1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * PCI Peer 2 Peer DMA support. 4 * 5 * Copyright (c) 2016-2018, Logan Gunthorpe 6 * Copyright (c) 2016-2017, Microsemi Corporation 7 * Copyright (c) 2017, Christoph Hellwig 8 * Copyright (c) 2018, Eideticom Inc. 9 */ 10 11 #define pr_fmt(fmt) "pci-p2pdma: " fmt 12 #include <linux/ctype.h> 13 #include <linux/dma-map-ops.h> 14 #include <linux/pci-p2pdma.h> 15 #include <linux/module.h> 16 #include <linux/slab.h> 17 #include <linux/genalloc.h> 18 #include <linux/memremap.h> 19 #include <linux/percpu-refcount.h> 20 #include <linux/random.h> 21 #include <linux/seq_buf.h> 22 #include <linux/xarray.h> 23 24 struct pci_p2pdma { 25 struct gen_pool *pool; 26 bool p2pmem_published; 27 struct xarray map_types; 28 struct p2pdma_provider mem[PCI_STD_NUM_BARS]; 29 }; 30 31 struct pci_p2pdma_pagemap { 32 struct dev_pagemap pgmap; 33 struct p2pdma_provider *mem; 34 }; 35 36 static struct pci_p2pdma_pagemap *to_p2p_pgmap(struct dev_pagemap *pgmap) 37 { 38 return container_of(pgmap, struct pci_p2pdma_pagemap, pgmap); 39 } 40 41 static ssize_t size_show(struct device *dev, struct device_attribute *attr, 42 char *buf) 43 { 44 struct pci_dev *pdev = to_pci_dev(dev); 45 struct pci_p2pdma *p2pdma; 46 size_t size = 0; 47 48 rcu_read_lock(); 49 p2pdma = rcu_dereference(pdev->p2pdma); 50 if (p2pdma && p2pdma->pool) 51 size = gen_pool_size(p2pdma->pool); 52 rcu_read_unlock(); 53 54 return sysfs_emit(buf, "%zd\n", size); 55 } 56 static DEVICE_ATTR_RO(size); 57 58 static ssize_t available_show(struct device *dev, struct device_attribute *attr, 59 char *buf) 60 { 61 struct pci_dev *pdev = to_pci_dev(dev); 62 struct pci_p2pdma *p2pdma; 63 size_t avail = 0; 64 65 rcu_read_lock(); 66 p2pdma = rcu_dereference(pdev->p2pdma); 67 if (p2pdma && p2pdma->pool) 68 avail = gen_pool_avail(p2pdma->pool); 69 rcu_read_unlock(); 70 71 return sysfs_emit(buf, "%zd\n", avail); 72 } 73 static DEVICE_ATTR_RO(available); 74 75 static ssize_t published_show(struct device *dev, struct device_attribute *attr, 76 char *buf) 77 { 78 struct pci_dev *pdev = to_pci_dev(dev); 79 struct pci_p2pdma *p2pdma; 80 bool published = false; 81 82 rcu_read_lock(); 83 p2pdma = rcu_dereference(pdev->p2pdma); 84 if (p2pdma) 85 published = p2pdma->p2pmem_published; 86 rcu_read_unlock(); 87 88 return sysfs_emit(buf, "%d\n", published); 89 } 90 static DEVICE_ATTR_RO(published); 91 92 static int p2pmem_alloc_mmap(struct file *filp, struct kobject *kobj, 93 const struct bin_attribute *attr, struct vm_area_struct *vma) 94 { 95 struct pci_dev *pdev = to_pci_dev(kobj_to_dev(kobj)); 96 size_t len = vma->vm_end - vma->vm_start; 97 struct pci_p2pdma *p2pdma; 98 struct percpu_ref *ref; 99 unsigned long vaddr; 100 void *kaddr; 101 int ret; 102 103 /* prevent private mappings from being established */ 104 if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) { 105 pci_info_ratelimited(pdev, 106 "%s: fail, attempted private mapping\n", 107 current->comm); 108 return -EINVAL; 109 } 110 111 if (vma->vm_pgoff) { 112 pci_info_ratelimited(pdev, 113 "%s: fail, attempted mapping with non-zero offset\n", 114 current->comm); 115 return -EINVAL; 116 } 117 118 rcu_read_lock(); 119 p2pdma = rcu_dereference(pdev->p2pdma); 120 if (!p2pdma) { 121 ret = -ENODEV; 122 goto out; 123 } 124 125 kaddr = (void *)gen_pool_alloc_owner(p2pdma->pool, len, (void **)&ref); 126 if (!kaddr) { 127 ret = -ENOMEM; 128 goto out; 129 } 130 131 /* 132 * vm_insert_page() can sleep, so a reference is taken to mapping 133 * such that rcu_read_unlock() can be done before inserting the 134 * pages 135 */ 136 if (unlikely(!percpu_ref_tryget_live_rcu(ref))) { 137 ret = -ENODEV; 138 goto out_free_mem; 139 } 140 rcu_read_unlock(); 141 142 for (vaddr = vma->vm_start; vaddr < vma->vm_end; vaddr += PAGE_SIZE) { 143 struct page *page = virt_to_page(kaddr); 144 145 /* 146 * Initialise the refcount for the freshly allocated page. As 147 * we have just allocated the page no one else should be 148 * using it. 149 */ 150 VM_WARN_ON_ONCE_PAGE(page_ref_count(page), page); 151 set_page_count(page, 1); 152 ret = vm_insert_page(vma, vaddr, page); 153 if (ret) { 154 gen_pool_free(p2pdma->pool, (uintptr_t)kaddr, len); 155 156 /* 157 * Reset the page count. We don't use put_page() 158 * because we don't want to trigger the 159 * p2pdma_folio_free() path. 160 */ 161 set_page_count(page, 0); 162 percpu_ref_put(ref); 163 return ret; 164 } 165 percpu_ref_get(ref); 166 put_page(page); 167 kaddr += PAGE_SIZE; 168 len -= PAGE_SIZE; 169 } 170 171 percpu_ref_put(ref); 172 173 return 0; 174 out_free_mem: 175 gen_pool_free(p2pdma->pool, (uintptr_t)kaddr, len); 176 out: 177 rcu_read_unlock(); 178 return ret; 179 } 180 181 static const struct bin_attribute p2pmem_alloc_attr = { 182 .attr = { .name = "allocate", .mode = 0660 }, 183 .mmap = p2pmem_alloc_mmap, 184 /* 185 * Some places where we want to call mmap (ie. python) will check 186 * that the file size is greater than the mmap size before allowing 187 * the mmap to continue. To work around this, just set the size 188 * to be very large. 