xref: /openbmc/linux/drivers/infiniband/core/device.c (revision 4da722ca)
1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/module.h>
35 #include <linux/string.h>
36 #include <linux/errno.h>
37 #include <linux/kernel.h>
38 #include <linux/slab.h>
39 #include <linux/init.h>
40 #include <linux/mutex.h>
41 #include <linux/netdevice.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <rdma/rdma_netlink.h>
45 #include <rdma/ib_addr.h>
46 #include <rdma/ib_cache.h>
47 
48 #include "core_priv.h"
49 
50 MODULE_AUTHOR("Roland Dreier");
51 MODULE_DESCRIPTION("core kernel InfiniBand API");
52 MODULE_LICENSE("Dual BSD/GPL");
53 
54 struct ib_client_data {
55 	struct list_head  list;
56 	struct ib_client *client;
57 	void *            data;
58 	/* The device or client is going down. Do not call client or device
59 	 * callbacks other than remove(). */
60 	bool		  going_down;
61 };
62 
63 struct workqueue_struct *ib_comp_wq;
64 struct workqueue_struct *ib_wq;
65 EXPORT_SYMBOL_GPL(ib_wq);
66 
67 /* The device_list and client_list contain devices and clients after their
68  * registration has completed, and the devices and clients are removed
69  * during unregistration. */
70 static LIST_HEAD(device_list);
71 static LIST_HEAD(client_list);
72 
73 /*
74  * device_mutex and lists_rwsem protect access to both device_list and
75  * client_list.  device_mutex protects writer access by device and client
76  * registration / de-registration.  lists_rwsem protects reader access to
77  * these lists.  Iterators of these lists must lock it for read, while updates
78  * to the lists must be done with a write lock. A special case is when the
79  * device_mutex is locked. In this case locking the lists for read access is
80  * not necessary as the device_mutex implies it.
81  *
82  * lists_rwsem also protects access to the client data list.
83  */
84 static DEFINE_MUTEX(device_mutex);
85 static DECLARE_RWSEM(lists_rwsem);
86 
87 static int ib_security_change(struct notifier_block *nb, unsigned long event,
88 			      void *lsm_data);
89 static void ib_policy_change_task(struct work_struct *work);
90 static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
91 
92 static struct notifier_block ibdev_lsm_nb = {
93 	.notifier_call = ib_security_change,
94 };
95 
96 static int ib_device_check_mandatory(struct ib_device *device)
97 {
98 #define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device, x), #x }
99 	static const struct {
100 		size_t offset;
101 		char  *name;
102 	} mandatory_table[] = {
103 		IB_MANDATORY_FUNC(query_device),
104 		IB_MANDATORY_FUNC(query_port),
105 		IB_MANDATORY_FUNC(query_pkey),
106 		IB_MANDATORY_FUNC(query_gid),
107 		IB_MANDATORY_FUNC(alloc_pd),
108 		IB_MANDATORY_FUNC(dealloc_pd),
109 		IB_MANDATORY_FUNC(create_ah),
110 		IB_MANDATORY_FUNC(destroy_ah),
111 		IB_MANDATORY_FUNC(create_qp),
112 		IB_MANDATORY_FUNC(modify_qp),
113 		IB_MANDATORY_FUNC(destroy_qp),
114 		IB_MANDATORY_FUNC(post_send),
115 		IB_MANDATORY_FUNC(post_recv),
116 		IB_MANDATORY_FUNC(create_cq),
117 		IB_MANDATORY_FUNC(destroy_cq),
118 		IB_MANDATORY_FUNC(poll_cq),
119 		IB_MANDATORY_FUNC(req_notify_cq),
120 		IB_MANDATORY_FUNC(get_dma_mr),
121 		IB_MANDATORY_FUNC(dereg_mr),
122 		IB_MANDATORY_FUNC(get_port_immutable)
123 	};
124 	int i;
125 
126 	for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
127 		if (!*(void **) ((void *) device + mandatory_table[i].offset)) {
128 			pr_warn("Device %s is missing mandatory function %s\n",
129 				device->name, mandatory_table[i].name);
130 			return -EINVAL;
131 		}
132 	}
133 
134 	return 0;
135 }
136 
137 static struct ib_device *__ib_device_get_by_name(const char *name)
138 {
139 	struct ib_device *device;
140 
141 	list_for_each_entry(device, &device_list, core_list)
142 		if (!strncmp(name, device->name, IB_DEVICE_NAME_MAX))
143 			return device;
144 
145 	return NULL;
146 }
147 
148 
149 static int alloc_name(char *name)
150 {
151 	unsigned long *inuse;
152 	char buf[IB_DEVICE_NAME_MAX];
153 	struct ib_device *device;
154 	int i;
155 
156 	inuse = (unsigned long *) get_zeroed_page(GFP_KERNEL);
157 	if (!inuse)
158 		return -ENOMEM;
159 
160 	list_for_each_entry(device, &device_list, core_list) {
161 		if (!sscanf(device->name, name, &i))
162 			continue;
163 		if (i < 0 || i >= PAGE_SIZE * 8)
164 			continue;
165 		snprintf(buf, sizeof buf, name, i);
166 		if (!strncmp(buf, device->name, IB_DEVICE_NAME_MAX))
167 			set_bit(i, inuse);
168 	}
169 
170 	i = find_first_zero_bit(inuse, PAGE_SIZE * 8);
171 	free_page((unsigned long) inuse);
172 	snprintf(buf, sizeof buf, name, i);
173 
174 	if (__ib_device_get_by_name(buf))
175 		return -ENFILE;
176 
177 	strlcpy(name, buf, IB_DEVICE_NAME_MAX);
178 	return 0;
179 }
180 
181 static void ib_device_release(struct device *device)
182 {
183 	struct ib_device *dev = container_of(device, struct ib_device, dev);
184 
185 	WARN_ON(dev->reg_state == IB_DEV_REGISTERED);
186 	if (dev->reg_state == IB_DEV_UNREGISTERED) {
187 		/*
188 		 * In IB_DEV_UNINITIALIZED state, cache or port table
189 		 * is not even created. Free cache and port table only when
190 		 * device reaches UNREGISTERED state.
