2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
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:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
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.
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
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>
50 #include "core_priv.h"
53 MODULE_AUTHOR("Roland Dreier");
54 MODULE_DESCRIPTION("core kernel InfiniBand API");
55 MODULE_LICENSE("Dual BSD/GPL");
57 struct workqueue_struct *ib_comp_wq;
58 struct workqueue_struct *ib_comp_unbound_wq;
59 struct workqueue_struct *ib_wq;
60 EXPORT_SYMBOL_GPL(ib_wq);
63 * Each of the three rwsem locks (devices, clients, client_data) protects the
64 * xarray of the same name. Specifically it allows the caller to assert that
65 * the MARK will/will not be changing under the lock, and for devices and
66 * clients, that the value in the xarray is still a valid pointer. Change of
67 * the MARK is linked to the object state, so holding the lock and testing the
68 * MARK also asserts that the contained object is in a certain state.
70 * This is used to build a two stage register/unregister flow where objects
71 * can continue to be in the xarray even though they are still in progress to
72 * register/unregister.
74 * The xarray itself provides additional locking, and restartable iteration,
75 * which is also relied on.
77 * Locks should not be nested, with the exception of client_data, which is
78 * allowed to nest under the read side of the other two locks.
80 * The devices_rwsem also protects the device name list, any change or
81 * assignment of device name must also hold the write side to guarantee unique
86 * devices contains devices that have had their names assigned. The
87 * devices may not be registered. Users that care about the registration
88 * status need to call ib_device_try_get() on the device to ensure it is
89 * registered, and keep it registered, for the required duration.
92 static DEFINE_XARRAY_FLAGS(devices, XA_FLAGS_ALLOC);
93 static DECLARE_RWSEM(devices_rwsem);
94 #define DEVICE_REGISTERED XA_MARK_1
96 static u32 highest_client_id;
97 #define CLIENT_REGISTERED XA_MARK_1
98 static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
99 static DECLARE_RWSEM(clients_rwsem);
101 static void ib_client_put(struct ib_client *client)
103 if (refcount_dec_and_test(&client->uses))
104 complete(&client->uses_zero);
108 * If client_data is registered then the corresponding client must also still
111 #define CLIENT_DATA_REGISTERED XA_MARK_1
113 unsigned int rdma_dev_net_id;
116 * A list of net namespaces is maintained in an xarray. This is necessary
117 * because we can't get the locking right using the existing net ns list. We
118 * would require a init_net callback after the list is updated.
120 static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
122 * rwsem to protect accessing the rdma_nets xarray entries.
124 static DECLARE_RWSEM(rdma_nets_rwsem);
126 bool ib_devices_shared_netns = true;
127 module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
128 MODULE_PARM_DESC(netns_mode,
129 "Share device among net namespaces; default=1 (shared)");
131 * rdma_dev_access_netns() - Return whether an rdma device can be accessed
132 * from a specified net namespace or not.
133 * @dev: Pointer to rdma device which needs to be checked
134 * @net: Pointer to net namesapce for which access to be checked
136 * When the rdma device is in shared mode, it ignores the net namespace.
137 * When the rdma device is exclusive to a net namespace, rdma device net
138 * namespace is checked against the specified one.
140 bool rdma_dev_access_netns(const struct ib_device *dev, const struct net *net)
142 return (ib_devices_shared_netns ||
143 net_eq(read_pnet(&dev->coredev.rdma_net), net));
145 EXPORT_SYMBOL(rdma_dev_access_netns);
148 * xarray has this behavior where it won't iterate over NULL values stored in
149 * allocated arrays. So we need our own iterator to see all values stored in
150 * the array. This does the same thing as xa_for_each except that it also
151 * returns NULL valued entries if the array is allocating. Simplified to only
152 * work on simple xarrays.
154 static void *xan_find_marked(struct xarray *xa, unsigned long *indexp,
157 XA_STATE(xas, xa, *indexp);
162 entry = xas_find_marked(&xas, ULONG_MAX, filter);
163 if (xa_is_zero(entry))
165 } while (xas_retry(&xas, entry));
169 *indexp = xas.xa_index;
170 if (xa_is_zero(entry))
174 return XA_ERROR(-ENOENT);
176 #define xan_for_each_marked(xa, index, entry, filter) \
177 for (index = 0, entry = xan_find_marked(xa, &(index), filter); \
179 (index)++, entry = xan_find_marked(xa, &(index), filter))
181 /* RCU hash table mapping netdevice pointers to struct ib_port_data */
182 static DEFINE_SPINLOCK(ndev_hash_lock);
183 static DECLARE_HASHTABLE(ndev_hash, 5);
185 static void free_netdevs(struct ib_device *ib_dev);
186 static void ib_unregister_work(struct work_struct *work);
187 static void __ib_unregister_device(struct ib_device *device);
188 static int ib_security_change(struct notifier_block *nb, unsigned long event,
190 static void ib_policy_change_task(struct work_struct *work);
191 static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
193 static void __ibdev_printk(const char *level, const struct ib_device *ibdev,
194 struct va_format *vaf)
196 if (ibdev && ibdev->dev.parent)
197 dev_printk_emit(level[1] - '0',
200 dev_driver_string(ibdev->dev.parent),
201 dev_name(ibdev->dev.parent),
202 dev_name(&ibdev->dev),
206 level, dev_name(&ibdev->dev), vaf);
208 printk("%s(NULL ib_device): %pV", level, vaf);
211 void ibdev_printk(const char *level, const struct ib_device *ibdev,
212 const char *format, ...)
214 struct va_format vaf;
217 va_start(args, format);
222 __ibdev_printk(level, ibdev, &vaf);
226 EXPORT_SYMBOL(ibdev_printk);
228 #define define_ibdev_printk_level(func, level) \
229 void func(const struct ib_device *ibdev, const char *fmt, ...) \
231 struct va_format vaf; \
234 va_start(args, fmt); \
239 __ibdev_printk(level, ibdev, &vaf); \
245 define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
246 define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
247 define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
248 define_ibdev_printk_level(ibdev_err, KERN_ERR);
249 define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
250 define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
251 define_ibdev_printk_level(ibdev_info, KERN_INFO);
253 static struct notifier_block ibdev_lsm_nb = {
254 .notifier_call = ib_security_change,
257 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
260 /* Pointer to the RCU head at the start of the ib_port_data array */
261 struct ib_port_data_rcu {
262 struct rcu_head rcu_head;
263 struct ib_port_data pdata[];
266 static void ib_device_check_mandatory(struct ib_device *device)
268 #define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
269 static const struct {
272 } mandatory_table[] = {
273 IB_MANDATORY_FUNC(query_device),
274 IB_MANDATORY_FUNC(query_port),
275 IB_MANDATORY_FUNC(alloc_pd),
276 IB_MANDATORY_FUNC(dealloc_pd),
277 IB_MANDATORY_FUNC(create_qp),
278 IB_MANDATORY_FUNC(modify_qp),
279 IB_MANDATORY_FUNC(destroy_qp),
280 IB_MANDATORY_FUNC(post_send),
281 IB_MANDATORY_FUNC(post_recv),
282 IB_MANDATORY_FUNC(create_cq),
283 IB_MANDATORY_FUNC(destroy_cq),
284 IB_MANDATORY_FUNC(poll_cq),
285 IB_MANDATORY_FUNC(req_notify_cq),
286 IB_MANDATORY_FUNC(get_dma_mr),
287 IB_MANDATORY_FUNC(dereg_mr),
288 IB_MANDATORY_FUNC(get_port_immutable)
292 device->kverbs_provider = true;
293 for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
294 if (!*(void **) ((void *) &device->ops +
295 mandatory_table[i].offset)) {
296 device->kverbs_provider = false;
303 * Caller must perform ib_device_put() to return the device reference count
304 * when ib_device_get_by_index() returns valid device pointer.
306 struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
308 struct ib_device *device;
310 down_read(&devices_rwsem);
311 device = xa_load(&devices, index);
313 if (!rdma_dev_access_netns(device, net)) {
318 if (!ib_device_try_get(device))
322 up_read(&devices_rwsem);
327 * ib_device_put - Release IB device reference
328 * @device: device whose reference to be released
330 * ib_device_put() releases reference to the IB device to allow it to be
331 * unregistered and eventually free.
333 void ib_device_put(struct ib_device *device)
335 if (refcount_dec_and_test(&device->refcount))
336 complete(&device->unreg_completion);
338 EXPORT_SYMBOL(ib_device_put);
340 static struct ib_device *__ib_device_get_by_name(const char *name)
342 struct ib_device *device;
345 xa_for_each (&devices, index, device)
346 if (!strcmp(name, dev_name(&device->dev)))
353 * ib_device_get_by_name - Find an IB device by name
354 * @name: The name to look for
355 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
357 * Find and hold an ib_device by its name. The caller must call
358 * ib_device_put() on the returned pointer.