189 */ 190 .size = SZ_1T, 191 }; 192 193 static struct attribute *p2pmem_attrs[] = { 194 &dev_attr_size.attr, 195 &dev_attr_available.attr, 196 &dev_attr_published.attr, 197 NULL, 198 }; 199 200 static const struct bin_attribute *const p2pmem_bin_attrs[] = { 201 &p2pmem_alloc_attr, 202 NULL, 203 }; 204 205 static const struct attribute_group p2pmem_group = { 206 .attrs = p2pmem_attrs, 207 .bin_attrs = p2pmem_bin_attrs, 208 .name = "p2pmem", 209 }; 210 211 static void p2pdma_folio_free(struct folio *folio) 212 { 213 struct page *page = &folio->page; 214 struct pci_p2pdma_pagemap *pgmap = to_p2p_pgmap(page_pgmap(page)); 215 /* safe to dereference while a reference is held to the percpu ref */ 216 struct pci_p2pdma *p2pdma = rcu_dereference_protected( 217 to_pci_dev(pgmap->mem->owner)->p2pdma, 1); 218 struct percpu_ref *ref; 219 220 gen_pool_free_owner(p2pdma->pool, (uintptr_t)page_to_virt(page), 221 PAGE_SIZE, (void **)&ref); 222 percpu_ref_put(ref); 223 } 224 225 static const struct dev_pagemap_ops p2pdma_pgmap_ops = { 226 .folio_free = p2pdma_folio_free, 227 }; 228 229 static void pci_p2pdma_release(void *data) 230 { 231 struct pci_dev *pdev = data; 232 struct pci_p2pdma *p2pdma; 233 234 p2pdma = rcu_dereference_protected(pdev->p2pdma, 1); 235 if (!p2pdma) 236 return; 237 238 /* Flush and disable pci_alloc_p2p_mem() */ 239 RCU_INIT_POINTER(pdev->p2pdma, NULL); 240 synchronize_rcu(); 241 xa_destroy(&p2pdma->map_types); 242 243 if (!p2pdma->pool) 244 return; 245 246 gen_pool_destroy(p2pdma->pool); 247 sysfs_remove_group(&pdev->dev.kobj, &p2pmem_group); 248 } 249 250 /** 251 * pcim_p2pdma_init - Initialise peer-to-peer DMA providers 252 * @pdev: The PCI device to enable P2PDMA for 253 * 254 * This function initializes the peer-to-peer DMA infrastructure 255 * for a PCI device. It allocates and sets up the necessary data 256 * structures to support P2PDMA operations, including mapping type 257 * tracking. 258 */ 259 int pcim_p2pdma_init(struct pci_dev *pdev) 260 { 261 struct pci_p2pdma *p2p; 262 int i, ret; 263 264 if (pdev->non_mappable_bars) 265 return -EOPNOTSUPP; 266 267 p2p = rcu_dereference_protected(pdev->p2pdma, 1); 268 if (p2p) 269 return 0; 270 271 p2p = devm_kzalloc(&pdev->dev, sizeof(*p2p), GFP_KERNEL); 272 if (!p2p) 273 return -ENOMEM; 274 275 xa_init(&p2p->map_types); 276 /* 277 * Iterate over all standard PCI BARs and record only those that 278 * correspond to MMIO regions. Skip non-memory resources (e.g. I/O 279 * port BARs) since they cannot be used for peer-to-peer (P2P) 280 * transactions. 281 */ 282 for (i = 0; i < PCI_STD_NUM_BARS; i++) { 283 if (!(pci_resource_flags(pdev, i) & IORESOURCE_MEM)) 284 continue; 285 286 p2p->mem[i].owner = &pdev->dev; 287 p2p->mem[i].bus_offset = 288 pci_bus_address(pdev, i) - pci_resource_start(pdev, i); 289 } 290 291 ret = devm_add_action_or_reset(&pdev->dev, pci_p2pdma_release, pdev); 292 if (ret) 293 goto out_p2p; 294 295 rcu_assign_pointer(pdev->p2pdma, p2p); 296 return 0; 297 298 out_p2p: 299 devm_kfree(&pdev->dev, p2p); 300 return ret; 301 } 302 EXPORT_SYMBOL_GPL(pcim_p2pdma_init); 303 304 /** 305 * pcim_p2pdma_provider - Get peer-to-peer DMA provider 306 * @pdev: The PCI device to enable P2PDMA for 307 * @bar: BAR index to get provider 308 * 309 * This function gets peer-to-peer DMA provider for a PCI device. The lifetime 310 * of the provider (and of course the MMIO) is bound to the lifetime of the 311 * driver. A driver calling this function must ensure that all references to the 312 * provider, and any DMA mappings created for any MMIO, are all cleaned up 313 * before the driver remove() completes. 314 * 315 * Since P2P is almost always shared with a second driver this means some system 316 * to notify, invalidate and revoke the MMIO's DMA must be in place to use this 317 * function. For example a revoke can be built using DMABUF. 