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(query_pkey),
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(dereg_mr),
289 IB_MANDATORY_FUNC(get_port_immutable)
293 device->kverbs_provider = true;
294 for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
295 if (!*(void **) ((void *) &device->ops +
296 mandatory_table[i].offset)) {
297 device->kverbs_provider = false;
304 * Caller must perform ib_device_put() to return the device reference count
305 * when ib_device_get_by_index() returns valid device pointer.
307 struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
309 struct ib_device *device;
311 down_read(&devices_rwsem);
312 device = xa_load(&devices, index);
314 if (!rdma_dev_access_netns(device, net)) {
319 if (!ib_device_try_get(device))
323 up_read(&devices_rwsem);
328 * ib_device_put - Release IB device reference
329 * @device: device whose reference to be released
331 * ib_device_put() releases reference to the IB device to allow it to be
332 * unregistered and eventually free.
334 void ib_device_put(struct ib_device *device)
336 if (refcount_dec_and_test(&device->refcount))
337 complete(&device->unreg_completion);
339 EXPORT_SYMBOL(ib_device_put);
341 static struct ib_device *__ib_device_get_by_name(const char *name)
343 struct ib_device *device;
346 xa_for_each (&devices, index, device)
347 if (!strcmp(name, dev_name(&device->dev)))
354 * ib_device_get_by_name - Find an IB device by name
355 * @name: The name to look for
356 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
358 * Find and hold an ib_device by its name. The caller must call
359 * ib_device_put() on the returned pointer.
361 struct ib_device *ib_device_get_by_name(const char *name,
362 enum rdma_driver_id driver_id)
364 struct ib_device *device;
366 down_read(&devices_rwsem);
367 device = __ib_device_get_by_name(name);
368 if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
369 device->ops.driver_id != driver_id)
373 if (!ib_device_try_get(device))
376 up_read(&devices_rwsem);
379 EXPORT_SYMBOL(ib_device_get_by_name);
381 static int rename_compat_devs(struct ib_device *device)
383 struct ib_core_device *cdev;
387 mutex_lock(&device->compat_devs_mutex);
388 xa_for_each (&device->compat_devs, index, cdev) {
389 ret = device_rename(&cdev->dev, dev_name(&device->dev));
392 "Fail to rename compatdev to new name %s\n",
393 dev_name(&device->dev));
397 mutex_unlock(&device->compat_devs_mutex);
401 int ib_device_rename(struct ib_device *ibdev, const char *name)
407 down_write(&devices_rwsem);
408 if (!strcmp(name, dev_name(&ibdev->dev))) {
409 up_write(&devices_rwsem);
413 if (__ib_device_get_by_name(name)) {
414 up_write(&devices_rwsem);
418 ret = device_rename(&ibdev->dev, name);
420 up_write(&devices_rwsem);
424 strlcpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
425 ret = rename_compat_devs(ibdev);
427 downgrade_write(&devices_rwsem);
428 down_read(&ibdev->client_data_rwsem);
429 xan_for_each_marked(&ibdev->client_data, index, client_data,
430 CLIENT_DATA_REGISTERED) {
431 struct ib_client *client = xa_load(&clients, index);
433 if (!client || !client->rename)
436 client->rename(ibdev, client_data);
438 up_read(&ibdev->client_data_rwsem);
439 up_read(&devices_rwsem);
443 int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
447 ibdev->use_cq_dim = use_dim;
452 static int alloc_name(struct ib_device *ibdev, const char *name)
454 struct ib_device *device;
460 lockdep_assert_held_write(&devices_rwsem);
462 xa_for_each (&devices, index, device) {
463 char buf[IB_DEVICE_NAME_MAX];
465 if (sscanf(dev_name(&device->dev), name, &i) != 1)
467 if (i < 0 || i >= INT_MAX)
469 snprintf(buf, sizeof buf, name, i);
470 if (strcmp(buf, dev_name(&device->dev)) != 0)
473 rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
478 rc = ida_alloc(&inuse, GFP_KERNEL);
482 rc = dev_set_name(&ibdev->dev, name, rc);
488 static void ib_device_release(struct device *device)
490 struct ib_device *dev = container_of(device, struct ib_device, dev);
493 WARN_ON(refcount_read(&dev->refcount));
494 if (dev->port_data) {
495 ib_cache_release_one(dev);
496 ib_security_release_port_pkey_list(dev);
497 rdma_counter_release(dev);
498 kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
503 mutex_destroy(&dev->unregistration_lock);
504 mutex_destroy(&dev->compat_devs_mutex);
506 xa_destroy(&dev->compat_devs);
507 xa_destroy(&dev->client_data);
508 kfree_rcu(dev, rcu_head);
511 static int ib_device_uevent(struct device *device,
512 struct kobj_uevent_env *env)
514 if (add_uevent_var(env, "NAME=%s", dev_name(device)))
518 * It would be nice to pass the node GUID with the event...
524 static const void *net_namespace(struct device *d)
526 struct ib_core_device *coredev =
527 container_of(d, struct ib_core_device, dev);
529 return read_pnet(&coredev->rdma_net);
532 static struct class ib_class = {
533 .name = "infiniband",
534 .dev_release = ib_device_release,
535 .dev_uevent = ib_device_uevent,
536 .ns_type = &net_ns_type_operations,
537 .namespace = net_namespace,
540 static void rdma_init_coredev(struct ib_core_device *coredev,
541 struct ib_device *dev, struct net *net)
543 /* This BUILD_BUG_ON is intended to catch layout change
544 * of union of ib_core_device and device.
545 * dev must be the first element as ib_core and providers
546 * driver uses it. Adding anything in ib_core_device before
547 * device will break this assumption.
549 BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
550 offsetof(struct ib_device, dev));
552 coredev->dev.class = &ib_class;
553 coredev->dev.groups = dev->groups;
554 device_initialize(&coredev->dev);
555 coredev->owner = dev;
556 INIT_LIST_HEAD(&coredev->port_list);
557 write_pnet(&coredev->rdma_net, net);
561 * _ib_alloc_device - allocate an IB device struct
562 * @size:size of structure to allocate
564 * Low-level drivers should use ib_alloc_device() to allocate &struct
565 * ib_device. @size is the size of the structure to be allocated,
566 * including any private data used by the low-level driver.
567 * ib_dealloc_device() must be used to free structures allocated with
570 struct ib_device *_ib_alloc_device(size_t size)
572 struct ib_device *device;
574 if (WARN_ON(size < sizeof(struct ib_device)))
577 device = kzalloc(size, GFP_KERNEL);
581 if (rdma_restrack_init(device)) {
586 device->groups[0] = &ib_dev_attr_group;
587 rdma_init_coredev(&device->coredev, device, &init_net);
589 INIT_LIST_HEAD(&device->event_handler_list);
590 spin_lock_init(&device->event_handler_lock);
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);
682 /* Readers shall not see a client until add has been completed */
683 xa_set_mark(&device->client_data, client->client_id,
684 CLIENT_DATA_REGISTERED);
685 up_read(&device->client_data_rwsem);
689 ib_device_put(device);
690 ib_client_put(client);
692 up_write(&device->client_data_rwsem);
696 static void remove_client_context(struct ib_device *device,
697 unsigned int client_id)
699 struct ib_client *client;
702 down_write(&device->client_data_rwsem);
703 if (!xa_get_mark(&device->client_data, client_id,
704 CLIENT_DATA_REGISTERED)) {
705 up_write(&device->client_data_rwsem);
708 client_data = xa_load(&device->client_data, client_id);
709 xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
710 client = xa_load(&clients, client_id);
711 up_write(&device->client_data_rwsem);
714 * Notice we cannot be holding any exclusive locks when calling the
715 * remove callback as the remove callback can recurse back into any
716 * public functions in this module and thus try for any locks those
719 * For this reason clients and drivers should not call the
720 * unregistration functions will holdling any locks.
