Loading...
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/netdevice.h>
41#include <net/net_namespace.h>
42#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/hashtable.h>
45#include <rdma/rdma_netlink.h>
46#include <rdma/ib_addr.h>
47#include <rdma/ib_cache.h>
48#include <rdma/rdma_counter.h>
49
50#include "core_priv.h"
51#include "restrack.h"
52
53MODULE_AUTHOR("Roland Dreier");
54MODULE_DESCRIPTION("core kernel InfiniBand API");
55MODULE_LICENSE("Dual BSD/GPL");
56
57struct workqueue_struct *ib_comp_wq;
58struct workqueue_struct *ib_comp_unbound_wq;
59struct workqueue_struct *ib_wq;
60EXPORT_SYMBOL_GPL(ib_wq);
61static struct workqueue_struct *ib_unreg_wq;
62
63/*
64 * Each of the three rwsem locks (devices, clients, client_data) protects the
65 * xarray of the same name. Specifically it allows the caller to assert that
66 * the MARK will/will not be changing under the lock, and for devices and
67 * clients, that the value in the xarray is still a valid pointer. Change of
68 * the MARK is linked to the object state, so holding the lock and testing the
69 * MARK also asserts that the contained object is in a certain state.
70 *
71 * This is used to build a two stage register/unregister flow where objects
72 * can continue to be in the xarray even though they are still in progress to
73 * register/unregister.
74 *
75 * The xarray itself provides additional locking, and restartable iteration,
76 * which is also relied on.
77 *
78 * Locks should not be nested, with the exception of client_data, which is
79 * allowed to nest under the read side of the other two locks.
80 *
81 * The devices_rwsem also protects the device name list, any change or
82 * assignment of device name must also hold the write side to guarantee unique
83 * names.
84 */
85
86/*
87 * devices contains devices that have had their names assigned. The
88 * devices may not be registered. Users that care about the registration
89 * status need to call ib_device_try_get() on the device to ensure it is
90 * registered, and keep it registered, for the required duration.
91 *
92 */
93static DEFINE_XARRAY_FLAGS(devices, XA_FLAGS_ALLOC);
94static DECLARE_RWSEM(devices_rwsem);
95#define DEVICE_REGISTERED XA_MARK_1
96
97static u32 highest_client_id;
98#define CLIENT_REGISTERED XA_MARK_1
99static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
100static DECLARE_RWSEM(clients_rwsem);
101
102static void ib_client_put(struct ib_client *client)
103{
104 if (refcount_dec_and_test(&client->uses))
105 complete(&client->uses_zero);
106}
107
108/*
109 * If client_data is registered then the corresponding client must also still
110 * be registered.
111 */
112#define CLIENT_DATA_REGISTERED XA_MARK_1
113
114unsigned int rdma_dev_net_id;
115
116/*
117 * A list of net namespaces is maintained in an xarray. This is necessary
118 * because we can't get the locking right using the existing net ns list. We
119 * would require a init_net callback after the list is updated.
120 */
121static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
122/*
123 * rwsem to protect accessing the rdma_nets xarray entries.
124 */
125static DECLARE_RWSEM(rdma_nets_rwsem);
126
127bool ib_devices_shared_netns = true;
128module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
129MODULE_PARM_DESC(netns_mode,
130 "Share device among net namespaces; default=1 (shared)");
131/**
132 * rdma_dev_access_netns() - Return whether an rdma device can be accessed
133 * from a specified net namespace or not.
134 * @dev: Pointer to rdma device which needs to be checked
135 * @net: Pointer to net namesapce for which access to be checked
136 *
137 * When the rdma device is in shared mode, it ignores the net namespace.
138 * When the rdma device is exclusive to a net namespace, rdma device net
139 * namespace is checked against the specified one.
140 */
141bool rdma_dev_access_netns(const struct ib_device *dev, const struct net *net)
142{
143 return (ib_devices_shared_netns ||
144 net_eq(read_pnet(&dev->coredev.rdma_net), net));
145}
146EXPORT_SYMBOL(rdma_dev_access_netns);
147
148/*
149 * xarray has this behavior where it won't iterate over NULL values stored in
150 * allocated arrays. So we need our own iterator to see all values stored in
151 * the array. This does the same thing as xa_for_each except that it also
152 * returns NULL valued entries if the array is allocating. Simplified to only
153 * work on simple xarrays.
154 */
155static void *xan_find_marked(struct xarray *xa, unsigned long *indexp,
156 xa_mark_t filter)
157{
158 XA_STATE(xas, xa, *indexp);
159 void *entry;
160
161 rcu_read_lock();
162 do {
163 entry = xas_find_marked(&xas, ULONG_MAX, filter);
164 if (xa_is_zero(entry))
165 break;
166 } while (xas_retry(&xas, entry));
167 rcu_read_unlock();
168
169 if (entry) {
170 *indexp = xas.xa_index;
171 if (xa_is_zero(entry))
172 return NULL;
173 return entry;
174 }
175 return XA_ERROR(-ENOENT);
176}
177#define xan_for_each_marked(xa, index, entry, filter) \
178 for (index = 0, entry = xan_find_marked(xa, &(index), filter); \
179 !xa_is_err(entry); \
180 (index)++, entry = xan_find_marked(xa, &(index), filter))
181
182/* RCU hash table mapping netdevice pointers to struct ib_port_data */
183static DEFINE_SPINLOCK(ndev_hash_lock);
184static DECLARE_HASHTABLE(ndev_hash, 5);
185
186static void free_netdevs(struct ib_device *ib_dev);
187static void ib_unregister_work(struct work_struct *work);
188static void __ib_unregister_device(struct ib_device *device);
189static int ib_security_change(struct notifier_block *nb, unsigned long event,
190 void *lsm_data);
191static void ib_policy_change_task(struct work_struct *work);
192static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
193
194static void __ibdev_printk(const char *level, const struct ib_device *ibdev,
195 struct va_format *vaf)
196{
197 if (ibdev && ibdev->dev.parent)
198 dev_printk_emit(level[1] - '0',
199 ibdev->dev.parent,
200 "%s %s %s: %pV",
201 dev_driver_string(ibdev->dev.parent),
202 dev_name(ibdev->dev.parent),
203 dev_name(&ibdev->dev),
204 vaf);
205 else if (ibdev)
206 printk("%s%s: %pV",
207 level, dev_name(&ibdev->dev), vaf);
208 else
209 printk("%s(NULL ib_device): %pV", level, vaf);
210}
211
212void ibdev_printk(const char *level, const struct ib_device *ibdev,
213 const char *format, ...)
214{
215 struct va_format vaf;
216 va_list args;
217
218 va_start(args, format);
219
220 vaf.fmt = format;
221 vaf.va = &args;
222
223 __ibdev_printk(level, ibdev, &vaf);
224
225 va_end(args);
226}
227EXPORT_SYMBOL(ibdev_printk);
228
229#define define_ibdev_printk_level(func, level) \
230void func(const struct ib_device *ibdev, const char *fmt, ...) \
231{ \
232 struct va_format vaf; \
233 va_list args; \
234 \
235 va_start(args, fmt); \
236 \
237 vaf.fmt = fmt; \
238 vaf.va = &args; \
239 \
240 __ibdev_printk(level, ibdev, &vaf); \
241 \
242 va_end(args); \
243} \
244EXPORT_SYMBOL(func);
245
246define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
247define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
248define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
249define_ibdev_printk_level(ibdev_err, KERN_ERR);
250define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
251define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
252define_ibdev_printk_level(ibdev_info, KERN_INFO);
253
254static struct notifier_block ibdev_lsm_nb = {
255 .notifier_call = ib_security_change,
256};
257
258static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
259 struct net *net);
260
261/* Pointer to the RCU head at the start of the ib_port_data array */
262struct ib_port_data_rcu {
263 struct rcu_head rcu_head;
264 struct ib_port_data pdata[];
265};
266
267static void ib_device_check_mandatory(struct ib_device *device)
268{
269#define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
270 static const struct {
271 size_t offset;
272 char *name;
273 } mandatory_table[] = {
274 IB_MANDATORY_FUNC(query_device),
275 IB_MANDATORY_FUNC(query_port),
276 IB_MANDATORY_FUNC(alloc_pd),
277 IB_MANDATORY_FUNC(dealloc_pd),
278 IB_MANDATORY_FUNC(create_qp),
279 IB_MANDATORY_FUNC(modify_qp),
280 IB_MANDATORY_FUNC(destroy_qp),
281 IB_MANDATORY_FUNC(post_send),
282 IB_MANDATORY_FUNC(post_recv),
283 IB_MANDATORY_FUNC(create_cq),
284 IB_MANDATORY_FUNC(destroy_cq),
285 IB_MANDATORY_FUNC(poll_cq),
286 IB_MANDATORY_FUNC(req_notify_cq),
287 IB_MANDATORY_FUNC(get_dma_mr),
288 IB_MANDATORY_FUNC(reg_user_mr),
289 IB_MANDATORY_FUNC(dereg_mr),
290 IB_MANDATORY_FUNC(get_port_immutable)
291 };
292 int i;
293
294 device->kverbs_provider = true;
295 for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
296 if (!*(void **) ((void *) &device->ops +
297 mandatory_table[i].offset)) {
298 device->kverbs_provider = false;
299 break;
300 }
301 }
302}
303
304/*
305 * Caller must perform ib_device_put() to return the device reference count
306 * when ib_device_get_by_index() returns valid device pointer.
307 */
308struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
309{
310 struct ib_device *device;
311
312 down_read(&devices_rwsem);
313 device = xa_load(&devices, index);
314 if (device) {
315 if (!rdma_dev_access_netns(device, net)) {
316 device = NULL;
317 goto out;
318 }
319
320 if (!ib_device_try_get(device))
321 device = NULL;
322 }
323out:
324 up_read(&devices_rwsem);
325 return device;
326}
327
328/**
329 * ib_device_put - Release IB device reference
330 * @device: device whose reference to be released
331 *
332 * ib_device_put() releases reference to the IB device to allow it to be
333 * unregistered and eventually free.
334 */
335void ib_device_put(struct ib_device *device)
336{
337 if (refcount_dec_and_test(&device->refcount))
338 complete(&device->unreg_completion);
339}
340EXPORT_SYMBOL(ib_device_put);
341
342static struct ib_device *__ib_device_get_by_name(const char *name)
343{
344 struct ib_device *device;
345 unsigned long index;
346
347 xa_for_each (&devices, index, device)
348 if (!strcmp(name, dev_name(&device->dev)))
349 return device;
350
351 return NULL;
352}
353
354/**
355 * ib_device_get_by_name - Find an IB device by name
356 * @name: The name to look for
357 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
358 *
359 * Find and hold an ib_device by its name. The caller must call
360 * ib_device_put() on the returned pointer.
361 */
362struct ib_device *ib_device_get_by_name(const char *name,
363 enum rdma_driver_id driver_id)
364{
365 struct ib_device *device;
366
367 down_read(&devices_rwsem);
368 device = __ib_device_get_by_name(name);
369 if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
370 device->ops.driver_id != driver_id)
371 device = NULL;
372
373 if (device) {
374 if (!ib_device_try_get(device))
375 device = NULL;
376 }
377 up_read(&devices_rwsem);
378 return device;
379}
380EXPORT_SYMBOL(ib_device_get_by_name);
381
382static int rename_compat_devs(struct ib_device *device)
383{
384 struct ib_core_device *cdev;
385 unsigned long index;
386 int ret = 0;
387
388 mutex_lock(&device->compat_devs_mutex);
389 xa_for_each (&device->compat_devs, index, cdev) {
390 ret = device_rename(&cdev->dev, dev_name(&device->dev));
391 if (ret) {
392 dev_warn(&cdev->dev,
393 "Fail to rename compatdev to new name %s\n",
394 dev_name(&device->dev));
395 break;
396 }
397 }
398 mutex_unlock(&device->compat_devs_mutex);
399 return ret;
400}
401
402int ib_device_rename(struct ib_device *ibdev, const char *name)
403{
404 unsigned long index;
405 void *client_data;
406 int ret;
407
408 down_write(&devices_rwsem);
409 if (!strcmp(name, dev_name(&ibdev->dev))) {
410 up_write(&devices_rwsem);
411 return 0;
412 }
413
414 if (__ib_device_get_by_name(name)) {
415 up_write(&devices_rwsem);
416 return -EEXIST;
417 }
418
419 ret = device_rename(&ibdev->dev, name);
420 if (ret) {
421 up_write(&devices_rwsem);
422 return ret;
423 }
424
425 strscpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
426 ret = rename_compat_devs(ibdev);
427
428 downgrade_write(&devices_rwsem);
429 down_read(&ibdev->client_data_rwsem);
430 xan_for_each_marked(&ibdev->client_data, index, client_data,
431 CLIENT_DATA_REGISTERED) {
432 struct ib_client *client = xa_load(&clients, index);
433
434 if (!client || !client->rename)
435 continue;
436
437 client->rename(ibdev, client_data);
438 }
439 up_read(&ibdev->client_data_rwsem);
440 up_read(&devices_rwsem);
441 return 0;
442}
443
444int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
445{
446 if (use_dim > 1)
447 return -EINVAL;
448 ibdev->use_cq_dim = use_dim;
449
450 return 0;
451}
452
453static int alloc_name(struct ib_device *ibdev, const char *name)
454{
455 struct ib_device *device;
456 unsigned long index;
457 struct ida inuse;
458 int rc;
459 int i;
460
461 lockdep_assert_held_write(&devices_rwsem);
462 ida_init(&inuse);
463 xa_for_each (&devices, index, device) {
464 char buf[IB_DEVICE_NAME_MAX];
465
466 if (sscanf(dev_name(&device->dev), name, &i) != 1)
467 continue;
468 if (i < 0 || i >= INT_MAX)
469 continue;
470 snprintf(buf, sizeof buf, name, i);
471 if (strcmp(buf, dev_name(&device->dev)) != 0)
472 continue;
473
474 rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
475 if (rc < 0)
476 goto out;
477 }
478
479 rc = ida_alloc(&inuse, GFP_KERNEL);
480 if (rc < 0)
481 goto out;
482
483 rc = dev_set_name(&ibdev->dev, name, rc);
484out:
485 ida_destroy(&inuse);
486 return rc;
487}
488
489static void ib_device_release(struct device *device)
490{
491 struct ib_device *dev = container_of(device, struct ib_device, dev);
492
493 free_netdevs(dev);
494 WARN_ON(refcount_read(&dev->refcount));
495 if (dev->hw_stats_data)
496 ib_device_release_hw_stats(dev->hw_stats_data);
497 if (dev->port_data) {
498 ib_cache_release_one(dev);
499 ib_security_release_port_pkey_list(dev);
500 rdma_counter_release(dev);
501 kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
502 pdata[0]),
503 rcu_head);
504 }
505
506 mutex_destroy(&dev->unregistration_lock);
507 mutex_destroy(&dev->compat_devs_mutex);
508
509 xa_destroy(&dev->compat_devs);
510 xa_destroy(&dev->client_data);
511 kfree_rcu(dev, rcu_head);
512}
513
514static int ib_device_uevent(const struct device *device,
515 struct kobj_uevent_env *env)
516{
517 if (add_uevent_var(env, "NAME=%s", dev_name(device)))
518 return -ENOMEM;
519
520 /*
521 * It would be nice to pass the node GUID with the event...
522 */
523
524 return 0;
525}
526
527static const void *net_namespace(const struct device *d)
528{
529 const struct ib_core_device *coredev =
530 container_of(d, struct ib_core_device, dev);
531
532 return read_pnet(&coredev->rdma_net);
533}
534
535static struct class ib_class = {
536 .name = "infiniband",
537 .dev_release = ib_device_release,
538 .dev_uevent = ib_device_uevent,
539 .ns_type = &net_ns_type_operations,
540 .namespace = net_namespace,
541};
542
543static void rdma_init_coredev(struct ib_core_device *coredev,
544 struct ib_device *dev, struct net *net)
545{
546 /* This BUILD_BUG_ON is intended to catch layout change
547 * of union of ib_core_device and device.
548 * dev must be the first element as ib_core and providers
549 * driver uses it. Adding anything in ib_core_device before
550 * device will break this assumption.
551 */
552 BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
553 offsetof(struct ib_device, dev));
554
555 coredev->dev.class = &ib_class;
556 coredev->dev.groups = dev->groups;
557 device_initialize(&coredev->dev);
558 coredev->owner = dev;
559 INIT_LIST_HEAD(&coredev->port_list);
560 write_pnet(&coredev->rdma_net, net);
561}
562
563/**
564 * _ib_alloc_device - allocate an IB device struct
565 * @size:size of structure to allocate
566 *
567 * Low-level drivers should use ib_alloc_device() to allocate &struct
568 * ib_device. @size is the size of the structure to be allocated,
569 * including any private data used by the low-level driver.
570 * ib_dealloc_device() must be used to free structures allocated with
571 * ib_alloc_device().
572 */
573struct ib_device *_ib_alloc_device(size_t size)
574{
575 struct ib_device *device;
576 unsigned int i;
577
578 if (WARN_ON(size < sizeof(struct ib_device)))
579 return NULL;
580
581 device = kzalloc(size, GFP_KERNEL);
582 if (!device)
583 return NULL;
584
585 if (rdma_restrack_init(device)) {
586 kfree(device);
587 return NULL;
588 }
589
590 rdma_init_coredev(&device->coredev, device, &init_net);
591
592 INIT_LIST_HEAD(&device->event_handler_list);
593 spin_lock_init(&device->qp_open_list_lock);
594 init_rwsem(&device->event_handler_rwsem);
595 mutex_init(&device->unregistration_lock);
596 /*
597 * client_data needs to be alloc because we don't want our mark to be
598 * destroyed if the user stores NULL in the client data.
599 */
600 xa_init_flags(&device->client_data, XA_FLAGS_ALLOC);
601 init_rwsem(&device->client_data_rwsem);
602 xa_init_flags(&device->compat_devs, XA_FLAGS_ALLOC);
603 mutex_init(&device->compat_devs_mutex);
604 init_completion(&device->unreg_completion);
605 INIT_WORK(&device->unregistration_work, ib_unregister_work);
606
607 spin_lock_init(&device->cq_pools_lock);
608 for (i = 0; i < ARRAY_SIZE(device->cq_pools); i++)
609 INIT_LIST_HEAD(&device->cq_pools[i]);
610
611 rwlock_init(&device->cache_lock);
612
613 device->uverbs_cmd_mask =
614 BIT_ULL(IB_USER_VERBS_CMD_ALLOC_MW) |
615 BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD) |
616 BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST) |
617 BIT_ULL(IB_USER_VERBS_CMD_CLOSE_XRCD) |
618 BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH) |
619 BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
620 BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ) |
621 BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP) |
622 BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ) |
623 BIT_ULL(IB_USER_VERBS_CMD_CREATE_XSRQ) |
624 BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_MW) |
625 BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD) |
626 BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR) |
627 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH) |
628 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ) |
629 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP) |
630 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ) |
631 BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST) |
632 BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT) |
633 BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP) |
634 BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ) |
635 BIT_ULL(IB_USER_VERBS_CMD_OPEN_QP) |
636 BIT_ULL(IB_USER_VERBS_CMD_OPEN_XRCD) |
637 BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE) |
638 BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT) |
639 BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP) |
640 BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ) |
641 BIT_ULL(IB_USER_VERBS_CMD_REG_MR) |
642 BIT_ULL(IB_USER_VERBS_CMD_REREG_MR) |
643 BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ);
644 return device;
645}
646EXPORT_SYMBOL(_ib_alloc_device);
647
648/**
649 * ib_dealloc_device - free an IB device struct
650 * @device:structure to free
651 *
652 * Free a structure allocated with ib_alloc_device().
653 */
654void ib_dealloc_device(struct ib_device *device)
655{
656 if (device->ops.dealloc_driver)
657 device->ops.dealloc_driver(device);
658
659 /*
660 * ib_unregister_driver() requires all devices to remain in the xarray
661 * while their ops are callable. The last op we call is dealloc_driver
662 * above. This is needed to create a fence on op callbacks prior to
663 * allowing the driver module to unload.
664 */
665 down_write(&devices_rwsem);
666 if (xa_load(&devices, device->index) == device)
667 xa_erase(&devices, device->index);
668 up_write(&devices_rwsem);
669
670 /* Expedite releasing netdev references */
671 free_netdevs(device);
672
673 WARN_ON(!xa_empty(&device->compat_devs));
674 WARN_ON(!xa_empty(&device->client_data));
675 WARN_ON(refcount_read(&device->refcount));
676 rdma_restrack_clean(device);
677 /* Balances with device_initialize */
678 put_device(&device->dev);
679}
680EXPORT_SYMBOL(ib_dealloc_device);
681
682/*
683 * add_client_context() and remove_client_context() must be safe against
684 * parallel calls on the same device - registration/unregistration of both the
685 * device and client can be occurring in parallel.
686 *
687 * The routines need to be a fence, any caller must not return until the add
688 * or remove is fully completed.
689 */
690static int add_client_context(struct ib_device *device,
691 struct ib_client *client)
692{
693 int ret = 0;
694
695 if (!device->kverbs_provider && !client->no_kverbs_req)
696 return 0;
697
698 down_write(&device->client_data_rwsem);
699 /*
700 * So long as the client is registered hold both the client and device
701 * unregistration locks.
702 */
703 if (!refcount_inc_not_zero(&client->uses))
704 goto out_unlock;
705 refcount_inc(&device->refcount);
706
707 /*
708 * Another caller to add_client_context got here first and has already
709 * completely initialized context.
710 */
711 if (xa_get_mark(&device->client_data, client->client_id,
712 CLIENT_DATA_REGISTERED))
713 goto out;
714
715 ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
716 GFP_KERNEL));
717 if (ret)
718 goto out;
719 downgrade_write(&device->client_data_rwsem);
720 if (client->add) {
721 if (client->add(device)) {
722 /*
723 * If a client fails to add then the error code is
724 * ignored, but we won't call any more ops on this
725 * client.
726 */
727 xa_erase(&device->client_data, client->client_id);
728 up_read(&device->client_data_rwsem);
729 ib_device_put(device);
730 ib_client_put(client);
731 return 0;
732 }
733 }
734
735 /* Readers shall not see a client until add has been completed */
736 xa_set_mark(&device->client_data, client->client_id,
737 CLIENT_DATA_REGISTERED);
738 up_read(&device->client_data_rwsem);
739 return 0;
740
741out:
742 ib_device_put(device);
743 ib_client_put(client);
744out_unlock:
745 up_write(&device->client_data_rwsem);
746 return ret;
747}
748
749static void remove_client_context(struct ib_device *device,
750 unsigned int client_id)
751{
752 struct ib_client *client;
753 void *client_data;
754
755 down_write(&device->client_data_rwsem);
756 if (!xa_get_mark(&device->client_data, client_id,
757 CLIENT_DATA_REGISTERED)) {
758 up_write(&device->client_data_rwsem);
759 return;
760 }
761 client_data = xa_load(&device->client_data, client_id);
762 xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
763 client = xa_load(&clients, client_id);
764 up_write(&device->client_data_rwsem);
765
766 /*
767 * Notice we cannot be holding any exclusive locks when calling the
768 * remove callback as the remove callback can recurse back into any
769 * public functions in this module and thus try for any locks those
770 * functions take.
771 *
772 * For this reason clients and drivers should not call the
773 * unregistration functions will holdling any locks.
774 */
775 if (client->remove)
776 client->remove(device, client_data);
777
778 xa_erase(&device->client_data, client_id);
779 ib_device_put(device);
780 ib_client_put(client);
781}
782
783static int alloc_port_data(struct ib_device *device)
784{
785 struct ib_port_data_rcu *pdata_rcu;
786 u32 port;
787
788 if (device->port_data)
789 return 0;
790
791 /* This can only be called once the physical port range is defined */
792 if (WARN_ON(!device->phys_port_cnt))
793 return -EINVAL;
794
795 /* Reserve U32_MAX so the logic to go over all the ports is sane */
796 if (WARN_ON(device->phys_port_cnt == U32_MAX))
797 return -EINVAL;
798
799 /*
800 * device->port_data is indexed directly by the port number to make
801 * access to this data as efficient as possible.