191 		 */
192 		ib_cache_release_one(dev);
193 		kfree(dev->port_immutable);
194 	}
195 	kfree(dev);
196 }
197 
198 static int ib_device_uevent(struct device *device,
199 			    struct kobj_uevent_env *env)
200 {
201 	struct ib_device *dev = container_of(device, struct ib_device, dev);
202 
203 	if (add_uevent_var(env, "NAME=%s", dev->name))
204 		return -ENOMEM;
205 
206 	/*
207 	 * It would be nice to pass the node GUID with the event...
208 	 */
209 
210 	return 0;
211 }
212 
213 static struct class ib_class = {
214 	.name    = "infiniband",
215 	.dev_release = ib_device_release,
216 	.dev_uevent = ib_device_uevent,
217 };
218 
219 /**
220  * ib_alloc_device - allocate an IB device struct
221  * @size:size of structure to allocate
222  *
223  * Low-level drivers should use ib_alloc_device() to allocate &struct
224  * ib_device.  @size is the size of the structure to be allocated,
225  * including any private data used by the low-level driver.
226  * ib_dealloc_device() must be used to free structures allocated with
227  * ib_alloc_device().
228  */
229 struct ib_device *ib_alloc_device(size_t size)
230 {
231 	struct ib_device *device;
232 
233 	if (WARN_ON(size < sizeof(struct ib_device)))
234 		return NULL;
235 
236 	device = kzalloc(size, GFP_KERNEL);
237 	if (!device)
238 		return NULL;
239 
240 	device->dev.class = &ib_class;
241 	device_initialize(&device->dev);
242 
243 	dev_set_drvdata(&device->dev, device);
244 
245 	INIT_LIST_HEAD(&device->event_handler_list);
246 	spin_lock_init(&device->event_handler_lock);
247 	spin_lock_init(&device->client_data_lock);
248 	INIT_LIST_HEAD(&device->client_data_list);
249 	INIT_LIST_HEAD(&device->port_list);
250 
251 	return device;
252 }
253 EXPORT_SYMBOL(ib_alloc_device);
254 
255 /**
256  * ib_dealloc_device - free an IB device struct
257  * @device:structure to free
258  *
259  * Free a structure allocated with ib_alloc_device().
260  */
261 void ib_dealloc_device(struct ib_device *device)
262 {
263 	WARN_ON(device->reg_state != IB_DEV_UNREGISTERED &&
264 		device->reg_state != IB_DEV_UNINITIALIZED);
265 	kobject_put(&device->dev.kobj);
266 }
267 EXPORT_SYMBOL(ib_dealloc_device);
268 
269 static int add_client_context(struct ib_device *device, struct ib_client *client)
270 {
271 	struct ib_client_data *context;
272 	unsigned long flags;
273 
274 	context = kmalloc(sizeof *context, GFP_KERNEL);
275 	if (!context)
276 		return -ENOMEM;
277 
278 	context->client = client;
279 	context->data   = NULL;
280 	context->going_down = false;
281 
282 	down_write(&lists_rwsem);
283 	spin_lock_irqsave(&device->client_data_lock, flags);
284 	list_add(&context->list, &device->client_data_list);
285 	spin_unlock_irqrestore(&device->client_data_lock, flags);
286 	up_write(&lists_rwsem);
287 
288 	return 0;
289 }
290 
291 static int verify_immutable(const struct ib_device *dev, u8 port)
292 {
293 	return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
294 			    rdma_max_mad_size(dev, port) != 0);
295 }
296 
297 static int read_port_immutable(struct ib_device *device)
298 {
299 	int ret;
300 	u8 start_port = rdma_start_port(device);
301 	u8 end_port = rdma_end_port(device);
302 	u8 port;
303 
304 	/**
305 	 * device->port_immutable is indexed directly by the port number to make
306 	 * access to this data as efficient as possible.
307 	 *
308 	 * Therefore port_immutable is declared as a 1 based array with
309 	 * potential empty slots at the beginning.
