 */ 820 821 /* immu_devi may not be set. */ 822 immu_devi = immu_devi_get(pdip); 823 if (immu_devi == NULL) { 824 if (immu_devi_set(pdip, dvp->dva_flags) != DDI_SUCCESS) { 825 dvp->dva_error = DDI_FAILURE; 826 return (DDI_WALK_TERMINATE); 827 } 828 } 829 830 immu_devi = immu_devi_get(pdip); 831 ASSERT(immu_devi); 832 immu = immu_devi->imd_immu; 833 if (immu == NULL) { 834 immu = immu_dvma_get_immu(pdip, dvp->dva_flags); 835 ASSERT(immu); 836 } 837 838 /* 839 * If we encounter a PCIE_PCIE bridge *ANCESTOR* we need to 840 * terminate the walk (since the device under the PCIE bridge 841 * is a PCIE device and has an independent entry in the 842 * root/context table) 843 */ 844 if (dvp->dva_rdip != pdip && 845 immu_devi->imd_pcib_type == IMMU_PCIB_PCIE_PCIE) { 846 return (DDI_WALK_TERMINATE); 847 } 848 849 /* 850 * In order to be a domain-dim, it must be a PCI device i.e. 851 * must have valid BDF. This also eliminates the root complex. 852 */ 853 if (immu_devi->imd_pcib_type != IMMU_PCIB_BAD && 854 immu_devi->imd_pcib_type != IMMU_PCIB_NOBDF) { 855 ASSERT(immu_devi->imd_bus >= 0); 856 ASSERT(immu_devi->imd_devfunc >= 0); 857 dvp->dva_ddip = pdip; 858 } 859 860 if (immu_devi->imd_display == B_TRUE || 861 (dvp->dva_flags & IMMU_FLAGS_UNITY)) { 862 dvp->dva_domain = immu->immu_unity_domain; 863 /* continue walking to find ddip */ 864 return (DDI_WALK_CONTINUE); 865 } 866 867 mutex_enter(&(DEVI(pdip)->devi_lock)); 868 domain = immu_devi->imd_domain; 869 ddip = immu_devi->imd_ddip; 870 mutex_exit(&(DEVI(pdip)->devi_lock)); 871 872 if (domain && ddip) { 873 /* if domain is set, it must be the same */ 874 if (dvp->dva_domain) { 875 ASSERT(domain == dvp->dva_domain); 876 } 877 dvp->dva_domain = domain; 878 dvp->dva_ddip = ddip; 879 return (DDI_WALK_TERMINATE); 880 } 881 882 /* immu_devi either has both set or both clear */ 883 ASSERT(domain == NULL); 884 ASSERT(ddip == NULL); 885 886 /* Domain may already be set, continue walking so that ddip gets set */ 887 if (dvp->dva_domain) { 888 return (DDI_WALK_CONTINUE); 889 } 890 891 /* domain is not set in either immu_devi or dvp */ 892 domain = bdf_domain_lookup(immu_devi); 893 if (domain == NULL) { 894 return (DDI_WALK_CONTINUE); 895 } 896 897 /* ok, the BDF hash had a domain for this BDF. */ 898 899 /* Grab lock again to check if something else set immu_devi fields */ 900 mutex_enter(&(DEVI(pdip)->devi_lock)); 901 if (immu_devi->imd_domain != NULL) { 902 ASSERT(immu_devi->imd_domain == domain); 903 dvp->dva_domain = domain; 904 } else { 905 dvp->dva_domain = domain; 906 } 907 mutex_exit(&(DEVI(pdip)->devi_lock)); 908 909 /* 910 * walk upwards until the topmost PCI bridge is found 911 */ 912 return (DDI_WALK_CONTINUE); 913 914 } 915 916 static void 917 map_unity_domain(domain_t *domain) 918 { 919 struct memlist *mp; 920 uint64_t start; 921 uint64_t npages; 922 dcookie_t dcookies[1] = {0}; 923 int dcount = 0; 924 925 ASSERT(domain); 926 ASSERT(domain->dom_did == IMMU_UNITY_DID); 927 928 /* 929 * We call into routines that grab the lock so we should 930 * not be called with the lock held. This does not matter 931 * much since, no else has a reference to this domain 932 */ 933 ASSERT(!rw_lock_held(&(domain->dom_pgtable_rwlock))); 934 935 /* 936 * UNITY arenas are a mirror of the physical memory 937 * installed on the system. 938 */ 939 940 #ifdef BUGGY_DRIVERS 941 /* 942 * Dont skip page0. Some broken HW/FW access it. 943 */ 944 dcookies[0].dck_paddr = 0; 945 dcookies[0].dck_npages = 1; 946 dcount = 1; 947 (void) dvma_map(domain->dom_immu, domain, 0, 1, dcookies, dcount, NULL, 948 IMMU_FLAGS_READ | IMMU_FLAGS_WRITE | IMMU_FLAGS_PAGE1); 949 #endif 950 951 memlist_read_lock(); 952 953 mp = phys_install; 954 955 if (mp->ml_address == 0) { 956 /* since we already mapped page1 above */ 957 start = IMMU_PAGESIZE; 958 } else { 959 start = mp->ml_address; 960 } 961 npages = mp->ml_size/IMMU_PAGESIZE + 1; 962 963 dcookies[0].dck_paddr = start; 964 dcookies[0].dck_npages = npages; 965 dcount = 1; 966 (void) dvma_map(domain->dom_immu, domain, start, npages, dcookies, 967 dcount, NULL, IMMU_FLAGS_READ | IMMU_FLAGS_WRITE); 968 969 ddi_err(DER_LOG, NULL, "IMMU: mapping PHYS span [0x%" PRIx64 970 " - 0x%" PRIx64 "]", start, start + mp->ml_size); 971 972 mp = mp->ml_next; 973 while (mp) { 974 ddi_err(DER_LOG, NULL, "IMMU: mapping PHYS span [0x%" PRIx64 975 " - 0x%" PRIx64 "]", mp->ml_address, 976 mp->ml_address + mp->ml_size); 977 978 start = mp->ml_address; 979 npages = mp->ml_size/IMMU_PAGESIZE + 1; 980 981 dcookies[0].dck_paddr = start; 982 dcookies[0].dck_npages = npages; 983 dcount = 1; 984 (void) dvma_map(domain->dom_immu, domain, start, npages, 985 dcookies, dcount, NULL, IMMU_FLAGS_READ | IMMU_FLAGS_WRITE); 986 mp = mp->ml_next; 987 } 988 989 mp = bios_rsvd; 990 while (mp) { 991 ddi_err(DER_LOG, NULL, "IMMU: mapping PHYS span [0x%" PRIx64 992 " - 0x%" PRIx64 "]", mp->ml_address, 993 mp->ml_address + mp->ml_size); 994 995 start = mp->ml_address; 996 npages = mp->ml_size/IMMU_PAGESIZE + 1; 997 998 dcookies[0].dck_paddr = start; 999 dcookies[0].dck_npages = npages; 1000 dcount = 1; 1001 (void) dvma_map(domain->dom_immu, domain, start, npages, 1002 dcookies, dcount, NULL, IMMU_FLAGS_READ | IMMU_FLAGS_WRITE); 1003 1004 mp = mp->ml_next; 1005 } 1006 1007 memlist_read_unlock(); 1008 } 1009 1010 /* 1011 * create_xlate_arena() 1012 * Create the dvma arena for a domain with translation 1013 * mapping 1014 */ 1015 static void 1016 create_xlate_arena(immu_t *immu, domain_t *domain, 1017 dev_info_t *rdip, immu_flags_t immu_flags) 1018 { 1019 char *arena_name; 1020 struct memlist *mp; 1021 int vmem_flags; 1022 uint64_t start; 1023 uint_t mgaw; 1024 uint64_t size; 1025 uint64_t maxaddr; 1026 void *vmem_ret; 1027 1028 arena_name = domain->dom_dvma_arena_name; 1029 1030 /* Note, don't do sizeof (arena_name) - it is just a pointer */ 1031 (void) snprintf(arena_name, 1032 sizeof (domain->dom_dvma_arena_name), 1033 "%s-domain-%d-xlate-DVMA-arena", immu->immu_name, 1034 domain->dom_did); 1035 1036 vmem_flags = (immu_flags & IMMU_FLAGS_NOSLEEP) ? VM_NOSLEEP : VM_SLEEP; 1037 1038 /* 1039 * No one else has access to this domain. 1040 * So no domain locks needed 1041 */ 1042 ASSERT(!rw_lock_held(&(domain->dom_pgtable_rwlock))); 1043 1044 /* Restrict mgaddr (max guest addr) to MGAW */ 1045 mgaw = IMMU_CAP_MGAW(immu->immu_regs_cap); 1046 1047 /* 1048 * To ensure we avoid ioapic and PCI MMIO ranges we just 1049 * use the physical memory address range of the system as the 1050 * range 1051 */ 1052 maxaddr = ((uint64_t)1 << mgaw); 1053 1054 memlist_read_lock(); 1055 1056 mp = phys_install; 1057 1058 if (mp->ml_address == 0) 1059 start = MMU_PAGESIZE; 1060 else 1061 start = mp->ml_address; 1062 1063 if (start + mp->ml_size > maxaddr) 1064 size = maxaddr - start; 1065 else 1066 size = mp->ml_size; 1067 1068 ddi_err(DER_VERB, rdip, 1069 "%s: Creating dvma vmem arena [0x%" PRIx64 1070 " - 0x%" PRIx64 "]", arena_name, start, start + size); 1071 1072 ASSERT(domain->dom_dvma_arena == NULL); 1073 1074 /* 1075 * We always allocate in quanta of IMMU_PAGESIZE 1076 */ 1077 domain->dom_dvma_arena = vmem_create(arena_name, 1078 (void *)(uintptr_t)start, /* start addr */ 1079 size, /* size */ 1080 IMMU_PAGESIZE, /* quantum */ 1081 NULL, /* afunc */ 1082 NULL, /* ffunc */ 1083 NULL, /* source */ 1084 0, /* qcache_max */ 1085 vmem_flags); 1086 1087 if (domain->dom_dvma_arena == NULL) { 1088 ddi_err(DER_PANIC, rdip, 1089 "Failed to allocate DVMA arena(%s) " 1090 "for domain ID (%d)", arena_name, domain->dom_did); 1091 /*NOTREACHED*/ 1092 } 1093 1094 mp = mp->ml_next; 1095 while (mp) { 1096 1097 if (mp->ml_address == 0) 1098 start = MMU_PAGESIZE; 1099 else 1100 start = mp->ml_address; 1101 1102 if (start + mp->ml_size > maxaddr) 1103 size = maxaddr - start; 1104 else 1105 size = mp->ml_size; 1106 1107 ddi_err(DER_VERB, rdip, 1108 "%s: Adding dvma vmem span [0x%" PRIx64 1109 " - 0x%" PRIx64 "]", arena_name, start, 1110 start + size); 1111 1112 vmem_ret = vmem_add(domain->dom_dvma_arena, 1113 (void *)(uintptr_t)start, size, vmem_flags); 1114 1115 if (vmem_ret == NULL) { 1116 ddi_err(DER_PANIC, rdip, 1117 "Failed to allocate DVMA arena(%s) " 1118 "for domain ID (%d)", 1119 arena_name, domain->dom_did); 1120 /*NOTREACHED*/ 1121 } 1122 mp = mp->ml_next; 1123 } 1124 memlist_read_unlock(); 1125 } 1126 1127 /* ################################### DOMAIN CODE ######################### */ 1128 1129 /* 1130 * Set the domain and domain-dip for a dip 1131 */ 1132 static void 1133 set_domain( 1134 dev_info_t *dip, 1135 dev_info_t *ddip, 1136 domain_t *domain) 1137 { 1138 immu_devi_t *immu_devi; 1139 domain_t *fdomain; 1140 dev_info_t *fddip; 1141 1142 ASSERT(dip); 1143 ASSERT(ddip); 1144 ASSERT(domain); 1145 ASSERT(domain->dom_did > 0); /* must be an initialized domain */ 1146 1147 immu_devi = immu_devi_get(dip); 1148 ASSERT(immu_devi); 1149 1150 mutex_enter(&(DEVI(dip)->devi_lock)); 1151 fddip = immu_devi->imd_ddip; 1152 fdomain = immu_devi->imd_domain; 1153 1154 if (fddip) { 1155 ASSERT(fddip == ddip); 1156 } else { 1157 immu_devi->imd_ddip = ddip; 1158 } 1159 1160 if (fdomain) { 1161 ASSERT(fdomain == domain); 1162 } else { 1163 immu_devi->imd_domain = domain; 1164 } 1165 mutex_exit(&(DEVI(dip)->devi_lock)); 1166 } 1167 1168 /* 1169 * device_domain() 1170 * Get domain for a device. The domain may be global in which case it 1171 * is shared between all IOMMU units. Due to potential AGAW differences 1172 * between IOMMU units, such global domains *have to be* UNITY mapping 1173 * domains. Alternatively, the domain may be local to a IOMMU unit. 1174 * Local domains may be shared or immu_devi, although the 1175 * scope of sharing 1176 * is restricted to devices controlled by the IOMMU unit to 1177 * which the domain 1178 * belongs. If shared, they (currently) have to be UNITY domains. If 1179 * immu_devi a domain may be either UNITY or translation (XLATE) domain. 