Linux Audio

Check our new training course

Loading...
v5.4
  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);
v6.2
   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/dma-map-ops.h>
  14#include <linux/pci-p2pdma.h>
  15#include <linux/module.h>
  16#include <linux/slab.h>
  17#include <linux/genalloc.h>
  18#include <linux/memremap.h>
  19#include <linux/percpu-refcount.h>
  20#include <linux/random.h>
  21#include <linux/seq_buf.h>
  22#include <linux/xarray.h>
  23
 
 
 
 
 
 
 
  24struct pci_p2pdma {
  25	struct gen_pool *pool;
  26	bool p2pmem_published;
  27	struct xarray map_types;
  28};
  29
  30struct pci_p2pdma_pagemap {
  31	struct dev_pagemap pgmap;
  32	struct pci_dev *provider;
  33	u64 bus_offset;
  34};
  35
  36static struct pci_p2pdma_pagemap *to_p2p_pgmap(struct dev_pagemap *pgmap)
  37{
  38	return container_of(pgmap, struct pci_p2pdma_pagemap, pgmap);
  39}
  40
  41static ssize_t size_show(struct device *dev, struct device_attribute *attr,
  42			 char *buf)
  43{
  44	struct pci_dev *pdev = to_pci_dev(dev);
  45	struct pci_p2pdma *p2pdma;
  46	size_t size = 0;
  47
  48	rcu_read_lock();
  49	p2pdma = rcu_dereference(pdev->p2pdma);
  50	if (p2pdma && p2pdma->pool)
  51		size = gen_pool_size(p2pdma->pool);
  52	rcu_read_unlock();
  53
  54	return sysfs_emit(buf, "%zd\n", size);
  55}
  56static DEVICE_ATTR_RO(size);
  57
  58static ssize_t available_show(struct device *dev, struct device_attribute *attr,
  59			      char *buf)
  60{
  61	struct pci_dev *pdev = to_pci_dev(dev);
  62	struct pci_p2pdma *p2pdma;
  63	size_t avail = 0;
  64
  65	rcu_read_lock();
  66	p2pdma = rcu_dereference(pdev->p2pdma);
  67	if (p2pdma && p2pdma->pool)
  68		avail = gen_pool_avail(p2pdma->pool);
  69	rcu_read_unlock();
  70
  71	return sysfs_emit(buf, "%zd\n", avail);
  72}
  73static DEVICE_ATTR_RO(available);
  74
  75static ssize_t published_show(struct device *dev, struct device_attribute *attr,
  76			      char *buf)
  77{
  78	struct pci_dev *pdev = to_pci_dev(dev);
  79	struct pci_p2pdma *p2pdma;
  80	bool published = false;
  81
  82	rcu_read_lock();
  83	p2pdma = rcu_dereference(pdev->p2pdma);
  84	if (p2pdma)
  85		published = p2pdma->p2pmem_published;
  86	rcu_read_unlock();
  87
  88	return sysfs_emit(buf, "%d\n", published);
 