802 *
803 * Therefore port_data is declared as a 1 based array with potential
804 * empty slots at the beginning.
805 */
806 pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
807 size_add(rdma_end_port(device), 1)),
808 GFP_KERNEL);
809 if (!pdata_rcu)
810 return -ENOMEM;
811 /*
812 * The rcu_head is put in front of the port data array and the stored
813 * pointer is adjusted since we never need to see that member until
814 * kfree_rcu.
815 */
816 device->port_data = pdata_rcu->pdata;
817
818 rdma_for_each_port (device, port) {
819 struct ib_port_data *pdata = &device->port_data[port];
820
821 pdata->ib_dev = device;
822 spin_lock_init(&pdata->pkey_list_lock);
823 INIT_LIST_HEAD(&pdata->pkey_list);
824 spin_lock_init(&pdata->netdev_lock);
825 INIT_HLIST_NODE(&pdata->ndev_hash_link);
826 }
827 return 0;
828}
829
830static int verify_immutable(const struct ib_device *dev, u32 port)
831{
832 return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
833 rdma_max_mad_size(dev, port) != 0);
834}
835
836static int setup_port_data(struct ib_device *device)
837{
838 u32 port;
839 int ret;
840
841 ret = alloc_port_data(device);
842 if (ret)
843 return ret;
844
845 rdma_for_each_port (device, port) {
846 struct ib_port_data *pdata = &device->port_data[port];
847
848 ret = device->ops.get_port_immutable(device, port,
849 &pdata->immutable);
850 if (ret)
851 return ret;
852
853 if (verify_immutable(device, port))
854 return -EINVAL;
855 }
856 return 0;
857}
858
859/**
860 * ib_port_immutable_read() - Read rdma port's immutable data
861 * @dev: IB device
862 * @port: port number whose immutable data to read. It starts with index 1 and
863 * valid upto including rdma_end_port().
864 */
865const struct ib_port_immutable*
866ib_port_immutable_read(struct ib_device *dev, unsigned int port)
867{
868 WARN_ON(!rdma_is_port_valid(dev, port));
869 return &dev->port_data[port].immutable;
870}
871EXPORT_SYMBOL(ib_port_immutable_read);
872
873void ib_get_device_fw_str(struct ib_device *dev, char *str)
874{
875 if (dev->ops.get_dev_fw_str)
876 dev->ops.get_dev_fw_str(dev, str);
877 else
878 str[0] = '\0';
879}
880EXPORT_SYMBOL(ib_get_device_fw_str);
881
882static void ib_policy_change_task(struct work_struct *work)
883{
884 struct ib_device *dev;
885 unsigned long index;
886
887 down_read(&devices_rwsem);
888 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
889 unsigned int i;
890
891 rdma_for_each_port (dev, i) {
892 u64 sp;
893 ib_get_cached_subnet_prefix(dev, i, &sp);
894 ib_security_cache_change(dev, i, sp);
895 }
896 }
897 up_read(&devices_rwsem);
898}
899
900static int ib_security_change(struct notifier_block *nb, unsigned long event,
901 void *lsm_data)
902{
903 if (event != LSM_POLICY_CHANGE)
904 return NOTIFY_DONE;
905
906 schedule_work(&ib_policy_change_work);
907 ib_mad_agent_security_change();
908
909 return NOTIFY_OK;
910}
911
912static void compatdev_release(struct device *dev)
913{
914 struct ib_core_device *cdev =
915 container_of(dev, struct ib_core_device, dev);
916
917 kfree(cdev);
918}
919
920static int add_one_compat_dev(struct ib_device *device,
921 struct rdma_dev_net *rnet)
922{
923 struct ib_core_device *cdev;
924 int ret;
925
926 lockdep_assert_held(&rdma_nets_rwsem);
927 if (!ib_devices_shared_netns)
928 return 0;
929
930 /*
931 * Create and add compat device in all namespaces other than where it
932 * is currently bound to.
933 */
934 if (net_eq(read_pnet(&rnet->net),
935 read_pnet(&device->coredev.rdma_net)))
936 return 0;
937
938 /*
939 * The first of init_net() or ib_register_device() to take the
940 * compat_devs_mutex wins and gets to add the device. Others will wait
941 * for completion here.
942 */
943 mutex_lock(&device->compat_devs_mutex);
944 cdev = xa_load(&device->compat_devs, rnet->id);
945 if (cdev) {
946 ret = 0;
947 goto done;
948 }
949 ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
950 if (ret)
951 goto done;
952
953 cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
954 if (!cdev) {
955 ret = -ENOMEM;
956 goto cdev_err;
957 }
958
959 cdev->dev.parent = device->dev.parent;
960 rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
961 cdev->dev.release = compatdev_release;
962 ret = dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
963 if (ret)
964 goto add_err;
965
966 ret = device_add(&cdev->dev);
967 if (ret)
968 goto add_err;
969 ret = ib_setup_port_attrs(cdev);
970 if (ret)
971 goto port_err;
972
973 ret = xa_err(xa_store(&device->compat_devs, rnet->id,
974 cdev, GFP_KERNEL));
975 if (ret)
976 goto insert_err;
977
978 mutex_unlock(&device->compat_devs_mutex);
979 return 0;
980
981insert_err:
982 ib_free_port_attrs(cdev);
983port_err:
984 device_del(&cdev->dev);
985add_err:
986 put_device(&cdev->dev);
987cdev_err:
988 xa_release(&device->compat_devs, rnet->id);
989done:
990 mutex_unlock(&device->compat_devs_mutex);
991 return ret;
992}
993
994static void remove_one_compat_dev(struct ib_device *device, u32 id)
995{
996 struct ib_core_device *cdev;
997
998 mutex_lock(&device->compat_devs_mutex);
999 cdev = xa_erase(&device->compat_devs, id);
1000 mutex_unlock(&device->compat_devs_mutex);
1001 if (cdev) {
1002 ib_free_port_attrs(cdev);
1003 device_del(&cdev->dev);
1004 put_device(&cdev->dev);
1005 }
1006}
1007
1008static void remove_compat_devs(struct ib_device *device)
1009{
1010 struct ib_core_device *cdev;
1011 unsigned long index;
1012
1013 xa_for_each (&device->compat_devs, index, cdev)
1014 remove_one_compat_dev(device, index);
1015}
1016
1017static int add_compat_devs(struct ib_device *device)
1018{
1019 struct rdma_dev_net *rnet;
1020 unsigned long index;
1021 int ret = 0;
1022
1023 lockdep_assert_held(&devices_rwsem);
1024
1025 down_read(&rdma_nets_rwsem);
1026 xa_for_each (&rdma_nets, index, rnet) {
1027 ret = add_one_compat_dev(device, rnet);
1028 if (ret)
1029 break;
1030 }
1031 up_read(&rdma_nets_rwsem);
1032 return ret;
1033}
1034
1035static void remove_all_compat_devs(void)
1036{
1037 struct ib_compat_device *cdev;
1038 struct ib_device *dev;
1039 unsigned long index;
1040
1041 down_read(&devices_rwsem);
1042 xa_for_each (&devices, index, dev) {
1043 unsigned long c_index = 0;
1044
1045 /* Hold nets_rwsem so that any other thread modifying this
1046 * system param can sync with this thread.
1047 */
1048 down_read(&rdma_nets_rwsem);
1049 xa_for_each (&dev->compat_devs, c_index, cdev)
1050 remove_one_compat_dev(dev, c_index);
1051 up_read(&rdma_nets_rwsem);
1052 }
1053 up_read(&devices_rwsem);
1054}
1055
1056static int add_all_compat_devs(void)
1057{
1058 struct rdma_dev_net *rnet;
1059 struct ib_device *dev;
1060 unsigned long index;
1061 int ret = 0;
1062
1063 down_read(&devices_rwsem);
1064 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1065 unsigned long net_index = 0;
1066
1067 /* Hold nets_rwsem so that any other thread modifying this
1068 * system param can sync with this thread.
1069 */
1070 down_read(&rdma_nets_rwsem);
1071 xa_for_each (&rdma_nets, net_index, rnet) {
1072 ret = add_one_compat_dev(dev, rnet);
1073 if (ret)
1074 break;
1075 }
1076 up_read(&rdma_nets_rwsem);
1077 }
1078 up_read(&devices_rwsem);
1079 if (ret)
1080 remove_all_compat_devs();
1081 return ret;
1082}
1083
1084int rdma_compatdev_set(u8 enable)
1085{
1086 struct rdma_dev_net *rnet;
1087 unsigned long index;
1088 int ret = 0;
1089
1090 down_write(&rdma_nets_rwsem);
1091 if (ib_devices_shared_netns == enable) {
1092 up_write(&rdma_nets_rwsem);
1093 return 0;
1094 }
1095
1096 /* enable/disable of compat devices is not supported
1097 * when more than default init_net exists.
1098 */
1099 xa_for_each (&rdma_nets, index, rnet) {
1100 ret++;
1101 break;
1102 }
1103 if (!ret)
1104 ib_devices_shared_netns = enable;
1105 up_write(&rdma_nets_rwsem);
1106 if (ret)
1107 return -EBUSY;
1108
1109 if (enable)
1110 ret = add_all_compat_devs();
1111 else
1112 remove_all_compat_devs();
1113 return ret;
1114}
1115
1116static void rdma_dev_exit_net(struct net *net)
1117{
1118 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1119 struct ib_device *dev;
1120 unsigned long index;
1121 int ret;
1122
1123 down_write(&rdma_nets_rwsem);
1124 /*
1125 * Prevent the ID from being re-used and hide the id from xa_for_each.
1126 */
1127 ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1128 WARN_ON(ret);
1129 up_write(&rdma_nets_rwsem);
1130
1131 down_read(&devices_rwsem);
1132 xa_for_each (&devices, index, dev) {
1133 get_device(&dev->dev);
1134 /*
1135 * Release the devices_rwsem so that pontentially blocking
1136 * device_del, doesn't hold the devices_rwsem for too long.
1137 */
1138 up_read(&devices_rwsem);
1139
1140 remove_one_compat_dev(dev, rnet->id);
1141
1142 /*
1143 * If the real device is in the NS then move it back to init.
1144 */
1145 rdma_dev_change_netns(dev, net, &init_net);
1146
1147 put_device(&dev->dev);
1148 down_read(&devices_rwsem);
1149 }
1150 up_read(&devices_rwsem);
1151
1152 rdma_nl_net_exit(rnet);
1153 xa_erase(&rdma_nets, rnet->id);
1154}
1155
1156static __net_init int rdma_dev_init_net(struct net *net)
1157{
1158 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1159 unsigned long index;
1160 struct ib_device *dev;
1161 int ret;
1162
1163 write_pnet(&rnet->net, net);
1164
1165 ret = rdma_nl_net_init(rnet);
1166 if (ret)
1167 return ret;
1168
1169 /* No need to create any compat devices in default init_net. */
1170 if (net_eq(net, &init_net))
1171 return 0;
1172
1173 ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1174 if (ret) {
1175 rdma_nl_net_exit(rnet);
1176 return ret;
1177 }
1178
1179 down_read(&devices_rwsem);
1180 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1181 /* Hold nets_rwsem so that netlink command cannot change
1182 * system configuration for device sharing mode.
1183 */
1184 down_read(&rdma_nets_rwsem);
1185 ret = add_one_compat_dev(dev, rnet);
1186 up_read(&rdma_nets_rwsem);
1187 if (ret)
1188 break;
1189 }
1190 up_read(&devices_rwsem);
1191
1192 if (ret)
1193 rdma_dev_exit_net(net);
1194
1195 return ret;
1196}
1197
1198/*
1199 * Assign the unique string device name and the unique device index. This is
1200 * undone by ib_dealloc_device.
1201 */
1202static int assign_name(struct ib_device *device, const char *name)
1203{
1204 static u32 last_id;
1205 int ret;
1206
1207 down_write(&devices_rwsem);
1208 /* Assign a unique name to the device */
1209 if (strchr(name, '%'))
1210 ret = alloc_name(device, name);
1211 else
1212 ret = dev_set_name(&device->dev, name);
1213 if (ret)
1214 goto out;
1215
1216 if (__ib_device_get_by_name(dev_name(&device->dev))) {
1217 ret = -ENFILE;
1218 goto out;
1219 }
1220 strscpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1221
1222 ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1223 &last_id, GFP_KERNEL);
1224 if (ret > 0)
1225 ret = 0;
1226
1227out:
1228 up_write(&devices_rwsem);
1229 return ret;
1230}
1231
1232/*
1233 * setup_device() allocates memory and sets up data that requires calling the
1234 * device ops, this is the only reason these actions are not done during
1235 * ib_alloc_device. It is undone by ib_dealloc_device().
1236 */
1237static int setup_device(struct ib_device *device)
1238{
1239 struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1240 int ret;
1241
1242 ib_device_check_mandatory(device);
1243
1244 ret = setup_port_data(device);
1245 if (ret) {
1246 dev_warn(&device->dev, "Couldn't create per-port data\n");
1247 return ret;
1248 }
1249
1250 memset(&device->attrs, 0, sizeof(device->attrs));
1251 ret = device->ops.query_device(device, &device->attrs, &uhw);
1252 if (ret) {
1253 dev_warn(&device->dev,
1254 "Couldn't query the device attributes\n");
1255 return ret;
1256 }
1257
1258 return 0;
1259}
1260
1261static void disable_device(struct ib_device *device)
1262{
1263 u32 cid;
1264
1265 WARN_ON(!refcount_read(&device->refcount));
1266
1267 down_write(&devices_rwsem);
1268 xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1269 up_write(&devices_rwsem);
1270
1271 /*
1272 * Remove clients in LIFO order, see assign_client_id. This could be
1273 * more efficient if xarray learns to reverse iterate. Since no new
1274 * clients can be added to this ib_device past this point we only need
1275 * the maximum possible client_id value here.
1276 */
1277 down_read(&clients_rwsem);
1278 cid = highest_client_id;
1279 up_read(&clients_rwsem);
1280 while (cid) {
1281 cid--;
1282 remove_client_context(device, cid);
1283 }
1284
1285 ib_cq_pool_cleanup(device);
1286
1287 /* Pairs with refcount_set in enable_device */
1288 ib_device_put(device);
1289 wait_for_completion(&device->unreg_completion);
1290
1291 /*
1292 * compat devices must be removed after device refcount drops to zero.
1293 * Otherwise init_net() may add more compatdevs after removing compat
1294 * devices and before device is disabled.
1295 */
1296 remove_compat_devs(device);
1297}
1298
1299/*
1300 * An enabled device is visible to all clients and to all the public facing
1301 * APIs that return a device pointer. This always returns with a new get, even
1302 * if it fails.
1303 */
1304static int enable_device_and_get(struct ib_device *device)
1305{
1306 struct ib_client *client;
1307 unsigned long index;
1308 int ret = 0;
1309
1310 /*
1311 * One ref belongs to the xa and the other belongs to this
1312 * thread. This is needed to guard against parallel unregistration.
1313 */
1314 refcount_set(&device->refcount, 2);
1315 down_write(&devices_rwsem);
1316 xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1317
1318 /*
1319 * By using downgrade_write() we ensure that no other thread can clear
1320 * DEVICE_REGISTERED while we are completing the client setup.
1321 */
1322 downgrade_write(&devices_rwsem);
1323
1324 if (device->ops.enable_driver) {
1325 ret = device->ops.enable_driver(device);
1326 if (ret)
1327 goto out;
1328 }
1329
1330 down_read(&clients_rwsem);
1331 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1332 ret = add_client_context(device, client);
1333 if (ret)
1334 break;
1335 }
1336 up_read(&clients_rwsem);
1337 if (!ret)
1338 ret = add_compat_devs(device);
1339out:
1340 up_read(&devices_rwsem);
1341 return ret;
1342}
1343
1344static void prevent_dealloc_device(struct ib_device *ib_dev)
1345{
1346}
1347
1348/**
1349 * ib_register_device - Register an IB device with IB core
1350 * @device: Device to register
1351 * @name: unique string device name. This may include a '%' which will
1352 * cause a unique index to be added to the passed device name.
1353 * @dma_device: pointer to a DMA-capable device. If %NULL, then the IB
1354 * device will be used. In this case the caller should fully
1355 * setup the ibdev for DMA. This usually means using dma_virt_ops.
1356 *
1357 * Low-level drivers use ib_register_device() to register their
1358 * devices with the IB core. All registered clients will receive a
1359 * callback for each device that is added. @device must be allocated
1360 * with ib_alloc_device().
1361 *
1362 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1363 * asynchronously then the device pointer may become freed as soon as this
1364 * function returns.
1365 */
1366int ib_register_device(struct ib_device *device, const char *name,
1367 struct device *dma_device)
1368{
1369 int ret;
1370
1371 ret = assign_name(device, name);
1372 if (ret)
1373 return ret;
1374
1375 /*
1376 * If the caller does not provide a DMA capable device then the IB core
1377 * will set up ib_sge and scatterlist structures that stash the kernel
1378 * virtual address into the address field.
1379 */
1380 WARN_ON(dma_device && !dma_device->dma_parms);
1381 device->dma_device = dma_device;
1382
1383 ret = setup_device(device);
1384 if (ret)
1385 return ret;
1386
1387 ret = ib_cache_setup_one(device);
1388 if (ret) {
1389 dev_warn(&device->dev,
1390 "Couldn't set up InfiniBand P_Key/GID cache\n");
1391 return ret;
1392 }
1393
1394 device->groups[0] = &ib_dev_attr_group;
1395 device->groups[1] = device->ops.device_group;
1396 ret = ib_setup_device_attrs(device);
1397 if (ret)
1398 goto cache_cleanup;
1399
1400 ib_device_register_rdmacg(device);
1401
1402 rdma_counter_init(device);
1403
1404 /*
1405 * Ensure that ADD uevent is not fired because it
1406 * is too early amd device is not initialized yet.
1407 */
1408 dev_set_uevent_suppress(&device->dev, true);
1409 ret = device_add(&device->dev);
1410 if (ret)
1411 goto cg_cleanup;
1412
1413 ret = ib_setup_port_attrs(&device->coredev);
1414 if (ret) {
1415 dev_warn(&device->dev,
1416 "Couldn't register device with driver model\n");
1417 goto dev_cleanup;
1418 }
1419
1420 ret = enable_device_and_get(device);
1421 if (ret) {
1422 void (*dealloc_fn)(struct ib_device *);
1423
1424 /*
1425 * If we hit this error flow then we don't want to
1426 * automatically dealloc the device since the caller is
1427 * expected to call ib_dealloc_device() after
1428 * ib_register_device() fails. This is tricky due to the
1429 * possibility for a parallel unregistration along with this
1430 * error flow. Since we have a refcount here we know any
1431 * parallel flow is stopped in disable_device and will see the
1432 * special dealloc_driver pointer, causing the responsibility to
1433 * ib_dealloc_device() to revert back to this thread.
1434 */
1435 dealloc_fn = device->ops.dealloc_driver;
1436 device->ops.dealloc_driver = prevent_dealloc_device;
1437 ib_device_put(device);
1438 __ib_unregister_device(device);
1439 device->ops.dealloc_driver = dealloc_fn;
1440 dev_set_uevent_suppress(&device->dev, false);
1441 return ret;
1442 }
1443 dev_set_uevent_suppress(&device->dev, false);
1444 /* Mark for userspace that device is ready */
1445 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1446 ib_device_put(device);
1447
1448 return 0;
1449
1450dev_cleanup:
1451 device_del(&device->dev);
1452cg_cleanup:
1453 dev_set_uevent_suppress(&device->dev, false);
1454 ib_device_unregister_rdmacg(device);
1455cache_cleanup:
1456 ib_cache_cleanup_one(device);
1457 return ret;
1458}
1459EXPORT_SYMBOL(ib_register_device);
1460
1461/* Callers must hold a get on the device. */
1462static void __ib_unregister_device(struct ib_device *ib_dev)
1463{
1464 /*
1465 * We have a registration lock so that all the calls to unregister are
1466 * fully fenced, once any unregister returns the device is truely
1467 * unregistered even if multiple callers are unregistering it at the
1468 * same time. This also interacts with the registration flow and
1469 * provides sane semantics if register and unregister are racing.
1470 */
1471 mutex_lock(&ib_dev->unregistration_lock);
1472 if (!refcount_read(&ib_dev->refcount))
1473 goto out;
1474
1475 disable_device(ib_dev);
1476
1477 /* Expedite removing unregistered pointers from the hash table */
1478 free_netdevs(ib_dev);
1479
1480 ib_free_port_attrs(&ib_dev->coredev);
1481 device_del(&ib_dev->dev);
1482 ib_device_unregister_rdmacg(ib_dev);
1483 ib_cache_cleanup_one(ib_dev);
1484
1485 /*
1486 * Drivers using the new flow may not call ib_dealloc_device except
1487 * in error unwind prior to registration success.
1488 */
1489 if (ib_dev->ops.dealloc_driver &&
1490 ib_dev->ops.dealloc_driver != prevent_dealloc_device) {
1491 WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1492 ib_dealloc_device(ib_dev);
1493 }
1494out:
1495 mutex_unlock(&ib_dev->unregistration_lock);
1496}
1497
1498/**
1499 * ib_unregister_device - Unregister an IB device
1500 * @ib_dev: The device to unregister
1501 *
1502 * Unregister an IB device. All clients will receive a remove callback.
1503 *
1504 * Callers should call this routine only once, and protect against races with
1505 * registration. Typically it should only be called as part of a remove
1506 * callback in an implementation of driver core's struct device_driver and
1507 * related.
1508 *
1509 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1510 * this function.
1511 */
1512void ib_unregister_device(struct ib_device *ib_dev)
1513{
1514 get_device(&ib_dev->dev);
1515 __ib_unregister_device(ib_dev);
1516 put_device(&ib_dev->dev);
1517}
1518EXPORT_SYMBOL(ib_unregister_device);
1519
1520/**
1521 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1522 * @ib_dev: The device to unregister
1523 *
1524 * This is the same as ib_unregister_device(), except it includes an internal
1525 * ib_device_put() that should match a 'get' obtained by the caller.
1526 *
1527 * It is safe to call this routine concurrently from multiple threads while
1528 * holding the 'get'. When the function returns the device is fully
1529 * unregistered.
1530 *
1531 * Drivers using this flow MUST use the driver_unregister callback to clean up
1532 * their resources associated with the device and dealloc it.
1533 */
1534void ib_unregister_device_and_put(struct ib_device *ib_dev)
1535{
1536 WARN_ON(!ib_dev->ops.dealloc_driver);
1537 get_device(&ib_dev->dev);
1538 ib_device_put(ib_dev);
1539 __ib_unregister_device(ib_dev);
1540 put_device(&ib_dev->dev);
1541}
1542EXPORT_SYMBOL(ib_unregister_device_and_put);
1543
1544/**
1545 * ib_unregister_driver - Unregister all IB devices for a driver
1546 * @driver_id: The driver to unregister
1547 *
1548 * This implements a fence for device unregistration. It only returns once all
1549 * devices associated with the driver_id have fully completed their
1550 * unregistration and returned from ib_unregister_device*().
1551 *
1552 * If device's are not yet unregistered it goes ahead and starts unregistering
1553 * them.
1554 *
1555 * This does not block creation of new devices with the given driver_id, that
1556 * is the responsibility of the caller.