723 client->remove(device, client_data);
725 xa_erase(&device->client_data, client_id);
726 ib_device_put(device);
727 ib_client_put(client);
730 static int alloc_port_data(struct ib_device *device)
732 struct ib_port_data_rcu *pdata_rcu;
735 if (device->port_data)
738 /* This can only be called once the physical port range is defined */
739 if (WARN_ON(!device->phys_port_cnt))
743 * device->port_data is indexed directly by the port number to make
744 * access to this data as efficient as possible.
746 * Therefore port_data is declared as a 1 based array with potential
747 * empty slots at the beginning.
749 pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
750 rdma_end_port(device) + 1),
755 * The rcu_head is put in front of the port data array and the stored
756 * pointer is adjusted since we never need to see that member until
759 device->port_data = pdata_rcu->pdata;
761 rdma_for_each_port (device, port) {
762 struct ib_port_data *pdata = &device->port_data[port];
764 pdata->ib_dev = device;
765 spin_lock_init(&pdata->pkey_list_lock);
766 INIT_LIST_HEAD(&pdata->pkey_list);
767 spin_lock_init(&pdata->netdev_lock);
768 INIT_HLIST_NODE(&pdata->ndev_hash_link);
773 static int verify_immutable(const struct ib_device *dev, u8 port)
775 return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
776 rdma_max_mad_size(dev, port) != 0);
779 static int setup_port_data(struct ib_device *device)
784 ret = alloc_port_data(device);
788 rdma_for_each_port (device, port) {
789 struct ib_port_data *pdata = &device->port_data[port];
791 ret = device->ops.get_port_immutable(device, port,
796 if (verify_immutable(device, port))
802 void ib_get_device_fw_str(struct ib_device *dev, char *str)
804 if (dev->ops.get_dev_fw_str)
805 dev->ops.get_dev_fw_str(dev, str);
809 EXPORT_SYMBOL(ib_get_device_fw_str);
811 static void ib_policy_change_task(struct work_struct *work)
813 struct ib_device *dev;
816 down_read(&devices_rwsem);
817 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
820 rdma_for_each_port (dev, i) {
822 int ret = ib_get_cached_subnet_prefix(dev,
827 "ib_get_cached_subnet_prefix err: %d, this should never happen here\n",
830 ib_security_cache_change(dev, i, sp);
833 up_read(&devices_rwsem);
836 static int ib_security_change(struct notifier_block *nb, unsigned long event,
839 if (event != LSM_POLICY_CHANGE)
842 schedule_work(&ib_policy_change_work);
843 ib_mad_agent_security_change();
848 static void compatdev_release(struct device *dev)
850 struct ib_core_device *cdev =
851 container_of(dev, struct ib_core_device, dev);
856 static int add_one_compat_dev(struct ib_device *device,
857 struct rdma_dev_net *rnet)
859 struct ib_core_device *cdev;
862 lockdep_assert_held(&rdma_nets_rwsem);
863 if (!ib_devices_shared_netns)
867 * Create and add compat device in all namespaces other than where it
868 * is currently bound to.
870 if (net_eq(read_pnet(&rnet->net),
871 read_pnet(&device->coredev.rdma_net)))
875 * The first of init_net() or ib_register_device() to take the
876 * compat_devs_mutex wins and gets to add the device. Others will wait
877 * for completion here.
879 mutex_lock(&device->compat_devs_mutex);
880 cdev = xa_load(&device->compat_devs, rnet->id);
885 ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
889 cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
895 cdev->dev.parent = device->dev.parent;
896 rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
897 cdev->dev.release = compatdev_release;
898 dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
900 ret = device_add(&cdev->dev);
903 ret = ib_setup_port_attrs(cdev);
907 ret = xa_err(xa_store(&device->compat_devs, rnet->id,
912 mutex_unlock(&device->compat_devs_mutex);
916 ib_free_port_attrs(cdev);
918 device_del(&cdev->dev);
920 put_device(&cdev->dev);
922 xa_release(&device->compat_devs, rnet->id);
924 mutex_unlock(&device->compat_devs_mutex);
928 static void remove_one_compat_dev(struct ib_device *device, u32 id)
930 struct ib_core_device *cdev;
932 mutex_lock(&device->compat_devs_mutex);
933 cdev = xa_erase(&device->compat_devs, id);
934 mutex_unlock(&device->compat_devs_mutex);
936 ib_free_port_attrs(cdev);
937 device_del(&cdev->dev);
938 put_device(&cdev->dev);
942 static void remove_compat_devs(struct ib_device *device)
944 struct ib_core_device *cdev;
947 xa_for_each (&device->compat_devs, index, cdev)
948 remove_one_compat_dev(device, index);
951 static int add_compat_devs(struct ib_device *device)
953 struct rdma_dev_net *rnet;
957 lockdep_assert_held(&devices_rwsem);
959 down_read(&rdma_nets_rwsem);
960 xa_for_each (&rdma_nets, index, rnet) {
961 ret = add_one_compat_dev(device, rnet);
965 up_read(&rdma_nets_rwsem);
969 static void remove_all_compat_devs(void)
971 struct ib_compat_device *cdev;
972 struct ib_device *dev;
975 down_read(&devices_rwsem);
976 xa_for_each (&devices, index, dev) {
977 unsigned long c_index = 0;
979 /* Hold nets_rwsem so that any other thread modifying this
980 * system param can sync with this thread.
982 down_read(&rdma_nets_rwsem);
983 xa_for_each (&dev->compat_devs, c_index, cdev)
984 remove_one_compat_dev(dev, c_index);
985 up_read(&rdma_nets_rwsem);
987 up_read(&devices_rwsem);
990 static int add_all_compat_devs(void)
992 struct rdma_dev_net *rnet;
993 struct ib_device *dev;
997 down_read(&devices_rwsem);
998 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
999 unsigned long net_index = 0;
1001 /* Hold nets_rwsem so that any other thread modifying this
1002 * system param can sync with this thread.
1004 down_read(&rdma_nets_rwsem);
1005 xa_for_each (&rdma_nets, net_index, rnet) {
1006 ret = add_one_compat_dev(dev, rnet);
1010 up_read(&rdma_nets_rwsem);
1012 up_read(&devices_rwsem);
1014 remove_all_compat_devs();
1018 int rdma_compatdev_set(u8 enable)
1020 struct rdma_dev_net *rnet;
1021 unsigned long index;
1024 down_write(&rdma_nets_rwsem);
1025 if (ib_devices_shared_netns == enable) {
1026 up_write(&rdma_nets_rwsem);
1030 /* enable/disable of compat devices is not supported
1031 * when more than default init_net exists.
1033 xa_for_each (&rdma_nets, index, rnet) {
1038 ib_devices_shared_netns = enable;
1039 up_write(&rdma_nets_rwsem);
1044 ret = add_all_compat_devs();
1046 remove_all_compat_devs();
1050 static void rdma_dev_exit_net(struct net *net)
1052 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1053 struct ib_device *dev;
1054 unsigned long index;
1057 down_write(&rdma_nets_rwsem);
1059 * Prevent the ID from being re-used and hide the id from xa_for_each.
1061 ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1063 up_write(&rdma_nets_rwsem);
1065 down_read(&devices_rwsem);
1066 xa_for_each (&devices, index, dev) {
1067 get_device(&dev->dev);
1069 * Release the devices_rwsem so that pontentially blocking
1070 * device_del, doesn't hold the devices_rwsem for too long.
1072 up_read(&devices_rwsem);
1074 remove_one_compat_dev(dev, rnet->id);
1077 * If the real device is in the NS then move it back to init.