1180 */ 1181 static domain_t * 1182 device_domain(dev_info_t *rdip, dev_info_t **ddipp, immu_flags_t immu_flags) 1183 { 1184 dev_info_t *ddip; /* topmost dip in domain i.e. domain owner */ 1185 immu_t *immu; 1186 domain_t *domain; 1187 dvma_arg_t dvarg = {0}; 1188 int level; 1189 1190 ASSERT(rdip); 1191 1192 *ddipp = NULL; 1193 1194 /* 1195 * Check if the domain is already set. This is usually true 1196 * if this is not the first DVMA transaction. 1197 */ 1198 ddip = NULL; 1199 domain = immu_devi_domain(rdip, &ddip); 1200 if (domain) { 1201 ASSERT(domain->dom_did > 0); 1202 ASSERT(ddip); 1203 *ddipp = ddip; 1204 return (domain); 1205 } 1206 1207 immu = immu_dvma_get_immu(rdip, immu_flags); 1208 if (immu == NULL) { 1209 /* 1210 * possible that there is no IOMMU unit for this device 1211 * - BIOS bugs are one example. 1212 */ 1213 ddi_err(DER_WARN, rdip, "No IMMU unit found for device"); 1214 return (NULL); 1215 } 1216 1217 immu_flags |= immu_devi_get(rdip)->imd_dvma_flags; 1218 1219 dvarg.dva_rdip = rdip; 1220 dvarg.dva_ddip = NULL; 1221 dvarg.dva_domain = NULL; 1222 dvarg.dva_flags = immu_flags; 1223 level = 0; 1224 if (immu_walk_ancestor(rdip, NULL, get_branch_domain, 1225 &dvarg, &level, immu_flags) != DDI_SUCCESS) { 1226 /* 1227 * maybe low memory. return error, 1228 * so driver tries again later 1229 */ 1230 return (NULL); 1231 } 1232 1233 /* should have walked at least 1 dip (i.e. edip) */ 1234 ASSERT(level > 0); 1235 1236 ddip = dvarg.dva_ddip; /* must be present */ 1237 domain = dvarg.dva_domain; /* may be NULL */ 1238 1239 /* 1240 * We may find the domain during our ancestor walk on any one of our 1241 * ancestor dips, If the domain is found then the domain-dip 1242 * (i.e. ddip) will also be found in the same immu_devi struct. 1243 * The domain-dip is the highest ancestor dip which shares the 1244 * same domain with edip. 1245 * The domain may or may not be found, but the domain dip must 1246 * be found. 1247 */ 1248 if (ddip == NULL) { 1249 ddi_err(DER_MODE, rdip, "Cannot find domain dip for device."); 1250 return (NULL); 1251 } 1252 1253 /* 1254 * Did we find a domain ? 1255 */ 1256 if (domain) { 1257 goto found; 1258 } 1259 1260 /* nope, so allocate */ 1261 domain = domain_create(immu, ddip, rdip, immu_flags); 1262 if (domain == NULL) { 1263 return (NULL); 1264 } 1265 ASSERT(domain->dom_did > 0); 1266 1267 /*FALLTHROUGH*/ 1268 found: 1269 /* 1270 * We know *domain *is* the right domain, so panic if 1271 * another domain is set for either the request-dip or 1272 * effective dip. 1273 */ 1274 set_domain(ddip, ddip, domain); 1275 set_domain(rdip, ddip, domain); 1276 1277 *ddipp = ddip; 1278 return (domain); 1279 } 1280 1281 static void 1282 create_unity_domain(immu_t *immu) 1283 { 1284 domain_t *domain; 1285 1286 /* 0 is reserved by Vt-d */ 1287 /*LINTED*/ 1288 ASSERT(IMMU_UNITY_DID > 0); 1289 1290 /* domain created during boot and always use sleep flag */ 1291 domain = kmem_zalloc(sizeof (domain_t), KM_SLEEP); 1292 1293 rw_init(&(domain->dom_pgtable_rwlock), NULL, RW_DEFAULT, NULL); 1294 1295 domain->dom_did = IMMU_UNITY_DID; 1296 domain->dom_maptype = IMMU_MAPTYPE_UNITY; 1297 1298 domain->dom_immu = immu; 1299 immu->immu_unity_domain = domain; 1300 1301 /* 1302 * Setup the domain's initial page table 1303 * should never fail. 1304 */ 1305 domain->dom_pgtable_root = pgtable_alloc(immu, IMMU_FLAGS_SLEEP); 1306 ASSERT(domain->dom_pgtable_root); 1307 pgtable_zero(immu, domain->dom_pgtable_root); 1308 1309 map_unity_domain(domain); 1310 1311 /* 1312 * put it on the system-wide UNITY domain list 1313 */ 1314 mutex_enter(&(immu_domain_lock)); 1315 list_insert_tail(&immu_unity_domain_list, domain); 1316 mutex_exit(&(immu_domain_lock)); 1317 } 1318 1319 /* 1320 * ddip is the domain-dip - the topmost dip in a domain 1321 * rdip is the requesting-dip - the device which is 1322 * requesting DVMA setup 1323 * if domain is a non-shared domain rdip == ddip 1324 */ 1325 static domain_t * 1326 domain_create(immu_t *immu, dev_info_t *ddip, dev_info_t *rdip, 1327 immu_flags_t immu_flags) 1328 { 1329 int kmflags; 1330 domain_t *domain; 1331 char mod_hash_name[128]; 1332 immu_devi_t *immu_devi; 1333 int did; 1334 dcookie_t dcookies[1] = {0}; 1335 int dcount = 0; 1336 1337 ASSERT(immu); 1338 ASSERT(ddip); 1339 1340 immu_devi = immu_devi_get(rdip); 1341 1342 ASSERT(immu_devi); 1343 1344 /* 1345 * First allocate a domainid. 1346 * This routine will never fail, since if we run out 1347 * of domains the unity domain will be allocated. 1348 */ 1349 did = did_alloc(immu, rdip, ddip, immu_flags); 1350 ASSERT(did > 0); 1351 if (did == IMMU_UNITY_DID) { 1352 /* domain overflow */ 1353 ASSERT(immu->immu_unity_domain); 1354 return (immu->immu_unity_domain); 1355 } 1356 1357 kmflags = (immu_flags & IMMU_FLAGS_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP; 1358 domain = kmem_zalloc(sizeof (domain_t), kmflags); 1359 if (domain == NULL) { 1360 ddi_err(DER_PANIC, rdip, "Failed to alloc DVMA domain " 1361 "structure for device. IOMMU unit: %s", immu->immu_name); 1362 /*NOTREACHED*/ 1363 } 1364 1365 rw_init(&(domain->dom_pgtable_rwlock), NULL, RW_DEFAULT, NULL); 1366 1367 (void) snprintf(mod_hash_name, sizeof (mod_hash_name), 1368 "immu%s-domain%d-pava-hash", immu->immu_name, did); 1369 1370 domain->dom_did = did; 1371 domain->dom_immu = immu; 1372 domain->dom_maptype = IMMU_MAPTYPE_XLATE; 1373 1374 /* 1375 * Create xlate DVMA arena for this domain. 1376 */ 1377 create_xlate_arena(immu, domain, rdip, immu_flags); 1378 1379 /* 1380 * Setup the domain's initial page table 1381 */ 1382 domain->dom_pgtable_root = pgtable_alloc(immu, immu_flags); 1383 if (domain->dom_pgtable_root == NULL) { 1384 ddi_err(DER_PANIC, rdip, "Failed to alloc root " 1385 "pgtable for domain (%d). IOMMU unit: %s", 1386 domain->dom_did, immu->immu_name); 1387 /*NOTREACHED*/ 1388 } 1389 pgtable_zero(immu, domain->dom_pgtable_root); 1390 1391 /* 1392 * Since this is a immu unit-specific domain, put it on 1393 * the per-immu domain list. 1394 */ 1395 mutex_enter(&(immu->immu_lock)); 1396 list_insert_head(&immu->immu_domain_list, domain); 1397 mutex_exit(&(immu->immu_lock)); 1398 1399 /* 1400 * Also put it on the system-wide xlate domain list 1401 */ 1402 mutex_enter(&(immu_domain_lock)); 1403 list_insert_head(&immu_xlate_domain_list, domain); 1404 mutex_exit(&(immu_domain_lock)); 1405 1406 bdf_domain_insert(immu_devi, domain); 1407 1408 #ifdef BUGGY_DRIVERS 1409 /* 1410 * Map page0. Some broken HW/FW access it. 1411 */ 1412 dcookies[0].dck_paddr = 0; 1413 dcookies[0].dck_npages = 1; 1414 dcount = 1; 1415 (void) dvma_map(domain->dom_immu, domain, 0, 1, dcookies, dcount, NULL, 1416 IMMU_FLAGS_READ | IMMU_FLAGS_WRITE | IMMU_FLAGS_PAGE1); 1417 #endif 1418 return (domain); 1419 } 1420 1421 /* 1422 * Create domainid arena. 1423 * Domainid 0 is reserved by Vt-d spec and cannot be used by 1424 * system software. 