1557 */
1558void ib_unregister_driver(enum rdma_driver_id driver_id)
1559{
1560 struct ib_device *ib_dev;
1561 unsigned long index;
1562
1563 down_read(&devices_rwsem);
1564 xa_for_each (&devices, index, ib_dev) {
1565 if (ib_dev->ops.driver_id != driver_id)
1566 continue;
1567
1568 get_device(&ib_dev->dev);
1569 up_read(&devices_rwsem);
1570
1571 WARN_ON(!ib_dev->ops.dealloc_driver);
1572 __ib_unregister_device(ib_dev);
1573
1574 put_device(&ib_dev->dev);
1575 down_read(&devices_rwsem);
1576 }
1577 up_read(&devices_rwsem);
1578}
1579EXPORT_SYMBOL(ib_unregister_driver);
1580
1581static void ib_unregister_work(struct work_struct *work)
1582{
1583 struct ib_device *ib_dev =
1584 container_of(work, struct ib_device, unregistration_work);
1585
1586 __ib_unregister_device(ib_dev);
1587 put_device(&ib_dev->dev);
1588}
1589
1590/**
1591 * ib_unregister_device_queued - Unregister a device using a work queue
1592 * @ib_dev: The device to unregister
1593 *
1594 * This schedules an asynchronous unregistration using a WQ for the device. A
1595 * driver should use this to avoid holding locks while doing unregistration,
1596 * such as holding the RTNL lock.
1597 *
1598 * Drivers using this API must use ib_unregister_driver before module unload
1599 * to ensure that all scheduled unregistrations have completed.
1600 */
1601void ib_unregister_device_queued(struct ib_device *ib_dev)
1602{
1603 WARN_ON(!refcount_read(&ib_dev->refcount));
1604 WARN_ON(!ib_dev->ops.dealloc_driver);
1605 get_device(&ib_dev->dev);
1606 if (!queue_work(ib_unreg_wq, &ib_dev->unregistration_work))
1607 put_device(&ib_dev->dev);
1608}
1609EXPORT_SYMBOL(ib_unregister_device_queued);
1610
1611/*
1612 * The caller must pass in a device that has the kref held and the refcount
1613 * released. If the device is in cur_net and still registered then it is moved
1614 * into net.
1615 */
1616static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1617 struct net *net)
1618{
1619 int ret2 = -EINVAL;
1620 int ret;
1621
1622 mutex_lock(&device->unregistration_lock);
1623
1624 /*
1625 * If a device not under ib_device_get() or if the unregistration_lock
1626 * is not held, the namespace can be changed, or it can be unregistered.
1627 * Check again under the lock.
1628 */
1629 if (refcount_read(&device->refcount) == 0 ||
1630 !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1631 ret = -ENODEV;
1632 goto out;
1633 }
1634
1635 kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1636 disable_device(device);
1637
1638 /*
1639 * At this point no one can be using the device, so it is safe to
1640 * change the namespace.
1641 */
1642 write_pnet(&device->coredev.rdma_net, net);
1643
1644 down_read(&devices_rwsem);
1645 /*
1646 * Currently rdma devices are system wide unique. So the device name
1647 * is guaranteed free in the new namespace. Publish the new namespace
1648 * at the sysfs level.
1649 */
1650 ret = device_rename(&device->dev, dev_name(&device->dev));
1651 up_read(&devices_rwsem);
1652 if (ret) {
1653 dev_warn(&device->dev,
1654 "%s: Couldn't rename device after namespace change\n",
1655 __func__);
1656 /* Try and put things back and re-enable the device */
1657 write_pnet(&device->coredev.rdma_net, cur_net);
1658 }
1659
1660 ret2 = enable_device_and_get(device);
1661 if (ret2) {
1662 /*
1663 * This shouldn't really happen, but if it does, let the user
1664 * retry at later point. So don't disable the device.
1665 */
1666 dev_warn(&device->dev,
1667 "%s: Couldn't re-enable device after namespace change\n",
1668 __func__);
1669 }
1670 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1671
1672 ib_device_put(device);
1673out:
1674 mutex_unlock(&device->unregistration_lock);
1675 if (ret)
1676 return ret;
1677 return ret2;
1678}
1679
1680int ib_device_set_netns_put(struct sk_buff *skb,
1681 struct ib_device *dev, u32 ns_fd)
1682{
1683 struct net *net;
1684 int ret;
1685
1686 net = get_net_ns_by_fd(ns_fd);
1687 if (IS_ERR(net)) {
1688 ret = PTR_ERR(net);
1689 goto net_err;
1690 }
1691
1692 if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1693 ret = -EPERM;
1694 goto ns_err;
1695 }
1696
1697 /*
1698 * All the ib_clients, including uverbs, are reset when the namespace is
1699 * changed and this cannot be blocked waiting for userspace to do
1700 * something, so disassociation is mandatory.
1701 */
1702 if (!dev->ops.disassociate_ucontext || ib_devices_shared_netns) {
1703 ret = -EOPNOTSUPP;
1704 goto ns_err;
1705 }
1706
1707 get_device(&dev->dev);
1708 ib_device_put(dev);
1709 ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1710 put_device(&dev->dev);
1711
1712 put_net(net);
1713 return ret;
1714
1715ns_err:
1716 put_net(net);
1717net_err:
1718 ib_device_put(dev);
1719 return ret;
1720}
1721
1722static struct pernet_operations rdma_dev_net_ops = {
1723 .init = rdma_dev_init_net,
1724 .exit = rdma_dev_exit_net,
1725 .id = &rdma_dev_net_id,
1726 .size = sizeof(struct rdma_dev_net),
1727};
1728
1729static int assign_client_id(struct ib_client *client)
1730{
1731 int ret;
1732
1733 down_write(&clients_rwsem);
1734 /*
1735 * The add/remove callbacks must be called in FIFO/LIFO order. To
1736 * achieve this we assign client_ids so they are sorted in
1737 * registration order.
1738 */
1739 client->client_id = highest_client_id;
1740 ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1741 if (ret)
1742 goto out;
1743
1744 highest_client_id++;
1745 xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1746
1747out:
1748 up_write(&clients_rwsem);
1749 return ret;
1750}
1751
1752static void remove_client_id(struct ib_client *client)
1753{
1754 down_write(&clients_rwsem);
1755 xa_erase(&clients, client->client_id);
1756 for (; highest_client_id; highest_client_id--)
1757 if (xa_load(&clients, highest_client_id - 1))
1758 break;
1759 up_write(&clients_rwsem);
1760}
1761
1762/**
1763 * ib_register_client - Register an IB client
1764 * @client:Client to register
1765 *
1766 * Upper level users of the IB drivers can use ib_register_client() to
1767 * register callbacks for IB device addition and removal. When an IB
1768 * device is added, each registered client's add method will be called
1769 * (in the order the clients were registered), and when a device is
1770 * removed, each client's remove method will be called (in the reverse
1771 * order that clients were registered). In addition, when
1772 * ib_register_client() is called, the client will receive an add
1773 * callback for all devices already registered.
1774 */
1775int ib_register_client(struct ib_client *client)
1776{
1777 struct ib_device *device;
1778 unsigned long index;
1779 int ret;
1780
1781 refcount_set(&client->uses, 1);
1782 init_completion(&client->uses_zero);
1783 ret = assign_client_id(client);
1784 if (ret)
1785 return ret;
1786
1787 down_read(&devices_rwsem);
1788 xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1789 ret = add_client_context(device, client);
1790 if (ret) {
1791 up_read(&devices_rwsem);
1792 ib_unregister_client(client);
1793 return ret;
1794 }
1795 }
1796 up_read(&devices_rwsem);
1797 return 0;
1798}
1799EXPORT_SYMBOL(ib_register_client);
1800
1801/**
1802 * ib_unregister_client - Unregister an IB client
1803 * @client:Client to unregister
1804 *
1805 * Upper level users use ib_unregister_client() to remove their client
1806 * registration. When ib_unregister_client() is called, the client
1807 * will receive a remove callback for each IB device still registered.
1808 *
1809 * This is a full fence, once it returns no client callbacks will be called,
1810 * or are running in another thread.
1811 */
1812void ib_unregister_client(struct ib_client *client)
1813{
1814 struct ib_device *device;
1815 unsigned long index;
1816
1817 down_write(&clients_rwsem);
1818 ib_client_put(client);
1819 xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1820 up_write(&clients_rwsem);
1821
1822 /* We do not want to have locks while calling client->remove() */
1823 rcu_read_lock();
1824 xa_for_each (&devices, index, device) {
1825 if (!ib_device_try_get(device))
1826 continue;
1827 rcu_read_unlock();
1828
1829 remove_client_context(device, client->client_id);
1830
1831 ib_device_put(device);
1832 rcu_read_lock();
1833 }
1834 rcu_read_unlock();
1835
1836 /*
1837 * remove_client_context() is not a fence, it can return even though a
1838 * removal is ongoing. Wait until all removals are completed.
1839 */
1840 wait_for_completion(&client->uses_zero);
1841 remove_client_id(client);
1842}
1843EXPORT_SYMBOL(ib_unregister_client);
1844
1845static int __ib_get_global_client_nl_info(const char *client_name,
1846 struct ib_client_nl_info *res)
1847{
1848 struct ib_client *client;
1849 unsigned long index;
1850 int ret = -ENOENT;
1851
1852 down_read(&clients_rwsem);
1853 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1854 if (strcmp(client->name, client_name) != 0)
1855 continue;
1856 if (!client->get_global_nl_info) {
1857 ret = -EOPNOTSUPP;
1858 break;
1859 }
1860 ret = client->get_global_nl_info(res);
1861 if (WARN_ON(ret == -ENOENT))
1862 ret = -EINVAL;
1863 if (!ret && res->cdev)
1864 get_device(res->cdev);
1865 break;
1866 }
1867 up_read(&clients_rwsem);
1868 return ret;
1869}
1870
1871static int __ib_get_client_nl_info(struct ib_device *ibdev,
1872 const char *client_name,
1873 struct ib_client_nl_info *res)
1874{
1875 unsigned long index;
1876 void *client_data;
1877 int ret = -ENOENT;
1878
1879 down_read(&ibdev->client_data_rwsem);
1880 xan_for_each_marked (&ibdev->client_data, index, client_data,
1881 CLIENT_DATA_REGISTERED) {
1882 struct ib_client *client = xa_load(&clients, index);
1883
1884 if (!client || strcmp(client->name, client_name) != 0)
1885 continue;
1886 if (!client->get_nl_info) {
1887 ret = -EOPNOTSUPP;
1888 break;
1889 }
1890 ret = client->get_nl_info(ibdev, client_data, res);
1891 if (WARN_ON(ret == -ENOENT))
1892 ret = -EINVAL;
1893
1894 /*
1895 * The cdev is guaranteed valid as long as we are inside the
1896 * client_data_rwsem as remove_one can't be called. Keep it
1897 * valid for the caller.
1898 */
1899 if (!ret && res->cdev)
1900 get_device(res->cdev);
1901 break;
1902 }
1903 up_read(&ibdev->client_data_rwsem);
1904
1905 return ret;
1906}
1907
1908/**
1909 * ib_get_client_nl_info - Fetch the nl_info from a client
1910 * @ibdev: IB device
1911 * @client_name: Name of the client
1912 * @res: Result of the query
1913 */
1914int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1915 struct ib_client_nl_info *res)
1916{
1917 int ret;
1918
1919 if (ibdev)
1920 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1921 else
1922 ret = __ib_get_global_client_nl_info(client_name, res);
1923#ifdef CONFIG_MODULES
1924 if (ret == -ENOENT) {
1925 request_module("rdma-client-%s", client_name);
1926 if (ibdev)
1927 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1928 else
1929 ret = __ib_get_global_client_nl_info(client_name, res);
1930 }
1931#endif
1932 if (ret) {
1933 if (ret == -ENOENT)
1934 return -EOPNOTSUPP;
1935 return ret;
1936 }
1937
1938 if (WARN_ON(!res->cdev))
1939 return -EINVAL;
1940 return 0;
1941}
1942
1943/**
1944 * ib_set_client_data - Set IB client context
1945 * @device:Device to set context for
1946 * @client:Client to set context for
1947 * @data:Context to set
1948 *
1949 * ib_set_client_data() sets client context data that can be retrieved with
1950 * ib_get_client_data(). This can only be called while the client is
1951 * registered to the device, once the ib_client remove() callback returns this
1952 * cannot be called.
1953 */
1954void ib_set_client_data(struct ib_device *device, struct ib_client *client,
1955 void *data)
1956{
1957 void *rc;
1958
1959 if (WARN_ON(IS_ERR(data)))
1960 data = NULL;
1961
1962 rc = xa_store(&device->client_data, client->client_id, data,
1963 GFP_KERNEL);
1964 WARN_ON(xa_is_err(rc));
1965}
1966EXPORT_SYMBOL(ib_set_client_data);
1967
1968/**
1969 * ib_register_event_handler - Register an IB event handler
1970 * @event_handler:Handler to register
1971 *
1972 * ib_register_event_handler() registers an event handler that will be
1973 * called back when asynchronous IB events occur (as defined in
1974 * chapter 11 of the InfiniBand Architecture Specification). This
1975 * callback occurs in workqueue context.
1976 */
1977void ib_register_event_handler(struct ib_event_handler *event_handler)
1978{
1979 down_write(&event_handler->device->event_handler_rwsem);
1980 list_add_tail(&event_handler->list,
1981 &event_handler->device->event_handler_list);
1982 up_write(&event_handler->device->event_handler_rwsem);
1983}
1984EXPORT_SYMBOL(ib_register_event_handler);
1985
1986/**
1987 * ib_unregister_event_handler - Unregister an event handler
1988 * @event_handler:Handler to unregister
1989 *
1990 * Unregister an event handler registered with
1991 * ib_register_event_handler().
1992 */
1993void ib_unregister_event_handler(struct ib_event_handler *event_handler)
1994{
1995 down_write(&event_handler->device->event_handler_rwsem);
1996 list_del(&event_handler->list);
1997 up_write(&event_handler->device->event_handler_rwsem);
1998}
1999EXPORT_SYMBOL(ib_unregister_event_handler);
2000
2001void ib_dispatch_event_clients(struct ib_event *event)
2002{
2003 struct ib_event_handler *handler;
2004
2005 down_read(&event->device->event_handler_rwsem);
2006
2007 list_for_each_entry(handler, &event->device->event_handler_list, list)
2008 handler->handler(handler, event);
2009
2010 up_read(&event->device->event_handler_rwsem);
2011}
2012
2013static int iw_query_port(struct ib_device *device,
2014 u32 port_num,
2015 struct ib_port_attr *port_attr)
2016{
2017 struct in_device *inetdev;
2018 struct net_device *netdev;
2019
2020 memset(port_attr, 0, sizeof(*port_attr));
2021
2022 netdev = ib_device_get_netdev(device, port_num);
2023 if (!netdev)
2024 return -ENODEV;
2025
2026 port_attr->max_mtu = IB_MTU_4096;
2027 port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
2028
2029 if (!netif_carrier_ok(netdev)) {
2030 port_attr->state = IB_PORT_DOWN;
2031 port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
2032 } else {
2033 rcu_read_lock();
2034 inetdev = __in_dev_get_rcu(netdev);
2035
2036 if (inetdev && inetdev->ifa_list) {
2037 port_attr->state = IB_PORT_ACTIVE;
2038 port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
2039 } else {
2040 port_attr->state = IB_PORT_INIT;
2041 port_attr->phys_state =
2042 IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
2043 }
2044
2045 rcu_read_unlock();
2046 }
2047
2048 dev_put(netdev);
2049 return device->ops.query_port(device, port_num, port_attr);
2050}
2051
2052static int __ib_query_port(struct ib_device *device,
2053 u32 port_num,
2054 struct ib_port_attr *port_attr)
2055{
2056 int err;
2057
2058 memset(port_attr, 0, sizeof(*port_attr));
2059
2060 err = device->ops.query_port(device, port_num, port_attr);
2061 if (err || port_attr->subnet_prefix)
2062 return err;
2063
2064 if (rdma_port_get_link_layer(device, port_num) !=
2065 IB_LINK_LAYER_INFINIBAND)
2066 return 0;
2067
2068 ib_get_cached_subnet_prefix(device, port_num,
2069 &port_attr->subnet_prefix);
2070 return 0;
2071}
2072
2073/**
2074 * ib_query_port - Query IB port attributes
2075 * @device:Device to query
2076 * @port_num:Port number to query
2077 * @port_attr:Port attributes
2078 *
2079 * ib_query_port() returns the attributes of a port through the
2080 * @port_attr pointer.
2081 */
2082int ib_query_port(struct ib_device *device,
2083 u32 port_num,
2084 struct ib_port_attr *port_attr)
2085{
2086 if (!rdma_is_port_valid(device, port_num))
2087 return -EINVAL;
2088
2089 if (rdma_protocol_iwarp(device, port_num))
2090 return iw_query_port(device, port_num, port_attr);
2091 else
2092 return __ib_query_port(device, port_num, port_attr);
2093}
2094EXPORT_SYMBOL(ib_query_port);
2095
2096static void add_ndev_hash(struct ib_port_data *pdata)
2097{
2098 unsigned long flags;
2099
2100 might_sleep();
2101
2102 spin_lock_irqsave(&ndev_hash_lock, flags);
2103 if (hash_hashed(&pdata->ndev_hash_link)) {
2104 hash_del_rcu(&pdata->ndev_hash_link);
2105 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2106 /*
2107 * We cannot do hash_add_rcu after a hash_del_rcu until the
2108 * grace period
2109 */
2110 synchronize_rcu();
2111 spin_lock_irqsave(&ndev_hash_lock, flags);
2112 }
2113 if (pdata->netdev)
2114 hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
2115 (uintptr_t)pdata->netdev);
2116 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2117}
2118
2119/**
2120 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2121 * @ib_dev: Device to modify
2122 * @ndev: net_device to affiliate, may be NULL
2123 * @port: IB port the net_device is connected to
2124 *
2125 * Drivers should use this to link the ib_device to a netdev so the netdev
2126 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2127 * affiliated with any port.
2128 *
2129 * The caller must ensure that the given ndev is not unregistered or
2130 * unregistering, and that either the ib_device is unregistered or
2131 * ib_device_set_netdev() is called with NULL when the ndev sends a
2132 * NETDEV_UNREGISTER event.
2133 */
2134int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2135 u32 port)
2136{
2137 struct net_device *old_ndev;
2138 struct ib_port_data *pdata;
2139 unsigned long flags;
2140 int ret;
2141
2142 /*
2143 * Drivers wish to call this before ib_register_driver, so we have to
2144 * setup the port data early.
2145 */
2146 ret = alloc_port_data(ib_dev);
2147 if (ret)
2148 return ret;
2149
2150 if (!rdma_is_port_valid(ib_dev, port))
2151 return -EINVAL;
2152
2153 pdata = &ib_dev->port_data[port];
2154 spin_lock_irqsave(&pdata->netdev_lock, flags);
2155 old_ndev = rcu_dereference_protected(
2156 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2157 if (old_ndev == ndev) {
2158 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2159 return 0;
2160 }
2161
2162 if (old_ndev)
2163 netdev_tracker_free(ndev, &pdata->netdev_tracker);
2164 if (ndev)
2165 netdev_hold(ndev, &pdata->netdev_tracker, GFP_ATOMIC);
2166 rcu_assign_pointer(pdata->netdev, ndev);
2167 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2168
2169 add_ndev_hash(pdata);
2170 if (old_ndev)
2171 __dev_put(old_ndev);
2172
2173 return 0;
2174}
2175EXPORT_SYMBOL(ib_device_set_netdev);
2176
2177static void free_netdevs(struct ib_device *ib_dev)
2178{
2179 unsigned long flags;
2180 u32 port;
2181
2182 if (!ib_dev->port_data)
2183 return;
2184
2185 rdma_for_each_port (ib_dev, port) {
2186 struct ib_port_data *pdata = &ib_dev->port_data[port];
2187 struct net_device *ndev;
2188
2189 spin_lock_irqsave(&pdata->netdev_lock, flags);
2190 ndev = rcu_dereference_protected(
2191 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2192 if (ndev) {
2193 spin_lock(&ndev_hash_lock);
2194 hash_del_rcu(&pdata->ndev_hash_link);
2195 spin_unlock(&ndev_hash_lock);
2196
2197 /*
2198 * If this is the last dev_put there is still a
2199 * synchronize_rcu before the netdev is kfreed, so we
2200 * can continue to rely on unlocked pointer
2201 * comparisons after the put
2202 */
2203 rcu_assign_pointer(pdata->netdev, NULL);
2204 netdev_put(ndev, &pdata->netdev_tracker);
2205 }
2206 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2207 }
2208}
2209
2210struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2211 u32 port)
2212{
2213 struct ib_port_data *pdata;
2214 struct net_device *res;
2215
2216 if (!rdma_is_port_valid(ib_dev, port))
2217 return NULL;
2218
2219 pdata = &ib_dev->port_data[port];
2220
2221 /*
2222 * New drivers should use ib_device_set_netdev() not the legacy
2223 * get_netdev().
2224 */
2225 if (ib_dev->ops.get_netdev)
2226 res = ib_dev->ops.get_netdev(ib_dev, port);
2227 else {
2228 spin_lock(&pdata->netdev_lock);
2229 res = rcu_dereference_protected(
2230 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2231 if (res)
2232 dev_hold(res);
2233 spin_unlock(&pdata->netdev_lock);
2234 }
2235
2236 /*
2237 * If we are starting to unregister expedite things by preventing
2238 * propagation of an unregistering netdev.
2239 */
2240 if (res && res->reg_state != NETREG_REGISTERED) {
2241 dev_put(res);
2242 return NULL;
2243 }
2244
2245 return res;
2246}
2247
2248/**
2249 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2250 * @ndev: netdev to locate
2251 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2252 *
2253 * Find and hold an ib_device that is associated with a netdev via
2254 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2255 * returned pointer.
2256 */
2257struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2258 enum rdma_driver_id driver_id)
2259{
2260 struct ib_device *res = NULL;
2261 struct ib_port_data *cur;
2262
2263 rcu_read_lock();
2264 hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2265 (uintptr_t)ndev) {
2266 if (rcu_access_pointer(cur->netdev) == ndev &&
2267 (driver_id == RDMA_DRIVER_UNKNOWN ||
2268 cur->ib_dev->ops.driver_id == driver_id) &&
2269 ib_device_try_get(cur->ib_dev)) {
2270 res = cur->ib_dev;
2271 break;
2272 }
2273 }
2274 rcu_read_unlock();
2275
2276 return res;
2277}
2278EXPORT_SYMBOL(ib_device_get_by_netdev);
2279
2280/**
2281 * ib_enum_roce_netdev - enumerate all RoCE ports
2282 * @ib_dev : IB device we want to query
2283 * @filter: Should we call the callback?
2284 * @filter_cookie: Cookie passed to filter
2285 * @cb: Callback to call for each found RoCE ports
2286 * @cookie: Cookie passed back to the callback
2287 *
2288 * Enumerates all of the physical RoCE ports of ib_dev
2289 * which are related to netdevice and calls callback() on each
2290 * device for which filter() function returns non zero.
2291 */
2292void ib_enum_roce_netdev(struct ib_device *ib_dev,
2293 roce_netdev_filter filter,
2294 void *filter_cookie,
2295 roce_netdev_callback cb,
2296 void *cookie)
2297{
2298 u32 port;
2299
2300 rdma_for_each_port (ib_dev, port)
2301 if (rdma_protocol_roce(ib_dev, port)) {
2302 struct net_device *idev =
2303 ib_device_get_netdev(ib_dev, port);
2304
2305 if (filter(ib_dev, port, idev, filter_cookie))
2306 cb(ib_dev, port, idev, cookie);
2307
2308 if (idev)
2309 dev_put(idev);
2310 }
2311}
2312
2313/**
2314 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2315 * @filter: Should we call the callback?
2316 * @filter_cookie: Cookie passed to filter
2317 * @cb: Callback to call for each found RoCE ports
2318 * @cookie: Cookie passed back to the callback
2319 *
2320 * Enumerates all RoCE devices' physical ports which are related
2321 * to netdevices and calls callback() on each device for which
2322 * filter() function returns non zero.
2323 */
2324void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2325 void *filter_cookie,
2326 roce_netdev_callback cb,
2327 void *cookie)
2328{
2329 struct ib_device *dev;
2330 unsigned long index;
2331
2332 down_read(&devices_rwsem);
2333 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2334 ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2335 up_read(&devices_rwsem);
2336}
2337
2338/*
2339 * ib_enum_all_devs - enumerate all ib_devices
2340 * @cb: Callback to call for each found ib_device
2341 *
2342 * Enumerates all ib_devices and calls callback() on each device.
