xref: /openbmc/linux/drivers/pci/p2pdma.c (revision 2f5947df)
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/iommu.h>
22 
23 struct pci_p2pdma {
24 	struct gen_pool *pool;
25 	bool p2pmem_published;
26 };
27 
28 struct p2pdma_pagemap {
29 	struct dev_pagemap pgmap;
30 	struct percpu_ref ref;
31 	struct completion ref_done;
32 };
33 
34 static ssize_t size_show(struct device *dev, struct device_attribute *attr,
35 			 char *buf)
36 {
37 	struct pci_dev *pdev = to_pci_dev(dev);
38 	size_t size = 0;
39 
40 	if (pdev->p2pdma->pool)
41 		size = gen_pool_size(pdev->p2pdma->pool);
42 
43 	return snprintf(buf, PAGE_SIZE, "%zd\n", size);
44 }
45 static DEVICE_ATTR_RO(size);
46 
47 static ssize_t available_show(struct device *dev, struct device_attribute *attr,
48 			      char *buf)
49 {
50 	struct pci_dev *pdev = to_pci_dev(dev);
51 	size_t avail = 0;
52 
53 	if (pdev->p2pdma->pool)
54 		avail = gen_pool_avail(pdev->p2pdma->pool);
55 
56 	return snprintf(buf, PAGE_SIZE, "%zd\n", avail);
57 }
58 static DEVICE_ATTR_RO(available);
59 
60 static ssize_t published_show(struct device *dev, struct device_attribute *attr,
61 			      char *buf)
62 {
63 	struct pci_dev *pdev = to_pci_dev(dev);
64 
65 	return snprintf(buf, PAGE_SIZE, "%d\n",
66 			pdev->p2pdma->p2pmem_published);
67 }
68 static DEVICE_ATTR_RO(published);
69 
70 static struct attribute *p2pmem_attrs[] = {
71 	&dev_attr_size.attr,
72 	&dev_attr_available.attr,
73 	&dev_attr_published.attr,
74 	NULL,
75 };
76 
77 static const struct attribute_group p2pmem_group = {
78 	.attrs = p2pmem_attrs,
79 	.name = "p2pmem",
80 };
81 
82 static struct p2pdma_pagemap *to_p2p_pgmap(struct percpu_ref *ref)
83 {
84 	return container_of(ref, struct p2pdma_pagemap, ref);
85 }
86 
87 static void pci_p2pdma_percpu_release(struct percpu_ref *ref)
88 {
89 	struct p2pdma_pagemap *p2p_pgmap = to_p2p_pgmap(ref);
90 
91 	complete(&p2p_pgmap->ref_done);
92 }
93 
94 static void pci_p2pdma_percpu_kill(struct percpu_ref *ref)
95 {
96 	percpu_ref_kill(ref);
97 }
98 
99 static void pci_p2pdma_percpu_cleanup(struct percpu_ref *ref)
100 {
101 	struct p2pdma_pagemap *p2p_pgmap = to_p2p_pgmap(ref);
102 
103 	wait_for_completion(&p2p_pgmap->ref_done);
104 	percpu_ref_exit(&p2p_pgmap->ref);
105 }
106 
107 static void pci_p2pdma_release(void *data)
108 {
109 	struct pci_dev *pdev = data;
110 	struct pci_p2pdma *p2pdma = pdev->p2pdma;
111 
112 	if (!p2pdma)
113 		return;
114 
115 	/* Flush and disable pci_alloc_p2p_mem() */
116 	pdev->p2pdma = NULL;
117 	synchronize_rcu();
118 
119 	gen_pool_destroy(p2pdma->pool);
120 	sysfs_remove_group(&pdev->dev.kobj, &p2pmem_group);
121 }
122 
123 static int pci_p2pdma_setup(struct pci_dev *pdev)
124 {
125 	int error = -ENOMEM;
126 	struct pci_p2pdma *p2p;
127 
128 	p2p = devm_kzalloc(&pdev->dev, sizeof(*p2p), GFP_KERNEL);
129 	if (!p2p)
130 		return -ENOMEM;
131 
132 	p2p->pool = gen_pool_create(PAGE_SHIFT, dev_to_node(&pdev->dev));
133 	if (!p2p->pool)
134 		goto out;
135 
136 	error = devm_add_action_or_reset(&pdev->dev, pci_p2pdma_release, pdev);