318 */ 319 struct p2pdma_provider *pcim_p2pdma_provider(struct pci_dev *pdev, int bar) 320 { 321 struct pci_p2pdma *p2p; 322 323 if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM) || 324 pdev->non_mappable_bars) 325 return NULL; 326 327 p2p = rcu_dereference_protected(pdev->p2pdma, 1); 328 if (WARN_ON(!p2p)) 329 /* Someone forgot to call to pcim_p2pdma_init() before */ 330 return NULL; 331 332 return &p2p->mem[bar]; 333 } 334 EXPORT_SYMBOL_GPL(pcim_p2pdma_provider); 335 336 static int pci_p2pdma_setup_pool(struct pci_dev *pdev) 337 { 338 struct pci_p2pdma *p2pdma; 339 int ret; 340 341 p2pdma = rcu_dereference_protected(pdev->p2pdma, 1); 342 if (p2pdma->pool) 343 /* We already setup pools, do nothing, */ 344 return 0; 345 346 p2pdma->pool = gen_pool_create(PAGE_SHIFT, dev_to_node(&pdev->dev)); 347 if (!p2pdma->pool) 348 return -ENOMEM; 349 350 ret = sysfs_create_group(&pdev->dev.kobj, &p2pmem_group); 351 if (ret) 352 goto out_pool_destroy; 353 354 return 0; 355 356 out_pool_destroy: 357 gen_pool_destroy(p2pdma->pool); 358 p2pdma->pool = NULL; 359 return ret; 360 } 361 362 static void pci_p2pdma_unmap_mappings(void *data) 363 { 364 struct pci_p2pdma_pagemap *p2p_pgmap = data; 365 366 /* 367 * Removing the alloc attribute from sysfs will call 368 * unmap_mapping_range() on the inode, teardown any existing userspace 369 * mappings and prevent new ones from being created. 370 */ 371 sysfs_remove_file_from_group(&p2p_pgmap->mem->owner->kobj, 372 &p2pmem_alloc_attr.attr, 373 p2pmem_group.name); 374 } 375 376 /** 377 * pci_p2pdma_add_resource - add memory for use as p2p memory 378 * @pdev: the device to add the memory to 379 * @bar: PCI BAR to add 380 * @size: size of the memory to add, may be zero to use the whole BAR 381 * @offset: offset into the PCI BAR 382 * 383 * The memory will be given ZONE_DEVICE struct pages so that it may 384 * be used with any DMA request. 385 */ 386 int pci_p2pdma_add_resource(struct pci_dev *pdev, int bar, size_t size, 387 u64 offset) 388 { 389 struct pci_p2pdma_pagemap *p2p_pgmap; 390 struct p2pdma_provider *mem; 391 struct dev_pagemap *pgmap; 392 struct pci_p2pdma *p2pdma; 393 void *addr; 394 int error; 395 396 if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM)) 397 return -EINVAL; 398 399 if (offset >= pci_resource_len(pdev, bar)) 400 return -EINVAL; 401 402 if (!size) 403 size = pci_resource_len(pdev, bar) - offset; 404 405 if (size + offset > pci_resource_len(pdev, bar)) 406 return -EINVAL; 407 408 error = pcim_p2pdma_init(pdev); 409 if (error) 410 return error; 411 412 error = pci_p2pdma_setup_pool(pdev); 413 if (error) 414 return error; 415 416 mem = pcim_p2pdma_provider(pdev, bar); 417 /* 418 * We checked validity of BAR prior to call 419 * to pcim_p2pdma_provider. It should never return NULL. 420 */ 421 if (WARN_ON(!mem)) 422 return -EINVAL; 423 424 p2p_pgmap = devm_kzalloc(&pdev->dev, sizeof(*p2p_pgmap), GFP_KERNEL); 425 if (!p2p_pgmap) 426 return -ENOMEM; 427 428 pgmap = &p2p_pgmap->pgmap; 429 pgmap->range.start = pci_resource_start(pdev, bar) + offset; 430 pgmap->range.end = pgmap->range.start + size - 1; 431 pgmap->nr_range = 1; 432 pgmap->type = MEMORY_DEVICE_PCI_P2PDMA; 433 pgmap->ops = &p2pdma_pgmap_ops; 434 p2p_pgmap->mem = mem; 435 436 addr = devm_memremap_pages(&pdev->dev, pgmap); 437 if (IS_ERR(addr)) { 438 error = PTR_ERR(addr); 439 goto pgmap_free; 440 } 441 442 error = devm_add_action(&pdev->dev, pci_p2pdma_unmap_mappings, 443 p2p_pgmap); 444 if (error) 445 goto pages_free; 446 447 p2pdma = rcu_dereference_protected(pdev->p2pdma, 1); 448 error = gen_pool_add_owner(p2pdma->pool, (unsigned long)addr, 449 pci_bus_address(pdev, bar) + offset, 450 range_len(&pgmap->range), dev_to_node(&pdev->dev), 451 &pgmap->ref); 452 if (error) 453 goto mappings_remove; 454 455 pci_info(pdev, "added peer-to-peer DMA memory %#llx-%#llx\n", 456 pgmap->range.start, pgmap->range.end); 457 458 return 0; 459 460 mappings_remove: 461 devm_remove_action(&pdev->dev, pci_p2pdma_unmap_mappings, p2p_pgmap); 462 pages_free: 463 devm_memunmap_pages(&pdev->dev, pgmap); 464 pgmap_free: 465 devm_kfree(&pdev->dev, p2p_pgmap); 466 return error; 467 } 468 EXPORT_SYMBOL_GPL(pci_p2pdma_add_resource); 469 470 /* 471 * Note this function returns the parent PCI device with a 472 * reference taken. It is the caller's responsibility to drop 473 * the reference. 474 */ 475 static struct pci_dev *find_parent_pci_dev(struct device *dev) 476 { 477 struct device *parent; 478 479 dev = get_device(dev); 480 481 while (dev) { 482 if (dev_is_pci(dev)) 483 return to_pci_dev(dev); 484 485 parent = get_device(dev->parent); 486 put_device(dev); 487 dev = parent; 488 } 489 490 return NULL; 491 } 492 493 /* 494 * Check if a PCI bridge has its ACS redirection bits set to redirect P2P 495 * TLPs upstream via ACS. Returns 1 if the packets will be redirected 496 * upstream, 0 otherwise. 