1079 rdma_dev_change_netns(dev, net, &init_net);
1081 put_device(&dev->dev);
1082 down_read(&devices_rwsem);
1084 up_read(&devices_rwsem);
1086 rdma_nl_net_exit(rnet);
1087 xa_erase(&rdma_nets, rnet->id);
1090 static __net_init int rdma_dev_init_net(struct net *net)
1092 struct rdma_dev_net *rnet = rdma_net_to_dev_net(net);
1093 unsigned long index;
1094 struct ib_device *dev;
1097 write_pnet(&rnet->net, net);
1099 ret = rdma_nl_net_init(rnet);
1103 /* No need to create any compat devices in default init_net. */
1104 if (net_eq(net, &init_net))
1107 ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1109 rdma_nl_net_exit(rnet);
1113 down_read(&devices_rwsem);
1114 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1115 /* Hold nets_rwsem so that netlink command cannot change
1116 * system configuration for device sharing mode.
1118 down_read(&rdma_nets_rwsem);
1119 ret = add_one_compat_dev(dev, rnet);
1120 up_read(&rdma_nets_rwsem);
1124 up_read(&devices_rwsem);
1127 rdma_dev_exit_net(net);
1133 * Assign the unique string device name and the unique device index. This is
1134 * undone by ib_dealloc_device.
1136 static int assign_name(struct ib_device *device, const char *name)
1141 down_write(&devices_rwsem);
1142 /* Assign a unique name to the device */
1143 if (strchr(name, '%'))
1144 ret = alloc_name(device, name);
1146 ret = dev_set_name(&device->dev, name);
1150 if (__ib_device_get_by_name(dev_name(&device->dev))) {
1154 strlcpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1156 ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1157 &last_id, GFP_KERNEL);
1162 up_write(&devices_rwsem);
1166 static void setup_dma_device(struct ib_device *device)
1168 struct device *parent = device->dev.parent;
1170 WARN_ON_ONCE(device->dma_device);
1171 if (device->dev.dma_ops) {
1173 * The caller provided custom DMA operations. Copy the
1174 * DMA-related fields that are used by e.g. dma_alloc_coherent()
1177 device->dma_device = &device->dev;
1178 if (!device->dev.dma_mask) {
1180 device->dev.dma_mask = parent->dma_mask;
1184 if (!device->dev.coherent_dma_mask) {
1186 device->dev.coherent_dma_mask =
1187 parent->coherent_dma_mask;
1193 * The caller did not provide custom DMA operations. Use the
1194 * DMA mapping operations of the parent device.
1196 WARN_ON_ONCE(!parent);
1197 device->dma_device = parent;
1200 if (!device->dev.dma_parms) {
1203 * The caller did not provide DMA parameters, so
1204 * 'parent' probably represents a PCI device. The PCI
1205 * core sets the maximum segment size to 64
1206 * KB. Increase this parameter to 2 GB.
1208 device->dev.dma_parms = parent->dma_parms;
1209 dma_set_max_seg_size(device->dma_device, SZ_2G);
1217 * setup_device() allocates memory and sets up data that requires calling the
1218 * device ops, this is the only reason these actions are not done during
1219 * ib_alloc_device. It is undone by ib_dealloc_device().
1221 static int setup_device(struct ib_device *device)
1223 struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1226 setup_dma_device(device);
1227 ib_device_check_mandatory(device);
1229 ret = setup_port_data(device);
1231 dev_warn(&device->dev, "Couldn't create per-port data\n");
1235 memset(&device->attrs, 0, sizeof(device->attrs));
1236 ret = device->ops.query_device(device, &device->attrs, &uhw);
1238 dev_warn(&device->dev,
1239 "Couldn't query the device attributes\n");
1246 static void disable_device(struct ib_device *device)
1250 WARN_ON(!refcount_read(&device->refcount));
1252 down_write(&devices_rwsem);
1253 xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1254 up_write(&devices_rwsem);
1257 * Remove clients in LIFO order, see assign_client_id. This could be
1258 * more efficient if xarray learns to reverse iterate. Since no new
1259 * clients can be added to this ib_device past this point we only need
1260 * the maximum possible client_id value here.
1262 down_read(&clients_rwsem);
1263 cid = highest_client_id;
1264 up_read(&clients_rwsem);
1267 remove_client_context(device, cid);
1270 /* Pairs with refcount_set in enable_device */
1271 ib_device_put(device);
1272 wait_for_completion(&device->unreg_completion);
1275 * compat devices must be removed after device refcount drops to zero.
1276 * Otherwise init_net() may add more compatdevs after removing compat
1277 * devices and before device is disabled.
1279 remove_compat_devs(device);
1283 * An enabled device is visible to all clients and to all the public facing
1284 * APIs that return a device pointer. This always returns with a new get, even
1287 static int enable_device_and_get(struct ib_device *device)
1289 struct ib_client *client;
1290 unsigned long index;
1294 * One ref belongs to the xa and the other belongs to this
1295 * thread. This is needed to guard against parallel unregistration.
1297 refcount_set(&device->refcount, 2);
1298 down_write(&devices_rwsem);
1299 xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1302 * By using downgrade_write() we ensure that no other thread can clear
1303 * DEVICE_REGISTERED while we are completing the client setup.
1305 downgrade_write(&devices_rwsem);
1307 if (device->ops.enable_driver) {
1308 ret = device->ops.enable_driver(device);
1313 down_read(&clients_rwsem);
1314 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1315 ret = add_client_context(device, client);
1319 up_read(&clients_rwsem);
1321 ret = add_compat_devs(device);
1323 up_read(&devices_rwsem);
1328 * ib_register_device - Register an IB device with IB core
1329 * @device: Device to register
1330 * @name: unique string device name. This may include a '%' which will
1331 * cause a unique index to be added to the passed device name.
1333 * Low-level drivers use ib_register_device() to register their
1334 * devices with the IB core. All registered clients will receive a
1335 * callback for each device that is added. @device must be allocated
1336 * with ib_alloc_device().
1338 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1339 * asynchronously then the device pointer may become freed as soon as this
1342 int ib_register_device(struct ib_device *device, const char *name)
1346 ret = assign_name(device, name);
1350 ret = setup_device(device);
1354 ret = ib_cache_setup_one(device);
1356 dev_warn(&device->dev,
1357 "Couldn't set up InfiniBand P_Key/GID cache\n");
1361 ib_device_register_rdmacg(device);
1363 rdma_counter_init(device);
1366 * Ensure that ADD uevent is not fired because it
1367 * is too early amd device is not initialized yet.
1369 dev_set_uevent_suppress(&device->dev, true);
1370 ret = device_add(&device->dev);
1374 ret = ib_device_register_sysfs(device);
1376 dev_warn(&device->dev,
1377 "Couldn't register device with driver model\n");
1381 ret = enable_device_and_get(device);
1382 dev_set_uevent_suppress(&device->dev, false);
1383 /* Mark for userspace that device is ready */
1384 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1386 void (*dealloc_fn)(struct ib_device *);
1389 * If we hit this error flow then we don't want to
1390 * automatically dealloc the device since the caller is
1391 * expected to call ib_dealloc_device() after
1392 * ib_register_device() fails. This is tricky due to the
1393 * possibility for a parallel unregistration along with this
1394 * error flow. Since we have a refcount here we know any
1395 * parallel flow is stopped in disable_device and will see the
1396 * NULL pointers, causing the responsibility to
1397 * ib_dealloc_device() to revert back to this thread.
