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 static ssize_t size_show(struct device *dev, struct device_attribute *attr,
31 struct pci_dev *pdev = to_pci_dev(dev);
34 if (pdev->p2pdma->pool)
35 size = gen_pool_size(pdev->p2pdma->pool);
37 return snprintf(buf, PAGE_SIZE, "%zd\n", size);
39 static DEVICE_ATTR_RO(size);
41 static ssize_t available_show(struct device *dev, struct device_attribute *attr,
44 struct pci_dev *pdev = to_pci_dev(dev);
47 if (pdev->p2pdma->pool)
48 avail = gen_pool_avail(pdev->p2pdma->pool);
50 return snprintf(buf, PAGE_SIZE, "%zd\n", avail);
52 static DEVICE_ATTR_RO(available);
54 static ssize_t published_show(struct device *dev, struct device_attribute *attr,
57 struct pci_dev *pdev = to_pci_dev(dev);
59 return snprintf(buf, PAGE_SIZE, "%d\n",
60 pdev->p2pdma->p2pmem_published);
62 static DEVICE_ATTR_RO(published);
64 static struct attribute *p2pmem_attrs[] = {
66 &dev_attr_available.attr,
67 &dev_attr_published.attr,
71 static const struct attribute_group p2pmem_group = {
72 .attrs = p2pmem_attrs,
76 static void pci_p2pdma_release(void *data)
78 struct pci_dev *pdev = data;
79 struct pci_p2pdma *p2pdma = pdev->p2pdma;
84 /* Flush and disable pci_alloc_p2p_mem() */
88 gen_pool_destroy(p2pdma->pool);
89 sysfs_remove_group(&pdev->dev.kobj, &p2pmem_group);
92 static int pci_p2pdma_setup(struct pci_dev *pdev)
95 struct pci_p2pdma *p2p;
97 p2p = devm_kzalloc(&pdev->dev, sizeof(*p2p), GFP_KERNEL);
101 p2p->pool = gen_pool_create(PAGE_SHIFT, dev_to_node(&pdev->dev));
105 error = devm_add_action_or_reset(&pdev->dev, pci_p2pdma_release, pdev);
107 goto out_pool_destroy;
111 error = sysfs_create_group(&pdev->dev.kobj, &p2pmem_group);
113 goto out_pool_destroy;
119 gen_pool_destroy(p2p->pool);
121 devm_kfree(&pdev->dev, p2p);
126 * pci_p2pdma_add_resource - add memory for use as p2p memory
127 * @pdev: the device to add the memory to
128 * @bar: PCI BAR to add
129 * @size: size of the memory to add, may be zero to use the whole BAR
130 * @offset: offset into the PCI BAR
132 * The memory will be given ZONE_DEVICE struct pages so that it may
133 * be used with any DMA request.
135 int pci_p2pdma_add_resource(struct pci_dev *pdev, int bar, size_t size,
138 struct dev_pagemap *pgmap;
142 if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM))
145 if (offset >= pci_resource_len(pdev, bar))
149 size = pci_resource_len(pdev, bar) - offset;
151 if (size + offset > pci_resource_len(pdev, bar))
155 error = pci_p2pdma_setup(pdev);
160 pgmap = devm_kzalloc(&pdev->dev, sizeof(*pgmap), GFP_KERNEL);
163 pgmap->res.start = pci_resource_start(pdev, bar) + offset;
164 pgmap->res.end = pgmap->res.start + size - 1;
165 pgmap->res.flags = pci_resource_flags(pdev, bar);
166 pgmap->type = MEMORY_DEVICE_PCI_P2PDMA;
167 pgmap->pci_p2pdma_bus_offset = pci_bus_address(pdev, bar) -
168 pci_resource_start(pdev, bar);
170 addr = devm_memremap_pages(&pdev->dev, pgmap);
172 error = PTR_ERR(addr);
176 error = gen_pool_add_owner(pdev->p2pdma->pool, (unsigned long)addr,
177 pci_bus_address(pdev, bar) + offset,
178 resource_size(&pgmap->res), dev_to_node(&pdev->dev),
183 pci_info(pdev, "added peer-to-peer DMA memory %pR\n",
189 devm_memunmap_pages(&pdev->dev, pgmap);
191 devm_kfree(&pdev->dev, pgmap);
194 EXPORT_SYMBOL_GPL(pci_p2pdma_add_resource);
197 * Note this function returns the parent PCI device with a
198 * reference taken. It is the caller's responsibility to drop
201 static struct pci_dev *find_parent_pci_dev(struct device *dev)
203 struct device *parent;
205 dev = get_device(dev);
209 return to_pci_dev(dev);
211 parent = get_device(dev->parent);
220 * Check if a PCI bridge has its ACS redirection bits set to redirect P2P
221 * TLPs upstream via ACS. Returns 1 if the packets will be redirected
222 * upstream, 0 otherwise.
