1 // SPDX-License-Identifier: GPL-2.0
3 * PCI Peer 2 Peer DMA support.
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.
11 #define pr_fmt(fmt) "pci-p2pdma: " fmt
12 #include <linux/ctype.h>
13 #include <linux/pci-p2pdma.h>
14 #include <linux/module.h>
15 #include <linux/slab.h>
16 #include <linux/genalloc.h>
17 #include <linux/memremap.h>
18 #include <linux/percpu-refcount.h>
19 #include <linux/random.h>
20 #include <linux/seq_buf.h>
21 #include <linux/iommu.h>
24 struct gen_pool *pool;
25 bool p2pmem_published;
28 struct p2pdma_pagemap {
29 struct dev_pagemap pgmap;
30 struct percpu_ref ref;
31 struct completion ref_done;
34 static ssize_t size_show(struct device *dev, struct device_attribute *attr,
37 struct pci_dev *pdev = to_pci_dev(dev);
40 if (pdev->p2pdma->pool)
41 size = gen_pool_size(pdev->p2pdma->pool);
43 return snprintf(buf, PAGE_SIZE, "%zd\n", size);
45 static DEVICE_ATTR_RO(size);
47 static ssize_t available_show(struct device *dev, struct device_attribute *attr,
50 struct pci_dev *pdev = to_pci_dev(dev);
53 if (pdev->p2pdma->pool)
54 avail = gen_pool_avail(pdev->p2pdma->pool);
56 return snprintf(buf, PAGE_SIZE, "%zd\n", avail);
58 static DEVICE_ATTR_RO(available);
60 static ssize_t published_show(struct device *dev, struct device_attribute *attr,
63 struct pci_dev *pdev = to_pci_dev(dev);
65 return snprintf(buf, PAGE_SIZE, "%d\n",
66 pdev->p2pdma->p2pmem_published);
68 static DEVICE_ATTR_RO(published);
70 static struct attribute *p2pmem_attrs[] = {
72 &dev_attr_available.attr,
73 &dev_attr_published.attr,
77 static const struct attribute_group p2pmem_group = {
78 .attrs = p2pmem_attrs,
82 static struct p2pdma_pagemap *to_p2p_pgmap(struct percpu_ref *ref)
84 return container_of(ref, struct p2pdma_pagemap, ref);
87 static void pci_p2pdma_percpu_release(struct percpu_ref *ref)
89 struct p2pdma_pagemap *p2p_pgmap = to_p2p_pgmap(ref);
91 complete(&p2p_pgmap->ref_done);
94 static void pci_p2pdma_percpu_kill(struct percpu_ref *ref)
99 static void pci_p2pdma_percpu_cleanup(struct percpu_ref *ref)
101 struct p2pdma_pagemap *p2p_pgmap = to_p2p_pgmap(ref);
103 wait_for_completion(&p2p_pgmap->ref_done);
104 percpu_ref_exit(&p2p_pgmap->ref);
107 static void pci_p2pdma_release(void *data)
109 struct pci_dev *pdev = data;
110 struct pci_p2pdma *p2pdma = pdev->p2pdma;
115 /* Flush and disable pci_alloc_p2p_mem() */
119 gen_pool_destroy(p2pdma->pool);
120 sysfs_remove_group(&pdev->dev.kobj, &p2pmem_group);
123 static int pci_p2pdma_setup(struct pci_dev *pdev)
126 struct pci_p2pdma *p2p;
128 p2p = devm_kzalloc(&pdev->dev, sizeof(*p2p), GFP_KERNEL);
132 p2p->pool = gen_pool_create(PAGE_SHIFT, dev_to_node(&pdev->dev));
136 error = devm_add_action_or_reset(&pdev->dev, pci_p2pdma_release, pdev);
138 goto out_pool_destroy;
142 error = sysfs_create_group(&pdev->dev.kobj, &p2pmem_group);
144 goto out_pool_destroy;
150 gen_pool_destroy(p2p->pool);
152 devm_kfree(&pdev->dev, p2p);
157 * pci_p2pdma_add_resource - add memory for use as p2p memory
158 * @pdev: the device to add the memory to
159 * @bar: PCI BAR to add
160 * @size: size of the memory to add, may be zero to use the whole BAR
161 * @offset: offset into the PCI BAR
163 * The memory will be given ZONE_DEVICE struct pages so that it may
164 * be used with any DMA request.