  89}
  90static DEVICE_ATTR_RO(published);
  91
  92static int p2pmem_alloc_mmap(struct file *filp, struct kobject *kobj,
  93		struct bin_attribute *attr, struct vm_area_struct *vma)
  94{
  95	struct pci_dev *pdev = to_pci_dev(kobj_to_dev(kobj));
  96	size_t len = vma->vm_end - vma->vm_start;
  97	struct pci_p2pdma *p2pdma;
  98	struct percpu_ref *ref;
  99	unsigned long vaddr;
 100	void *kaddr;
 101	int ret;
 102
 103	/* prevent private mappings from being established */
 104	if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
 105		pci_info_ratelimited(pdev,
 106				     "%s: fail, attempted private mapping\n",
 107				     current->comm);
 108		return -EINVAL;
 109	}
 110
 111	if (vma->vm_pgoff) {
 112		pci_info_ratelimited(pdev,
 113				     "%s: fail, attempted mapping with non-zero offset\n",
 114				     current->comm);
 115		return -EINVAL;
 116	}
 117
 118	rcu_read_lock();
 119	p2pdma = rcu_dereference(pdev->p2pdma);
 120	if (!p2pdma) {
 121		ret = -ENODEV;
 122		goto out;
 123	}
 124
 125	kaddr = (void *)gen_pool_alloc_owner(p2pdma->pool, len, (void **)&ref);
 126	if (!kaddr) {
 127		ret = -ENOMEM;
 128		goto out;
 129	}
 130
 131	/*
 132	 * vm_insert_page() can sleep, so a reference is taken to mapping
 133	 * such that rcu_read_unlock() can be done before inserting the
 134	 * pages
 135	 */
 136	if (unlikely(!percpu_ref_tryget_live_rcu(ref))) {
 137		ret = -ENODEV;
 138		goto out_free_mem;
 139	}
 140	rcu_read_unlock();
 141
 142	for (vaddr = vma->vm_start; vaddr < vma->vm_end; vaddr += PAGE_SIZE) {
 143		ret = vm_insert_page(vma, vaddr, virt_to_page(kaddr));
 144		if (ret) {
 145			gen_pool_free(p2pdma->pool, (uintptr_t)kaddr, len);
 146			return ret;
 147		}
 148		percpu_ref_get(ref);
 149		put_page(virt_to_page(kaddr));
 150		kaddr += PAGE_SIZE;
 151		len -= PAGE_SIZE;
 152	}
 153
 154	percpu_ref_put(ref);
 155
 156	return 0;
 157out_free_mem:
 158	gen_pool_free(p2pdma->pool, (uintptr_t)kaddr, len);
 159out:
 160	rcu_read_unlock();
 161	return ret;
 162}
 163
 164static struct bin_attribute p2pmem_alloc_attr = {
 165	.attr = { .name = "allocate", .mode = 0660 },
 166	.mmap = p2pmem_alloc_mmap,
 167	/*
 168	 * Some places where we want to call mmap (ie. python) will check
 169	 * that the file size is greater than the mmap size before allowing
 170	 * the mmap to continue. To work around this, just set the size
 171	 * to be very large.
 172	 */
 173	.size = SZ_1T,
 174};
 175
 176static struct attribute *p2pmem_attrs[] = {
 177	&dev_attr_size.attr,
 178	&dev_attr_available.attr,
 179	&dev_attr_published.attr,
 180	NULL,
 181};
 182
 183static struct bin_attribute *p2pmem_bin_attrs[] = {
 184	&p2pmem_alloc_attr,
 185	NULL,
 186};
 187
 188static const struct attribute_group p2pmem_group = {
 189	.attrs = p2pmem_attrs,
 190	.bin_attrs = p2pmem_bin_attrs,
 191	.name = "p2pmem",
 192};
 193
 194static void p2pdma_page_free(struct page *page)
 195{
 196	struct pci_p2pdma_pagemap *pgmap = to_p2p_pgmap(page->pgmap);
 197	struct percpu_ref *ref;
 198
 199	gen_pool_free_owner(pgmap->provider->p2pdma->pool,
 200			    (uintptr_t)page_to_virt(page), PAGE_SIZE,
 201			    (void **)&ref);
 202	percpu_ref_put(ref);
 203}
 204
 205static const struct dev_pagemap_ops p2pdma_pgmap_ops = {
 206	.page_free = p2pdma_page_free,
 207};
 208
 209static void pci_p2pdma_release(void *data)
 210{
 211	struct pci_dev *pdev = data;
 212	struct pci_p2pdma *p2pdma;
 213
 214	p2pdma = rcu_dereference_protected(pdev->p2pdma, 1);
 215	if (!p2pdma)
 216		return;
 217
 218	/* Flush and disable pci_alloc_p2p_mem() */
 219	pdev->p2pdma = NULL;
 220	synchronize_rcu();
 221
 222	gen_pool_destroy(p2pdma->pool);
 223	sysfs_remove_group(&pdev->dev.kobj, &p2pmem_group);
 224	xa_destroy(&p2pdma->map_types);
 225}
 226
 227static int pci_p2pdma_setup(struct pci_dev *pdev)
 228{
 229	int error = -ENOMEM;
 230	struct pci_p2pdma *p2p;
 231
 232	p2p = devm_kzalloc(&pdev->dev, sizeof(*p2p), GFP_KERNEL);
 233	if (!p2p)
 234		return -ENOMEM;
 235
 236	xa_init(&p2p->map_types);
 237
 238	p2p->pool = gen_pool_create(PAGE_SHIFT, dev_to_node(&pdev->dev));
 239	if (!p2p->pool)
 240		goto out;
 241
 242	error = devm_add_action_or_reset(&pdev->dev, pci_p2pdma_release, pdev);
 243	if (error)
 244		goto out_pool_destroy;
 245
 
 
 246	error = sysfs_create_group(&pdev->dev.kobj, &p2pmem_group);
 247	if (error)
 248		goto out_pool_destroy;
 249
 250	rcu_assign_pointer(pdev->p2pdma, p2p);
 251	return 0;
 252
 253out_pool_destroy:
 