360 struct ib_device *ib_device_get_by_name(const char *name,
361 enum rdma_driver_id driver_id)
363 struct ib_device *device;
365 down_read(&devices_rwsem);
366 device = __ib_device_get_by_name(name);
367 if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
368 device->ops.driver_id != driver_id)
372 if (!ib_device_try_get(device))
375 up_read(&devices_rwsem);
378 EXPORT_SYMBOL(ib_device_get_by_name);
380 static int rename_compat_devs(struct ib_device *device)
382 struct ib_core_device *cdev;
386 mutex_lock(&device->compat_devs_mutex);
387 xa_for_each (&device->compat_devs, index, cdev) {
388 ret = device_rename(&cdev->dev, dev_name(&device->dev));
391 "Fail to rename compatdev to new name %s\n",
392 dev_name(&device->dev));
396 mutex_unlock(&device->compat_devs_mutex);
400 int ib_device_rename(struct ib_device *ibdev, const char *name)
406 down_write(&devices_rwsem);
407 if (!strcmp(name, dev_name(&ibdev->dev))) {
408 up_write(&devices_rwsem);
412 if (__ib_device_get_by_name(name)) {
413 up_write(&devices_rwsem);
417 ret = device_rename(&ibdev->dev, name);
419 up_write(&devices_rwsem);
423 strlcpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
424 ret = rename_compat_devs(ibdev);
426 downgrade_write(&devices_rwsem);
427 down_read(&ibdev->client_data_rwsem);
428 xan_for_each_marked(&ibdev->client_data, index, client_data,
429 CLIENT_DATA_REGISTERED) {
430 struct ib_client *client = xa_load(&clients, index);
432 if (!client || !client->rename)
435 client->rename(ibdev, client_data);
437 up_read(&ibdev->client_data_rwsem);
438 up_read(&devices_rwsem);
442 int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
446 ibdev->use_cq_dim = use_dim;
451 static int alloc_name(struct ib_device *ibdev, const char *name)
453 struct ib_device *device;
459 lockdep_assert_held_write(&devices_rwsem);
461 xa_for_each (&devices, index, device) {
462 char buf[IB_DEVICE_NAME_MAX];
464 if (sscanf(dev_name(&device->dev), name, &i) != 1)
466 if (i < 0 || i >= INT_MAX)
468 snprintf(buf, sizeof buf, name, i);
469 if (strcmp(buf, dev_name(&device->dev)) != 0)
472 rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
477 rc = ida_alloc(&inuse, GFP_KERNEL);
481 rc = dev_set_name(&ibdev->dev, name, rc);
487 static void ib_device_release(struct device *device)
489 struct ib_device *dev = container_of(device, struct ib_device, dev);
492 WARN_ON(refcount_read(&dev->refcount));
493 if (dev->port_data) {
494 ib_cache_release_one(dev);
495 ib_security_release_port_pkey_list(dev);
496 rdma_counter_release(dev);
497 kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
502 mutex_destroy(&dev->unregistration_lock);
503 mutex_destroy(&dev->compat_devs_mutex);
505 xa_destroy(&dev->compat_devs);
506 xa_destroy(&dev->client_data);
507 kfree_rcu(dev, rcu_head);
510 static int ib_device_uevent(struct device *device,
511 struct kobj_uevent_env *env)
513 if (add_uevent_var(env, "NAME=%s", dev_name(device)))
517 * It would be nice to pass the node GUID with the event...
523 static const void *net_namespace(struct device *d)
525 struct ib_core_device *coredev =
526 container_of(d, struct ib_core_device, dev);
528 return read_pnet(&coredev->rdma_net);
531 static struct class ib_class = {
532 .name = "infiniband",
533 .dev_release = ib_device_release,
534 .dev_uevent = ib_device_uevent,
535 .ns_type = &net_ns_type_operations,
536 .namespace = net_namespace,
539 static void rdma_init_coredev(struct ib_core_device *coredev,
540 struct ib_device *dev, struct net *net)
542 /* This BUILD_BUG_ON is intended to catch layout change
543 * of union of ib_core_device and device.
544 * dev must be the first element as ib_core and providers
545 * driver uses it. Adding anything in ib_core_device before
546 * device will break this assumption.
548 BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
549 offsetof(struct ib_device, dev));
551 coredev->dev.class = &ib_class;
552 coredev->dev.groups = dev->groups;
553 device_initialize(&coredev->dev);
554 coredev->owner = dev;
555 INIT_LIST_HEAD(&coredev->port_list);
556 write_pnet(&coredev->rdma_net, net);
560 * _ib_alloc_device - allocate an IB device struct
561 * @size:size of structure to allocate
563 * Low-level drivers should use ib_alloc_device() to allocate &struct
564 * ib_device. @size is the size of the structure to be allocated,
565 * including any private data used by the low-level driver.
566 * ib_dealloc_device() must be used to free structures allocated with
569 struct ib_device *_ib_alloc_device(size_t size)
571 struct ib_device *device;
573 if (WARN_ON(size < sizeof(struct ib_device)))
576 device = kzalloc(size, GFP_KERNEL);
580 if (rdma_restrack_init(device)) {
585 device->groups[0] = &ib_dev_attr_group;
586 rdma_init_coredev(&device->coredev, device, &init_net);
588 INIT_LIST_HEAD(&device->event_handler_list);
589 spin_lock_init(&device->qp_open_list_lock);
590 init_rwsem(&device->event_handler_rwsem);
591 mutex_init(&device->unregistration_lock);
593 * client_data needs to be alloc because we don't want our mark to be
594 * destroyed if the user stores NULL in the client data.
596 xa_init_flags(&device->client_data, XA_FLAGS_ALLOC);
597 init_rwsem(&device->client_data_rwsem);
598 xa_init_flags(&device->compat_devs, XA_FLAGS_ALLOC);
599 mutex_init(&device->compat_devs_mutex);
600 init_completion(&device->unreg_completion);
601 INIT_WORK(&device->unregistration_work, ib_unregister_work);
605 EXPORT_SYMBOL(_ib_alloc_device);
608 * ib_dealloc_device - free an IB device struct
609 * @device:structure to free
611 * Free a structure allocated with ib_alloc_device().
613 void ib_dealloc_device(struct ib_device *device)
615 if (device->ops.dealloc_driver)
616 device->ops.dealloc_driver(device);
619 * ib_unregister_driver() requires all devices to remain in the xarray
620 * while their ops are callable. The last op we call is dealloc_driver
621 * above. This is needed to create a fence on op callbacks prior to
622 * allowing the driver module to unload.
624 down_write(&devices_rwsem);
625 if (xa_load(&devices, device->index) == device)
626 xa_erase(&devices, device->index);
627 up_write(&devices_rwsem);
629 /* Expedite releasing netdev references */
630 free_netdevs(device);
632 WARN_ON(!xa_empty(&device->compat_devs));
633 WARN_ON(!xa_empty(&device->client_data));
634 WARN_ON(refcount_read(&device->refcount));
635 rdma_restrack_clean(device);
636 /* Balances with device_initialize */
637 put_device(&device->dev);
639 EXPORT_SYMBOL(ib_dealloc_device);
642 * add_client_context() and remove_client_context() must be safe against
643 * parallel calls on the same device - registration/unregistration of both the
644 * device and client can be occurring in parallel.
646 * The routines need to be a fence, any caller must not return until the add
647 * or remove is fully completed.
649 static int add_client_context(struct ib_device *device,
650 struct ib_client *client)
654 if (!device->kverbs_provider && !client->no_kverbs_req)
657 down_write(&device->client_data_rwsem);
659 * So long as the client is registered hold both the client and device
660 * unregistration locks.
662 if (!refcount_inc_not_zero(&client->uses))
664 refcount_inc(&device->refcount);
667 * Another caller to add_client_context got here first and has already
668 * completely initialized context.
670 if (xa_get_mark(&device->client_data, client->client_id,
671 CLIENT_DATA_REGISTERED))
674 ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
678 downgrade_write(&device->client_data_rwsem);
680 if (client->add(device)) {
682 * If a client fails to add then the error code is
683 * ignored, but we won't call any more ops on this
686 xa_erase(&device->client_data, client->client_id);
687 up_read(&device->client_data_rwsem);
688 ib_device_put(device);
689 ib_client_put(client);
694 /* Readers shall not see a client until add has been completed */
695 xa_set_mark(&device->client_data, client->client_id,
696 CLIENT_DATA_REGISTERED);
697 up_read(&device->client_data_rwsem);
701 ib_device_put(device);
702 ib_client_put(client);
704 up_write(&device->client_data_rwsem);
708 static void remove_client_context(struct ib_device *device,
709 unsigned int client_id)
711 struct ib_client *client;
714 down_write(&device->client_data_rwsem);
715 if (!xa_get_mark(&device->client_data, client_id,
716 CLIENT_DATA_REGISTERED)) {
717 up_write(&device->client_data_rwsem);
720 client_data = xa_load(&device->client_data, client_id);
721 xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
722 client = xa_load(&clients, client_id);
723 up_write(&device->client_data_rwsem);
726 * Notice we cannot be holding any exclusive locks when calling the
727 * remove callback as the remove callback can recurse back into any
728 * public functions in this module and thus try for any locks those
731 * For this reason clients and drivers should not call the
732 * unregistration functions will holdling any locks.