2343 */
2344int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2345 struct netlink_callback *cb)
2346{
2347 unsigned long index;
2348 struct ib_device *dev;
2349 unsigned int idx = 0;
2350 int ret = 0;
2351
2352 down_read(&devices_rwsem);
2353 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2354 if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2355 continue;
2356
2357 ret = nldev_cb(dev, skb, cb, idx);
2358 if (ret)
2359 break;
2360 idx++;
2361 }
2362 up_read(&devices_rwsem);
2363 return ret;
2364}
2365
2366/**
2367 * ib_query_pkey - Get P_Key table entry
2368 * @device:Device to query
2369 * @port_num:Port number to query
2370 * @index:P_Key table index to query
2371 * @pkey:Returned P_Key
2372 *
2373 * ib_query_pkey() fetches the specified P_Key table entry.
2374 */
2375int ib_query_pkey(struct ib_device *device,
2376 u32 port_num, u16 index, u16 *pkey)
2377{
2378 if (!rdma_is_port_valid(device, port_num))
2379 return -EINVAL;
2380
2381 if (!device->ops.query_pkey)
2382 return -EOPNOTSUPP;
2383
2384 return device->ops.query_pkey(device, port_num, index, pkey);
2385}
2386EXPORT_SYMBOL(ib_query_pkey);
2387
2388/**
2389 * ib_modify_device - Change IB device attributes
2390 * @device:Device to modify
2391 * @device_modify_mask:Mask of attributes to change
2392 * @device_modify:New attribute values
2393 *
2394 * ib_modify_device() changes a device's attributes as specified by
2395 * the @device_modify_mask and @device_modify structure.
2396 */
2397int ib_modify_device(struct ib_device *device,
2398 int device_modify_mask,
2399 struct ib_device_modify *device_modify)
2400{
2401 if (!device->ops.modify_device)
2402 return -EOPNOTSUPP;
2403
2404 return device->ops.modify_device(device, device_modify_mask,
2405 device_modify);
2406}
2407EXPORT_SYMBOL(ib_modify_device);
2408
2409/**
2410 * ib_modify_port - Modifies the attributes for the specified port.
2411 * @device: The device to modify.
2412 * @port_num: The number of the port to modify.
2413 * @port_modify_mask: Mask used to specify which attributes of the port
2414 * to change.
2415 * @port_modify: New attribute values for the port.
2416 *
2417 * ib_modify_port() changes a port's attributes as specified by the
2418 * @port_modify_mask and @port_modify structure.
2419 */
2420int ib_modify_port(struct ib_device *device,
2421 u32 port_num, int port_modify_mask,
2422 struct ib_port_modify *port_modify)
2423{
2424 int rc;
2425
2426 if (!rdma_is_port_valid(device, port_num))
2427 return -EINVAL;
2428
2429 if (device->ops.modify_port)
2430 rc = device->ops.modify_port(device, port_num,
2431 port_modify_mask,
2432 port_modify);
2433 else if (rdma_protocol_roce(device, port_num) &&
2434 ((port_modify->set_port_cap_mask & ~IB_PORT_CM_SUP) == 0 ||
2435 (port_modify->clr_port_cap_mask & ~IB_PORT_CM_SUP) == 0))
2436 rc = 0;
2437 else
2438 rc = -EOPNOTSUPP;
2439 return rc;
2440}
2441EXPORT_SYMBOL(ib_modify_port);
2442
2443/**
2444 * ib_find_gid - Returns the port number and GID table index where
2445 * a specified GID value occurs. Its searches only for IB link layer.
2446 * @device: The device to query.
2447 * @gid: The GID value to search for.
2448 * @port_num: The port number of the device where the GID value was found.
2449 * @index: The index into the GID table where the GID was found. This
2450 * parameter may be NULL.
2451 */
2452int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2453 u32 *port_num, u16 *index)
2454{
2455 union ib_gid tmp_gid;
2456 u32 port;
2457 int ret, i;
2458
2459 rdma_for_each_port (device, port) {
2460 if (!rdma_protocol_ib(device, port))
2461 continue;
2462
2463 for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2464 ++i) {
2465 ret = rdma_query_gid(device, port, i, &tmp_gid);
2466 if (ret)
2467 continue;
2468
2469 if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2470 *port_num = port;
2471 if (index)
2472 *index = i;
2473 return 0;
2474 }
2475 }
2476 }
2477
2478 return -ENOENT;
2479}
2480EXPORT_SYMBOL(ib_find_gid);
2481
2482/**
2483 * ib_find_pkey - Returns the PKey table index where a specified
2484 * PKey value occurs.
2485 * @device: The device to query.
2486 * @port_num: The port number of the device to search for the PKey.
2487 * @pkey: The PKey value to search for.
2488 * @index: The index into the PKey table where the PKey was found.
2489 */
2490int ib_find_pkey(struct ib_device *device,
2491 u32 port_num, u16 pkey, u16 *index)
2492{
2493 int ret, i;
2494 u16 tmp_pkey;
2495 int partial_ix = -1;
2496
2497 for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2498 ++i) {
2499 ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2500 if (ret)
2501 return ret;
2502 if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2503 /* if there is full-member pkey take it.*/
2504 if (tmp_pkey & 0x8000) {
2505 *index = i;
2506 return 0;
2507 }
2508 if (partial_ix < 0)
2509 partial_ix = i;
2510 }
2511 }
2512
2513 /*no full-member, if exists take the limited*/
2514 if (partial_ix >= 0) {
2515 *index = partial_ix;
2516 return 0;
2517 }
2518 return -ENOENT;
2519}
2520EXPORT_SYMBOL(ib_find_pkey);
2521
2522/**
2523 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2524 * for a received CM request
2525 * @dev: An RDMA device on which the request has been received.
2526 * @port: Port number on the RDMA device.
2527 * @pkey: The Pkey the request came on.
2528 * @gid: A GID that the net_dev uses to communicate.
2529 * @addr: Contains the IP address that the request specified as its
2530 * destination.
2531 *
2532 */
2533struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2534 u32 port,
2535 u16 pkey,
2536 const union ib_gid *gid,
2537 const struct sockaddr *addr)
2538{
2539 struct net_device *net_dev = NULL;
2540 unsigned long index;
2541 void *client_data;
2542
2543 if (!rdma_protocol_ib(dev, port))
2544 return NULL;
2545
2546 /*
2547 * Holding the read side guarantees that the client will not become
2548 * unregistered while we are calling get_net_dev_by_params()
2549 */
2550 down_read(&dev->client_data_rwsem);
2551 xan_for_each_marked (&dev->client_data, index, client_data,
2552 CLIENT_DATA_REGISTERED) {
2553 struct ib_client *client = xa_load(&clients, index);
2554
2555 if (!client || !client->get_net_dev_by_params)
2556 continue;
2557
2558 net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2559 addr, client_data);
2560 if (net_dev)
2561 break;
2562 }
2563 up_read(&dev->client_data_rwsem);
2564
2565 return net_dev;
2566}
2567EXPORT_SYMBOL(ib_get_net_dev_by_params);
2568
2569void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2570{
2571 struct ib_device_ops *dev_ops = &dev->ops;
2572#define SET_DEVICE_OP(ptr, name) \
2573 do { \
2574 if (ops->name) \
2575 if (!((ptr)->name)) \
2576 (ptr)->name = ops->name; \
2577 } while (0)
2578
2579#define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2580
2581 if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2582 WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2583 dev_ops->driver_id != ops->driver_id);
2584 dev_ops->driver_id = ops->driver_id;
2585 }
2586 if (ops->owner) {
2587 WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2588 dev_ops->owner = ops->owner;
2589 }
2590 if (ops->uverbs_abi_ver)
2591 dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2592
2593 dev_ops->uverbs_no_driver_id_binding |=
2594 ops->uverbs_no_driver_id_binding;
2595
2596 SET_DEVICE_OP(dev_ops, add_gid);
2597 SET_DEVICE_OP(dev_ops, advise_mr);
2598 SET_DEVICE_OP(dev_ops, alloc_dm);
2599 SET_DEVICE_OP(dev_ops, alloc_hw_device_stats);
2600 SET_DEVICE_OP(dev_ops, alloc_hw_port_stats);
2601 SET_DEVICE_OP(dev_ops, alloc_mr);
2602 SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2603 SET_DEVICE_OP(dev_ops, alloc_mw);
2604 SET_DEVICE_OP(dev_ops, alloc_pd);
2605 SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2606 SET_DEVICE_OP(dev_ops, alloc_ucontext);
2607 SET_DEVICE_OP(dev_ops, alloc_xrcd);
2608 SET_DEVICE_OP(dev_ops, attach_mcast);
2609 SET_DEVICE_OP(dev_ops, check_mr_status);
2610 SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2611 SET_DEVICE_OP(dev_ops, counter_bind_qp);
2612 SET_DEVICE_OP(dev_ops, counter_dealloc);
2613 SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2614 SET_DEVICE_OP(dev_ops, counter_update_stats);
2615 SET_DEVICE_OP(dev_ops, create_ah);
2616 SET_DEVICE_OP(dev_ops, create_counters);
2617 SET_DEVICE_OP(dev_ops, create_cq);
2618 SET_DEVICE_OP(dev_ops, create_flow);
2619 SET_DEVICE_OP(dev_ops, create_qp);
2620 SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2621 SET_DEVICE_OP(dev_ops, create_srq);
2622 SET_DEVICE_OP(dev_ops, create_user_ah);
2623 SET_DEVICE_OP(dev_ops, create_wq);
2624 SET_DEVICE_OP(dev_ops, dealloc_dm);
2625 SET_DEVICE_OP(dev_ops, dealloc_driver);
2626 SET_DEVICE_OP(dev_ops, dealloc_mw);
2627 SET_DEVICE_OP(dev_ops, dealloc_pd);
2628 SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2629 SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2630 SET_DEVICE_OP(dev_ops, del_gid);
2631 SET_DEVICE_OP(dev_ops, dereg_mr);
2632 SET_DEVICE_OP(dev_ops, destroy_ah);
2633 SET_DEVICE_OP(dev_ops, destroy_counters);
2634 SET_DEVICE_OP(dev_ops, destroy_cq);
2635 SET_DEVICE_OP(dev_ops, destroy_flow);
2636 SET_DEVICE_OP(dev_ops, destroy_flow_action);
2637 SET_DEVICE_OP(dev_ops, destroy_qp);
2638 SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2639 SET_DEVICE_OP(dev_ops, destroy_srq);
2640 SET_DEVICE_OP(dev_ops, destroy_wq);
2641 SET_DEVICE_OP(dev_ops, device_group);
2642 SET_DEVICE_OP(dev_ops, detach_mcast);
2643 SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2644 SET_DEVICE_OP(dev_ops, drain_rq);
2645 SET_DEVICE_OP(dev_ops, drain_sq);
2646 SET_DEVICE_OP(dev_ops, enable_driver);
2647 SET_DEVICE_OP(dev_ops, fill_res_cm_id_entry);
2648 SET_DEVICE_OP(dev_ops, fill_res_cq_entry);
2649 SET_DEVICE_OP(dev_ops, fill_res_cq_entry_raw);
2650 SET_DEVICE_OP(dev_ops, fill_res_mr_entry);
2651 SET_DEVICE_OP(dev_ops, fill_res_mr_entry_raw);
2652 SET_DEVICE_OP(dev_ops, fill_res_qp_entry);
2653 SET_DEVICE_OP(dev_ops, fill_res_qp_entry_raw);
2654 SET_DEVICE_OP(dev_ops, fill_res_srq_entry);
2655 SET_DEVICE_OP(dev_ops, fill_res_srq_entry_raw);
2656 SET_DEVICE_OP(dev_ops, fill_stat_mr_entry);
2657 SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2658 SET_DEVICE_OP(dev_ops, get_dma_mr);
2659 SET_DEVICE_OP(dev_ops, get_hw_stats);
2660 SET_DEVICE_OP(dev_ops, get_link_layer);
2661 SET_DEVICE_OP(dev_ops, get_netdev);
2662 SET_DEVICE_OP(dev_ops, get_numa_node);
2663 SET_DEVICE_OP(dev_ops, get_port_immutable);
2664 SET_DEVICE_OP(dev_ops, get_vector_affinity);
2665 SET_DEVICE_OP(dev_ops, get_vf_config);
2666 SET_DEVICE_OP(dev_ops, get_vf_guid);
2667 SET_DEVICE_OP(dev_ops, get_vf_stats);
2668 SET_DEVICE_OP(dev_ops, iw_accept);
2669 SET_DEVICE_OP(dev_ops, iw_add_ref);
2670 SET_DEVICE_OP(dev_ops, iw_connect);
2671 SET_DEVICE_OP(dev_ops, iw_create_listen);
2672 SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2673 SET_DEVICE_OP(dev_ops, iw_get_qp);
2674 SET_DEVICE_OP(dev_ops, iw_reject);
2675 SET_DEVICE_OP(dev_ops, iw_rem_ref);
2676 SET_DEVICE_OP(dev_ops, map_mr_sg);
2677 SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
2678 SET_DEVICE_OP(dev_ops, mmap);
2679 SET_DEVICE_OP(dev_ops, mmap_free);
2680 SET_DEVICE_OP(dev_ops, modify_ah);
2681 SET_DEVICE_OP(dev_ops, modify_cq);
2682 SET_DEVICE_OP(dev_ops, modify_device);
2683 SET_DEVICE_OP(dev_ops, modify_hw_stat);
2684 SET_DEVICE_OP(dev_ops, modify_port);
2685 SET_DEVICE_OP(dev_ops, modify_qp);
2686 SET_DEVICE_OP(dev_ops, modify_srq);
2687 SET_DEVICE_OP(dev_ops, modify_wq);
2688 SET_DEVICE_OP(dev_ops, peek_cq);
2689 SET_DEVICE_OP(dev_ops, poll_cq);
2690 SET_DEVICE_OP(dev_ops, port_groups);
2691 SET_DEVICE_OP(dev_ops, post_recv);
2692 SET_DEVICE_OP(dev_ops, post_send);
2693 SET_DEVICE_OP(dev_ops, post_srq_recv);
2694 SET_DEVICE_OP(dev_ops, process_mad);
2695 SET_DEVICE_OP(dev_ops, query_ah);
2696 SET_DEVICE_OP(dev_ops, query_device);
2697 SET_DEVICE_OP(dev_ops, query_gid);
2698 SET_DEVICE_OP(dev_ops, query_pkey);
2699 SET_DEVICE_OP(dev_ops, query_port);
2700 SET_DEVICE_OP(dev_ops, query_qp);
2701 SET_DEVICE_OP(dev_ops, query_srq);
2702 SET_DEVICE_OP(dev_ops, query_ucontext);
2703 SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2704 SET_DEVICE_OP(dev_ops, read_counters);
2705 SET_DEVICE_OP(dev_ops, reg_dm_mr);
2706 SET_DEVICE_OP(dev_ops, reg_user_mr);
2707 SET_DEVICE_OP(dev_ops, reg_user_mr_dmabuf);
2708 SET_DEVICE_OP(dev_ops, req_notify_cq);
2709 SET_DEVICE_OP(dev_ops, rereg_user_mr);
2710 SET_DEVICE_OP(dev_ops, resize_cq);
2711 SET_DEVICE_OP(dev_ops, set_vf_guid);
2712 SET_DEVICE_OP(dev_ops, set_vf_link_state);
2713
2714 SET_OBJ_SIZE(dev_ops, ib_ah);
2715 SET_OBJ_SIZE(dev_ops, ib_counters);
2716 SET_OBJ_SIZE(dev_ops, ib_cq);
2717 SET_OBJ_SIZE(dev_ops, ib_mw);
2718 SET_OBJ_SIZE(dev_ops, ib_pd);
2719 SET_OBJ_SIZE(dev_ops, ib_qp);
2720 SET_OBJ_SIZE(dev_ops, ib_rwq_ind_table);
2721 SET_OBJ_SIZE(dev_ops, ib_srq);
2722 SET_OBJ_SIZE(dev_ops, ib_ucontext);
2723 SET_OBJ_SIZE(dev_ops, ib_xrcd);
2724}
2725EXPORT_SYMBOL(ib_set_device_ops);
2726
2727#ifdef CONFIG_INFINIBAND_VIRT_DMA
2728int ib_dma_virt_map_sg(struct ib_device *dev, struct scatterlist *sg, int nents)
2729{
2730 struct scatterlist *s;
2731 int i;
2732
2733 for_each_sg(sg, s, nents, i) {
2734 sg_dma_address(s) = (uintptr_t)sg_virt(s);
2735 sg_dma_len(s) = s->length;
2736 }
2737 return nents;
2738}
2739EXPORT_SYMBOL(ib_dma_virt_map_sg);
2740#endif /* CONFIG_INFINIBAND_VIRT_DMA */
2741
2742static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2743 [RDMA_NL_LS_OP_RESOLVE] = {
2744 .doit = ib_nl_handle_resolve_resp,
2745 .flags = RDMA_NL_ADMIN_PERM,
2746 },
2747 [RDMA_NL_LS_OP_SET_TIMEOUT] = {
2748 .doit = ib_nl_handle_set_timeout,
2749 .flags = RDMA_NL_ADMIN_PERM,
2750 },
2751 [RDMA_NL_LS_OP_IP_RESOLVE] = {
2752 .doit = ib_nl_handle_ip_res_resp,
2753 .flags = RDMA_NL_ADMIN_PERM,
2754 },
2755};
2756
2757static int __init ib_core_init(void)
2758{
2759 int ret = -ENOMEM;
2760
2761 ib_wq = alloc_workqueue("infiniband", 0, 0);
2762 if (!ib_wq)
2763 return -ENOMEM;
2764
2765 ib_unreg_wq = alloc_workqueue("ib-unreg-wq", WQ_UNBOUND,
2766 WQ_UNBOUND_MAX_ACTIVE);
2767 if (!ib_unreg_wq)
2768 goto err;
2769
2770 ib_comp_wq = alloc_workqueue("ib-comp-wq",
2771 WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2772 if (!ib_comp_wq)
2773 goto err_unbound;
2774
2775 ib_comp_unbound_wq =
2776 alloc_workqueue("ib-comp-unb-wq",
2777 WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2778 WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2779 if (!ib_comp_unbound_wq)
2780 goto err_comp;
2781
2782 ret = class_register(&ib_class);
2783 if (ret) {
2784 pr_warn("Couldn't create InfiniBand device class\n");
2785 goto err_comp_unbound;
2786 }
2787
2788 rdma_nl_init();
2789
2790 ret = addr_init();
2791 if (ret) {
2792 pr_warn("Couldn't init IB address resolution\n");
2793 goto err_ibnl;
2794 }
2795
2796 ret = ib_mad_init();
2797 if (ret) {
2798 pr_warn("Couldn't init IB MAD\n");
2799 goto err_addr;
2800 }
2801
2802 ret = ib_sa_init();
2803 if (ret) {
2804 pr_warn("Couldn't init SA\n");
2805 goto err_mad;
2806 }
2807
2808 ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2809 if (ret) {
2810 pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2811 goto err_sa;
2812 }
2813
2814 ret = register_pernet_device(&rdma_dev_net_ops);
2815 if (ret) {
2816 pr_warn("Couldn't init compat dev. ret %d\n", ret);
2817 goto err_compat;
2818 }
2819
2820 nldev_init();
2821 rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2822 ret = roce_gid_mgmt_init();
2823 if (ret) {
2824 pr_warn("Couldn't init RoCE GID management\n");
2825 goto err_parent;
2826 }
2827
2828 return 0;
2829
2830err_parent:
2831 rdma_nl_unregister(RDMA_NL_LS);
2832 nldev_exit();
2833 unregister_pernet_device(&rdma_dev_net_ops);
2834err_compat:
2835 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2836err_sa:
2837 ib_sa_cleanup();
2838err_mad:
2839 ib_mad_cleanup();
2840err_addr:
2841 addr_cleanup();
2842err_ibnl:
2843 class_unregister(&ib_class);
2844err_comp_unbound:
2845 destroy_workqueue(ib_comp_unbound_wq);
2846err_comp:
2847 destroy_workqueue(ib_comp_wq);
2848err_unbound:
2849 destroy_workqueue(ib_unreg_wq);
2850err:
2851 destroy_workqueue(ib_wq);
2852 return ret;
2853}
2854
2855static void __exit ib_core_cleanup(void)
2856{
2857 roce_gid_mgmt_cleanup();
2858 rdma_nl_unregister(RDMA_NL_LS);
2859 nldev_exit();
2860 unregister_pernet_device(&rdma_dev_net_ops);
2861 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2862 ib_sa_cleanup();
2863 ib_mad_cleanup();
2864 addr_cleanup();
2865 rdma_nl_exit();
2866 class_unregister(&ib_class);
2867 destroy_workqueue(ib_comp_unbound_wq);
2868 destroy_workqueue(ib_comp_wq);
2869 /* Make sure that any pending umem accounting work is done. */
2870 destroy_workqueue(ib_wq);
2871 destroy_workqueue(ib_unreg_wq);
2872 WARN_ON(!xa_empty(&clients));
2873 WARN_ON(!xa_empty(&devices));
2874}
2875
2876MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
2877
2878/* ib core relies on netdev stack to first register net_ns_type_operations
2879 * ns kobject type before ib_core initialization.
2880 */
2881fs_initcall(ib_core_init);
2882module_exit(ib_core_cleanup);
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/netdevice.h>
41#include <net/net_namespace.h>
42#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/hashtable.h>
45#include <rdma/rdma_netlink.h>
46#include <rdma/ib_addr.h>
47#include <rdma/ib_cache.h>
48#include <rdma/rdma_counter.h>
49
50#include "core_priv.h"
51#include "restrack.h"
52
53MODULE_AUTHOR("Roland Dreier");
54MODULE_DESCRIPTION("core kernel InfiniBand API");
55MODULE_LICENSE("Dual BSD/GPL");
56
57struct workqueue_struct *ib_comp_wq;
58struct workqueue_struct *ib_comp_unbound_wq;
59struct workqueue_struct *ib_wq;
60EXPORT_SYMBOL_GPL(ib_wq);
61static struct workqueue_struct *ib_unreg_wq;
62
63/*
64 * Each of the three rwsem locks (devices, clients, client_data) protects the
65 * xarray of the same name. Specifically it allows the caller to assert that
66 * the MARK will/will not be changing under the lock, and for devices and
67 * clients, that the value in the xarray is still a valid pointer. Change of
68 * the MARK is linked to the object state, so holding the lock and testing the
69 * MARK also asserts that the contained object is in a certain state.
70 *
71 * This is used to build a two stage register/unregister flow where objects
72 * can continue to be in the xarray even though they are still in progress to
73 * register/unregister.
74 *
75 * The xarray itself provides additional locking, and restartable iteration,
76 * which is also relied on.
77 *
78 * Locks should not be nested, with the exception of client_data, which is
79 * allowed to nest under the read side of the other two locks.
80 *
81 * The devices_rwsem also protects the device name list, any change or
82 * assignment of device name must also hold the write side to guarantee unique
83 * names.
84 */
85
86/*
87 * devices contains devices that have had their names assigned. The
88 * devices may not be registered. Users that care about the registration
89 * status need to call ib_device_try_get() on the device to ensure it is
90 * registered, and keep it registered, for the required duration.
91 *
92 */
93static DEFINE_XARRAY_FLAGS(devices, XA_FLAGS_ALLOC);
94static DECLARE_RWSEM(devices_rwsem);
95#define DEVICE_REGISTERED XA_MARK_1
96
97static u32 highest_client_id;
98#define CLIENT_REGISTERED XA_MARK_1
99static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
100static DECLARE_RWSEM(clients_rwsem);
101
102static void ib_client_put(struct ib_client *client)
103{
104 if (refcount_dec_and_test(&client->uses))
105 complete(&client->uses_zero);
106}
107
108/*
109 * If client_data is registered then the corresponding client must also still
110 * be registered.