 254	gen_pool_destroy(p2p->pool);
 255out:
 256	devm_kfree(&pdev->dev, p2p);
 257	return error;
 258}
 259
 260static void pci_p2pdma_unmap_mappings(void *data)
 261{
 262	struct pci_dev *pdev = data;
 263
 264	/*
 265	 * Removing the alloc attribute from sysfs will call
 266	 * unmap_mapping_range() on the inode, teardown any existing userspace
 267	 * mappings and prevent new ones from being created.
 268	 */
 269	sysfs_remove_file_from_group(&pdev->dev.kobj, &p2pmem_alloc_attr.attr,
 270				     p2pmem_group.name);
 271}
 272
 273/**
 274 * pci_p2pdma_add_resource - add memory for use as p2p memory
 275 * @pdev: the device to add the memory to
 276 * @bar: PCI BAR to add
 277 * @size: size of the memory to add, may be zero to use the whole BAR
 278 * @offset: offset into the PCI BAR
 279 *
 280 * The memory will be given ZONE_DEVICE struct pages so that it may
 281 * be used with any DMA request.
 282 */
 283int pci_p2pdma_add_resource(struct pci_dev *pdev, int bar, size_t size,
 284			    u64 offset)
 285{
 286	struct pci_p2pdma_pagemap *p2p_pgmap;
 287	struct dev_pagemap *pgmap;
 288	struct pci_p2pdma *p2pdma;
 289	void *addr;
 290	int error;
 291
 292	if (!(pci_resource_flags(pdev, bar) & IORESOURCE_MEM))
 293		return -EINVAL;
 294
 295	if (offset >= pci_resource_len(pdev, bar))
 296		return -EINVAL;
 297
 298	if (!size)
 299		size = pci_resource_len(pdev, bar) - offset;
 300
 301	if (size + offset > pci_resource_len(pdev, bar))
 302		return -EINVAL;
 303
 304	if (!pdev->p2pdma) {
 305		error = pci_p2pdma_setup(pdev);
 306		if (error)
 307			return error;
 308	}
 309
 310	p2p_pgmap = devm_kzalloc(&pdev->dev, sizeof(*p2p_pgmap), GFP_KERNEL);
 311	if (!p2p_pgmap)
 312		return -ENOMEM;
 313
 314	pgmap = &p2p_pgmap->pgmap;
 315	pgmap->range.start = pci_resource_start(pdev, bar) + offset;
 316	pgmap->range.end = pgmap->range.start + size - 1;
 317	pgmap->nr_range = 1;
 318	pgmap->type = MEMORY_DEVICE_PCI_P2PDMA;
 319	pgmap->ops = &p2pdma_pgmap_ops;
 320
 321	p2p_pgmap->provider = pdev;
 322	p2p_pgmap->bus_offset = pci_bus_address(pdev, bar) -
 323		pci_resource_start(pdev, bar);
 324
 325	addr = devm_memremap_pages(&pdev->dev, pgmap);
 326	if (IS_ERR(addr)) {
 327		error = PTR_ERR(addr);
 328		goto pgmap_free;
 329	}
 330
 331	error = devm_add_action_or_reset(&pdev->dev, pci_p2pdma_unmap_mappings,
 332					 pdev);
 333	if (error)
 334		goto pages_free;
 335
 336	p2pdma = rcu_dereference_protected(pdev->p2pdma, 1);
 337	error = gen_pool_add_owner(p2pdma->pool, (unsigned long)addr,
 338			pci_bus_address(pdev, bar) + offset,
 339			range_len(&pgmap->range), dev_to_node(&pdev->dev),
 340			&pgmap->ref);
 341	if (error)
 342		goto pages_free;
 343
 344	pci_info(pdev, "added peer-to-peer DMA memory %#llx-%#llx\n",
 345		 pgmap->range.start, pgmap->range.end);
 346
 347	return 0;
 348
 349pages_free:
 350	devm_memunmap_pages(&pdev->dev, pgmap);
 351pgmap_free:
 352	devm_kfree(&pdev->dev, pgmap);
 353	return error;
 354}
 355EXPORT_SYMBOL_GPL(pci_p2pdma_add_resource);
 356
 357/*
 358 * Note this function returns the parent PCI device with a
 359 * reference taken. It is the caller's responsibility to drop
 360 * the reference.
 361 */
 362static struct pci_dev *find_parent_pci_dev(struct device *dev)
 363{
 364	struct device *parent;
 365
 366	dev = get_device(dev);
 367
 368	while (dev) {
 369		if (dev_is_pci(dev))
 370			return to_pci_dev(dev);
 371
 372		parent = get_device(dev->parent);
 373		put_device(dev);
 374		dev = parent;
 375	}
 376
 377	return NULL;
 378}
 379
 380/*
 381 * Check if a PCI bridge has its ACS redirection bits set to redirect P2P
 382 * TLPs upstream via ACS. Returns 1 if the packets will be redirected
 383 * upstream, 0 otherwise.
 384 */
 385static int pci_bridge_has_acs_redir(struct pci_dev *pdev)
 386{
 387	int pos;
 388	u16 ctrl;
 389
 390	pos = pdev->acs_cap;
 391	if (!pos)
 392		return 0;
 393
 394	pci_read_config_word(pdev, pos + PCI_ACS_CTRL, &ctrl);
 395
 396	if (ctrl & (PCI_ACS_RR | PCI_ACS_CR | PCI_ACS_EC))
 397		return 1;
 398
 399	return 0;
 400}
 401
 402static void seq_buf_print_bus_devfn(struct seq_buf *buf, struct pci_dev *pdev)
 403{
 404	if (!buf)
 405		return;
 406
 407	seq_buf_printf(buf, "%s;", pci_name(pdev));
 408}
 409
 410static bool cpu_supports_p2pdma(void)
 411{
 412#ifdef CONFIG_X86
 413	struct cpuinfo_x86 *c = &cpu_data(0);
 414
 415	/* Any AMD CPU whose family ID is Zen or newer supports p2pdma */
 416	if (c->x86_vendor == X86_VENDOR_AMD && c->x86 >= 0x17)
 417		return true;
 418#endif
 419
 420	return false;
 421}
 422
 423static const struct pci_p2pdma_whitelist_entry {
 424	unsigned short vendor;
 425	unsigned short device;
 426	enum {
 427		REQ_SAME_HOST_BRIDGE	= 1 << 0,
 428	} flags;
 429} pci_p2pdma_whitelist[] = {
 
 
 