1399 dealloc_fn = device->ops.dealloc_driver;
1400 device->ops.dealloc_driver = NULL;
1401 ib_device_put(device);
1402 __ib_unregister_device(device);
1403 device->ops.dealloc_driver = dealloc_fn;
1406 ib_device_put(device);
1411 device_del(&device->dev);
1413 dev_set_uevent_suppress(&device->dev, false);
1414 ib_device_unregister_rdmacg(device);
1415 ib_cache_cleanup_one(device);
1418 EXPORT_SYMBOL(ib_register_device);
1420 /* Callers must hold a get on the device. */
1421 static void __ib_unregister_device(struct ib_device *ib_dev)
1424 * We have a registration lock so that all the calls to unregister are
1425 * fully fenced, once any unregister returns the device is truely
1426 * unregistered even if multiple callers are unregistering it at the
1427 * same time. This also interacts with the registration flow and
1428 * provides sane semantics if register and unregister are racing.
1430 mutex_lock(&ib_dev->unregistration_lock);
1431 if (!refcount_read(&ib_dev->refcount))
1434 disable_device(ib_dev);
1436 /* Expedite removing unregistered pointers from the hash table */
1437 free_netdevs(ib_dev);
1439 ib_device_unregister_sysfs(ib_dev);
1440 device_del(&ib_dev->dev);
1441 ib_device_unregister_rdmacg(ib_dev);
1442 ib_cache_cleanup_one(ib_dev);
1445 * Drivers using the new flow may not call ib_dealloc_device except
1446 * in error unwind prior to registration success.
1448 if (ib_dev->ops.dealloc_driver) {
1449 WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1450 ib_dealloc_device(ib_dev);
1453 mutex_unlock(&ib_dev->unregistration_lock);
1457 * ib_unregister_device - Unregister an IB device
1458 * @ib_dev: The device to unregister
1460 * Unregister an IB device. All clients will receive a remove callback.
1462 * Callers should call this routine only once, and protect against races with
1463 * registration. Typically it should only be called as part of a remove
1464 * callback in an implementation of driver core's struct device_driver and
1467 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1470 void ib_unregister_device(struct ib_device *ib_dev)
1472 get_device(&ib_dev->dev);
1473 __ib_unregister_device(ib_dev);
1474 put_device(&ib_dev->dev);
1476 EXPORT_SYMBOL(ib_unregister_device);
1479 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1480 * @ib_dev: The device to unregister
1482 * This is the same as ib_unregister_device(), except it includes an internal
1483 * ib_device_put() that should match a 'get' obtained by the caller.
1485 * It is safe to call this routine concurrently from multiple threads while
1486 * holding the 'get'. When the function returns the device is fully
1489 * Drivers using this flow MUST use the driver_unregister callback to clean up
1490 * their resources associated with the device and dealloc it.
1492 void ib_unregister_device_and_put(struct ib_device *ib_dev)
1494 WARN_ON(!ib_dev->ops.dealloc_driver);
1495 get_device(&ib_dev->dev);
1496 ib_device_put(ib_dev);
1497 __ib_unregister_device(ib_dev);
1498 put_device(&ib_dev->dev);
1500 EXPORT_SYMBOL(ib_unregister_device_and_put);
1503 * ib_unregister_driver - Unregister all IB devices for a driver
1504 * @driver_id: The driver to unregister
1506 * This implements a fence for device unregistration. It only returns once all
1507 * devices associated with the driver_id have fully completed their
1508 * unregistration and returned from ib_unregister_device*().
1510 * If device's are not yet unregistered it goes ahead and starts unregistering
1513 * This does not block creation of new devices with the given driver_id, that
1514 * is the responsibility of the caller.
1516 void ib_unregister_driver(enum rdma_driver_id driver_id)
1518 struct ib_device *ib_dev;
1519 unsigned long index;
1521 down_read(&devices_rwsem);
1522 xa_for_each (&devices, index, ib_dev) {
1523 if (ib_dev->ops.driver_id != driver_id)
1526 get_device(&ib_dev->dev);
1527 up_read(&devices_rwsem);
1529 WARN_ON(!ib_dev->ops.dealloc_driver);
1530 __ib_unregister_device(ib_dev);
1532 put_device(&ib_dev->dev);
1533 down_read(&devices_rwsem);
1535 up_read(&devices_rwsem);
1537 EXPORT_SYMBOL(ib_unregister_driver);
1539 static void ib_unregister_work(struct work_struct *work)
1541 struct ib_device *ib_dev =
1542 container_of(work, struct ib_device, unregistration_work);
1544 __ib_unregister_device(ib_dev);
1545 put_device(&ib_dev->dev);
1549 * ib_unregister_device_queued - Unregister a device using a work queue
1550 * @ib_dev: The device to unregister
1552 * This schedules an asynchronous unregistration using a WQ for the device. A
1553 * driver should use this to avoid holding locks while doing unregistration,
1554 * such as holding the RTNL lock.
1556 * Drivers using this API must use ib_unregister_driver before module unload
1557 * to ensure that all scheduled unregistrations have completed.
1559 void ib_unregister_device_queued(struct ib_device *ib_dev)
1561 WARN_ON(!refcount_read(&ib_dev->refcount));
1562 WARN_ON(!ib_dev->ops.dealloc_driver);
1563 get_device(&ib_dev->dev);
1564 if (!queue_work(system_unbound_wq, &ib_dev->unregistration_work))
1565 put_device(&ib_dev->dev);
1567 EXPORT_SYMBOL(ib_unregister_device_queued);
1570 * The caller must pass in a device that has the kref held and the refcount
1571 * released. If the device is in cur_net and still registered then it is moved
1574 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1580 mutex_lock(&device->unregistration_lock);
1583 * If a device not under ib_device_get() or if the unregistration_lock
1584 * is not held, the namespace can be changed, or it can be unregistered.
1585 * Check again under the lock.
1587 if (refcount_read(&device->refcount) == 0 ||
1588 !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1593 kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1594 disable_device(device);
1597 * At this point no one can be using the device, so it is safe to
1598 * change the namespace.
1600 write_pnet(&device->coredev.rdma_net, net);
1602 down_read(&devices_rwsem);
1604 * Currently rdma devices are system wide unique. So the device name
1605 * is guaranteed free in the new namespace. Publish the new namespace
1606 * at the sysfs level.
1608 ret = device_rename(&device->dev, dev_name(&device->dev));
1609 up_read(&devices_rwsem);
1611 dev_warn(&device->dev,
1612 "%s: Couldn't rename device after namespace change\n",
1614 /* Try and put things back and re-enable the device */
1615 write_pnet(&device->coredev.rdma_net, cur_net);
1618 ret2 = enable_device_and_get(device);
1621 * This shouldn't really happen, but if it does, let the user
1622 * retry at later point. So don't disable the device.
1624 dev_warn(&device->dev,
1625 "%s: Couldn't re-enable device after namespace change\n",
1628 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1630 ib_device_put(device);
1632 mutex_unlock(&device->unregistration_lock);
1638 int ib_device_set_netns_put(struct sk_buff *skb,
1639 struct ib_device *dev, u32 ns_fd)
1644 net = get_net_ns_by_fd(ns_fd);
1650 if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1656 * Currently supported only for those providers which support
1657 * disassociation and don't do port specific sysfs init. Once a
1658 * port_cleanup infrastructure is implemented, this limitation will be
1661 if (!dev->ops.disassociate_ucontext || dev->ops.init_port ||
1662 ib_devices_shared_netns) {
1667 get_device(&dev->dev);
1669 ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1670 put_device(&dev->dev);
1682 static struct pernet_operations rdma_dev_net_ops = {
1683 .init = rdma_dev_init_net,
1684 .exit = rdma_dev_exit_net,
1685 .id = &rdma_dev_net_id,
1686 .size = sizeof(struct rdma_dev_net),
1689 static int assign_client_id(struct ib_client *client)
1693 down_write(&clients_rwsem);
1695 * The add/remove callbacks must be called in FIFO/LIFO order. To
1696 * achieve this we assign client_ids so they are sorted in
1697 * registration order.