1425 * Domainid 1 is reserved by solaris and used for *all* of the following: 1426 * as the "uninitialized" domain - For devices not yet controlled 1427 * by Solaris 1428 * as the "unity" domain - For devices that will always belong 1429 * to the unity domain 1430 * as the "overflow" domain - Used for any new device after we 1431 * run out of domains 1432 * All of the above domains map into a single domain with 1433 * domainid 1 and UNITY DVMA mapping 1434 * Each IMMU unity has its own unity/uninit/overflow domain 1435 */ 1436 static void 1437 did_init(immu_t *immu) 1438 { 1439 (void) snprintf(immu->immu_did_arena_name, 1440 sizeof (immu->immu_did_arena_name), 1441 "%s_domainid_arena", immu->immu_name); 1442 1443 ddi_err(DER_VERB, NULL, "%s: Creating domainid arena %s", 1444 immu->immu_name, immu->immu_did_arena_name); 1445 1446 immu->immu_did_arena = vmem_create( 1447 immu->immu_did_arena_name, 1448 (void *)(uintptr_t)(IMMU_UNITY_DID + 1), /* start addr */ 1449 immu->immu_max_domains - IMMU_UNITY_DID, 1450 1, /* quantum */ 1451 NULL, /* afunc */ 1452 NULL, /* ffunc */ 1453 NULL, /* source */ 1454 0, /* qcache_max */ 1455 VM_SLEEP); 1456 1457 /* Even with SLEEP flag, vmem_create() can fail */ 1458 if (immu->immu_did_arena == NULL) { 1459 ddi_err(DER_PANIC, NULL, "%s: Failed to create Intel " 1460 "IOMMU domainid allocator: %s", immu->immu_name, 1461 immu->immu_did_arena_name); 1462 } 1463 } 1464 1465 /* ######################### CONTEXT CODE ################################# */ 1466 1467 static void 1468 context_set(immu_t *immu, domain_t *domain, pgtable_t *root_table, 1469 int bus, int devfunc) 1470 { 1471 pgtable_t *context; 1472 pgtable_t *pgtable_root; 1473 pgtable_t *unity_pgtable_root; 1474 hw_rce_t *hw_rent; 1475 hw_rce_t *hw_cent; 1476 hw_rce_t *ctxp; 1477 int sid; 1478 krw_t rwtype; 1479 boolean_t fill_root; 1480 boolean_t fill_ctx; 1481 1482 ASSERT(immu); 1483 ASSERT(domain); 1484 ASSERT(root_table); 1485 ASSERT(bus >= 0); 1486 ASSERT(devfunc >= 0); 1487 ASSERT(domain->dom_pgtable_root); 1488 1489 pgtable_root = domain->dom_pgtable_root; 1490 1491 ctxp = (hw_rce_t *)(root_table->swpg_next_array); 1492 context = *(pgtable_t **)(ctxp + bus); 1493 hw_rent = (hw_rce_t *)(root_table->hwpg_vaddr) + bus; 1494 1495 fill_root = B_FALSE; 1496 fill_ctx = B_FALSE; 1497 1498 /* Check the most common case first with reader lock */ 1499 rw_enter(&(immu->immu_ctx_rwlock), RW_READER); 1500 rwtype = RW_READER; 1501 again: 1502 if (ROOT_GET_P(hw_rent)) { 1503 ASSERT(ROOT_GET_CONT(hw_rent) == context->hwpg_paddr); 1504 hw_cent = (hw_rce_t *)(context->hwpg_vaddr) + devfunc; 1505 if (CONT_GET_AVAIL(hw_cent) == IMMU_CONT_INITED) { 1506 ASSERT(CONT_GET_P(hw_cent)); 1507 ASSERT(CONT_GET_DID(hw_cent) == domain->dom_did); 1508 ASSERT(CONT_GET_AW(hw_cent) == immu->immu_dvma_agaw); 1509 ASSERT(CONT_GET_TTYPE(hw_cent) == TTYPE_XLATE_ONLY); 1510 ASSERT(CONT_GET_ASR(hw_cent) == 1511 pgtable_root->hwpg_paddr); 1512 rw_exit(&(immu->immu_ctx_rwlock)); 1513 return; 1514 } else { 1515 fill_ctx = B_TRUE; 1516 } 1517 } else { 1518 fill_root = B_TRUE; 1519 fill_ctx = B_TRUE; 1520 } 1521 1522 if (rwtype == RW_READER && 1523 rw_tryupgrade(&(immu->immu_ctx_rwlock)) == 0) { 1524 rw_exit(&(immu->immu_ctx_rwlock)); 1525 rw_enter(&(immu->immu_ctx_rwlock), RW_WRITER); 1526 rwtype = RW_WRITER; 1527 goto again; 1528 } 1529 rwtype = RW_WRITER; 1530 1531 if (fill_root == B_TRUE) { 1532 ROOT_SET_CONT(hw_rent, context->hwpg_paddr); 1533 ROOT_SET_P(hw_rent); 1534 immu_regs_cpu_flush(immu, (caddr_t)hw_rent, sizeof (hw_rce_t)); 1535 } 1536 1537 if (fill_ctx == B_TRUE) { 1538 hw_cent = (hw_rce_t *)(context->hwpg_vaddr) + devfunc; 1539 unity_pgtable_root = immu->immu_unity_domain->dom_pgtable_root; 1540 ASSERT(CONT_GET_AVAIL(hw_cent) == IMMU_CONT_UNINITED); 1541 ASSERT(CONT_GET_P(hw_cent)); 1542 ASSERT(CONT_GET_DID(hw_cent) == 1543 immu->immu_unity_domain->dom_did); 1544 ASSERT(CONT_GET_AW(hw_cent) == immu->immu_dvma_agaw); 1545 ASSERT(CONT_GET_TTYPE(hw_cent) == TTYPE_XLATE_ONLY); 1546 ASSERT(CONT_GET_ASR(hw_cent) == 1547 unity_pgtable_root->hwpg_paddr); 1548 1549 /* need to disable context entry before reprogramming it */ 1550 bzero(hw_cent, sizeof (hw_rce_t)); 1551 1552 /* flush caches */ 1553 immu_regs_cpu_flush(immu, (caddr_t)hw_cent, sizeof (hw_rce_t)); 1554 ASSERT(rw_write_held(&(immu->immu_ctx_rwlock))); 1555 1556 sid = ((bus << 8) | devfunc); 1557 immu_flush_context_fsi(immu, 0, sid, domain->dom_did); 1558 1559 immu_regs_wbf_flush(immu); 1560 1561 CONT_SET_AVAIL(hw_cent, IMMU_CONT_INITED); 1562 CONT_SET_DID(hw_cent, domain->dom_did); 1563 CONT_SET_AW(hw_cent, immu->immu_dvma_agaw); 1564 CONT_SET_ASR(hw_cent, pgtable_root->hwpg_paddr); 1565 /*LINTED*/ 1566 CONT_SET_TTYPE(hw_cent, TTYPE_XLATE_ONLY); 1567 CONT_SET_P(hw_cent); 1568 immu_regs_cpu_flush(immu, (caddr_t)hw_cent, sizeof (hw_rce_t)); 1569 } 1570 rw_exit(&(immu->immu_ctx_rwlock)); 1571 } 1572 1573 static pgtable_t * 1574 context_create(immu_t *immu) 1575 { 1576 int bus; 1577 int devfunc; 1578 pgtable_t *root_table; 1579 pgtable_t *context; 1580 pgtable_t *pgtable_root; 1581 hw_rce_t *ctxp; 1582 hw_rce_t *hw_rent; 1583 hw_rce_t *hw_cent; 1584 1585 /* Allocate a zeroed root table (4K 256b entries) */ 1586 root_table = pgtable_alloc(immu, IMMU_FLAGS_SLEEP); 1587 pgtable_zero(immu, root_table); 1588 1589 /* 1590 * Setup context tables for all possible root table entries. 1591 * Start out with unity domains for all entries. 1592 */ 1593 ctxp = (hw_rce_t *)(root_table->swpg_next_array); 1594 hw_rent = (hw_rce_t *)(root_table->hwpg_vaddr); 1595 for (bus = 0; bus < IMMU_ROOT_NUM; bus++, ctxp++, hw_rent++) { 1596 context = pgtable_alloc(immu, IMMU_FLAGS_SLEEP); 1597 pgtable_zero(immu, context); 1598 ASSERT(ROOT_GET_P(hw_rent) == 0); 1599 ROOT_SET_P(hw_rent); 1600 ROOT_SET_CONT(hw_rent, context->hwpg_paddr); 1601 hw_cent = (hw_rce_t *)(context->hwpg_vaddr); 1602 for (devfunc = 0; devfunc < IMMU_CONT_NUM; 1603 devfunc++, hw_cent++) { 1604 ASSERT(CONT_GET_P(hw_cent) == 0); 1605 pgtable_root = 1606 immu->immu_unity_domain->dom_pgtable_root; 1607 CONT_SET_DID(hw_cent, 1608 immu->immu_unity_domain->dom_did); 1609 CONT_SET_AW(hw_cent, immu->immu_dvma_agaw); 1610 CONT_SET_ASR(hw_cent, pgtable_root->hwpg_paddr); 1611 /*LINTED*/ 1612 CONT_SET_TTYPE(hw_cent, TTYPE_XLATE_ONLY); 1613 CONT_SET_AVAIL(hw_cent, IMMU_CONT_UNINITED); 1614 CONT_SET_P(hw_cent); 1615 } 1616 immu_regs_cpu_flush(immu, context->hwpg_vaddr, IMMU_PAGESIZE); 1617 *((pgtable_t **)ctxp) = context; 1618 } 1619 immu_regs_cpu_flush(immu, root_table->hwpg_vaddr, IMMU_PAGESIZE); 1620 1621 return (root_table); 1622 } 1623 1624 /* 1625 * Called during rootnex attach, so no locks needed 1626 */ 1627 static void 1628 context_init(immu_t *immu) 1629 { 1630 ASSERT(immu); 1631 ASSERT(immu->immu_ctx_root == NULL); 1632 1633 rw_init(&(immu->immu_ctx_rwlock), NULL, RW_DEFAULT, NULL); 1634 1635 immu_regs_wbf_flush(immu); 1636 1637 immu->immu_ctx_root = context_create(immu); 1638 1639 immu_regs_set_root_table(immu); 1640 1641 rw_enter(&(immu->immu_ctx_rwlock), RW_WRITER); 1642 immu_flush_context_gbl(immu); 1643 rw_exit(&(immu->immu_ctx_rwlock)); 1644 immu_flush_iotlb_gbl(immu); 1645 immu_regs_wbf_flush(immu); 1646 } 1647 1648 1649 /* 1650 * Find top pcib 1651 */ 1652 static int 1653 find_top_pcib(dev_info_t *dip, void *arg) 1654 { 1655 immu_devi_t *immu_devi; 1656 dev_info_t **pcibdipp = (dev_info_t **)arg; 1657 1658 ASSERT(dip); 1659 1660 immu_devi = immu_devi_get(dip); 1661 ASSERT(immu_devi); 1662 1663 if (immu_devi->imd_pcib_type == IMMU_PCIB_PCI_PCI) { 1664 *pcibdipp = dip; 1665 } 1666 1667 return (DDI_WALK_CONTINUE); 1668 } 1669 1670 static int 1671 immu_context_update(immu_t *immu, domain_t *domain, dev_info_t *ddip, 1672 dev_info_t *rdip, immu_flags_t immu_flags) 1673 { 1674 immu_devi_t *r_immu_devi; 1675 immu_devi_t *d_immu_devi; 1676 int r_bus; 1677 int d_bus; 1678 int r_devfunc; 1679 int d_devfunc; 1680 immu_pcib_t d_pcib_type; 1681 immu_pcib_t r_pcib_type; 1682 dev_info_t *pcibdip; 1683 1684 if (ddip == NULL || rdip == NULL || 1685 ddip == root_devinfo || rdip == root_devinfo) { 1686 ddi_err(DER_MODE, rdip, "immu_contexts_update: domain-dip or " 1687 "request-dip are NULL or are root devinfo"); 1688 return (DDI_FAILURE); 1689 } 1690 1691 /* 1692 * We need to set the context fields 1693 * based on what type of device rdip and ddip are. 1694 * To do that we need the immu_devi field. 1695 * Set the immu_devi field (if not already set) 1696 */ 1697 if (immu_devi_set(ddip, immu_flags) == DDI_FAILURE) { 1698 ddi_err(DER_MODE, rdip, 1699 "immu_context_update: failed to set immu_devi for ddip"); 1700 return (DDI_FAILURE); 1701 } 1702 1703 if (immu_devi_set(rdip, immu_flags) == DDI_FAILURE) { 1704 ddi_err(DER_MODE, rdip, 1705 "immu_context_update: failed to set immu_devi for rdip"); 1706 return (DDI_FAILURE); 1707 } 1708 1709 d_immu_devi = immu_devi_get(ddip); 1710 r_immu_devi = immu_devi_get(rdip); 1711 ASSERT(r_immu_devi); 1712 ASSERT(d_immu_devi); 1713 1714 d_bus = d_immu_devi->imd_bus; 1715 d_devfunc = d_immu_devi->imd_devfunc; 1716 d_pcib_type = d_immu_devi->imd_pcib_type; 1717 r_bus = r_immu_devi->imd_bus; 1718 r_devfunc = r_immu_devi->imd_devfunc; 1719 r_pcib_type = r_immu_devi->imd_pcib_type; 1720 1721 ASSERT(d_bus >= 0); 1722 1723 if (rdip == ddip) { 1724 ASSERT(d_pcib_type == IMMU_PCIB_ENDPOINT || 1725 d_pcib_type == IMMU_PCIB_PCIE_PCIE); 1726 ASSERT(r_bus >= 0); 1727 ASSERT(r_devfunc >= 0); 1728 /* rdip is a PCIE device. set context for it only */ 1729 context_set(immu, domain, immu->immu_ctx_root, r_bus, 1730 r_devfunc); 1731 #ifdef BUGGY_DRIVERS 1732 } else if (r_immu_devi == d_immu_devi) { 1733 #ifdef TEST 1734 ddi_err(DER_WARN, rdip, "Driver bug: Devices 0x%lx and " 1735 "0x%lx are identical", rdip, ddip); 1736 #endif 1737 ASSERT(d_pcib_type == IMMU_PCIB_ENDPOINT); 1738 ASSERT(r_bus >= 0); 1739 ASSERT(r_devfunc >= 0); 1740 /* rdip is a PCIE device. set context for it only */ 1741 context_set(immu, domain, immu->immu_ctx_root, r_bus, 1742 r_devfunc); 1743 #endif 1744 } else if (d_pcib_type == IMMU_PCIB_PCIE_PCI) { 1745 /* 