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>
23 struct percpu_ref devmap_ref
;
24 struct completion devmap_ref_done
;
25 struct gen_pool
*pool
;
26 bool p2pmem_published
;
29 static ssize_t
size_show(struct device
*dev
, struct device_attribute
*attr
,
32 struct pci_dev
*pdev
= to_pci_dev(dev
);
35 if (pdev
->p2pdma
->pool
)
36 size
= gen_pool_size(pdev
->p2pdma
->pool
);
38 return snprintf(buf
, PAGE_SIZE
, "%zd\n", size
);
40 static DEVICE_ATTR_RO(size
);
42 static ssize_t
available_show(struct device
*dev
, struct device_attribute
*attr
,
45 struct pci_dev
*pdev
= to_pci_dev(dev
);
48 if (pdev
->p2pdma
->pool
)
49 avail
= gen_pool_avail(pdev
->p2pdma
->pool
);
51 return snprintf(buf
, PAGE_SIZE
, "%zd\n", avail
);
53 static DEVICE_ATTR_RO(available
);
55 static ssize_t
published_show(struct device
*dev
, struct device_attribute
*attr
,
58 struct pci_dev
*pdev
= to_pci_dev(dev
);
60 return snprintf(buf
, PAGE_SIZE
, "%d\n",
61 pdev
->p2pdma
->p2pmem_published
);
63 static DEVICE_ATTR_RO(published
);
65 static struct attribute
*p2pmem_attrs
[] = {
67 &dev_attr_available
.attr
,
68 &dev_attr_published
.attr
,
72 static const struct attribute_group p2pmem_group
= {
73 .attrs
= p2pmem_attrs
,
77 static void pci_p2pdma_percpu_release(struct percpu_ref
*ref
)
79 struct pci_p2pdma
*p2p
=
80 container_of(ref
, struct pci_p2pdma
, devmap_ref
);
82 complete_all(&p2p
->devmap_ref_done
);
85 static void pci_p2pdma_percpu_kill(struct percpu_ref
*ref
)
88 * pci_p2pdma_add_resource() may be called multiple times
89 * by a driver and may register the percpu_kill devm action multiple
90 * times. We only want the first action to actually kill the
93 if (percpu_ref_is_dying(ref
))
99 static void pci_p2pdma_release(void *data
)
101 struct pci_dev
*pdev
= data
;
106 wait_for_completion(&pdev
->p2pdma
->devmap_ref_done
);
107 percpu_ref_exit(&pdev
->p2pdma
->devmap_ref
);
109 gen_pool_destroy(pdev
->p2pdma
->pool
);
110 sysfs_remove_group(&pdev
->dev
.kobj
, &p2pmem_group
);
114 static int pci_p2pdma_setup(struct pci_dev
*pdev
)
117 struct pci_p2pdma
*p2p
;
119 p2p
= devm_kzalloc(&pdev
->dev
, sizeof(*p2p
), GFP_KERNEL
);
123 p2p
->pool
= gen_pool_create(PAGE_SHIFT
, dev_to_node(&pdev
->dev
));
127 init_completion(&p2p
->devmap_ref_done
);
128 error
= percpu_ref_init(&p2p
->devmap_ref
,
129 pci_p2pdma_percpu_release
, 0, GFP_KERNEL
);
131 goto out_pool_destroy
;
133 error
= devm_add_action_or_reset(&pdev
->dev
, pci_p2pdma_release
, pdev
);
135 goto out_pool_destroy
;
139 error
= sysfs_create_group(&pdev
->dev
.kobj
, &p2pmem_group
);
141 goto out_pool_destroy
;
147 gen_pool_destroy(p2p
->pool
);
149 devm_kfree(&pdev
->dev
, p2p
);
154 * pci_p2pdma_add_resource - add memory for use as p2p memory
155 * @pdev: the device to add the memory to
156 * @bar: PCI BAR to add
157 * @size: size of the memory to add, may be zero to use the whole BAR
158 * @offset: offset into the PCI BAR
160 * The memory will be given ZONE_DEVICE struct pages so that it may
161 * be used with any DMA request.