224 static int pci_bridge_has_acs_redir(struct pci_dev *pdev)
229 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ACS);
233 pci_read_config_word(pdev, pos + PCI_ACS_CTRL, &ctrl);
235 if (ctrl & (PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_EC))
241 static void seq_buf_print_bus_devfn(struct seq_buf *buf, struct pci_dev *pdev)
246 seq_buf_printf(buf, "%s;", pci_name(pdev));
250 * If we can't find a common upstream bridge take a look at the root
251 * complex and compare it to a whitelist of known good hardware.
253 static bool root_complex_whitelist(struct pci_dev *dev)
255 struct pci_host_bridge *host = pci_find_host_bridge(dev->bus);
256 struct pci_dev *root = pci_get_slot(host->bus, PCI_DEVFN(0, 0));
257 unsigned short vendor, device;
259 if (iommu_present(dev->dev.bus))
265 vendor = root->vendor;
266 device = root->device;
269 /* AMD ZEN host bridges can do peer to peer */
270 if (vendor == PCI_VENDOR_ID_AMD && device == 0x1450)
277 * Find the distance through the nearest common upstream bridge between
280 * If the two devices are the same device then 0 will be returned.
282 * If there are two virtual functions of the same device behind the same
283 * bridge port then 2 will be returned (one step down to the PCIe switch,
284 * then one step back to the same device).
286 * In the case where two devices are connected to the same PCIe switch, the
287 * value 4 will be returned. This corresponds to the following PCI tree:
290 * \+ Switch Upstream Port
291 * +-+ Switch Downstream Port
293 * \-+ Switch Downstream Port
296 * The distance is 4 because we traverse from Device A through the downstream
297 * port of the switch, to the common upstream port, back up to the second
298 * downstream port and then to Device B.
300 * Any two devices that don't have a common upstream bridge will return -1.
301 * In this way devices on separate PCIe root ports will be rejected, which
302 * is what we want for peer-to-peer seeing each PCIe root port defines a
303 * separate hierarchy domain and there's no way to determine whether the root
304 * complex supports forwarding between them.
306 * In the case where two devices are connected to different PCIe switches,
307 * this function will still return a positive distance as long as both
308 * switches eventually have a common upstream bridge. Note this covers
309 * the case of using multiple PCIe switches to achieve a desired level of
310 * fan-out from a root port. The exact distance will be a function of the
311 * number of switches between Device A and Device B.
313 * If a bridge which has any ACS redirection bits set is in the path
314 * then this functions will return -2. This is so we reject any
315 * cases where the TLPs are forwarded up into the root complex.
316 * In this case, a list of all infringing bridge addresses will be
317 * populated in acs_list (assuming it's non-null) for printk purposes.
319 static int upstream_bridge_distance(struct pci_dev *provider,
320 struct pci_dev *client,
321 struct seq_buf *acs_list)
323 struct pci_dev *a = provider, *b = client, *bb;
329 * Note, we don't need to take references to devices returned by
330 * pci_upstream_bridge() seeing we hold a reference to a child
331 * device which will already hold a reference to the upstream bridge.
337 if (pci_bridge_has_acs_redir(a)) {
338 seq_buf_print_bus_devfn(acs_list, a);
346 goto check_b_path_acs;
348 bb = pci_upstream_bridge(bb);
352 a = pci_upstream_bridge(a);
357 * Allow the connection if both devices are on a whitelisted root
358 * complex, but add an arbitrary large value to the distance.