735 client->remove(device, client_data);
737 xa_erase(&device->client_data, client_id);
738 ib_device_put(device);
739 ib_client_put(client);
742 static int alloc_port_data(struct ib_device *device)
744 struct ib_port_data_rcu *pdata_rcu;
747 if (device->port_data)
750 /* This can only be called once the physical port range is defined */
751 if (WARN_ON(!device->phys_port_cnt))
755 * device->port_data is indexed directly by the port number to make
756 * access to this data as efficient as possible.
758 * Therefore port_data is declared as a 1 based array with potential
759 * empty slots at the beginning.
761 pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
762 rdma_end_port(device) + 1),
767 * The rcu_head is put in front of the port data array and the stored
768 * pointer is adjusted since we never need to see that member until
771 device->port_data = pdata_rcu->pdata;
773 rdma_for_each_port (device, port) {
774 struct ib_port_data *pdata = &device->port_data[port];
776 pdata->ib_dev = device;
777 spin_lock_init(&pdata->pkey_list_lock);
778 INIT_LIST_HEAD(&pdata->pkey_list);
779 spin_lock_init(&pdata->netdev_lock);
780 INIT_HLIST_NODE(&pdata->ndev_hash_link);
785 static int verify_immutable(const struct ib_device *dev, u8 port)
787 return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
788 rdma_max_mad_size(dev, port) != 0);
791 static int setup_port_data(struct ib_device *device)
796 ret = alloc_port_data(device);
800 rdma_for_each_port (device, port) {
801 struct ib_port_data *pdata = &device->port_data[port];
803 ret = device->ops.get_port_immutable(device, port,
808 if (verify_immutable(device, port))
814 void ib_get_device_fw_str(struct ib_device *dev, char *str)
816 if (dev->ops.get_dev_fw_str)
817 dev->ops.get_dev_fw_str(dev, str);
821 EXPORT_SYMBOL(ib_get_device_fw_str);
823 static void ib_policy_change_task(struct work_struct *work)
825 struct ib_device *dev;
828 down_read(&devices_rwsem);
829 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
832 rdma_for_each_port (dev, i) {
834 int ret = ib_get_cached_subnet_prefix(dev,
839 "ib_get_cached_subnet_prefix err: %d, this should never happen here\n",
842 ib_security_cache_change(dev, i, sp);
845 up_read(&devices_rwsem);
848 static int ib_security_change(struct notifier_block *nb, unsigned long event,
851 if (event != LSM_POLICY_CHANGE)
854 schedule_work(&ib_policy_change_work);
855 ib_mad_agent_security_change();
860 static void compatdev_release(struct device *dev)
862 struct ib_core_device *cdev =
863 container_of(dev, struct ib_core_device, dev);
868 static int add_one_compat_dev(struct ib_device *device,
869 struct rdma_dev_net *rnet)
871 struct ib_core_device *cdev;
874 lockdep_assert_held(&rdma_nets_rwsem);
875 if (!ib_devices_shared_netns)
879 * Create and add compat device in all namespaces other than where it
880 * is currently bound to.
882 if (net_eq(read_pnet(&rnet->net),
883 read_pnet(&device->coredev.rdma_net)))
887 * The first of init_net() or ib_register_device() to take the
888 * compat_devs_mutex wins and gets to add the device. Others will wait
889 * for completion here.
891 mutex_lock(&device->compat_devs_mutex);
892 cdev = xa_load(&device->compat_devs, rnet->id);
897 ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
901 cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
907 cdev->dev.parent = device->dev.parent;
908 rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
909 cdev->dev.release = compatdev_release;
910 ret = dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
914 ret = device_add(&cdev->dev);
917 ret = ib_setup_port_attrs(cdev);
921 ret = xa_err(xa_store(&device->compat_devs, rnet->id,
926 mutex_unlock(&device->compat_devs_mutex);
930 ib_free_port_attrs(cdev);
932 device_del(&cdev->dev);
934 put_device(&cdev->dev);
936 xa_release(&device->compat_devs, rnet->id);
938 mutex_unlock(&device->compat_devs_mutex);
942 static void remove_one_compat_dev(struct ib_device *device, u32 id)
944 struct ib_core_device *cdev;
946 mutex_lock(&device->compat_devs_mutex);
947 cdev = xa_erase(&device->compat_devs, id);
948 mutex_unlock(&device->compat_devs_mutex);
950 ib_free_port_attrs(cdev);
951 device_del(&cdev->dev);
952 put_device(&cdev->dev);
956 static void remove_compat_devs(struct ib_device *device)
958 struct ib_core_device *cdev;
961 xa_for_each (&device->compat_devs, index, cdev)
962 remove_one_compat_dev(device, index);
965 static int add_compat_devs(struct ib_device *device)
967 struct rdma_dev_net *rnet;
971 lockdep_assert_held(&devices_rwsem);
973 down_read(&rdma_nets_rwsem);
974 xa_for_each (&rdma_nets, index, rnet) {
975 ret = add_one_compat_dev(device, rnet);
979 up_read(&rdma_nets_rwsem);
983 static void remove_all_compat_devs(void)
985 struct ib_compat_device *cdev;
986 struct ib_device *dev;
989 down_read(&devices_rwsem);
990 xa_for_each (&devices, index, dev) {
991 unsigned long c_index = 0;
993 /* Hold nets_rwsem so that any other thread modifying this
994 * system param can sync with this thread.
996 down_read(&rdma_nets_rwsem);
997 xa_for_each (&dev->compat_devs, c_index, cdev)
998 remove_one_compat_dev(dev, c_index);
999 up_read(&rdma_nets_rwsem);
1001 up_read(&devices_rwsem);
1004 static int add_all_compat_devs(void)
1006 struct rdma_dev_net *rnet;
1007 struct ib_device *dev;
1008 unsigned long index;
1011 down_read(&devices_rwsem);
1012 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1013 unsigned long net_index = 0;
1015 /* Hold nets_rwsem so that any other thread modifying this
1016 * system param can sync with this thread.
1018 down_read(&rdma_nets_rwsem);
1019 xa_for_each (&rdma_nets, net_index, rnet) {
1020 ret = add_one_compat_dev(dev, rnet);
1024 up_read(&rdma_nets_rwsem);
1026 up_read(&devices_rwsem);
1028 remove_all_compat_devs();
1032 int rdma_compatdev_set(u8 enable)
1034 struct rdma_dev_net *rnet;
1035 unsigned long index;
1038 down_write(&rdma_nets_rwsem);
1039 if (ib_devices_shared_netns == enable) {
1040 up_write(&rdma_nets_rwsem);
1044 /* enable/disable of compat devices is not supported
1045 * when more than default init_net exists.
1047 xa_for_each (&rdma_nets, index, rnet) {
1052 ib_devices_shared_netns = enable;
1053 up_write(&rdma_nets_rwsem);
1058 ret = add_all_compat_devs();
1060 remove_all_compat_devs();
1064 static void rdma_dev_exit_net(struct net *net)
1066 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1067 struct ib_device *dev;
1068 unsigned long index;
1071 down_write(&rdma_nets_rwsem);
1073 * Prevent the ID from being re-used and hide the id from xa_for_each.
1075 ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1077 up_write(&rdma_nets_rwsem);
1079 down_read(&devices_rwsem);
1080 xa_for_each (&devices, index, dev) {
1081 get_device(&dev->dev);
1083 * Release the devices_rwsem so that pontentially blocking
1084 * device_del, doesn't hold the devices_rwsem for too long.
1086 up_read(&devices_rwsem);
1088 remove_one_compat_dev(dev, rnet->id);
1091 * If the real device is in the NS then move it back to init.
1093 rdma_dev_change_netns(dev, net, &init_net);
1095 put_device(&dev->dev);
1096 down_read(&devices_rwsem);
1098 up_read(&devices_rwsem);
1100 rdma_nl_net_exit(rnet);
1101 xa_erase(&rdma_nets, rnet->id);
1104 static __net_init int rdma_dev_init_net(struct net *net)
1106 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1107 unsigned long index;
1108 struct ib_device *dev;
1111 write_pnet(&rnet->net, net);
1113 ret = rdma_nl_net_init(rnet);
1117 /* No need to create any compat devices in default init_net. */
1118 if (net_eq(net, &init_net))
1121 ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1123 rdma_nl_net_exit(rnet);
1127 down_read(&devices_rwsem);
1128 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1129 /* Hold nets_rwsem so that netlink command cannot change
1130 * system configuration for device sharing mode.
1132 down_read(&rdma_nets_rwsem);
1133 ret = add_one_compat_dev(dev, rnet);
1134 up_read(&rdma_nets_rwsem);
1138 up_read(&devices_rwsem);
1141 rdma_dev_exit_net(net);
1147 * Assign the unique string device name and the unique device index. This is
1148 * undone by ib_dealloc_device.