310 	 */
311 	device->port_immutable = kzalloc(sizeof(*device->port_immutable)
312 					 * (end_port + 1),
313 					 GFP_KERNEL);
314 	if (!device->port_immutable)
315 		return -ENOMEM;
316 
317 	for (port = start_port; port <= end_port; ++port) {
318 		ret = device->get_port_immutable(device, port,
319 						 &device->port_immutable[port]);
320 		if (ret)
321 			return ret;
322 
323 		if (verify_immutable(device, port))
324 			return -EINVAL;
325 	}
326 	return 0;
327 }
328 
329 void ib_get_device_fw_str(struct ib_device *dev, char *str, size_t str_len)
330 {
331 	if (dev->get_dev_fw_str)
332 		dev->get_dev_fw_str(dev, str, str_len);
333 	else
334 		str[0] = '\0';
335 }
336 EXPORT_SYMBOL(ib_get_device_fw_str);
337 
338 static int setup_port_pkey_list(struct ib_device *device)
339 {
340 	int i;
341 
342 	/**
343 	 * device->port_pkey_list is indexed directly by the port number,
344 	 * Therefore it is declared as a 1 based array with potential empty
345 	 * slots at the beginning.
346 	 */
347 	device->port_pkey_list = kcalloc(rdma_end_port(device) + 1,
348 					 sizeof(*device->port_pkey_list),
349 					 GFP_KERNEL);
350 
351 	if (!device->port_pkey_list)
352 		return -ENOMEM;
353 
354 	for (i = 0; i < (rdma_end_port(device) + 1); i++) {
355 		spin_lock_init(&device->port_pkey_list[i].list_lock);
356 		INIT_LIST_HEAD(&device->port_pkey_list[i].pkey_list);
357 	}
358 
359 	return 0;
360 }
361 
362 static void ib_policy_change_task(struct work_struct *work)
363 {
364 	struct ib_device *dev;
365 
366 	down_read(&lists_rwsem);
367 	list_for_each_entry(dev, &device_list, core_list) {
368 		int i;
369 
370 		for (i = rdma_start_port(dev); i <= rdma_end_port(dev); i++) {
371 			u64 sp;
372 			int ret = ib_get_cached_subnet_prefix(dev,
373 							      i,
374 							      &sp);
375 
376 			WARN_ONCE(ret,
377 				  "ib_get_cached_subnet_prefix err: %d, this should never happen here\n",
378 				  ret);
379 			if (!ret)
380 				ib_security_cache_change(dev, i, sp);
381 		}
382 	}
383 	up_read(&lists_rwsem);
384 }
385 
386 static int ib_security_change(struct notifier_block *nb, unsigned long event,
387 			      void *lsm_data)
388 {
389 	if (event != LSM_POLICY_CHANGE)
390 		return NOTIFY_DONE;
391 
392 	schedule_work(&ib_policy_change_work);
393 
394 	return NOTIFY_OK;
395 }
396 
397 /**
398  * ib_register_device - Register an IB device with IB core
399  * @device:Device to register
400  *
401  * Low-level drivers use ib_register_device() to register their
402  * devices with the IB core.  All registered clients will receive a
403  * callback for each device that is added. @device must be allocated
404  * with ib_alloc_device().
405  */
406 int ib_register_device(struct ib_device *device,
407 		       int (*port_callback)(struct ib_device *,
408 					    u8, struct kobject *))
409 {
410 	int ret;
411 	struct ib_client *client;
412 	struct ib_udata uhw = {.outlen = 0, .inlen = 0};
413 	struct device *parent = device->dev.parent;
414 
415 	WARN_ON_ONCE(!parent);
416 	WARN_ON_ONCE(device->dma_device);
417 	if (device->dev.dma_ops) {
418 		/*
419 		 * The caller provided custom DMA operations. Copy the
420 		 * DMA-related fields that are used by e.g. dma_alloc_coherent()
421 		 * into device->dev.
422 		 */
423 		device->dma_device = &device->dev;
424 		if (!device->dev.dma_mask)
425 			device->dev.dma_mask = parent->dma_mask;
426 		if (!device->dev.coherent_dma_mask)
427 			device->dev.coherent_dma_mask =
428 				parent->coherent_dma_mask;
429 	} else {
430 		/*
431 		 * The caller did not provide custom DMA operations. Use the
432 		 * DMA mapping operations of the parent device.