497 */ 498 static int pci_bridge_has_acs_redir(struct pci_dev *pdev) 499 { 500 int pos; 501 u16 ctrl; 502 503 pos = pdev->acs_cap; 504 if (!pos) 505 return 0; 506 507 pci_read_config_word(pdev, pos + PCI_ACS_CTRL, &ctrl); 508 509 if (ctrl & (PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_EC)) 510 return 1; 511 512 return 0; 513 } 514 515 static void seq_buf_print_bus_devfn(struct seq_buf *buf, struct pci_dev *pdev) 516 { 517 if (!buf) 518 return; 519 520 seq_buf_printf(buf, "%s;", pci_name(pdev)); 521 } 522 523 static bool cpu_supports_p2pdma(void) 524 { 525 #ifdef CONFIG_X86 526 struct cpuinfo_x86 *c = &cpu_data(0); 527 528 /* Any AMD CPU whose family ID is Zen or newer supports p2pdma */ 529 if (c->x86_vendor == X86_VENDOR_AMD && c->x86 >= 0x17) 530 return true; 531 #endif 532 533 return false; 534 } 535 536 static const struct pci_p2pdma_whitelist_entry { 537 unsigned short vendor; 538 int device; 539 enum { 540 REQ_SAME_HOST_BRIDGE = 1 << 0, 541 } flags; 542 } pci_p2pdma_whitelist[] = { 543 /* Intel Xeon E5/Core i7 */ 544 {PCI_VENDOR_ID_INTEL, 0x3c00, REQ_SAME_HOST_BRIDGE}, 545 {PCI_VENDOR_ID_INTEL, 0x3c01, REQ_SAME_HOST_BRIDGE}, 546 /* Intel Xeon E7 v4/Xeon E5 v4/Xeon E3 v4/Xeon D */ 547 {PCI_VENDOR_ID_INTEL, 0x6f00, REQ_SAME_HOST_BRIDGE}, 548 /* Intel Xeon E7 v3/Xeon E5 v3/Core i7 */ 549 {PCI_VENDOR_ID_INTEL, 0x2f00, REQ_SAME_HOST_BRIDGE}, 550 {PCI_VENDOR_ID_INTEL, 0x2f01, REQ_SAME_HOST_BRIDGE}, 551 /* Intel Haswell (client) */ 552 {PCI_VENDOR_ID_INTEL, 0x0c08, REQ_SAME_HOST_BRIDGE}, 553 /* Intel Skylake-E */ 554 {PCI_VENDOR_ID_INTEL, 0x2030, 0}, 555 {PCI_VENDOR_ID_INTEL, 0x2031, 0}, 556 {PCI_VENDOR_ID_INTEL, 0x2032, 0}, 557 {PCI_VENDOR_ID_INTEL, 0x2033, 0}, 558 {PCI_VENDOR_ID_INTEL, 0x2020, 0}, 559 {PCI_VENDOR_ID_INTEL, 0x09a2, 0}, 560 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_DSA_SPR0, 0}, 561 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IAX_SPR0, 0}, 562 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_DSA_GNRD, 0}, 563 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_DSA_DMR, 0}, 564 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IAA_DMR, 0}, 565 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_QAT_4XXX, 0}, 566 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_QAT_401XX, 0}, 567 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_QAT_402XX, 0}, 568 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_QAT_420XX, 0}, 569 {PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_QAT_6XXX, 0}, 570 /* Google SoCs. */ 571 {PCI_VENDOR_ID_GOOGLE, PCI_ANY_ID, 0}, 572 /* Nvidia CPUs */ 573 {PCI_VENDOR_ID_NVIDIA, 0x2f95, 0}, 574 {PCI_VENDOR_ID_NVIDIA, 0x2f96, 0}, 575 {PCI_VENDOR_ID_NVIDIA, 0x2f97, 0}, 576 {PCI_VENDOR_ID_NVIDIA, 0x2f98, 0}, 577 /* Zhaoxin KX-6000/KH-40000/KX-6000G/KX-7000/KH-50000 */ 578 {PCI_VENDOR_ID_ZHAOXIN, 0x1003, REQ_SAME_HOST_BRIDGE}, 579 {PCI_VENDOR_ID_ZHAOXIN, 0x1005, REQ_SAME_HOST_BRIDGE}, 580 {PCI_VENDOR_ID_ZHAOXIN, 0x1006, REQ_SAME_HOST_BRIDGE}, 581 {PCI_VENDOR_ID_ZHAOXIN, 0x1007, REQ_SAME_HOST_BRIDGE}, 582 {PCI_VENDOR_ID_ZHAOXIN, 0x1008, 0}, 583 /* Alibaba T-HEAD Yitian 710 CPU */ 584 {PCI_VENDOR_ID_ALIBABA, 0x8000, 0}, 585 {} 586 }; 587 588 /* 589 * If the first device on host's root bus is either devfn 00.0 or a PCIe 590 * Root Port, return it. Otherwise return NULL. 591 * 592 * We often use a devfn 00.0 "host bridge" in the pci_p2pdma_whitelist[] 593 * (though there is no PCI/PCIe requirement for such a device). On some 594 * platforms, e.g., Intel Skylake, there is no such host bridge device, and 595 * pci_p2pdma_whitelist[] may contain a Root Port at any devfn. 596 * 597 * This function is similar to pci_get_slot(host->bus, 0), but it does 598 * not take the pci_bus_sem lock since __host_bridge_whitelist() must not 599 * sleep. 600 * 601 * For this to be safe, the caller should hold a reference to a device on the 602 * bridge, which should ensure the host_bridge device will not be freed 603 * or removed from the head of the devices list. 604 */ 605 static struct pci_dev *pci_host_bridge_dev(struct pci_host_bridge *host) 606 { 607 struct pci_dev *root; 608 609 root = list_first_entry_or_null(&host->bus->devices, 610 struct pci_dev, bus_list); 611 612 if (!root) 613 return NULL; 614 615 if (root->devfn == PCI_DEVFN(0, 0)) 616 return root; 617 618 if (pci_pcie_type(root) == PCI_EXP_TYPE_ROOT_PORT) 619 return root; 620 621 return NULL; 622 } 623 624 static bool __host_bridge_whitelist(struct pci_host_bridge *host, 625 bool same_host_bridge, bool warn) 626 { 627 struct pci_dev *root = pci_host_bridge_dev(host); 628 const struct pci_p2pdma_whitelist_entry *entry; 629 unsigned short vendor, device; 630 631 if (!root) 632 return false; 633 634 vendor = root->vendor; 635 device = root->device; 636 637 for (entry = pci_p2pdma_whitelist; entry->vendor; entry++) { 638 if (vendor != entry->vendor) 639 continue; 640 641 if (entry->device != PCI_ANY_ID && device != entry->device) 642 continue; 643 644 if (entry->flags & REQ_SAME_HOST_BRIDGE && !same_host_bridge) 645 return false; 646 647 return true; 648 } 649 650 if (warn) 651 pci_warn(root, "Host bridge not in P2PDMA whitelist: %04x:%04x\n", 652 vendor, device); 653 654 return false; 655 } 656 657 /* 658 * If we can't find a common upstream bridge take a look at the root 659 * complex and compare it to a whitelist of known good hardware. 