1150 static int assign_name(struct ib_device *device, const char *name)
1155 down_write(&devices_rwsem);
1156 /* Assign a unique name to the device */
1157 if (strchr(name, '%'))
1158 ret = alloc_name(device, name);
1160 ret = dev_set_name(&device->dev, name);
1164 if (__ib_device_get_by_name(dev_name(&device->dev))) {
1168 strlcpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1170 ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1171 &last_id, GFP_KERNEL);
1176 up_write(&devices_rwsem);
1180 static void setup_dma_device(struct ib_device *device)
1182 struct device *parent = device->dev.parent;
1184 WARN_ON_ONCE(device->dma_device);
1186 #ifdef CONFIG_DMA_OPS
1187 if (device->dev.dma_ops) {
1189 * The caller provided custom DMA operations. Copy the
1190 * DMA-related fields that are used by e.g. dma_alloc_coherent()
1193 device->dma_device = &device->dev;
1194 if (!device->dev.dma_mask) {
1196 device->dev.dma_mask = parent->dma_mask;
1200 if (!device->dev.coherent_dma_mask) {
1202 device->dev.coherent_dma_mask =
1203 parent->coherent_dma_mask;
1208 #endif /* CONFIG_DMA_OPS */
1211 * The caller did not provide custom DMA operations. Use the
1212 * DMA mapping operations of the parent device.
1214 WARN_ON_ONCE(!parent);
1215 device->dma_device = parent;
1218 if (!device->dev.dma_parms) {
1221 * The caller did not provide DMA parameters, so
1222 * 'parent' probably represents a PCI device. The PCI
1223 * core sets the maximum segment size to 64
1224 * KB. Increase this parameter to 2 GB.
1226 device->dev.dma_parms = parent->dma_parms;
1227 dma_set_max_seg_size(device->dma_device, SZ_2G);
1235 * setup_device() allocates memory and sets up data that requires calling the
1236 * device ops, this is the only reason these actions are not done during
1237 * ib_alloc_device. It is undone by ib_dealloc_device().
1239 static int setup_device(struct ib_device *device)
1241 struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1244 setup_dma_device(device);
1245 ib_device_check_mandatory(device);
1247 ret = setup_port_data(device);
1249 dev_warn(&device->dev, "Couldn't create per-port data\n");
1253 memset(&device->attrs, 0, sizeof(device->attrs));
1254 ret = device->ops.query_device(device, &device->attrs, &uhw);
1256 dev_warn(&device->dev,
1257 "Couldn't query the device attributes\n");
1264 static void disable_device(struct ib_device *device)
1268 WARN_ON(!refcount_read(&device->refcount));
1270 down_write(&devices_rwsem);
1271 xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1272 up_write(&devices_rwsem);
1275 * Remove clients in LIFO order, see assign_client_id. This could be
1276 * more efficient if xarray learns to reverse iterate. Since no new
1277 * clients can be added to this ib_device past this point we only need
1278 * the maximum possible client_id value here.
1280 down_read(&clients_rwsem);
1281 cid = highest_client_id;
1282 up_read(&clients_rwsem);
1285 remove_client_context(device, cid);
1288 /* Pairs with refcount_set in enable_device */
1289 ib_device_put(device);
1290 wait_for_completion(&device->unreg_completion);
1293 * compat devices must be removed after device refcount drops to zero.
1294 * Otherwise init_net() may add more compatdevs after removing compat
1295 * devices and before device is disabled.
1297 remove_compat_devs(device);
1301 * An enabled device is visible to all clients and to all the public facing
1302 * APIs that return a device pointer. This always returns with a new get, even
1305 static int enable_device_and_get(struct ib_device *device)
1307 struct ib_client *client;
1308 unsigned long index;
1312 * One ref belongs to the xa and the other belongs to this
1313 * thread. This is needed to guard against parallel unregistration.
1315 refcount_set(&device->refcount, 2);
1316 down_write(&devices_rwsem);
1317 xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1320 * By using downgrade_write() we ensure that no other thread can clear
1321 * DEVICE_REGISTERED while we are completing the client setup.
1323 downgrade_write(&devices_rwsem);
1325 if (device->ops.enable_driver) {
1326 ret = device->ops.enable_driver(device);
1331 down_read(&clients_rwsem);
1332 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1333 ret = add_client_context(device, client);
1337 up_read(&clients_rwsem);
1339 ret = add_compat_devs(device);
1341 up_read(&devices_rwsem);
1345 static void prevent_dealloc_device(struct ib_device *ib_dev)
1350 * ib_register_device - Register an IB device with IB core
1351 * @device: Device to register
1352 * @name: unique string device name. This may include a '%' which will
1353 * cause a unique index to be added to the passed device name.
1355 * Low-level drivers use ib_register_device() to register their
1356 * devices with the IB core. All registered clients will receive a
1357 * callback for each device that is added. @device must be allocated
1358 * with ib_alloc_device().
1360 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1361 * asynchronously then the device pointer may become freed as soon as this
1364 int ib_register_device(struct ib_device *device, const char *name)
1368 ret = assign_name(device, name);
1372 ret = setup_device(device);
1376 ret = ib_cache_setup_one(device);
1378 dev_warn(&device->dev,
1379 "Couldn't set up InfiniBand P_Key/GID cache\n");
1383 ib_device_register_rdmacg(device);
1385 rdma_counter_init(device);
1388 * Ensure that ADD uevent is not fired because it
1389 * is too early amd device is not initialized yet.
1391 dev_set_uevent_suppress(&device->dev, true);
1392 ret = device_add(&device->dev);
1396 ret = ib_device_register_sysfs(device);
1398 dev_warn(&device->dev,
1399 "Couldn't register device with driver model\n");
1403 ib_cq_pool_init(device);
1404 ret = enable_device_and_get(device);
1405 dev_set_uevent_suppress(&device->dev, false);
1406 /* Mark for userspace that device is ready */
1407 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1409 void (*dealloc_fn)(struct ib_device *);
1412 * If we hit this error flow then we don't want to
1413 * automatically dealloc the device since the caller is
1414 * expected to call ib_dealloc_device() after
1415 * ib_register_device() fails. This is tricky due to the
1416 * possibility for a parallel unregistration along with this
1417 * error flow. Since we have a refcount here we know any
1418 * parallel flow is stopped in disable_device and will see the
1419 * special dealloc_driver pointer, causing the responsibility to
1420 * ib_dealloc_device() to revert back to this thread.
1422 dealloc_fn = device->ops.dealloc_driver;
1423 device->ops.dealloc_driver = prevent_dealloc_device;
1424 ib_device_put(device);
1425 __ib_unregister_device(device);
1426 device->ops.dealloc_driver = dealloc_fn;
1429 ib_device_put(device);
1434 device_del(&device->dev);
1436 dev_set_uevent_suppress(&device->dev, false);
1437 ib_device_unregister_rdmacg(device);
1438 ib_cache_cleanup_one(device);
1441 EXPORT_SYMBOL(ib_register_device);
1443 /* Callers must hold a get on the device. */
1444 static void __ib_unregister_device(struct ib_device *ib_dev)
1447 * We have a registration lock so that all the calls to unregister are
1448 * fully fenced, once any unregister returns the device is truely
1449 * unregistered even if multiple callers are unregistering it at the
1450 * same time. This also interacts with the registration flow and
1451 * provides sane semantics if register and unregister are racing.
1453 mutex_lock(&ib_dev->unregistration_lock);
1454 if (!refcount_read(&ib_dev->refcount))
1457 disable_device(ib_dev);
1458 ib_cq_pool_destroy(ib_dev);
1460 /* Expedite removing unregistered pointers from the hash table */
1461 free_netdevs(ib_dev);
1463 ib_device_unregister_sysfs(ib_dev);
1464 device_del(&ib_dev->dev);
1465 ib_device_unregister_rdmacg(ib_dev);
1466 ib_cache_cleanup_one(ib_dev);
1469 * Drivers using the new flow may not call ib_dealloc_device except
1470 * in error unwind prior to registration success.
1472 if (ib_dev->ops.dealloc_driver &&
1473 ib_dev->ops.dealloc_driver != prevent_dealloc_device) {
1474 WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1475 ib_dealloc_device(ib_dev);
1478 mutex_unlock(&ib_dev->unregistration_lock);
1482 * ib_unregister_device - Unregister an IB device
1483 * @ib_dev: The device to unregister
1485 * Unregister an IB device. All clients will receive a remove callback.
1487 * Callers should call this routine only once, and protect against races with
1488 * registration. Typically it should only be called as part of a remove
1489 * callback in an implementation of driver core's struct device_driver and
1492 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1495 void ib_unregister_device(struct ib_device *ib_dev)
1497 get_device(&ib_dev->dev);
1498 __ib_unregister_device(ib_dev);
1499 put_device(&ib_dev->dev);
1501 EXPORT_SYMBOL(ib_unregister_device);
1504 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1505 * @ib_dev: The device to unregister
1507 * This is the same as ib_unregister_device(), except it includes an internal
1508 * ib_device_put() that should match a 'get' obtained by the caller.
1510 * It is safe to call this routine concurrently from multiple threads while
1511 * holding the 'get'. When the function returns the device is fully
1514 * Drivers using this flow MUST use the driver_unregister callback to clean up
1515 * their resources associated with the device and dealloc it.