433 		 */
434 		device->dma_device = parent;
435 	}
436 
437 	mutex_lock(&device_mutex);
438 
439 	if (strchr(device->name, '%')) {
440 		ret = alloc_name(device->name);
441 		if (ret)
442 			goto out;
443 	}
444 
445 	if (ib_device_check_mandatory(device)) {
446 		ret = -EINVAL;
447 		goto out;
448 	}
449 
450 	ret = read_port_immutable(device);
451 	if (ret) {
452 		pr_warn("Couldn't create per port immutable data %s\n",
453 			device->name);
454 		goto out;
455 	}
456 
457 	ret = setup_port_pkey_list(device);
458 	if (ret) {
459 		pr_warn("Couldn't create per port_pkey_list\n");
460 		goto out;
461 	}
462 
463 	ret = ib_cache_setup_one(device);
464 	if (ret) {
465 		pr_warn("Couldn't set up InfiniBand P_Key/GID cache\n");
466 		goto port_cleanup;
467 	}
468 
469 	ret = ib_device_register_rdmacg(device);
470 	if (ret) {
471 		pr_warn("Couldn't register device with rdma cgroup\n");
472 		goto cache_cleanup;
473 	}
474 
475 	memset(&device->attrs, 0, sizeof(device->attrs));
476 	ret = device->query_device(device, &device->attrs, &uhw);
477 	if (ret) {
478 		pr_warn("Couldn't query the device attributes\n");
479 		goto cache_cleanup;
480 	}
481 
482 	ret = ib_device_register_sysfs(device, port_callback);
483 	if (ret) {
484 		pr_warn("Couldn't register device %s with driver model\n",
485 			device->name);
486 		goto cache_cleanup;
487 	}
488 
489 	device->reg_state = IB_DEV_REGISTERED;
490 
491 	list_for_each_entry(client, &client_list, list)
492 		if (client->add && !add_client_context(device, client))
493 			client->add(device);
494 
495 	down_write(&lists_rwsem);
496 	list_add_tail(&device->core_list, &device_list);
497 	up_write(&lists_rwsem);
498 	mutex_unlock(&device_mutex);
499 	return 0;
500 
501 cache_cleanup:
502 	ib_cache_cleanup_one(device);
503 	ib_cache_release_one(device);
504 port_cleanup:
505 	kfree(device->port_immutable);
506 out:
507 	mutex_unlock(&device_mutex);
508 	return ret;
509 }
510 EXPORT_SYMBOL(ib_register_device);
511 
512 /**
513  * ib_unregister_device - Unregister an IB device
514  * @device:Device to unregister
515  *
516  * Unregister an IB device.  All clients will receive a remove callback.
517  */
518 void ib_unregister_device(struct ib_device *device)
519 {
520 	struct ib_client_data *context, *tmp;
521 	unsigned long flags;
522 
523 	mutex_lock(&device_mutex);
524 
525 	down_write(&lists_rwsem);
526 	list_del(&device->core_list);
527 	spin_lock_irqsave(&device->client_data_lock, flags);
528 	list_for_each_entry_safe(context, tmp, &device->client_data_list, list)
529 		context->going_down = true;
530 	spin_unlock_irqrestore(&device->client_data_lock, flags);
531 	downgrade_write(&lists_rwsem);
532 
533 	list_for_each_entry_safe(context, tmp, &device->client_data_list,
534 				 list) {
535 		if (context->client->remove)
536 			context->client->remove(device, context->data);
537 	}
538 	up_read(&lists_rwsem);
539 
540 	mutex_unlock(&device_mutex);
541 
542 	ib_device_unregister_rdmacg(device);
543 	ib_device_unregister_sysfs(device);
544 	ib_cache_cleanup_one(device);
545 
546 	ib_security_destroy_port_pkey_list(device);
547 	kfree(device->port_pkey_list);
548 
549 	down_write(&lists_rwsem);
550 	spin_lock_irqsave(&device->client_data_lock, flags);
551 	list_for_each_entry_safe(context, tmp, &device->client_data_list, list)
552 		kfree(context);
553 	spin_unlock_irqrestore(&device->client_data_lock, flags);
554 	up_write(&lists_rwsem);
555 
556 	device->reg_state = IB_DEV_UNREGISTERED;
557 }
558 EXPORT_SYMBOL(ib_unregister_device);
559 
560 /**
561  * ib_register_client - Register an IB client
562  * @client:Client to register
563  *
564  * Upper level users of the IB drivers can use ib_register_client() to
565  * register callbacks for IB device addition and removal.  When an IB
566  * device is added, each registered client's add method will be called
567  * (in the order the clients were registered), and when a device is
568  * removed, each client's remove method will be called (in the reverse
569  * order that clients were registered).  In addition, when
570  * ib_register_client() is called, the client will receive an add
571  * callback for all devices already registered.
572  */
573 int ib_register_client(struct ib_client *client)
574 {
575 	struct ib_device *device;
576 
577 	mutex_lock(&device_mutex);
578 
579 	list_for_each_entry(device, &device_list, core_list)
580 		if (client->add && !add_client_context(device, client))
581 			client->add(device);
582 
583 	down_write(&lists_rwsem);
584 	list_add_tail(&client->list, &client_list);
585 	up_write(&lists_rwsem);
586 
587 	mutex_unlock(&device_mutex);
588 
589 	return 0;
590 }
591 EXPORT_SYMBOL(ib_register_client);
592 
593 /**
594  * ib_unregister_client - Unregister an IB client
595  * @client:Client to unregister
596  *
597  * Upper level users use ib_unregister_client() to remove their client
598  * registration.  When ib_unregister_client() is called, the client
599  * will receive a remove callback for each IB device still registered.