660 */ 661 static bool host_bridge_whitelist(struct pci_dev *a, struct pci_dev *b, 662 bool warn) 663 { 664 struct pci_host_bridge *host_a = pci_find_host_bridge(a->bus); 665 struct pci_host_bridge *host_b = pci_find_host_bridge(b->bus); 666 667 if (host_a == host_b) 668 return __host_bridge_whitelist(host_a, true, warn); 669 670 if (__host_bridge_whitelist(host_a, false, warn) && 671 __host_bridge_whitelist(host_b, false, warn)) 672 return true; 673 674 return false; 675 } 676 677 static unsigned long map_types_idx(struct pci_dev *client) 678 { 679 return (pci_domain_nr(client->bus) << 16) | pci_dev_id(client); 680 } 681 682 /* 683 * Calculate the P2PDMA mapping type and distance between two PCI devices. 684 * 685 * If the two devices are the same PCI function, return 686 * PCI_P2PDMA_MAP_BUS_ADDR and a distance of 0. 687 * 688 * If they are two functions of the same device, return 689 * PCI_P2PDMA_MAP_BUS_ADDR and a distance of 2 (one hop up to the bridge, 690 * then one hop back down to another function of the same device). 691 * 692 * In the case where two devices are connected to the same PCIe switch, 693 * return a distance of 4. This corresponds to the following PCI tree: 694 * 695 * -+ Root Port 696 * \+ Switch Upstream Port 697 * +-+ Switch Downstream Port 0 698 * + \- Device A 699 * \-+ Switch Downstream Port 1 700 * \- Device B 701 * 702 * The distance is 4 because we traverse from Device A to Downstream Port 0 703 * to the common Switch Upstream Port, back down to Downstream Port 1 and 704 * then to Device B. The mapping type returned depends on the ACS 705 * redirection setting of the ports along the path. 706 * 707 * If ACS redirect is set on any port in the path, traffic between the 708 * devices will go through the host bridge, so return 709 * PCI_P2PDMA_MAP_THRU_HOST_BRIDGE; otherwise return 710 * PCI_P2PDMA_MAP_BUS_ADDR. 711 * 712 * Any two devices that have a data path that goes through the host bridge 713 * will consult a whitelist. If the host bridge is in the whitelist, return 714 * PCI_P2PDMA_MAP_THRU_HOST_BRIDGE with the distance set to the number of 715 * ports per above. If the device is not in the whitelist, return 716 * PCI_P2PDMA_MAP_NOT_SUPPORTED. 717 */ 718 static enum pci_p2pdma_map_type 719 calc_map_type_and_dist(struct pci_dev *provider, struct pci_dev *client, 720 int *dist, bool verbose) 721 { 722 enum pci_p2pdma_map_type map_type = PCI_P2PDMA_MAP_THRU_HOST_BRIDGE; 723 struct pci_dev *a = provider, *b = client, *bb; 724 struct pci_p2pdma *p2pdma; 725 struct seq_buf acs_list; 726 int acs_cnt = 0; 727 int dist_a = 0; 728 int dist_b = 0; 729 char buf[128]; 730 731 seq_buf_init(&acs_list, buf, sizeof(buf)); 732 733 /* 734 * Note, we don't need to take references to devices returned by 735 * pci_upstream_bridge() seeing we hold a reference to a child 736 * device which will already hold a reference to the upstream bridge. 737 */ 738 while (a) { 739 dist_b = 0; 740 741 if (pci_bridge_has_acs_redir(a)) { 742 seq_buf_print_bus_devfn(&acs_list, a); 743 acs_cnt++; 744 } 745 746 bb = b; 747 748 while (bb) { 749 if (a == bb) 750 goto check_b_path_acs; 751 752 bb = pci_upstream_bridge(bb); 753 dist_b++; 754 } 755 756 a = pci_upstream_bridge(a); 757 dist_a++; 758 } 759 760 *dist = dist_a + dist_b; 761 goto map_through_host_bridge; 762 763 check_b_path_acs: 764 bb = b; 765 766 while (bb) { 767 if (a == bb) 768 break; 769 770 if (pci_bridge_has_acs_redir(bb)) { 771 seq_buf_print_bus_devfn(&acs_list, bb); 772 acs_cnt++; 773 } 774 775 bb = pci_upstream_bridge(bb); 776 } 777 778 *dist = dist_a + dist_b; 779 780 if (!acs_cnt) { 781 map_type = PCI_P2PDMA_MAP_BUS_ADDR; 782 goto done; 783 } 784 785 if (verbose) { 786 /* Drop the final semicolon; the list is not empty here */ 787 if (!seq_buf_has_overflowed(&acs_list)) 788 acs_list.buffer[acs_list.len - 1] = '\0'; 789 pci_warn(client, "ACS redirect is set between the client and provider (%s)\n", 790 pci_name(provider)); 791 pci_warn(client, "to disable ACS redirect for this path, add the kernel parameter: pci=disable_acs_redir=%s\n", 792 seq_buf_str(&acs_list)); 793 } 794 795 map_through_host_bridge: 796 if (!cpu_supports_p2pdma() && 797 !host_bridge_whitelist(provider, client, verbose)) { 798 if (verbose) 