360 if (root_complex_whitelist(provider) &&
361 root_complex_whitelist(client))
362 return 0x1000 + dist_a + dist_b;
373 if (pci_bridge_has_acs_redir(bb)) {
374 seq_buf_print_bus_devfn(acs_list, bb);
378 bb = pci_upstream_bridge(bb);
384 return dist_a + dist_b;
387 static int upstream_bridge_distance_warn(struct pci_dev *provider,
388 struct pci_dev *client)
390 struct seq_buf acs_list;
393 seq_buf_init(&acs_list, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
394 if (!acs_list.buffer)
397 ret = upstream_bridge_distance(provider, client, &acs_list);
399 pci_warn(client, "cannot be used for peer-to-peer DMA as ACS redirect is set between the client and provider (%s)\n",
401 /* Drop final semicolon */
402 acs_list.buffer[acs_list.len-1] = 0;
403 pci_warn(client, "to disable ACS redirect for this path, add the kernel parameter: pci=disable_acs_redir=%s\n",
406 } else if (ret < 0) {
407 pci_warn(client, "cannot be used for peer-to-peer DMA as the client and provider (%s) do not share an upstream bridge\n",
411 kfree(acs_list.buffer);
417 * pci_p2pdma_distance_many - Determine the cumulative distance between
418 * a p2pdma provider and the clients in use.
419 * @provider: p2pdma provider to check against the client list
420 * @clients: array of devices to check (NULL-terminated)
421 * @num_clients: number of clients in the array
422 * @verbose: if true, print warnings for devices when we return -1
424 * Returns -1 if any of the clients are not compatible (behind the same
425 * root port as the provider), otherwise returns a positive number where
426 * a lower number is the preferable choice. (If there's one client
427 * that's the same as the provider it will return 0, which is best choice).
429 * For now, "compatible" means the provider and the clients are all behind
430 * the same PCI root port. This cuts out cases that may work but is safest
431 * for the user. Future work can expand this to white-list root complexes that
432 * can safely forward between each ports.
434 int pci_p2pdma_distance_many(struct pci_dev *provider, struct device **clients,
435 int num_clients, bool verbose)
437 bool not_supported = false;
438 struct pci_dev *pci_client;
442 if (num_clients == 0)
445 for (i = 0; i < num_clients; i++) {
446 if (IS_ENABLED(CONFIG_DMA_VIRT_OPS) &&
447 clients[i]->dma_ops == &dma_virt_ops) {
450 "cannot be used for peer-to-peer DMA because the driver makes use of dma_virt_ops\n");
454 pci_client = find_parent_pci_dev(clients[i]);
458 "cannot be used for peer-to-peer DMA as it is not a PCI device\n");
463 ret = upstream_bridge_distance_warn(provider,
466 ret = upstream_bridge_distance(provider, pci_client,
469 pci_dev_put(pci_client);
472 not_supported = true;
474 if (not_supported && !verbose)
485 EXPORT_SYMBOL_GPL(pci_p2pdma_distance_many);
488 * pci_has_p2pmem - check if a given PCI device has published any p2pmem
489 * @pdev: PCI device to check
491 bool pci_has_p2pmem(struct pci_dev *pdev)
493 return pdev->p2pdma && pdev->p2pdma->p2pmem_published;
495 EXPORT_SYMBOL_GPL(pci_has_p2pmem);
498 * pci_p2pmem_find - find a peer-to-peer DMA memory device compatible with
499 * the specified list of clients and shortest distance (as determined
500 * by pci_p2pmem_dma())
501 * @clients: array of devices to check (NULL-terminated)
502 * @num_clients: number of client devices in the list
504 * If multiple devices are behind the same switch, the one "closest" to the
505 * client devices in use will be chosen first. (So if one of the providers is
506 * the same as one of the clients, that provider will be used ahead of any
507 * other providers that are unrelated). If multiple providers are an equal
508 * distance away, one will be chosen at random.
510 * Returns a pointer to the PCI device with a reference taken (use pci_dev_put
511 * to return the reference) or NULL if no compatible device is found. The
512 * found provider will also be assigned to the client list.