137 	if (error)
138 		goto out_pool_destroy;
139 
140 	pdev->p2pdma = p2p;
141 
142 	error = sysfs_create_group(&pdev->dev.kobj, &p2pmem_group);
143 	if (error)
144 		goto out_pool_destroy;
145 
146 	return 0;
147 
148 out_pool_destroy:
149 	pdev->p2pdma = NULL;
150 	gen_pool_destroy(p2p->pool);
151 out:
152 	devm_kfree(&pdev->dev, p2p);
153 	return error;
154 }
155 
156 /**
157  * pci_p2pdma_add_resource - add memory for use as p2p memory
158  * @pdev: the device to add the memory to
159  * @bar: PCI BAR to add
160  * @size: size of the memory to add, may be zero to use the whole BAR
161  * @offset: offset into the PCI BAR
162  *
163  * The memory will be given ZONE_DEVICE struct pages so that it may
164  * be used with any DMA request.
165  */
166 int pci_p2pdma_add_resource(struct pci_dev *pdev, int bar, size_t size,
167 			    u64 offset)
168 {
169 	struct p2pdma_pagemap *p2p_pgmap;
170 	struct dev_pagemap *pgmap;
171 	void *addr;
172 	int error;
173 
174 	if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM))
175 		return -EINVAL;
176 
177 	if (offset >= pci_resource_len(pdev, bar))
178 		return -EINVAL;
179 
180 	if (!size)
181 		size = pci_resource_len(pdev, bar) - offset;
182 
183 	if (size + offset > pci_resource_len(pdev, bar))
184 		return -EINVAL;
185 
186 	if (!pdev->p2pdma) {
187 		error = pci_p2pdma_setup(pdev);
188 		if (error)
189 			return error;
190 	}
191 
192 	p2p_pgmap = devm_kzalloc(&pdev->dev, sizeof(*p2p_pgmap), GFP_KERNEL);
193 	if (!p2p_pgmap)
194 		return -ENOMEM;
195 
196 	init_completion(&p2p_pgmap->ref_done);
197 	error = percpu_ref_init(&p2p_pgmap->ref,
198 			pci_p2pdma_percpu_release, 0, GFP_KERNEL);
199 	if (error)
200 		goto pgmap_free;
201 
202 	pgmap = &p2p_pgmap->pgmap;
203 
204 	pgmap->res.start = pci_resource_start(pdev, bar) + offset;
205 	pgmap->res.end = pgmap->res.start + size - 1;
206 	pgmap->res.flags = pci_resource_flags(pdev, bar);
207 	pgmap->ref = &p2p_pgmap->ref;
208 	pgmap->type = MEMORY_DEVICE_PCI_P2PDMA;
209 	pgmap->pci_p2pdma_bus_offset = pci_bus_address(pdev, bar) -
210 		pci_resource_start(pdev, bar);
211 	pgmap->kill = pci_p2pdma_percpu_kill;
212 	pgmap->cleanup = pci_p2pdma_percpu_cleanup;
213 
214 	addr = devm_memremap_pages(&pdev->dev, pgmap);
215 	if (IS_ERR(addr)) {
216 		error = PTR_ERR(addr);
217 		goto pgmap_free;
218 	}
219 
220 	error = gen_pool_add_owner(pdev->p2pdma->pool, (unsigned long)addr,
221 			pci_bus_address(pdev, bar) + offset,
222 			resource_size(&pgmap->res), dev_to_node(&pdev->dev),
223 			&p2p_pgmap->ref);
224 	if (error)
225 		goto pages_free;
226 
227 	pci_info(pdev, "added peer-to-peer DMA memory %pR\n",
228 		 &pgmap->res);
229 
230 	return 0;
231 
232 pages_free:
233 	devm_memunmap_pages(&pdev->dev, pgmap);
234 pgmap_free:
235 	devm_kfree(&pdev->dev, p2p_pgmap);
236 	return error;
237 }
238 EXPORT_SYMBOL_GPL(pci_p2pdma_add_resource);
239 
240 /*
241  * Note this function returns the parent PCI device with a
242  * reference taken. It is the caller's responsibily to drop
243  * the reference.