799 pci_warn(client, "cannot be used for peer-to-peer DMA as the client and provider (%s) do not share an upstream bridge or whitelisted host bridge\n", 800 pci_name(provider)); 801 map_type = PCI_P2PDMA_MAP_NOT_SUPPORTED; 802 } 803 done: 804 rcu_read_lock(); 805 p2pdma = rcu_dereference(provider->p2pdma); 806 if (p2pdma) 807 xa_store(&p2pdma->map_types, map_types_idx(client), 808 xa_mk_value(map_type), GFP_ATOMIC); 809 rcu_read_unlock(); 810 return map_type; 811 } 812 813 /** 814 * pci_p2pdma_distance_many - Determine the cumulative distance between 815 * a p2pdma provider and the clients in use. 816 * @provider: p2pdma provider to check against the client list 817 * @clients: array of devices to check (NULL-terminated) 818 * @num_clients: number of clients in the array 819 * @verbose: if true, print warnings for devices when we return -1 820 * 821 * Returns -1 if any of the clients are not compatible, otherwise returns a 822 * positive number where a lower number is the preferable choice. (If there's 823 * one client that's the same as the provider it will return 0, which is best 824 * choice). 825 * 826 * "compatible" means the provider and the clients are either all behind 827 * the same PCI root port or the host bridges connected to each of the devices 828 * are listed in the 'pci_p2pdma_whitelist'. 829 */ 830 int pci_p2pdma_distance_many(struct pci_dev *provider, struct device **clients, 831 int num_clients, bool verbose) 832 { 833 enum pci_p2pdma_map_type map; 834 bool not_supported = false; 835 struct pci_dev *pci_client; 836 int total_dist = 0; 837 int i, distance; 838 839 if (num_clients == 0) 840 return -1; 841 842 for (i = 0; i < num_clients; i++) { 843 pci_client = find_parent_pci_dev(clients[i]); 844 if (!pci_client) { 845 if (verbose) 846 dev_warn(clients[i], 847 "cannot be used for peer-to-peer DMA as it is not a PCI device\n"); 848 return -1; 849 } 850 851 map = calc_map_type_and_dist(provider, pci_client, &distance, 852 verbose); 853 854 pci_dev_put(pci_client); 855 856 if (map == PCI_P2PDMA_MAP_NOT_SUPPORTED) 857 not_supported = true; 858 859 if (not_supported && !verbose) 860 break; 861 862 total_dist += distance; 863 } 864 865 if (not_supported) 866 return -1; 867 868 return total_dist; 869 } 870 EXPORT_SYMBOL_GPL(pci_p2pdma_distance_many); 871 872 /** 873 * pci_has_p2pmem - check if a given PCI device has published any p2pmem 874 * @pdev: PCI device to check 875 */ 876 static bool pci_has_p2pmem(struct pci_dev *pdev) 877 { 878 struct pci_p2pdma *p2pdma; 879 bool res; 880 881 rcu_read_lock(); 882 p2pdma = rcu_dereference(pdev->p2pdma); 883 884 /* 885 * The callers hand the result to pci_alloc_p2pmem(), so only a 886 * provider backed by a pool is of any use here. pcim_p2pdma_init() 887 * creates providers without one. 888 */ 889 res = p2pdma && p2pdma->pool && p2pdma->p2pmem_published; 890 rcu_read_unlock(); 891 892 return res; 893 } 894 895 /** 896 * pci_p2pmem_find_many - find a peer-to-peer DMA memory device compatible with 897 * the specified list of clients and shortest distance 898 * @clients: array of devices to check (NULL-terminated) 899 * @num_clients: number of client devices in the list 900 * 901 * If multiple devices are behind the same switch, the one "closest" to the 902 * client devices in use will be chosen first. (So if one of the providers is 903 * the same as one of the clients, that provider will be used ahead of any 904 * other providers that are unrelated). If multiple providers are an equal 905 * distance away, one will be chosen at random. 906 * 907 * Returns a pointer to the PCI device with a reference taken (use pci_dev_put 908 * to return the reference) or NULL if no compatible device is found. The 909 * found provider will also be assigned to the client list. 910 */ 911 struct pci_dev *pci_p2pmem_find_many(struct device **clients, int num_clients) 912 { 913 struct pci_dev *pdev = NULL; 914 int distance; 915 int closest_distance = INT_MAX; 916 struct pci_dev **closest_pdevs; 917 int dev_cnt = 0; 918 const int max_devs = PAGE_SIZE / sizeof(*closest_pdevs); 919 int i; 920 921 closest_pdevs = kmalloc(PAGE_SIZE, GFP_KERNEL); 922 if (!closest_pdevs) 923 return NULL; 924 925 for_each_pci_dev(pdev) { 926 if (!pci_has_p2pmem(pdev)) 927 continue; 928 929 distance = pci_p2pdma_distance_many(pdev, clients, 930 num_clients, false); 931 if (distance < 0 || distance > closest_distance) 932 continue; 933 934 if (distance == closest_distance && dev_cnt >= max_devs) 935 continue; 936 937 if (distance < closest_distance) { 938 for (i = 0; i < dev_cnt; i++) 939 pci_dev_put(closest_pdevs[i]); 940 941 dev_cnt = 0; 942 closest_distance = distance; 943 } 944 945 closest_pdevs[dev_cnt++] = pci_dev_get(pdev); 946 } 947 948 if (dev_cnt) 949 pdev = pci_dev_get(closest_pdevs[get_random_u32_below(dev_cnt)]); 950 951 for (i = 0; i < dev_cnt; i++) 952 pci_dev_put(closest_pdevs[i]); 953 954 kfree(closest_pdevs); 955 return pdev; 956 } 957 EXPORT_SYMBOL_GPL(pci_p2pmem_find_many); 958 959 /** 960 * pci_alloc_p2pmem - allocate peer-to-peer DMA memory 961 * @pdev: the device to allocate memory from 962 * @size: number of bytes to allocate 963 * 964 * Returns the allocated memory or NULL on error. 