514 struct pci_dev *pci_p2pmem_find_many(struct device **clients, int num_clients)
516 struct pci_dev *pdev = NULL;
518 int closest_distance = INT_MAX;
519 struct pci_dev **closest_pdevs;
521 const int max_devs = PAGE_SIZE / sizeof(*closest_pdevs);
524 closest_pdevs = kmalloc(PAGE_SIZE, GFP_KERNEL);
528 while ((pdev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, pdev))) {
529 if (!pci_has_p2pmem(pdev))
532 distance = pci_p2pdma_distance_many(pdev, clients,
534 if (distance < 0 || distance > closest_distance)
537 if (distance == closest_distance && dev_cnt >= max_devs)
540 if (distance < closest_distance) {
541 for (i = 0; i < dev_cnt; i++)
542 pci_dev_put(closest_pdevs[i]);
545 closest_distance = distance;
548 closest_pdevs[dev_cnt++] = pci_dev_get(pdev);
552 pdev = pci_dev_get(closest_pdevs[prandom_u32_max(dev_cnt)]);
554 for (i = 0; i < dev_cnt; i++)
555 pci_dev_put(closest_pdevs[i]);
557 kfree(closest_pdevs);
560 EXPORT_SYMBOL_GPL(pci_p2pmem_find_many);
563 * pci_alloc_p2p_mem - allocate peer-to-peer DMA memory
564 * @pdev: the device to allocate memory from
565 * @size: number of bytes to allocate
567 * Returns the allocated memory or NULL on error.
569 void *pci_alloc_p2pmem(struct pci_dev *pdev, size_t size)
572 struct percpu_ref *ref;
575 * Pairs with synchronize_rcu() in pci_p2pdma_release() to
576 * ensure pdev->p2pdma is non-NULL for the duration of the
580 if (unlikely(!pdev->p2pdma))
583 ret = (void *)gen_pool_alloc_owner(pdev->p2pdma->pool, size,
588 if (unlikely(!percpu_ref_tryget_live(ref))) {
589 gen_pool_free(pdev->p2pdma->pool, (unsigned long) ret, size);
597 EXPORT_SYMBOL_GPL(pci_alloc_p2pmem);
600 * pci_free_p2pmem - free peer-to-peer DMA memory
601 * @pdev: the device the memory was allocated from
602 * @addr: address of the memory that was allocated
603 * @size: number of bytes that were allocated
605 void pci_free_p2pmem(struct pci_dev *pdev, void *addr, size_t size)
607 struct percpu_ref *ref;
609 gen_pool_free_owner(pdev->p2pdma->pool, (uintptr_t)addr, size,
613 EXPORT_SYMBOL_GPL(pci_free_p2pmem);
616 * pci_virt_to_bus - return the PCI bus address for a given virtual
617 * address obtained with pci_alloc_p2pmem()
618 * @pdev: the device the memory was allocated from
619 * @addr: address of the memory that was allocated
621 pci_bus_addr_t pci_p2pmem_virt_to_bus(struct pci_dev *pdev, void *addr)
629 * Note: when we added the memory to the pool we used the PCI
630 * bus address as the physical address. So gen_pool_virt_to_phys()
631 * actually returns the bus address despite the misleading name.
633 return gen_pool_virt_to_phys(pdev->p2pdma->pool, (unsigned long)addr);
635 EXPORT_SYMBOL_GPL(pci_p2pmem_virt_to_bus);
638 * pci_p2pmem_alloc_sgl - allocate peer-to-peer DMA memory in a scatterlist
639 * @pdev: the device to allocate memory from
640 * @nents: the number of SG entries in the list
641 * @length: number of bytes to allocate
643 * Return: %NULL on error or &struct scatterlist pointer and @nents on success
645 struct scatterlist *pci_p2pmem_alloc_sgl(struct pci_dev *pdev,
646 unsigned int *nents, u32 length)
648 struct scatterlist *sg;
651 sg = kzalloc(sizeof(*sg), GFP_KERNEL);
655 sg_init_table(sg, 1);
657 addr = pci_alloc_p2pmem(pdev, length);
661 sg_set_buf(sg, addr, length);
669 EXPORT_SYMBOL_GPL(pci_p2pmem_alloc_sgl);
672 * pci_p2pmem_free_sgl - free a scatterlist allocated by pci_p2pmem_alloc_sgl()
673 * @pdev: the device to allocate memory from
674 * @sgl: the allocated scatterlist
676 void pci_p2pmem_free_sgl(struct pci_dev *pdev, struct scatterlist *sgl)
678 struct scatterlist *sg;
681 for_each_sg(sgl, sg, INT_MAX, count) {
685 pci_free_p2pmem(pdev, sg_virt(sg), sg->length);
689 EXPORT_SYMBOL_GPL(pci_p2pmem_free_sgl);
692 * pci_p2pmem_publish - publish the peer-to-peer DMA memory for use by
693 * other devices with pci_p2pmem_find()
694 * @pdev: the device with peer-to-peer DMA memory to publish
695 * @publish: set to true to publish the memory, false to unpublish it
697 * Published memory can be used by other PCI device drivers for
698 * peer-2-peer DMA operations. Non-published memory is reserved for
699 * exclusive use of the device driver that registers the peer-to-peer
702 void pci_p2pmem_publish(struct pci_dev *pdev, bool publish)
705 pdev->p2pdma->p2pmem_published = publish;
707 EXPORT_SYMBOL_GPL(pci_p2pmem_publish);
710 * pci_p2pdma_map_sg - map a PCI peer-to-peer scatterlist for DMA
711 * @dev: device doing the DMA request
712 * @sg: scatter list to map
713 * @nents: elements in the scatterlist
714 * @dir: DMA direction
716 * Scatterlists mapped with this function should not be unmapped in any way.