163 int pci_p2pdma_add_resource(struct pci_dev
*pdev
, int bar
, size_t size
,
166 struct dev_pagemap
*pgmap
;
170 if (!(pci_resource_flags(pdev
, bar
) & IORESOURCE_MEM
))
173 if (offset
>= pci_resource_len(pdev
, bar
))
177 size
= pci_resource_len(pdev
, bar
) - offset
;
179 if (size
+ offset
> pci_resource_len(pdev
, bar
))
183 error
= pci_p2pdma_setup(pdev
);
188 pgmap
= devm_kzalloc(&pdev
->dev
, sizeof(*pgmap
), GFP_KERNEL
);
192 pgmap
->res
.start
= pci_resource_start(pdev
, bar
) + offset
;
193 pgmap
->res
.end
= pgmap
->res
.start
+ size
- 1;
194 pgmap
->res
.flags
= pci_resource_flags(pdev
, bar
);
195 pgmap
->ref
= &pdev
->p2pdma
->devmap_ref
;
196 pgmap
->type
= MEMORY_DEVICE_PCI_P2PDMA
;
197 pgmap
->pci_p2pdma_bus_offset
= pci_bus_address(pdev
, bar
) -
198 pci_resource_start(pdev
, bar
);
199 pgmap
->kill
= pci_p2pdma_percpu_kill
;
201 addr
= devm_memremap_pages(&pdev
->dev
, pgmap
);
203 error
= PTR_ERR(addr
);
207 error
= gen_pool_add_virt(pdev
->p2pdma
->pool
, (unsigned long)addr
,
208 pci_bus_address(pdev
, bar
) + offset
,
209 resource_size(&pgmap
->res
), dev_to_node(&pdev
->dev
));
213 pci_info(pdev
, "added peer-to-peer DMA memory %pR\n",
219 devm_kfree(&pdev
->dev
, pgmap
);
222 EXPORT_SYMBOL_GPL(pci_p2pdma_add_resource
);
225 * Note this function returns the parent PCI device with a
226 * reference taken. It is the caller's responsibily to drop
229 static struct pci_dev
*find_parent_pci_dev(struct device
*dev
)
231 struct device
*parent
;
233 dev
= get_device(dev
);
237 return to_pci_dev(dev
);
239 parent
= get_device(dev
->parent
);
248 * Check if a PCI bridge has its ACS redirection bits set to redirect P2P
249 * TLPs upstream via ACS. Returns 1 if the packets will be redirected
250 * upstream, 0 otherwise.
252 static int pci_bridge_has_acs_redir(struct pci_dev
*pdev
)
257 pos
= pci_find_ext_capability(pdev
, PCI_EXT_CAP_ID_ACS
);
261 pci_read_config_word(pdev
, pos
+ PCI_ACS_CTRL
, &ctrl
);
263 if (ctrl
& (PCI_ACS_RR
| PCI_ACS_CR
| PCI_ACS_EC
))
269 static void seq_buf_print_bus_devfn(struct seq_buf
*buf
, struct pci_dev
*pdev
)
274 seq_buf_printf(buf
, "%s;", pci_name(pdev
));
278 * Find the distance through the nearest common upstream bridge between
281 * If the two devices are the same device then 0 will be returned.
283 * If there are two virtual functions of the same device behind the same
284 * bridge port then 2 will be returned (one step down to the PCIe switch,
285 * then one step back to the same device).
287 * In the case where two devices are connected to the same PCIe switch, the
288 * value 4 will be returned. This corresponds to the following PCI tree:
291 * \+ Switch Upstream Port
292 * +-+ Switch Downstream Port
294 * \-+ Switch Downstream Port
297 * The distance is 4 because we traverse from Device A through the downstream
298 * port of the switch, to the common upstream port, back up to the second
299 * downstream port and then to Device B.
301 * Any two devices that don't have a common upstream bridge will return -1.
302 * In this way devices on separate PCIe root ports will be rejected, which
303 * is what we want for peer-to-peer seeing each PCIe root port defines a
304 * separate hierarchy domain and there's no way to determine whether the root
305 * complex supports forwarding between them.
307 * In the case where two devices are connected to different PCIe switches,
308 * this function will still return a positive distance as long as both
309 * switches eventually have a common upstream bridge. Note this covers
310 * the case of using multiple PCIe switches to achieve a desired level of
311 * fan-out from a root port. The exact distance will be a function of the
312 * number of switches between Device A and Device B.
314 * If a bridge which has any ACS redirection bits set is in the path
315 * then this functions will return -2. This is so we reject any
316 * cases where the TLPs are forwarded up into the root complex.
317 * In this case, a list of all infringing bridge addresses will be
318 * populated in acs_list (assuming it's non-null) for printk purposes.
320 static int upstream_bridge_distance(struct pci_dev
*a
,
322 struct seq_buf
*acs_list
)
326 struct pci_dev
*bb
= NULL
;
330 * Note, we don't need to take references to devices returned by
331 * pci_upstream_bridge() seeing we hold a reference to a child
332 * device which will already hold a reference to the upstream bridge.