600  */
601 void ib_unregister_client(struct ib_client *client)
602 {
603 	struct ib_client_data *context, *tmp;
604 	struct ib_device *device;
605 	unsigned long flags;
606 
607 	mutex_lock(&device_mutex);
608 
609 	down_write(&lists_rwsem);
610 	list_del(&client->list);
611 	up_write(&lists_rwsem);
612 
613 	list_for_each_entry(device, &device_list, core_list) {
614 		struct ib_client_data *found_context = NULL;
615 
616 		down_write(&lists_rwsem);
617 		spin_lock_irqsave(&device->client_data_lock, flags);
618 		list_for_each_entry_safe(context, tmp, &device->client_data_list, list)
619 			if (context->client == client) {
620 				context->going_down = true;
621 				found_context = context;
622 				break;
623 			}
624 		spin_unlock_irqrestore(&device->client_data_lock, flags);
625 		up_write(&lists_rwsem);
626 
627 		if (client->remove)
628 			client->remove(device, found_context ?
629 					       found_context->data : NULL);
630 
631 		if (!found_context) {
632 			pr_warn("No client context found for %s/%s\n",
633 				device->name, client->name);
634 			continue;
635 		}
636 
637 		down_write(&lists_rwsem);
638 		spin_lock_irqsave(&device->client_data_lock, flags);
639 		list_del(&found_context->list);
640 		kfree(found_context);
641 		spin_unlock_irqrestore(&device->client_data_lock, flags);
642 		up_write(&lists_rwsem);
643 	}
644 
645 	mutex_unlock(&device_mutex);
646 }
647 EXPORT_SYMBOL(ib_unregister_client);
648 
649 /**
650  * ib_get_client_data - Get IB client context
651  * @device:Device to get context for
652  * @client:Client to get context for
653  *
654  * ib_get_client_data() returns client context set with
655  * ib_set_client_data().
656  */
657 void *ib_get_client_data(struct ib_device *device, struct ib_client *client)
658 {
659 	struct ib_client_data *context;
660 	void *ret = NULL;
661 	unsigned long flags;
662 
663 	spin_lock_irqsave(&device->client_data_lock, flags);
664 	list_for_each_entry(context, &device->client_data_list, list)
665 		if (context->client == client) {
666 			ret = context->data;
667 			break;
668 		}
669 	spin_unlock_irqrestore(&device->client_data_lock, flags);
670 
671 	return ret;
672 }
673 EXPORT_SYMBOL(ib_get_client_data);
674 
675 /**
676  * ib_set_client_data - Set IB client context
677  * @device:Device to set context for
678  * @client:Client to set context for
679  * @data:Context to set
680  *
681  * ib_set_client_data() sets client context that can be retrieved with
682  * ib_get_client_data().
683  */
684 void ib_set_client_data(struct ib_device *device, struct ib_client *client,
685 			void *data)
686 {
687 	struct ib_client_data *context;
688 	unsigned long flags;
689 
690 	spin_lock_irqsave(&device->client_data_lock, flags);
691 	list_for_each_entry(context, &device->client_data_list, list)
692 		if (context->client == client) {
693 			context->data = data;
694 			goto out;
695 		}
696 
697 	pr_warn("No client context found for %s/%s\n",
698 		device->name, client->name);
699 
700 out:
701 	spin_unlock_irqrestore(&device->client_data_lock, flags);
702 }
703 EXPORT_SYMBOL(ib_set_client_data);
704 
705 /**
706  * ib_register_event_handler - Register an IB event handler
707  * @event_handler:Handler to register
708  *
709  * ib_register_event_handler() registers an event handler that will be
710  * called back when asynchronous IB events occur (as defined in
711  * chapter 11 of the InfiniBand Architecture Specification).  This
712  * callback may occur in interrupt context.
713  */
714 int ib_register_event_handler  (struct ib_event_handler *event_handler)
715 {
716 	unsigned long flags;
717 
718 	spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
719 	list_add_tail(&event_handler->list,
720 		      &event_handler->device->event_handler_list);
721 	spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
722 
723 	return 0;
724 }
725 EXPORT_SYMBOL(ib_register_event_handler);
726 
727 /**
728  * ib_unregister_event_handler - Unregister an event handler
729  * @event_handler:Handler to unregister
730  *
731  * Unregister an event handler registered with
732  * ib_register_event_handler().
733  */
734 int ib_unregister_event_handler(struct ib_event_handler *event_handler)
735 {
736 	unsigned long flags;
737 
738 	spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
739 	list_del(&event_handler->list);
740 	spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
741 
742 	return 0;
743 }
744 EXPORT_SYMBOL(ib_unregister_event_handler);
745 
746 /**
747  * ib_dispatch_event - Dispatch an asynchronous event
748  * @event:Event to dispatch
749  *
750  * Low-level drivers must call ib_dispatch_event() to dispatch the
751  * event to all registered event handlers when an asynchronous event
752  * occurs.
753  */
754 void ib_dispatch_event(struct ib_event *event)
755 {
756 	unsigned long flags;
757 	struct ib_event_handler *handler;
758 
759 	spin_lock_irqsave(&event->device->event_handler_lock, flags);
760 
761 	list_for_each_entry(handler, &event->device->event_handler_list, list)
762 		handler->handler(handler, event);
763 
764 	spin_unlock_irqrestore(&event->device->event_handler_lock, flags);
765 }
766 EXPORT_SYMBOL(ib_dispatch_event);
767 
768 /**
769  * ib_query_port - Query IB port attributes
770  * @device:Device to query
771  * @port_num:Port number to query
772  * @port_attr:Port attributes
773  *
774  * ib_query_port() returns the attributes of a port through the
775  * @port_attr pointer.