718 * Returns the number of SG entries mapped or 0 on error.
720 int pci_p2pdma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
721 enum dma_data_direction dir)
723 struct dev_pagemap *pgmap;
724 struct scatterlist *s;
729 * p2pdma mappings are not compatible with devices that use
730 * dma_virt_ops. If the upper layers do the right thing
731 * this should never happen because it will be prevented
732 * by the check in pci_p2pdma_distance_many()
734 if (WARN_ON_ONCE(IS_ENABLED(CONFIG_DMA_VIRT_OPS) &&
735 dev->dma_ops == &dma_virt_ops))
738 for_each_sg(sg, s, nents, i) {
739 pgmap = sg_page(s)->pgmap;
742 s->dma_address = paddr - pgmap->pci_p2pdma_bus_offset;
743 sg_dma_len(s) = s->length;
748 EXPORT_SYMBOL_GPL(pci_p2pdma_map_sg);
751 * pci_p2pdma_enable_store - parse a configfs/sysfs attribute store
753 * @page: contents of the value to be stored
754 * @p2p_dev: returns the PCI device that was selected to be used
755 * (if one was specified in the stored value)
756 * @use_p2pdma: returns whether to enable p2pdma or not
758 * Parses an attribute value to decide whether to enable p2pdma.
759 * The value can select a PCI device (using its full BDF device
760 * name) or a boolean (in any format strtobool() accepts). A false
761 * value disables p2pdma, a true value expects the caller
762 * to automatically find a compatible device and specifying a PCI device
763 * expects the caller to use the specific provider.
765 * pci_p2pdma_enable_show() should be used as the show operation for
768 * Returns 0 on success
770 int pci_p2pdma_enable_store(const char *page, struct pci_dev **p2p_dev,
775 dev = bus_find_device_by_name(&pci_bus_type, NULL, page);
778 *p2p_dev = to_pci_dev(dev);
780 if (!pci_has_p2pmem(*p2p_dev)) {
782 "PCI device has no peer-to-peer memory: %s\n",
784 pci_dev_put(*p2p_dev);
789 } else if ((page[0] == '0' || page[0] == '1') && !iscntrl(page[1])) {
791 * If the user enters a PCI device that doesn't exist
792 * like "0000:01:00.1", we don't want strtobool to think
793 * it's a '0' when it's clearly not what the user wanted.
794 * So we require 0's and 1's to be exactly one character.
796 } else if (!strtobool(page, use_p2pdma)) {
800 pr_err("No such PCI device: %.*s\n", (int)strcspn(page, "\n"), page);
803 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_store);
806 * pci_p2pdma_enable_show - show a configfs/sysfs attribute indicating
807 * whether p2pdma is enabled
808 * @page: contents of the stored value
809 * @p2p_dev: the selected p2p device (NULL if no device is selected)
810 * @use_p2pdma: whether p2pdma has been enabled
812 * Attributes that use pci_p2pdma_enable_store() should use this function
813 * to show the value of the attribute.
815 * Returns 0 on success
817 ssize_t pci_p2pdma_enable_show(char *page, struct pci_dev *p2p_dev,
821 return sprintf(page, "0\n");
824 return sprintf(page, "1\n");
826 return sprintf(page, "%s\n", pci_name(p2p_dev));
828 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show);