1517 void ib_unregister_device_and_put(struct ib_device *ib_dev)
1519 WARN_ON(!ib_dev->ops.dealloc_driver);
1520 get_device(&ib_dev->dev);
1521 ib_device_put(ib_dev);
1522 __ib_unregister_device(ib_dev);
1523 put_device(&ib_dev->dev);
1525 EXPORT_SYMBOL(ib_unregister_device_and_put);
1528 * ib_unregister_driver - Unregister all IB devices for a driver
1529 * @driver_id: The driver to unregister
1531 * This implements a fence for device unregistration. It only returns once all
1532 * devices associated with the driver_id have fully completed their
1533 * unregistration and returned from ib_unregister_device*().
1535 * If device's are not yet unregistered it goes ahead and starts unregistering
1538 * This does not block creation of new devices with the given driver_id, that
1539 * is the responsibility of the caller.
1541 void ib_unregister_driver(enum rdma_driver_id driver_id)
1543 struct ib_device *ib_dev;
1544 unsigned long index;
1546 down_read(&devices_rwsem);
1547 xa_for_each (&devices, index, ib_dev) {
1548 if (ib_dev->ops.driver_id != driver_id)
1551 get_device(&ib_dev->dev);
1552 up_read(&devices_rwsem);
1554 WARN_ON(!ib_dev->ops.dealloc_driver);
1555 __ib_unregister_device(ib_dev);
1557 put_device(&ib_dev->dev);
1558 down_read(&devices_rwsem);
1560 up_read(&devices_rwsem);
1562 EXPORT_SYMBOL(ib_unregister_driver);
1564 static void ib_unregister_work(struct work_struct *work)
1566 struct ib_device *ib_dev =
1567 container_of(work, struct ib_device, unregistration_work);
1569 __ib_unregister_device(ib_dev);
1570 put_device(&ib_dev->dev);
1574 * ib_unregister_device_queued - Unregister a device using a work queue
1575 * @ib_dev: The device to unregister
1577 * This schedules an asynchronous unregistration using a WQ for the device. A
1578 * driver should use this to avoid holding locks while doing unregistration,
1579 * such as holding the RTNL lock.
1581 * Drivers using this API must use ib_unregister_driver before module unload
1582 * to ensure that all scheduled unregistrations have completed.
1584 void ib_unregister_device_queued(struct ib_device *ib_dev)
1586 WARN_ON(!refcount_read(&ib_dev->refcount));
1587 WARN_ON(!ib_dev->ops.dealloc_driver);
1588 get_device(&ib_dev->dev);
1589 if (!queue_work(system_unbound_wq, &ib_dev->unregistration_work))
1590 put_device(&ib_dev->dev);
1592 EXPORT_SYMBOL(ib_unregister_device_queued);
1595 * The caller must pass in a device that has the kref held and the refcount
1596 * released. If the device is in cur_net and still registered then it is moved
1599 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1605 mutex_lock(&device->unregistration_lock);
1608 * If a device not under ib_device_get() or if the unregistration_lock
1609 * is not held, the namespace can be changed, or it can be unregistered.
1610 * Check again under the lock.
1612 if (refcount_read(&device->refcount) == 0 ||
1613 !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1618 kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1619 disable_device(device);
1622 * At this point no one can be using the device, so it is safe to
1623 * change the namespace.
1625 write_pnet(&device->coredev.rdma_net, net);
1627 down_read(&devices_rwsem);
1629 * Currently rdma devices are system wide unique. So the device name
1630 * is guaranteed free in the new namespace. Publish the new namespace
1631 * at the sysfs level.
1633 ret = device_rename(&device->dev, dev_name(&device->dev));
1634 up_read(&devices_rwsem);
1636 dev_warn(&device->dev,
1637 "%s: Couldn't rename device after namespace change\n",
1639 /* Try and put things back and re-enable the device */
1640 write_pnet(&device->coredev.rdma_net, cur_net);
1643 ret2 = enable_device_and_get(device);
1646 * This shouldn't really happen, but if it does, let the user
1647 * retry at later point. So don't disable the device.
1649 dev_warn(&device->dev,
1650 "%s: Couldn't re-enable device after namespace change\n",
1653 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1655 ib_device_put(device);
1657 mutex_unlock(&device->unregistration_lock);
1663 int ib_device_set_netns_put(struct sk_buff *skb,
1664 struct ib_device *dev, u32 ns_fd)
1669 net = get_net_ns_by_fd(ns_fd);
1675 if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1681 * Currently supported only for those providers which support
1682 * disassociation and don't do port specific sysfs init. Once a
1683 * port_cleanup infrastructure is implemented, this limitation will be
1686 if (!dev->ops.disassociate_ucontext || dev->ops.init_port ||
1687 ib_devices_shared_netns) {
1692 get_device(&dev->dev);
1694 ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1695 put_device(&dev->dev);
1707 static struct pernet_operations rdma_dev_net_ops = {
1708 .init = rdma_dev_init_net,
1709 .exit = rdma_dev_exit_net,
1710 .id = &rdma_dev_net_id,
1711 .size = sizeof(struct rdma_dev_net),
1714 static int assign_client_id(struct ib_client *client)
1718 down_write(&clients_rwsem);
1720 * The add/remove callbacks must be called in FIFO/LIFO order. To
1721 * achieve this we assign client_ids so they are sorted in
1722 * registration order.
1724 client->client_id = highest_client_id;
1725 ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1729 highest_client_id++;
1730 xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1733 up_write(&clients_rwsem);
1737 static void remove_client_id(struct ib_client *client)
1739 down_write(&clients_rwsem);
1740 xa_erase(&clients, client->client_id);
1741 for (; highest_client_id; highest_client_id--)
1742 if (xa_load(&clients, highest_client_id - 1))
1744 up_write(&clients_rwsem);
1748 * ib_register_client - Register an IB client
1749 * @client:Client to register
1751 * Upper level users of the IB drivers can use ib_register_client() to
1752 * register callbacks for IB device addition and removal. When an IB
1753 * device is added, each registered client's add method will be called
1754 * (in the order the clients were registered), and when a device is
1755 * removed, each client's remove method will be called (in the reverse
1756 * order that clients were registered). In addition, when
1757 * ib_register_client() is called, the client will receive an add
1758 * callback for all devices already registered.
1760 int ib_register_client(struct ib_client *client)
1762 struct ib_device *device;
1763 unsigned long index;
1766 refcount_set(&client->uses, 1);
1767 init_completion(&client->uses_zero);
1768 ret = assign_client_id(client);
1772 down_read(&devices_rwsem);
1773 xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1774 ret = add_client_context(device, client);
1776 up_read(&devices_rwsem);
1777 ib_unregister_client(client);
1781 up_read(&devices_rwsem);
1784 EXPORT_SYMBOL(ib_register_client);
1787 * ib_unregister_client - Unregister an IB client
1788 * @client:Client to unregister
1790 * Upper level users use ib_unregister_client() to remove their client
1791 * registration. When ib_unregister_client() is called, the client
1792 * will receive a remove callback for each IB device still registered.
1794 * This is a full fence, once it returns no client callbacks will be called,
1795 * or are running in another thread.
1797 void ib_unregister_client(struct ib_client *client)
1799 struct ib_device *device;
1800 unsigned long index;
1802 down_write(&clients_rwsem);
1803 ib_client_put(client);
1804 xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1805 up_write(&clients_rwsem);
1807 /* We do not want to have locks while calling client->remove() */
1809 xa_for_each (&devices, index, device) {
1810 if (!ib_device_try_get(device))
1814 remove_client_context(device, client->client_id);
1816 ib_device_put(device);
1822 * remove_client_context() is not a fence, it can return even though a
1823 * removal is ongoing. Wait until all removals are completed.
1825 wait_for_completion(&client->uses_zero);
1826 remove_client_id(client);
1828 EXPORT_SYMBOL(ib_unregister_client);
1830 static int __ib_get_global_client_nl_info(const char *client_name,
1831 struct ib_client_nl_info *res)
1833 struct ib_client *client;
1834 unsigned long index;
1837 down_read(&clients_rwsem);
1838 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1839 if (strcmp(client->name, client_name) != 0)
1841 if (!client->get_global_nl_info) {
1845 ret = client->get_global_nl_info(res);
1846 if (WARN_ON(ret == -ENOENT))
1848 if (!ret && res->cdev)
1849 get_device(res->cdev);
1852 up_read(&clients_rwsem);
1856 static int __ib_get_client_nl_info(struct ib_device *ibdev,
1857 const char *client_name,
1858 struct ib_client_nl_info *res)
1860 unsigned long index;
1864 down_read(&ibdev->client_data_rwsem);
1865 xan_for_each_marked (&ibdev->client_data, index, client_data,
1866 CLIENT_DATA_REGISTERED) {
1867 struct ib_client *client = xa_load(&clients, index);
1869 if (!client || strcmp(client->name, client_name) != 0)
1871 if (!client->get_nl_info) {
1875 ret = client->get_nl_info(ibdev, client_data, res);
1876 if (WARN_ON(ret == -ENOENT))
1880 * The cdev is guaranteed valid as long as we are inside the
1881 * client_data_rwsem as remove_one can't be called. Keep it
1882 * valid for the caller.
