Loading...
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/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/xarray.h>
22
23enum pci_p2pdma_map_type {
24 PCI_P2PDMA_MAP_UNKNOWN = 0,
25 PCI_P2PDMA_MAP_NOT_SUPPORTED,
26 PCI_P2PDMA_MAP_BUS_ADDR,
27 PCI_P2PDMA_MAP_THRU_HOST_BRIDGE,
28};
29
30struct pci_p2pdma {
31 struct gen_pool *pool;
32 bool p2pmem_published;
33 struct xarray map_types;
34};
35
36struct pci_p2pdma_pagemap {
37 struct dev_pagemap pgmap;
38 struct pci_dev *provider;
39 u64 bus_offset;
40};
41
42static struct pci_p2pdma_pagemap *to_p2p_pgmap(struct dev_pagemap *pgmap)
43{
44 return container_of(pgmap, struct pci_p2pdma_pagemap, pgmap);
45}
46
47static ssize_t size_show(struct device *dev, struct device_attribute *attr,
48 char *buf)
49{
50 struct pci_dev *pdev = to_pci_dev(dev);
51 size_t size = 0;
52
53 if (pdev->p2pdma->pool)
54 size = gen_pool_size(pdev->p2pdma->pool);
55
56 return snprintf(buf, PAGE_SIZE, "%zd\n", size);
57}
58static DEVICE_ATTR_RO(size);
59
60static ssize_t available_show(struct device *dev, struct device_attribute *attr,
61 char *buf)
62{
63 struct pci_dev *pdev = to_pci_dev(dev);
64 size_t avail = 0;
65
66 if (pdev->p2pdma->pool)
67 avail = gen_pool_avail(pdev->p2pdma->pool);
68
69 return snprintf(buf, PAGE_SIZE, "%zd\n", avail);
70}
71static DEVICE_ATTR_RO(available);
72
73static ssize_t published_show(struct device *dev, struct device_attribute *attr,
74 char *buf)
75{
76 struct pci_dev *pdev = to_pci_dev(dev);
77
78 return snprintf(buf, PAGE_SIZE, "%d\n",
79 pdev->p2pdma->p2pmem_published);
80}
81static DEVICE_ATTR_RO(published);
82
83static struct attribute *p2pmem_attrs[] = {
84 &dev_attr_size.attr,
85 &dev_attr_available.attr,
86 &dev_attr_published.attr,
87 NULL,
88};
89
90static const struct attribute_group p2pmem_group = {
91 .attrs = p2pmem_attrs,
92 .name = "p2pmem",
93};
94
95static void pci_p2pdma_release(void *data)
96{
97 struct pci_dev *pdev = data;
98 struct pci_p2pdma *p2pdma = pdev->p2pdma;
99
100 if (!p2pdma)
101 return;
102
103 /* Flush and disable pci_alloc_p2p_mem() */
104 pdev->p2pdma = NULL;
105 synchronize_rcu();
106
107 gen_pool_destroy(p2pdma->pool);
108 sysfs_remove_group(&pdev->dev.kobj, &p2pmem_group);
109 xa_destroy(&p2pdma->map_types);
110}
111
112static int pci_p2pdma_setup(struct pci_dev *pdev)
113{
114 int error = -ENOMEM;
115 struct pci_p2pdma *p2p;
116
117 p2p = devm_kzalloc(&pdev->dev, sizeof(*p2p), GFP_KERNEL);
118 if (!p2p)
119 return -ENOMEM;
120
121 xa_init(&p2p->map_types);
122
123 p2p->pool = gen_pool_create(PAGE_SHIFT, dev_to_node(&pdev->dev));
124 if (!p2p->pool)
125 goto out;
126
127 error = devm_add_action_or_reset(&pdev->dev, pci_p2pdma_release, pdev);
128 if (error)
129 goto out_pool_destroy;
130
131 pdev->p2pdma = p2p;
132
133 error = sysfs_create_group(&pdev->dev.kobj, &p2pmem_group);
134 if (error)
135 goto out_pool_destroy;
136
137 return 0;
138
139out_pool_destroy:
140 pdev->p2pdma = NULL;
141 gen_pool_destroy(p2p->pool);
142out:
143 devm_kfree(&pdev->dev, p2p);
144 return error;
145}
146
147/**
148 * pci_p2pdma_add_resource - add memory for use as p2p memory
149 * @pdev: the device to add the memory to
150 * @bar: PCI BAR to add
151 * @size: size of the memory to add, may be zero to use the whole BAR
152 * @offset: offset into the PCI BAR
153 *
154 * The memory will be given ZONE_DEVICE struct pages so that it may
155 * be used with any DMA request.
156 */
157int pci_p2pdma_add_resource(struct pci_dev *pdev, int bar, size_t size,
158 u64 offset)
159{
160 struct pci_p2pdma_pagemap *p2p_pgmap;
161 struct dev_pagemap *pgmap;
162 void *addr;
163 int error;
164
165 if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM))
166 return -EINVAL;
167
168 if (offset >= pci_resource_len(pdev, bar))
169 return -EINVAL;
170
171 if (!size)
172 size = pci_resource_len(pdev, bar) - offset;
173
174 if (size + offset > pci_resource_len(pdev, bar))
175 return -EINVAL;
176
177 if (!pdev->p2pdma) {
178 error = pci_p2pdma_setup(pdev);
179 if (error)
180 return error;
181 }
182
183 p2p_pgmap = devm_kzalloc(&pdev->dev, sizeof(*p2p_pgmap), GFP_KERNEL);
184 if (!p2p_pgmap)
185 return -ENOMEM;
186
187 pgmap = &p2p_pgmap->pgmap;
188 pgmap->res.start = pci_resource_start(pdev, bar) + offset;
189 pgmap->res.end = pgmap->res.start + size - 1;
190 pgmap->res.flags = pci_resource_flags(pdev, bar);
191 pgmap->type = MEMORY_DEVICE_PCI_P2PDMA;
192
193 p2p_pgmap->provider = pdev;
194 p2p_pgmap->bus_offset = pci_bus_address(pdev, bar) -
195 pci_resource_start(pdev, bar);
196
197 addr = devm_memremap_pages(&pdev->dev, pgmap);
198 if (IS_ERR(addr)) {
199 error = PTR_ERR(addr);
200 goto pgmap_free;
201 }
202
203 error = gen_pool_add_owner(pdev->p2pdma->pool, (unsigned long)addr,
204 pci_bus_address(pdev, bar) + offset,
205 resource_size(&pgmap->res), dev_to_node(&pdev->dev),
206 pgmap->ref);
207 if (error)
208 goto pages_free;
209
210 pci_info(pdev, "added peer-to-peer DMA memory %pR\n",
211 &pgmap->res);
212
213 return 0;
214
215pages_free:
216 devm_memunmap_pages(&pdev->dev, pgmap);
217pgmap_free:
218 devm_kfree(&pdev->dev, pgmap);
219 return error;
220}
221EXPORT_SYMBOL_GPL(pci_p2pdma_add_resource);
222
223/*
224 * Note this function returns the parent PCI device with a
225 * reference taken. It is the caller's responsibility to drop
226 * the reference.
227 */
228static struct pci_dev *find_parent_pci_dev(struct device *dev)
229{
230 struct device *parent;
231
232 dev = get_device(dev);
233
234 while (dev) {
235 if (dev_is_pci(dev))
236 return to_pci_dev(dev);
237
238 parent = get_device(dev->parent);
239 put_device(dev);
240 dev = parent;
241 }
242
243 return NULL;
244}
245
246/*
247 * Check if a PCI bridge has its ACS redirection bits set to redirect P2P
248 * TLPs upstream via ACS. Returns 1 if the packets will be redirected
249 * upstream, 0 otherwise.