338 if (pci_bridge_has_acs_redir(a
)) {
339 seq_buf_print_bus_devfn(acs_list
, a
);
347 goto check_b_path_acs
;
349 bb
= pci_upstream_bridge(bb
);
353 a
= pci_upstream_bridge(a
);
366 if (pci_bridge_has_acs_redir(bb
)) {
367 seq_buf_print_bus_devfn(acs_list
, bb
);
371 bb
= pci_upstream_bridge(bb
);
377 return dist_a
+ dist_b
;
380 static int upstream_bridge_distance_warn(struct pci_dev
*provider
,
381 struct pci_dev
*client
)
383 struct seq_buf acs_list
;
386 seq_buf_init(&acs_list
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
387 if (!acs_list
.buffer
)
390 ret
= upstream_bridge_distance(provider
, client
, &acs_list
);
392 pci_warn(client
, "cannot be used for peer-to-peer DMA as ACS redirect is set between the client and provider (%s)\n",
394 /* Drop final semicolon */
395 acs_list
.buffer
[acs_list
.len
-1] = 0;
396 pci_warn(client
, "to disable ACS redirect for this path, add the kernel parameter: pci=disable_acs_redir=%s\n",
399 } else if (ret
< 0) {
400 pci_warn(client
, "cannot be used for peer-to-peer DMA as the client and provider (%s) do not share an upstream bridge\n",
404 kfree(acs_list
.buffer
);
410 * pci_p2pdma_distance_many - Determive the cumulative distance between
411 * a p2pdma provider and the clients in use.
412 * @provider: p2pdma provider to check against the client list
413 * @clients: array of devices to check (NULL-terminated)
414 * @num_clients: number of clients in the array
415 * @verbose: if true, print warnings for devices when we return -1
417 * Returns -1 if any of the clients are not compatible (behind the same
418 * root port as the provider), otherwise returns a positive number where
419 * a lower number is the preferable choice. (If there's one client
420 * that's the same as the provider it will return 0, which is best choice).
422 * For now, "compatible" means the provider and the clients are all behind
423 * the same PCI root port. This cuts out cases that may work but is safest
424 * for the user. Future work can expand this to white-list root complexes that
425 * can safely forward between each ports.
427 int pci_p2pdma_distance_many(struct pci_dev
*provider
, struct device
**clients
,
428 int num_clients
, bool verbose
)
430 bool not_supported
= false;
431 struct pci_dev
*pci_client
;
435 if (num_clients
== 0)
438 for (i
= 0; i
< num_clients
; i
++) {
439 pci_client
= find_parent_pci_dev(clients
[i
]);
443 "cannot be used for peer-to-peer DMA as it is not a PCI device\n");
448 ret
= upstream_bridge_distance_warn(provider
,
451 ret
= upstream_bridge_distance(provider
, pci_client
,
454 pci_dev_put(pci_client
);
457 not_supported
= true;
459 if (not_supported
&& !verbose
)
470 EXPORT_SYMBOL_GPL(pci_p2pdma_distance_many
);
473 * pci_has_p2pmem - check if a given PCI device has published any p2pmem
474 * @pdev: PCI device to check
476 bool pci_has_p2pmem(struct pci_dev
*pdev
)
478 return pdev
->p2pdma
&& pdev
->p2pdma
->p2pmem_published
;
480 EXPORT_SYMBOL_GPL(pci_has_p2pmem
);
483 * pci_p2pmem_find - find a peer-to-peer DMA memory device compatible with
484 * the specified list of clients and shortest distance (as determined
485 * by pci_p2pmem_dma())
486 * @clients: array of devices to check (NULL-terminated)
487 * @num_clients: number of client devices in the list
489 * If multiple devices are behind the same switch, the one "closest" to the
490 * client devices in use will be chosen first. (So if one of the providers is
491 * the same as one of the clients, that provider will be used ahead of any
492 * other providers that are unrelated). If multiple providers are an equal
493 * distance away, one will be chosen at random.
495 * Returns a pointer to the PCI device with a reference taken (use pci_dev_put
496 * to return the reference) or NULL if no compatible device is found. The
497 * found provider will also be assigned to the client list.