965 */ 966 void *pci_alloc_p2pmem(struct pci_dev *pdev, size_t size) 967 { 968 void *ret = NULL; 969 struct percpu_ref *ref; 970 struct pci_p2pdma *p2pdma; 971 972 /* 973 * Pairs with synchronize_rcu() in pci_p2pdma_release() to 974 * ensure pdev->p2pdma is non-NULL for the duration of the 975 * read-lock. 976 */ 977 rcu_read_lock(); 978 p2pdma = rcu_dereference(pdev->p2pdma); 979 if (unlikely(!p2pdma)) 980 goto out; 981 982 ret = (void *)gen_pool_alloc_owner(p2pdma->pool, size, (void **) &ref); 983 if (!ret) 984 goto out; 985 986 if (unlikely(!percpu_ref_tryget_live_rcu(ref))) { 987 gen_pool_free(p2pdma->pool, (unsigned long) ret, size); 988 ret = NULL; 989 } 990 out: 991 rcu_read_unlock(); 992 return ret; 993 } 994 EXPORT_SYMBOL_GPL(pci_alloc_p2pmem); 995 996 /** 997 * pci_free_p2pmem - free peer-to-peer DMA memory 998 * @pdev: the device the memory was allocated from 999 * @addr: address of the memory that was allocated 1000 * @size: number of bytes that were allocated 1001 */ 1002 void pci_free_p2pmem(struct pci_dev *pdev, void *addr, size_t size) 1003 { 1004 struct percpu_ref *ref; 1005 struct pci_p2pdma *p2pdma = rcu_dereference_protected(pdev->p2pdma, 1); 1006 1007 gen_pool_free_owner(p2pdma->pool, (uintptr_t)addr, size, 1008 (void **) &ref); 1009 percpu_ref_put(ref); 1010 } 1011 EXPORT_SYMBOL_GPL(pci_free_p2pmem); 1012 1013 /** 1014 * pci_p2pmem_virt_to_bus - return the PCI bus address for a given virtual 1015 * address obtained with pci_alloc_p2pmem() 1016 * @pdev: the device the memory was allocated from 1017 * @addr: address of the memory that was allocated 1018 */ 1019 pci_bus_addr_t pci_p2pmem_virt_to_bus(struct pci_dev *pdev, void *addr) 1020 { 1021 struct pci_p2pdma *p2pdma; 1022 1023 if (!addr) 1024 return 0; 1025 1026 p2pdma = rcu_dereference_protected(pdev->p2pdma, 1); 1027 if (!p2pdma) 1028 return 0; 1029 1030 /* 1031 * Note: when we added the memory to the pool we used the PCI 1032 * bus address as the physical address. So gen_pool_virt_to_phys() 1033 * actually returns the bus address despite the misleading name. 1034 */ 1035 return gen_pool_virt_to_phys(p2pdma->pool, (unsigned long)addr); 1036 } 1037 EXPORT_SYMBOL_GPL(pci_p2pmem_virt_to_bus); 1038 1039 /** 1040 * pci_p2pmem_alloc_sgl - allocate peer-to-peer DMA memory in a scatterlist 1041 * @pdev: the device to allocate memory from 1042 * @nents: the number of SG entries in the list 1043 * @length: number of bytes to allocate 1044 * 1045 * Return: %NULL on error or &struct scatterlist pointer and @nents on success 1046 */ 1047 struct scatterlist *pci_p2pmem_alloc_sgl(struct pci_dev *pdev, 1048 unsigned int *nents, u32 length) 1049 { 1050 struct scatterlist *sg; 1051 void *addr; 1052 1053 sg = kmalloc_obj(*sg); 1054 if (!sg) 1055 return NULL; 1056 1057 sg_init_table(sg, 1); 1058 1059 addr = pci_alloc_p2pmem(pdev, length); 1060 if (!addr) 1061 goto out_free_sg; 1062 1063 sg_set_buf(sg, addr, length); 1064 *nents = 1; 1065 return sg; 1066 1067 out_free_sg: 1068 kfree(sg); 1069 return NULL; 1070 } 1071 EXPORT_SYMBOL_GPL(pci_p2pmem_alloc_sgl); 1072 1073 /** 1074 * pci_p2pmem_free_sgl - free a scatterlist allocated by pci_p2pmem_alloc_sgl() 1075 * @pdev: the device to allocate memory from 1076 * @sgl: the allocated scatterlist 1077 */ 1078 void pci_p2pmem_free_sgl(struct pci_dev *pdev, struct scatterlist *sgl) 1079 { 1080 struct scatterlist *sg; 1081 int count; 1082 1083 for_each_sg(sgl, sg, INT_MAX, count) { 1084 if (!sg) 1085 break; 1086 1087 pci_free_p2pmem(pdev, sg_virt(sg), sg->length); 1088 } 1089 kfree(sgl); 1090 } 1091 EXPORT_SYMBOL_GPL(pci_p2pmem_free_sgl); 1092 1093 /** 1094 * pci_p2pmem_publish - publish the peer-to-peer DMA memory for use by 1095 * other devices with pci_p2pmem_find() 1096 * @pdev: the device with peer-to-peer DMA memory to publish 1097 * @publish: set to true to publish the memory, false to unpublish it 1098 * 1099 * Published memory can be used by other PCI device drivers for 1100 * peer-2-peer DMA operations. Non-published memory is reserved for 1101 * exclusive use of the device driver that registers the peer-to-peer 1102 * memory. 