 430	/* Intel Xeon E5/Core i7 */
 431	{PCI_VENDOR_ID_INTEL,	0x3c00, REQ_SAME_HOST_BRIDGE},
 432	{PCI_VENDOR_ID_INTEL,	0x3c01, REQ_SAME_HOST_BRIDGE},
 433	/* Intel Xeon E7 v3/Xeon E5 v3/Core i7 */
 434	{PCI_VENDOR_ID_INTEL,	0x2f00, REQ_SAME_HOST_BRIDGE},
 435	{PCI_VENDOR_ID_INTEL,	0x2f01, REQ_SAME_HOST_BRIDGE},
 436	/* Intel SkyLake-E */
 437	{PCI_VENDOR_ID_INTEL,	0x2030, 0},
 438	{PCI_VENDOR_ID_INTEL,	0x2031, 0},
 439	{PCI_VENDOR_ID_INTEL,	0x2032, 0},
 440	{PCI_VENDOR_ID_INTEL,	0x2033, 0},
 441	{PCI_VENDOR_ID_INTEL,	0x2020, 0},
 442	{PCI_VENDOR_ID_INTEL,	0x09a2, 0},
 443	{}
 444};
 445
 446/*
 447 * If the first device on host's root bus is either devfn 00.0 or a PCIe
 448 * Root Port, return it.  Otherwise return NULL.
 449 *
 450 * We often use a devfn 00.0 "host bridge" in the pci_p2pdma_whitelist[]
 451 * (though there is no PCI/PCIe requirement for such a device).  On some
 452 * platforms, e.g., Intel Skylake, there is no such host bridge device, and
 453 * pci_p2pdma_whitelist[] may contain a Root Port at any devfn.
 454 *
 455 * This function is similar to pci_get_slot(host->bus, 0), but it does
 456 * not take the pci_bus_sem lock since __host_bridge_whitelist() must not
 457 * sleep.
 458 *
 459 * For this to be safe, the caller should hold a reference to a device on the
 460 * bridge, which should ensure the host_bridge device will not be freed
 461 * or removed from the head of the devices list.
 462 */
 463static struct pci_dev *pci_host_bridge_dev(struct pci_host_bridge *host)
 464{
 465	struct pci_dev *root;
 466
 467	root = list_first_entry_or_null(&host->bus->devices,
 468					struct pci_dev, bus_list);
 469
 470	if (!root)
 471		return NULL;
 472
 473	if (root->devfn == PCI_DEVFN(0, 0))
 474		return root;
 475
 476	if (pci_pcie_type(root) == PCI_EXP_TYPE_ROOT_PORT)
 477		return root;
 478
 479	return NULL;
 480}
 481
 482static bool __host_bridge_whitelist(struct pci_host_bridge *host,
 483				    bool same_host_bridge, bool warn)
 484{
 485	struct pci_dev *root = pci_host_bridge_dev(host);
 486	const struct pci_p2pdma_whitelist_entry *entry;
 487	unsigned short vendor, device;
 488
 489	if (!root)
 490		return false;
 491
 492	vendor = root->vendor;
 493	device = root->device;
 
 494
 495	for (entry = pci_p2pdma_whitelist; entry->vendor; entry++) {
 496		if (vendor != entry->vendor || device != entry->device)
 497			continue;
 498		if (entry->flags & REQ_SAME_HOST_BRIDGE && !same_host_bridge)
 499			return false;
 500
 501		return true;
 502	}
 503
 504	if (warn)
 505		pci_warn(root, "Host bridge not in P2PDMA whitelist: %04x:%04x\n",
 506			 vendor, device);
 507
 508	return false;
 509}
 510
 511/*
 512 * If we can't find a common upstream bridge take a look at the root
 513 * complex and compare it to a whitelist of known good hardware.
 514 */
 515static bool host_bridge_whitelist(struct pci_dev *a, struct pci_dev *b,
 516				  bool warn)
 517{
 518	struct pci_host_bridge *host_a = pci_find_host_bridge(a->bus);
 519	struct pci_host_bridge *host_b = pci_find_host_bridge(b->bus);
 520
 521	if (host_a == host_b)
 522		return __host_bridge_whitelist(host_a, true, warn);
 523
 524	if (__host_bridge_whitelist(host_a, false, warn) &&
 525	    __host_bridge_whitelist(host_b, false, warn))
 526		return true;
 527
 528	return false;
 529}
 530
 531static unsigned long map_types_idx(struct pci_dev *client)
 532{
 533	return (pci_domain_nr(client->bus) << 16) |
 534		(client->bus->number << 8) | client->devfn;
 535}
 536
 537/*
 538 * Calculate the P2PDMA mapping type and distance between two PCI devices.
 539 *
 540 * If the two devices are the same PCI function, return
 541 * PCI_P2PDMA_MAP_BUS_ADDR and a distance of 0.
 542 *
 543 * If they are two functions of the same device, return
 544 * PCI_P2PDMA_MAP_BUS_ADDR and a distance of 2 (one hop up to the bridge,
 545 * then one hop back down to another function of the same device).
 546 *
 547 * In the case where two devices are connected to the same PCIe switch,
 548 * return a distance of 4. This corresponds to the following PCI tree:
 549 *
 550 *     -+  Root Port
 551 *      \+ Switch Upstream Port
 552 *       +-+ Switch Downstream Port 0
 553 *       + \- Device A
 554 *       \-+ Switch Downstream Port 1
 555 *         \- Device B
 556 *
 557 * The distance is 4 because we traverse from Device A to Downstream Port 0
 558 * to the common Switch Upstream Port, back down to Downstream Port 1 and
 559 * then to Device B. The mapping type returned depends on the ACS
 560 * redirection setting of the ports along the path.
 561 *
 562 * If ACS redirect is set on any port in the path, traffic between the
 563 * devices will go through the host bridge, so return
 564 * PCI_P2PDMA_MAP_THRU_HOST_BRIDGE; otherwise return
 565 * PCI_P2PDMA_MAP_BUS_ADDR.
 566 *
 567 * Any two devices that have a data path that goes through the host bridge
 568 * will consult a whitelist. If the host bridge is in the whitelist, return
 569 * PCI_P2PDMA_MAP_THRU_HOST_BRIDGE with the distance set to the number of
 570 * ports per above. If the device is not in the whitelist, return
 571 * PCI_P2PDMA_MAP_NOT_SUPPORTED.
 572 */
 573static enum pci_p2pdma_map_type
 574calc_map_type_and_dist(struct pci_dev *provider, struct pci_dev *client,
 575		int *dist, bool verbose)
 576{
 577	enum pci_p2pdma_map_type map_type = PCI_P2PDMA_MAP_THRU_HOST_BRIDGE;
 578	struct pci_dev *a = provider, *b = client, *bb;
 579	bool acs_redirects = false;
 580	struct pci_p2pdma *p2pdma;
 581	struct seq_buf acs_list;
 582	int acs_cnt = 0;
 583	int dist_a = 0;
 584	int dist_b = 0;
 585	char buf[128];
 586
 587	seq_buf_init(&acs_list, buf, sizeof(buf));
 