166 int pci_p2pdma_add_resource(struct pci_dev *pdev, int bar, size_t size,
169 struct p2pdma_pagemap *p2p_pgmap;
170 struct dev_pagemap *pgmap;
174 if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM))
177 if (offset >= pci_resource_len(pdev, bar))
181 size = pci_resource_len(pdev, bar) - offset;
183 if (size + offset > pci_resource_len(pdev, bar))
187 error = pci_p2pdma_setup(pdev);
192 p2p_pgmap = devm_kzalloc(&pdev->dev, sizeof(*p2p_pgmap), GFP_KERNEL);
196 init_completion(&p2p_pgmap->ref_done);
197 error = percpu_ref_init(&p2p_pgmap->ref,
198 pci_p2pdma_percpu_release, 0, GFP_KERNEL);
202 pgmap = &p2p_pgmap->pgmap;
204 pgmap->res.start = pci_resource_start(pdev, bar) + offset;
205 pgmap->res.end = pgmap->res.start + size - 1;
206 pgmap->res.flags = pci_resource_flags(pdev, bar);
207 pgmap->ref = &p2p_pgmap->ref;
208 pgmap->type = MEMORY_DEVICE_PCI_P2PDMA;
209 pgmap->pci_p2pdma_bus_offset = pci_bus_address(pdev, bar) -
210 pci_resource_start(pdev, bar);
211 pgmap->kill = pci_p2pdma_percpu_kill;
212 pgmap->cleanup = pci_p2pdma_percpu_cleanup;
214 addr = devm_memremap_pages(&pdev->dev, pgmap);
216 error = PTR_ERR(addr);
220 error = gen_pool_add_owner(pdev->p2pdma->pool, (unsigned long)addr,
221 pci_bus_address(pdev, bar) + offset,
222 resource_size(&pgmap->res), dev_to_node(&pdev->dev),
227 pci_info(pdev, "added peer-to-peer DMA memory %pR\n",
233 devm_memunmap_pages(&pdev->dev, pgmap);
235 devm_kfree(&pdev->dev, p2p_pgmap);
238 EXPORT_SYMBOL_GPL(pci_p2pdma_add_resource);
241 * Note this function returns the parent PCI device with a
242 * reference taken. It is the caller's responsibily to drop
245 static struct pci_dev *find_parent_pci_dev(struct device *dev)
247 struct device *parent;
249 dev = get_device(dev);
253 return to_pci_dev(dev);
255 parent = get_device(dev->parent);
264 * Check if a PCI bridge has its ACS redirection bits set to redirect P2P
265 * TLPs upstream via ACS. Returns 1 if the packets will be redirected
266 * upstream, 0 otherwise.
268 static int pci_bridge_has_acs_redir(struct pci_dev *pdev)
273 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ACS);
277 pci_read_config_word(pdev, pos + PCI_ACS_CTRL, &ctrl);
279 if (ctrl & (PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_EC))
285 static void seq_buf_print_bus_devfn(struct seq_buf *buf, struct pci_dev *pdev)
290 seq_buf_printf(buf, "%s;", pci_name(pdev));
294 * If we can't find a common upstream bridge take a look at the root
295 * complex and compare it to a whitelist of known good hardware.