250 */
251static int pci_bridge_has_acs_redir(struct pci_dev *pdev)
252{
253 int pos;
254 u16 ctrl;
255
256 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ACS);
257 if (!pos)
258 return 0;
259
260 pci_read_config_word(pdev, pos + PCI_ACS_CTRL, &ctrl);
261
262 if (ctrl & (PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_EC))
263 return 1;
264
265 return 0;
266}
267
268static void seq_buf_print_bus_devfn(struct seq_buf *buf, struct pci_dev *pdev)
269{
270 if (!buf)
271 return;
272
273 seq_buf_printf(buf, "%s;", pci_name(pdev));
274}
275
276static const struct pci_p2pdma_whitelist_entry {
277 unsigned short vendor;
278 unsigned short device;
279 enum {
280 REQ_SAME_HOST_BRIDGE = 1 << 0,
281 } flags;
282} pci_p2pdma_whitelist[] = {
283 /* AMD ZEN */
284 {PCI_VENDOR_ID_AMD, 0x1450, 0},
285
286 /* Intel Xeon E5/Core i7 */
287 {PCI_VENDOR_ID_INTEL, 0x3c00, REQ_SAME_HOST_BRIDGE},
288 {PCI_VENDOR_ID_INTEL, 0x3c01, REQ_SAME_HOST_BRIDGE},
289 /* Intel Xeon E7 v3/Xeon E5 v3/Core i7 */
290 {PCI_VENDOR_ID_INTEL, 0x2f00, REQ_SAME_HOST_BRIDGE},
291 {PCI_VENDOR_ID_INTEL, 0x2f01, REQ_SAME_HOST_BRIDGE},
292 {}
293};
294
295static bool __host_bridge_whitelist(struct pci_host_bridge *host,
296 bool same_host_bridge)
297{
298 struct pci_dev *root = pci_get_slot(host->bus, PCI_DEVFN(0, 0));
299 const struct pci_p2pdma_whitelist_entry *entry;
300 unsigned short vendor, device;
301
302 if (!root)
303 return false;
304
305 vendor = root->vendor;
306 device = root->device;
307 pci_dev_put(root);
308
309 for (entry = pci_p2pdma_whitelist; entry->vendor; entry++) {
310 if (vendor != entry->vendor || device != entry->device)
311 continue;
312 if (entry->flags & REQ_SAME_HOST_BRIDGE && !same_host_bridge)
313 return false;
314
315 return true;
316 }
317
318 return false;
319}
320
321/*
322 * If we can't find a common upstream bridge take a look at the root
323 * complex and compare it to a whitelist of known good hardware.
324 */
325static bool host_bridge_whitelist(struct pci_dev *a, struct pci_dev *b)
326{
327 struct pci_host_bridge *host_a = pci_find_host_bridge(a->bus);
328 struct pci_host_bridge *host_b = pci_find_host_bridge(b->bus);
329
330 if (host_a == host_b)
331 return __host_bridge_whitelist(host_a, true);
332
333 if (__host_bridge_whitelist(host_a, false) &&
334 __host_bridge_whitelist(host_b, false))
335 return true;
336
337 return false;
338}
339
340static enum pci_p2pdma_map_type
341__upstream_bridge_distance(struct pci_dev *provider, struct pci_dev *client,
342 int *dist, bool *acs_redirects, struct seq_buf *acs_list)
343{
344 struct pci_dev *a = provider, *b = client, *bb;
345 int dist_a = 0;
346 int dist_b = 0;
347 int acs_cnt = 0;
348
349 if (acs_redirects)
350 *acs_redirects = false;
351
352 /*
353 * Note, we don't need to take references to devices returned by
354 * pci_upstream_bridge() seeing we hold a reference to a child
355 * device which will already hold a reference to the upstream bridge.
356 */
357
358 while (a) {
359 dist_b = 0;
360
361 if (pci_bridge_has_acs_redir(a)) {
362 seq_buf_print_bus_devfn(acs_list, a);
363 acs_cnt++;
364 }
365
366 bb = b;
367
368 while (bb) {
369 if (a == bb)
370 goto check_b_path_acs;
371
372 bb = pci_upstream_bridge(bb);
373 dist_b++;
374 }
375
376 a = pci_upstream_bridge(a);
377 dist_a++;
378 }
379
380 if (dist)
381 *dist = dist_a + dist_b;
382
383 return PCI_P2PDMA_MAP_THRU_HOST_BRIDGE;
384
385check_b_path_acs:
386 bb = b;
387
388 while (bb) {
389 if (a == bb)
390 break;
391
392 if (pci_bridge_has_acs_redir(bb)) {
393 seq_buf_print_bus_devfn(acs_list, bb);
394 acs_cnt++;
395 }
396
397 bb = pci_upstream_bridge(bb);
398 }
399
400 if (dist)
401 *dist = dist_a + dist_b;
402
403 if (acs_cnt) {
404 if (acs_redirects)
405 *acs_redirects = true;
406
407 return PCI_P2PDMA_MAP_THRU_HOST_BRIDGE;
408 }
409
410 return PCI_P2PDMA_MAP_BUS_ADDR;
411}
412
413static unsigned long map_types_idx(struct pci_dev *client)
414{
415 return (pci_domain_nr(client->bus) << 16) |
416 (client->bus->number << 8) | client->devfn;
417}
418
419/*
420 * Find the distance through the nearest common upstream bridge between
421 * two PCI devices.
422 *
423 * If the two devices are the same device then 0 will be returned.
424 *
425 * If there are two virtual functions of the same device behind the same
426 * bridge port then 2 will be returned (one step down to the PCIe switch,
427 * then one step back to the same device).
428 *
429 * In the case where two devices are connected to the same PCIe switch, the
430 * value 4 will be returned. This corresponds to the following PCI tree:
431 *
432 * -+ Root Port
433 * \+ Switch Upstream Port
434 * +-+ Switch Downstream Port
435 * + \- Device A
436 * \-+ Switch Downstream Port
437 * \- Device B
438 *
439 * The distance is 4 because we traverse from Device A through the downstream
440 * port of the switch, to the common upstream port, back up to the second
441 * downstream port and then to Device B.
442 *
443 * Any two devices that cannot communicate using p2pdma will return
444 * PCI_P2PDMA_MAP_NOT_SUPPORTED.
445 *
446 * Any two devices that have a data path that goes through the host bridge
447 * will consult a whitelist. If the host bridges are on the whitelist,
448 * this function will return PCI_P2PDMA_MAP_THRU_HOST_BRIDGE.
449 *
450 * If either bridge is not on the whitelist this function returns
451 * PCI_P2PDMA_MAP_NOT_SUPPORTED.
452 *
453 * If a bridge which has any ACS redirection bits set is in the path,
454 * acs_redirects will be set to true. In this case, a list of all infringing
455 * bridge addresses will be populated in acs_list (assuming it's non-null)
456 * for printk purposes.
457 */
458static enum pci_p2pdma_map_type
459upstream_bridge_distance(struct pci_dev *provider, struct pci_dev *client,
460 int *dist, bool *acs_redirects, struct seq_buf *acs_list)
461{
462 enum pci_p2pdma_map_type map_type;
463
464 map_type = __upstream_bridge_distance(provider, client, dist,
465 acs_redirects, acs_list);
466
467 if (map_type == PCI_P2PDMA_MAP_THRU_HOST_BRIDGE) {
468 if (!host_bridge_whitelist(provider, client))
469 map_type = PCI_P2PDMA_MAP_NOT_SUPPORTED;
470 }
471
472 if (provider->p2pdma)
473 xa_store(&provider->p2pdma->map_types, map_types_idx(client),
474 xa_mk_value(map_type), GFP_KERNEL);
475
476 return map_type;
477}
478
479static enum pci_p2pdma_map_type
480upstream_bridge_distance_warn(struct pci_dev *provider, struct pci_dev *client,
481 int *dist)
482{
483 struct seq_buf acs_list;
484 bool acs_redirects;
485 int ret;
486
487 seq_buf_init(&acs_list, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
488 if (!acs_list.buffer)
489 return -ENOMEM;
490
491 ret = upstream_bridge_distance(provider, client, dist, &acs_redirects,
492 &acs_list);
493 if (acs_redirects) {
494 pci_warn(client, "ACS redirect is set between the client and provider (%s)\n",
495 pci_name(provider));
496 /* Drop final semicolon */
497 acs_list.buffer[acs_list.len-1] = 0;
498 pci_warn(client, "to disable ACS redirect for this path, add the kernel parameter: pci=disable_acs_redir=%s\n",
499 acs_list.buffer);
500 }
501
502 if (ret == PCI_P2PDMA_MAP_NOT_SUPPORTED) {
503 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",
504 pci_name(provider));
505 }
506
507 kfree(acs_list.buffer);
508
509 return ret;
510}
511
512/**
513 * pci_p2pdma_distance_many - Determine the cumulative distance between
514 * a p2pdma provider and the clients in use.
515 * @provider: p2pdma provider to check against the client list
516 * @clients: array of devices to check (NULL-terminated)
517 * @num_clients: number of clients in the array
518 * @verbose: if true, print warnings for devices when we return -1
519 *
520 * Returns -1 if any of the clients are not compatible, otherwise returns a
521 * positive number where a lower number is the preferable choice. (If there's
522 * one client that's the same as the provider it will return 0, which is best
523 * choice).
524 *
525 * "compatible" means the provider and the clients are either all behind
526 * the same PCI root port or the host bridges connected to each of the devices
527 * are listed in the 'pci_p2pdma_whitelist'.