 588
 589	/*
 590	 * Note, we don't need to take references to devices returned by
 591	 * pci_upstream_bridge() seeing we hold a reference to a child
 592	 * device which will already hold a reference to the upstream bridge.
 593	 */
 
 594	while (a) {
 595		dist_b = 0;
 596
 597		if (pci_bridge_has_acs_redir(a)) {
 598			seq_buf_print_bus_devfn(&acs_list, a);
 599			acs_cnt++;
 600		}
 601
 602		bb = b;
 603
 604		while (bb) {
 605			if (a == bb)
 606				goto check_b_path_acs;
 607
 608			bb = pci_upstream_bridge(bb);
 609			dist_b++;
 610		}
 611
 612		a = pci_upstream_bridge(a);
 613		dist_a++;
 614	}
 615
 616	*dist = dist_a + dist_b;
 617	goto map_through_host_bridge;
 
 
 618
 619check_b_path_acs:
 620	bb = b;
 621
 622	while (bb) {
 623		if (a == bb)
 624			break;
 625
 626		if (pci_bridge_has_acs_redir(bb)) {
 627			seq_buf_print_bus_devfn(&acs_list, bb);
 628			acs_cnt++;
 629		}
 630
 631		bb = pci_upstream_bridge(bb);
 632	}
 633
 634	*dist = dist_a + dist_b;
 
 635
 636	if (!acs_cnt) {
 637		map_type = PCI_P2PDMA_MAP_BUS_ADDR;
 638		goto done;
 
 
 639	}
 640
 641	if (verbose) {
 642		acs_list.buffer[acs_list.len-1] = 0; /* drop final semicolon */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 643		pci_warn(client, "ACS redirect is set between the client and provider (%s)\n",
 644			 pci_name(provider));
 
 
 645		pci_warn(client, "to disable ACS redirect for this path, add the kernel parameter: pci=disable_acs_redir=%s\n",
 646			 acs_list.buffer);
 647	}
 648	acs_redirects = true;
 649
 650map_through_host_bridge:
 651	if (!cpu_supports_p2pdma() &&
 652	    !host_bridge_whitelist(provider, client, acs_redirects)) {
 653		if (verbose)
 654			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",
 655				 pci_name(provider));
 656		map_type = PCI_P2PDMA_MAP_NOT_SUPPORTED;
 657	}
 658done:
 659	rcu_read_lock();
 660	p2pdma = rcu_dereference(provider->p2pdma);
 661	if (p2pdma)
 662		xa_store(&p2pdma->map_types, map_types_idx(client),
 663			 xa_mk_value(map_type), GFP_KERNEL);
 664	rcu_read_unlock();
 665	return map_type;
 666}
 667
 668/**
 669 * pci_p2pdma_distance_many - Determine the cumulative distance between
 670 *	a p2pdma provider and the clients in use.
 671 * @provider: p2pdma provider to check against the client list
 672 * @clients: array of devices to check (NULL-terminated)
 673 * @num_clients: number of clients in the array
 674 * @verbose: if true, print warnings for devices when we return -1
 675 *
 676 * Returns -1 if any of the clients are not compatible, otherwise returns a
 677 * positive number where a lower number is the preferable choice. (If there's
 678 * one client that's the same as the provider it will return 0, which is best
 679 * choice).
 680 *
 681 * "compatible" means the provider and the clients are either all behind
 682 * the same PCI root port or the host bridges connected to each of the devices
 683 * are listed in the 'pci_p2pdma_whitelist'.
 684 */
 685int pci_p2pdma_distance_many(struct pci_dev *provider, struct device **clients,
 686			     int num_clients, bool verbose)
 687{
 688	enum pci_p2pdma_map_type map;
 689	bool not_supported = false;
 690	struct pci_dev *pci_client;
 691	int total_dist = 0;
 692	int i, distance;
 
 693
 694	if (num_clients == 0)
 695		return -1;
 696
 697	for (i = 0; i < num_clients; i++) {
 
 
 
 
 
 
 
 
 698		pci_client = find_parent_pci_dev(clients[i]);
 699		if (!pci_client) {
 700			if (verbose)
 701				dev_warn(clients[i],
 702					 "cannot be used for peer-to-peer DMA as it is not a PCI device\n");
 703			return -1;
 704		}
 705
 706		map = calc_map_type_and_dist(provider, pci_client, &distance,
 707					     verbose);
 
 
 