1746 * ddip is a PCIE_PCI bridge. Set context for ddip's 1747 * secondary bus. If rdip is on ddip's secondary 1748 * bus, set context for rdip. Else, set context 1749 * for rdip's PCI bridge on ddip's secondary bus. 1750 */ 1751 context_set(immu, domain, immu->immu_ctx_root, 1752 d_immu_devi->imd_sec, 0); 1753 if (d_immu_devi->imd_sec == r_bus) { 1754 context_set(immu, domain, immu->immu_ctx_root, 1755 r_bus, r_devfunc); 1756 } else { 1757 pcibdip = NULL; 1758 if (immu_walk_ancestor(rdip, ddip, find_top_pcib, 1759 &pcibdip, NULL, immu_flags) == DDI_SUCCESS && 1760 pcibdip != NULL) { 1761 ASSERT(pcibdip); 1762 r_immu_devi = immu_devi_get(pcibdip); 1763 ASSERT(d_immu_devi); 1764 ASSERT(d_immu_devi->imd_pcib_type == 1765 IMMU_PCIB_PCI_PCI); 1766 r_bus = r_immu_devi->imd_bus; 1767 r_devfunc = r_immu_devi->imd_devfunc; 1768 context_set(immu, domain, immu->immu_ctx_root, 1769 r_bus, r_devfunc); 1770 } else { 1771 ddi_err(DER_PANIC, rdip, "Failed to find PCI " 1772 " bridge for PCI device"); 1773 /*NOTREACHED*/ 1774 } 1775 } 1776 } else if (d_pcib_type == IMMU_PCIB_PCI_PCI) { 1777 context_set(immu, domain, immu->immu_ctx_root, d_bus, 1778 d_devfunc); 1779 } else if (d_pcib_type == IMMU_PCIB_ENDPOINT) { 1780 ASSERT(r_pcib_type == IMMU_PCIB_NOBDF); 1781 /* 1782 * ddip is a PCIE device which has a non-PCI device under it 1783 * i.e. it is a PCI-nonPCI bridge. Example: pciicde-ata 1784 */ 1785 context_set(immu, domain, immu->immu_ctx_root, d_bus, 1786 d_devfunc); 1787 } else { 1788 ddi_err(DER_PANIC, rdip, "unknown device type. Cannot " 1789 "set IMMU context."); 1790 /*NOTREACHED*/ 1791 } 1792 1793 /* XXX do we need a membar_producer() here */ 1794 return (DDI_SUCCESS); 1795 } 1796 1797 /* ##################### END CONTEXT CODE ################################## */ 1798 /* ##################### MAPPING CODE ################################## */ 1799 1800 1801 static boolean_t 1802 PDTE_check(immu_t *immu, hw_pdte_t pdte, pgtable_t *next, paddr_t paddr, 1803 dev_info_t *rdip, immu_flags_t immu_flags) 1804 { 1805 if (immu_flags & IMMU_FLAGS_PAGE1) { 1806 ASSERT(paddr == 0); 1807 } else { 1808 ASSERT((next == NULL) ^ (paddr == 0)); 1809 } 1810 1811 /* The PDTE must be set i.e. present bit is set */ 1812 if (!PDTE_P(pdte)) { 1813 ddi_err(DER_MODE, rdip, "No present flag"); 1814 return (B_FALSE); 1815 } 1816 1817 /* 1818 * Just assert to check most significant system software field 1819 * (PDTE_SW4) as it is same as present bit and we 1820 * checked that above 1821 */ 1822 ASSERT(PDTE_SW4(pdte)); 1823 1824 /* 1825 * TM field should be clear if not reserved. 1826 * non-leaf is always reserved 1827 */ 1828 if (next == NULL && immu->immu_TM_reserved == B_FALSE) { 1829 if (PDTE_TM(pdte)) { 1830 ddi_err(DER_MODE, rdip, "TM flag set"); 1831 return (B_FALSE); 1832 } 1833 } 1834 1835 /* 1836 * The SW3 field is not used and must be clear 1837 */ 1838 if (PDTE_SW3(pdte)) { 1839 ddi_err(DER_MODE, rdip, "SW3 set"); 1840 return (B_FALSE); 1841 } 1842 1843 /* 1844 * PFN (for PTE) or next level pgtable-paddr (for PDE) must be set 1845 */ 1846 if (next == NULL) { 1847 ASSERT(paddr % IMMU_PAGESIZE == 0); 1848 if (PDTE_PADDR(pdte) != paddr) { 1849 ddi_err(DER_MODE, rdip, 1850 "PTE paddr mismatch: %lx != %lx", 1851 PDTE_PADDR(pdte), paddr); 1852 return (B_FALSE); 1853 } 1854 } else { 1855 if (PDTE_PADDR(pdte) != next->hwpg_paddr) { 1856 ddi_err(DER_MODE, rdip, 1857 "PDE paddr mismatch: %lx != %lx", 1858 PDTE_PADDR(pdte), next->hwpg_paddr); 1859 return (B_FALSE); 1860 } 1861 } 1862 1863 /* 1864 * SNP field should be clear if not reserved. 1865 * non-leaf is always reserved 1866 */ 1867 if (next == NULL && immu->immu_SNP_reserved == B_FALSE) { 1868 if (PDTE_SNP(pdte)) { 1869 ddi_err(DER_MODE, rdip, "SNP set"); 1870 return (B_FALSE); 1871 } 1872 } 1873 1874 /* second field available for system software should be clear */ 1875 if (PDTE_SW2(pdte)) { 1876 ddi_err(DER_MODE, rdip, "SW2 set"); 1877 return (B_FALSE); 1878 } 1879 1880 /* Super pages field should be clear */ 1881 if (PDTE_SP(pdte)) { 1882 ddi_err(DER_MODE, rdip, "SP set"); 1883 return (B_FALSE); 1884 } 1885 1886 /* 1887 * least significant field available for 1888 * system software should be clear 1889 */ 1890 if (PDTE_SW1(pdte)) { 1891 ddi_err(DER_MODE, rdip, "SW1 set"); 1892 return (B_FALSE); 1893 } 1894 1895 if ((immu_flags & IMMU_FLAGS_READ) && !PDTE_READ(pdte)) { 1896 ddi_err(DER_MODE, rdip, "READ not set"); 1897 return (B_FALSE); 1898 } 1899 1900 if ((immu_flags & IMMU_FLAGS_WRITE) && !PDTE_WRITE(pdte)) { 1901 ddi_err(DER_MODE, rdip, "WRITE not set"); 1902 return (B_FALSE); 1903 } 1904 1905 return (B_TRUE); 1906 } 1907 /*ARGSUSED*/ 1908 static void 1909 PTE_clear_all(immu_t *immu, domain_t *domain, xlate_t *xlate, 1910 uint64_t *dvma_ptr, uint64_t *npages_ptr, dev_info_t *rdip) 1911 { 1912 uint64_t npages; 1913 uint64_t dvma; 1914 pgtable_t *pgtable; 1915 hw_pdte_t *hwp; 1916 hw_pdte_t *shwp; 1917 int idx; 1918 hw_pdte_t pte; 1919 1920 ASSERT(xlate->xlt_level == 1); 1921 1922 pgtable = xlate->xlt_pgtable; 1923 idx = xlate->xlt_idx; 1924 1925 ASSERT(pgtable); 1926 ASSERT(idx <= IMMU_PGTABLE_MAXIDX); 1927 1928 dvma = *dvma_ptr; 1929 npages = *npages_ptr; 1930 1931 ASSERT(dvma); 1932 ASSERT(dvma % IMMU_PAGESIZE == 0); 1933 ASSERT(npages); 1934 1935 /* 1936 * since a caller gets a unique dvma for a physical address, 1937 * no other concurrent thread will be writing to the same 1938 * PTE even if it has the same paddr. So no locks needed. 1939 */ 1940 shwp = (hw_pdte_t *)(pgtable->hwpg_vaddr) + idx; 1941 1942 hwp = shwp; 1943 for (; npages > 0 && idx <= IMMU_PGTABLE_MAXIDX; idx++, hwp++) { 1944 1945 pte = *hwp; 1946 1947 /* Cannot clear a HW PTE that is aleady clear */ 1948 ASSERT(PDTE_P(pte)); 1949 PDTE_CLEAR_P(pte); 1950 *hwp = pte; 1951 1952 dvma += IMMU_PAGESIZE; 1953 npages--; 1954 } 1955 1956 1957 #ifdef TEST 1958 /* dont need to flush write during unmap */ 1959 immu_regs_cpu_flush(immu, (caddr_t)shwp, 1960 (hwp - shwp) * sizeof (hw_pdte_t)); 1961 #endif 1962 1963 *dvma_ptr = dvma; 1964 *npages_ptr = npages; 1965 1966 xlate->xlt_idx = idx; 1967 } 1968 1969 /*ARGSUSED*/ 1970 static void 1971 xlate_setup(immu_t *immu, uint64_t dvma, xlate_t *xlate, 1972 int nlevels, dev_info_t *rdip) 1973 { 1974 int level; 1975 uint64_t offbits; 1976 1977 /* level 0 is never used. Sanity check */ 1978 ASSERT(xlate->xlt_level == 0); 1979 ASSERT(xlate->xlt_idx == 0); 1980 ASSERT(xlate->xlt_pgtable == NULL); 1981 ASSERT(dvma % IMMU_PAGESIZE == 0); 1982 1983 /* 1984 * Skip the first 12 bits which is the offset into 1985 * 4K PFN (phys page frame based on IMMU_PAGESIZE) 1986 */ 1987 offbits = dvma >> IMMU_PAGESHIFT; 1988 1989 /* skip to level 1 i.e. leaf PTE */ 1990 for (level = 1, xlate++; level <= nlevels; level++, xlate++) { 1991 xlate->xlt_level = level; 1992 xlate->xlt_idx = (offbits & IMMU_PGTABLE_LEVEL_MASK); 1993 ASSERT(xlate->xlt_idx <= IMMU_PGTABLE_MAXIDX); 1994 xlate->xlt_pgtable = NULL; 1995 offbits >>= IMMU_PGTABLE_LEVEL_STRIDE; 1996 } 1997 } 1998 1999 /* 2000 * Read the pgtables 2001 */ 2002 static void 2003 PDE_lookup(immu_t *immu, domain_t *domain, xlate_t *xlate, int nlevels, 2004 dev_info_t *rdip) 2005 { 2006 pgtable_t *pgtable; 2007 pgtable_t *next; 2008 hw_pdte_t pde; 2009 uint_t idx; 2010 2011 /* xlate should be at level 0 */ 2012 ASSERT(xlate->xlt_level == 0); 2013 ASSERT(xlate->xlt_idx == 0); 2014 2015 /* start with highest level pgtable i.e. root */ 2016 xlate += nlevels; 2017 ASSERT(xlate->xlt_level == nlevels); 2018 2019 if (xlate->xlt_pgtable == NULL) { 2020 xlate->xlt_pgtable = domain->dom_pgtable_root; 2021 } 2022 2023 for (; xlate->xlt_level > 1; xlate--) { 2024 2025 idx = xlate->xlt_idx; 2026 pgtable = xlate->xlt_pgtable; 2027 2028 ASSERT(pgtable); 2029 ASSERT(idx <= IMMU_PGTABLE_MAXIDX); 2030 2031 if ((xlate - 1)->xlt_pgtable) { 2032 continue; 2033 } 2034 2035 /* xlate's leafier level is not set, set it now */ 2036 2037 /* Lock the pgtable in read mode */ 2038 rw_enter(&(pgtable->swpg_rwlock), RW_READER); 2039 2040 /* 2041 * since we are unmapping, the pgtable should 2042 * already point to a leafier pgtable. 