1884 if (!ret && res->cdev)
1885 get_device(res->cdev);
1888 up_read(&ibdev->client_data_rwsem);
1894 * ib_get_client_nl_info - Fetch the nl_info from a client
1895 * @device - IB device
1896 * @client_name - Name of the client
1897 * @res - Result of the query
1899 int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1900 struct ib_client_nl_info *res)
1905 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1907 ret = __ib_get_global_client_nl_info(client_name, res);
1908 #ifdef CONFIG_MODULES
1909 if (ret == -ENOENT) {
1910 request_module("rdma-client-%s", client_name);
1912 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1914 ret = __ib_get_global_client_nl_info(client_name, res);
1923 if (WARN_ON(!res->cdev))
1929 * ib_set_client_data - Set IB client context
1930 * @device:Device to set context for
1931 * @client:Client to set context for
1932 * @data:Context to set
1934 * ib_set_client_data() sets client context data that can be retrieved with
1935 * ib_get_client_data(). This can only be called while the client is
1936 * registered to the device, once the ib_client remove() callback returns this
1939 void ib_set_client_data(struct ib_device *device, struct ib_client *client,
1944 if (WARN_ON(IS_ERR(data)))
1947 rc = xa_store(&device->client_data, client->client_id, data,
1949 WARN_ON(xa_is_err(rc));
1951 EXPORT_SYMBOL(ib_set_client_data);
1954 * ib_register_event_handler - Register an IB event handler
1955 * @event_handler:Handler to register
1957 * ib_register_event_handler() registers an event handler that will be
1958 * called back when asynchronous IB events occur (as defined in
1959 * chapter 11 of the InfiniBand Architecture Specification). This
1960 * callback occurs in workqueue context.
1962 void ib_register_event_handler(struct ib_event_handler *event_handler)
1964 down_write(&event_handler->device->event_handler_rwsem);
1965 list_add_tail(&event_handler->list,
1966 &event_handler->device->event_handler_list);
1967 up_write(&event_handler->device->event_handler_rwsem);
1969 EXPORT_SYMBOL(ib_register_event_handler);
1972 * ib_unregister_event_handler - Unregister an event handler
1973 * @event_handler:Handler to unregister
1975 * Unregister an event handler registered with
1976 * ib_register_event_handler().
1978 void ib_unregister_event_handler(struct ib_event_handler *event_handler)
1980 down_write(&event_handler->device->event_handler_rwsem);
1981 list_del(&event_handler->list);
1982 up_write(&event_handler->device->event_handler_rwsem);
1984 EXPORT_SYMBOL(ib_unregister_event_handler);
1986 void ib_dispatch_event_clients(struct ib_event *event)
1988 struct ib_event_handler *handler;
1990 down_read(&event->device->event_handler_rwsem);
1992 list_for_each_entry(handler, &event->device->event_handler_list, list)
1993 handler->handler(handler, event);
1995 up_read(&event->device->event_handler_rwsem);
1998 static int iw_query_port(struct ib_device *device,
2000 struct ib_port_attr *port_attr)
2002 struct in_device *inetdev;
2003 struct net_device *netdev;
2005 memset(port_attr, 0, sizeof(*port_attr));
2007 netdev = ib_device_get_netdev(device, port_num);
2011 port_attr->max_mtu = IB_MTU_4096;
2012 port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
2014 if (!netif_carrier_ok(netdev)) {
2015 port_attr->state = IB_PORT_DOWN;
2016 port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
2019 inetdev = __in_dev_get_rcu(netdev);
2021 if (inetdev && inetdev->ifa_list) {
2022 port_attr->state = IB_PORT_ACTIVE;
2023 port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
2025 port_attr->state = IB_PORT_INIT;
2026 port_attr->phys_state =
2027 IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
2034 return device->ops.query_port(device, port_num, port_attr);
2037 static int __ib_query_port(struct ib_device *device,
2039 struct ib_port_attr *port_attr)
2041 union ib_gid gid = {};
2044 memset(port_attr, 0, sizeof(*port_attr));
2046 err = device->ops.query_port(device, port_num, port_attr);
2047 if (err || port_attr->subnet_prefix)
2050 if (rdma_port_get_link_layer(device, port_num) !=
2051 IB_LINK_LAYER_INFINIBAND)
2054 err = device->ops.query_gid(device, port_num, 0, &gid);
2058 port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
2063 * ib_query_port - Query IB port attributes
2064 * @device:Device to query
2065 * @port_num:Port number to query
2066 * @port_attr:Port attributes
2068 * ib_query_port() returns the attributes of a port through the
2069 * @port_attr pointer.
2071 int ib_query_port(struct ib_device *device,
2073 struct ib_port_attr *port_attr)
2075 if (!rdma_is_port_valid(device, port_num))
2078 if (rdma_protocol_iwarp(device, port_num))
2079 return iw_query_port(device, port_num, port_attr);
2081 return __ib_query_port(device, port_num, port_attr);
2083 EXPORT_SYMBOL(ib_query_port);
2085 static void add_ndev_hash(struct ib_port_data *pdata)
2087 unsigned long flags;
2091 spin_lock_irqsave(&ndev_hash_lock, flags);
2092 if (hash_hashed(&pdata->ndev_hash_link)) {
2093 hash_del_rcu(&pdata->ndev_hash_link);
2094 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2096 * We cannot do hash_add_rcu after a hash_del_rcu until the
2100 spin_lock_irqsave(&ndev_hash_lock, flags);
2103 hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
2104 (uintptr_t)pdata->netdev);
2105 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2109 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2110 * @ib_dev: Device to modify
2111 * @ndev: net_device to affiliate, may be NULL
2112 * @port: IB port the net_device is connected to
2114 * Drivers should use this to link the ib_device to a netdev so the netdev
2115 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2116 * affiliated with any port.
2118 * The caller must ensure that the given ndev is not unregistered or
2119 * unregistering, and that either the ib_device is unregistered or
2120 * ib_device_set_netdev() is called with NULL when the ndev sends a
2121 * NETDEV_UNREGISTER event.
2123 int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2126 struct net_device *old_ndev;
2127 struct ib_port_data *pdata;
2128 unsigned long flags;
2132 * Drivers wish to call this before ib_register_driver, so we have to
2133 * setup the port data early.
2135 ret = alloc_port_data(ib_dev);
2139 if (!rdma_is_port_valid(ib_dev, port))
2142 pdata = &ib_dev->port_data[port];
2143 spin_lock_irqsave(&pdata->netdev_lock, flags);
2144 old_ndev = rcu_dereference_protected(
2145 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2146 if (old_ndev == ndev) {
2147 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2153 rcu_assign_pointer(pdata->netdev, ndev);
2154 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2156 add_ndev_hash(pdata);
2162 EXPORT_SYMBOL(ib_device_set_netdev);
2164 static void free_netdevs(struct ib_device *ib_dev)
2166 unsigned long flags;
2169 if (!ib_dev->port_data)
2172 rdma_for_each_port (ib_dev, port) {
2173 struct ib_port_data *pdata = &ib_dev->port_data[port];
2174 struct net_device *ndev;
2176 spin_lock_irqsave(&pdata->netdev_lock, flags);
2177 ndev = rcu_dereference_protected(
2178 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2180 spin_lock(&ndev_hash_lock);
2181 hash_del_rcu(&pdata->ndev_hash_link);
2182 spin_unlock(&ndev_hash_lock);
2185 * If this is the last dev_put there is still a
2186 * synchronize_rcu before the netdev is kfreed, so we
2187 * can continue to rely on unlocked pointer
2188 * comparisons after the put
2190 rcu_assign_pointer(pdata->netdev, NULL);
2193 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2197 struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2200 struct ib_port_data *pdata;
2201 struct net_device *res;
2203 if (!rdma_is_port_valid(ib_dev, port))
2206 pdata = &ib_dev->port_data[port];
2209 * New drivers should use ib_device_set_netdev() not the legacy
2212 if (ib_dev->ops.get_netdev)
2213 res = ib_dev->ops.get_netdev(ib_dev, port);
2215 spin_lock(&pdata->netdev_lock);
2216 res = rcu_dereference_protected(
2217 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2220 spin_unlock(&pdata->netdev_lock);
2224 * If we are starting to unregister expedite things by preventing
2225 * propagation of an unregistering netdev.
2227 if (res && res->reg_state != NETREG_REGISTERED) {
2236 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2237 * @ndev: netdev to locate
2238 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2240 * Find and hold an ib_device that is associated with a netdev via
2241 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2244 struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2245 enum rdma_driver_id driver_id)
2247 struct ib_device *res = NULL;
2248 struct ib_port_data *cur;
2251 hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2253 if (rcu_access_pointer(cur->netdev) == ndev &&
2254 (driver_id == RDMA_DRIVER_UNKNOWN ||
2255 cur->ib_dev->ops.driver_id == driver_id) &&
2256 ib_device_try_get(cur->ib_dev)) {
2265 EXPORT_SYMBOL(ib_device_get_by_netdev);
2268 * ib_enum_roce_netdev - enumerate all RoCE ports
2269 * @ib_dev : IB device we want to query
2270 * @filter: Should we call the callback?