1699 client->client_id = highest_client_id;
1700 ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1704 highest_client_id++;
1705 xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1708 up_write(&clients_rwsem);
1712 static void remove_client_id(struct ib_client *client)
1714 down_write(&clients_rwsem);
1715 xa_erase(&clients, client->client_id);
1716 for (; highest_client_id; highest_client_id--)
1717 if (xa_load(&clients, highest_client_id - 1))
1719 up_write(&clients_rwsem);
1723 * ib_register_client - Register an IB client
1724 * @client:Client to register
1726 * Upper level users of the IB drivers can use ib_register_client() to
1727 * register callbacks for IB device addition and removal. When an IB
1728 * device is added, each registered client's add method will be called
1729 * (in the order the clients were registered), and when a device is
1730 * removed, each client's remove method will be called (in the reverse
1731 * order that clients were registered). In addition, when
1732 * ib_register_client() is called, the client will receive an add
1733 * callback for all devices already registered.
1735 int ib_register_client(struct ib_client *client)
1737 struct ib_device *device;
1738 unsigned long index;
1741 refcount_set(&client->uses, 1);
1742 init_completion(&client->uses_zero);
1743 ret = assign_client_id(client);
1747 down_read(&devices_rwsem);
1748 xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1749 ret = add_client_context(device, client);
1751 up_read(&devices_rwsem);
1752 ib_unregister_client(client);
1756 up_read(&devices_rwsem);
1759 EXPORT_SYMBOL(ib_register_client);
1762 * ib_unregister_client - Unregister an IB client
1763 * @client:Client to unregister
1765 * Upper level users use ib_unregister_client() to remove their client
1766 * registration. When ib_unregister_client() is called, the client
1767 * will receive a remove callback for each IB device still registered.
1769 * This is a full fence, once it returns no client callbacks will be called,
1770 * or are running in another thread.
1772 void ib_unregister_client(struct ib_client *client)
1774 struct ib_device *device;
1775 unsigned long index;
1777 down_write(&clients_rwsem);
1778 ib_client_put(client);
1779 xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1780 up_write(&clients_rwsem);
1782 /* We do not want to have locks while calling client->remove() */
1784 xa_for_each (&devices, index, device) {
1785 if (!ib_device_try_get(device))
1789 remove_client_context(device, client->client_id);
1791 ib_device_put(device);
1797 * remove_client_context() is not a fence, it can return even though a
1798 * removal is ongoing. Wait until all removals are completed.
1800 wait_for_completion(&client->uses_zero);
1801 remove_client_id(client);
1803 EXPORT_SYMBOL(ib_unregister_client);
1805 static int __ib_get_global_client_nl_info(const char *client_name,
1806 struct ib_client_nl_info *res)
1808 struct ib_client *client;
1809 unsigned long index;
1812 down_read(&clients_rwsem);
1813 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1814 if (strcmp(client->name, client_name) != 0)
1816 if (!client->get_global_nl_info) {
1820 ret = client->get_global_nl_info(res);
1821 if (WARN_ON(ret == -ENOENT))
1823 if (!ret && res->cdev)
1824 get_device(res->cdev);
1827 up_read(&clients_rwsem);
1831 static int __ib_get_client_nl_info(struct ib_device *ibdev,
1832 const char *client_name,
1833 struct ib_client_nl_info *res)
1835 unsigned long index;
1839 down_read(&ibdev->client_data_rwsem);
1840 xan_for_each_marked (&ibdev->client_data, index, client_data,
1841 CLIENT_DATA_REGISTERED) {
1842 struct ib_client *client = xa_load(&clients, index);
1844 if (!client || strcmp(client->name, client_name) != 0)
1846 if (!client->get_nl_info) {
1850 ret = client->get_nl_info(ibdev, client_data, res);
1851 if (WARN_ON(ret == -ENOENT))
1855 * The cdev is guaranteed valid as long as we are inside the
1856 * client_data_rwsem as remove_one can't be called. Keep it
1857 * valid for the caller.
1859 if (!ret && res->cdev)
1860 get_device(res->cdev);
1863 up_read(&ibdev->client_data_rwsem);
1869 * ib_get_client_nl_info - Fetch the nl_info from a client
1870 * @device - IB device
1871 * @client_name - Name of the client
1872 * @res - Result of the query
1874 int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1875 struct ib_client_nl_info *res)
1880 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1882 ret = __ib_get_global_client_nl_info(client_name, res);
1883 #ifdef CONFIG_MODULES
1884 if (ret == -ENOENT) {
1885 request_module("rdma-client-%s", client_name);
1887 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1889 ret = __ib_get_global_client_nl_info(client_name, res);
1898 if (WARN_ON(!res->cdev))
1904 * ib_set_client_data - Set IB client context
1905 * @device:Device to set context for
1906 * @client:Client to set context for
1907 * @data:Context to set
1909 * ib_set_client_data() sets client context data that can be retrieved with
1910 * ib_get_client_data(). This can only be called while the client is
1911 * registered to the device, once the ib_client remove() callback returns this
1914 void ib_set_client_data(struct ib_device *device, struct ib_client *client,
1919 if (WARN_ON(IS_ERR(data)))
1922 rc = xa_store(&device->client_data, client->client_id, data,
1924 WARN_ON(xa_is_err(rc));
1926 EXPORT_SYMBOL(ib_set_client_data);
1929 * ib_register_event_handler - Register an IB event handler
1930 * @event_handler:Handler to register
1932 * ib_register_event_handler() registers an event handler that will be
1933 * called back when asynchronous IB events occur (as defined in
1934 * chapter 11 of the InfiniBand Architecture Specification). This
1935 * callback may occur in interrupt context.
1937 void ib_register_event_handler(struct ib_event_handler *event_handler)
1939 unsigned long flags;
1941 spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
1942 list_add_tail(&event_handler->list,
1943 &event_handler->device->event_handler_list);
1944 spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
1946 EXPORT_SYMBOL(ib_register_event_handler);
1949 * ib_unregister_event_handler - Unregister an event handler
1950 * @event_handler:Handler to unregister
1952 * Unregister an event handler registered with
1953 * ib_register_event_handler().
1955 void ib_unregister_event_handler(struct ib_event_handler *event_handler)
1957 unsigned long flags;
1959 spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
1960 list_del(&event_handler->list);
1961 spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
1963 EXPORT_SYMBOL(ib_unregister_event_handler);
1966 * ib_dispatch_event - Dispatch an asynchronous event
1967 * @event:Event to dispatch
1969 * Low-level drivers must call ib_dispatch_event() to dispatch the
1970 * event to all registered event handlers when an asynchronous event
1973 void ib_dispatch_event(struct ib_event *event)
1975 unsigned long flags;
1976 struct ib_event_handler *handler;
1978 spin_lock_irqsave(&event->device->event_handler_lock, flags);
1980 list_for_each_entry(handler, &event->device->event_handler_list, list)
1981 handler->handler(handler, event);
1983 spin_unlock_irqrestore(&event->device->event_handler_lock, flags);
1985 EXPORT_SYMBOL(ib_dispatch_event);
1987 static int iw_query_port(struct ib_device *device,
1989 struct ib_port_attr *port_attr)
1991 struct in_device *inetdev;
1992 struct net_device *netdev;
1995 memset(port_attr, 0, sizeof(*port_attr));
1997 netdev = ib_device_get_netdev(device, port_num);
2001 port_attr->max_mtu = IB_MTU_4096;
2002 port_attr->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
2004 if (!netif_carrier_ok(netdev)) {
2005 port_attr->state = IB_PORT_DOWN;
2006 port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
2009 inetdev = __in_dev_get_rcu(netdev);
2011 if (inetdev && inetdev->ifa_list) {
2012 port_attr->state = IB_PORT_ACTIVE;
2013 port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
2015 port_attr->state = IB_PORT_INIT;
2016 port_attr->phys_state =
2017 IB_PORT_PHYS_STATE_PORT_CONFIGURATION_TRAINING;
2024 err = device->ops.query_port(device, port_num, port_attr);
2031 static int __ib_query_port(struct ib_device *device,
2033 struct ib_port_attr *port_attr)
2035 union ib_gid gid = {};
2038 memset(port_attr, 0, sizeof(*port_attr));
2040 err = device->ops.query_port(device, port_num, port_attr);
2041 if (err || port_attr->subnet_prefix)
2044 if (rdma_port_get_link_layer(device, port_num) !=
2045 IB_LINK_LAYER_INFINIBAND)
2048 err = device->ops.query_gid(device, port_num, 0, &gid);
2052 port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
2057 * ib_query_port - Query IB port attributes
2058 * @device:Device to query
2059 * @port_num:Port number to query
2060 * @port_attr:Port attributes
2062 * ib_query_port() returns the attributes of a port through the
2063 * @port_attr pointer.