244  */
245 static struct pci_dev *find_parent_pci_dev(struct device *dev)
246 {
247 	struct device *parent;
248 
249 	dev = get_device(dev);
250 
251 	while (dev) {
252 		if (dev_is_pci(dev))
253 			return to_pci_dev(dev);
254 
255 		parent = get_device(dev->parent);
256 		put_device(dev);
257 		dev = parent;
258 	}
259 
260 	return NULL;
261 }
262 
263 /*
264  * Check if a PCI bridge has its ACS redirection bits set to redirect P2P
265  * TLPs upstream via ACS. Returns 1 if the packets will be redirected
266  * upstream, 0 otherwise.
267  */
268 static int pci_bridge_has_acs_redir(struct pci_dev *pdev)
269 {
270 	int pos;
271 	u16 ctrl;
272 
273 	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ACS);
274 	if (!pos)
275 		return 0;
276 
277 	pci_read_config_word(pdev, pos + PCI_ACS_CTRL, &ctrl);
278 
279 	if (ctrl & (PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_EC))
280 		return 1;
281 
282 	return 0;
283 }
284 
285 static void seq_buf_print_bus_devfn(struct seq_buf *buf, struct pci_dev *pdev)
286 {
287 	if (!buf)
288 		return;
289 
290 	seq_buf_printf(buf, "%s;", pci_name(pdev));
291 }
292 
293 /*
294  * If we can't find a common upstream bridge take a look at the root
295  * complex and compare it to a whitelist of known good hardware.
296  */
297 static bool root_complex_whitelist(struct pci_dev *dev)
298 {
299 	struct pci_host_bridge *host = pci_find_host_bridge(dev->bus);
300 	struct pci_dev *root = pci_get_slot(host->bus, PCI_DEVFN(0, 0));
301 	unsigned short vendor, device;
302 
303 	if (iommu_present(dev->dev.bus))
304 		return false;
305 
306 	if (!root)
307 		return false;
308 
309 	vendor = root->vendor;
310 	device = root->device;
311 	pci_dev_put(root);
312 
313 	/* AMD ZEN host bridges can do peer to peer */
314 	if (vendor == PCI_VENDOR_ID_AMD && device == 0x1450)
315 		return true;
316 
317 	return false;
318 }
319 
320 /*
321  * Find the distance through the nearest common upstream bridge between
322  * two PCI devices.
323  *
324  * If the two devices are the same device then 0 will be returned.
325  *
326  * If there are two virtual functions of the same device behind the same
327  * bridge port then 2 will be returned (one step down to the PCIe switch,
328  * then one step back to the same device).
329  *
330  * In the case where two devices are connected to the same PCIe switch, the
331  * value 4 will be returned. This corresponds to the following PCI tree:
332  *
333  *     -+  Root Port
334  *      \+ Switch Upstream Port
335  *       +-+ Switch Downstream Port
336  *       + \- Device A
337  *       \-+ Switch Downstream Port
338  *         \- Device B
339  *
340  * The distance is 4 because we traverse from Device A through the downstream
341  * port of the switch, to the common upstream port, back up to the second
342  * downstream port and then to Device B.
343  *
344  * Any two devices that don't have a common upstream bridge will return -1.