 
 708
 709		pci_dev_put(pci_client);
 710
 711		if (map == PCI_P2PDMA_MAP_NOT_SUPPORTED)
 712			not_supported = true;
 713
 714		if (not_supported && !verbose)
 715			break;
 716
 717		total_dist += distance;
 718	}
 719
 720	if (not_supported)
 721		return -1;
 722
 723	return total_dist;
 724}
 725EXPORT_SYMBOL_GPL(pci_p2pdma_distance_many);
 726
 727/**
 728 * pci_has_p2pmem - check if a given PCI device has published any p2pmem
 729 * @pdev: PCI device to check
 730 */
 731bool pci_has_p2pmem(struct pci_dev *pdev)
 732{
 733	struct pci_p2pdma *p2pdma;
 734	bool res;
 735
 736	rcu_read_lock();
 737	p2pdma = rcu_dereference(pdev->p2pdma);
 738	res = p2pdma && p2pdma->p2pmem_published;
 739	rcu_read_unlock();
 740
 741	return res;
 742}
 743EXPORT_SYMBOL_GPL(pci_has_p2pmem);
 744
 745/**
 746 * pci_p2pmem_find_many - find a peer-to-peer DMA memory device compatible with
 747 *	the specified list of clients and shortest distance (as determined
 748 *	by pci_p2pmem_dma())
 749 * @clients: array of devices to check (NULL-terminated)
 750 * @num_clients: number of client devices in the list
 751 *
 752 * If multiple devices are behind the same switch, the one "closest" to the
 753 * client devices in use will be chosen first. (So if one of the providers is
 754 * the same as one of the clients, that provider will be used ahead of any
 755 * other providers that are unrelated). If multiple providers are an equal
 756 * distance away, one will be chosen at random.
 757 *
 758 * Returns a pointer to the PCI device with a reference taken (use pci_dev_put
 759 * to return the reference) or NULL if no compatible device is found. The
 760 * found provider will also be assigned to the client list.
 761 */
 762struct pci_dev *pci_p2pmem_find_many(struct device **clients, int num_clients)
 763{
 764	struct pci_dev *pdev = NULL;
 765	int distance;
 766	int closest_distance = INT_MAX;
 767	struct pci_dev **closest_pdevs;
 768	int dev_cnt = 0;
 769	const int max_devs = PAGE_SIZE / sizeof(*closest_pdevs);
 770	int i;
 771
 772	closest_pdevs = kmalloc(PAGE_SIZE, GFP_KERNEL);
 773	if (!closest_pdevs)
 774		return NULL;
 775
 776	for_each_pci_dev(pdev) {
 777		if (!pci_has_p2pmem(pdev))
 778			continue;
 779
 780		distance = pci_p2pdma_distance_many(pdev, clients,
 781						    num_clients, false);
 782		if (distance < 0 || distance > closest_distance)
 783			continue;
 784
 785		if (distance == closest_distance && dev_cnt >= max_devs)
 786			continue;
 787
 788		if (distance < closest_distance) {
 789			for (i = 0; i < dev_cnt; i++)
 790				pci_dev_put(closest_pdevs[i]);
 791
 792			dev_cnt = 0;
 793			closest_distance = distance;
 794		}
 795
 796		closest_pdevs[dev_cnt++] = pci_dev_get(pdev);
 797	}
 798
 799	if (dev_cnt)
 800		pdev = pci_dev_get(closest_pdevs[get_random_u32_below(dev_cnt)]);
 801
 802	for (i = 0; i < dev_cnt; i++)
 803		pci_dev_put(closest_pdevs[i]);
 804
 805	kfree(closest_pdevs);
 806	return pdev;
 807}
 808EXPORT_SYMBOL_GPL(pci_p2pmem_find_many);
 809
 810/**
 811 * pci_alloc_p2pmem - allocate peer-to-peer DMA memory
 812 * @pdev: the device to allocate memory from
 813 * @size: number of bytes to allocate
 814 *
 815 * Returns the allocated memory or NULL on error.
 816 */
 817void *pci_alloc_p2pmem(struct pci_dev *pdev, size_t size)
 818{
 819	void *ret = NULL;
 820	struct percpu_ref *ref;
 821	struct pci_p2pdma *p2pdma;
 822
 823	/*
 824	 * Pairs with synchronize_rcu() in pci_p2pdma_release() to
 825	 * ensure pdev->p2pdma is non-NULL for the duration of the
 826	 * read-lock.
 827	 */
 828	rcu_read_lock();
 829	p2pdma = rcu_dereference(pdev->p2pdma);
 830	if (unlikely(!p2pdma))
 831		goto out;
 832
 833	ret = (void *)gen_pool_alloc_owner(p2pdma->pool, size, (void **) &ref);
 