2043 */ 2044 next = *(pgtable->swpg_next_array + idx); 2045 ASSERT(next); 2046 2047 pde = *((hw_pdte_t *)(pgtable->hwpg_vaddr) + idx); 2048 2049 ASSERT(PDTE_check(immu, pde, next, 0, rdip, 0) == B_TRUE); 2050 2051 (xlate - 1)->xlt_pgtable = next; 2052 2053 rw_exit(&(pgtable->swpg_rwlock)); 2054 } 2055 } 2056 2057 /*ARGSUSED*/ 2058 static void 2059 PTE_set_one(immu_t *immu, hw_pdte_t *hwp, paddr_t paddr, 2060 dev_info_t *rdip, immu_flags_t immu_flags) 2061 { 2062 hw_pdte_t pte; 2063 2064 pte = *hwp; 2065 2066 #ifndef DEBUG 2067 /* Set paddr */ 2068 ASSERT(paddr % IMMU_PAGESIZE == 0); 2069 pte = 0; 2070 PDTE_SET_PADDR(pte, paddr); 2071 PDTE_SET_READ(pte); 2072 PDTE_SET_WRITE(pte); 2073 *hwp = pte; 2074 #else 2075 2076 if (PDTE_P(pte)) { 2077 if (PDTE_PADDR(pte) != paddr) { 2078 ddi_err(DER_MODE, rdip, "PTE paddr %lx != paddr %lx", 2079 PDTE_PADDR(pte), paddr); 2080 } 2081 #ifdef BUGGY_DRIVERS 2082 return; 2083 #else 2084 goto out; 2085 #endif 2086 } 2087 2088 /* Don't touch SW4. It is the present field */ 2089 2090 /* clear TM field if not reserved */ 2091 if (immu->immu_TM_reserved == B_FALSE) { 2092 PDTE_CLEAR_TM(pte); 2093 } 2094 2095 #ifdef DEBUG 2096 /* Clear 3rd field for system software - not used */ 2097 PDTE_CLEAR_SW3(pte); 2098 #endif 2099 2100 /* Set paddr */ 2101 ASSERT(paddr % IMMU_PAGESIZE == 0); 2102 PDTE_CLEAR_PADDR(pte); 2103 PDTE_SET_PADDR(pte, paddr); 2104 2105 /* clear SNP field if not reserved. */ 2106 if (immu->immu_SNP_reserved == B_FALSE) { 2107 PDTE_CLEAR_SNP(pte); 2108 } 2109 2110 #ifdef DEBUG 2111 /* Clear SW2 field available for software */ 2112 PDTE_CLEAR_SW2(pte); 2113 #endif 2114 2115 2116 #ifdef DEBUG 2117 /* SP is don't care for PTEs. Clear it for cleanliness */ 2118 PDTE_CLEAR_SP(pte); 2119 #endif 2120 2121 #ifdef DEBUG 2122 /* Clear SW1 field available for software */ 2123 PDTE_CLEAR_SW1(pte); 2124 #endif 2125 2126 /* 2127 * Now that we are done writing the PTE 2128 * set the "present" flag. Note this present 2129 * flag is a bit in the PDE/PTE that the 2130 * spec says is available for system software. 2131 * This is an implementation detail of Solaris 2132 * bare-metal Intel IOMMU. 2133 * The present field in a PDE/PTE is not defined 2134 * by the Vt-d spec 2135 */ 2136 2137 PDTE_SET_P(pte); 2138 2139 out: 2140 #ifdef BUGGY_DRIVERS 2141 PDTE_SET_READ(pte); 2142 PDTE_SET_WRITE(pte); 2143 #else 2144 if (immu_flags & IMMU_FLAGS_READ) 2145 PDTE_SET_READ(pte); 2146 if (immu_flags & IMMU_FLAGS_WRITE) 2147 PDTE_SET_WRITE(pte); 2148 #endif 2149 2150 *hwp = pte; 2151 #endif 2152 } 2153 2154 /*ARGSUSED*/ 2155 static void 2156 PTE_set_all(immu_t *immu, domain_t *domain, xlate_t *xlate, 2157 uint64_t *dvma_ptr, uint64_t *nvpages_ptr, dcookie_t *dcookies, 2158 int dcount, dev_info_t *rdip, immu_flags_t immu_flags) 2159 { 2160 paddr_t paddr; 2161 uint64_t nvpages; 2162 uint64_t nppages; 2163 uint64_t dvma; 2164 pgtable_t *pgtable; 2165 hw_pdte_t *hwp; 2166 hw_pdte_t *shwp; 2167 int idx; 2168 int j; 2169 2170 ASSERT(xlate->xlt_level == 1); 2171 2172 pgtable = xlate->xlt_pgtable; 2173 idx = xlate->xlt_idx; 2174 2175 ASSERT(idx <= IMMU_PGTABLE_MAXIDX); 2176 ASSERT(pgtable); 2177 2178 dvma = *dvma_ptr; 2179 nvpages = *nvpages_ptr; 2180 2181 ASSERT(dvma || (immu_flags & IMMU_FLAGS_PAGE1)); 2182 ASSERT(nvpages); 2183 2184 /* 2185 * since a caller gets a unique dvma for a physical address, 2186 * no other concurrent thread will be writing to the same 2187 * PTE even if it has the same paddr. So no locks needed. 2188 */ 2189 shwp = (hw_pdte_t *)(pgtable->hwpg_vaddr) + idx; 2190 2191 hwp = shwp; 2192 for (j = dcount - 1; j >= 0; j--) { 2193 if (nvpages <= dcookies[j].dck_npages) 2194 break; 2195 nvpages -= dcookies[j].dck_npages; 2196 } 2197 2198 ASSERT(j >= 0); 2199 ASSERT(nvpages); 2200 ASSERT(nvpages <= dcookies[j].dck_npages); 2201 nppages = nvpages; 2202 paddr = dcookies[j].dck_paddr + 2203 (dcookies[j].dck_npages - nppages) * IMMU_PAGESIZE; 2204 2205 nvpages = *nvpages_ptr; 2206 for (; nvpages > 0 && idx <= IMMU_PGTABLE_MAXIDX; idx++, hwp++) { 2207 2208 ASSERT(paddr || (immu_flags & IMMU_FLAGS_PAGE1)); 2209 2210 PTE_set_one(immu, hwp, paddr, rdip, immu_flags); 2211 2212 ASSERT(PDTE_check(immu, *hwp, NULL, paddr, rdip, immu_flags) 2213 == B_TRUE); 2214 nppages--; 2215 nvpages--; 2216 paddr += IMMU_PAGESIZE; 2217 dvma += IMMU_PAGESIZE; 2218 2219 if (nppages == 0) { 2220 j++; 2221 } 2222 2223 if (j == dcount) { 2224 ASSERT(nvpages == 0); 2225 break; 2226 } 2227 2228 ASSERT(nvpages); 2229 if (nppages == 0) { 2230 nppages = dcookies[j].dck_npages; 2231 paddr = dcookies[j].dck_paddr; 2232 } 2233 } 2234 2235 /* flush writes to HW PTE table */ 2236 immu_regs_cpu_flush(immu, (caddr_t)shwp, (hwp - shwp) * 2237 sizeof (hw_pdte_t)); 2238 2239 if (nvpages) { 2240 *dvma_ptr = dvma; 2241 *nvpages_ptr = nvpages; 2242 } else { 2243 *dvma_ptr = 0; 2244 *nvpages_ptr = 0; 2245 } 2246 2247 xlate->xlt_idx = idx; 2248 } 2249 2250 /*ARGSUSED*/ 2251 static void 2252 PDE_set_one(immu_t *immu, hw_pdte_t *hwp, pgtable_t *next, 2253 dev_info_t *rdip, immu_flags_t immu_flags) 2254 { 2255 hw_pdte_t pde; 2256 2257 pde = *hwp; 2258 2259 /* if PDE is already set, make sure it is correct */ 2260 if (PDTE_P(pde)) { 2261 ASSERT(PDTE_PADDR(pde) == next->hwpg_paddr); 2262 #ifdef BUGGY_DRIVERS 2263 return; 2264 #else 2265 goto out; 2266 #endif 2267 } 2268 2269 /* Dont touch SW4, it is the present bit */ 2270 2271 /* don't touch TM field it is reserved for PDEs */ 2272 2273 /* 3rd field available for system software is not used */ 2274 PDTE_CLEAR_SW3(pde); 2275 2276 /* Set next level pgtable-paddr for PDE */ 2277 ASSERT(next->hwpg_paddr % IMMU_PAGESIZE == 0); 2278 PDTE_CLEAR_PADDR(pde); 2279 PDTE_SET_PADDR(pde, next->hwpg_paddr); 2280 2281 /* don't touch SNP field it is reserved for PDEs */ 2282 2283 /* Clear second field available for system software */ 2284 PDTE_CLEAR_SW2(pde); 2285 2286 /* No super pages for PDEs */ 2287 PDTE_CLEAR_SP(pde); 2288 2289 /* Clear SW1 for software */ 2290 PDTE_CLEAR_SW1(pde); 2291 2292 /* 2293 * Now that we are done writing the PDE 2294 * set the "present" flag. Note this present 2295 * flag is a bit in the PDE/PTE that the 2296 * spec says is available for system software. 2297 * This is an implementation detail of Solaris 2298 * base-metal Intel IOMMU. 2299 * The present field in a PDE/PTE is not defined 2300 * by the Vt-d spec 2301 */ 2302 2303 out: 2304 #ifdef BUGGY_DRIVERS 2305 PDTE_SET_READ(pde); 2306 PDTE_SET_WRITE(pde); 2307 #else 2308 if (immu_flags & IMMU_FLAGS_READ) 2309 PDTE_SET_READ(pde); 2310 if (immu_flags & IMMU_FLAGS_WRITE) 2311 PDTE_SET_WRITE(pde); 2312 #endif 2313 2314 PDTE_SET_P(pde); 2315 2316 *hwp = pde; 2317 2318 immu_regs_cpu_flush(immu, (caddr_t)hwp, sizeof (hw_pdte_t)); 2319 } 2320 2321 /* 2322 * Used to set PDEs 2323 */ 2324 static boolean_t 2325 PDE_set_all(immu_t *immu, domain_t *domain, xlate_t *xlate, int nlevels, 2326 dev_info_t *rdip, immu_flags_t immu_flags) 2327 { 2328 pgtable_t *pgtable; 2329 pgtable_t *new; 2330 pgtable_t *next; 2331 hw_pdte_t *hwp; 2332 int level; 2333 uint_t idx; 2334 krw_t rwtype; 2335 boolean_t set = B_FALSE; 2336 2337 /* xlate should be at level 0 */ 2338 ASSERT(xlate->xlt_level == 0); 2339 ASSERT(xlate->xlt_idx == 0); 2340 2341 /* start with highest level pgtable i.e. root */ 2342 xlate += nlevels; 2343 ASSERT(xlate->xlt_level == nlevels); 2344 2345 new = NULL; 2346 xlate->xlt_pgtable = domain->dom_pgtable_root; 2347 for (level = nlevels; level > 1; level--, xlate--) { 2348 2349 ASSERT(xlate->xlt_level == level); 2350 2351 idx = xlate->xlt_idx; 2352 pgtable = xlate->xlt_pgtable; 2353 2354 ASSERT(pgtable); 2355 ASSERT(idx <= IMMU_PGTABLE_MAXIDX); 2356 2357 /* speculative alloc */ 2358 if (new == NULL) { 2359 new = pgtable_alloc(immu, immu_flags); 2360 if (new == NULL) { 2361 ddi_err(DER_PANIC, rdip, "pgtable alloc err"); 2362 } 2363 } 2364 2365 /* Lock the pgtable in READ mode first */ 2366 rw_enter(&(pgtable->swpg_rwlock), RW_READER); 2367 rwtype = RW_READER; 2368 again: 2369 hwp = (hw_pdte_t *)(pgtable->hwpg_vaddr) + idx; 2370 2371 ASSERT(pgtable->swpg_next_array); 2372 2373 next = (pgtable->swpg_next_array)[idx]; 2374 2375 /* 2376 * check if leafier level already has a pgtable 2377 * if yes, verify 2378 */ 2379 if (next == NULL) { 2380 /* Change to a write lock */ 2381 if (rwtype == RW_READER && 2382 rw_tryupgrade(&(pgtable->swpg_rwlock)) == 0) { 2383 rw_exit(&(pgtable->swpg_rwlock)); 2384 rw_enter(&(pgtable->swpg_rwlock), RW_WRITER); 2385 rwtype = RW_WRITER; 2386 goto again; 2387 } 2388 rwtype = RW_WRITER; 2389 pgtable_zero(immu, new); 2390 next = new; 2391 new = NULL; 2392 (pgtable->swpg_next_array)[idx] = next; 2393 PDE_set_one(immu, hwp, next, rdip, immu_flags); 2394 set = B_TRUE; 2395 rw_downgrade(&(pgtable->swpg_rwlock)); 2396 rwtype = RW_READER; 2397 } else { 2398 hw_pdte_t pde = *hwp; 2399 2400 #ifndef BUGGY_DRIVERS 2401 /* 2402 * If buggy driver we already set permission 2403 * READ+WRITE so nothing to do for that case 2404 * XXX Check that read writer perms change before 2405 * actually setting perms. Also need to hold lock 2406 */ 2407 if (immu_flags & IMMU_FLAGS_READ) 2408 PDTE_SET_READ(pde); 2409 if (immu_flags & IMMU_FLAGS_WRITE) 2410 PDTE_SET_WRITE(pde); 2411 2412 #endif 2413 2414 *hwp = pde; 2415 } 2416 2417 ASSERT(PDTE_check(immu, *hwp, next, 0, rdip, immu_flags) 2418 == B_TRUE); 2419 2420 (xlate - 1)->xlt_pgtable = next; 2421 ASSERT(rwtype == RW_READER); 2422 rw_exit(&(pgtable->swpg_rwlock)); 2423 } 2424 2425 if (new) { 2426 pgtable_free(immu, new); 2427 } 2428 2429 return (set); 2430 } 2431 2432 /* 2433 * dvma_map() 2434 * map a contiguous range of DVMA pages 2435 * 2436 * immu: IOMMU unit for which we are generating DVMA cookies 2437 * domain: domain 2438 * sdvma: Starting dvma 2439 * spaddr: Starting paddr 2440 * npages: Number of pages 