345  * In this way devices on separate PCIe root ports will be rejected, which
346  * is what we want for peer-to-peer seeing each PCIe root port defines a
347  * separate hierarchy domain and there's no way to determine whether the root
348  * complex supports forwarding between them.
349  *
350  * In the case where two devices are connected to different PCIe switches,
351  * this function will still return a positive distance as long as both
352  * switches eventually have a common upstream bridge. Note this covers
353  * the case of using multiple PCIe switches to achieve a desired level of
354  * fan-out from a root port. The exact distance will be a function of the
355  * number of switches between Device A and Device B.
356  *
357  * If a bridge which has any ACS redirection bits set is in the path
358  * then this functions will return -2. This is so we reject any
359  * cases where the TLPs are forwarded up into the root complex.
360  * In this case, a list of all infringing bridge addresses will be
361  * populated in acs_list (assuming it's non-null) for printk purposes.
362  */
363 static int upstream_bridge_distance(struct pci_dev *provider,
364 				    struct pci_dev *client,
365 				    struct seq_buf *acs_list)
366 {
367 	struct pci_dev *a = provider, *b = client, *bb;
368 	int dist_a = 0;
369 	int dist_b = 0;
370 	int acs_cnt = 0;
371 
372 	/*
373 	 * Note, we don't need to take references to devices returned by
374 	 * pci_upstream_bridge() seeing we hold a reference to a child
375 	 * device which will already hold a reference to the upstream bridge.
376 	 */
377 
378 	while (a) {
379 		dist_b = 0;
380 
381 		if (pci_bridge_has_acs_redir(a)) {
382 			seq_buf_print_bus_devfn(acs_list, a);
383 			acs_cnt++;
384 		}
385 
386 		bb = b;
387 
388 		while (bb) {
389 			if (a == bb)
390 				goto check_b_path_acs;
391 
392 			bb = pci_upstream_bridge(bb);
393 			dist_b++;
394 		}
395 
396 		a = pci_upstream_bridge(a);
397 		dist_a++;
398 	}
399 
400 	/*
401 	 * Allow the connection if both devices are on a whitelisted root
402 	 * complex, but add an arbitary large value to the distance.
403 	 */
404 	if (root_complex_whitelist(provider) &&
405 	    root_complex_whitelist(client))
406 		return 0x1000 + dist_a + dist_b;
407 
408 	return -1;
409 
410 check_b_path_acs:
411 	bb = b;
412 
413 	while (bb) {
414 		if (a == bb)
415 			break;
416 
417 		if (pci_bridge_has_acs_redir(bb)) {
418 			seq_buf_print_bus_devfn(acs_list, bb);
419 			acs_cnt++;
420 		}
421 
422 		bb = pci_upstream_bridge(bb);
423 	}
424 
425 	if (acs_cnt)
426 		return -2;
427 
428 	return dist_a + dist_b;
429 }
430 
431 static int upstream_bridge_distance_warn(struct pci_dev *provider,
432 					 struct pci_dev *client)
433 {
434 	struct seq_buf acs_list;
435 	int ret;
436 
437 	seq_buf_init(&acs_list, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
438 	if (!acs_list.buffer)
439 		return -ENOMEM;
440 
441 	ret = upstream_bridge_distance(provider, client, &acs_list);
442 	if (ret == -2) {
443 		pci_warn(client, "cannot be used for peer-to-peer DMA as ACS redirect is set between the client and provider (%s)\n",
444 			 pci_name(provider));
445 		/* Drop final semicolon */
446 		acs_list.buffer[acs_list.len-1] = 0;
447 		pci_warn(client, "to disable ACS redirect for this path, add the kernel parameter: pci=disable_acs_redir=%s\n",
448 			 acs_list.buffer);
449 
450 	} else if (ret < 0) {
451 		pci_warn(client, "cannot be used for peer-to-peer DMA as the client and provider (%s) do not share an upstream bridge\n",
452 			 pci_name(provider));
453 	}
454 
455 	kfree(acs_list.buffer);
456 
457 	return ret;
458 }
459 
460 /**
461  * pci_p2pdma_distance_many - Determive the cumulative distance between
462  *	a p2pdma provider and the clients in use.