499 struct pci_dev
*pci_p2pmem_find_many(struct device
**clients
, int num_clients
)
501 struct pci_dev
*pdev
= NULL
;
503 int closest_distance
= INT_MAX
;
504 struct pci_dev
**closest_pdevs
;
506 const int max_devs
= PAGE_SIZE
/ sizeof(*closest_pdevs
);
509 closest_pdevs
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
513 while ((pdev
= pci_get_device(PCI_ANY_ID
, PCI_ANY_ID
, pdev
))) {
514 if (!pci_has_p2pmem(pdev
))
517 distance
= pci_p2pdma_distance_many(pdev
, clients
,
519 if (distance
< 0 || distance
> closest_distance
)
522 if (distance
== closest_distance
&& dev_cnt
>= max_devs
)
525 if (distance
< closest_distance
) {
526 for (i
= 0; i
< dev_cnt
; i
++)
527 pci_dev_put(closest_pdevs
[i
]);
530 closest_distance
= distance
;
533 closest_pdevs
[dev_cnt
++] = pci_dev_get(pdev
);
537 pdev
= pci_dev_get(closest_pdevs
[prandom_u32_max(dev_cnt
)]);
539 for (i
= 0; i
< dev_cnt
; i
++)
540 pci_dev_put(closest_pdevs
[i
]);
542 kfree(closest_pdevs
);
545 EXPORT_SYMBOL_GPL(pci_p2pmem_find_many
);
548 * pci_alloc_p2p_mem - allocate peer-to-peer DMA memory
549 * @pdev: the device to allocate memory from
550 * @size: number of bytes to allocate
552 * Returns the allocated memory or NULL on error.
554 void *pci_alloc_p2pmem(struct pci_dev
*pdev
, size_t size
)
558 if (unlikely(!pdev
->p2pdma
))
561 if (unlikely(!percpu_ref_tryget_live(&pdev
->p2pdma
->devmap_ref
)))
564 ret
= (void *)gen_pool_alloc(pdev
->p2pdma
->pool
, size
);
567 percpu_ref_put(&pdev
->p2pdma
->devmap_ref
);
571 EXPORT_SYMBOL_GPL(pci_alloc_p2pmem
);
574 * pci_free_p2pmem - free peer-to-peer DMA memory
575 * @pdev: the device the memory was allocated from
576 * @addr: address of the memory that was allocated
577 * @size: number of bytes that were allocated
579 void pci_free_p2pmem(struct pci_dev
*pdev
, void *addr
, size_t size
)
581 gen_pool_free(pdev
->p2pdma
->pool
, (uintptr_t)addr
, size
);
582 percpu_ref_put(&pdev
->p2pdma
->devmap_ref
);
584 EXPORT_SYMBOL_GPL(pci_free_p2pmem
);
587 * pci_virt_to_bus - return the PCI bus address for a given virtual
588 * address obtained with pci_alloc_p2pmem()
589 * @pdev: the device the memory was allocated from
590 * @addr: address of the memory that was allocated
592 pci_bus_addr_t
pci_p2pmem_virt_to_bus(struct pci_dev
*pdev
, void *addr
)
600 * Note: when we added the memory to the pool we used the PCI
601 * bus address as the physical address. So gen_pool_virt_to_phys()
602 * actually returns the bus address despite the misleading name.
604 return gen_pool_virt_to_phys(pdev
->p2pdma
->pool
, (unsigned long)addr
);
606 EXPORT_SYMBOL_GPL(pci_p2pmem_virt_to_bus
);
609 * pci_p2pmem_alloc_sgl - allocate peer-to-peer DMA memory in a scatterlist
610 * @pdev: the device to allocate memory from
611 * @nents: the number of SG entries in the list
612 * @length: number of bytes to allocate
614 * Return: %NULL on error or &struct scatterlist pointer and @nents on success
616 struct scatterlist
*pci_p2pmem_alloc_sgl(struct pci_dev
*pdev
,
617 unsigned int *nents
, u32 length
)
619 struct scatterlist
*sg
;
622 sg
= kzalloc(sizeof(*sg
), GFP_KERNEL
);
626 sg_init_table(sg
, 1);
628 addr
= pci_alloc_p2pmem(pdev
, length
);
632 sg_set_buf(sg
, addr
, length
);
640 EXPORT_SYMBOL_GPL(pci_p2pmem_alloc_sgl
);
643 * pci_p2pmem_free_sgl - free a scatterlist allocated by pci_p2pmem_alloc_sgl()
644 * @pdev: the device to allocate memory from
645 * @sgl: the allocated scatterlist
647 void pci_p2pmem_free_sgl(struct pci_dev
*pdev
, struct scatterlist
*sgl
)
649 struct scatterlist
*sg
;
652 for_each_sg(sgl
, sg
, INT_MAX
, count
) {
656 pci_free_p2pmem(pdev
, sg_virt(sg
), sg
->length
);
660 EXPORT_SYMBOL_GPL(pci_p2pmem_free_sgl
);
663 * pci_p2pmem_publish - publish the peer-to-peer DMA memory for use by
664 * other devices with pci_p2pmem_find()
665 * @pdev: the device with peer-to-peer DMA memory to publish
666 * @publish: set to true to publish the memory, false to unpublish it
668 * Published memory can be used by other PCI device drivers for
669 * peer-2-peer DMA operations. Non-published memory is reserved for
670 * exclusive use of the device driver that registers the peer-to-peer
673 void pci_p2pmem_publish(struct pci_dev
*pdev
, bool publish
)
676 pdev
->p2pdma
->p2pmem_published
= publish
;
678 EXPORT_SYMBOL_GPL(pci_p2pmem_publish
);
681 * pci_p2pdma_map_sg - map a PCI peer-to-peer scatterlist for DMA
682 * @dev: device doing the DMA request
683 * @sg: scatter list to map
684 * @nents: elements in the scatterlist
685 * @dir: DMA direction
687 * Scatterlists mapped with this function should not be unmapped in any way.