111 */
112#define CLIENT_DATA_REGISTERED XA_MARK_1
113
114unsigned int rdma_dev_net_id;
115
116/*
117 * A list of net namespaces is maintained in an xarray. This is necessary
118 * because we can't get the locking right using the existing net ns list. We
119 * would require a init_net callback after the list is updated.
120 */
121static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
122/*
123 * rwsem to protect accessing the rdma_nets xarray entries.
124 */
125static DECLARE_RWSEM(rdma_nets_rwsem);
126
127bool ib_devices_shared_netns = true;
128module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
129MODULE_PARM_DESC(netns_mode,
130 "Share device among net namespaces; default=1 (shared)");
131/**
132 * rdma_dev_access_netns() - Return whether an rdma device can be accessed
133 * from a specified net namespace or not.
134 * @dev: Pointer to rdma device which needs to be checked
135 * @net: Pointer to net namesapce for which access to be checked
136 *
137 * When the rdma device is in shared mode, it ignores the net namespace.
138 * When the rdma device is exclusive to a net namespace, rdma device net
139 * namespace is checked against the specified one.
140 */
141bool rdma_dev_access_netns(const struct ib_device *dev, const struct net *net)
142{
143 return (ib_devices_shared_netns ||
144 net_eq(read_pnet(&dev->coredev.rdma_net), net));
145}
146EXPORT_SYMBOL(rdma_dev_access_netns);
147
148/*
149 * xarray has this behavior where it won't iterate over NULL values stored in
150 * allocated arrays. So we need our own iterator to see all values stored in
151 * the array. This does the same thing as xa_for_each except that it also
152 * returns NULL valued entries if the array is allocating. Simplified to only
153 * work on simple xarrays.
154 */
155static void *xan_find_marked(struct xarray *xa, unsigned long *indexp,
156 xa_mark_t filter)
157{
158 XA_STATE(xas, xa, *indexp);
159 void *entry;
160
161 rcu_read_lock();
162 do {
163 entry = xas_find_marked(&xas, ULONG_MAX, filter);
164 if (xa_is_zero(entry))
165 break;
166 } while (xas_retry(&xas, entry));
167 rcu_read_unlock();
168
169 if (entry) {
170 *indexp = xas.xa_index;
171 if (xa_is_zero(entry))
172 return NULL;
173 return entry;
174 }
175 return XA_ERROR(-ENOENT);
176}
177#define xan_for_each_marked(xa, index, entry, filter) \
178 for (index = 0, entry = xan_find_marked(xa, &(index), filter); \
179 !xa_is_err(entry); \
180 (index)++, entry = xan_find_marked(xa, &(index), filter))
181
182/* RCU hash table mapping netdevice pointers to struct ib_port_data */
183static DEFINE_SPINLOCK(ndev_hash_lock);
184static DECLARE_HASHTABLE(ndev_hash, 5);
185
186static void free_netdevs(struct ib_device *ib_dev);
187static void ib_unregister_work(struct work_struct *work);
188static void __ib_unregister_device(struct ib_device *device);
189static int ib_security_change(struct notifier_block *nb, unsigned long event,
190 void *lsm_data);
191static void ib_policy_change_task(struct work_struct *work);
192static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
193
194static void __ibdev_printk(const char *level, const struct ib_device *ibdev,
195 struct va_format *vaf)
196{
197 if (ibdev && ibdev->dev.parent)
198 dev_printk_emit(level[1] - '0',
199 ibdev->dev.parent,
200 "%s %s %s: %pV",
201 dev_driver_string(ibdev->dev.parent),
202 dev_name(ibdev->dev.parent),
203 dev_name(&ibdev->dev),
204 vaf);
205 else if (ibdev)
206 printk("%s%s: %pV",
207 level, dev_name(&ibdev->dev), vaf);
208 else
209 printk("%s(NULL ib_device): %pV", level, vaf);
210}
211
212void ibdev_printk(const char *level, const struct ib_device *ibdev,
213 const char *format, ...)
214{
215 struct va_format vaf;
216 va_list args;
217
218 va_start(args, format);
219
220 vaf.fmt = format;
221 vaf.va = &args;
222
223 __ibdev_printk(level, ibdev, &vaf);
224
225 va_end(args);
226}
227EXPORT_SYMBOL(ibdev_printk);
228
229#define define_ibdev_printk_level(func, level) \
230void func(const struct ib_device *ibdev, const char *fmt, ...) \
231{ \
232 struct va_format vaf; \
233 va_list args; \
234 \
235 va_start(args, fmt); \
236 \
237 vaf.fmt = fmt; \
238 vaf.va = &args; \
239 \
240 __ibdev_printk(level, ibdev, &vaf); \
241 \
242 va_end(args); \
243} \
244EXPORT_SYMBOL(func);
245
246define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
247define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
248define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
249define_ibdev_printk_level(ibdev_err, KERN_ERR);
250define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
251define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
252define_ibdev_printk_level(ibdev_info, KERN_INFO);
253
254static struct notifier_block ibdev_lsm_nb = {
255 .notifier_call = ib_security_change,
256};
257
258static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
259 struct net *net);
260
261/* Pointer to the RCU head at the start of the ib_port_data array */
262struct ib_port_data_rcu {
263 struct rcu_head rcu_head;
264 struct ib_port_data pdata[];
265};
266
267static void ib_device_check_mandatory(struct ib_device *device)
268{
269#define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
270 static const struct {
271 size_t offset;
272 char *name;
273 } mandatory_table[] = {
274 IB_MANDATORY_FUNC(query_device),
275 IB_MANDATORY_FUNC(query_port),
276 IB_MANDATORY_FUNC(alloc_pd),
277 IB_MANDATORY_FUNC(dealloc_pd),
278 IB_MANDATORY_FUNC(create_qp),
279 IB_MANDATORY_FUNC(modify_qp),
280 IB_MANDATORY_FUNC(destroy_qp),
281 IB_MANDATORY_FUNC(post_send),
282 IB_MANDATORY_FUNC(post_recv),
283 IB_MANDATORY_FUNC(create_cq),
284 IB_MANDATORY_FUNC(destroy_cq),
285 IB_MANDATORY_FUNC(poll_cq),
286 IB_MANDATORY_FUNC(req_notify_cq),
287 IB_MANDATORY_FUNC(get_dma_mr),
288 IB_MANDATORY_FUNC(reg_user_mr),
289 IB_MANDATORY_FUNC(dereg_mr),
290 IB_MANDATORY_FUNC(get_port_immutable)
291 };
292 int i;
293
294 device->kverbs_provider = true;
295 for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
296 if (!*(void **) ((void *) &device->ops +
297 mandatory_table[i].offset)) {
298 device->kverbs_provider = false;
299 break;
300 }
301 }
302}
303
304/*
305 * Caller must perform ib_device_put() to return the device reference count
306 * when ib_device_get_by_index() returns valid device pointer.
307 */
308struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
309{
310 struct ib_device *device;
311
312 down_read(&devices_rwsem);
313 device = xa_load(&devices, index);
314 if (device) {
315 if (!rdma_dev_access_netns(device, net)) {
316 device = NULL;
317 goto out;
318 }
319
320 if (!ib_device_try_get(device))
321 device = NULL;
322 }
323out:
324 up_read(&devices_rwsem);
325 return device;
326}
327
328/**
329 * ib_device_put - Release IB device reference
330 * @device: device whose reference to be released
331 *
332 * ib_device_put() releases reference to the IB device to allow it to be
333 * unregistered and eventually free.
334 */
335void ib_device_put(struct ib_device *device)
336{
337 if (refcount_dec_and_test(&device->refcount))
338 complete(&device->unreg_completion);
339}
340EXPORT_SYMBOL(ib_device_put);
341
342static struct ib_device *__ib_device_get_by_name(const char *name)
343{
344 struct ib_device *device;
345 unsigned long index;
346
347 xa_for_each (&devices, index, device)
348 if (!strcmp(name, dev_name(&device->dev)))
349 return device;
350
351 return NULL;
352}
353
354/**
355 * ib_device_get_by_name - Find an IB device by name
356 * @name: The name to look for
357 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
358 *
359 * Find and hold an ib_device by its name. The caller must call
360 * ib_device_put() on the returned pointer.
361 */
362struct ib_device *ib_device_get_by_name(const char *name,
363 enum rdma_driver_id driver_id)
364{
365 struct ib_device *device;
366
367 down_read(&devices_rwsem);
368 device = __ib_device_get_by_name(name);
369 if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
370 device->ops.driver_id != driver_id)
371 device = NULL;
372
373 if (device) {
374 if (!ib_device_try_get(device))
375 device = NULL;
376 }
377 up_read(&devices_rwsem);
378 return device;
379}
380EXPORT_SYMBOL(ib_device_get_by_name);
381
382static int rename_compat_devs(struct ib_device *device)
383{
384 struct ib_core_device *cdev;
385 unsigned long index;
386 int ret = 0;
387
388 mutex_lock(&device->compat_devs_mutex);
389 xa_for_each (&device->compat_devs, index, cdev) {
390 ret = device_rename(&cdev->dev, dev_name(&device->dev));
391 if (ret) {
392 dev_warn(&cdev->dev,
393 "Fail to rename compatdev to new name %s\n",
394 dev_name(&device->dev));
395 break;
396 }
397 }
398 mutex_unlock(&device->compat_devs_mutex);
399 return ret;
400}
401
402int ib_device_rename(struct ib_device *ibdev, const char *name)
403{
404 unsigned long index;
405 void *client_data;
406 int ret;
407
408 down_write(&devices_rwsem);
409 if (!strcmp(name, dev_name(&ibdev->dev))) {
410 up_write(&devices_rwsem);
411 return 0;
412 }
413
414 if (__ib_device_get_by_name(name)) {
415 up_write(&devices_rwsem);
416 return -EEXIST;
417 }
418
419 ret = device_rename(&ibdev->dev, name);
420 if (ret) {
421 up_write(&devices_rwsem);
422 return ret;
423 }
424
425 strscpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
426 ret = rename_compat_devs(ibdev);
427
428 downgrade_write(&devices_rwsem);
429 down_read(&ibdev->client_data_rwsem);
430 xan_for_each_marked(&ibdev->client_data, index, client_data,
431 CLIENT_DATA_REGISTERED) {
432 struct ib_client *client = xa_load(&clients, index);
433
434 if (!client || !client->rename)
435 continue;
436
437 client->rename(ibdev, client_data);
438 }
439 up_read(&ibdev->client_data_rwsem);
440 rdma_nl_notify_event(ibdev, 0, RDMA_RENAME_EVENT);
441 up_read(&devices_rwsem);
442 return 0;
443}
444
445int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
446{
447 if (use_dim > 1)
448 return -EINVAL;
449 ibdev->use_cq_dim = use_dim;
450
451 return 0;
452}
453
454static int alloc_name(struct ib_device *ibdev, const char *name)
455{
456 struct ib_device *device;
457 unsigned long index;
458 struct ida inuse;
459 int rc;
460 int i;
461
462 lockdep_assert_held_write(&devices_rwsem);
463 ida_init(&inuse);
464 xa_for_each (&devices, index, device) {
465 char buf[IB_DEVICE_NAME_MAX];
466
467 if (sscanf(dev_name(&device->dev), name, &i) != 1)
468 continue;
469 if (i < 0 || i >= INT_MAX)
470 continue;
471 snprintf(buf, sizeof buf, name, i);
472 if (strcmp(buf, dev_name(&device->dev)) != 0)
473 continue;
474
475 rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
476 if (rc < 0)
477 goto out;
478 }
479
480 rc = ida_alloc(&inuse, GFP_KERNEL);
481 if (rc < 0)
482 goto out;
483
484 rc = dev_set_name(&ibdev->dev, name, rc);
485out:
486 ida_destroy(&inuse);
487 return rc;
488}
489
490static void ib_device_release(struct device *device)
491{
492 struct ib_device *dev = container_of(device, struct ib_device, dev);
493
494 free_netdevs(dev);
495 WARN_ON(refcount_read(&dev->refcount));
496 if (dev->hw_stats_data)
497 ib_device_release_hw_stats(dev->hw_stats_data);
498 if (dev->port_data) {
499 ib_cache_release_one(dev);
500 ib_security_release_port_pkey_list(dev);
501 rdma_counter_release(dev);
502 kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
503 pdata[0]),
504 rcu_head);
505 }
506
507 mutex_destroy(&dev->subdev_lock);
508 mutex_destroy(&dev->unregistration_lock);
509 mutex_destroy(&dev->compat_devs_mutex);
510
511 xa_destroy(&dev->compat_devs);
512 xa_destroy(&dev->client_data);
513 kfree_rcu(dev, rcu_head);
514}
515
516static int ib_device_uevent(const struct device *device,
517 struct kobj_uevent_env *env)
518{
519 if (add_uevent_var(env, "NAME=%s", dev_name(device)))
520 return -ENOMEM;
521
522 /*
523 * It would be nice to pass the node GUID with the event...
524 */
525
526 return 0;
527}
528
529static const void *net_namespace(const struct device *d)
530{
531 const struct ib_core_device *coredev =
532 container_of(d, struct ib_core_device, dev);
533
534 return read_pnet(&coredev->rdma_net);
535}
536
537static struct class ib_class = {
538 .name = "infiniband",
539 .dev_release = ib_device_release,
540 .dev_uevent = ib_device_uevent,
541 .ns_type = &net_ns_type_operations,
542 .namespace = net_namespace,
543};
544
545static void rdma_init_coredev(struct ib_core_device *coredev,
546 struct ib_device *dev, struct net *net)
547{
548 /* This BUILD_BUG_ON is intended to catch layout change
549 * of union of ib_core_device and device.
550 * dev must be the first element as ib_core and providers
551 * driver uses it. Adding anything in ib_core_device before
552 * device will break this assumption.
553 */
554 BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
555 offsetof(struct ib_device, dev));
556
557 coredev->dev.class = &ib_class;
558 coredev->dev.groups = dev->groups;
559 device_initialize(&coredev->dev);
560 coredev->owner = dev;
561 INIT_LIST_HEAD(&coredev->port_list);
562 write_pnet(&coredev->rdma_net, net);
563}
564
565/**
566 * _ib_alloc_device - allocate an IB device struct
567 * @size:size of structure to allocate
568 *
569 * Low-level drivers should use ib_alloc_device() to allocate &struct
570 * ib_device. @size is the size of the structure to be allocated,
571 * including any private data used by the low-level driver.
572 * ib_dealloc_device() must be used to free structures allocated with
573 * ib_alloc_device().
574 */
575struct ib_device *_ib_alloc_device(size_t size)
576{
577 struct ib_device *device;
578 unsigned int i;
579
580 if (WARN_ON(size < sizeof(struct ib_device)))
581 return NULL;
582
583 device = kzalloc(size, GFP_KERNEL);
584 if (!device)
585 return NULL;
586
587 if (rdma_restrack_init(device)) {
588 kfree(device);
589 return NULL;
590 }
591
592 rdma_init_coredev(&device->coredev, device, &init_net);
593
594 INIT_LIST_HEAD(&device->event_handler_list);
595 spin_lock_init(&device->qp_open_list_lock);
596 init_rwsem(&device->event_handler_rwsem);
597 mutex_init(&device->unregistration_lock);
598 /*
599 * client_data needs to be alloc because we don't want our mark to be
600 * destroyed if the user stores NULL in the client data.
601 */
602 xa_init_flags(&device->client_data, XA_FLAGS_ALLOC);
603 init_rwsem(&device->client_data_rwsem);
604 xa_init_flags(&device->compat_devs, XA_FLAGS_ALLOC);
605 mutex_init(&device->compat_devs_mutex);
606 init_completion(&device->unreg_completion);
607 INIT_WORK(&device->unregistration_work, ib_unregister_work);
608
609 spin_lock_init(&device->cq_pools_lock);
610 for (i = 0; i < ARRAY_SIZE(device->cq_pools); i++)
611 INIT_LIST_HEAD(&device->cq_pools[i]);
612
613 rwlock_init(&device->cache_lock);
614
615 device->uverbs_cmd_mask =
616 BIT_ULL(IB_USER_VERBS_CMD_ALLOC_MW) |
617 BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD) |
618 BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST) |
619 BIT_ULL(IB_USER_VERBS_CMD_CLOSE_XRCD) |
620 BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH) |
621 BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
622 BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ) |
623 BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP) |
624 BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ) |
625 BIT_ULL(IB_USER_VERBS_CMD_CREATE_XSRQ) |
626 BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_MW) |
627 BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD) |
628 BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR) |
629 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH) |
630 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ) |
631 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP) |
632 BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ) |
633 BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST) |
634 BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT) |
635 BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP) |
636 BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ) |
637 BIT_ULL(IB_USER_VERBS_CMD_OPEN_QP) |
638 BIT_ULL(IB_USER_VERBS_CMD_OPEN_XRCD) |
639 BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE) |
640 BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT) |
641 BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP) |
642 BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ) |
643 BIT_ULL(IB_USER_VERBS_CMD_REG_MR) |
644 BIT_ULL(IB_USER_VERBS_CMD_REREG_MR) |
645 BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ);
646
647 mutex_init(&device->subdev_lock);
648 INIT_LIST_HEAD(&device->subdev_list_head);
649 INIT_LIST_HEAD(&device->subdev_list);
650
651 return device;
652}
653EXPORT_SYMBOL(_ib_alloc_device);
654
655/**
656 * ib_dealloc_device - free an IB device struct
657 * @device:structure to free
658 *
659 * Free a structure allocated with ib_alloc_device().
660 */
661void ib_dealloc_device(struct ib_device *device)
662{
663 if (device->ops.dealloc_driver)
664 device->ops.dealloc_driver(device);
665
666 /*
667 * ib_unregister_driver() requires all devices to remain in the xarray
668 * while their ops are callable. The last op we call is dealloc_driver
669 * above. This is needed to create a fence on op callbacks prior to
670 * allowing the driver module to unload.
671 */
672 down_write(&devices_rwsem);
673 if (xa_load(&devices, device->index) == device)
674 xa_erase(&devices, device->index);
675 up_write(&devices_rwsem);
676
677 /* Expedite releasing netdev references */
678 free_netdevs(device);
679
680 WARN_ON(!xa_empty(&device->compat_devs));
681 WARN_ON(!xa_empty(&device->client_data));
682 WARN_ON(refcount_read(&device->refcount));
683 rdma_restrack_clean(device);
684 /* Balances with device_initialize */
685 put_device(&device->dev);
686}
687EXPORT_SYMBOL(ib_dealloc_device);
688
689/*
690 * add_client_context() and remove_client_context() must be safe against
691 * parallel calls on the same device - registration/unregistration of both the
692 * device and client can be occurring in parallel.
693 *
694 * The routines need to be a fence, any caller must not return until the add
695 * or remove is fully completed.
696 */
697static int add_client_context(struct ib_device *device,
698 struct ib_client *client)
699{
700 int ret = 0;
701
702 if (!device->kverbs_provider && !client->no_kverbs_req)
703 return 0;
704
705 down_write(&device->client_data_rwsem);
706 /*
707 * So long as the client is registered hold both the client and device
708 * unregistration locks.
709 */
710 if (!refcount_inc_not_zero(&client->uses))
711 goto out_unlock;
712 refcount_inc(&device->refcount);
713
714 /*
715 * Another caller to add_client_context got here first and has already
716 * completely initialized context.
717 */
718 if (xa_get_mark(&device->client_data, client->client_id,
719 CLIENT_DATA_REGISTERED))
720 goto out;
721
722 ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
723 GFP_KERNEL));
724 if (ret)
725 goto out;
726 downgrade_write(&device->client_data_rwsem);
727 if (client->add) {
728 if (client->add(device)) {
729 /*
730 * If a client fails to add then the error code is
731 * ignored, but we won't call any more ops on this
732 * client.
733 */
734 xa_erase(&device->client_data, client->client_id);
735 up_read(&device->client_data_rwsem);
736 ib_device_put(device);
737 ib_client_put(client);
738 return 0;
739 }
740 }
741
742 /* Readers shall not see a client until add has been completed */
743 xa_set_mark(&device->client_data, client->client_id,
744 CLIENT_DATA_REGISTERED);
745 up_read(&device->client_data_rwsem);
746 return 0;
747
748out:
749 ib_device_put(device);
750 ib_client_put(client);
751out_unlock:
752 up_write(&device->client_data_rwsem);
753 return ret;
754}
755
756static void remove_client_context(struct ib_device *device,
757 unsigned int client_id)
758{
759 struct ib_client *client;
760 void *client_data;
761
762 down_write(&device->client_data_rwsem);
763 if (!xa_get_mark(&device->client_data, client_id,
764 CLIENT_DATA_REGISTERED)) {
765 up_write(&device->client_data_rwsem);
766 return;
767 }
768 client_data = xa_load(&device->client_data, client_id);
769 xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
770 client = xa_load(&clients, client_id);
771 up_write(&device->client_data_rwsem);
772
773 /*
774 * Notice we cannot be holding any exclusive locks when calling the
775 * remove callback as the remove callback can recurse back into any
776 * public functions in this module and thus try for any locks those
777 * functions take.
778 *
779 * For this reason clients and drivers should not call the
780 * unregistration functions will holdling any locks.
781 */
782 if (client->remove)
783 client->remove(device, client_data);
784
785 xa_erase(&device->client_data, client_id);
786 ib_device_put(device);
787 ib_client_put(client);
788}
789
790static int alloc_port_data(struct ib_device *device)
791{
792 struct ib_port_data_rcu *pdata_rcu;
793 u32 port;
794
795 if (device->port_data)
796 return 0;
797
798 /* This can only be called once the physical port range is defined */
799 if (WARN_ON(!device->phys_port_cnt))
800 return -EINVAL;
801
802 /* Reserve U32_MAX so the logic to go over all the ports is sane */
803 if (WARN_ON(device->phys_port_cnt == U32_MAX))
804 return -EINVAL;
805
806 /*
807 * device->port_data is indexed directly by the port number to make
808 * access to this data as efficient as possible.
809 *
810 * Therefore port_data is declared as a 1 based array with potential
811 * empty slots at the beginning.
812 */
813 pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
814 size_add(rdma_end_port(device), 1)),
815 GFP_KERNEL);
816 if (!pdata_rcu)
817 return -ENOMEM;
818 /*
819 * The rcu_head is put in front of the port data array and the stored
820 * pointer is adjusted since we never need to see that member until
821 * kfree_rcu.
822 */
823 device->port_data = pdata_rcu->pdata;
824
825 rdma_for_each_port (device, port) {
826 struct ib_port_data *pdata = &device->port_data[port];
827
828 pdata->ib_dev = device;
829 spin_lock_init(&pdata->pkey_list_lock);
830 INIT_LIST_HEAD(&pdata->pkey_list);
831 spin_lock_init(&pdata->netdev_lock);
832 INIT_HLIST_NODE(&pdata->ndev_hash_link);
833 }
834 return 0;
835}
836
837static int verify_immutable(const struct ib_device *dev, u32 port)
838{
839 return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
840 rdma_max_mad_size(dev, port) != 0);
841}
842
843static int setup_port_data(struct ib_device *device)
844{
845 u32 port;
846 int ret;
847
848 ret = alloc_port_data(device);
849 if (ret)
850 return ret;
851
852 rdma_for_each_port (device, port) {
853 struct ib_port_data *pdata = &device->port_data[port];
854
855 ret = device->ops.get_port_immutable(device, port,
856 &pdata->immutable);
857 if (ret)
858 return ret;
859
860 if (verify_immutable(device, port))
861 return -EINVAL;
862 }
863 return 0;
864}
865
866/**
867 * ib_port_immutable_read() - Read rdma port's immutable data
868 * @dev: IB device
869 * @port: port number whose immutable data to read. It starts with index 1 and
870 * valid upto including rdma_end_port().