463  * @provider: p2pdma provider to check against the client list
464  * @clients: array of devices to check (NULL-terminated)
465  * @num_clients: number of clients in the array
466  * @verbose: if true, print warnings for devices when we return -1
467  *
468  * Returns -1 if any of the clients are not compatible (behind the same
469  * root port as the provider), otherwise returns a positive number where
470  * a lower number is the preferable choice. (If there's one client
471  * that's the same as the provider it will return 0, which is best choice).
472  *
473  * For now, "compatible" means the provider and the clients are all behind
474  * the same PCI root port. This cuts out cases that may work but is safest
475  * for the user. Future work can expand this to white-list root complexes that
476  * can safely forward between each ports.
477  */
478 int pci_p2pdma_distance_many(struct pci_dev *provider, struct device **clients,
479 			     int num_clients, bool verbose)
480 {
481 	bool not_supported = false;
482 	struct pci_dev *pci_client;
483 	int distance = 0;
484 	int i, ret;
485 
486 	if (num_clients == 0)
487 		return -1;
488 
489 	for (i = 0; i < num_clients; i++) {
490 		pci_client = find_parent_pci_dev(clients[i]);
491 		if (!pci_client) {
492 			if (verbose)
493 				dev_warn(clients[i],
494 					 "cannot be used for peer-to-peer DMA as it is not a PCI device\n");
495 			return -1;
496 		}
497 
498 		if (verbose)
499 			ret = upstream_bridge_distance_warn(provider,
500 							    pci_client);
501 		else
502 			ret = upstream_bridge_distance(provider, pci_client,
503 						       NULL);
504 
505 		pci_dev_put(pci_client);
506 
507 		if (ret < 0)
508 			not_supported = true;
509 
510 		if (not_supported && !verbose)
511 			break;
512 
513 		distance += ret;
514 	}
515 
516 	if (not_supported)
517 		return -1;
518 
519 	return distance;
520 }
521 EXPORT_SYMBOL_GPL(pci_p2pdma_distance_many);
522 
523 /**
524  * pci_has_p2pmem - check if a given PCI device has published any p2pmem
525  * @pdev: PCI device to check
526  */
527 bool pci_has_p2pmem(struct pci_dev *pdev)
528 {
529 	return pdev->p2pdma && pdev->p2pdma->p2pmem_published;
530 }
531 EXPORT_SYMBOL_GPL(pci_has_p2pmem);
532 
533 /**
534  * pci_p2pmem_find - find a peer-to-peer DMA memory device compatible with
535  *	the specified list of clients and shortest distance (as determined
536  *	by pci_p2pmem_dma())
537  * @clients: array of devices to check (NULL-terminated)
538  * @num_clients: number of client devices in the list
539  *
540  * If multiple devices are behind the same switch, the one "closest" to the
541  * client devices in use will be chosen first. (So if one of the providers is
542  * the same as one of the clients, that provider will be used ahead of any
543  * other providers that are unrelated). If multiple providers are an equal
544  * distance away, one will be chosen at random.
545  *
546  * Returns a pointer to the PCI device with a reference taken (use pci_dev_put
547  * to return the reference) or NULL if no compatible device is found. The
548  * found provider will also be assigned to the client list.