297 static bool root_complex_whitelist(struct pci_dev *dev)
299 struct pci_host_bridge *host = pci_find_host_bridge(dev->bus);
300 struct pci_dev *root = pci_get_slot(host->bus, PCI_DEVFN(0, 0));
301 unsigned short vendor, device;
303 if (iommu_present(dev->dev.bus))
309 vendor = root->vendor;
310 device = root->device;
313 /* AMD ZEN host bridges can do peer to peer */
314 if (vendor == PCI_VENDOR_ID_AMD && device == 0x1450)
321 * Find the distance through the nearest common upstream bridge between
324 * If the two devices are the same device then 0 will be returned.
326 * If there are two virtual functions of the same device behind the same
327 * bridge port then 2 will be returned (one step down to the PCIe switch,
328 * then one step back to the same device).
330 * In the case where two devices are connected to the same PCIe switch, the
331 * value 4 will be returned. This corresponds to the following PCI tree:
334 * \+ Switch Upstream Port
335 * +-+ Switch Downstream Port
337 * \-+ Switch Downstream Port
340 * The distance is 4 because we traverse from Device A through the downstream
341 * port of the switch, to the common upstream port, back up to the second
342 * downstream port and then to Device B.
344 * Any two devices that don't have a common upstream bridge will return -1.
345 * In this way devices on separate PCIe root ports will be rejected, which
346 * is what we want for peer-to-peer seeing each PCIe root port defines a
347 * separate hierarchy domain and there's no way to determine whether the root
348 * complex supports forwarding between them.
350 * In the case where two devices are connected to different PCIe switches,
351 * this function will still return a positive distance as long as both
352 * switches eventually have a common upstream bridge. Note this covers
353 * the case of using multiple PCIe switches to achieve a desired level of
354 * fan-out from a root port. The exact distance will be a function of the
355 * number of switches between Device A and Device B.
357 * If a bridge which has any ACS redirection bits set is in the path
358 * then this functions will return -2. This is so we reject any
359 * cases where the TLPs are forwarded up into the root complex.
360 * In this case, a list of all infringing bridge addresses will be
361 * populated in acs_list (assuming it's non-null) for printk purposes.
363 static int upstream_bridge_distance(struct pci_dev *provider,
364 struct pci_dev *client,
365 struct seq_buf *acs_list)
367 struct pci_dev *a = provider, *b = client, *bb;
373 * Note, we don't need to take references to devices returned by
374 * pci_upstream_bridge() seeing we hold a reference to a child
375 * device which will already hold a reference to the upstream bridge.
381 if (pci_bridge_has_acs_redir(a)) {
382 seq_buf_print_bus_devfn(acs_list, a);
390 goto check_b_path_acs;
392 bb = pci_upstream_bridge(bb);
396 a = pci_upstream_bridge(a);
401 * Allow the connection if both devices are on a whitelisted root
402 * complex, but add an arbitary large value to the distance.
404 if (root_complex_whitelist(provider) &&
405 root_complex_whitelist(client))
406 return 0x1000 + dist_a + dist_b;
417 if (pci_bridge_has_acs_redir(bb)) {
418 seq_buf_print_bus_devfn(acs_list, bb);
422 bb = pci_upstream_bridge(bb);
428 return dist_a + dist_b;
431 static int upstream_bridge_distance_warn(struct pci_dev *provider,
432 struct pci_dev *client)
434 struct seq_buf acs_list;
437 seq_buf_init(&acs_list, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
438 if (!acs_list.buffer)
441 ret = upstream_bridge_distance(provider, client, &acs_list);
443 pci_warn(client, "cannot be used for peer-to-peer DMA as ACS redirect is set between the client and provider (%s)\n",
445 /* Drop final semicolon */
446 acs_list.buffer[acs_list.len-1] = 0;
447 pci_warn(client, "to disable ACS redirect for this path, add the kernel parameter: pci=disable_acs_redir=%s\n",
450 } else if (ret < 0) {
451 pci_warn(client, "cannot be used for peer-to-peer DMA as the client and provider (%s) do not share an upstream bridge\n",
455 kfree(acs_list.buffer);
461 * pci_p2pdma_distance_many - Determive the cumulative distance between
462 * a p2pdma provider and the clients in use.