871 */
872const struct ib_port_immutable*
873ib_port_immutable_read(struct ib_device *dev, unsigned int port)
874{
875 WARN_ON(!rdma_is_port_valid(dev, port));
876 return &dev->port_data[port].immutable;
877}
878EXPORT_SYMBOL(ib_port_immutable_read);
879
880void ib_get_device_fw_str(struct ib_device *dev, char *str)
881{
882 if (dev->ops.get_dev_fw_str)
883 dev->ops.get_dev_fw_str(dev, str);
884 else
885 str[0] = '\0';
886}
887EXPORT_SYMBOL(ib_get_device_fw_str);
888
889static void ib_policy_change_task(struct work_struct *work)
890{
891 struct ib_device *dev;
892 unsigned long index;
893
894 down_read(&devices_rwsem);
895 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
896 unsigned int i;
897
898 rdma_for_each_port (dev, i) {
899 u64 sp;
900 ib_get_cached_subnet_prefix(dev, i, &sp);
901 ib_security_cache_change(dev, i, sp);
902 }
903 }
904 up_read(&devices_rwsem);
905}
906
907static int ib_security_change(struct notifier_block *nb, unsigned long event,
908 void *lsm_data)
909{
910 if (event != LSM_POLICY_CHANGE)
911 return NOTIFY_DONE;
912
913 schedule_work(&ib_policy_change_work);
914 ib_mad_agent_security_change();
915
916 return NOTIFY_OK;
917}
918
919static void compatdev_release(struct device *dev)
920{
921 struct ib_core_device *cdev =
922 container_of(dev, struct ib_core_device, dev);
923
924 kfree(cdev);
925}
926
927static int add_one_compat_dev(struct ib_device *device,
928 struct rdma_dev_net *rnet)
929{
930 struct ib_core_device *cdev;
931 int ret;
932
933 lockdep_assert_held(&rdma_nets_rwsem);
934 if (!ib_devices_shared_netns)
935 return 0;
936
937 /*
938 * Create and add compat device in all namespaces other than where it
939 * is currently bound to.
940 */
941 if (net_eq(read_pnet(&rnet->net),
942 read_pnet(&device->coredev.rdma_net)))
943 return 0;
944
945 /*
946 * The first of init_net() or ib_register_device() to take the
947 * compat_devs_mutex wins and gets to add the device. Others will wait
948 * for completion here.
949 */
950 mutex_lock(&device->compat_devs_mutex);
951 cdev = xa_load(&device->compat_devs, rnet->id);
952 if (cdev) {
953 ret = 0;
954 goto done;
955 }
956 ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
957 if (ret)
958 goto done;
959
960 cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
961 if (!cdev) {
962 ret = -ENOMEM;
963 goto cdev_err;
964 }
965
966 cdev->dev.parent = device->dev.parent;
967 rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
968 cdev->dev.release = compatdev_release;
969 ret = dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
970 if (ret)
971 goto add_err;
972
973 ret = device_add(&cdev->dev);
974 if (ret)
975 goto add_err;
976 ret = ib_setup_port_attrs(cdev);
977 if (ret)
978 goto port_err;
979
980 ret = xa_err(xa_store(&device->compat_devs, rnet->id,
981 cdev, GFP_KERNEL));
982 if (ret)
983 goto insert_err;
984
985 mutex_unlock(&device->compat_devs_mutex);
986 return 0;
987
988insert_err:
989 ib_free_port_attrs(cdev);
990port_err:
991 device_del(&cdev->dev);
992add_err:
993 put_device(&cdev->dev);
994cdev_err:
995 xa_release(&device->compat_devs, rnet->id);
996done:
997 mutex_unlock(&device->compat_devs_mutex);
998 return ret;
999}
1000
1001static void remove_one_compat_dev(struct ib_device *device, u32 id)
1002{
1003 struct ib_core_device *cdev;
1004
1005 mutex_lock(&device->compat_devs_mutex);
1006 cdev = xa_erase(&device->compat_devs, id);
1007 mutex_unlock(&device->compat_devs_mutex);
1008 if (cdev) {
1009 ib_free_port_attrs(cdev);
1010 device_del(&cdev->dev);
1011 put_device(&cdev->dev);
1012 }
1013}
1014
1015static void remove_compat_devs(struct ib_device *device)
1016{
1017 struct ib_core_device *cdev;
1018 unsigned long index;
1019
1020 xa_for_each (&device->compat_devs, index, cdev)
1021 remove_one_compat_dev(device, index);
1022}
1023
1024static int add_compat_devs(struct ib_device *device)
1025{
1026 struct rdma_dev_net *rnet;
1027 unsigned long index;
1028 int ret = 0;
1029
1030 lockdep_assert_held(&devices_rwsem);
1031
1032 down_read(&rdma_nets_rwsem);
1033 xa_for_each (&rdma_nets, index, rnet) {
1034 ret = add_one_compat_dev(device, rnet);
1035 if (ret)
1036 break;
1037 }
1038 up_read(&rdma_nets_rwsem);
1039 return ret;
1040}
1041
1042static void remove_all_compat_devs(void)
1043{
1044 struct ib_compat_device *cdev;
1045 struct ib_device *dev;
1046 unsigned long index;
1047
1048 down_read(&devices_rwsem);
1049 xa_for_each (&devices, index, dev) {
1050 unsigned long c_index = 0;
1051
1052 /* Hold nets_rwsem so that any other thread modifying this
1053 * system param can sync with this thread.
1054 */
1055 down_read(&rdma_nets_rwsem);
1056 xa_for_each (&dev->compat_devs, c_index, cdev)
1057 remove_one_compat_dev(dev, c_index);
1058 up_read(&rdma_nets_rwsem);
1059 }
1060 up_read(&devices_rwsem);
1061}
1062
1063static int add_all_compat_devs(void)
1064{
1065 struct rdma_dev_net *rnet;
1066 struct ib_device *dev;
1067 unsigned long index;
1068 int ret = 0;
1069
1070 down_read(&devices_rwsem);
1071 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1072 unsigned long net_index = 0;
1073
1074 /* Hold nets_rwsem so that any other thread modifying this
1075 * system param can sync with this thread.
1076 */
1077 down_read(&rdma_nets_rwsem);
1078 xa_for_each (&rdma_nets, net_index, rnet) {
1079 ret = add_one_compat_dev(dev, rnet);
1080 if (ret)
1081 break;
1082 }
1083 up_read(&rdma_nets_rwsem);
1084 }
1085 up_read(&devices_rwsem);
1086 if (ret)
1087 remove_all_compat_devs();
1088 return ret;
1089}
1090
1091int rdma_compatdev_set(u8 enable)
1092{
1093 struct rdma_dev_net *rnet;
1094 unsigned long index;
1095 int ret = 0;
1096
1097 down_write(&rdma_nets_rwsem);
1098 if (ib_devices_shared_netns == enable) {
1099 up_write(&rdma_nets_rwsem);
1100 return 0;
1101 }
1102
1103 /* enable/disable of compat devices is not supported
1104 * when more than default init_net exists.
1105 */
1106 xa_for_each (&rdma_nets, index, rnet) {
1107 ret++;
1108 break;
1109 }
1110 if (!ret)
1111 ib_devices_shared_netns = enable;
1112 up_write(&rdma_nets_rwsem);
1113 if (ret)
1114 return -EBUSY;
1115
1116 if (enable)
1117 ret = add_all_compat_devs();
1118 else
1119 remove_all_compat_devs();
1120 return ret;
1121}
1122
1123static void rdma_dev_exit_net(struct net *net)
1124{
1125 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1126 struct ib_device *dev;
1127 unsigned long index;
1128 int ret;
1129
1130 down_write(&rdma_nets_rwsem);
1131 /*
1132 * Prevent the ID from being re-used and hide the id from xa_for_each.
1133 */
1134 ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1135 WARN_ON(ret);
1136 up_write(&rdma_nets_rwsem);
1137
1138 down_read(&devices_rwsem);
1139 xa_for_each (&devices, index, dev) {
1140 get_device(&dev->dev);
1141 /*
1142 * Release the devices_rwsem so that pontentially blocking
1143 * device_del, doesn't hold the devices_rwsem for too long.
1144 */
1145 up_read(&devices_rwsem);
1146
1147 remove_one_compat_dev(dev, rnet->id);
1148
1149 /*
1150 * If the real device is in the NS then move it back to init.
1151 */
1152 rdma_dev_change_netns(dev, net, &init_net);
1153
1154 put_device(&dev->dev);
1155 down_read(&devices_rwsem);
1156 }
1157 up_read(&devices_rwsem);
1158
1159 rdma_nl_net_exit(rnet);
1160 xa_erase(&rdma_nets, rnet->id);
1161}
1162
1163static __net_init int rdma_dev_init_net(struct net *net)
1164{
1165 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1166 unsigned long index;
1167 struct ib_device *dev;
1168 int ret;
1169
1170 write_pnet(&rnet->net, net);
1171
1172 ret = rdma_nl_net_init(rnet);
1173 if (ret)
1174 return ret;
1175
1176 /* No need to create any compat devices in default init_net. */
1177 if (net_eq(net, &init_net))
1178 return 0;
1179
1180 ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1181 if (ret) {
1182 rdma_nl_net_exit(rnet);
1183 return ret;
1184 }
1185
1186 down_read(&devices_rwsem);
1187 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1188 /* Hold nets_rwsem so that netlink command cannot change
1189 * system configuration for device sharing mode.
1190 */
1191 down_read(&rdma_nets_rwsem);
1192 ret = add_one_compat_dev(dev, rnet);
1193 up_read(&rdma_nets_rwsem);
1194 if (ret)
1195 break;
1196 }
1197 up_read(&devices_rwsem);
1198
1199 if (ret)
1200 rdma_dev_exit_net(net);
1201
1202 return ret;
1203}
1204
1205/*
1206 * Assign the unique string device name and the unique device index. This is
1207 * undone by ib_dealloc_device.
1208 */
1209static int assign_name(struct ib_device *device, const char *name)
1210{
1211 static u32 last_id;
1212 int ret;
1213
1214 down_write(&devices_rwsem);
1215 /* Assign a unique name to the device */
1216 if (strchr(name, '%'))
1217 ret = alloc_name(device, name);
1218 else
1219 ret = dev_set_name(&device->dev, name);
1220 if (ret)
1221 goto out;
1222
1223 if (__ib_device_get_by_name(dev_name(&device->dev))) {
1224 ret = -ENFILE;
1225 goto out;
1226 }
1227 strscpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1228
1229 ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1230 &last_id, GFP_KERNEL);
1231 if (ret > 0)
1232 ret = 0;
1233
1234out:
1235 up_write(&devices_rwsem);
1236 return ret;
1237}
1238
1239/*
1240 * setup_device() allocates memory and sets up data that requires calling the
1241 * device ops, this is the only reason these actions are not done during
1242 * ib_alloc_device. It is undone by ib_dealloc_device().
1243 */
1244static int setup_device(struct ib_device *device)
1245{
1246 struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1247 int ret;
1248
1249 ib_device_check_mandatory(device);
1250
1251 ret = setup_port_data(device);
1252 if (ret) {
1253 dev_warn(&device->dev, "Couldn't create per-port data\n");
1254 return ret;
1255 }
1256
1257 memset(&device->attrs, 0, sizeof(device->attrs));
1258 ret = device->ops.query_device(device, &device->attrs, &uhw);
1259 if (ret) {
1260 dev_warn(&device->dev,
1261 "Couldn't query the device attributes\n");
1262 return ret;
1263 }
1264
1265 return 0;
1266}
1267
1268static void disable_device(struct ib_device *device)
1269{
1270 u32 cid;
1271
1272 WARN_ON(!refcount_read(&device->refcount));
1273
1274 down_write(&devices_rwsem);
1275 xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1276 up_write(&devices_rwsem);
1277
1278 /*
1279 * Remove clients in LIFO order, see assign_client_id. This could be
1280 * more efficient if xarray learns to reverse iterate. Since no new
1281 * clients can be added to this ib_device past this point we only need
1282 * the maximum possible client_id value here.
1283 */
1284 down_read(&clients_rwsem);
1285 cid = highest_client_id;
1286 up_read(&clients_rwsem);
1287 while (cid) {
1288 cid--;
1289 remove_client_context(device, cid);
1290 }
1291
1292 ib_cq_pool_cleanup(device);
1293
1294 /* Pairs with refcount_set in enable_device */
1295 ib_device_put(device);
1296 wait_for_completion(&device->unreg_completion);
1297
1298 /*
1299 * compat devices must be removed after device refcount drops to zero.
1300 * Otherwise init_net() may add more compatdevs after removing compat
1301 * devices and before device is disabled.
1302 */
1303 remove_compat_devs(device);
1304}
1305
1306/*
1307 * An enabled device is visible to all clients and to all the public facing
1308 * APIs that return a device pointer. This always returns with a new get, even
1309 * if it fails.
1310 */
1311static int enable_device_and_get(struct ib_device *device)
1312{
1313 struct ib_client *client;
1314 unsigned long index;
1315 int ret = 0;
1316
1317 /*
1318 * One ref belongs to the xa and the other belongs to this
1319 * thread. This is needed to guard against parallel unregistration.
1320 */
1321 refcount_set(&device->refcount, 2);
1322 down_write(&devices_rwsem);
1323 xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1324
1325 /*
1326 * By using downgrade_write() we ensure that no other thread can clear
1327 * DEVICE_REGISTERED while we are completing the client setup.
1328 */
1329 downgrade_write(&devices_rwsem);
1330
1331 if (device->ops.enable_driver) {
1332 ret = device->ops.enable_driver(device);
1333 if (ret)
1334 goto out;
1335 }
1336
1337 down_read(&clients_rwsem);
1338 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1339 ret = add_client_context(device, client);
1340 if (ret)
1341 break;
1342 }
1343 up_read(&clients_rwsem);
1344 if (!ret)
1345 ret = add_compat_devs(device);
1346out:
1347 up_read(&devices_rwsem);
1348 return ret;
1349}
1350
1351static void prevent_dealloc_device(struct ib_device *ib_dev)
1352{
1353}
1354
1355static void ib_device_notify_register(struct ib_device *device)
1356{
1357 struct net_device *netdev;
1358 u32 port;
1359 int ret;
1360
1361 ret = rdma_nl_notify_event(device, 0, RDMA_REGISTER_EVENT);
1362 if (ret)
1363 return;
1364
1365 rdma_for_each_port(device, port) {
1366 netdev = ib_device_get_netdev(device, port);
1367 if (!netdev)
1368 continue;
1369
1370 ret = rdma_nl_notify_event(device, port,
1371 RDMA_NETDEV_ATTACH_EVENT);
1372 dev_put(netdev);
1373 if (ret)
1374 return;
1375 }
1376}
1377
1378/**
1379 * ib_register_device - Register an IB device with IB core
1380 * @device: Device to register
1381 * @name: unique string device name. This may include a '%' which will
1382 * cause a unique index to be added to the passed device name.
1383 * @dma_device: pointer to a DMA-capable device. If %NULL, then the IB
1384 * device will be used. In this case the caller should fully
1385 * setup the ibdev for DMA. This usually means using dma_virt_ops.
1386 *
1387 * Low-level drivers use ib_register_device() to register their
1388 * devices with the IB core. All registered clients will receive a
1389 * callback for each device that is added. @device must be allocated
1390 * with ib_alloc_device().
1391 *
1392 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1393 * asynchronously then the device pointer may become freed as soon as this
1394 * function returns.
1395 */
1396int ib_register_device(struct ib_device *device, const char *name,
1397 struct device *dma_device)
1398{
1399 int ret;
1400
1401 ret = assign_name(device, name);
1402 if (ret)
1403 return ret;
1404
1405 /*
1406 * If the caller does not provide a DMA capable device then the IB core
1407 * will set up ib_sge and scatterlist structures that stash the kernel
1408 * virtual address into the address field.
1409 */
1410 WARN_ON(dma_device && !dma_device->dma_parms);
1411 device->dma_device = dma_device;
1412
1413 ret = setup_device(device);
1414 if (ret)
1415 return ret;
1416
1417 ret = ib_cache_setup_one(device);
1418 if (ret) {
1419 dev_warn(&device->dev,
1420 "Couldn't set up InfiniBand P_Key/GID cache\n");
1421 return ret;
1422 }
1423
1424 device->groups[0] = &ib_dev_attr_group;
1425 device->groups[1] = device->ops.device_group;
1426 ret = ib_setup_device_attrs(device);
1427 if (ret)
1428 goto cache_cleanup;
1429
1430 ib_device_register_rdmacg(device);
1431
1432 rdma_counter_init(device);
1433
1434 /*
1435 * Ensure that ADD uevent is not fired because it
1436 * is too early amd device is not initialized yet.
1437 */
1438 dev_set_uevent_suppress(&device->dev, true);
1439 ret = device_add(&device->dev);
1440 if (ret)
1441 goto cg_cleanup;
1442
1443 ret = ib_setup_port_attrs(&device->coredev);
1444 if (ret) {
1445 dev_warn(&device->dev,
1446 "Couldn't register device with driver model\n");
1447 goto dev_cleanup;
1448 }
1449
1450 ret = enable_device_and_get(device);
1451 if (ret) {
1452 void (*dealloc_fn)(struct ib_device *);
1453
1454 /*
1455 * If we hit this error flow then we don't want to
1456 * automatically dealloc the device since the caller is
1457 * expected to call ib_dealloc_device() after
1458 * ib_register_device() fails. This is tricky due to the
1459 * possibility for a parallel unregistration along with this
1460 * error flow. Since we have a refcount here we know any
1461 * parallel flow is stopped in disable_device and will see the
1462 * special dealloc_driver pointer, causing the responsibility to
1463 * ib_dealloc_device() to revert back to this thread.
1464 */
1465 dealloc_fn = device->ops.dealloc_driver;
1466 device->ops.dealloc_driver = prevent_dealloc_device;
1467 ib_device_put(device);
1468 __ib_unregister_device(device);
1469 device->ops.dealloc_driver = dealloc_fn;
1470 dev_set_uevent_suppress(&device->dev, false);
1471 return ret;
1472 }
1473 dev_set_uevent_suppress(&device->dev, false);
1474 /* Mark for userspace that device is ready */
1475 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1476
1477 ib_device_notify_register(device);
1478 ib_device_put(device);
1479
1480 return 0;
1481
1482dev_cleanup:
1483 device_del(&device->dev);
1484cg_cleanup:
1485 dev_set_uevent_suppress(&device->dev, false);
1486 ib_device_unregister_rdmacg(device);
1487cache_cleanup:
1488 ib_cache_cleanup_one(device);
1489 return ret;
1490}
1491EXPORT_SYMBOL(ib_register_device);
1492
1493/* Callers must hold a get on the device. */
1494static void __ib_unregister_device(struct ib_device *ib_dev)
1495{
1496 struct ib_device *sub, *tmp;
1497
1498 mutex_lock(&ib_dev->subdev_lock);
1499 list_for_each_entry_safe_reverse(sub, tmp,
1500 &ib_dev->subdev_list_head,
1501 subdev_list) {
1502 list_del(&sub->subdev_list);
1503 ib_dev->ops.del_sub_dev(sub);
1504 ib_device_put(ib_dev);
1505 }
1506 mutex_unlock(&ib_dev->subdev_lock);
1507
1508 /*
1509 * We have a registration lock so that all the calls to unregister are
1510 * fully fenced, once any unregister returns the device is truely
1511 * unregistered even if multiple callers are unregistering it at the
1512 * same time. This also interacts with the registration flow and
1513 * provides sane semantics if register and unregister are racing.
1514 */
1515 mutex_lock(&ib_dev->unregistration_lock);
1516 if (!refcount_read(&ib_dev->refcount))
1517 goto out;
1518
1519 disable_device(ib_dev);
1520 rdma_nl_notify_event(ib_dev, 0, RDMA_UNREGISTER_EVENT);
1521
1522 /* Expedite removing unregistered pointers from the hash table */
1523 free_netdevs(ib_dev);
1524
1525 ib_free_port_attrs(&ib_dev->coredev);
1526 device_del(&ib_dev->dev);
1527 ib_device_unregister_rdmacg(ib_dev);
1528 ib_cache_cleanup_one(ib_dev);
1529
1530 /*
1531 * Drivers using the new flow may not call ib_dealloc_device except
1532 * in error unwind prior to registration success.
1533 */
1534 if (ib_dev->ops.dealloc_driver &&
1535 ib_dev->ops.dealloc_driver != prevent_dealloc_device) {
1536 WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1537 ib_dealloc_device(ib_dev);
1538 }
1539out:
1540 mutex_unlock(&ib_dev->unregistration_lock);
1541}
1542
1543/**
1544 * ib_unregister_device - Unregister an IB device
1545 * @ib_dev: The device to unregister
1546 *
1547 * Unregister an IB device. All clients will receive a remove callback.
1548 *
1549 * Callers should call this routine only once, and protect against races with
1550 * registration. Typically it should only be called as part of a remove
1551 * callback in an implementation of driver core's struct device_driver and
1552 * related.
1553 *
1554 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1555 * this function.
1556 */
1557void ib_unregister_device(struct ib_device *ib_dev)
1558{
1559 get_device(&ib_dev->dev);
1560 __ib_unregister_device(ib_dev);
1561 put_device(&ib_dev->dev);
1562}
1563EXPORT_SYMBOL(ib_unregister_device);
1564
1565/**
1566 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1567 * @ib_dev: The device to unregister
1568 *
1569 * This is the same as ib_unregister_device(), except it includes an internal
1570 * ib_device_put() that should match a 'get' obtained by the caller.
1571 *
1572 * It is safe to call this routine concurrently from multiple threads while
1573 * holding the 'get'. When the function returns the device is fully
1574 * unregistered.
1575 *
1576 * Drivers using this flow MUST use the driver_unregister callback to clean up
1577 * their resources associated with the device and dealloc it.
1578 */
1579void ib_unregister_device_and_put(struct ib_device *ib_dev)
1580{
1581 WARN_ON(!ib_dev->ops.dealloc_driver);
1582 get_device(&ib_dev->dev);
1583 ib_device_put(ib_dev);
1584 __ib_unregister_device(ib_dev);
1585 put_device(&ib_dev->dev);
1586}
1587EXPORT_SYMBOL(ib_unregister_device_and_put);
1588
1589/**
1590 * ib_unregister_driver - Unregister all IB devices for a driver
1591 * @driver_id: The driver to unregister
1592 *
1593 * This implements a fence for device unregistration. It only returns once all
1594 * devices associated with the driver_id have fully completed their
1595 * unregistration and returned from ib_unregister_device*().
1596 *
1597 * If device's are not yet unregistered it goes ahead and starts unregistering
1598 * them.
1599 *
1600 * This does not block creation of new devices with the given driver_id, that
1601 * is the responsibility of the caller.
1602 */
1603void ib_unregister_driver(enum rdma_driver_id driver_id)
1604{
1605 struct ib_device *ib_dev;
1606 unsigned long index;
1607
1608 down_read(&devices_rwsem);
1609 xa_for_each (&devices, index, ib_dev) {
1610 if (ib_dev->ops.driver_id != driver_id)
1611 continue;
1612
1613 get_device(&ib_dev->dev);
1614 up_read(&devices_rwsem);
1615
1616 WARN_ON(!ib_dev->ops.dealloc_driver);
1617 __ib_unregister_device(ib_dev);
1618
1619 put_device(&ib_dev->dev);
1620 down_read(&devices_rwsem);
1621 }
1622 up_read(&devices_rwsem);
1623}
1624EXPORT_SYMBOL(ib_unregister_driver);
1625
1626static void ib_unregister_work(struct work_struct *work)
1627{
1628 struct ib_device *ib_dev =
1629 container_of(work, struct ib_device, unregistration_work);
1630
1631 __ib_unregister_device(ib_dev);
1632 put_device(&ib_dev->dev);
1633}
1634
1635/**
1636 * ib_unregister_device_queued - Unregister a device using a work queue
1637 * @ib_dev: The device to unregister
1638 *
1639 * This schedules an asynchronous unregistration using a WQ for the device. A
1640 * driver should use this to avoid holding locks while doing unregistration,
1641 * such as holding the RTNL lock.
1642 *
1643 * Drivers using this API must use ib_unregister_driver before module unload
1644 * to ensure that all scheduled unregistrations have completed.