528 */
529int pci_p2pdma_distance_many(struct pci_dev *provider, struct device **clients,
530 int num_clients, bool verbose)
531{
532 bool not_supported = false;
533 struct pci_dev *pci_client;
534 int total_dist = 0;
535 int distance;
536 int i, ret;
537
538 if (num_clients == 0)
539 return -1;
540
541 for (i = 0; i < num_clients; i++) {
542 if (IS_ENABLED(CONFIG_DMA_VIRT_OPS) &&
543 clients[i]->dma_ops == &dma_virt_ops) {
544 if (verbose)
545 dev_warn(clients[i],
546 "cannot be used for peer-to-peer DMA because the driver makes use of dma_virt_ops\n");
547 return -1;
548 }
549
550 pci_client = find_parent_pci_dev(clients[i]);
551 if (!pci_client) {
552 if (verbose)
553 dev_warn(clients[i],
554 "cannot be used for peer-to-peer DMA as it is not a PCI device\n");
555 return -1;
556 }
557
558 if (verbose)
559 ret = upstream_bridge_distance_warn(provider,
560 pci_client, &distance);
561 else
562 ret = upstream_bridge_distance(provider, pci_client,
563 &distance, NULL, NULL);
564
565 pci_dev_put(pci_client);
566
567 if (ret == PCI_P2PDMA_MAP_NOT_SUPPORTED)
568 not_supported = true;
569
570 if (not_supported && !verbose)
571 break;
572
573 total_dist += distance;
574 }
575
576 if (not_supported)
577 return -1;
578
579 return total_dist;
580}
581EXPORT_SYMBOL_GPL(pci_p2pdma_distance_many);
582
583/**
584 * pci_has_p2pmem - check if a given PCI device has published any p2pmem
585 * @pdev: PCI device to check
586 */
587bool pci_has_p2pmem(struct pci_dev *pdev)
588{
589 return pdev->p2pdma && pdev->p2pdma->p2pmem_published;
590}
591EXPORT_SYMBOL_GPL(pci_has_p2pmem);
592
593/**
594 * pci_p2pmem_find - find a peer-to-peer DMA memory device compatible with
595 * the specified list of clients and shortest distance (as determined
596 * by pci_p2pmem_dma())
597 * @clients: array of devices to check (NULL-terminated)
598 * @num_clients: number of client devices in the list
599 *
600 * If multiple devices are behind the same switch, the one "closest" to the
601 * client devices in use will be chosen first. (So if one of the providers is
602 * the same as one of the clients, that provider will be used ahead of any
603 * other providers that are unrelated). If multiple providers are an equal
604 * distance away, one will be chosen at random.
605 *
606 * Returns a pointer to the PCI device with a reference taken (use pci_dev_put
607 * to return the reference) or NULL if no compatible device is found. The
608 * found provider will also be assigned to the client list.
609 */
610struct pci_dev *pci_p2pmem_find_many(struct device **clients, int num_clients)
611{
612 struct pci_dev *pdev = NULL;
613 int distance;
614 int closest_distance = INT_MAX;
615 struct pci_dev **closest_pdevs;
616 int dev_cnt = 0;
617 const int max_devs = PAGE_SIZE / sizeof(*closest_pdevs);
618 int i;
619
620 closest_pdevs = kmalloc(PAGE_SIZE, GFP_KERNEL);
621 if (!closest_pdevs)
622 return NULL;
623
624 while ((pdev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, pdev))) {
625 if (!pci_has_p2pmem(pdev))
626 continue;
627
628 distance = pci_p2pdma_distance_many(pdev, clients,
629 num_clients, false);
630 if (distance < 0 || distance > closest_distance)
631 continue;
632
633 if (distance == closest_distance && dev_cnt >= max_devs)
634 continue;
635
636 if (distance < closest_distance) {
637 for (i = 0; i < dev_cnt; i++)
638 pci_dev_put(closest_pdevs[i]);
639
640 dev_cnt = 0;
641 closest_distance = distance;
642 }
643
644 closest_pdevs[dev_cnt++] = pci_dev_get(pdev);
645 }
646
647 if (dev_cnt)
648 pdev = pci_dev_get(closest_pdevs[prandom_u32_max(dev_cnt)]);
649
650 for (i = 0; i < dev_cnt; i++)
651 pci_dev_put(closest_pdevs[i]);
652
653 kfree(closest_pdevs);
654 return pdev;
655}
656EXPORT_SYMBOL_GPL(pci_p2pmem_find_many);
657
658/**
659 * pci_alloc_p2p_mem - allocate peer-to-peer DMA memory
660 * @pdev: the device to allocate memory from
661 * @size: number of bytes to allocate
662 *
663 * Returns the allocated memory or NULL on error.
664 */
665void *pci_alloc_p2pmem(struct pci_dev *pdev, size_t size)
666{
667 void *ret = NULL;
668 struct percpu_ref *ref;
669
670 /*
671 * Pairs with synchronize_rcu() in pci_p2pdma_release() to
672 * ensure pdev->p2pdma is non-NULL for the duration of the
673 * read-lock.
674 */
675 rcu_read_lock();
676 if (unlikely(!pdev->p2pdma))
677 goto out;
678
679 ret = (void *)gen_pool_alloc_owner(pdev->p2pdma->pool, size,
680 (void **) &ref);
681 if (!ret)
682 goto out;
683
684 if (unlikely(!percpu_ref_tryget_live(ref))) {
685 gen_pool_free(pdev->p2pdma->pool, (unsigned long) ret, size);
686 ret = NULL;
687 goto out;
688 }
689out:
690 rcu_read_unlock();
691 return ret;
692}
693EXPORT_SYMBOL_GPL(pci_alloc_p2pmem);
694
695/**
696 * pci_free_p2pmem - free peer-to-peer DMA memory
697 * @pdev: the device the memory was allocated from
698 * @addr: address of the memory that was allocated
699 * @size: number of bytes that were allocated
700 */
701void pci_free_p2pmem(struct pci_dev *pdev, void *addr, size_t size)
702{
703 struct percpu_ref *ref;
704
705 gen_pool_free_owner(pdev->p2pdma->pool, (uintptr_t)addr, size,
706 (void **) &ref);
707 percpu_ref_put(ref);
708}
709EXPORT_SYMBOL_GPL(pci_free_p2pmem);
710
711/**
712 * pci_virt_to_bus - return the PCI bus address for a given virtual
713 * address obtained with pci_alloc_p2pmem()
714 * @pdev: the device the memory was allocated from
715 * @addr: address of the memory that was allocated
716 */
717pci_bus_addr_t pci_p2pmem_virt_to_bus(struct pci_dev *pdev, void *addr)
718{
719 if (!addr)
720 return 0;
721 if (!pdev->p2pdma)
722 return 0;
723
724 /*
725 * Note: when we added the memory to the pool we used the PCI
726 * bus address as the physical address. So gen_pool_virt_to_phys()
727 * actually returns the bus address despite the misleading name.
728 */
729 return gen_pool_virt_to_phys(pdev->p2pdma->pool, (unsigned long)addr);
730}
731EXPORT_SYMBOL_GPL(pci_p2pmem_virt_to_bus);
732
733/**
734 * pci_p2pmem_alloc_sgl - allocate peer-to-peer DMA memory in a scatterlist
735 * @pdev: the device to allocate memory from
736 * @nents: the number of SG entries in the list
737 * @length: number of bytes to allocate
738 *
739 * Return: %NULL on error or &struct scatterlist pointer and @nents on success
740 */
741struct scatterlist *pci_p2pmem_alloc_sgl(struct pci_dev *pdev,
742 unsigned int *nents, u32 length)
743{
744 struct scatterlist *sg;
745 void *addr;
746
747 sg = kzalloc(sizeof(*sg), GFP_KERNEL);
748 if (!sg)
749 return NULL;
750
751 sg_init_table(sg, 1);
752
753 addr = pci_alloc_p2pmem(pdev, length);
754 if (!addr)
755 goto out_free_sg;
756
757 sg_set_buf(sg, addr, length);
758 *nents = 1;
759 return sg;
760
761out_free_sg:
762 kfree(sg);
763 return NULL;
764}
765EXPORT_SYMBOL_GPL(pci_p2pmem_alloc_sgl);
766
767/**
768 * pci_p2pmem_free_sgl - free a scatterlist allocated by pci_p2pmem_alloc_sgl()
769 * @pdev: the device to allocate memory from
770 * @sgl: the allocated scatterlist
771 */
772void pci_p2pmem_free_sgl(struct pci_dev *pdev, struct scatterlist *sgl)
773{
774 struct scatterlist *sg;
775 int count;
776
777 for_each_sg(sgl, sg, INT_MAX, count) {
778 if (!sg)
779 break;
780
781 pci_free_p2pmem(pdev, sg_virt(sg), sg->length);
782 }
783 kfree(sgl);
784}
785EXPORT_SYMBOL_GPL(pci_p2pmem_free_sgl);
786
787/**
788 * pci_p2pmem_publish - publish the peer-to-peer DMA memory for use by
789 * other devices with pci_p2pmem_find()
790 * @pdev: the device with peer-to-peer DMA memory to publish
791 * @publish: set to true to publish the memory, false to unpublish it
792 *
793 * Published memory can be used by other PCI device drivers for
794 * peer-2-peer DMA operations. Non-published memory is reserved for
795 * exclusive use of the device driver that registers the peer-to-peer
796 * memory.