 834	if (!ret)
 835		goto out;
 836
 837	if (unlikely(!percpu_ref_tryget_live_rcu(ref))) {
 838		gen_pool_free(p2pdma->pool, (unsigned long) ret, size);
 839		ret = NULL;
 840		goto out;
 841	}
 842out:
 843	rcu_read_unlock();
 844	return ret;
 845}
 846EXPORT_SYMBOL_GPL(pci_alloc_p2pmem);
 847
 848/**
 849 * pci_free_p2pmem - free peer-to-peer DMA memory
 850 * @pdev: the device the memory was allocated from
 851 * @addr: address of the memory that was allocated
 852 * @size: number of bytes that were allocated
 853 */
 854void pci_free_p2pmem(struct pci_dev *pdev, void *addr, size_t size)
 855{
 856	struct percpu_ref *ref;
 857	struct pci_p2pdma *p2pdma = rcu_dereference_protected(pdev->p2pdma, 1);
 858
 859	gen_pool_free_owner(p2pdma->pool, (uintptr_t)addr, size,
 860			(void **) &ref);
 861	percpu_ref_put(ref);
 862}
 863EXPORT_SYMBOL_GPL(pci_free_p2pmem);
 864
 865/**
 866 * pci_p2pmem_virt_to_bus - return the PCI bus address for a given virtual
 867 *	address obtained with pci_alloc_p2pmem()
 868 * @pdev: the device the memory was allocated from
 869 * @addr: address of the memory that was allocated
 870 */
 871pci_bus_addr_t pci_p2pmem_virt_to_bus(struct pci_dev *pdev, void *addr)
 872{
 873	struct pci_p2pdma *p2pdma;
 874
 875	if (!addr)
 876		return 0;
 877
 878	p2pdma = rcu_dereference_protected(pdev->p2pdma, 1);
 879	if (!p2pdma)
 880		return 0;
 881
 882	/*
 883	 * Note: when we added the memory to the pool we used the PCI
 884	 * bus address as the physical address. So gen_pool_virt_to_phys()
 885	 * actually returns the bus address despite the misleading name.
 886	 */
 887	return gen_pool_virt_to_phys(p2pdma->pool, (unsigned long)addr);
 888}
 889EXPORT_SYMBOL_GPL(pci_p2pmem_virt_to_bus);
 890
 891/**
 892 * pci_p2pmem_alloc_sgl - allocate peer-to-peer DMA memory in a scatterlist
 893 * @pdev: the device to allocate memory from
 894 * @nents: the number of SG entries in the list
 895 * @length: number of bytes to allocate
 896 *
 897 * Return: %NULL on error or &struct scatterlist pointer and @nents on success
 898 */
 899struct scatterlist *pci_p2pmem_alloc_sgl(struct pci_dev *pdev,
 900					 unsigned int *nents, u32 length)
 901{
 902	struct scatterlist *sg;
 903	void *addr;
 904
 905	sg = kmalloc(sizeof(*sg), GFP_KERNEL);
 906	if (!sg)
 907		return NULL;
 908
 909	sg_init_table(sg, 1);
 910
 911	addr = pci_alloc_p2pmem(pdev, length);
 912	if (!addr)
 913		goto out_free_sg;
 914
 915	sg_set_buf(sg, addr, length);
 916	*nents = 1;
 917	return sg;
 918
 919out_free_sg:
 920	kfree(sg);
 921	return NULL;
 922}
 923EXPORT_SYMBOL_GPL(pci_p2pmem_alloc_sgl);
 924
 925/**
 926 * pci_p2pmem_free_sgl - free a scatterlist allocated by pci_p2pmem_alloc_sgl()
 927 * @pdev: the device to allocate memory from
 928 * @sgl: the allocated scatterlist
 929 */
 930void pci_p2pmem_free_sgl(struct pci_dev *pdev, struct scatterlist *sgl)
 931{
 932	struct scatterlist *sg;
 933	int count;
 934
 935	for_each_sg(sgl, sg, INT_MAX, count) {
 936		if (!sg)
 937			break;
 938
 939		pci_free_p2pmem(pdev, sg_virt(sg), sg->length);
 940	}
 941	kfree(sgl);
 942}
 943EXPORT_SYMBOL_GPL(pci_p2pmem_free_sgl);
 944
 945/**
 946 * pci_p2pmem_publish - publish the peer-to-peer DMA memory for use by
 947 *	other devices with pci_p2pmem_find()
 948 * @pdev: the device with peer-to-peer DMA memory to publish
 949 * @publish: set to true to publish the memory, false to unpublish it
 950 *
 951 * Published memory can be used by other PCI device drivers for
 952 * peer-2-peer DMA operations. Non-published memory is reserved for
 953 * exclusive use of the device driver that registers the peer-to-peer
 954 * memory.
 955 */
 956void pci_p2pmem_publish(struct pci_dev *pdev, bool publish)
 957{
 958	struct pci_p2pdma *p2pdma;
 959
 960	rcu_read_lock();
 961	p2pdma = rcu_dereference(pdev->p2pdma);
 962	if (p2pdma)
 963		p2pdma->p2pmem_published = publish;
 964	rcu_read_unlock();
 965}
 966EXPORT_SYMBOL_GPL(pci_p2pmem_publish);
 967
 968static enum pci_p2pdma_map_type pci_p2pdma_map_type(struct dev_pagemap *pgmap,
 969						    struct device *dev)
 970{
 971	enum pci_p2pdma_map_type type = PCI_P2PDMA_MAP_NOT_SUPPORTED;
 972	struct pci_dev *provider = to_p2p_pgmap(pgmap)->provider;
 973	struct pci_dev *client;
 974	struct pci_p2pdma *p2pdma;
 975	int dist;
 976
 977	if (!provider->p2pdma)
 978		return PCI_P2PDMA_MAP_NOT_SUPPORTED;
 979
 980	if (!dev_is_pci(dev))
 981		return PCI_P2PDMA_MAP_NOT_SUPPORTED;
 
 982
 983	client = to_pci_dev(dev);
 
 
 
 
 
 984
 985	rcu_read_lock();
 986	p2pdma = rcu_dereference(provider->p2pdma);
 
 
 
 
 
 
 