2271 * @filter_cookie: Cookie passed to filter
2272 * @cb: Callback to call for each found RoCE ports
2273 * @cookie: Cookie passed back to the callback
2275 * Enumerates all of the physical RoCE ports of ib_dev
2276 * which are related to netdevice and calls callback() on each
2277 * device for which filter() function returns non zero.
2279 void ib_enum_roce_netdev(struct ib_device *ib_dev,
2280 roce_netdev_filter filter,
2281 void *filter_cookie,
2282 roce_netdev_callback cb,
2287 rdma_for_each_port (ib_dev, port)
2288 if (rdma_protocol_roce(ib_dev, port)) {
2289 struct net_device *idev =
2290 ib_device_get_netdev(ib_dev, port);
2292 if (filter(ib_dev, port, idev, filter_cookie))
2293 cb(ib_dev, port, idev, cookie);
2301 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2302 * @filter: Should we call the callback?
2303 * @filter_cookie: Cookie passed to filter
2304 * @cb: Callback to call for each found RoCE ports
2305 * @cookie: Cookie passed back to the callback
2307 * Enumerates all RoCE devices' physical ports which are related
2308 * to netdevices and calls callback() on each device for which
2309 * filter() function returns non zero.
2311 void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2312 void *filter_cookie,
2313 roce_netdev_callback cb,
2316 struct ib_device *dev;
2317 unsigned long index;
2319 down_read(&devices_rwsem);
2320 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2321 ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2322 up_read(&devices_rwsem);
2326 * ib_enum_all_devs - enumerate all ib_devices
2327 * @cb: Callback to call for each found ib_device
2329 * Enumerates all ib_devices and calls callback() on each device.
2331 int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2332 struct netlink_callback *cb)
2334 unsigned long index;
2335 struct ib_device *dev;
2336 unsigned int idx = 0;
2339 down_read(&devices_rwsem);
2340 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2341 if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2344 ret = nldev_cb(dev, skb, cb, idx);
2349 up_read(&devices_rwsem);
2354 * ib_query_pkey - Get P_Key table entry
2355 * @device:Device to query
2356 * @port_num:Port number to query
2357 * @index:P_Key table index to query
2358 * @pkey:Returned P_Key
2360 * ib_query_pkey() fetches the specified P_Key table entry.
2362 int ib_query_pkey(struct ib_device *device,
2363 u8 port_num, u16 index, u16 *pkey)
2365 if (!rdma_is_port_valid(device, port_num))
2368 if (!device->ops.query_pkey)
2371 return device->ops.query_pkey(device, port_num, index, pkey);
2373 EXPORT_SYMBOL(ib_query_pkey);
2376 * ib_modify_device - Change IB device attributes
2377 * @device:Device to modify
2378 * @device_modify_mask:Mask of attributes to change
2379 * @device_modify:New attribute values
2381 * ib_modify_device() changes a device's attributes as specified by
2382 * the @device_modify_mask and @device_modify structure.
2384 int ib_modify_device(struct ib_device *device,
2385 int device_modify_mask,
2386 struct ib_device_modify *device_modify)
2388 if (!device->ops.modify_device)
2391 return device->ops.modify_device(device, device_modify_mask,
2394 EXPORT_SYMBOL(ib_modify_device);
2397 * ib_modify_port - Modifies the attributes for the specified port.
2398 * @device: The device to modify.
2399 * @port_num: The number of the port to modify.
2400 * @port_modify_mask: Mask used to specify which attributes of the port
2402 * @port_modify: New attribute values for the port.
2404 * ib_modify_port() changes a port's attributes as specified by the
2405 * @port_modify_mask and @port_modify structure.
2407 int ib_modify_port(struct ib_device *device,
2408 u8 port_num, int port_modify_mask,
2409 struct ib_port_modify *port_modify)
2413 if (!rdma_is_port_valid(device, port_num))
2416 if (device->ops.modify_port)
2417 rc = device->ops.modify_port(device, port_num,
2420 else if (rdma_protocol_roce(device, port_num) &&
2421 ((port_modify->set_port_cap_mask & ~IB_PORT_CM_SUP) == 0 ||
2422 (port_modify->clr_port_cap_mask & ~IB_PORT_CM_SUP) == 0))
2428 EXPORT_SYMBOL(ib_modify_port);
2431 * ib_find_gid - Returns the port number and GID table index where
2432 * a specified GID value occurs. Its searches only for IB link layer.
2433 * @device: The device to query.
2434 * @gid: The GID value to search for.
2435 * @port_num: The port number of the device where the GID value was found.
2436 * @index: The index into the GID table where the GID was found. This
2437 * parameter may be NULL.
2439 int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2440 u8 *port_num, u16 *index)
2442 union ib_gid tmp_gid;
2446 rdma_for_each_port (device, port) {
2447 if (!rdma_protocol_ib(device, port))
2450 for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2452 ret = rdma_query_gid(device, port, i, &tmp_gid);
2455 if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2466 EXPORT_SYMBOL(ib_find_gid);
2469 * ib_find_pkey - Returns the PKey table index where a specified
2470 * PKey value occurs.
2471 * @device: The device to query.
2472 * @port_num: The port number of the device to search for the PKey.
2473 * @pkey: The PKey value to search for.
2474 * @index: The index into the PKey table where the PKey was found.
2476 int ib_find_pkey(struct ib_device *device,
2477 u8 port_num, u16 pkey, u16 *index)
2481 int partial_ix = -1;
2483 for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2485 ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2488 if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2489 /* if there is full-member pkey take it.*/
2490 if (tmp_pkey & 0x8000) {
2499 /*no full-member, if exists take the limited*/
2500 if (partial_ix >= 0) {
2501 *index = partial_ix;
2506 EXPORT_SYMBOL(ib_find_pkey);
2509 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2510 * for a received CM request
2511 * @dev: An RDMA device on which the request has been received.
2512 * @port: Port number on the RDMA device.
2513 * @pkey: The Pkey the request came on.
2514 * @gid: A GID that the net_dev uses to communicate.
2515 * @addr: Contains the IP address that the request specified as its
2519 struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2522 const union ib_gid *gid,
2523 const struct sockaddr *addr)
2525 struct net_device *net_dev = NULL;
2526 unsigned long index;
2529 if (!rdma_protocol_ib(dev, port))
2533 * Holding the read side guarantees that the client will not become
2534 * unregistered while we are calling get_net_dev_by_params()
2536 down_read(&dev->client_data_rwsem);
2537 xan_for_each_marked (&dev->client_data, index, client_data,
2538 CLIENT_DATA_REGISTERED) {
2539 struct ib_client *client = xa_load(&clients, index);
2541 if (!client || !client->get_net_dev_by_params)
2544 net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2549 up_read(&dev->client_data_rwsem);
2553 EXPORT_SYMBOL(ib_get_net_dev_by_params);
2555 void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2557 struct ib_device_ops *dev_ops = &dev->ops;
2558 #define SET_DEVICE_OP(ptr, name) \
2561 if (!((ptr)->name)) \
2562 (ptr)->name = ops->name; \
2565 #define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2567 if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2568 WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2569 dev_ops->driver_id != ops->driver_id);
2570 dev_ops->driver_id = ops->driver_id;
2573 WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2574 dev_ops->owner = ops->owner;
2576 if (ops->uverbs_abi_ver)
2577 dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2579 dev_ops->uverbs_no_driver_id_binding |=
2580 ops->uverbs_no_driver_id_binding;
2582 SET_DEVICE_OP(dev_ops, add_gid);