463 * @provider: p2pdma provider to check against the client list
464 * @clients: array of devices to check (NULL-terminated)
465 * @num_clients: number of clients in the array
466 * @verbose: if true, print warnings for devices when we return -1
468 * Returns -1 if any of the clients are not compatible (behind the same
469 * root port as the provider), otherwise returns a positive number where
470 * a lower number is the preferable choice. (If there's one client
471 * that's the same as the provider it will return 0, which is best choice).
473 * For now, "compatible" means the provider and the clients are all behind
474 * the same PCI root port. This cuts out cases that may work but is safest
475 * for the user. Future work can expand this to white-list root complexes that
476 * can safely forward between each ports.
478 int pci_p2pdma_distance_many(struct pci_dev *provider, struct device **clients,
479 int num_clients, bool verbose)
481 bool not_supported = false;
482 struct pci_dev *pci_client;
486 if (num_clients == 0)
489 for (i = 0; i < num_clients; i++) {
490 pci_client = find_parent_pci_dev(clients[i]);
494 "cannot be used for peer-to-peer DMA as it is not a PCI device\n");
499 ret = upstream_bridge_distance_warn(provider,
502 ret = upstream_bridge_distance(provider, pci_client,
505 pci_dev_put(pci_client);
508 not_supported = true;
510 if (not_supported && !verbose)
521 EXPORT_SYMBOL_GPL(pci_p2pdma_distance_many);
524 * pci_has_p2pmem - check if a given PCI device has published any p2pmem
525 * @pdev: PCI device to check
527 bool pci_has_p2pmem(struct pci_dev *pdev)
529 return pdev->p2pdma && pdev->p2pdma->p2pmem_published;
531 EXPORT_SYMBOL_GPL(pci_has_p2pmem);
534 * pci_p2pmem_find - find a peer-to-peer DMA memory device compatible with
535 * the specified list of clients and shortest distance (as determined
536 * by pci_p2pmem_dma())
537 * @clients: array of devices to check (NULL-terminated)
538 * @num_clients: number of client devices in the list
540 * If multiple devices are behind the same switch, the one "closest" to the
541 * client devices in use will be chosen first. (So if one of the providers is
542 * the same as one of the clients, that provider will be used ahead of any
543 * other providers that are unrelated). If multiple providers are an equal
544 * distance away, one will be chosen at random.
546 * Returns a pointer to the PCI device with a reference taken (use pci_dev_put
547 * to return the reference) or NULL if no compatible device is found. The
548 * found provider will also be assigned to the client list.
550 struct pci_dev *pci_p2pmem_find_many(struct device **clients, int num_clients)
552 struct pci_dev *pdev = NULL;
554 int closest_distance = INT_MAX;
555 struct pci_dev **closest_pdevs;
557 const int max_devs = PAGE_SIZE / sizeof(*closest_pdevs);
560 closest_pdevs = kmalloc(PAGE_SIZE, GFP_KERNEL);
564 while ((pdev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, pdev))) {
565 if (!pci_has_p2pmem(pdev))
568 distance = pci_p2pdma_distance_many(pdev, clients,
570 if (distance < 0 || distance > closest_distance)
573 if (distance == closest_distance && dev_cnt >= max_devs)
576 if (distance < closest_distance) {
577 for (i = 0; i < dev_cnt; i++)
578 pci_dev_put(closest_pdevs[i]);
581 closest_distance = distance;
584 closest_pdevs[dev_cnt++] = pci_dev_get(pdev);
588 pdev = pci_dev_get(closest_pdevs[prandom_u32_max(dev_cnt)]);
590 for (i = 0; i < dev_cnt; i++)
591 pci_dev_put(closest_pdevs[i]);
593 kfree(closest_pdevs);
596 EXPORT_SYMBOL_GPL(pci_p2pmem_find_many);
599 * pci_alloc_p2p_mem - allocate peer-to-peer DMA memory
600 * @pdev: the device to allocate memory from
601 * @size: number of bytes to allocate
603 * Returns the allocated memory or NULL on error.