549  */
550 struct pci_dev *pci_p2pmem_find_many(struct device **clients, int num_clients)
551 {
552 	struct pci_dev *pdev = NULL;
553 	int distance;
554 	int closest_distance = INT_MAX;
555 	struct pci_dev **closest_pdevs;
556 	int dev_cnt = 0;
557 	const int max_devs = PAGE_SIZE / sizeof(*closest_pdevs);
558 	int i;
559 
560 	closest_pdevs = kmalloc(PAGE_SIZE, GFP_KERNEL);
561 	if (!closest_pdevs)
562 		return NULL;
563 
564 	while ((pdev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, pdev))) {
565 		if (!pci_has_p2pmem(pdev))
566 			continue;
567 
568 		distance = pci_p2pdma_distance_many(pdev, clients,
569 						    num_clients, false);
570 		if (distance < 0 || distance > closest_distance)
571 			continue;
572 
573 		if (distance == closest_distance && dev_cnt >= max_devs)
574 			continue;
575 
576 		if (distance < closest_distance) {
577 			for (i = 0; i < dev_cnt; i++)
578 				pci_dev_put(closest_pdevs[i]);
579 
580 			dev_cnt = 0;
581 			closest_distance = distance;
582 		}
583 
584 		closest_pdevs[dev_cnt++] = pci_dev_get(pdev);
585 	}
586 
587 	if (dev_cnt)
588 		pdev = pci_dev_get(closest_pdevs[prandom_u32_max(dev_cnt)]);
589 
590 	for (i = 0; i < dev_cnt; i++)
591 		pci_dev_put(closest_pdevs[i]);
592 
593 	kfree(closest_pdevs);
594 	return pdev;
595 }
596 EXPORT_SYMBOL_GPL(pci_p2pmem_find_many);
597 
598 /**
599  * pci_alloc_p2p_mem - allocate peer-to-peer DMA memory
600  * @pdev: the device to allocate memory from
601  * @size: number of bytes to allocate
602  *
603  * Returns the allocated memory or NULL on error.
604  */
605 void *pci_alloc_p2pmem(struct pci_dev *pdev, size_t size)
606 {
607 	void *ret = NULL;
608 	struct percpu_ref *ref;
609 
610 	/*
611 	 * Pairs with synchronize_rcu() in pci_p2pdma_release() to
612 	 * ensure pdev->p2pdma is non-NULL for the duration of the
613 	 * read-lock.
614 	 */
615 	rcu_read_lock();
616 	if (unlikely(!pdev->p2pdma))
617 		goto out;
618 
619 	ret = (void *)gen_pool_alloc_owner(pdev->p2pdma->pool, size,
620 			(void **) &ref);
621 	if (!ret)
622 		goto out;
623 
624 	if (unlikely(!percpu_ref_tryget_live(ref))) {
625 		gen_pool_free(pdev->p2pdma->pool, (unsigned long) ret, size);
626 		ret = NULL;
627 		goto out;
628 	}
629 out:
630 	rcu_read_unlock();
631 	return ret;
632 }
633 EXPORT_SYMBOL_GPL(pci_alloc_p2pmem);
634 
635 /**
636  * pci_free_p2pmem - free peer-to-peer DMA memory
637  * @pdev: the device the memory was allocated from
638  * @addr: address of the memory that was allocated
639  * @size: number of bytes that were allocated
640  */
641 void pci_free_p2pmem(struct pci_dev *pdev, void *addr, size_t size)
642 {
643 	struct percpu_ref *ref;
644 
645 	gen_pool_free_owner(pdev->p2pdma->pool, (uintptr_t)addr, size,
646 			(void **) &ref);
647 	percpu_ref_put(ref);
648 }
649 EXPORT_SYMBOL_GPL(pci_free_p2pmem);
650 
651 /**
652  * pci_virt_to_bus - return the PCI bus address for a given virtual
653  *	address obtained with pci_alloc_p2pmem()
654  * @pdev: the device the memory was allocated from
655  * @addr: address of the memory that was allocated
656  */
657 pci_bus_addr_t pci_p2pmem_virt_to_bus(struct pci_dev *pdev, void *addr)
658 {
659 	if (!addr)
660 		return 0;
661 	if (!pdev->p2pdma)
662 		return 0;
663 
664 	/*
665 	 * Note: when we added the memory to the pool we used the PCI
666 	 * bus address as the physical address. So gen_pool_virt_to_phys()
667 	 * actually returns the bus address despite the misleading name.