776  */
777 int ib_query_port(struct ib_device *device,
778 		  u8 port_num,
779 		  struct ib_port_attr *port_attr)
780 {
781 	union ib_gid gid;
782 	int err;
783 
784 	if (!rdma_is_port_valid(device, port_num))
785 		return -EINVAL;
786 
787 	memset(port_attr, 0, sizeof(*port_attr));
788 	err = device->query_port(device, port_num, port_attr);
789 	if (err || port_attr->subnet_prefix)
790 		return err;
791 
792 	if (rdma_port_get_link_layer(device, port_num) != IB_LINK_LAYER_INFINIBAND)
793 		return 0;
794 
795 	err = ib_query_gid(device, port_num, 0, &gid, NULL);
796 	if (err)
797 		return err;
798 
799 	port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
800 	return 0;
801 }
802 EXPORT_SYMBOL(ib_query_port);
803 
804 /**
805  * ib_query_gid - Get GID table entry
806  * @device:Device to query
807  * @port_num:Port number to query
808  * @index:GID table index to query
809  * @gid:Returned GID
810  * @attr: Returned GID attributes related to this GID index (only in RoCE).
811  *   NULL means ignore.
812  *
813  * ib_query_gid() fetches the specified GID table entry.
814  */
815 int ib_query_gid(struct ib_device *device,
816 		 u8 port_num, int index, union ib_gid *gid,
817 		 struct ib_gid_attr *attr)
818 {
819 	if (rdma_cap_roce_gid_table(device, port_num))
820 		return ib_get_cached_gid(device, port_num, index, gid, attr);
821 
822 	if (attr)
823 		return -EINVAL;
824 
825 	return device->query_gid(device, port_num, index, gid);
826 }
827 EXPORT_SYMBOL(ib_query_gid);
828 
829 /**
830  * ib_enum_roce_netdev - enumerate all RoCE ports
831  * @ib_dev : IB device we want to query
832  * @filter: Should we call the callback?
833  * @filter_cookie: Cookie passed to filter
834  * @cb: Callback to call for each found RoCE ports
835  * @cookie: Cookie passed back to the callback
836  *
837  * Enumerates all of the physical RoCE ports of ib_dev
838  * which are related to netdevice and calls callback() on each
839  * device for which filter() function returns non zero.
840  */
841 void ib_enum_roce_netdev(struct ib_device *ib_dev,
842 			 roce_netdev_filter filter,
843 			 void *filter_cookie,
844 			 roce_netdev_callback cb,
845 			 void *cookie)
846 {
847 	u8 port;
848 
849 	for (port = rdma_start_port(ib_dev); port <= rdma_end_port(ib_dev);
850 	     port++)
851 		if (rdma_protocol_roce(ib_dev, port)) {
852 			struct net_device *idev = NULL;
853 
854 			if (ib_dev->get_netdev)
855 				idev = ib_dev->get_netdev(ib_dev, port);
856 
857 			if (idev &&
858 			    idev->reg_state >= NETREG_UNREGISTERED) {
859 				dev_put(idev);
860 				idev = NULL;
861 			}
862 
863 			if (filter(ib_dev, port, idev, filter_cookie))
864 				cb(ib_dev, port, idev, cookie);
865 
866 			if (idev)
867 				dev_put(idev);
868 		}
869 }
870 
871 /**
872  * ib_enum_all_roce_netdevs - enumerate all RoCE devices
873  * @filter: Should we call the callback?
874  * @filter_cookie: Cookie passed to filter
875  * @cb: Callback to call for each found RoCE ports
876  * @cookie: Cookie passed back to the callback
877  *
878  * Enumerates all RoCE devices' physical ports which are related
879  * to netdevices and calls callback() on each device for which
880  * filter() function returns non zero.
881  */
882 void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
883 			      void *filter_cookie,
884 			      roce_netdev_callback cb,
885 			      void *cookie)
886 {
887 	struct ib_device *dev;
888 
889 	down_read(&lists_rwsem);
890 	list_for_each_entry(dev, &device_list, core_list)
891 		ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
892 	up_read(&lists_rwsem);
893 }
894 
895 /**
896  * ib_query_pkey - Get P_Key table entry
897  * @device:Device to query
898  * @port_num:Port number to query
899  * @index:P_Key table index to query
900  * @pkey:Returned P_Key
901  *
902  * ib_query_pkey() fetches the specified P_Key table entry.
903  */
904 int ib_query_pkey(struct ib_device *device,
905 		  u8 port_num, u16 index, u16 *pkey)
906 {
907 	return device->query_pkey(device, port_num, index, pkey);
908 }
909 EXPORT_SYMBOL(ib_query_pkey);
910 
911 /**
912  * ib_modify_device - Change IB device attributes
913  * @device:Device to modify
914  * @device_modify_mask:Mask of attributes to change
915  * @device_modify:New attribute values
916  *
917  * ib_modify_device() changes a device's attributes as specified by
918  * the @device_modify_mask and @device_modify structure.