605 void *pci_alloc_p2pmem(struct pci_dev *pdev, size_t size)
608 struct percpu_ref *ref;
611 * Pairs with synchronize_rcu() in pci_p2pdma_release() to
612 * ensure pdev->p2pdma is non-NULL for the duration of the
616 if (unlikely(!pdev->p2pdma))
619 ret = (void *)gen_pool_alloc_owner(pdev->p2pdma->pool, size,
624 if (unlikely(!percpu_ref_tryget_live(ref))) {
625 gen_pool_free(pdev->p2pdma->pool, (unsigned long) ret, size);
633 EXPORT_SYMBOL_GPL(pci_alloc_p2pmem);
636 * pci_free_p2pmem - free peer-to-peer DMA memory
637 * @pdev: the device the memory was allocated from
638 * @addr: address of the memory that was allocated
639 * @size: number of bytes that were allocated
641 void pci_free_p2pmem(struct pci_dev *pdev, void *addr, size_t size)
643 struct percpu_ref *ref;
645 gen_pool_free_owner(pdev->p2pdma->pool, (uintptr_t)addr, size,
649 EXPORT_SYMBOL_GPL(pci_free_p2pmem);
652 * pci_virt_to_bus - return the PCI bus address for a given virtual
653 * address obtained with pci_alloc_p2pmem()
654 * @pdev: the device the memory was allocated from
655 * @addr: address of the memory that was allocated
657 pci_bus_addr_t pci_p2pmem_virt_to_bus(struct pci_dev *pdev, void *addr)
665 * Note: when we added the memory to the pool we used the PCI
666 * bus address as the physical address. So gen_pool_virt_to_phys()
667 * actually returns the bus address despite the misleading name.
669 return gen_pool_virt_to_phys(pdev->p2pdma->pool, (unsigned long)addr);
671 EXPORT_SYMBOL_GPL(pci_p2pmem_virt_to_bus);
674 * pci_p2pmem_alloc_sgl - allocate peer-to-peer DMA memory in a scatterlist
675 * @pdev: the device to allocate memory from
676 * @nents: the number of SG entries in the list
677 * @length: number of bytes to allocate
679 * Return: %NULL on error or &struct scatterlist pointer and @nents on success
681 struct scatterlist *pci_p2pmem_alloc_sgl(struct pci_dev *pdev,
682 unsigned int *nents, u32 length)
684 struct scatterlist *sg;
687 sg = kzalloc(sizeof(*sg), GFP_KERNEL);
691 sg_init_table(sg, 1);
693 addr = pci_alloc_p2pmem(pdev, length);
697 sg_set_buf(sg, addr, length);
705 EXPORT_SYMBOL_GPL(pci_p2pmem_alloc_sgl);
708 * pci_p2pmem_free_sgl - free a scatterlist allocated by pci_p2pmem_alloc_sgl()
709 * @pdev: the device to allocate memory from
710 * @sgl: the allocated scatterlist
712 void pci_p2pmem_free_sgl(struct pci_dev *pdev, struct scatterlist *sgl)
714 struct scatterlist *sg;
717 for_each_sg(sgl, sg, INT_MAX, count) {
721 pci_free_p2pmem(pdev, sg_virt(sg), sg->length);
725 EXPORT_SYMBOL_GPL(pci_p2pmem_free_sgl);
728 * pci_p2pmem_publish - publish the peer-to-peer DMA memory for use by
729 * other devices with pci_p2pmem_find()
730 * @pdev: the device with peer-to-peer DMA memory to publish
731 * @publish: set to true to publish the memory, false to unpublish it
733 * Published memory can be used by other PCI device drivers for
734 * peer-2-peer DMA operations. Non-published memory is reserved for
735 * exclusive use of the device driver that registers the peer-to-peer
738 void pci_p2pmem_publish(struct pci_dev *pdev, bool publish)
741 pdev->p2pdma->p2pmem_published = publish;
743 EXPORT_SYMBOL_GPL(pci_p2pmem_publish);
746 * pci_p2pdma_map_sg - map a PCI peer-to-peer scatterlist for DMA
747 * @dev: device doing the DMA request
748 * @sg: scatter list to map
749 * @nents: elements in the scatterlist
750 * @dir: DMA direction
752 * Scatterlists mapped with this function should not be unmapped in any way.
