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