797 */
798void pci_p2pmem_publish(struct pci_dev *pdev, bool publish)
799{
800 if (pdev->p2pdma)
801 pdev->p2pdma->p2pmem_published = publish;
802}
803EXPORT_SYMBOL_GPL(pci_p2pmem_publish);
804
805static enum pci_p2pdma_map_type pci_p2pdma_map_type(struct pci_dev *provider,
806 struct pci_dev *client)
807{
808 if (!provider->p2pdma)
809 return PCI_P2PDMA_MAP_NOT_SUPPORTED;
810
811 return xa_to_value(xa_load(&provider->p2pdma->map_types,
812 map_types_idx(client)));
813}
814
815static int __pci_p2pdma_map_sg(struct pci_p2pdma_pagemap *p2p_pgmap,
816 struct device *dev, struct scatterlist *sg, int nents)
817{
818 struct scatterlist *s;
819 phys_addr_t paddr;
820 int i;
821
822 /*
823 * p2pdma mappings are not compatible with devices that use
824 * dma_virt_ops. If the upper layers do the right thing
825 * this should never happen because it will be prevented
826 * by the check in pci_p2pdma_distance_many()
827 */
828 if (WARN_ON_ONCE(IS_ENABLED(CONFIG_DMA_VIRT_OPS) &&
829 dev->dma_ops == &dma_virt_ops))
830 return 0;
831
832 for_each_sg(sg, s, nents, i) {
833 paddr = sg_phys(s);
834
835 s->dma_address = paddr - p2p_pgmap->bus_offset;
836 sg_dma_len(s) = s->length;
837 }
838
839 return nents;
840}
841
842/**
843 * pci_p2pdma_map_sg - map a PCI peer-to-peer scatterlist for DMA
844 * @dev: device doing the DMA request
845 * @sg: scatter list to map
846 * @nents: elements in the scatterlist
847 * @dir: DMA direction
848 * @attrs: DMA attributes passed to dma_map_sg() (if called)
849 *
850 * Scatterlists mapped with this function should be unmapped using
851 * pci_p2pdma_unmap_sg_attrs().
852 *
853 * Returns the number of SG entries mapped or 0 on error.
854 */
855int pci_p2pdma_map_sg_attrs(struct device *dev, struct scatterlist *sg,
856 int nents, enum dma_data_direction dir, unsigned long attrs)
857{
858 struct pci_p2pdma_pagemap *p2p_pgmap =
859 to_p2p_pgmap(sg_page(sg)->pgmap);
860 struct pci_dev *client;
861
862 if (WARN_ON_ONCE(!dev_is_pci(dev)))
863 return 0;
864
865 client = to_pci_dev(dev);
866
867 switch (pci_p2pdma_map_type(p2p_pgmap->provider, client)) {
868 case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
869 return dma_map_sg_attrs(dev, sg, nents, dir, attrs);
870 case PCI_P2PDMA_MAP_BUS_ADDR:
871 return __pci_p2pdma_map_sg(p2p_pgmap, dev, sg, nents);
872 default:
873 WARN_ON_ONCE(1);
874 return 0;
875 }
876}
877EXPORT_SYMBOL_GPL(pci_p2pdma_map_sg_attrs);
878
879/**
880 * pci_p2pdma_unmap_sg - unmap a PCI peer-to-peer scatterlist that was
881 * mapped with pci_p2pdma_map_sg()
882 * @dev: device doing the DMA request
883 * @sg: scatter list to map
884 * @nents: number of elements returned by pci_p2pdma_map_sg()
885 * @dir: DMA direction
886 * @attrs: DMA attributes passed to dma_unmap_sg() (if called)
887 */
888void pci_p2pdma_unmap_sg_attrs(struct device *dev, struct scatterlist *sg,
889 int nents, enum dma_data_direction dir, unsigned long attrs)
890{
891 struct pci_p2pdma_pagemap *p2p_pgmap =
892 to_p2p_pgmap(sg_page(sg)->pgmap);
893 enum pci_p2pdma_map_type map_type;
894 struct pci_dev *client;
895
896 if (WARN_ON_ONCE(!dev_is_pci(dev)))
897 return;
898
899 client = to_pci_dev(dev);
900
901 map_type = pci_p2pdma_map_type(p2p_pgmap->provider, client);
902
903 if (map_type == PCI_P2PDMA_MAP_THRU_HOST_BRIDGE)
904 dma_unmap_sg_attrs(dev, sg, nents, dir, attrs);
905}
906EXPORT_SYMBOL_GPL(pci_p2pdma_unmap_sg_attrs);
907
908/**
909 * pci_p2pdma_enable_store - parse a configfs/sysfs attribute store
910 * to enable p2pdma
911 * @page: contents of the value to be stored
912 * @p2p_dev: returns the PCI device that was selected to be used
913 * (if one was specified in the stored value)
914 * @use_p2pdma: returns whether to enable p2pdma or not
915 *
916 * Parses an attribute value to decide whether to enable p2pdma.
917 * The value can select a PCI device (using its full BDF device
918 * name) or a boolean (in any format strtobool() accepts). A false
919 * value disables p2pdma, a true value expects the caller
920 * to automatically find a compatible device and specifying a PCI device
921 * expects the caller to use the specific provider.
922 *
923 * pci_p2pdma_enable_show() should be used as the show operation for
924 * the attribute.
925 *
926 * Returns 0 on success
927 */
928int pci_p2pdma_enable_store(const char *page, struct pci_dev **p2p_dev,
929 bool *use_p2pdma)
930{
931 struct device *dev;
932
933 dev = bus_find_device_by_name(&pci_bus_type, NULL, page);
934 if (dev) {
935 *use_p2pdma = true;
936 *p2p_dev = to_pci_dev(dev);
937
938 if (!pci_has_p2pmem(*p2p_dev)) {
939 pci_err(*p2p_dev,
940 "PCI device has no peer-to-peer memory: %s\n",
941 page);
942 pci_dev_put(*p2p_dev);
943 return -ENODEV;
944 }
945
946 return 0;
947 } else if ((page[0] == '0' || page[0] == '1') && !iscntrl(page[1])) {
948 /*
949 * If the user enters a PCI device that doesn't exist
950 * like "0000:01:00.1", we don't want strtobool to think
951 * it's a '0' when it's clearly not what the user wanted.
952 * So we require 0's and 1's to be exactly one character.
953 */
954 } else if (!strtobool(page, use_p2pdma)) {
955 return 0;
956 }
957
958 pr_err("No such PCI device: %.*s\n", (int)strcspn(page, "\n"), page);
959 return -ENODEV;
960}
961EXPORT_SYMBOL_GPL(pci_p2pdma_enable_store);
962
963/**
964 * pci_p2pdma_enable_show - show a configfs/sysfs attribute indicating
965 * whether p2pdma is enabled
966 * @page: contents of the stored value
967 * @p2p_dev: the selected p2p device (NULL if no device is selected)
968 * @use_p2pdma: whether p2pdma has been enabled
969 *
970 * Attributes that use pci_p2pdma_enable_store() should use this function
971 * to show the value of the attribute.