1103 */ 1104 void pci_p2pmem_publish(struct pci_dev *pdev, bool publish) 1105 { 1106 struct pci_p2pdma *p2pdma; 1107 1108 rcu_read_lock(); 1109 p2pdma = rcu_dereference(pdev->p2pdma); 1110 if (p2pdma) 1111 p2pdma->p2pmem_published = publish; 1112 rcu_read_unlock(); 1113 } 1114 EXPORT_SYMBOL_GPL(pci_p2pmem_publish); 1115 1116 /** 1117 * pci_p2pdma_map_type - Determine the mapping type for P2PDMA transfers 1118 * @provider: P2PDMA provider structure 1119 * @dev: Target device for the transfer 1120 * 1121 * Determines how peer-to-peer DMA transfers should be mapped between 1122 * the provider and the target device. The mapping type indicates whether 1123 * the transfer can be done directly through PCI switches or must go 1124 * through the host bridge. 1125 */ 1126 enum pci_p2pdma_map_type pci_p2pdma_map_type(struct p2pdma_provider *provider, 1127 struct device *dev) 1128 { 1129 enum pci_p2pdma_map_type type = PCI_P2PDMA_MAP_NOT_SUPPORTED; 1130 struct pci_dev *pdev = to_pci_dev(provider->owner); 1131 struct pci_dev *client; 1132 struct pci_p2pdma *p2pdma; 1133 int dist; 1134 1135 if (!pdev->p2pdma) 1136 return PCI_P2PDMA_MAP_NOT_SUPPORTED; 1137 1138 if (!dev_is_pci(dev)) 1139 return PCI_P2PDMA_MAP_NOT_SUPPORTED; 1140 1141 client = to_pci_dev(dev); 1142 1143 rcu_read_lock(); 1144 p2pdma = rcu_dereference(pdev->p2pdma); 1145 1146 if (p2pdma) 1147 type = xa_to_value(xa_load(&p2pdma->map_types, 1148 map_types_idx(client))); 1149 rcu_read_unlock(); 1150 1151 if (type == PCI_P2PDMA_MAP_UNKNOWN) 1152 return calc_map_type_and_dist(pdev, client, &dist, true); 1153 1154 return type; 1155 } 1156 1157 void __pci_p2pdma_update_state(struct pci_p2pdma_map_state *state, 1158 struct device *dev, struct page *page) 1159 { 1160 struct pci_p2pdma_pagemap *p2p_pgmap = to_p2p_pgmap(page_pgmap(page)); 1161 1162 if (state->mem == p2p_pgmap->mem) 1163 return; 1164 1165 state->mem = p2p_pgmap->mem; 1166 state->map = pci_p2pdma_map_type(p2p_pgmap->mem, dev); 1167 } 1168 1169 /** 1170 * pci_p2pdma_enable_store - parse a configfs/sysfs attribute store 1171 * to enable p2pdma 1172 * @page: contents of the value to be stored 1173 * @p2p_dev: returns the PCI device that was selected to be used 1174 * (if one was specified in the stored value) 1175 * @use_p2pdma: returns whether to enable p2pdma or not 1176 * 1177 * Parses an attribute value to decide whether to enable p2pdma. 1178 * The value can select a PCI device (using its full BDF device 1179 * name) or a boolean (in any format kstrtobool() accepts). A false 1180 * value disables p2pdma, a true value expects the caller 1181 * to automatically find a compatible device and specifying a PCI device 1182 * expects the caller to use the specific provider. 1183 * 1184 * pci_p2pdma_enable_show() should be used as the show operation for 1185 * the attribute. 1186 * 1187 * Returns 0 on success 1188 */ 1189 int pci_p2pdma_enable_store(const char *page, struct pci_dev **p2p_dev, 1190 bool *use_p2pdma) 1191 { 1192 struct device *dev; 1193 1194 dev = bus_find_device_by_name(&pci_bus_type, NULL, page); 1195 if (dev) { 1196 *use_p2pdma = true; 1197 *p2p_dev = to_pci_dev(dev); 1198 1199 if (!pci_has_p2pmem(*p2p_dev)) { 1200 pci_err(*p2p_dev, 1201 "PCI device has no peer-to-peer memory: %s\n", 1202 page); 1203 pci_dev_put(*p2p_dev); 1204 return -ENODEV; 1205 } 1206 1207 return 0; 1208 } else if ((page[0] == '0' || page[0] == '1') && !iscntrl(page[1])) { 1209 /* 1210 * If the user enters a PCI device that doesn't exist 1211 * like "0000:01:00.1", we don't want kstrtobool to think 1212 * it's a '0' when it's clearly not what the user wanted. 1213 * So we require 0's and 1's to be exactly one character. 1214 */ 1215 } else if (!kstrtobool(page, use_p2pdma)) { 1216 return 0; 1217 } 1218 1219 pr_err("No such PCI device: %.*s\n", (int)strcspn(page, "\n"), page); 1220 return -ENODEV; 1221 } 1222 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_store); 1223 1224 /** 1225 * pci_p2pdma_enable_show - show a configfs/sysfs attribute indicating 1226 * whether p2pdma is enabled 1227 * @page: contents of the stored value 1228 * @p2p_dev: the selected p2p device (NULL if no device is selected) 1229 * @use_p2pdma: whether p2pdma has been enabled 1230 * 1231 * Attributes that use pci_p2pdma_enable_store() should use this function 1232 * to show the value of the attribute. 1233 * 1234 * Returns 0 on success 1235 */ 1236 ssize_t pci_p2pdma_enable_show(char *page, struct pci_dev *p2p_dev, 1237 bool use_p2pdma) 1238 { 1239 if (!use_p2pdma) 1240 return sprintf(page, "0\n"); 1241 1242 if (!p2p_dev) 1243 return sprintf(page, "1\n"); 1244 1245 return sprintf(page, "%s\n", pci_name(p2p_dev)); 1246 } 1247 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show); 1248