919  */
920 int ib_modify_device(struct ib_device *device,
921 		     int device_modify_mask,
922 		     struct ib_device_modify *device_modify)
923 {
924 	if (!device->modify_device)
925 		return -ENOSYS;
926 
927 	return device->modify_device(device, device_modify_mask,
928 				     device_modify);
929 }
930 EXPORT_SYMBOL(ib_modify_device);
931 
932 /**
933  * ib_modify_port - Modifies the attributes for the specified port.
934  * @device: The device to modify.
935  * @port_num: The number of the port to modify.
936  * @port_modify_mask: Mask used to specify which attributes of the port
937  *   to change.
938  * @port_modify: New attribute values for the port.
939  *
940  * ib_modify_port() changes a port's attributes as specified by the
941  * @port_modify_mask and @port_modify structure.
942  */
943 int ib_modify_port(struct ib_device *device,
944 		   u8 port_num, int port_modify_mask,
945 		   struct ib_port_modify *port_modify)
946 {
947 	if (!device->modify_port)
948 		return -ENOSYS;
949 
950 	if (!rdma_is_port_valid(device, port_num))
951 		return -EINVAL;
952 
953 	return device->modify_port(device, port_num, port_modify_mask,
954 				   port_modify);
955 }
956 EXPORT_SYMBOL(ib_modify_port);
957 
958 /**
959  * ib_find_gid - Returns the port number and GID table index where
960  *   a specified GID value occurs.
961  * @device: The device to query.
962  * @gid: The GID value to search for.
963  * @gid_type: Type of GID.
964  * @ndev: The ndev related to the GID to search for.
965  * @port_num: The port number of the device where the GID value was found.
966  * @index: The index into the GID table where the GID was found.  This
967  *   parameter may be NULL.
968  */
969 int ib_find_gid(struct ib_device *device, union ib_gid *gid,
970 		enum ib_gid_type gid_type, struct net_device *ndev,
971 		u8 *port_num, u16 *index)
972 {
973 	union ib_gid tmp_gid;
974 	int ret, port, i;
975 
976 	for (port = rdma_start_port(device); port <= rdma_end_port(device); ++port) {
977 		if (rdma_cap_roce_gid_table(device, port)) {
978 			if (!ib_find_cached_gid_by_port(device, gid, gid_type, port,
979 							ndev, index)) {
980 				*port_num = port;
981 				return 0;
982 			}
983 		}
984 
985 		if (gid_type != IB_GID_TYPE_IB)
986 			continue;
987 
988 		for (i = 0; i < device->port_immutable[port].gid_tbl_len; ++i) {
989 			ret = ib_query_gid(device, port, i, &tmp_gid, NULL);
990 			if (ret)
991 				return ret;
992 			if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
993 				*port_num = port;
994 				if (index)
995 					*index = i;
996 				return 0;
997 			}
998 		}
999 	}
1000 
1001 	return -ENOENT;
1002 }
1003 EXPORT_SYMBOL(ib_find_gid);
1004 
1005 /**
1006  * ib_find_pkey - Returns the PKey table index where a specified
1007  *   PKey value occurs.
1008  * @device: The device to query.
1009  * @port_num: The port number of the device to search for the PKey.
1010  * @pkey: The PKey value to search for.
1011  * @index: The index into the PKey table where the PKey was found.
1012  */
1013 int ib_find_pkey(struct ib_device *device,
1014 		 u8 port_num, u16 pkey, u16 *index)
1015 {
1016 	int ret, i;
1017 	u16 tmp_pkey;
1018 	int partial_ix = -1;
1019 
1020 	for (i = 0; i < device->port_immutable[port_num].pkey_tbl_len; ++i) {
1021 		ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
1022 		if (ret)
1023 			return ret;
1024 		if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
1025 			/* if there is full-member pkey take it.*/
1026 			if (tmp_pkey & 0x8000) {
1027 				*index = i;
1028 				return 0;
1029 			}
1030 			if (partial_ix < 0)
1031 				partial_ix = i;
1032 		}
1033 	}
1034 
1035 	/*no full-member, if exists take the limited*/
1036 	if (partial_ix >= 0) {
1037 		*index = partial_ix;
1038 		return 0;
1039 	}
1040 	return -ENOENT;
1041 }
1042 EXPORT_SYMBOL(ib_find_pkey);
1043 
1044 /**
1045  * ib_get_net_dev_by_params() - Return the appropriate net_dev
1046  * for a received CM request
1047  * @dev:	An RDMA device on which the request has been received.
1048  * @port:	Port number on the RDMA device.
1049  * @pkey:	The Pkey the request came on.
1050  * @gid:	A GID that the net_dev uses to communicate.
1051  * @addr:	Contains the IP address that the request specified as its
1052  *		destination.