972 *
973 * Returns 0 on success
974 */
975ssize_t pci_p2pdma_enable_show(char *page, struct pci_dev *p2p_dev,
976 bool use_p2pdma)
977{
978 if (!use_p2pdma)
979 return sprintf(page, "0\n");
980
981 if (!p2p_dev)
982 return sprintf(page, "1\n");
983
984 return sprintf(page, "%s\n", pci_name(p2p_dev));
985}
986EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show);
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/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/xarray.h>
22
23enum pci_p2pdma_map_type {
24 PCI_P2PDMA_MAP_UNKNOWN = 0,
25 PCI_P2PDMA_MAP_NOT_SUPPORTED,
26 PCI_P2PDMA_MAP_BUS_ADDR,
27 PCI_P2PDMA_MAP_THRU_HOST_BRIDGE,
28};
29
30struct pci_p2pdma {
31 struct gen_pool *pool;
32 bool p2pmem_published;
33 struct xarray map_types;
34};
35
36struct pci_p2pdma_pagemap {
37 struct dev_pagemap pgmap;
38 struct pci_dev *provider;
39 u64 bus_offset;
40};
41
42static struct pci_p2pdma_pagemap *to_p2p_pgmap(struct dev_pagemap *pgmap)
43{
44 return container_of(pgmap, struct pci_p2pdma_pagemap, pgmap);
45}
46
47static ssize_t size_show(struct device *dev, struct device_attribute *attr,
48 char *buf)
49{
50 struct pci_dev *pdev = to_pci_dev(dev);
51 struct pci_p2pdma *p2pdma;
52 size_t size = 0;
53
54 rcu_read_lock();
55 p2pdma = rcu_dereference(pdev->p2pdma);
56 if (p2pdma && p2pdma->pool)
57 size = gen_pool_size(p2pdma->pool);
58 rcu_read_unlock();
59
60 return sysfs_emit(buf, "%zd\n", size);
61}
62static DEVICE_ATTR_RO(size);
63
64static ssize_t available_show(struct device *dev, struct device_attribute *attr,
65 char *buf)
66{
67 struct pci_dev *pdev = to_pci_dev(dev);
68 struct pci_p2pdma *p2pdma;
69 size_t avail = 0;
70
71 rcu_read_lock();
72 p2pdma = rcu_dereference(pdev->p2pdma);
73 if (p2pdma && p2pdma->pool)
74 avail = gen_pool_avail(p2pdma->pool);
75 rcu_read_unlock();
76
77 return sysfs_emit(buf, "%zd\n", avail);
78}
79static DEVICE_ATTR_RO(available);
80
81static ssize_t published_show(struct device *dev, struct device_attribute *attr,
82 char *buf)
83{
84 struct pci_dev *pdev = to_pci_dev(dev);
85 struct pci_p2pdma *p2pdma;
86 bool published = false;
87
88 rcu_read_lock();
89 p2pdma = rcu_dereference(pdev->p2pdma);
90 if (p2pdma)
91 published = p2pdma->p2pmem_published;
92 rcu_read_unlock();
93
94 return sysfs_emit(buf, "%d\n", published);
95}
96static DEVICE_ATTR_RO(published);
97
98static struct attribute *p2pmem_attrs[] = {
99 &dev_attr_size.attr,
100 &dev_attr_available.attr,
101 &dev_attr_published.attr,
102 NULL,
103};
104
105static const struct attribute_group p2pmem_group = {
106 .attrs = p2pmem_attrs,
107 .name = "p2pmem",
108};
109
110static void pci_p2pdma_release(void *data)
111{
112 struct pci_dev *pdev = data;
113 struct pci_p2pdma *p2pdma;
114
115 p2pdma = rcu_dereference_protected(pdev->p2pdma, 1);
116 if (!p2pdma)
117 return;
118
119 /* Flush and disable pci_alloc_p2p_mem() */
120 pdev->p2pdma = NULL;
121 synchronize_rcu();
122
123 gen_pool_destroy(p2pdma->pool);
124 sysfs_remove_group(&pdev->dev.kobj, &p2pmem_group);
125 xa_destroy(&p2pdma->map_types);
126}
127
128static int pci_p2pdma_setup(struct pci_dev *pdev)
129{
130 int error = -ENOMEM;
131 struct pci_p2pdma *p2p;
132
133 p2p = devm_kzalloc(&pdev->dev, sizeof(*p2p), GFP_KERNEL);
134 if (!p2p)
135 return -ENOMEM;
136
137 xa_init(&p2p->map_types);
138
139 p2p->pool = gen_pool_create(PAGE_SHIFT, dev_to_node(&pdev->dev));
140 if (!p2p->pool)
141 goto out;
142
143 error = devm_add_action_or_reset(&pdev->dev, pci_p2pdma_release, pdev);
144 if (error)
145 goto out_pool_destroy;
146
147 error = sysfs_create_group(&pdev->dev.kobj, &p2pmem_group);
148 if (error)
149 goto out_pool_destroy;
150
151 rcu_assign_pointer(pdev->p2pdma, p2p);
152 return 0;
153
154out_pool_destroy:
155 gen_pool_destroy(p2p->pool);
156out:
157 devm_kfree(&pdev->dev, p2p);
158 return error;
159}
160
161/**
162 * pci_p2pdma_add_resource - add memory for use as p2p memory
163 * @pdev: the device to add the memory to
164 * @bar: PCI BAR to add
165 * @size: size of the memory to add, may be zero to use the whole BAR
166 * @offset: offset into the PCI BAR
167 *
168 * The memory will be given ZONE_DEVICE struct pages so that it may
169 * be used with any DMA request.
170 */
171int pci_p2pdma_add_resource(struct pci_dev *pdev, int bar, size_t size,
172 u64 offset)
173{
174 struct pci_p2pdma_pagemap *p2p_pgmap;
175 struct dev_pagemap *pgmap;
176 struct pci_p2pdma *p2pdma;
177 void *addr;
178 int error;
179
180 if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM))
181 return -EINVAL;
182
183 if (offset >= pci_resource_len(pdev, bar))
184 return -EINVAL;
185
186 if (!size)
187 size = pci_resource_len(pdev, bar) - offset;
188
189 if (size + offset > pci_resource_len(pdev, bar))
190 return -EINVAL;
191
192 if (!pdev->p2pdma) {
193 error = pci_p2pdma_setup(pdev);
194 if (error)
195 return error;
196 }
197
198 p2p_pgmap = devm_kzalloc(&pdev->dev, sizeof(*p2p_pgmap), GFP_KERNEL);
199 if (!p2p_pgmap)
200 return -ENOMEM;
201
202 pgmap = &p2p_pgmap->pgmap;
203 pgmap->range.start = pci_resource_start(pdev, bar) + offset;
204 pgmap->range.end = pgmap->range.start + size - 1;
205 pgmap->nr_range = 1;
206 pgmap->type = MEMORY_DEVICE_PCI_P2PDMA;
207
208 p2p_pgmap->provider = pdev;
209 p2p_pgmap->bus_offset = pci_bus_address(pdev, bar) -
210 pci_resource_start(pdev, bar);
211
212 addr = devm_memremap_pages(&pdev->dev, pgmap);
213 if (IS_ERR(addr)) {
214 error = PTR_ERR(addr);
215 goto pgmap_free;
216 }
217
218 p2pdma = rcu_dereference_protected(pdev->p2pdma, 1);
219 error = gen_pool_add_owner(p2pdma->pool, (unsigned long)addr,
220 pci_bus_address(pdev, bar) + offset,
221 range_len(&pgmap->range), dev_to_node(&pdev->dev),
222 pgmap->ref);
223 if (error)
224 goto pages_free;
225
226 pci_info(pdev, "added peer-to-peer DMA memory %#llx-%#llx\n",
227 pgmap->range.start, pgmap->range.end);
228
229 return 0;
230
231pages_free:
232 devm_memunmap_pages(&pdev->dev, pgmap);
233pgmap_free:
234 devm_kfree(&pdev->dev, pgmap);
235 return error;
236}
237EXPORT_SYMBOL_GPL(pci_p2pdma_add_resource);
238
239/*
240 * Note this function returns the parent PCI device with a
241 * reference taken. It is the caller's responsibility to drop
242 * the reference.
243 */
244static struct pci_dev *find_parent_pci_dev(struct device *dev)
245{
246 struct device *parent;
247
248 dev = get_device(dev);
249
250 while (dev) {
251 if (dev_is_pci(dev))
252 return to_pci_dev(dev);
253
254 parent = get_device(dev->parent);
255 put_device(dev);
256 dev = parent;
257 }
258
259 return NULL;
260}
261
262/*
263 * Check if a PCI bridge has its ACS redirection bits set to redirect P2P
264 * TLPs upstream via ACS. Returns 1 if the packets will be redirected
265 * upstream, 0 otherwise.
266 */
267static int pci_bridge_has_acs_redir(struct pci_dev *pdev)
268{
269 int pos;
270 u16 ctrl;
271
272 pos = pdev->acs_cap;
273 if (!pos)
274 return 0;
275
276 pci_read_config_word(pdev, pos + PCI_ACS_CTRL, &ctrl);
277
278 if (ctrl & (PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_EC))
279 return 1;
280
281 return 0;
282}
283
284static void seq_buf_print_bus_devfn(struct seq_buf *buf, struct pci_dev *pdev)
285{
286 if (!buf)
287 return;
288
289 seq_buf_printf(buf, "%s;", pci_name(pdev));
290}
291
292static bool cpu_supports_p2pdma(void)
293{
294#ifdef CONFIG_X86
295 struct cpuinfo_x86 *c = &cpu_data(0);
296
297 /* Any AMD CPU whose family ID is Zen or newer supports p2pdma */
298 if (c->x86_vendor == X86_VENDOR_AMD && c->x86 >= 0x17)
299 return true;
300#endif
301
302 return false;
303}
304
305static const struct pci_p2pdma_whitelist_entry {
306 unsigned short vendor;
307 unsigned short device;
308 enum {
309 REQ_SAME_HOST_BRIDGE = 1 << 0,
310 } flags;
311} pci_p2pdma_whitelist[] = {
312 /* Intel Xeon E5/Core i7 */
313 {PCI_VENDOR_ID_INTEL, 0x3c00, REQ_SAME_HOST_BRIDGE},
314 {PCI_VENDOR_ID_INTEL, 0x3c01, REQ_SAME_HOST_BRIDGE},
315 /* Intel Xeon E7 v3/Xeon E5 v3/Core i7 */
316 {PCI_VENDOR_ID_INTEL, 0x2f00, REQ_SAME_HOST_BRIDGE},
317 {PCI_VENDOR_ID_INTEL, 0x2f01, REQ_SAME_HOST_BRIDGE},
318 /* Intel SkyLake-E */
319 {PCI_VENDOR_ID_INTEL, 0x2030, 0},
320 {PCI_VENDOR_ID_INTEL, 0x2031, 0},
321 {PCI_VENDOR_ID_INTEL, 0x2032, 0},
322 {PCI_VENDOR_ID_INTEL, 0x2033, 0},
323 {PCI_VENDOR_ID_INTEL, 0x2020, 0},
324 {}
325};
326
327/*
328 * This lookup function tries to find the PCI device corresponding to a given
329 * host bridge.