2065 int ib_query_port(struct ib_device *device,
2067 struct ib_port_attr *port_attr)
2069 if (!rdma_is_port_valid(device, port_num))
2072 if (rdma_protocol_iwarp(device, port_num))
2073 return iw_query_port(device, port_num, port_attr);
2075 return __ib_query_port(device, port_num, port_attr);
2077 EXPORT_SYMBOL(ib_query_port);
2079 static void add_ndev_hash(struct ib_port_data *pdata)
2081 unsigned long flags;
2085 spin_lock_irqsave(&ndev_hash_lock, flags);
2086 if (hash_hashed(&pdata->ndev_hash_link)) {
2087 hash_del_rcu(&pdata->ndev_hash_link);
2088 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2090 * We cannot do hash_add_rcu after a hash_del_rcu until the
2094 spin_lock_irqsave(&ndev_hash_lock, flags);
2097 hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
2098 (uintptr_t)pdata->netdev);
2099 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2103 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2104 * @ib_dev: Device to modify
2105 * @ndev: net_device to affiliate, may be NULL
2106 * @port: IB port the net_device is connected to
2108 * Drivers should use this to link the ib_device to a netdev so the netdev
2109 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2110 * affiliated with any port.
2112 * The caller must ensure that the given ndev is not unregistered or
2113 * unregistering, and that either the ib_device is unregistered or
2114 * ib_device_set_netdev() is called with NULL when the ndev sends a
2115 * NETDEV_UNREGISTER event.
2117 int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2120 struct net_device *old_ndev;
2121 struct ib_port_data *pdata;
2122 unsigned long flags;
2126 * Drivers wish to call this before ib_register_driver, so we have to
2127 * setup the port data early.
2129 ret = alloc_port_data(ib_dev);
2133 if (!rdma_is_port_valid(ib_dev, port))
2136 pdata = &ib_dev->port_data[port];
2137 spin_lock_irqsave(&pdata->netdev_lock, flags);
2138 old_ndev = rcu_dereference_protected(
2139 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2140 if (old_ndev == ndev) {
2141 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2147 rcu_assign_pointer(pdata->netdev, ndev);
2148 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2150 add_ndev_hash(pdata);
2156 EXPORT_SYMBOL(ib_device_set_netdev);
2158 static void free_netdevs(struct ib_device *ib_dev)
2160 unsigned long flags;
2163 if (!ib_dev->port_data)
2166 rdma_for_each_port (ib_dev, port) {
2167 struct ib_port_data *pdata = &ib_dev->port_data[port];
2168 struct net_device *ndev;
2170 spin_lock_irqsave(&pdata->netdev_lock, flags);
2171 ndev = rcu_dereference_protected(
2172 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2174 spin_lock(&ndev_hash_lock);
2175 hash_del_rcu(&pdata->ndev_hash_link);
2176 spin_unlock(&ndev_hash_lock);
2179 * If this is the last dev_put there is still a
2180 * synchronize_rcu before the netdev is kfreed, so we
2181 * can continue to rely on unlocked pointer
2182 * comparisons after the put
2184 rcu_assign_pointer(pdata->netdev, NULL);
2187 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2191 struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2194 struct ib_port_data *pdata;
2195 struct net_device *res;
2197 if (!rdma_is_port_valid(ib_dev, port))
2200 pdata = &ib_dev->port_data[port];
2203 * New drivers should use ib_device_set_netdev() not the legacy
2206 if (ib_dev->ops.get_netdev)
2207 res = ib_dev->ops.get_netdev(ib_dev, port);
2209 spin_lock(&pdata->netdev_lock);
2210 res = rcu_dereference_protected(
2211 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2214 spin_unlock(&pdata->netdev_lock);
2218 * If we are starting to unregister expedite things by preventing
2219 * propagation of an unregistering netdev.
2221 if (res && res->reg_state != NETREG_REGISTERED) {
2230 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2231 * @ndev: netdev to locate
2232 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2234 * Find and hold an ib_device that is associated with a netdev via
2235 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2238 struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2239 enum rdma_driver_id driver_id)
2241 struct ib_device *res = NULL;
2242 struct ib_port_data *cur;
2245 hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2247 if (rcu_access_pointer(cur->netdev) == ndev &&
2248 (driver_id == RDMA_DRIVER_UNKNOWN ||
2249 cur->ib_dev->ops.driver_id == driver_id) &&
2250 ib_device_try_get(cur->ib_dev)) {
2259 EXPORT_SYMBOL(ib_device_get_by_netdev);
2262 * ib_enum_roce_netdev - enumerate all RoCE ports
2263 * @ib_dev : IB device we want to query
2264 * @filter: Should we call the callback?
2265 * @filter_cookie: Cookie passed to filter
2266 * @cb: Callback to call for each found RoCE ports
2267 * @cookie: Cookie passed back to the callback
2269 * Enumerates all of the physical RoCE ports of ib_dev
2270 * which are related to netdevice and calls callback() on each
2271 * device for which filter() function returns non zero.
2273 void ib_enum_roce_netdev(struct ib_device *ib_dev,
2274 roce_netdev_filter filter,
2275 void *filter_cookie,
2276 roce_netdev_callback cb,
2281 rdma_for_each_port (ib_dev, port)
2282 if (rdma_protocol_roce(ib_dev, port)) {
2283 struct net_device *idev =
2284 ib_device_get_netdev(ib_dev, port);
2286 if (filter(ib_dev, port, idev, filter_cookie))
2287 cb(ib_dev, port, idev, cookie);
2295 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2296 * @filter: Should we call the callback?
2297 * @filter_cookie: Cookie passed to filter
2298 * @cb: Callback to call for each found RoCE ports
2299 * @cookie: Cookie passed back to the callback
2301 * Enumerates all RoCE devices' physical ports which are related
2302 * to netdevices and calls callback() on each device for which
2303 * filter() function returns non zero.
2305 void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2306 void *filter_cookie,
2307 roce_netdev_callback cb,
2310 struct ib_device *dev;
2311 unsigned long index;
2313 down_read(&devices_rwsem);
2314 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2315 ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2316 up_read(&devices_rwsem);
2320 * ib_enum_all_devs - enumerate all ib_devices
2321 * @cb: Callback to call for each found ib_device
2323 * Enumerates all ib_devices and calls callback() on each device.
2325 int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2326 struct netlink_callback *cb)
2328 unsigned long index;
2329 struct ib_device *dev;
2330 unsigned int idx = 0;
2333 down_read(&devices_rwsem);
2334 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2335 if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2338 ret = nldev_cb(dev, skb, cb, idx);
2343 up_read(&devices_rwsem);
2348 * ib_query_pkey - Get P_Key table entry
2349 * @device:Device to query
2350 * @port_num:Port number to query
2351 * @index:P_Key table index to query
2352 * @pkey:Returned P_Key
2354 * ib_query_pkey() fetches the specified P_Key table entry.