689 * Returns the number of SG entries mapped or 0 on error.
691 int pci_p2pdma_map_sg(struct device
*dev
, struct scatterlist
*sg
, int nents
,
692 enum dma_data_direction dir
)
694 struct dev_pagemap
*pgmap
;
695 struct scatterlist
*s
;
700 * p2pdma mappings are not compatible with devices that use
701 * dma_virt_ops. If the upper layers do the right thing
702 * this should never happen because it will be prevented
703 * by the check in pci_p2pdma_add_client()
705 if (WARN_ON_ONCE(IS_ENABLED(CONFIG_DMA_VIRT_OPS
) &&
706 dev
->dma_ops
== &dma_virt_ops
))
709 for_each_sg(sg
, s
, nents
, i
) {
710 pgmap
= sg_page(s
)->pgmap
;
713 s
->dma_address
= paddr
- pgmap
->pci_p2pdma_bus_offset
;
714 sg_dma_len(s
) = s
->length
;
719 EXPORT_SYMBOL_GPL(pci_p2pdma_map_sg
);
722 * pci_p2pdma_enable_store - parse a configfs/sysfs attribute store
724 * @page: contents of the value to be stored
725 * @p2p_dev: returns the PCI device that was selected to be used
726 * (if one was specified in the stored value)
727 * @use_p2pdma: returns whether to enable p2pdma or not
729 * Parses an attribute value to decide whether to enable p2pdma.
730 * The value can select a PCI device (using its full BDF device
731 * name) or a boolean (in any format strtobool() accepts). A false
732 * value disables p2pdma, a true value expects the caller
733 * to automatically find a compatible device and specifying a PCI device
734 * expects the caller to use the specific provider.
736 * pci_p2pdma_enable_show() should be used as the show operation for
739 * Returns 0 on success
741 int pci_p2pdma_enable_store(const char *page
, struct pci_dev
**p2p_dev
,
746 dev
= bus_find_device_by_name(&pci_bus_type
, NULL
, page
);
749 *p2p_dev
= to_pci_dev(dev
);
751 if (!pci_has_p2pmem(*p2p_dev
)) {
753 "PCI device has no peer-to-peer memory: %s\n",
755 pci_dev_put(*p2p_dev
);
760 } else if ((page
[0] == '0' || page
[0] == '1') && !iscntrl(page
[1])) {
762 * If the user enters a PCI device that doesn't exist
763 * like "0000:01:00.1", we don't want strtobool to think
764 * it's a '0' when it's clearly not what the user wanted.
765 * So we require 0's and 1's to be exactly one character.
767 } else if (!strtobool(page
, use_p2pdma
)) {
771 pr_err("No such PCI device: %.*s\n", (int)strcspn(page
, "\n"), page
);
774 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_store
);
777 * pci_p2pdma_enable_show - show a configfs/sysfs attribute indicating
778 * whether p2pdma is enabled
779 * @page: contents of the stored value
780 * @p2p_dev: the selected p2p device (NULL if no device is selected)
781 * @use_p2pdma: whether p2pdma has been enabled
783 * Attributes that use pci_p2pdma_enable_store() should use this function
784 * to show the value of the attribute.
786 * Returns 0 on success
788 ssize_t
pci_p2pdma_enable_show(char *page
, struct pci_dev
*p2p_dev
,
792 return sprintf(page
, "0\n");
795 return sprintf(page
, "1\n");
797 return sprintf(page
, "%s\n", pci_name(p2p_dev
));
799 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show
);