668 	 */
669 	return gen_pool_virt_to_phys(pdev->p2pdma->pool, (unsigned long)addr);
670 }
671 EXPORT_SYMBOL_GPL(pci_p2pmem_virt_to_bus);
672 
673 /**
674  * pci_p2pmem_alloc_sgl - allocate peer-to-peer DMA memory in a scatterlist
675  * @pdev: the device to allocate memory from
676  * @nents: the number of SG entries in the list
677  * @length: number of bytes to allocate
678  *
679  * Return: %NULL on error or &struct scatterlist pointer and @nents on success
680  */
681 struct scatterlist *pci_p2pmem_alloc_sgl(struct pci_dev *pdev,
682 					 unsigned int *nents, u32 length)
683 {
684 	struct scatterlist *sg;
685 	void *addr;
686 
687 	sg = kzalloc(sizeof(*sg), GFP_KERNEL);
688 	if (!sg)
689 		return NULL;
690 
691 	sg_init_table(sg, 1);
692 
693 	addr = pci_alloc_p2pmem(pdev, length);
694 	if (!addr)
695 		goto out_free_sg;
696 
697 	sg_set_buf(sg, addr, length);
698 	*nents = 1;
699 	return sg;
700 
701 out_free_sg:
702 	kfree(sg);
703 	return NULL;
704 }
705 EXPORT_SYMBOL_GPL(pci_p2pmem_alloc_sgl);
706 
707 /**
708  * pci_p2pmem_free_sgl - free a scatterlist allocated by pci_p2pmem_alloc_sgl()
709  * @pdev: the device to allocate memory from
710  * @sgl: the allocated scatterlist
711  */
712 void pci_p2pmem_free_sgl(struct pci_dev *pdev, struct scatterlist *sgl)
713 {
714 	struct scatterlist *sg;
715 	int count;
716 
717 	for_each_sg(sgl, sg, INT_MAX, count) {
718 		if (!sg)
719 			break;
720 
721 		pci_free_p2pmem(pdev, sg_virt(sg), sg->length);
722 	}
723 	kfree(sgl);
724 }
725 EXPORT_SYMBOL_GPL(pci_p2pmem_free_sgl);
726 
727 /**
728  * pci_p2pmem_publish - publish the peer-to-peer DMA memory for use by
729  *	other devices with pci_p2pmem_find()
730  * @pdev: the device with peer-to-peer DMA memory to publish
731  * @publish: set to true to publish the memory, false to unpublish it
732  *
733  * Published memory can be used by other PCI device drivers for
734  * peer-2-peer DMA operations. Non-published memory is reserved for
735  * exclusive use of the device driver that registers the peer-to-peer
736  * memory.
737  */
738 void pci_p2pmem_publish(struct pci_dev *pdev, bool publish)
739 {
740 	if (pdev->p2pdma)
741 		pdev->p2pdma->p2pmem_published = publish;
742 }
743 EXPORT_SYMBOL_GPL(pci_p2pmem_publish);
744 
745 /**
746  * pci_p2pdma_map_sg - map a PCI peer-to-peer scatterlist for DMA
747  * @dev: device doing the DMA request
748  * @sg: scatter list to map
749  * @nents: elements in the scatterlist
750  * @dir: DMA direction
751  *
752  * Scatterlists mapped with this function should not be unmapped in any way.
753  *
754  * Returns the number of SG entries mapped or 0 on error.