2356 int ib_query_pkey(struct ib_device *device,
2357 u8 port_num, u16 index, u16 *pkey)
2359 if (!rdma_is_port_valid(device, port_num))
2362 return device->ops.query_pkey(device, port_num, index, pkey);
2364 EXPORT_SYMBOL(ib_query_pkey);
2367 * ib_modify_device - Change IB device attributes
2368 * @device:Device to modify
2369 * @device_modify_mask:Mask of attributes to change
2370 * @device_modify:New attribute values
2372 * ib_modify_device() changes a device's attributes as specified by
2373 * the @device_modify_mask and @device_modify structure.
2375 int ib_modify_device(struct ib_device *device,
2376 int device_modify_mask,
2377 struct ib_device_modify *device_modify)
2379 if (!device->ops.modify_device)
2382 return device->ops.modify_device(device, device_modify_mask,
2385 EXPORT_SYMBOL(ib_modify_device);
2388 * ib_modify_port - Modifies the attributes for the specified port.
2389 * @device: The device to modify.
2390 * @port_num: The number of the port to modify.
2391 * @port_modify_mask: Mask used to specify which attributes of the port
2393 * @port_modify: New attribute values for the port.
2395 * ib_modify_port() changes a port's attributes as specified by the
2396 * @port_modify_mask and @port_modify structure.
2398 int ib_modify_port(struct ib_device *device,
2399 u8 port_num, int port_modify_mask,
2400 struct ib_port_modify *port_modify)
2404 if (!rdma_is_port_valid(device, port_num))
2407 if (device->ops.modify_port)
2408 rc = device->ops.modify_port(device, port_num,
2411 else if (rdma_protocol_roce(device, port_num) &&
2412 ((port_modify->set_port_cap_mask & ~IB_PORT_CM_SUP) == 0 ||
2413 (port_modify->clr_port_cap_mask & ~IB_PORT_CM_SUP) == 0))
2419 EXPORT_SYMBOL(ib_modify_port);
2422 * ib_find_gid - Returns the port number and GID table index where
2423 * a specified GID value occurs. Its searches only for IB link layer.
2424 * @device: The device to query.
2425 * @gid: The GID value to search for.
2426 * @port_num: The port number of the device where the GID value was found.
2427 * @index: The index into the GID table where the GID was found. This
2428 * parameter may be NULL.
2430 int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2431 u8 *port_num, u16 *index)
2433 union ib_gid tmp_gid;
2437 rdma_for_each_port (device, port) {
2438 if (!rdma_protocol_ib(device, port))
2441 for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2443 ret = rdma_query_gid(device, port, i, &tmp_gid);
2446 if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2457 EXPORT_SYMBOL(ib_find_gid);
2460 * ib_find_pkey - Returns the PKey table index where a specified
2461 * PKey value occurs.
2462 * @device: The device to query.
2463 * @port_num: The port number of the device to search for the PKey.
2464 * @pkey: The PKey value to search for.
2465 * @index: The index into the PKey table where the PKey was found.
2467 int ib_find_pkey(struct ib_device *device,
2468 u8 port_num, u16 pkey, u16 *index)
2472 int partial_ix = -1;
2474 for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2476 ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2479 if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2480 /* if there is full-member pkey take it.*/
2481 if (tmp_pkey & 0x8000) {
2490 /*no full-member, if exists take the limited*/
2491 if (partial_ix >= 0) {
2492 *index = partial_ix;
2497 EXPORT_SYMBOL(ib_find_pkey);
2500 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2501 * for a received CM request
2502 * @dev: An RDMA device on which the request has been received.
2503 * @port: Port number on the RDMA device.
2504 * @pkey: The Pkey the request came on.
2505 * @gid: A GID that the net_dev uses to communicate.
2506 * @addr: Contains the IP address that the request specified as its
2510 struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2513 const union ib_gid *gid,
2514 const struct sockaddr *addr)
2516 struct net_device *net_dev = NULL;
2517 unsigned long index;
2520 if (!rdma_protocol_ib(dev, port))
2524 * Holding the read side guarantees that the client will not become
2525 * unregistered while we are calling get_net_dev_by_params()
2527 down_read(&dev->client_data_rwsem);
2528 xan_for_each_marked (&dev->client_data, index, client_data,
2529 CLIENT_DATA_REGISTERED) {
2530 struct ib_client *client = xa_load(&clients, index);
2532 if (!client || !client->get_net_dev_by_params)
2535 net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2540 up_read(&dev->client_data_rwsem);
2544 EXPORT_SYMBOL(ib_get_net_dev_by_params);
2546 void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2548 struct ib_device_ops *dev_ops = &dev->ops;
2549 #define SET_DEVICE_OP(ptr, name) \
2552 if (!((ptr)->name)) \
2553 (ptr)->name = ops->name; \
2556 #define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2558 if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2559 WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2560 dev_ops->driver_id != ops->driver_id);
2561 dev_ops->driver_id = ops->driver_id;
2564 WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2565 dev_ops->owner = ops->owner;
2567 if (ops->uverbs_abi_ver)
2568 dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2570 dev_ops->uverbs_no_driver_id_binding |=
2571 ops->uverbs_no_driver_id_binding;
2573 SET_DEVICE_OP(dev_ops, add_gid);
2574 SET_DEVICE_OP(dev_ops, advise_mr);
2575 SET_DEVICE_OP(dev_ops, alloc_dm);
2576 SET_DEVICE_OP(dev_ops, alloc_fmr);
2577 SET_DEVICE_OP(dev_ops, alloc_hw_stats);
2578 SET_DEVICE_OP(dev_ops, alloc_mr);
2579 SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2580 SET_DEVICE_OP(dev_ops, alloc_mw);
2581 SET_DEVICE_OP(dev_ops, alloc_pd);
2582 SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2583 SET_DEVICE_OP(dev_ops, alloc_ucontext);
2584 SET_DEVICE_OP(dev_ops, alloc_xrcd);
2585 SET_DEVICE_OP(dev_ops, attach_mcast);
2586 SET_DEVICE_OP(dev_ops, check_mr_status);
2587 SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2588 SET_DEVICE_OP(dev_ops, counter_bind_qp);
2589 SET_DEVICE_OP(dev_ops, counter_dealloc);
2590 SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2591 SET_DEVICE_OP(dev_ops, counter_update_stats);
2592 SET_DEVICE_OP(dev_ops, create_ah);
2593 SET_DEVICE_OP(dev_ops, create_counters);
2594 SET_DEVICE_OP(dev_ops, create_cq);
2595 SET_DEVICE_OP(dev_ops, create_flow);
2596 SET_DEVICE_OP(dev_ops, create_flow_action_esp);
2597 SET_DEVICE_OP(dev_ops, create_qp);