330 *
331 * It assumes the host bridge device is the first PCI device in the
332 * bus->devices list and that the devfn is 00.0. These assumptions should hold
333 * for all the devices in the whitelist above.
334 *
335 * This function is equivalent to pci_get_slot(host->bus, 0), however it does
336 * not take the pci_bus_sem lock seeing __host_bridge_whitelist() must not
337 * sleep.
338 *
339 * For this to be safe, the caller should hold a reference to a device on the
340 * bridge, which should ensure the host_bridge device will not be freed
341 * or removed from the head of the devices list.
342 */
343static struct pci_dev *pci_host_bridge_dev(struct pci_host_bridge *host)
344{
345 struct pci_dev *root;
346
347 root = list_first_entry_or_null(&host->bus->devices,
348 struct pci_dev, bus_list);
349
350 if (!root)
351 return NULL;
352 if (root->devfn != PCI_DEVFN(0, 0))
353 return NULL;
354
355 return root;
356}
357
358static bool __host_bridge_whitelist(struct pci_host_bridge *host,
359 bool same_host_bridge, bool warn)
360{
361 struct pci_dev *root = pci_host_bridge_dev(host);
362 const struct pci_p2pdma_whitelist_entry *entry;
363 unsigned short vendor, device;
364
365 if (!root)
366 return false;
367
368 vendor = root->vendor;
369 device = root->device;
370
371 for (entry = pci_p2pdma_whitelist; entry->vendor; entry++) {
372 if (vendor != entry->vendor || device != entry->device)
373 continue;
374 if (entry->flags & REQ_SAME_HOST_BRIDGE && !same_host_bridge)
375 return false;
376
377 return true;
378 }
379
380 if (warn)
381 pci_warn(root, "Host bridge not in P2PDMA whitelist: %04x:%04x\n",
382 vendor, device);
383
384 return false;
385}
386
387/*
388 * If we can't find a common upstream bridge take a look at the root
389 * complex and compare it to a whitelist of known good hardware.
390 */
391static bool host_bridge_whitelist(struct pci_dev *a, struct pci_dev *b,
392 bool warn)
393{
394 struct pci_host_bridge *host_a = pci_find_host_bridge(a->bus);
395 struct pci_host_bridge *host_b = pci_find_host_bridge(b->bus);
396
397 if (host_a == host_b)
398 return __host_bridge_whitelist(host_a, true, warn);
399
400 if (__host_bridge_whitelist(host_a, false, warn) &&
401 __host_bridge_whitelist(host_b, false, warn))
402 return true;
403
404 return false;
405}
406
407static unsigned long map_types_idx(struct pci_dev *client)
408{
409 return (pci_domain_nr(client->bus) << 16) |
410 (client->bus->number << 8) | client->devfn;
411}
412
413/*
414 * Calculate the P2PDMA mapping type and distance between two PCI devices.
415 *
416 * If the two devices are the same PCI function, return
417 * PCI_P2PDMA_MAP_BUS_ADDR and a distance of 0.
418 *
419 * If they are two functions of the same device, return
420 * PCI_P2PDMA_MAP_BUS_ADDR and a distance of 2 (one hop up to the bridge,
421 * then one hop back down to another function of the same device).
422 *
423 * In the case where two devices are connected to the same PCIe switch,
424 * return a distance of 4. This corresponds to the following PCI tree:
425 *
426 * -+ Root Port
427 * \+ Switch Upstream Port
428 * +-+ Switch Downstream Port 0
429 * + \- Device A
430 * \-+ Switch Downstream Port 1
431 * \- Device B
432 *
433 * The distance is 4 because we traverse from Device A to Downstream Port 0
434 * to the common Switch Upstream Port, back down to Downstream Port 1 and
435 * then to Device B. The mapping type returned depends on the ACS
436 * redirection setting of the ports along the path.
437 *
438 * If ACS redirect is set on any port in the path, traffic between the
439 * devices will go through the host bridge, so return
440 * PCI_P2PDMA_MAP_THRU_HOST_BRIDGE; otherwise return
441 * PCI_P2PDMA_MAP_BUS_ADDR.
442 *
443 * Any two devices that have a data path that goes through the host bridge
444 * will consult a whitelist. If the host bridge is in the whitelist, return
445 * PCI_P2PDMA_MAP_THRU_HOST_BRIDGE with the distance set to the number of
446 * ports per above. If the device is not in the whitelist, return
447 * PCI_P2PDMA_MAP_NOT_SUPPORTED.
448 */
449static enum pci_p2pdma_map_type
450calc_map_type_and_dist(struct pci_dev *provider, struct pci_dev *client,
451 int *dist, bool verbose)
452{
453 enum pci_p2pdma_map_type map_type = PCI_P2PDMA_MAP_THRU_HOST_BRIDGE;
454 struct pci_dev *a = provider, *b = client, *bb;
455 bool acs_redirects = false;
456 struct pci_p2pdma *p2pdma;
457 struct seq_buf acs_list;
458 int acs_cnt = 0;
459 int dist_a = 0;
460 int dist_b = 0;
461 char buf[128];
462
463 seq_buf_init(&acs_list, buf, sizeof(buf));
464
465 /*
466 * Note, we don't need to take references to devices returned by
467 * pci_upstream_bridge() seeing we hold a reference to a child
468 * device which will already hold a reference to the upstream bridge.
469 */
470 while (a) {
471 dist_b = 0;
472
473 if (pci_bridge_has_acs_redir(a)) {
474 seq_buf_print_bus_devfn(&acs_list, a);
475 acs_cnt++;
476 }
477
478 bb = b;
479
480 while (bb) {
481 if (a == bb)
482 goto check_b_path_acs;
483
484 bb = pci_upstream_bridge(bb);
485 dist_b++;
486 }
487
488 a = pci_upstream_bridge(a);
489 dist_a++;
490 }
491
492 *dist = dist_a + dist_b;
493 goto map_through_host_bridge;
494
495check_b_path_acs:
496 bb = b;
497
498 while (bb) {
499 if (a == bb)
500 break;
501
502 if (pci_bridge_has_acs_redir(bb)) {
503 seq_buf_print_bus_devfn(&acs_list, bb);
504 acs_cnt++;
505 }
506
507 bb = pci_upstream_bridge(bb);
508 }
509
510 *dist = dist_a + dist_b;
511
512 if (!acs_cnt) {
513 map_type = PCI_P2PDMA_MAP_BUS_ADDR;
514 goto done;
515 }
516
517 if (verbose) {
518 acs_list.buffer[acs_list.len-1] = 0; /* drop final semicolon */
519 pci_warn(client, "ACS redirect is set between the client and provider (%s)\n",
520 pci_name(provider));
521 pci_warn(client, "to disable ACS redirect for this path, add the kernel parameter: pci=disable_acs_redir=%s\n",
522 acs_list.buffer);
523 }
524 acs_redirects = true;
525
526map_through_host_bridge:
527 if (!cpu_supports_p2pdma() &&
528 !host_bridge_whitelist(provider, client, acs_redirects)) {
529 if (verbose)
530 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",
531 pci_name(provider));
532 map_type = PCI_P2PDMA_MAP_NOT_SUPPORTED;
533 }
534done:
535 rcu_read_lock();
536 p2pdma = rcu_dereference(provider->p2pdma);
537 if (p2pdma)
538 xa_store(&p2pdma->map_types, map_types_idx(client),
539 xa_mk_value(map_type), GFP_KERNEL);
540 rcu_read_unlock();
541 return map_type;
542}
543
544/**
545 * pci_p2pdma_distance_many - Determine the cumulative distance between
546 * a p2pdma provider and the clients in use.