2441 * rdip: requesting device 2442 * immu_flags: flags 2443 */ 2444 static boolean_t 2445 dvma_map(immu_t *immu, domain_t *domain, uint64_t sdvma, uint64_t snvpages, 2446 dcookie_t *dcookies, int dcount, dev_info_t *rdip, immu_flags_t immu_flags) 2447 { 2448 uint64_t dvma; 2449 uint64_t n; 2450 int nlevels = immu->immu_dvma_nlevels; 2451 xlate_t xlate[IMMU_PGTABLE_MAX_LEVELS + 1] = {0}; 2452 boolean_t pde_set = B_FALSE; 2453 2454 ASSERT(nlevels <= IMMU_PGTABLE_MAX_LEVELS); 2455 ASSERT(sdvma % IMMU_PAGESIZE == 0); 2456 ASSERT(snvpages); 2457 2458 n = snvpages; 2459 dvma = sdvma; 2460 2461 while (n > 0) { 2462 xlate_setup(immu, dvma, xlate, nlevels, rdip); 2463 2464 /* Lookup or allocate PGDIRs and PGTABLEs if necessary */ 2465 if (PDE_set_all(immu, domain, xlate, nlevels, rdip, immu_flags) 2466 == B_TRUE) { 2467 pde_set = B_TRUE; 2468 } 2469 2470 /* set all matching ptes that fit into this leaf pgtable */ 2471 PTE_set_all(immu, domain, &xlate[1], &dvma, &n, dcookies, 2472 dcount, rdip, immu_flags); 2473 } 2474 2475 return (pde_set); 2476 } 2477 2478 /* 2479 * dvma_unmap() 2480 * unmap a range of DVMAs 2481 * 2482 * immu: IOMMU unit state 2483 * domain: domain for requesting device 2484 * ddip: domain-dip 2485 * dvma: starting DVMA 2486 * npages: Number of IMMU pages to be unmapped 2487 * rdip: requesting device 2488 */ 2489 static void 2490 dvma_unmap(immu_t *immu, domain_t *domain, uint64_t sdvma, uint64_t snpages, 2491 dev_info_t *rdip) 2492 { 2493 int nlevels = immu->immu_dvma_nlevels; 2494 xlate_t xlate[IMMU_PGTABLE_MAX_LEVELS + 1] = {0}; 2495 uint64_t n; 2496 uint64_t dvma; 2497 2498 ASSERT(nlevels <= IMMU_PGTABLE_MAX_LEVELS); 2499 ASSERT(sdvma != 0); 2500 ASSERT(sdvma % IMMU_PAGESIZE == 0); 2501 ASSERT(snpages); 2502 2503 dvma = sdvma; 2504 n = snpages; 2505 2506 while (n > 0) { 2507 /* setup the xlate array */ 2508 xlate_setup(immu, dvma, xlate, nlevels, rdip); 2509 2510 /* just lookup existing pgtables. Should never fail */ 2511 PDE_lookup(immu, domain, xlate, nlevels, rdip); 2512 2513 /* clear all matching ptes that fit into this leaf pgtable */ 2514 PTE_clear_all(immu, domain, &xlate[1], &dvma, &n, rdip); 2515 } 2516 2517 /* No need to flush IOTLB after unmap */ 2518 } 2519 2520 static uint64_t 2521 dvma_alloc(ddi_dma_impl_t *hp, domain_t *domain, uint_t npages) 2522 { 2523 ddi_dma_attr_t *dma_attr; 2524 uint64_t dvma; 2525 size_t xsize, align; 2526 uint64_t minaddr, maxaddr; 2527 2528 ASSERT(domain->dom_maptype != IMMU_MAPTYPE_UNITY); 2529 2530 /* shotcuts */ 2531 dma_attr = &(hp->dmai_attr); 2532 2533 /* parameters */ 2534 xsize = npages * IMMU_PAGESIZE; 2535 align = MAX((size_t)(dma_attr->dma_attr_align), IMMU_PAGESIZE); 2536 minaddr = dma_attr->dma_attr_addr_lo; 2537 maxaddr = dma_attr->dma_attr_addr_hi + 1; 2538 /* nocross is checked in cookie_update() */ 2539 2540 /* handle the rollover cases */ 2541 if (maxaddr < dma_attr->dma_attr_addr_hi) { 2542 maxaddr = dma_attr->dma_attr_addr_hi; 2543 } 2544 2545 /* 2546 * allocate from vmem arena. 2547 */ 2548 dvma = (uint64_t)(uintptr_t)vmem_xalloc(domain->dom_dvma_arena, 2549 xsize, align, 0, 0, (void *)(uintptr_t)minaddr, 2550 (void *)(uintptr_t)maxaddr, VM_NOSLEEP); 2551 2552 ASSERT(dvma); 2553 ASSERT(dvma >= minaddr); 2554 ASSERT(dvma + xsize - 1 < maxaddr); 2555 2556 return (dvma); 2557 } 2558 2559 static void 2560 dvma_free(domain_t *domain, uint64_t dvma, uint64_t npages) 2561 { 2562 uint64_t size = npages * IMMU_PAGESIZE; 2563 2564 ASSERT(domain); 2565 ASSERT(domain->dom_did > 0); 2566 ASSERT(dvma); 2567 ASSERT(npages); 2568 2569 if (domain->dom_maptype != IMMU_MAPTYPE_XLATE) { 2570 ASSERT(domain->dom_maptype == IMMU_MAPTYPE_UNITY); 2571 return; 2572 } 2573 2574 vmem_free(domain->dom_dvma_arena, (void *)(uintptr_t)dvma, size); 2575 } 2576 /*ARGSUSED*/ 2577 static void 2578 cookie_free(rootnex_dma_t *dma, immu_t *immu, domain_t *domain, 2579 dev_info_t *rdip) 2580 { 2581 int i; 2582 uint64_t dvma; 2583 uint64_t npages; 2584 dvcookie_t *dvcookies = dma->dp_dvcookies; 2585 2586 ASSERT(dma->dp_max_cookies); 2587 ASSERT(dma->dp_max_dcookies); 2588 ASSERT(dma->dp_dvmax < dma->dp_max_cookies); 2589 ASSERT(dma->dp_dmax < dma->dp_max_dcookies); 2590 2591 /* 2592 * we allocated DVMA in a single chunk. Calculate total number 2593 * of pages 2594 */ 2595 for (i = 0, npages = 0; i <= dma->dp_dvmax; i++) { 2596 npages += dvcookies[i].dvck_npages; 2597 } 2598 dvma = dvcookies[0].dvck_dvma; 2599 #ifdef DEBUG 2600 /* Unmap only in DEBUG mode */ 2601 dvma_unmap(immu, domain, dvma, npages, rdip); 2602 #endif 2603 dvma_free(domain, dvma, npages); 2604 2605 kmem_free(dma->dp_dvcookies, sizeof (dvcookie_t) * dma->dp_max_cookies); 2606 dma->dp_dvcookies = NULL; 2607 kmem_free(dma->dp_dcookies, sizeof (dcookie_t) * dma->dp_max_dcookies); 2608 dma->dp_dcookies = NULL; 2609 if (dma->dp_need_to_free_cookie == B_TRUE) { 2610 kmem_free(dma->dp_cookies, sizeof (ddi_dma_cookie_t) * 2611 dma->dp_max_cookies); 2612 dma->dp_dcookies = NULL; 2613 dma->dp_need_to_free_cookie = B_FALSE; 2614 } 2615 2616 dma->dp_max_cookies = 0; 2617 dma->dp_max_dcookies = 0; 2618 dma->dp_cookie_size = 0; 2619 dma->dp_dvmax = 0; 2620 dma->dp_dmax = 0; 2621 } 2622 2623 /* 2624 * cookie_alloc() 2625 */ 2626 static int 2627 cookie_alloc(rootnex_dma_t *dma, struct ddi_dma_req *dmareq, 2628 ddi_dma_attr_t *attr, uint_t prealloc) 2629 { 2630 int kmflag; 2631 rootnex_sglinfo_t *sinfo = &(dma->dp_sglinfo); 2632 dvcookie_t *dvcookies = dma->dp_dvcookies; 2633 dcookie_t *dcookies = dma->dp_dcookies; 2634 ddi_dma_cookie_t *cookies = dma->dp_cookies; 2635 uint64_t max_cookies; 2636 uint64_t max_dcookies; 2637 uint64_t cookie_size; 2638 2639 /* we need to allocate new array */ 2640 if (dmareq->dmar_fp == DDI_DMA_SLEEP) { 2641 kmflag = KM_SLEEP; 2642 } else { 2643 kmflag = KM_NOSLEEP; 2644 } 2645 2646 /* 2647 * XXX make sure cookies size doen't exceed sinfo->si_max_cookie_size; 2648 */ 2649 2650 /* 2651 * figure out the rough estimate of array size 2652 * At a minimum, each cookie must hold 1 page. 2653 * At a maximum, it cannot exceed dma_attr_sgllen 2654 */ 2655 max_dcookies = dmareq->dmar_object.dmao_size + IMMU_PAGEOFFSET; 2656 max_dcookies /= IMMU_PAGESIZE; 2657 max_dcookies++; 2658 max_cookies = MIN(max_dcookies, attr->dma_attr_sgllen); 2659 2660 /* allocate the dvma cookie array */ 2661 dvcookies = kmem_zalloc(sizeof (dvcookie_t) * max_cookies, kmflag); 2662 if (dvcookies == NULL) { 2663 return (DDI_FAILURE); 2664 } 2665 2666 /* allocate the "phys" cookie array */ 2667 dcookies = kmem_zalloc(sizeof (dcookie_t) * max_dcookies, kmflag); 2668 if (dcookies == NULL) { 2669 kmem_free(dvcookies, sizeof (dvcookie_t) * max_cookies); 2670 dvcookies = NULL; 2671 return (DDI_FAILURE); 2672 } 2673 2674 /* allocate the "real" cookie array - the one given to users */ 2675 cookie_size = sizeof (ddi_dma_cookie_t) * max_cookies; 2676 if (max_cookies > prealloc) { 2677 cookies = kmem_zalloc(cookie_size, kmflag); 2678 if (cookies == NULL) { 2679 kmem_free(dvcookies, sizeof (dvcookie_t) * max_cookies); 2680 kmem_free(dcookies, sizeof (dcookie_t) * max_dcookies); 2681 goto fail; 2682 } 2683 dma->dp_need_to_free_cookie = B_TRUE; 2684 } else { 2685 /* the preallocated buffer fits this size */ 2686 cookies = (ddi_dma_cookie_t *)dma->dp_prealloc_buffer; 2687 bzero(cookies, sizeof (ddi_dma_cookie_t)* max_cookies); 2688 dma->dp_need_to_free_cookie = B_FALSE; 2689 } 2690 2691 dma->dp_dvcookies = dvcookies; 2692 dma->dp_dcookies = dcookies; 2693 dma->dp_cookies = cookies; 2694 dma->dp_cookie_size = cookie_size; 2695 dma->dp_max_cookies = max_cookies; 2696 dma->dp_max_dcookies = max_dcookies; 2697 dma->dp_dvmax = 0; 2698 dma->dp_dmax = 0; 2699 sinfo->si_max_pages = dma->dp_max_cookies; 2700 2701 return (DDI_SUCCESS); 2702 2703 fail: 2704 dma->dp_dvcookies = NULL; 2705 dma->dp_dcookies = NULL; 2706 dma->dp_cookies = NULL; 2707 dma->dp_cookie_size = 0; 2708 dma->dp_max_cookies = 0; 2709 dma->dp_max_dcookies = 0; 2710 dma->dp_dvmax = 0; 2711 dma->dp_dmax = 0; 2712 dma->dp_need_to_free_cookie = B_FALSE; 2713 sinfo->si_max_pages = 0; 2714 2715 return (DDI_FAILURE); 2716 } 2717 2718 /*ARGSUSED*/ 2719 static void 2720 cookie_update(domain_t *domain, rootnex_dma_t *dma, paddr_t paddr, 2721 int64_t psize, uint64_t maxseg, size_t nocross) 2722 { 2723 dvcookie_t *dvcookies = dma->dp_dvcookies; 2724 dcookie_t *dcookies = dma->dp_dcookies; 2725 ddi_dma_cookie_t *cookies = dma->dp_cookies; 2726 uint64_t dvmax = dma->dp_dvmax; 2727 uint64_t dmax = dma->dp_dmax; 2728 2729 ASSERT(dvmax < dma->dp_max_cookies); 2730 ASSERT(dmax < dma->dp_max_dcookies); 2731 2732 paddr &= IMMU_PAGEMASK; 2733 2734 ASSERT(paddr); 2735 ASSERT(psize); 2736 ASSERT(maxseg); 2737 2738 /* 2739 * check to see if this page would put us 2740 * over the max cookie size. 2741 */ 2742 if (cookies[dvmax].dmac_size + psize > maxseg) { 2743 dvmax++; /* use the next dvcookie */ 2744 dmax++; /* also means we use the next dcookie */ 2745 ASSERT(dvmax < dma->dp_max_cookies); 2746 ASSERT(dmax < dma->dp_max_dcookies); 2747 } 2748 2749 /* 2750 * check to see if this page would make us larger than 2751 * the nocross boundary. If yes, create a new cookie 2752 * otherwise we will fail later with vmem_xalloc() 2753 * due to overconstrained alloc requests 2754 * nocross == 0 implies no nocross constraint. 