1645 */
1646void ib_unregister_device_queued(struct ib_device *ib_dev)
1647{
1648 WARN_ON(!refcount_read(&ib_dev->refcount));
1649 WARN_ON(!ib_dev->ops.dealloc_driver);
1650 get_device(&ib_dev->dev);
1651 if (!queue_work(ib_unreg_wq, &ib_dev->unregistration_work))
1652 put_device(&ib_dev->dev);
1653}
1654EXPORT_SYMBOL(ib_unregister_device_queued);
1655
1656/*
1657 * The caller must pass in a device that has the kref held and the refcount
1658 * released. If the device is in cur_net and still registered then it is moved
1659 * into net.
1660 */
1661static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1662 struct net *net)
1663{
1664 int ret2 = -EINVAL;
1665 int ret;
1666
1667 mutex_lock(&device->unregistration_lock);
1668
1669 /*
1670 * If a device not under ib_device_get() or if the unregistration_lock
1671 * is not held, the namespace can be changed, or it can be unregistered.
1672 * Check again under the lock.
1673 */
1674 if (refcount_read(&device->refcount) == 0 ||
1675 !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1676 ret = -ENODEV;
1677 goto out;
1678 }
1679
1680 kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1681 disable_device(device);
1682
1683 /*
1684 * At this point no one can be using the device, so it is safe to
1685 * change the namespace.
1686 */
1687 write_pnet(&device->coredev.rdma_net, net);
1688
1689 down_read(&devices_rwsem);
1690 /*
1691 * Currently rdma devices are system wide unique. So the device name
1692 * is guaranteed free in the new namespace. Publish the new namespace
1693 * at the sysfs level.
1694 */
1695 ret = device_rename(&device->dev, dev_name(&device->dev));
1696 up_read(&devices_rwsem);
1697 if (ret) {
1698 dev_warn(&device->dev,
1699 "%s: Couldn't rename device after namespace change\n",
1700 __func__);
1701 /* Try and put things back and re-enable the device */
1702 write_pnet(&device->coredev.rdma_net, cur_net);
1703 }
1704
1705 ret2 = enable_device_and_get(device);
1706 if (ret2) {
1707 /*
1708 * This shouldn't really happen, but if it does, let the user
1709 * retry at later point. So don't disable the device.
1710 */
1711 dev_warn(&device->dev,
1712 "%s: Couldn't re-enable device after namespace change\n",
1713 __func__);
1714 }
1715 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1716
1717 ib_device_put(device);
1718out:
1719 mutex_unlock(&device->unregistration_lock);
1720 if (ret)
1721 return ret;
1722 return ret2;
1723}
1724
1725int ib_device_set_netns_put(struct sk_buff *skb,
1726 struct ib_device *dev, u32 ns_fd)
1727{
1728 struct net *net;
1729 int ret;
1730
1731 net = get_net_ns_by_fd(ns_fd);
1732 if (IS_ERR(net)) {
1733 ret = PTR_ERR(net);
1734 goto net_err;
1735 }
1736
1737 if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1738 ret = -EPERM;
1739 goto ns_err;
1740 }
1741
1742 /*
1743 * All the ib_clients, including uverbs, are reset when the namespace is
1744 * changed and this cannot be blocked waiting for userspace to do
1745 * something, so disassociation is mandatory.
1746 */
1747 if (!dev->ops.disassociate_ucontext || ib_devices_shared_netns) {
1748 ret = -EOPNOTSUPP;
1749 goto ns_err;
1750 }
1751
1752 get_device(&dev->dev);
1753 ib_device_put(dev);
1754 ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1755 put_device(&dev->dev);
1756
1757 put_net(net);
1758 return ret;
1759
1760ns_err:
1761 put_net(net);
1762net_err:
1763 ib_device_put(dev);
1764 return ret;
1765}
1766
1767static struct pernet_operations rdma_dev_net_ops = {
1768 .init = rdma_dev_init_net,
1769 .exit = rdma_dev_exit_net,
1770 .id = &rdma_dev_net_id,
1771 .size = sizeof(struct rdma_dev_net),
1772};
1773
1774static int assign_client_id(struct ib_client *client)
1775{
1776 int ret;
1777
1778 lockdep_assert_held(&clients_rwsem);
1779 /*
1780 * The add/remove callbacks must be called in FIFO/LIFO order. To
1781 * achieve this we assign client_ids so they are sorted in
1782 * registration order.
1783 */
1784 client->client_id = highest_client_id;
1785 ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1786 if (ret)
1787 return ret;
1788
1789 highest_client_id++;
1790 xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1791 return 0;
1792}
1793
1794static void remove_client_id(struct ib_client *client)
1795{
1796 down_write(&clients_rwsem);
1797 xa_erase(&clients, client->client_id);
1798 for (; highest_client_id; highest_client_id--)
1799 if (xa_load(&clients, highest_client_id - 1))
1800 break;
1801 up_write(&clients_rwsem);
1802}
1803
1804/**
1805 * ib_register_client - Register an IB client
1806 * @client:Client to register
1807 *
1808 * Upper level users of the IB drivers can use ib_register_client() to
1809 * register callbacks for IB device addition and removal. When an IB
1810 * device is added, each registered client's add method will be called
1811 * (in the order the clients were registered), and when a device is
1812 * removed, each client's remove method will be called (in the reverse
1813 * order that clients were registered). In addition, when
1814 * ib_register_client() is called, the client will receive an add
1815 * callback for all devices already registered.
1816 */
1817int ib_register_client(struct ib_client *client)
1818{
1819 struct ib_device *device;
1820 unsigned long index;
1821 bool need_unreg = false;
1822 int ret;
1823
1824 refcount_set(&client->uses, 1);
1825 init_completion(&client->uses_zero);
1826
1827 /*
1828 * The devices_rwsem is held in write mode to ensure that a racing
1829 * ib_register_device() sees a consisent view of clients and devices.
1830 */
1831 down_write(&devices_rwsem);
1832 down_write(&clients_rwsem);
1833 ret = assign_client_id(client);
1834 if (ret)
1835 goto out;
1836
1837 need_unreg = true;
1838 xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1839 ret = add_client_context(device, client);
1840 if (ret)
1841 goto out;
1842 }
1843 ret = 0;
1844out:
1845 up_write(&clients_rwsem);
1846 up_write(&devices_rwsem);
1847 if (need_unreg && ret)
1848 ib_unregister_client(client);
1849 return ret;
1850}
1851EXPORT_SYMBOL(ib_register_client);
1852
1853/**
1854 * ib_unregister_client - Unregister an IB client
1855 * @client:Client to unregister
1856 *
1857 * Upper level users use ib_unregister_client() to remove their client
1858 * registration. When ib_unregister_client() is called, the client
1859 * will receive a remove callback for each IB device still registered.
1860 *
1861 * This is a full fence, once it returns no client callbacks will be called,
1862 * or are running in another thread.
1863 */
1864void ib_unregister_client(struct ib_client *client)
1865{
1866 struct ib_device *device;
1867 unsigned long index;
1868
1869 down_write(&clients_rwsem);
1870 ib_client_put(client);
1871 xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1872 up_write(&clients_rwsem);
1873
1874 /* We do not want to have locks while calling client->remove() */
1875 rcu_read_lock();
1876 xa_for_each (&devices, index, device) {
1877 if (!ib_device_try_get(device))
1878 continue;
1879 rcu_read_unlock();
1880
1881 remove_client_context(device, client->client_id);
1882
1883 ib_device_put(device);
1884 rcu_read_lock();
1885 }
1886 rcu_read_unlock();
1887
1888 /*
1889 * remove_client_context() is not a fence, it can return even though a
1890 * removal is ongoing. Wait until all removals are completed.
1891 */
1892 wait_for_completion(&client->uses_zero);
1893 remove_client_id(client);
1894}
1895EXPORT_SYMBOL(ib_unregister_client);
1896
1897static int __ib_get_global_client_nl_info(const char *client_name,
1898 struct ib_client_nl_info *res)
1899{
1900 struct ib_client *client;
1901 unsigned long index;
1902 int ret = -ENOENT;
1903
1904 down_read(&clients_rwsem);
1905 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1906 if (strcmp(client->name, client_name) != 0)
1907 continue;
1908 if (!client->get_global_nl_info) {
1909 ret = -EOPNOTSUPP;
1910 break;
1911 }
1912 ret = client->get_global_nl_info(res);
1913 if (WARN_ON(ret == -ENOENT))
1914 ret = -EINVAL;
1915 if (!ret && res->cdev)
1916 get_device(res->cdev);
1917 break;
1918 }
1919 up_read(&clients_rwsem);
1920 return ret;
1921}
1922
1923static int __ib_get_client_nl_info(struct ib_device *ibdev,
1924 const char *client_name,
1925 struct ib_client_nl_info *res)
1926{
1927 unsigned long index;
1928 void *client_data;
1929 int ret = -ENOENT;
1930
1931 down_read(&ibdev->client_data_rwsem);
1932 xan_for_each_marked (&ibdev->client_data, index, client_data,
1933 CLIENT_DATA_REGISTERED) {
1934 struct ib_client *client = xa_load(&clients, index);
1935
1936 if (!client || strcmp(client->name, client_name) != 0)
1937 continue;
1938 if (!client->get_nl_info) {
1939 ret = -EOPNOTSUPP;
1940 break;
1941 }
1942 ret = client->get_nl_info(ibdev, client_data, res);
1943 if (WARN_ON(ret == -ENOENT))
1944 ret = -EINVAL;
1945
1946 /*
1947 * The cdev is guaranteed valid as long as we are inside the
1948 * client_data_rwsem as remove_one can't be called. Keep it
1949 * valid for the caller.
1950 */
1951 if (!ret && res->cdev)
1952 get_device(res->cdev);
1953 break;
1954 }
1955 up_read(&ibdev->client_data_rwsem);
1956
1957 return ret;
1958}
1959
1960/**
1961 * ib_get_client_nl_info - Fetch the nl_info from a client
1962 * @ibdev: IB device
1963 * @client_name: Name of the client
1964 * @res: Result of the query
1965 */
1966int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1967 struct ib_client_nl_info *res)
1968{
1969 int ret;
1970
1971 if (ibdev)
1972 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1973 else
1974 ret = __ib_get_global_client_nl_info(client_name, res);
1975#ifdef CONFIG_MODULES
1976 if (ret == -ENOENT) {
1977 request_module("rdma-client-%s", client_name);
1978 if (ibdev)
1979 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1980 else
1981 ret = __ib_get_global_client_nl_info(client_name, res);
1982 }
1983#endif
1984 if (ret) {
1985 if (ret == -ENOENT)
1986 return -EOPNOTSUPP;
1987 return ret;
1988 }
1989
1990 if (WARN_ON(!res->cdev))
1991 return -EINVAL;
1992 return 0;
1993}
1994
1995/**
1996 * ib_set_client_data - Set IB client context
1997 * @device:Device to set context for
1998 * @client:Client to set context for
1999 * @data:Context to set
2000 *
2001 * ib_set_client_data() sets client context data that can be retrieved with
2002 * ib_get_client_data(). This can only be called while the client is
2003 * registered to the device, once the ib_client remove() callback returns this
2004 * cannot be called.
2005 */
2006void ib_set_client_data(struct ib_device *device, struct ib_client *client,
2007 void *data)
2008{
2009 void *rc;
2010
2011 if (WARN_ON(IS_ERR(data)))
2012 data = NULL;
2013
2014 rc = xa_store(&device->client_data, client->client_id, data,
2015 GFP_KERNEL);
2016 WARN_ON(xa_is_err(rc));
2017}
2018EXPORT_SYMBOL(ib_set_client_data);
2019
2020/**
2021 * ib_register_event_handler - Register an IB event handler
2022 * @event_handler:Handler to register
2023 *
2024 * ib_register_event_handler() registers an event handler that will be
2025 * called back when asynchronous IB events occur (as defined in
2026 * chapter 11 of the InfiniBand Architecture Specification). This
2027 * callback occurs in workqueue context.
2028 */
2029void ib_register_event_handler(struct ib_event_handler *event_handler)
2030{
2031 down_write(&event_handler->device->event_handler_rwsem);
2032 list_add_tail(&event_handler->list,
2033 &event_handler->device->event_handler_list);
2034 up_write(&event_handler->device->event_handler_rwsem);
2035}
2036EXPORT_SYMBOL(ib_register_event_handler);
2037
2038/**
2039 * ib_unregister_event_handler - Unregister an event handler
2040 * @event_handler:Handler to unregister
2041 *
2042 * Unregister an event handler registered with
2043 * ib_register_event_handler().
2044 */
2045void ib_unregister_event_handler(struct ib_event_handler *event_handler)
2046{
2047 down_write(&event_handler->device->event_handler_rwsem);
2048 list_del(&event_handler->list);
2049 up_write(&event_handler->device->event_handler_rwsem);
2050}
2051EXPORT_SYMBOL(ib_unregister_event_handler);
2052
2053void ib_dispatch_event_clients(struct ib_event *event)
2054{
2055 struct ib_event_handler *handler;
2056
2057 down_read(&event->device->event_handler_rwsem);
2058
2059 list_for_each_entry(handler, &event->device->event_handler_list, list)
2060 handler->handler(handler, event);
2061
2062 up_read(&event->device->event_handler_rwsem);
2063}
2064
2065static int iw_query_port(struct ib_device *device,
2066 u32 port_num,
2067 struct ib_port_attr *port_attr)
2068{
2069 struct in_device *inetdev;
2070 struct net_device *netdev;
2071
2072 memset(port_attr, 0, sizeof(*port_attr));
2073
2074 netdev = ib_device_get_netdev(device, port_num);
2075 if (!netdev)
2076 return -ENODEV;
2077
2078 port_attr->max_mtu = IB_MTU_4096;
2079 port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
2080
2081 if (!netif_carrier_ok(netdev)) {
2082 port_attr->state = IB_PORT_DOWN;
2083 port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
2084 } else {
2085 rcu_read_lock();
2086 inetdev = __in_dev_get_rcu(netdev);
2087
2088 if (inetdev && inetdev->ifa_list) {
2089 port_attr->state = IB_PORT_ACTIVE;
2090 port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
2091 } else {
2092 port_attr->state = IB_PORT_INIT;
2093 port_attr->phys_state =
2094 IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
2095 }
2096
2097 rcu_read_unlock();
2098 }
2099
2100 dev_put(netdev);
2101 return device->ops.query_port(device, port_num, port_attr);
2102}
2103
2104static int __ib_query_port(struct ib_device *device,
2105 u32 port_num,
2106 struct ib_port_attr *port_attr)
2107{
2108 int err;
2109
2110 memset(port_attr, 0, sizeof(*port_attr));
2111
2112 err = device->ops.query_port(device, port_num, port_attr);
2113 if (err || port_attr->subnet_prefix)
2114 return err;
2115
2116 if (rdma_port_get_link_layer(device, port_num) !=
2117 IB_LINK_LAYER_INFINIBAND)
2118 return 0;
2119
2120 ib_get_cached_subnet_prefix(device, port_num,
2121 &port_attr->subnet_prefix);
2122 return 0;
2123}
2124
2125/**
2126 * ib_query_port - Query IB port attributes
2127 * @device:Device to query
2128 * @port_num:Port number to query
2129 * @port_attr:Port attributes
2130 *
2131 * ib_query_port() returns the attributes of a port through the
2132 * @port_attr pointer.
2133 */
2134int ib_query_port(struct ib_device *device,
2135 u32 port_num,
2136 struct ib_port_attr *port_attr)
2137{
2138 if (!rdma_is_port_valid(device, port_num))
2139 return -EINVAL;
2140
2141 if (rdma_protocol_iwarp(device, port_num))
2142 return iw_query_port(device, port_num, port_attr);
2143 else
2144 return __ib_query_port(device, port_num, port_attr);
2145}
2146EXPORT_SYMBOL(ib_query_port);
2147
2148static void add_ndev_hash(struct ib_port_data *pdata)
2149{
2150 unsigned long flags;
2151
2152 might_sleep();
2153
2154 spin_lock_irqsave(&ndev_hash_lock, flags);
2155 if (hash_hashed(&pdata->ndev_hash_link)) {
2156 hash_del_rcu(&pdata->ndev_hash_link);
2157 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2158 /*
2159 * We cannot do hash_add_rcu after a hash_del_rcu until the
2160 * grace period
2161 */
2162 synchronize_rcu();
2163 spin_lock_irqsave(&ndev_hash_lock, flags);
2164 }
2165 if (pdata->netdev)
2166 hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
2167 (uintptr_t)pdata->netdev);
2168 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2169}
2170
2171/**
2172 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2173 * @ib_dev: Device to modify
2174 * @ndev: net_device to affiliate, may be NULL
2175 * @port: IB port the net_device is connected to
2176 *
2177 * Drivers should use this to link the ib_device to a netdev so the netdev
2178 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2179 * affiliated with any port.
2180 *
2181 * The caller must ensure that the given ndev is not unregistered or
2182 * unregistering, and that either the ib_device is unregistered or
2183 * ib_device_set_netdev() is called with NULL when the ndev sends a
2184 * NETDEV_UNREGISTER event.
2185 */
2186int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2187 u32 port)
2188{
2189 enum rdma_nl_notify_event_type etype;
2190 struct net_device *old_ndev;
2191 struct ib_port_data *pdata;
2192 unsigned long flags;
2193 int ret;
2194
2195 if (!rdma_is_port_valid(ib_dev, port))
2196 return -EINVAL;
2197
2198 /*
2199 * Drivers wish to call this before ib_register_driver, so we have to
2200 * setup the port data early.
2201 */
2202 ret = alloc_port_data(ib_dev);
2203 if (ret)
2204 return ret;
2205
2206 pdata = &ib_dev->port_data[port];
2207 spin_lock_irqsave(&pdata->netdev_lock, flags);
2208 old_ndev = rcu_dereference_protected(
2209 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2210 if (old_ndev == ndev) {
2211 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2212 return 0;
2213 }
2214
2215 rcu_assign_pointer(pdata->netdev, ndev);
2216 netdev_put(old_ndev, &pdata->netdev_tracker);
2217 netdev_hold(ndev, &pdata->netdev_tracker, GFP_ATOMIC);
2218 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2219
2220 add_ndev_hash(pdata);
2221
2222 /* Make sure that the device is registered before we send events */
2223 if (xa_load(&devices, ib_dev->index) != ib_dev)
2224 return 0;
2225
2226 etype = ndev ? RDMA_NETDEV_ATTACH_EVENT : RDMA_NETDEV_DETACH_EVENT;
2227 rdma_nl_notify_event(ib_dev, port, etype);
2228
2229 return 0;
2230}
2231EXPORT_SYMBOL(ib_device_set_netdev);
2232
2233static void free_netdevs(struct ib_device *ib_dev)
2234{
2235 unsigned long flags;
2236 u32 port;
2237
2238 if (!ib_dev->port_data)
2239 return;
2240
2241 rdma_for_each_port (ib_dev, port) {
2242 struct ib_port_data *pdata = &ib_dev->port_data[port];
2243 struct net_device *ndev;
2244
2245 spin_lock_irqsave(&pdata->netdev_lock, flags);
2246 ndev = rcu_dereference_protected(
2247 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2248 if (ndev) {
2249 spin_lock(&ndev_hash_lock);
2250 hash_del_rcu(&pdata->ndev_hash_link);
2251 spin_unlock(&ndev_hash_lock);
2252
2253 /*
2254 * If this is the last dev_put there is still a
2255 * synchronize_rcu before the netdev is kfreed, so we
2256 * can continue to rely on unlocked pointer
2257 * comparisons after the put
2258 */
2259 rcu_assign_pointer(pdata->netdev, NULL);
2260 netdev_put(ndev, &pdata->netdev_tracker);
2261 }
2262 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2263 }
2264}
2265
2266struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2267 u32 port)
2268{
2269 struct ib_port_data *pdata;
2270 struct net_device *res;
2271
2272 if (!rdma_is_port_valid(ib_dev, port))
2273 return NULL;
2274
2275 if (!ib_dev->port_data)
2276 return NULL;
2277
2278 pdata = &ib_dev->port_data[port];
2279
2280 /*
2281 * New drivers should use ib_device_set_netdev() not the legacy
2282 * get_netdev().
2283 */
2284 if (ib_dev->ops.get_netdev)
2285 res = ib_dev->ops.get_netdev(ib_dev, port);
2286 else {
2287 spin_lock(&pdata->netdev_lock);
2288 res = rcu_dereference_protected(
2289 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2290 dev_hold(res);
2291 spin_unlock(&pdata->netdev_lock);
2292 }
2293
2294 return res;
2295}
2296EXPORT_SYMBOL(ib_device_get_netdev);
2297
2298/**
2299 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2300 * @ndev: netdev to locate
2301 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2302 *
2303 * Find and hold an ib_device that is associated with a netdev via
2304 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2305 * returned pointer.
2306 */
2307struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2308 enum rdma_driver_id driver_id)
2309{
2310 struct ib_device *res = NULL;
2311 struct ib_port_data *cur;
2312
2313 rcu_read_lock();
2314 hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2315 (uintptr_t)ndev) {
2316 if (rcu_access_pointer(cur->netdev) == ndev &&
2317 (driver_id == RDMA_DRIVER_UNKNOWN ||
2318 cur->ib_dev->ops.driver_id == driver_id) &&
2319 ib_device_try_get(cur->ib_dev)) {
2320 res = cur->ib_dev;
2321 break;
2322 }
2323 }
2324 rcu_read_unlock();
2325
2326 return res;
2327}
2328EXPORT_SYMBOL(ib_device_get_by_netdev);
2329
2330/**
2331 * ib_enum_roce_netdev - enumerate all RoCE ports
2332 * @ib_dev : IB device we want to query
2333 * @filter: Should we call the callback?
2334 * @filter_cookie: Cookie passed to filter
2335 * @cb: Callback to call for each found RoCE ports
2336 * @cookie: Cookie passed back to the callback
2337 *
2338 * Enumerates all of the physical RoCE ports of ib_dev
2339 * which are related to netdevice and calls callback() on each
2340 * device for which filter() function returns non zero.
2341 */
2342void ib_enum_roce_netdev(struct ib_device *ib_dev,
2343 roce_netdev_filter filter,
2344 void *filter_cookie,
2345 roce_netdev_callback cb,
2346 void *cookie)
2347{
2348 u32 port;
2349
2350 rdma_for_each_port (ib_dev, port)
2351 if (rdma_protocol_roce(ib_dev, port)) {
2352 struct net_device *idev =
2353 ib_device_get_netdev(ib_dev, port);
2354
2355 if (filter(ib_dev, port, idev, filter_cookie))
2356 cb(ib_dev, port, idev, cookie);
2357 dev_put(idev);
2358 }
2359}
2360
2361/**
2362 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2363 * @filter: Should we call the callback?
2364 * @filter_cookie: Cookie passed to filter
2365 * @cb: Callback to call for each found RoCE ports
2366 * @cookie: Cookie passed back to the callback
2367 *
2368 * Enumerates all RoCE devices' physical ports which are related
2369 * to netdevices and calls callback() on each device for which
2370 * filter() function returns non zero.
2371 */
2372void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2373 void *filter_cookie,
2374 roce_netdev_callback cb,
2375 void *cookie)
2376{
2377 struct ib_device *dev;
2378 unsigned long index;
2379
2380 down_read(&devices_rwsem);
2381 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2382 ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2383 up_read(&devices_rwsem);
2384}
2385
2386/*
2387 * ib_enum_all_devs - enumerate all ib_devices
2388 * @cb: Callback to call for each found ib_device
2389 *
2390 * Enumerates all ib_devices and calls callback() on each device.