2598 SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2599 SET_DEVICE_OP(dev_ops, create_srq);
2600 SET_DEVICE_OP(dev_ops, create_wq);
2601 SET_DEVICE_OP(dev_ops, dealloc_dm);
2602 SET_DEVICE_OP(dev_ops, dealloc_driver);
2603 SET_DEVICE_OP(dev_ops, dealloc_fmr);
2604 SET_DEVICE_OP(dev_ops, dealloc_mw);
2605 SET_DEVICE_OP(dev_ops, dealloc_pd);
2606 SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2607 SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2608 SET_DEVICE_OP(dev_ops, del_gid);
2609 SET_DEVICE_OP(dev_ops, dereg_mr);
2610 SET_DEVICE_OP(dev_ops, destroy_ah);
2611 SET_DEVICE_OP(dev_ops, destroy_counters);
2612 SET_DEVICE_OP(dev_ops, destroy_cq);
2613 SET_DEVICE_OP(dev_ops, destroy_flow);
2614 SET_DEVICE_OP(dev_ops, destroy_flow_action);
2615 SET_DEVICE_OP(dev_ops, destroy_qp);
2616 SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2617 SET_DEVICE_OP(dev_ops, destroy_srq);
2618 SET_DEVICE_OP(dev_ops, destroy_wq);
2619 SET_DEVICE_OP(dev_ops, detach_mcast);
2620 SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2621 SET_DEVICE_OP(dev_ops, drain_rq);
2622 SET_DEVICE_OP(dev_ops, drain_sq);
2623 SET_DEVICE_OP(dev_ops, enable_driver);
2624 SET_DEVICE_OP(dev_ops, fill_res_entry);
2625 SET_DEVICE_OP(dev_ops, fill_stat_entry);
2626 SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2627 SET_DEVICE_OP(dev_ops, get_dma_mr);
2628 SET_DEVICE_OP(dev_ops, get_hw_stats);
2629 SET_DEVICE_OP(dev_ops, get_link_layer);
2630 SET_DEVICE_OP(dev_ops, get_netdev);
2631 SET_DEVICE_OP(dev_ops, get_port_immutable);
2632 SET_DEVICE_OP(dev_ops, get_vector_affinity);
2633 SET_DEVICE_OP(dev_ops, get_vf_config);
2634 SET_DEVICE_OP(dev_ops, get_vf_guid);
2635 SET_DEVICE_OP(dev_ops, get_vf_stats);
2636 SET_DEVICE_OP(dev_ops, init_port);
2637 SET_DEVICE_OP(dev_ops, invalidate_range);
2638 SET_DEVICE_OP(dev_ops, iw_accept);
2639 SET_DEVICE_OP(dev_ops, iw_add_ref);
2640 SET_DEVICE_OP(dev_ops, iw_connect);
2641 SET_DEVICE_OP(dev_ops, iw_create_listen);
2642 SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2643 SET_DEVICE_OP(dev_ops, iw_get_qp);
2644 SET_DEVICE_OP(dev_ops, iw_reject);
2645 SET_DEVICE_OP(dev_ops, iw_rem_ref);
2646 SET_DEVICE_OP(dev_ops, map_mr_sg);
2647 SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
2648 SET_DEVICE_OP(dev_ops, map_phys_fmr);
2649 SET_DEVICE_OP(dev_ops, mmap);
2650 SET_DEVICE_OP(dev_ops, mmap_free);
2651 SET_DEVICE_OP(dev_ops, modify_ah);
2652 SET_DEVICE_OP(dev_ops, modify_cq);
2653 SET_DEVICE_OP(dev_ops, modify_device);
2654 SET_DEVICE_OP(dev_ops, modify_flow_action_esp);
2655 SET_DEVICE_OP(dev_ops, modify_port);
2656 SET_DEVICE_OP(dev_ops, modify_qp);
2657 SET_DEVICE_OP(dev_ops, modify_srq);
2658 SET_DEVICE_OP(dev_ops, modify_wq);
2659 SET_DEVICE_OP(dev_ops, peek_cq);
2660 SET_DEVICE_OP(dev_ops, poll_cq);
2661 SET_DEVICE_OP(dev_ops, post_recv);
2662 SET_DEVICE_OP(dev_ops, post_send);
2663 SET_DEVICE_OP(dev_ops, post_srq_recv);
2664 SET_DEVICE_OP(dev_ops, process_mad);
2665 SET_DEVICE_OP(dev_ops, query_ah);
2666 SET_DEVICE_OP(dev_ops, query_device);
2667 SET_DEVICE_OP(dev_ops, query_gid);
2668 SET_DEVICE_OP(dev_ops, query_pkey);
2669 SET_DEVICE_OP(dev_ops, query_port);
2670 SET_DEVICE_OP(dev_ops, query_qp);
2671 SET_DEVICE_OP(dev_ops, query_srq);
2672 SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2673 SET_DEVICE_OP(dev_ops, read_counters);
2674 SET_DEVICE_OP(dev_ops, reg_dm_mr);
2675 SET_DEVICE_OP(dev_ops, reg_user_mr);
2676 SET_DEVICE_OP(dev_ops, req_ncomp_notif);
2677 SET_DEVICE_OP(dev_ops, req_notify_cq);
2678 SET_DEVICE_OP(dev_ops, rereg_user_mr);
2679 SET_DEVICE_OP(dev_ops, resize_cq);
2680 SET_DEVICE_OP(dev_ops, set_vf_guid);
2681 SET_DEVICE_OP(dev_ops, set_vf_link_state);
2682 SET_DEVICE_OP(dev_ops, unmap_fmr);
2684 SET_OBJ_SIZE(dev_ops, ib_ah);
2685 SET_OBJ_SIZE(dev_ops, ib_cq);
2686 SET_OBJ_SIZE(dev_ops, ib_pd);
2687 SET_OBJ_SIZE(dev_ops, ib_srq);
2688 SET_OBJ_SIZE(dev_ops, ib_ucontext);
2690 EXPORT_SYMBOL(ib_set_device_ops);
2692 static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2693 [RDMA_NL_LS_OP_RESOLVE] = {
2694 .doit = ib_nl_handle_resolve_resp,
2695 .flags = RDMA_NL_ADMIN_PERM,
2697 [RDMA_NL_LS_OP_SET_TIMEOUT] = {
2698 .doit = ib_nl_handle_set_timeout,
2699 .flags = RDMA_NL_ADMIN_PERM,
2701 [RDMA_NL_LS_OP_IP_RESOLVE] = {
2702 .doit = ib_nl_handle_ip_res_resp,
2703 .flags = RDMA_NL_ADMIN_PERM,
2707 static int __init ib_core_init(void)
2711 ib_wq = alloc_workqueue("infiniband", 0, 0);
2715 ib_comp_wq = alloc_workqueue("ib-comp-wq",
2716 WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2722 ib_comp_unbound_wq =
2723 alloc_workqueue("ib-comp-unb-wq",
2724 WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2725 WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2726 if (!ib_comp_unbound_wq) {
2731 ret = class_register(&ib_class);
2733 pr_warn("Couldn't create InfiniBand device class\n");
2734 goto err_comp_unbound;
2741 pr_warn("Could't init IB address resolution\n");
2745 ret = ib_mad_init();
2747 pr_warn("Couldn't init IB MAD\n");
2753 pr_warn("Couldn't init SA\n");
2757 ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2759 pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2763 ret = register_pernet_device(&rdma_dev_net_ops);
2765 pr_warn("Couldn't init compat dev. ret %d\n", ret);
2770 rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2771 roce_gid_mgmt_init();
2776 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2784 class_unregister(&ib_class);
2786 destroy_workqueue(ib_comp_unbound_wq);
2788 destroy_workqueue(ib_comp_wq);
2790 destroy_workqueue(ib_wq);
2794 static void __exit ib_core_cleanup(void)
2796 roce_gid_mgmt_cleanup();
2798 rdma_nl_unregister(RDMA_NL_LS);
2799 unregister_pernet_device(&rdma_dev_net_ops);
2800 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2805 class_unregister(&ib_class);
2806 destroy_workqueue(ib_comp_unbound_wq);
2807 destroy_workqueue(ib_comp_wq);
2808 /* Make sure that any pending umem accounting work is done. */
2809 destroy_workqueue(ib_wq);
2810 flush_workqueue(system_unbound_wq);
2811 WARN_ON(!xa_empty(&clients));
2812 WARN_ON(!xa_empty(&devices));
2815 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
2817 /* ib core relies on netdev stack to first register net_ns_type_operations
2818 * ns kobject type before ib_core initialization.
2820 fs_initcall(ib_core_init);
2821 module_exit(ib_core_cleanup);