2755 */ 2756 if (nocross > 0) { 2757 ASSERT((dvcookies[dvmax].dvck_npages) * IMMU_PAGESIZE 2758 <= nocross); 2759 if ((dvcookies[dvmax].dvck_npages + 1) * IMMU_PAGESIZE 2760 > nocross) { 2761 dvmax++; /* use the next dvcookie */ 2762 dmax++; /* also means we use the next dcookie */ 2763 ASSERT(dvmax < dma->dp_max_cookies); 2764 ASSERT(dmax < dma->dp_max_dcookies); 2765 } 2766 ASSERT((dvcookies[dvmax].dvck_npages) * IMMU_PAGESIZE 2767 <= nocross); 2768 } 2769 2770 /* 2771 * If the cookie is empty 2772 */ 2773 if (dvcookies[dvmax].dvck_npages == 0) { 2774 ASSERT(cookies[dvmax].dmac_size == 0); 2775 ASSERT(dvcookies[dvmax].dvck_dvma == 0); 2776 ASSERT(dvcookies[dvmax].dvck_npages 2777 == 0); 2778 ASSERT(dcookies[dmax].dck_paddr == 0); 2779 ASSERT(dcookies[dmax].dck_npages == 0); 2780 2781 dvcookies[dvmax].dvck_dvma = 0; 2782 dvcookies[dvmax].dvck_npages = 1; 2783 dcookies[dmax].dck_paddr = paddr; 2784 dcookies[dmax].dck_npages = 1; 2785 cookies[dvmax].dmac_size = psize; 2786 } else { 2787 /* Cookie not empty. Add to it */ 2788 cookies[dma->dp_dvmax].dmac_size += psize; 2789 ASSERT(dvcookies[dma->dp_dvmax].dvck_dvma == 0); 2790 dvcookies[dma->dp_dvmax].dvck_npages++; 2791 ASSERT(dcookies[dmax].dck_paddr != 0); 2792 ASSERT(dcookies[dmax].dck_npages != 0); 2793 2794 /* Check if this paddr is contiguous */ 2795 if (IMMU_CONTIG_PADDR(dcookies[dmax], paddr)) { 2796 dcookies[dmax].dck_npages++; 2797 } else { 2798 /* No, we need a new dcookie */ 2799 dmax++; 2800 ASSERT(dcookies[dmax].dck_paddr == 0); 2801 ASSERT(dcookies[dmax].dck_npages == 0); 2802 dcookies[dmax].dck_paddr = paddr; 2803 dcookies[dmax].dck_npages = 1; 2804 } 2805 } 2806 2807 dma->dp_dvmax = dvmax; 2808 dma->dp_dmax = dmax; 2809 } 2810 2811 static void 2812 cookie_finalize(ddi_dma_impl_t *hp, immu_t *immu, domain_t *domain, 2813 dev_info_t *rdip, immu_flags_t immu_flags) 2814 { 2815 int i; 2816 rootnex_dma_t *dma = (rootnex_dma_t *)hp->dmai_private; 2817 dvcookie_t *dvcookies = dma->dp_dvcookies; 2818 dcookie_t *dcookies = dma->dp_dcookies; 2819 ddi_dma_cookie_t *cookies = dma->dp_cookies; 2820 uint64_t npages; 2821 uint64_t dvma; 2822 boolean_t pde_set; 2823 2824 /* First calculate the total number of pages required */ 2825 for (i = 0, npages = 0; i <= dma->dp_dvmax; i++) { 2826 npages += dvcookies[i].dvck_npages; 2827 } 2828 2829 /* Now allocate dvma */ 2830 dvma = dvma_alloc(hp, domain, npages); 2831 2832 /* Now map the dvma */ 2833 pde_set = dvma_map(immu, domain, dvma, npages, dcookies, 2834 dma->dp_dmax + 1, rdip, immu_flags); 2835 2836 /* Invalidate the IOTLB */ 2837 immu_flush_iotlb_psi(immu, domain->dom_did, dvma, npages, 2838 pde_set == B_TRUE ? TLB_IVA_WHOLE : TLB_IVA_LEAF); 2839 2840 /* Now setup dvcookies and real cookie addresses */ 2841 for (i = 0; i <= dma->dp_dvmax; i++) { 2842 dvcookies[i].dvck_dvma = dvma; 2843 cookies[i].dmac_laddress = dvma; 2844 ASSERT(cookies[i].dmac_size != 0); 2845 cookies[i].dmac_type = 0; 2846 dvma += (dvcookies[i].dvck_npages * IMMU_PAGESIZE); 2847 } 2848 2849 #ifdef TEST 2850 immu_flush_iotlb_dsi(immu, domain->dom_did); 2851 #endif 2852 } 2853 2854 /* 2855 * cookie_create() 2856 */ 2857 static int 2858 cookie_create(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq, 2859 ddi_dma_attr_t *a, immu_t *immu, domain_t *domain, dev_info_t *rdip, 2860 uint_t prealloc_count, immu_flags_t immu_flags) 2861 { 2862 ddi_dma_atyp_t buftype; 2863 uint64_t offset; 2864 page_t **pparray; 2865 uint64_t paddr; 2866 uint_t psize; 2867 uint_t size; 2868 uint64_t maxseg; 2869 caddr_t vaddr; 2870 uint_t pcnt; 2871 page_t *page; 2872 rootnex_sglinfo_t *sglinfo; 2873 ddi_dma_obj_t *dmar_object; 2874 rootnex_dma_t *dma; 2875 size_t nocross; 2876 2877 dma = (rootnex_dma_t *)hp->dmai_private; 2878 sglinfo = &(dma->dp_sglinfo); 2879 dmar_object = &(dmareq->dmar_object); 2880 maxseg = sglinfo->si_max_cookie_size; 2881 pparray = dmar_object->dmao_obj.virt_obj.v_priv; 2882 vaddr = dmar_object->dmao_obj.virt_obj.v_addr; 2883 buftype = dmar_object->dmao_type; 2884 size = dmar_object->dmao_size; 2885 nocross = (size_t)(a->dma_attr_seg + 1); 2886 2887 /* 2888 * Allocate cookie, dvcookie and dcookie 2889 */ 2890 if (cookie_alloc(dma, dmareq, a, prealloc_count) != DDI_SUCCESS) { 2891 return (DDI_FAILURE); 2892 } 2893 hp->dmai_cookie = dma->dp_cookies; 2894 2895 pcnt = 0; 2896 2897 /* retrieve paddr, psize, offset from dmareq */ 2898 if (buftype == DMA_OTYP_PAGES) { 2899 page = dmar_object->dmao_obj.pp_obj.pp_pp; 2900 ASSERT(!PP_ISFREE(page) && PAGE_LOCKED(page)); 2901 offset = dmar_object->dmao_obj.pp_obj.pp_offset & 2902 MMU_PAGEOFFSET; 2903 paddr = pfn_to_pa(page->p_pagenum) + offset; 2904 psize = MIN((MMU_PAGESIZE - offset), size); 2905 sglinfo->si_asp = NULL; 2906 page = page->p_next; 2907 } else { 2908 ASSERT((buftype == DMA_OTYP_VADDR) || 2909 (buftype == DMA_OTYP_BUFVADDR)); 2910 sglinfo->si_asp = dmar_object->dmao_obj.virt_obj.v_as; 2911 if (sglinfo->si_asp == NULL) { 2912 sglinfo->si_asp = &kas; 2913 } 2914 offset = (uintptr_t)vaddr & MMU_PAGEOFFSET; 2915 if (pparray != NULL) { 2916 ASSERT(!PP_ISFREE(pparray[pcnt])); 2917 paddr = pfn_to_pa(pparray[pcnt]->p_pagenum) + offset; 2918 psize = MIN((MMU_PAGESIZE - offset), size); 2919 pcnt++; 2920 } else { 2921 paddr = pfn_to_pa(hat_getpfnum(sglinfo->si_asp->a_hat, 2922 vaddr)) + offset; 2923 psize = MIN(size, (MMU_PAGESIZE - offset)); 2924 vaddr += psize; 2925 } 2926 } 2927 2928 /* save the iommu page offset */ 2929 sglinfo->si_buf_offset = offset & IMMU_PAGEOFFSET; 2930 2931 /* 2932 * setup dvcookie and dcookie for [paddr, paddr+psize) 2933 */ 2934 cookie_update(domain, dma, paddr, psize, maxseg, nocross); 2935 2936 size -= psize; 2937 while (size > 0) { 2938 /* get the size for this page (i.e. partial or full page) */ 2939 psize = MIN(size, MMU_PAGESIZE); 2940 if (buftype == DMA_OTYP_PAGES) { 2941 /* get the paddr from the page_t */ 2942 ASSERT(!PP_ISFREE(page) && PAGE_LOCKED(page)); 2943 paddr = pfn_to_pa(page->p_pagenum); 2944 page = page->p_next; 2945 } else if (pparray != NULL) { 2946 /* index into the array of page_t's to get the paddr */ 2947 ASSERT(!PP_ISFREE(pparray[pcnt])); 2948 paddr = pfn_to_pa(pparray[pcnt]->p_pagenum); 2949 pcnt++; 2950 } else { 2951 /* call into the VM to get the paddr */ 2952 paddr = pfn_to_pa(hat_getpfnum 2953 (sglinfo->si_asp->a_hat, vaddr)); 2954 vaddr += psize; 2955 } 2956 /* 2957 * set dvcookie and dcookie for [paddr, paddr+psize) 2958 */ 2959 cookie_update(domain, dma, paddr, psize, maxseg, nocross); 2960 size -= psize; 2961 } 2962 2963 cookie_finalize(hp, immu, domain, rdip, immu_flags); 2964 2965 /* take account in the offset into the first page */ 2966 dma->dp_cookies[0].dmac_laddress += sglinfo->si_buf_offset; 2967 2968 /* save away how many cookies we have */ 2969 sglinfo->si_sgl_size = dma->dp_dvmax + 1; 2970 2971 return (DDI_SUCCESS); 2972 } 2973 2974 /* ############################# Functions exported ######################## */ 2975 2976 /* 2977 * setup the DVMA subsystem 2978 * this code runs only for the first IOMMU unit 2979 */ 2980 void 2981 immu_dvma_setup(list_t *listp) 2982 { 2983 immu_t *immu; 2984 uint_t kval; 2985 size_t nchains; 2986 2987 /* locks */ 2988 mutex_init(&immu_domain_lock, NULL, MUTEX_DEFAULT, NULL); 2989 2990 /* Create lists */ 2991 list_create(&immu_unity_domain_list, sizeof (domain_t), 2992 offsetof(domain_t, dom_maptype_node)); 2993 list_create(&immu_xlate_domain_list, sizeof (domain_t), 2994 offsetof(domain_t, dom_maptype_node)); 2995 2996 /* Setup BDF domain hash */ 2997 nchains = 0xff; 2998 kval = mod_hash_iddata_gen(nchains); 2999 3000 bdf_domain_hash = mod_hash_create_extended("BDF-DOMAIN_HASH", 3001 nchains, mod_hash_null_keydtor, mod_hash_null_valdtor, 3002 mod_hash_byid, (void *)(uintptr_t)kval, mod_hash_idkey_cmp, 3003 KM_NOSLEEP); 3004 ASSERT(bdf_domain_hash); 3005 3006 immu = list_head(listp); 3007 for (; immu; immu = list_next(listp, immu)) { 3008 create_unity_domain(immu); 3009 did_init(immu); 3010 context_init(immu); 3011 immu->immu_dvma_setup = B_TRUE; 3012 } 3013 } 3014 3015 /* 3016 * Startup up one DVMA unit 3017 */ 3018 void 3019 immu_dvma_startup(immu_t *immu) 3020 { 3021 ASSERT(immu); 3022 ASSERT(immu->immu_dvma_running == B_FALSE); 3023 3024 if (immu_gfxdvma_enable == B_FALSE && 3025 immu->immu_dvma_gfx_only == B_TRUE) { 3026 return; 3027 } 3028 3029 /* 3030 * DVMA will start once IOMMU is "running" 3031 */ 3032 ASSERT(immu->immu_dvma_running == B_FALSE); 3033 immu->immu_dvma_running = B_TRUE; 3034 } 3035 3036 /* 3037 * immu_dvma_physmem_update() 3038 * called when the installed memory on a 3039 * system increases, to expand domain DVMA 