2391 */
2392int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2393 struct netlink_callback *cb)
2394{
2395 unsigned long index;
2396 struct ib_device *dev;
2397 unsigned int idx = 0;
2398 int ret = 0;
2399
2400 down_read(&devices_rwsem);
2401 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2402 if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2403 continue;
2404
2405 ret = nldev_cb(dev, skb, cb, idx);
2406 if (ret)
2407 break;
2408 idx++;
2409 }
2410 up_read(&devices_rwsem);
2411 return ret;
2412}
2413
2414/**
2415 * ib_query_pkey - Get P_Key table entry
2416 * @device:Device to query
2417 * @port_num:Port number to query
2418 * @index:P_Key table index to query
2419 * @pkey:Returned P_Key
2420 *
2421 * ib_query_pkey() fetches the specified P_Key table entry.
2422 */
2423int ib_query_pkey(struct ib_device *device,
2424 u32 port_num, u16 index, u16 *pkey)
2425{
2426 if (!rdma_is_port_valid(device, port_num))
2427 return -EINVAL;
2428
2429 if (!device->ops.query_pkey)
2430 return -EOPNOTSUPP;
2431
2432 return device->ops.query_pkey(device, port_num, index, pkey);
2433}
2434EXPORT_SYMBOL(ib_query_pkey);
2435
2436/**
2437 * ib_modify_device - Change IB device attributes
2438 * @device:Device to modify
2439 * @device_modify_mask:Mask of attributes to change
2440 * @device_modify:New attribute values
2441 *
2442 * ib_modify_device() changes a device's attributes as specified by
2443 * the @device_modify_mask and @device_modify structure.
2444 */
2445int ib_modify_device(struct ib_device *device,
2446 int device_modify_mask,
2447 struct ib_device_modify *device_modify)
2448{
2449 if (!device->ops.modify_device)
2450 return -EOPNOTSUPP;
2451
2452 return device->ops.modify_device(device, device_modify_mask,
2453 device_modify);
2454}
2455EXPORT_SYMBOL(ib_modify_device);
2456
2457/**
2458 * ib_modify_port - Modifies the attributes for the specified port.
2459 * @device: The device to modify.
2460 * @port_num: The number of the port to modify.
2461 * @port_modify_mask: Mask used to specify which attributes of the port
2462 * to change.
2463 * @port_modify: New attribute values for the port.
2464 *
2465 * ib_modify_port() changes a port's attributes as specified by the
2466 * @port_modify_mask and @port_modify structure.
2467 */
2468int ib_modify_port(struct ib_device *device,
2469 u32 port_num, int port_modify_mask,
2470 struct ib_port_modify *port_modify)
2471{
2472 int rc;
2473
2474 if (!rdma_is_port_valid(device, port_num))
2475 return -EINVAL;
2476
2477 if (device->ops.modify_port)
2478 rc = device->ops.modify_port(device, port_num,
2479 port_modify_mask,
2480 port_modify);
2481 else if (rdma_protocol_roce(device, port_num) &&
2482 ((port_modify->set_port_cap_mask & ~IB_PORT_CM_SUP) == 0 ||
2483 (port_modify->clr_port_cap_mask & ~IB_PORT_CM_SUP) == 0))
2484 rc = 0;
2485 else
2486 rc = -EOPNOTSUPP;
2487 return rc;
2488}
2489EXPORT_SYMBOL(ib_modify_port);
2490
2491/**
2492 * ib_find_gid - Returns the port number and GID table index where
2493 * a specified GID value occurs. Its searches only for IB link layer.
2494 * @device: The device to query.
2495 * @gid: The GID value to search for.
2496 * @port_num: The port number of the device where the GID value was found.
2497 * @index: The index into the GID table where the GID was found. This
2498 * parameter may be NULL.
2499 */
2500int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2501 u32 *port_num, u16 *index)
2502{
2503 union ib_gid tmp_gid;
2504 u32 port;
2505 int ret, i;
2506
2507 rdma_for_each_port (device, port) {
2508 if (!rdma_protocol_ib(device, port))
2509 continue;
2510
2511 for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2512 ++i) {
2513 ret = rdma_query_gid(device, port, i, &tmp_gid);
2514 if (ret)
2515 continue;
2516
2517 if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2518 *port_num = port;
2519 if (index)
2520 *index = i;
2521 return 0;
2522 }
2523 }
2524 }
2525
2526 return -ENOENT;
2527}
2528EXPORT_SYMBOL(ib_find_gid);
2529
2530/**
2531 * ib_find_pkey - Returns the PKey table index where a specified
2532 * PKey value occurs.
2533 * @device: The device to query.
2534 * @port_num: The port number of the device to search for the PKey.
2535 * @pkey: The PKey value to search for.
2536 * @index: The index into the PKey table where the PKey was found.
2537 */
2538int ib_find_pkey(struct ib_device *device,
2539 u32 port_num, u16 pkey, u16 *index)
2540{
2541 int ret, i;
2542 u16 tmp_pkey;
2543 int partial_ix = -1;
2544
2545 for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2546 ++i) {
2547 ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2548 if (ret)
2549 return ret;
2550 if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2551 /* if there is full-member pkey take it.*/
2552 if (tmp_pkey & 0x8000) {
2553 *index = i;
2554 return 0;
2555 }
2556 if (partial_ix < 0)
2557 partial_ix = i;
2558 }
2559 }
2560
2561 /*no full-member, if exists take the limited*/
2562 if (partial_ix >= 0) {
2563 *index = partial_ix;
2564 return 0;
2565 }
2566 return -ENOENT;
2567}
2568EXPORT_SYMBOL(ib_find_pkey);
2569
2570/**
2571 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2572 * for a received CM request
2573 * @dev: An RDMA device on which the request has been received.
2574 * @port: Port number on the RDMA device.
2575 * @pkey: The Pkey the request came on.
2576 * @gid: A GID that the net_dev uses to communicate.
2577 * @addr: Contains the IP address that the request specified as its
2578 * destination.
2579 *
2580 */
2581struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2582 u32 port,
2583 u16 pkey,
2584 const union ib_gid *gid,
2585 const struct sockaddr *addr)
2586{
2587 struct net_device *net_dev = NULL;
2588 unsigned long index;
2589 void *client_data;
2590
2591 if (!rdma_protocol_ib(dev, port))
2592 return NULL;
2593
2594 /*
2595 * Holding the read side guarantees that the client will not become
2596 * unregistered while we are calling get_net_dev_by_params()
2597 */
2598 down_read(&dev->client_data_rwsem);
2599 xan_for_each_marked (&dev->client_data, index, client_data,
2600 CLIENT_DATA_REGISTERED) {
2601 struct ib_client *client = xa_load(&clients, index);
2602
2603 if (!client || !client->get_net_dev_by_params)
2604 continue;
2605
2606 net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2607 addr, client_data);
2608 if (net_dev)
2609 break;
2610 }
2611 up_read(&dev->client_data_rwsem);
2612
2613 return net_dev;
2614}
2615EXPORT_SYMBOL(ib_get_net_dev_by_params);
2616
2617void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2618{
2619 struct ib_device_ops *dev_ops = &dev->ops;
2620#define SET_DEVICE_OP(ptr, name) \
2621 do { \
2622 if (ops->name) \
2623 if (!((ptr)->name)) \
2624 (ptr)->name = ops->name; \
2625 } while (0)
2626
2627#define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2628
2629 if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2630 WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2631 dev_ops->driver_id != ops->driver_id);
2632 dev_ops->driver_id = ops->driver_id;
2633 }
2634 if (ops->owner) {
2635 WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2636 dev_ops->owner = ops->owner;
2637 }
2638 if (ops->uverbs_abi_ver)
2639 dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2640
2641 dev_ops->uverbs_no_driver_id_binding |=
2642 ops->uverbs_no_driver_id_binding;
2643
2644 SET_DEVICE_OP(dev_ops, add_gid);
2645 SET_DEVICE_OP(dev_ops, add_sub_dev);
2646 SET_DEVICE_OP(dev_ops, advise_mr);
2647 SET_DEVICE_OP(dev_ops, alloc_dm);
2648 SET_DEVICE_OP(dev_ops, alloc_hw_device_stats);
2649 SET_DEVICE_OP(dev_ops, alloc_hw_port_stats);
2650 SET_DEVICE_OP(dev_ops, alloc_mr);
2651 SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2652 SET_DEVICE_OP(dev_ops, alloc_mw);
2653 SET_DEVICE_OP(dev_ops, alloc_pd);
2654 SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2655 SET_DEVICE_OP(dev_ops, alloc_ucontext);
2656 SET_DEVICE_OP(dev_ops, alloc_xrcd);
2657 SET_DEVICE_OP(dev_ops, attach_mcast);
2658 SET_DEVICE_OP(dev_ops, check_mr_status);
2659 SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2660 SET_DEVICE_OP(dev_ops, counter_bind_qp);
2661 SET_DEVICE_OP(dev_ops, counter_dealloc);
2662 SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2663 SET_DEVICE_OP(dev_ops, counter_update_stats);
2664 SET_DEVICE_OP(dev_ops, create_ah);
2665 SET_DEVICE_OP(dev_ops, create_counters);
2666 SET_DEVICE_OP(dev_ops, create_cq);
2667 SET_DEVICE_OP(dev_ops, create_flow);
2668 SET_DEVICE_OP(dev_ops, create_qp);
2669 SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2670 SET_DEVICE_OP(dev_ops, create_srq);
2671 SET_DEVICE_OP(dev_ops, create_user_ah);
2672 SET_DEVICE_OP(dev_ops, create_wq);
2673 SET_DEVICE_OP(dev_ops, dealloc_dm);
2674 SET_DEVICE_OP(dev_ops, dealloc_driver);
2675 SET_DEVICE_OP(dev_ops, dealloc_mw);
2676 SET_DEVICE_OP(dev_ops, dealloc_pd);
2677 SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2678 SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2679 SET_DEVICE_OP(dev_ops, del_gid);
2680 SET_DEVICE_OP(dev_ops, del_sub_dev);
2681 SET_DEVICE_OP(dev_ops, dereg_mr);
2682 SET_DEVICE_OP(dev_ops, destroy_ah);
2683 SET_DEVICE_OP(dev_ops, destroy_counters);
2684 SET_DEVICE_OP(dev_ops, destroy_cq);
2685 SET_DEVICE_OP(dev_ops, destroy_flow);
2686 SET_DEVICE_OP(dev_ops, destroy_flow_action);
2687 SET_DEVICE_OP(dev_ops, destroy_qp);
2688 SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2689 SET_DEVICE_OP(dev_ops, destroy_srq);
2690 SET_DEVICE_OP(dev_ops, destroy_wq);
2691 SET_DEVICE_OP(dev_ops, device_group);
2692 SET_DEVICE_OP(dev_ops, detach_mcast);
2693 SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2694 SET_DEVICE_OP(dev_ops, drain_rq);
2695 SET_DEVICE_OP(dev_ops, drain_sq);
2696 SET_DEVICE_OP(dev_ops, enable_driver);
2697 SET_DEVICE_OP(dev_ops, fill_res_cm_id_entry);
2698 SET_DEVICE_OP(dev_ops, fill_res_cq_entry);
2699 SET_DEVICE_OP(dev_ops, fill_res_cq_entry_raw);
2700 SET_DEVICE_OP(dev_ops, fill_res_mr_entry);
2701 SET_DEVICE_OP(dev_ops, fill_res_mr_entry_raw);
2702 SET_DEVICE_OP(dev_ops, fill_res_qp_entry);
2703 SET_DEVICE_OP(dev_ops, fill_res_qp_entry_raw);
2704 SET_DEVICE_OP(dev_ops, fill_res_srq_entry);
2705 SET_DEVICE_OP(dev_ops, fill_res_srq_entry_raw);
2706 SET_DEVICE_OP(dev_ops, fill_stat_mr_entry);
2707 SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2708 SET_DEVICE_OP(dev_ops, get_dma_mr);
2709 SET_DEVICE_OP(dev_ops, get_hw_stats);
2710 SET_DEVICE_OP(dev_ops, get_link_layer);
2711 SET_DEVICE_OP(dev_ops, get_netdev);
2712 SET_DEVICE_OP(dev_ops, get_numa_node);
2713 SET_DEVICE_OP(dev_ops, get_port_immutable);
2714 SET_DEVICE_OP(dev_ops, get_vector_affinity);
2715 SET_DEVICE_OP(dev_ops, get_vf_config);
2716 SET_DEVICE_OP(dev_ops, get_vf_guid);
2717 SET_DEVICE_OP(dev_ops, get_vf_stats);
2718 SET_DEVICE_OP(dev_ops, iw_accept);
2719 SET_DEVICE_OP(dev_ops, iw_add_ref);
2720 SET_DEVICE_OP(dev_ops, iw_connect);
2721 SET_DEVICE_OP(dev_ops, iw_create_listen);
2722 SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2723 SET_DEVICE_OP(dev_ops, iw_get_qp);
2724 SET_DEVICE_OP(dev_ops, iw_reject);
2725 SET_DEVICE_OP(dev_ops, iw_rem_ref);
2726 SET_DEVICE_OP(dev_ops, map_mr_sg);
2727 SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
2728 SET_DEVICE_OP(dev_ops, mmap);
2729 SET_DEVICE_OP(dev_ops, mmap_free);
2730 SET_DEVICE_OP(dev_ops, modify_ah);
2731 SET_DEVICE_OP(dev_ops, modify_cq);
2732 SET_DEVICE_OP(dev_ops, modify_device);
2733 SET_DEVICE_OP(dev_ops, modify_hw_stat);
2734 SET_DEVICE_OP(dev_ops, modify_port);
2735 SET_DEVICE_OP(dev_ops, modify_qp);
2736 SET_DEVICE_OP(dev_ops, modify_srq);
2737 SET_DEVICE_OP(dev_ops, modify_wq);
2738 SET_DEVICE_OP(dev_ops, peek_cq);
2739 SET_DEVICE_OP(dev_ops, poll_cq);
2740 SET_DEVICE_OP(dev_ops, port_groups);
2741 SET_DEVICE_OP(dev_ops, post_recv);
2742 SET_DEVICE_OP(dev_ops, post_send);
2743 SET_DEVICE_OP(dev_ops, post_srq_recv);
2744 SET_DEVICE_OP(dev_ops, process_mad);
2745 SET_DEVICE_OP(dev_ops, query_ah);
2746 SET_DEVICE_OP(dev_ops, query_device);
2747 SET_DEVICE_OP(dev_ops, query_gid);
2748 SET_DEVICE_OP(dev_ops, query_pkey);
2749 SET_DEVICE_OP(dev_ops, query_port);
2750 SET_DEVICE_OP(dev_ops, query_qp);
2751 SET_DEVICE_OP(dev_ops, query_srq);
2752 SET_DEVICE_OP(dev_ops, query_ucontext);
2753 SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2754 SET_DEVICE_OP(dev_ops, read_counters);
2755 SET_DEVICE_OP(dev_ops, reg_dm_mr);
2756 SET_DEVICE_OP(dev_ops, reg_user_mr);
2757 SET_DEVICE_OP(dev_ops, reg_user_mr_dmabuf);
2758 SET_DEVICE_OP(dev_ops, req_notify_cq);
2759 SET_DEVICE_OP(dev_ops, rereg_user_mr);
2760 SET_DEVICE_OP(dev_ops, resize_cq);
2761 SET_DEVICE_OP(dev_ops, set_vf_guid);
2762 SET_DEVICE_OP(dev_ops, set_vf_link_state);
2763 SET_DEVICE_OP(dev_ops, ufile_hw_cleanup);
2764
2765 SET_OBJ_SIZE(dev_ops, ib_ah);
2766 SET_OBJ_SIZE(dev_ops, ib_counters);
2767 SET_OBJ_SIZE(dev_ops, ib_cq);
2768 SET_OBJ_SIZE(dev_ops, ib_mw);
2769 SET_OBJ_SIZE(dev_ops, ib_pd);
2770 SET_OBJ_SIZE(dev_ops, ib_qp);
2771 SET_OBJ_SIZE(dev_ops, ib_rwq_ind_table);
2772 SET_OBJ_SIZE(dev_ops, ib_srq);
2773 SET_OBJ_SIZE(dev_ops, ib_ucontext);
2774 SET_OBJ_SIZE(dev_ops, ib_xrcd);
2775}
2776EXPORT_SYMBOL(ib_set_device_ops);
2777
2778int ib_add_sub_device(struct ib_device *parent,
2779 enum rdma_nl_dev_type type,
2780 const char *name)
2781{
2782 struct ib_device *sub;
2783 int ret = 0;
2784
2785 if (!parent->ops.add_sub_dev || !parent->ops.del_sub_dev)
2786 return -EOPNOTSUPP;
2787
2788 if (!ib_device_try_get(parent))
2789 return -EINVAL;
2790
2791 sub = parent->ops.add_sub_dev(parent, type, name);
2792 if (IS_ERR(sub)) {
2793 ib_device_put(parent);
2794 return PTR_ERR(sub);
2795 }
2796
2797 sub->type = type;
2798 sub->parent = parent;
2799
2800 mutex_lock(&parent->subdev_lock);
2801 list_add_tail(&parent->subdev_list_head, &sub->subdev_list);
2802 mutex_unlock(&parent->subdev_lock);
2803
2804 return ret;
2805}
2806EXPORT_SYMBOL(ib_add_sub_device);
2807
2808int ib_del_sub_device_and_put(struct ib_device *sub)
2809{
2810 struct ib_device *parent = sub->parent;
2811
2812 if (!parent)
2813 return -EOPNOTSUPP;
2814
2815 mutex_lock(&parent->subdev_lock);
2816 list_del(&sub->subdev_list);
2817 mutex_unlock(&parent->subdev_lock);
2818
2819 ib_device_put(sub);
2820 parent->ops.del_sub_dev(sub);
2821 ib_device_put(parent);
2822
2823 return 0;
2824}
2825EXPORT_SYMBOL(ib_del_sub_device_and_put);
2826
2827#ifdef CONFIG_INFINIBAND_VIRT_DMA
2828int ib_dma_virt_map_sg(struct ib_device *dev, struct scatterlist *sg, int nents)
2829{
2830 struct scatterlist *s;
2831 int i;
2832
2833 for_each_sg(sg, s, nents, i) {
2834 sg_dma_address(s) = (uintptr_t)sg_virt(s);
2835 sg_dma_len(s) = s->length;
2836 }
2837 return nents;
2838}
2839EXPORT_SYMBOL(ib_dma_virt_map_sg);
2840#endif /* CONFIG_INFINIBAND_VIRT_DMA */
2841
2842static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2843 [RDMA_NL_LS_OP_RESOLVE] = {
2844 .doit = ib_nl_handle_resolve_resp,
2845 .flags = RDMA_NL_ADMIN_PERM,
2846 },
2847 [RDMA_NL_LS_OP_SET_TIMEOUT] = {
2848 .doit = ib_nl_handle_set_timeout,
2849 .flags = RDMA_NL_ADMIN_PERM,
2850 },
2851 [RDMA_NL_LS_OP_IP_RESOLVE] = {
2852 .doit = ib_nl_handle_ip_res_resp,
2853 .flags = RDMA_NL_ADMIN_PERM,
2854 },
2855};
2856
2857static int ib_netdevice_event(struct notifier_block *this,
2858 unsigned long event, void *ptr)
2859{
2860 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
2861 struct net_device *ib_ndev;
2862 struct ib_device *ibdev;
2863 u32 port;
2864
2865 switch (event) {
2866 case NETDEV_CHANGENAME:
2867 ibdev = ib_device_get_by_netdev(ndev, RDMA_DRIVER_UNKNOWN);
2868 if (!ibdev)
2869 return NOTIFY_DONE;
2870
2871 rdma_for_each_port(ibdev, port) {
2872 ib_ndev = ib_device_get_netdev(ibdev, port);
2873 if (ndev == ib_ndev)
2874 rdma_nl_notify_event(ibdev, port,
2875 RDMA_NETDEV_RENAME_EVENT);
2876 dev_put(ib_ndev);
2877 }
2878 ib_device_put(ibdev);
2879 break;
2880 default:
2881 break;
2882 }
2883
2884 return NOTIFY_DONE;
2885}
2886
2887static struct notifier_block nb_netdevice = {
2888 .notifier_call = ib_netdevice_event,
2889};
2890
2891static int __init ib_core_init(void)
2892{
2893 int ret = -ENOMEM;
2894
2895 ib_wq = alloc_workqueue("infiniband", 0, 0);
2896 if (!ib_wq)
2897 return -ENOMEM;
2898
2899 ib_unreg_wq = alloc_workqueue("ib-unreg-wq", WQ_UNBOUND,
2900 WQ_UNBOUND_MAX_ACTIVE);
2901 if (!ib_unreg_wq)
2902 goto err;
2903
2904 ib_comp_wq = alloc_workqueue("ib-comp-wq",
2905 WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2906 if (!ib_comp_wq)
2907 goto err_unbound;
2908
2909 ib_comp_unbound_wq =
2910 alloc_workqueue("ib-comp-unb-wq",
2911 WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2912 WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2913 if (!ib_comp_unbound_wq)
2914 goto err_comp;
2915
2916 ret = class_register(&ib_class);
2917 if (ret) {
2918 pr_warn("Couldn't create InfiniBand device class\n");
2919 goto err_comp_unbound;
2920 }
2921
2922 rdma_nl_init();
2923
2924 ret = addr_init();
2925 if (ret) {
2926 pr_warn("Couldn't init IB address resolution\n");
2927 goto err_ibnl;
2928 }
2929
2930 ret = ib_mad_init();
2931 if (ret) {
2932 pr_warn("Couldn't init IB MAD\n");
2933 goto err_addr;
2934 }
2935
2936 ret = ib_sa_init();
2937 if (ret) {
2938 pr_warn("Couldn't init SA\n");
2939 goto err_mad;
2940 }
2941
2942 ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2943 if (ret) {
2944 pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2945 goto err_sa;
2946 }
2947
2948 ret = register_pernet_device(&rdma_dev_net_ops);
2949 if (ret) {
2950 pr_warn("Couldn't init compat dev. ret %d\n", ret);
2951 goto err_compat;
2952 }
2953
2954 nldev_init();
2955 rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2956 ret = roce_gid_mgmt_init();
2957 if (ret) {
2958 pr_warn("Couldn't init RoCE GID management\n");
2959 goto err_parent;
2960 }
2961
2962 register_netdevice_notifier(&nb_netdevice);
2963
2964 return 0;
2965
2966err_parent:
2967 rdma_nl_unregister(RDMA_NL_LS);
2968 nldev_exit();
2969 unregister_pernet_device(&rdma_dev_net_ops);
2970err_compat:
2971 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2972err_sa:
2973 ib_sa_cleanup();
2974err_mad:
2975 ib_mad_cleanup();
2976err_addr:
2977 addr_cleanup();
2978err_ibnl:
2979 class_unregister(&ib_class);
2980err_comp_unbound:
2981 destroy_workqueue(ib_comp_unbound_wq);
2982err_comp:
2983 destroy_workqueue(ib_comp_wq);
2984err_unbound:
2985 destroy_workqueue(ib_unreg_wq);
2986err:
2987 destroy_workqueue(ib_wq);
2988 return ret;
2989}
2990
2991static void __exit ib_core_cleanup(void)
2992{
2993 unregister_netdevice_notifier(&nb_netdevice);
2994 roce_gid_mgmt_cleanup();
2995 rdma_nl_unregister(RDMA_NL_LS);
2996 nldev_exit();
2997 unregister_pernet_device(&rdma_dev_net_ops);
2998 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2999 ib_sa_cleanup();
3000 ib_mad_cleanup();
3001 addr_cleanup();
3002 rdma_nl_exit();
3003 class_unregister(&ib_class);
3004 destroy_workqueue(ib_comp_unbound_wq);
3005 destroy_workqueue(ib_comp_wq);
3006 /* Make sure that any pending umem accounting work is done. */
3007 destroy_workqueue(ib_wq);
3008 destroy_workqueue(ib_unreg_wq);
3009 WARN_ON(!xa_empty(&clients));
3010 WARN_ON(!xa_empty(&devices));
3011}
3012
3013MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
3014
3015/* ib core relies on netdev stack to first register net_ns_type_operations
3016 * ns kobject type before ib_core initialization.
3017 */
3018fs_initcall(ib_core_init);
3019module_exit(ib_core_cleanup);