547 * @provider: p2pdma provider to check against the client list
548 * @clients: array of devices to check (NULL-terminated)
549 * @num_clients: number of clients in the array
550 * @verbose: if true, print warnings for devices when we return -1
551 *
552 * Returns -1 if any of the clients are not compatible, otherwise returns a
553 * positive number where a lower number is the preferable choice. (If there's
554 * one client that's the same as the provider it will return 0, which is best
555 * choice).
556 *
557 * "compatible" means the provider and the clients are either all behind
558 * the same PCI root port or the host bridges connected to each of the devices
559 * are listed in the 'pci_p2pdma_whitelist'.
560 */
561int pci_p2pdma_distance_many(struct pci_dev *provider, struct device **clients,
562 int num_clients, bool verbose)
563{
564 enum pci_p2pdma_map_type map;
565 bool not_supported = false;
566 struct pci_dev *pci_client;
567 int total_dist = 0;
568 int i, distance;
569
570 if (num_clients == 0)
571 return -1;
572
573 for (i = 0; i < num_clients; i++) {
574 pci_client = find_parent_pci_dev(clients[i]);
575 if (!pci_client) {
576 if (verbose)
577 dev_warn(clients[i],
578 "cannot be used for peer-to-peer DMA as it is not a PCI device\n");
579 return -1;
580 }
581
582 map = calc_map_type_and_dist(provider, pci_client, &distance,
583 verbose);
584
585 pci_dev_put(pci_client);
586
587 if (map == PCI_P2PDMA_MAP_NOT_SUPPORTED)
588 not_supported = true;
589
590 if (not_supported && !verbose)
591 break;
592
593 total_dist += distance;
594 }
595
596 if (not_supported)
597 return -1;
598
599 return total_dist;
600}
601EXPORT_SYMBOL_GPL(pci_p2pdma_distance_many);
602
603/**
604 * pci_has_p2pmem - check if a given PCI device has published any p2pmem
605 * @pdev: PCI device to check
606 */
607bool pci_has_p2pmem(struct pci_dev *pdev)
608{
609 struct pci_p2pdma *p2pdma;
610 bool res;
611
612 rcu_read_lock();
613 p2pdma = rcu_dereference(pdev->p2pdma);
614 res = p2pdma && p2pdma->p2pmem_published;
615 rcu_read_unlock();
616
617 return res;
618}
619EXPORT_SYMBOL_GPL(pci_has_p2pmem);
620
621/**
622 * pci_p2pmem_find_many - find a peer-to-peer DMA memory device compatible with
623 * the specified list of clients and shortest distance (as determined
624 * by pci_p2pmem_dma())
625 * @clients: array of devices to check (NULL-terminated)
626 * @num_clients: number of client devices in the list
627 *
628 * If multiple devices are behind the same switch, the one "closest" to the
629 * client devices in use will be chosen first. (So if one of the providers is
630 * the same as one of the clients, that provider will be used ahead of any
631 * other providers that are unrelated). If multiple providers are an equal
632 * distance away, one will be chosen at random.
633 *
634 * Returns a pointer to the PCI device with a reference taken (use pci_dev_put
635 * to return the reference) or NULL if no compatible device is found. The
636 * found provider will also be assigned to the client list.
637 */
638struct pci_dev *pci_p2pmem_find_many(struct device **clients, int num_clients)
639{
640 struct pci_dev *pdev = NULL;
641 int distance;
642 int closest_distance = INT_MAX;
643 struct pci_dev **closest_pdevs;
644 int dev_cnt = 0;
645 const int max_devs = PAGE_SIZE / sizeof(*closest_pdevs);
646 int i;
647
648 closest_pdevs = kmalloc(PAGE_SIZE, GFP_KERNEL);
649 if (!closest_pdevs)
650 return NULL;
651
652 while ((pdev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, pdev))) {
653 if (!pci_has_p2pmem(pdev))
654 continue;
655
656 distance = pci_p2pdma_distance_many(pdev, clients,
657 num_clients, false);
658 if (distance < 0 || distance > closest_distance)
659 continue;
660
661 if (distance == closest_distance && dev_cnt >= max_devs)
662 continue;
663
664 if (distance < closest_distance) {
665 for (i = 0; i < dev_cnt; i++)
666 pci_dev_put(closest_pdevs[i]);
667
668 dev_cnt = 0;
669 closest_distance = distance;
670 }
671
672 closest_pdevs[dev_cnt++] = pci_dev_get(pdev);
673 }
674
675 if (dev_cnt)
676 pdev = pci_dev_get(closest_pdevs[prandom_u32_max(dev_cnt)]);
677
678 for (i = 0; i < dev_cnt; i++)
679 pci_dev_put(closest_pdevs[i]);
680
681 kfree(closest_pdevs);
682 return pdev;
683}
684EXPORT_SYMBOL_GPL(pci_p2pmem_find_many);
685
686/**
687 * pci_alloc_p2pmem - allocate peer-to-peer DMA memory
688 * @pdev: the device to allocate memory from
689 * @size: number of bytes to allocate
690 *
691 * Returns the allocated memory or NULL on error.
692 */
693void *pci_alloc_p2pmem(struct pci_dev *pdev, size_t size)
694{
695 void *ret = NULL;
696 struct percpu_ref *ref;
697 struct pci_p2pdma *p2pdma;
698
699 /*
700 * Pairs with synchronize_rcu() in pci_p2pdma_release() to
701 * ensure pdev->p2pdma is non-NULL for the duration of the
702 * read-lock.
703 */
704 rcu_read_lock();
705 p2pdma = rcu_dereference(pdev->p2pdma);
706 if (unlikely(!p2pdma))
707 goto out;
708
709 ret = (void *)gen_pool_alloc_owner(p2pdma->pool, size, (void **) &ref);
710 if (!ret)
711 goto out;
712
713 if (unlikely(!percpu_ref_tryget_live(ref))) {
714 gen_pool_free(p2pdma->pool, (unsigned long) ret, size);
715 ret = NULL;
716 goto out;
717 }
718out:
719 rcu_read_unlock();
720 return ret;
721}
722EXPORT_SYMBOL_GPL(pci_alloc_p2pmem);
723
724/**
725 * pci_free_p2pmem - free peer-to-peer DMA memory
726 * @pdev: the device the memory was allocated from
727 * @addr: address of the memory that was allocated
728 * @size: number of bytes that were allocated
729 */
730void pci_free_p2pmem(struct pci_dev *pdev, void *addr, size_t size)
731{
732 struct percpu_ref *ref;
733 struct pci_p2pdma *p2pdma = rcu_dereference_protected(pdev->p2pdma, 1);
734
735 gen_pool_free_owner(p2pdma->pool, (uintptr_t)addr, size,
736 (void **) &ref);
737 percpu_ref_put(ref);
738}
739EXPORT_SYMBOL_GPL(pci_free_p2pmem);
740
741/**
742 * pci_p2pmem_virt_to_bus - return the PCI bus address for a given virtual
743 * address obtained with pci_alloc_p2pmem()
744 * @pdev: the device the memory was allocated from
745 * @addr: address of the memory that was allocated
746 */
747pci_bus_addr_t pci_p2pmem_virt_to_bus(struct pci_dev *pdev, void *addr)
748{
749 struct pci_p2pdma *p2pdma;
750
751 if (!addr)
752 return 0;
753
754 p2pdma = rcu_dereference_protected(pdev->p2pdma, 1);
755 if (!p2pdma)
756 return 0;
757
758 /*
759 * Note: when we added the memory to the pool we used the PCI
760 * bus address as the physical address. So gen_pool_virt_to_phys()
761 * actually returns the bus address despite the misleading name.