755  */
756 int pci_p2pdma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
757 		      enum dma_data_direction dir)
758 {
759 	struct dev_pagemap *pgmap;
760 	struct scatterlist *s;
761 	phys_addr_t paddr;
762 	int i;
763 
764 	/*
765 	 * p2pdma mappings are not compatible with devices that use
766 	 * dma_virt_ops. If the upper layers do the right thing
767 	 * this should never happen because it will be prevented
768 	 * by the check in pci_p2pdma_add_client()
769 	 */
770 	if (WARN_ON_ONCE(IS_ENABLED(CONFIG_DMA_VIRT_OPS) &&
771 			 dev->dma_ops == &dma_virt_ops))
772 		return 0;
773 
774 	for_each_sg(sg, s, nents, i) {
775 		pgmap = sg_page(s)->pgmap;
776 		paddr = sg_phys(s);
777 
778 		s->dma_address = paddr - pgmap->pci_p2pdma_bus_offset;
779 		sg_dma_len(s) = s->length;
780 	}
781 
782 	return nents;
783 }
784 EXPORT_SYMBOL_GPL(pci_p2pdma_map_sg);
785 
786 /**
787  * pci_p2pdma_enable_store - parse a configfs/sysfs attribute store
788  *		to enable p2pdma
789  * @page: contents of the value to be stored
790  * @p2p_dev: returns the PCI device that was selected to be used
791  *		(if one was specified in the stored value)
792  * @use_p2pdma: returns whether to enable p2pdma or not
793  *
794  * Parses an attribute value to decide whether to enable p2pdma.
795  * The value can select a PCI device (using its full BDF device
796  * name) or a boolean (in any format strtobool() accepts). A false
797  * value disables p2pdma, a true value expects the caller
798  * to automatically find a compatible device and specifying a PCI device
799  * expects the caller to use the specific provider.
800  *
801  * pci_p2pdma_enable_show() should be used as the show operation for
802  * the attribute.
803  *
804  * Returns 0 on success
805  */
806 int pci_p2pdma_enable_store(const char *page, struct pci_dev **p2p_dev,
807 			    bool *use_p2pdma)
808 {
809 	struct device *dev;
810 
811 	dev = bus_find_device_by_name(&pci_bus_type, NULL, page);
812 	if (dev) {
813 		*use_p2pdma = true;
814 		*p2p_dev = to_pci_dev(dev);
815 
816 		if (!pci_has_p2pmem(*p2p_dev)) {
817 			pci_err(*p2p_dev,
818 				"PCI device has no peer-to-peer memory: %s\n",
819 				page);
820 			pci_dev_put(*p2p_dev);
821 			return -ENODEV;
822 		}
823 
824 		return 0;
825 	} else if ((page[0] == '0' || page[0] == '1') && !iscntrl(page[1])) {
826 		/*
827 		 * If the user enters a PCI device that  doesn't exist
828 		 * like "0000:01:00.1", we don't want strtobool to think
829 		 * it's a '0' when it's clearly not what the user wanted.
830 		 * So we require 0's and 1's to be exactly one character.
831 		 */
832 	} else if (!strtobool(page, use_p2pdma)) {
833 		return 0;
834 	}
835 
836 	pr_err("No such PCI device: %.*s\n", (int)strcspn(page, "\n"), page);
837 	return -ENODEV;
838 }
839 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_store);
840 
841 /**
842  * pci_p2pdma_enable_show - show a configfs/sysfs attribute indicating
843  *		whether p2pdma is enabled
844  * @page: contents of the stored value
845  * @p2p_dev: the selected p2p device (NULL if no device is selected)
846  * @use_p2pdma: whether p2pdma has been enabled
847  *
848  * Attributes that use pci_p2pdma_enable_store() should use this function
849  * to show the value of the attribute.
850  *
851  * Returns 0 on success
852  */
853 ssize_t pci_p2pdma_enable_show(char *page, struct pci_dev *p2p_dev,
854 			       bool use_p2pdma)
855 {
856 	if (!use_p2pdma)
857 		return sprintf(page, "0\n");
858 
859 	if (!p2p_dev)
860 		return sprintf(page, "1\n");
861 
862 	return sprintf(page, "%s\n", pci_name(p2p_dev));
863 }
864 EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show);
865