 987
 988	if (p2pdma)
 989		type = xa_to_value(xa_load(&p2pdma->map_types,
 990					   map_types_idx(client)));
 991	rcu_read_unlock();
 992
 993	if (type == PCI_P2PDMA_MAP_UNKNOWN)
 994		return calc_map_type_and_dist(provider, client, &dist, true);
 
 995
 996	return type;
 997}
 998
 999/**
1000 * pci_p2pdma_map_segment - map an sg segment determining the mapping type
1001 * @state: State structure that should be declared outside of the for_each_sg()
1002 *	loop and initialized to zero.
1003 * @dev: DMA device that's doing the mapping operation
1004 * @sg: scatterlist segment to map
1005 *
1006 * This is a helper to be used by non-IOMMU dma_map_sg() implementations where
1007 * the sg segment is the same for the page_link and the dma_address.
1008 *
1009 * Attempt to map a single segment in an SGL with the PCI bus address.
1010 * The segment must point to a PCI P2PDMA page and thus must be
1011 * wrapped in a is_pci_p2pdma_page(sg_page(sg)) check.
1012 *
1013 * Returns the type of mapping used and maps the page if the type is
1014 * PCI_P2PDMA_MAP_BUS_ADDR.
 
 
1015 */
1016enum pci_p2pdma_map_type
1017pci_p2pdma_map_segment(struct pci_p2pdma_map_state *state, struct device *dev,
1018		       struct scatterlist *sg)
1019{
1020	if (state->pgmap != sg_page(sg)->pgmap) {
1021		state->pgmap = sg_page(sg)->pgmap;
1022		state->map = pci_p2pdma_map_type(state->pgmap, dev);
1023		state->bus_off = to_p2p_pgmap(state->pgmap)->bus_offset;
 
 
 
 
 
 
 
 
 
 
 
 
1024	}
 
 
1025
1026	if (state->map == PCI_P2PDMA_MAP_BUS_ADDR) {
1027		sg->dma_address = sg_phys(sg) + state->bus_off;
1028		sg_dma_len(sg) = sg->length;
1029		sg_dma_mark_bus_address(sg);
1030	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1031
1032	return state->map;
 
1033}
 
1034
1035/**
1036 * pci_p2pdma_enable_store - parse a configfs/sysfs attribute store
1037 *		to enable p2pdma
1038 * @page: contents of the value to be stored
1039 * @p2p_dev: returns the PCI device that was selected to be used
1040 *		(if one was specified in the stored value)
1041 * @use_p2pdma: returns whether to enable p2pdma or not
1042 *
1043 * Parses an attribute value to decide whether to enable p2pdma.
1044 * The value can select a PCI device (using its full BDF device
1045 * name) or a boolean (in any format kstrtobool() accepts). A false
1046 * value disables p2pdma, a true value expects the caller
1047 * to automatically find a compatible device and specifying a PCI device
1048 * expects the caller to use the specific provider.
1049 *
1050 * pci_p2pdma_enable_show() should be used as the show operation for
1051 * the attribute.
1052 *
1053 * Returns 0 on success
1054 */
1055int pci_p2pdma_enable_store(const char *page, struct pci_dev **p2p_dev,
1056			    bool *use_p2pdma)
1057{
1058	struct device *dev;
1059
1060	dev = bus_find_device_by_name(&pci_bus_type, NULL, page);
1061	if (dev) {
1062		*use_p2pdma = true;
1063		*p2p_dev = to_pci_dev(dev);
1064
1065		if (!pci_has_p2pmem(*p2p_dev)) {
1066			pci_err(*p2p_dev,
1067				"PCI device has no peer-to-peer memory: %s\n",
1068				page);
1069			pci_dev_put(*p2p_dev);
1070			return -ENODEV;
1071		}
1072
1073		return 0;
1074	} else if ((page[0] == '0' || page[0] == '1') && !iscntrl(page[1])) {
1075		/*
1076		 * If the user enters a PCI device that  doesn't exist
1077		 * like "0000:01:00.1", we don't want kstrtobool to think
1078		 * it's a '0' when it's clearly not what the user wanted.
1079		 * So we require 0's and 1's to be exactly one character.
1080		 */
1081	} else if (!kstrtobool(page, use_p2pdma)) {
1082		return 0;
1083	}
1084
1085	pr_err("No such PCI device: %.*s\n", (int)strcspn(page, "\n"), page);
1086	return -ENODEV;
1087}
1088EXPORT_SYMBOL_GPL(pci_p2pdma_enable_store);
1089
1090/**
1091 * pci_p2pdma_enable_show - show a configfs/sysfs attribute indicating
1092 *		whether p2pdma is enabled
1093 * @page: contents of the stored value
1094 * @p2p_dev: the selected p2p device (NULL if no device is selected)
1095 * @use_p2pdma: whether p2pdma has been enabled
1096 *
1097 * Attributes that use pci_p2pdma_enable_store() should use this function
1098 * to show the value of the attribute.
1099 *
1100 * Returns 0 on success
1101 */
1102ssize_t pci_p2pdma_enable_show(char *page, struct pci_dev *p2p_dev,
1103			       bool use_p2pdma)
1104{
1105	if (!use_p2pdma)
1106		return sprintf(page, "0\n");
1107
1108	if (!p2p_dev)
1109		return sprintf(page, "1\n");
1110
1111	return sprintf(page, "%s\n", pci_name(p2p_dev));
1112}
1113EXPORT_SYMBOL_GPL(pci_p2pdma_enable_show);