1053  */
1054 struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
1055 					    u8 port,
1056 					    u16 pkey,
1057 					    const union ib_gid *gid,
1058 					    const struct sockaddr *addr)
1059 {
1060 	struct net_device *net_dev = NULL;
1061 	struct ib_client_data *context;
1062 
1063 	if (!rdma_protocol_ib(dev, port))
1064 		return NULL;
1065 
1066 	down_read(&lists_rwsem);
1067 
1068 	list_for_each_entry(context, &dev->client_data_list, list) {
1069 		struct ib_client *client = context->client;
1070 
1071 		if (context->going_down)
1072 			continue;
1073 
1074 		if (client->get_net_dev_by_params) {
1075 			net_dev = client->get_net_dev_by_params(dev, port, pkey,
1076 								gid, addr,
1077 								context->data);
1078 			if (net_dev)
1079 				break;
1080 		}
1081 	}
1082 
1083 	up_read(&lists_rwsem);
1084 
1085 	return net_dev;
1086 }
1087 EXPORT_SYMBOL(ib_get_net_dev_by_params);
1088 
1089 static struct ibnl_client_cbs ibnl_ls_cb_table[] = {
1090 	[RDMA_NL_LS_OP_RESOLVE] = {
1091 		.dump = ib_nl_handle_resolve_resp,
1092 		.module = THIS_MODULE },
1093 	[RDMA_NL_LS_OP_SET_TIMEOUT] = {
1094 		.dump = ib_nl_handle_set_timeout,
1095 		.module = THIS_MODULE },
1096 	[RDMA_NL_LS_OP_IP_RESOLVE] = {
1097 		.dump = ib_nl_handle_ip_res_resp,
1098 		.module = THIS_MODULE },
1099 };
1100 
1101 static int ib_add_ibnl_clients(void)
1102 {
1103 	return ibnl_add_client(RDMA_NL_LS, ARRAY_SIZE(ibnl_ls_cb_table),
1104 			       ibnl_ls_cb_table);
1105 }
1106 
1107 static void ib_remove_ibnl_clients(void)
1108 {
1109 	ibnl_remove_client(RDMA_NL_LS);
1110 }
1111 
1112 static int __init ib_core_init(void)
1113 {
1114 	int ret;
1115 
1116 	ib_wq = alloc_workqueue("infiniband", 0, 0);
1117 	if (!ib_wq)
1118 		return -ENOMEM;
1119 
1120 	ib_comp_wq = alloc_workqueue("ib-comp-wq",
1121 			WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
1122 	if (!ib_comp_wq) {
1123 		ret = -ENOMEM;
1124 		goto err;
1125 	}
1126 
1127 	ret = class_register(&ib_class);
1128 	if (ret) {
1129 		pr_warn("Couldn't create InfiniBand device class\n");
1130 		goto err_comp;
1131 	}
1132 
1133 	ret = ibnl_init();
1134 	if (ret) {
1135 		pr_warn("Couldn't init IB netlink interface\n");
1136 		goto err_sysfs;
1137 	}
1138 
1139 	ret = addr_init();
1140 	if (ret) {
1141 		pr_warn("Could't init IB address resolution\n");
1142 		goto err_ibnl;
1143 	}
1144 
1145 	ret = ib_mad_init();
1146 	if (ret) {
1147 		pr_warn("Couldn't init IB MAD\n");
1148 		goto err_addr;
1149 	}
1150 
1151 	ret = ib_sa_init();
1152 	if (ret) {
1153 		pr_warn("Couldn't init SA\n");
1154 		goto err_mad;
1155 	}
1156 
1157 	ret = ib_add_ibnl_clients();
1158 	if (ret) {
1159 		pr_warn("Couldn't register ibnl clients\n");
1160 		goto err_sa;
1161 	}
1162 
1163 	ret = register_lsm_notifier(&ibdev_lsm_nb);
1164 	if (ret) {
1165 		pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
1166 		goto err_ibnl_clients;
1167 	}
1168 
1169 	ib_cache_setup();
1170 
1171 	return 0;
1172 
1173 err_ibnl_clients:
1174 	ib_remove_ibnl_clients();
1175 err_sa:
1176 	ib_sa_cleanup();
1177 err_mad:
1178 	ib_mad_cleanup();
1179 err_addr:
1180 	addr_cleanup();
1181 err_ibnl:
1182 	ibnl_cleanup();
1183 err_sysfs:
1184 	class_unregister(&ib_class);
1185 err_comp:
1186 	destroy_workqueue(ib_comp_wq);
1187 err:
1188 	destroy_workqueue(ib_wq);
1189 	return ret;
1190 }
1191 
1192 static void __exit ib_core_cleanup(void)
1193 {
1194 	unregister_lsm_notifier(&ibdev_lsm_nb);
1195 	ib_cache_cleanup();
1196 	ib_remove_ibnl_clients();
1197 	ib_sa_cleanup();
1198 	ib_mad_cleanup();
1199 	addr_cleanup();
1200 	ibnl_cleanup();
1201 	class_unregister(&ib_class);
1202 	destroy_workqueue(ib_comp_wq);
1203 	/* Make sure that any pending umem accounting work is done. */
1204 	destroy_workqueue(ib_wq);
1205 }
1206 
1207 module_init(ib_core_init);
1208 module_exit(ib_core_cleanup);
1209