3040 * for domains with UNITY mapping 3041 */ 3042 void 3043 immu_dvma_physmem_update(uint64_t addr, uint64_t size) 3044 { 3045 uint64_t start; 3046 uint64_t npages; 3047 int dcount; 3048 dcookie_t dcookies[1] = {0}; 3049 domain_t *domain; 3050 3051 /* 3052 * Just walk the system-wide list of domains with 3053 * UNITY mapping. Both the list of *all* domains 3054 * and *UNITY* domains is protected by the same 3055 * single lock 3056 */ 3057 mutex_enter(&immu_domain_lock); 3058 domain = list_head(&immu_unity_domain_list); 3059 for (; domain; domain = list_next(&immu_unity_domain_list, domain)) { 3060 3061 /* There is no vmem_arena for unity domains. Just map it */ 3062 ddi_err(DER_LOG, NULL, "IMMU: unity-domain: Adding map " 3063 "[0x%" PRIx64 " - 0x%" PRIx64 "]", addr, addr + size); 3064 3065 start = IMMU_ROUNDOWN(addr); 3066 npages = (IMMU_ROUNDUP(size) / IMMU_PAGESIZE) + 1; 3067 3068 dcookies[0].dck_paddr = start; 3069 dcookies[0].dck_npages = npages; 3070 dcount = 1; 3071 (void) dvma_map(domain->dom_immu, domain, start, npages, 3072 dcookies, dcount, NULL, IMMU_FLAGS_READ | IMMU_FLAGS_WRITE); 3073 3074 } 3075 mutex_exit(&immu_domain_lock); 3076 } 3077 3078 3079 int 3080 immu_dvma_map(ddi_dma_impl_t *hp, struct ddi_dma_req *dmareq, memrng_t *mrng, 3081 uint_t prealloc_count, dev_info_t *rdip, immu_flags_t immu_flags) 3082 { 3083 ddi_dma_attr_t *attr; 3084 dev_info_t *ddip; 3085 domain_t *domain; 3086 immu_t *immu; 3087 dcookie_t dcookies[1] = {0}; 3088 int dcount = 0; 3089 boolean_t pde_set = B_TRUE; 3090 int r = DDI_FAILURE; 3091 3092 ASSERT(immu_enable == B_TRUE); 3093 ASSERT(immu_running == B_TRUE || !(immu_flags & IMMU_FLAGS_DMAHDL)); 3094 ASSERT(hp || !(immu_flags & IMMU_FLAGS_DMAHDL)); 3095 3096 /* 3097 * Intel IOMMU will only be turned on if IOMMU 3098 * page size is a multiple of IOMMU page size 3099 */ 3100 3101 /*LINTED*/ 3102 ASSERT(MMU_PAGESIZE % IMMU_PAGESIZE == 0); 3103 3104 /* Can only do DVMA if dip is attached */ 3105 if (rdip == NULL) { 3106 ddi_err(DER_PANIC, rdip, "DVMA map: No device specified"); 3107 /*NOTREACHED*/ 3108 } 3109 3110 immu_flags |= dma_to_immu_flags(dmareq); 3111 3112 immu = immu_dvma_get_immu(rdip, immu_flags); 3113 if (immu == NULL) { 3114 /* 3115 * possible that there is no IOMMU unit for this device 3116 * - BIOS bugs are one example. 3117 */ 3118 ddi_err(DER_WARN, rdip, "No IMMU unit found for device"); 3119 return (DDI_DMA_NORESOURCES); 3120 } 3121 3122 /* 3123 * redirect isa devices attached under lpc to lpc dip 3124 */ 3125 if (strcmp(ddi_node_name(ddi_get_parent(rdip)), "isa") == 0) { 3126 rdip = get_lpc_devinfo(immu, rdip, immu_flags); 3127 if (rdip == NULL) { 3128 ddi_err(DER_PANIC, rdip, "IMMU redirect failed"); 3129 /*NOTREACHED*/ 3130 } 3131 } 3132 3133 /* Reset immu, as redirection can change IMMU */ 3134 immu = NULL; 3135 3136 /* 3137 * for gart, redirect to the real graphic devinfo 3138 */ 3139 if (strcmp(ddi_node_name(rdip), "agpgart") == 0) { 3140 rdip = get_gfx_devinfo(rdip); 3141 if (rdip == NULL) { 3142 ddi_err(DER_PANIC, rdip, "IMMU redirect failed"); 3143 /*NOTREACHED*/ 3144 } 3145 } 3146 3147 /* 3148 * Setup DVMA domain for the device. This does 3149 * work only the first time we do DVMA for a 3150 * device. 3151 */ 3152 ddip = NULL; 3153 domain = device_domain(rdip, &ddip, immu_flags); 3154 if (domain == NULL) { 3155 ASSERT(ddip == NULL); 3156 ddi_err(DER_MODE, rdip, "Intel IOMMU setup failed for device"); 3157 return (DDI_DMA_NORESOURCES); 3158 } 3159 3160 /* 3161 * If a domain is found, we must also have a domain dip 3162 * which is the topmost ancestor dip of rdip that shares 3163 * the same domain with rdip. 3164 */ 3165 if (domain->dom_did == 0 || ddip == NULL) { 3166 ddi_err(DER_MODE, rdip, "domain did 0(%d) or ddip NULL(%p)", 3167 domain->dom_did, ddip); 3168 return (DDI_DMA_NORESOURCES); 3169 } 3170 3171 immu = domain->dom_immu; 3172 ASSERT(immu); 3173 if (domain->dom_did == IMMU_UNITY_DID) { 3174 ASSERT(domain == immu->immu_unity_domain); 3175 /* mapping already done. Let rootnex create cookies */ 3176 r = DDI_DMA_USE_PHYSICAL; 3177 } else if (immu_flags & IMMU_FLAGS_DMAHDL) { 3178 3179 /* if we have a DMA handle, the IOMMUs must be running */ 3180 ASSERT(immu->immu_regs_running == B_TRUE); 3181 ASSERT(immu->immu_dvma_running == B_TRUE); 3182 3183 attr = &hp->dmai_attr; 3184 if (attr == NULL) { 3185 ddi_err(DER_PANIC, rdip, 3186 "DMA handle (%p): NULL attr", hp); 3187 /*NOTREACHED*/ 3188 } 3189 3190 if (cookie_create(hp, dmareq, attr, immu, domain, rdip, 3191 prealloc_count, immu_flags) != DDI_SUCCESS) { 3192 ddi_err(DER_MODE, rdip, "dvcookie_alloc: failed"); 3193 return (DDI_DMA_NORESOURCES); 3194 } 3195 r = DDI_DMA_MAPPED; 3196 } else if (immu_flags & IMMU_FLAGS_MEMRNG) { 3197 dcookies[0].dck_paddr = mrng->mrng_start; 3198 dcookies[0].dck_npages = mrng->mrng_npages; 3199 dcount = 1; 3200 pde_set = dvma_map(immu, domain, mrng->mrng_start, 3201 mrng->mrng_npages, dcookies, dcount, rdip, immu_flags); 3202 immu_flush_iotlb_psi(immu, domain->dom_did, mrng->mrng_start, 3203 mrng->mrng_npages, pde_set == B_TRUE ? 3204 TLB_IVA_WHOLE : TLB_IVA_LEAF); 3205 r = DDI_DMA_MAPPED; 3206 } else { 3207 ddi_err(DER_PANIC, rdip, "invalid flags for immu_dvma_map()"); 3208 /*NOTREACHED*/ 3209 } 3210 3211 /* 3212 * Update the root and context entries 3213 */ 3214 if (immu_context_update(immu, domain, ddip, rdip, immu_flags) 3215 != DDI_SUCCESS) { 3216 ddi_err(DER_MODE, rdip, "DVMA map: context update failed"); 3217 return (DDI_DMA_NORESOURCES); 3218 } 3219 3220 immu_regs_wbf_flush(immu); 3221 3222 return (r); 3223 } 3224 3225 int 3226 immu_dvma_unmap(ddi_dma_impl_t *hp, dev_info_t *rdip) 3227 { 3228 ddi_dma_attr_t *attr; 3229 rootnex_dma_t *dma; 3230 domain_t *domain; 3231 immu_t *immu; 3232 dev_info_t *ddip; 3233 immu_flags_t immu_flags; 3234 3235 ASSERT(immu_enable == B_TRUE); 3236 ASSERT(immu_running == B_TRUE); 3237 ASSERT(hp); 3238 3239 /* 3240 * Intel IOMMU will only be turned on if IOMMU 3241 * page size is same as MMU page size 3242 */ 3243 /*LINTED*/ 3244 ASSERT(MMU_PAGESIZE == IMMU_PAGESIZE); 3245 3246 /* rdip need not be attached */ 3247 if (rdip == NULL) { 3248 ddi_err(DER_PANIC, rdip, "DVMA unmap: No device specified"); 3249 return (DDI_DMA_NORESOURCES); 3250 } 3251 3252 /* 3253 * Get the device domain, this should always 3254 * succeed since there had to be a domain to 3255 * setup DVMA. 3256 */ 3257 dma = (rootnex_dma_t *)hp->dmai_private; 3258 attr = &hp->dmai_attr; 3259 if (attr == NULL) { 3260 ddi_err(DER_PANIC, rdip, "DMA handle (%p) has NULL attr", hp); 3261 /*NOTREACHED*/ 3262 } 3263 immu_flags = dma->dp_sleep_flags; 3264 3265 immu = immu_dvma_get_immu(rdip, immu_flags); 3266 if (immu == NULL) { 3267 /* 3268 * possible that there is no IOMMU unit for this device 3269 * - BIOS bugs are one example. 3270 */ 3271 ddi_err(DER_WARN, rdip, "No IMMU unit found for device"); 3272 return (DDI_DMA_NORESOURCES); 3273 } 3274 3275 3276 /* 3277 * redirect isa devices attached under lpc to lpc dip 3278 */ 3279 if (strcmp(ddi_node_name(ddi_get_parent(rdip)), "isa") == 0) { 3280 rdip = get_lpc_devinfo(immu, rdip, immu_flags); 3281 if (rdip == NULL) { 3282 ddi_err(DER_PANIC, rdip, "IMMU redirect failed"); 3283 /*NOTREACHED*/ 3284 } 3285 } 3286 3287 /* Reset immu, as redirection can change IMMU */ 3288 immu = NULL; 3289 3290 /* 3291 * for gart, redirect to the real graphic devinfo 3292 */ 3293 if (strcmp(ddi_node_name(rdip), "agpgart") == 0) { 3294 rdip = get_gfx_devinfo(rdip); 3295 if (rdip == NULL) { 3296 ddi_err(DER_PANIC, rdip, "IMMU redirect failed"); 3297 /*NOTREACHED*/ 3298 } 3299 } 3300 3301 ddip = NULL; 3302 domain = device_domain(rdip, &ddip, immu_flags); 3303 if (domain == NULL || domain->dom_did == 0 || ddip == NULL) { 3304 ddi_err(DER_MODE, rdip, "Attempt to unmap DVMA for " 3305 "a device without domain or with an uninitialized " 3306 "domain"); 3307 return (DDI_DMA_NORESOURCES); 3308 } 3309 3310 /* 3311 * immu must be set in the domain. 3312 */ 3313 immu = domain->dom_immu; 3314 ASSERT(immu); 3315 if (domain->dom_did == IMMU_UNITY_DID) { 3316 ASSERT(domain == immu->immu_unity_domain); 3317 /* 3318 * domain is unity, nothing to do here, let the rootnex 3319 * code free the cookies. 3320 */ 3321 return (DDI_DMA_USE_PHYSICAL); 3322 } 3323 3324 dma = hp->dmai_private; 3325 if (dma == NULL) { 3326 ddi_err(DER_PANIC, rdip, "DVMA unmap: DMA handle (%p) has " 3327 "no private dma structure", hp); 3328 /*NOTREACHED*/ 3329 } 3330 3331 cookie_free(dma, immu, domain, rdip); 3332 3333 /* No invalidation needed for unmap */ 3334 immu_regs_wbf_flush(immu); 3335 3336 return (DDI_SUCCESS); 3337 } 3338 3339 immu_devi_t * 3340 immu_devi_get(dev_info_t *rdip) 3341 { 3342 immu_devi_t *immu_devi; 3343 volatile uintptr_t *vptr = (uintptr_t *)&(DEVI(rdip)->devi_iommu); 3344 3345 /* Just want atomic reads. No need for lock */ 3346 immu_devi = (immu_devi_t *)(uintptr_t)atomic_or_64_nv((uint64_t *)vptr, 3347 0); 3348 return (immu_devi); 3349 } 3350