2583 SET_DEVICE_OP(dev_ops, advise_mr);
2584 SET_DEVICE_OP(dev_ops, alloc_dm);
2585 SET_DEVICE_OP(dev_ops, alloc_hw_stats);
2586 SET_DEVICE_OP(dev_ops, alloc_mr);
2587 SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2588 SET_DEVICE_OP(dev_ops, alloc_mw);
2589 SET_DEVICE_OP(dev_ops, alloc_pd);
2590 SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2591 SET_DEVICE_OP(dev_ops, alloc_ucontext);
2592 SET_DEVICE_OP(dev_ops, alloc_xrcd);
2593 SET_DEVICE_OP(dev_ops, attach_mcast);
2594 SET_DEVICE_OP(dev_ops, check_mr_status);
2595 SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2596 SET_DEVICE_OP(dev_ops, counter_bind_qp);
2597 SET_DEVICE_OP(dev_ops, counter_dealloc);
2598 SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2599 SET_DEVICE_OP(dev_ops, counter_update_stats);
2600 SET_DEVICE_OP(dev_ops, create_ah);
2601 SET_DEVICE_OP(dev_ops, create_counters);
2602 SET_DEVICE_OP(dev_ops, create_cq);
2603 SET_DEVICE_OP(dev_ops, create_flow);
2604 SET_DEVICE_OP(dev_ops, create_flow_action_esp);
2605 SET_DEVICE_OP(dev_ops, create_qp);
2606 SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2607 SET_DEVICE_OP(dev_ops, create_srq);
2608 SET_DEVICE_OP(dev_ops, create_wq);
2609 SET_DEVICE_OP(dev_ops, dealloc_dm);
2610 SET_DEVICE_OP(dev_ops, dealloc_driver);
2611 SET_DEVICE_OP(dev_ops, dealloc_mw);
2612 SET_DEVICE_OP(dev_ops, dealloc_pd);
2613 SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2614 SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2615 SET_DEVICE_OP(dev_ops, del_gid);
2616 SET_DEVICE_OP(dev_ops, dereg_mr);
2617 SET_DEVICE_OP(dev_ops, destroy_ah);
2618 SET_DEVICE_OP(dev_ops, destroy_counters);
2619 SET_DEVICE_OP(dev_ops, destroy_cq);
2620 SET_DEVICE_OP(dev_ops, destroy_flow);
2621 SET_DEVICE_OP(dev_ops, destroy_flow_action);
2622 SET_DEVICE_OP(dev_ops, destroy_qp);
2623 SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2624 SET_DEVICE_OP(dev_ops, destroy_srq);
2625 SET_DEVICE_OP(dev_ops, destroy_wq);
2626 SET_DEVICE_OP(dev_ops, detach_mcast);
2627 SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2628 SET_DEVICE_OP(dev_ops, drain_rq);
2629 SET_DEVICE_OP(dev_ops, drain_sq);
2630 SET_DEVICE_OP(dev_ops, enable_driver);
2631 SET_DEVICE_OP(dev_ops, fill_res_cm_id_entry);
2632 SET_DEVICE_OP(dev_ops, fill_res_cq_entry);
2633 SET_DEVICE_OP(dev_ops, fill_res_cq_entry_raw);
2634 SET_DEVICE_OP(dev_ops, fill_res_mr_entry);
2635 SET_DEVICE_OP(dev_ops, fill_res_mr_entry_raw);
2636 SET_DEVICE_OP(dev_ops, fill_res_qp_entry);
2637 SET_DEVICE_OP(dev_ops, fill_res_qp_entry_raw);
2638 SET_DEVICE_OP(dev_ops, fill_stat_mr_entry);
2639 SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2640 SET_DEVICE_OP(dev_ops, get_dma_mr);
2641 SET_DEVICE_OP(dev_ops, get_hw_stats);
2642 SET_DEVICE_OP(dev_ops, get_link_layer);
2643 SET_DEVICE_OP(dev_ops, get_netdev);
2644 SET_DEVICE_OP(dev_ops, get_port_immutable);
2645 SET_DEVICE_OP(dev_ops, get_vector_affinity);
2646 SET_DEVICE_OP(dev_ops, get_vf_config);
2647 SET_DEVICE_OP(dev_ops, get_vf_guid);
2648 SET_DEVICE_OP(dev_ops, get_vf_stats);
2649 SET_DEVICE_OP(dev_ops, init_port);
2650 SET_DEVICE_OP(dev_ops, iw_accept);
2651 SET_DEVICE_OP(dev_ops, iw_add_ref);
2652 SET_DEVICE_OP(dev_ops, iw_connect);
2653 SET_DEVICE_OP(dev_ops, iw_create_listen);
2654 SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2655 SET_DEVICE_OP(dev_ops, iw_get_qp);
2656 SET_DEVICE_OP(dev_ops, iw_reject);
2657 SET_DEVICE_OP(dev_ops, iw_rem_ref);
2658 SET_DEVICE_OP(dev_ops, map_mr_sg);
2659 SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
2660 SET_DEVICE_OP(dev_ops, mmap);
2661 SET_DEVICE_OP(dev_ops, mmap_free);
2662 SET_DEVICE_OP(dev_ops, modify_ah);
2663 SET_DEVICE_OP(dev_ops, modify_cq);
2664 SET_DEVICE_OP(dev_ops, modify_device);
2665 SET_DEVICE_OP(dev_ops, modify_flow_action_esp);
2666 SET_DEVICE_OP(dev_ops, modify_port);
2667 SET_DEVICE_OP(dev_ops, modify_qp);
2668 SET_DEVICE_OP(dev_ops, modify_srq);
2669 SET_DEVICE_OP(dev_ops, modify_wq);
2670 SET_DEVICE_OP(dev_ops, peek_cq);
2671 SET_DEVICE_OP(dev_ops, poll_cq);
2672 SET_DEVICE_OP(dev_ops, post_recv);
2673 SET_DEVICE_OP(dev_ops, post_send);
2674 SET_DEVICE_OP(dev_ops, post_srq_recv);
2675 SET_DEVICE_OP(dev_ops, process_mad);
2676 SET_DEVICE_OP(dev_ops, query_ah);
2677 SET_DEVICE_OP(dev_ops, query_device);
2678 SET_DEVICE_OP(dev_ops, query_gid);
2679 SET_DEVICE_OP(dev_ops, query_pkey);
2680 SET_DEVICE_OP(dev_ops, query_port);
2681 SET_DEVICE_OP(dev_ops, query_qp);
2682 SET_DEVICE_OP(dev_ops, query_srq);
2683 SET_DEVICE_OP(dev_ops, query_ucontext);
2684 SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2685 SET_DEVICE_OP(dev_ops, read_counters);
2686 SET_DEVICE_OP(dev_ops, reg_dm_mr);
2687 SET_DEVICE_OP(dev_ops, reg_user_mr);
2688 SET_DEVICE_OP(dev_ops, req_ncomp_notif);
2689 SET_DEVICE_OP(dev_ops, req_notify_cq);
2690 SET_DEVICE_OP(dev_ops, rereg_user_mr);
2691 SET_DEVICE_OP(dev_ops, resize_cq);
2692 SET_DEVICE_OP(dev_ops, set_vf_guid);
2693 SET_DEVICE_OP(dev_ops, set_vf_link_state);
2695 SET_OBJ_SIZE(dev_ops, ib_ah);
2696 SET_OBJ_SIZE(dev_ops, ib_counters);
2697 SET_OBJ_SIZE(dev_ops, ib_cq);
2698 SET_OBJ_SIZE(dev_ops, ib_pd);
2699 SET_OBJ_SIZE(dev_ops, ib_srq);
2700 SET_OBJ_SIZE(dev_ops, ib_ucontext);
2701 SET_OBJ_SIZE(dev_ops, ib_xrcd);
2703 EXPORT_SYMBOL(ib_set_device_ops);
2705 static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2706 [RDMA_NL_LS_OP_RESOLVE] = {
2707 .doit = ib_nl_handle_resolve_resp,
2708 .flags = RDMA_NL_ADMIN_PERM,
2710 [RDMA_NL_LS_OP_SET_TIMEOUT] = {
2711 .doit = ib_nl_handle_set_timeout,
2712 .flags = RDMA_NL_ADMIN_PERM,
2714 [RDMA_NL_LS_OP_IP_RESOLVE] = {
2715 .doit = ib_nl_handle_ip_res_resp,
2716 .flags = RDMA_NL_ADMIN_PERM,
2720 static int __init ib_core_init(void)
2724 ib_wq = alloc_workqueue("infiniband", 0, 0);
2728 ib_comp_wq = alloc_workqueue("ib-comp-wq",
2729 WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2735 ib_comp_unbound_wq =
2736 alloc_workqueue("ib-comp-unb-wq",
2737 WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2738 WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2739 if (!ib_comp_unbound_wq) {
2744 ret = class_register(&ib_class);
2746 pr_warn("Couldn't create InfiniBand device class\n");
2747 goto err_comp_unbound;
2754 pr_warn("Could't init IB address resolution\n");
2758 ret = ib_mad_init();
2760 pr_warn("Couldn't init IB MAD\n");
2766 pr_warn("Couldn't init SA\n");
2770 ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2772 pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2776 ret = register_pernet_device(&rdma_dev_net_ops);
2778 pr_warn("Couldn't init compat dev. ret %d\n", ret);
2783 rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2784 roce_gid_mgmt_init();
2789 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2797 class_unregister(&ib_class);
2799 destroy_workqueue(ib_comp_unbound_wq);
2801 destroy_workqueue(ib_comp_wq);
2803 destroy_workqueue(ib_wq);
2807 static void __exit ib_core_cleanup(void)
2809 roce_gid_mgmt_cleanup();
2811 rdma_nl_unregister(RDMA_NL_LS);
2812 unregister_pernet_device(&rdma_dev_net_ops);
2813 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2818 class_unregister(&ib_class);
2819 destroy_workqueue(ib_comp_unbound_wq);
2820 destroy_workqueue(ib_comp_wq);
2821 /* Make sure that any pending umem accounting work is done. */
2822 destroy_workqueue(ib_wq);
2823 flush_workqueue(system_unbound_wq);
2824 WARN_ON(!xa_empty(&clients));
2825 WARN_ON(!xa_empty(&devices));
2828 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
2830 /* ib core relies on netdev stack to first register net_ns_type_operations
2831 * ns kobject type before ib_core initialization.
2833 fs_initcall(ib_core_init);
2834 module_exit(ib_core_cleanup);