762 */
763 return gen_pool_virt_to_phys(p2pdma->pool, (unsigned long)addr);
764}
765EXPORT_SYMBOL_GPL(pci_p2pmem_virt_to_bus);
766
767/**
768 * pci_p2pmem_alloc_sgl - allocate peer-to-peer DMA memory in a scatterlist
769 * @pdev: the device to allocate memory from
770 * @nents: the number of SG entries in the list
771 * @length: number of bytes to allocate
772 *
773 * Return: %NULL on error or &struct scatterlist pointer and @nents on success
774 */
775struct scatterlist *pci_p2pmem_alloc_sgl(struct pci_dev *pdev,
776 unsigned int *nents, u32 length)
777{
778 struct scatterlist *sg;
779 void *addr;
780
781 sg = kmalloc(sizeof(*sg), GFP_KERNEL);
782 if (!sg)
783 return NULL;
784
785 sg_init_table(sg, 1);
786
787 addr = pci_alloc_p2pmem(pdev, length);
788 if (!addr)
789 goto out_free_sg;
790
791 sg_set_buf(sg, addr, length);
792 *nents = 1;
793 return sg;
794
795out_free_sg:
796 kfree(sg);
797 return NULL;
798}
799EXPORT_SYMBOL_GPL(pci_p2pmem_alloc_sgl);
800
801/**
802 * pci_p2pmem_free_sgl - free a scatterlist allocated by pci_p2pmem_alloc_sgl()
803 * @pdev: the device to allocate memory from
804 * @sgl: the allocated scatterlist
805 */
806void pci_p2pmem_free_sgl(struct pci_dev *pdev, struct scatterlist *sgl)
807{
808 struct scatterlist *sg;
809 int count;
810
811 for_each_sg(sgl, sg, INT_MAX, count) {
812 if (!sg)
813 break;
814
815 pci_free_p2pmem(pdev, sg_virt(sg), sg->length);
816 }
817 kfree(sgl);
818}
819EXPORT_SYMBOL_GPL(pci_p2pmem_free_sgl);
820
821/**
822 * pci_p2pmem_publish - publish the peer-to-peer DMA memory for use by
823 * other devices with pci_p2pmem_find()
824 * @pdev: the device with peer-to-peer DMA memory to publish
825 * @publish: set to true to publish the memory, false to unpublish it
826 *
827 * Published memory can be used by other PCI device drivers for
828 * peer-2-peer DMA operations. Non-published memory is reserved for
829 * exclusive use of the device driver that registers the peer-to-peer
830 * memory.
831 */
832void pci_p2pmem_publish(struct pci_dev *pdev, bool publish)
833{
834 struct pci_p2pdma *p2pdma;
835
836 rcu_read_lock();
837 p2pdma = rcu_dereference(pdev->p2pdma);
838 if (p2pdma)
839 p2pdma->p2pmem_published = publish;
840 rcu_read_unlock();
841}
842EXPORT_SYMBOL_GPL(pci_p2pmem_publish);
843
844static enum pci_p2pdma_map_type pci_p2pdma_map_type(struct dev_pagemap *pgmap,
845 struct device *dev)
846{
847 enum pci_p2pdma_map_type type = PCI_P2PDMA_MAP_NOT_SUPPORTED;
848 struct pci_dev *provider = to_p2p_pgmap(pgmap)->provider;
849 struct pci_dev *client;
850 struct pci_p2pdma *p2pdma;
851
852 if (!provider->p2pdma)
853 return PCI_P2PDMA_MAP_NOT_SUPPORTED;
854
855 if (!dev_is_pci(dev))
856 return PCI_P2PDMA_MAP_NOT_SUPPORTED;
857
858 client = to_pci_dev(dev);
859
860 rcu_read_lock();
861 p2pdma = rcu_dereference(provider->p2pdma);
862
863 if (p2pdma)
864 type = xa_to_value(xa_load(&p2pdma->map_types,
865 map_types_idx(client)));
866 rcu_read_unlock();
867 return type;
868}
869
870static int __pci_p2pdma_map_sg(struct pci_p2pdma_pagemap *p2p_pgmap,
871 struct device *dev, struct scatterlist *sg, int nents)
872{
873 struct scatterlist *s;
874 int i;
875
876 for_each_sg(sg, s, nents, i) {
877 s->dma_address = sg_phys(s) - p2p_pgmap->bus_offset;
878 sg_dma_len(s) = s->length;
879 }
880
881 return nents;
882}
883
884/**
885 * pci_p2pdma_map_sg_attrs - map a PCI peer-to-peer scatterlist for DMA
886 * @dev: device doing the DMA request
887 * @sg: scatter list to map
888 * @nents: elements in the scatterlist
889 * @dir: DMA direction
890 * @attrs: DMA attributes passed to dma_map_sg() (if called)
891 *
892 * Scatterlists mapped with this function should be unmapped using
893 * pci_p2pdma_unmap_sg_attrs().
894 *
895 * Returns the number of SG entries mapped or 0 on error.
896 */
897int pci_p2pdma_map_sg_attrs(struct device *dev, struct scatterlist *sg,
898 int nents, enum dma_data_direction dir, unsigned long attrs)
899{
900 struct pci_p2pdma_pagemap *p2p_pgmap =
901 to_p2p_pgmap(sg_page(sg)->pgmap);
902
903 switch (pci_p2pdma_map_type(sg_page(sg)->pgmap, dev)) {
904 case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
905 return dma_map_sg_attrs(dev, sg, nents, dir, attrs);
906 case PCI_P2PDMA_MAP_BUS_ADDR:
907 return __pci_p2pdma_map_sg(p2p_pgmap, dev, sg, nents);
908 default:
909 WARN_ON_ONCE(1);
910 return 0;
911 }
912}
913EXPORT_SYMBOL_GPL(pci_p2pdma_map_sg_attrs);
914
915/**
916 * pci_p2pdma_unmap_sg_attrs - unmap a PCI peer-to-peer scatterlist that was
917 * mapped with pci_p2pdma_map_sg()
918 * @dev: device doing the DMA request
919 * @sg: scatter list to map
920 * @nents: number of elements returned by pci_p2pdma_map_sg()
921 * @dir: DMA direction
922 * @attrs: DMA attributes passed to dma_unmap_sg() (if called)
923 */
924void pci_p2pdma_unmap_sg_attrs(struct device *dev, struct scatterlist *sg,
925 int nents, enum dma_data_direction dir, unsigned long attrs)
926{
927 enum pci_p2pdma_map_type map_type;
928
929 map_type = pci_p2pdma_map_type(sg_page(sg)->pgmap, dev);
930
931 if (map_type == PCI_P2PDMA_MAP_THRU_HOST_BRIDGE)
932 dma_unmap_sg_attrs(dev, sg, nents, dir, attrs);
933}
934EXPORT_SYMBOL_GPL(pci_p2pdma_unmap_sg_attrs);
935
936/**
937 * pci_p2pdma_enable_store - parse a configfs/sysfs attribute store
938 * to enable p2pdma
939 * @page: contents of the value to be stored
940 * @p2p_dev: returns the PCI device that was selected to be used
941 * (if one was specified in the stored value)
942 * @use_p2pdma: returns whether to enable p2pdma or not
943 *
944 * Parses an attribute value to decide whether to enable p2pdma.
945 * The value can select a PCI device (using its full BDF device
946 * name) or a boolean (in any format strtobool() accepts). A false
947 * value disables p2pdma, a true value expects the caller
948 * to automatically find a compatible device and specifying a PCI device
949 * expects the caller to use the specific provider.
950 *
951 * pci_p2pdma_enable_show() should be used as the show operation for
952 * the attribute.
953 *
954 * Returns 0 on success
955 */
956int pci_p2pdma_enable_store(const char *page, struct pci_dev **p2p_dev,
957 bool *use_p2pdma)
958{
959 struct device *dev;
960
961 dev = bus_find_device_by_name(&pci_bus_type, NULL, page);
962 if (dev) {
963 *use_p2pdma = true;
964 *p2p_dev = to_pci_dev(dev);
965
966 if (!pci_has_p2pmem(*p2p_dev)) {
967 pci_err(*p2p_dev,
968 "PCI device has no peer-to-peer memory: %s\n",
969 page);
970 pci_dev_put(*p2p_dev);
971 return -ENODEV;
972 }
973
974 return 0;
975 } else if ((page[0] == '0' || page[0] == '1') && !iscntrl(page[1])) {
976 /*
977 * If the user enters a PCI device that doesn't exist
978 * like "0000:01:00.1", we don't want strtobool to think
979 * it's a '0' when it's clearly not what the user wanted.
980 * So we require 0's and 1's to be exactly one character.
981 */
982 } else if (!strtobool(page, use_p2pdma)) {
983 return 0;
984 }
985
986 pr_err("No such PCI device: %.*s\n", (int)strcspn(page, "\n"), page);
987 return -ENODEV;
988}
989EXPORT_SYMBOL_GPL(pci_p2pdma_enable_store);
990
991/**
992 * pci_p2pdma_enable_show - show a configfs/sysfs attribute indicating
993 * whether p2pdma is enabled
994 * @page: contents of the stored value
995 * @p2p_dev: the selected p2p device (NULL if no device is selected)
996 * @use_p2pdma: whether p2pdma has been enabled
997 *
998 * Attributes that use pci_p2pdma_enable_store() should use this function
999 * to show the value of the attribute.
1000 *
1001 * Returns 0 on success
1002 */
1003ssize_t pci_p2pdma_enable_show(char *page, struct pci_dev *p2p_dev,
1004 bool use_p2pdma)
1005{
1006 if (!use_p2pdma)
1007 return sprintf(page, "0\n");
1008
1009 if (!p2p_dev)
1010 return sprintf(page, "1\n");
1011
1012 return sprintf(page, "%s\n", pci_name(p2p_dev));
1013}
1014EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show);