754 * Returns the number of SG entries mapped or 0 on error.
756 int pci_p2pdma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
757 enum dma_data_direction dir)
759 struct dev_pagemap *pgmap;
760 struct scatterlist *s;
765 * p2pdma mappings are not compatible with devices that use
766 * dma_virt_ops. If the upper layers do the right thing
767 * this should never happen because it will be prevented
768 * by the check in pci_p2pdma_add_client()
770 if (WARN_ON_ONCE(IS_ENABLED(CONFIG_DMA_VIRT_OPS) &&
771 dev->dma_ops == &dma_virt_ops))
774 for_each_sg(sg, s, nents, i) {
775 pgmap = sg_page(s)->pgmap;
778 s->dma_address = paddr - pgmap->pci_p2pdma_bus_offset;
779 sg_dma_len(s) = s->length;
784 EXPORT_SYMBOL_GPL(pci_p2pdma_map_sg);
787 * pci_p2pdma_enable_store - parse a configfs/sysfs attribute store
789 * @page: contents of the value to be stored
790 * @p2p_dev: returns the PCI device that was selected to be used
791 * (if one was specified in the stored value)
792 * @use_p2pdma: returns whether to enable p2pdma or not
794 * Parses an attribute value to decide whether to enable p2pdma.
795 * The value can select a PCI device (using its full BDF device
796 * name) or a boolean (in any format strtobool() accepts). A false
797 * value disables p2pdma, a true value expects the caller
798 * to automatically find a compatible device and specifying a PCI device
799 * expects the caller to use the specific provider.
801 * pci_p2pdma_enable_show() should be used as the show operation for
804 * Returns 0 on success
806 int pci_p2pdma_enable_store(const char *page, struct pci_dev **p2p_dev,
811 dev = bus_find_device_by_name(&pci_bus_type, NULL, page);
814 *p2p_dev = to_pci_dev(dev);
816 if (!pci_has_p2pmem(*p2p_dev)) {
818 "PCI device has no peer-to-peer memory: %s\n",
820 pci_dev_put(*p2p_dev);
825 } else if ((page[0] == '0' || page[0] == '1') && !iscntrl(page[1])) {
827 * If the user enters a PCI device that doesn't exist
828 * like "0000:01:00.1", we don't want strtobool to think
829 * it's a '0' when it's clearly not what the user wanted.
830 * So we require 0's and 1's to be exactly one character.
832 } else if (!strtobool(page, use_p2pdma)) {
836 pr_err("No such PCI device: %.*s\n", (int)strcspn(page, "\n"), page);
839 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_store);
842 * pci_p2pdma_enable_show - show a configfs/sysfs attribute indicating
843 * whether p2pdma is enabled
844 * @page: contents of the stored value
845 * @p2p_dev: the selected p2p device (NULL if no device is selected)
846 * @use_p2pdma: whether p2pdma has been enabled
848 * Attributes that use pci_p2pdma_enable_store() should use this function
849 * to show the value of the attribute.
851 * Returns 0 on success
853 ssize_t pci_p2pdma_enable_show(char *page, struct pci_dev *p2p_dev,
857 return sprintf(page, "0\n");
860 return sprintf(page, "1\n");
862 return sprintf(page, "%s\n", pci_name(p2p_dev));
864 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show);