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 <net/netns/generic.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/hashtable.h>
46 #include <rdma/rdma_netlink.h>
47 #include <rdma/ib_addr.h>
48 #include <rdma/ib_cache.h>
49 #include <rdma/rdma_counter.h>
51 #include "core_priv.h"
54 MODULE_AUTHOR("Roland Dreier");
55 MODULE_DESCRIPTION("core kernel InfiniBand API");
56 MODULE_LICENSE("Dual BSD/GPL");
58 struct workqueue_struct *ib_comp_wq;
59 struct workqueue_struct *ib_comp_unbound_wq;
60 struct workqueue_struct *ib_wq;
61 EXPORT_SYMBOL_GPL(ib_wq);
64 * Each of the three rwsem locks (devices, clients, client_data) protects the
65 * xarray of the same name. Specifically it allows the caller to assert that
66 * the MARK will/will not be changing under the lock, and for devices and
67 * clients, that the value in the xarray is still a valid pointer. Change of
68 * the MARK is linked to the object state, so holding the lock and testing the
69 * MARK also asserts that the contained object is in a certain state.
71 * This is used to build a two stage register/unregister flow where objects
72 * can continue to be in the xarray even though they are still in progress to
73 * register/unregister.
75 * The xarray itself provides additional locking, and restartable iteration,
76 * which is also relied on.
78 * Locks should not be nested, with the exception of client_data, which is
79 * allowed to nest under the read side of the other two locks.
81 * The devices_rwsem also protects the device name list, any change or
82 * assignment of device name must also hold the write side to guarantee unique
87 * devices contains devices that have had their names assigned. The
88 * devices may not be registered. Users that care about the registration
89 * status need to call ib_device_try_get() on the device to ensure it is
90 * registered, and keep it registered, for the required duration.
93 static DEFINE_XARRAY_FLAGS(devices, XA_FLAGS_ALLOC);
94 static DECLARE_RWSEM(devices_rwsem);
95 #define DEVICE_REGISTERED XA_MARK_1
97 static u32 highest_client_id;
98 #define CLIENT_REGISTERED XA_MARK_1
99 static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
100 static DECLARE_RWSEM(clients_rwsem);
102 static void ib_client_put(struct ib_client *client)
104 if (refcount_dec_and_test(&client->uses))
105 complete(&client->uses_zero);
109 * If client_data is registered then the corresponding client must also still
112 #define CLIENT_DATA_REGISTERED XA_MARK_1
115 * struct rdma_dev_net - rdma net namespace metadata for a net
116 * @net: Pointer to owner net namespace
117 * @id: xarray id to identify the net namespace.
119 struct rdma_dev_net {
124 static unsigned int rdma_dev_net_id;
127 * A list of net namespaces is maintained in an xarray. This is necessary
128 * because we can't get the locking right using the existing net ns list. We
129 * would require a init_net callback after the list is updated.
131 static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
133 * rwsem to protect accessing the rdma_nets xarray entries.
135 static DECLARE_RWSEM(rdma_nets_rwsem);
137 bool ib_devices_shared_netns = true;
138 module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
139 MODULE_PARM_DESC(netns_mode,
140 "Share device among net namespaces; default=1 (shared)");
142 * rdma_dev_access_netns() - Return whether a rdma device can be accessed
143 * from a specified net namespace or not.
144 * @device: Pointer to rdma device which needs to be checked
145 * @net: Pointer to net namesapce for which access to be checked
147 * rdma_dev_access_netns() - Return whether a rdma device can be accessed
148 * from a specified net namespace or not. When
149 * rdma device is in shared mode, it ignores the
150 * net namespace. When rdma device is exclusive
151 * to a net namespace, rdma device net namespace is
152 * checked against the specified one.
154 bool rdma_dev_access_netns(const struct ib_device *dev, const struct net *net)
156 return (ib_devices_shared_netns ||
157 net_eq(read_pnet(&dev->coredev.rdma_net), net));
159 EXPORT_SYMBOL(rdma_dev_access_netns);
162 * xarray has this behavior where it won't iterate over NULL values stored in
163 * allocated arrays. So we need our own iterator to see all values stored in
164 * the array. This does the same thing as xa_for_each except that it also
165 * returns NULL valued entries if the array is allocating. Simplified to only
166 * work on simple xarrays.
168 static void *xan_find_marked(struct xarray *xa, unsigned long *indexp,
171 XA_STATE(xas, xa, *indexp);
176 entry = xas_find_marked(&xas, ULONG_MAX, filter);
177 if (xa_is_zero(entry))
179 } while (xas_retry(&xas, entry));
183 *indexp = xas.xa_index;
184 if (xa_is_zero(entry))
188 return XA_ERROR(-ENOENT);
190 #define xan_for_each_marked(xa, index, entry, filter) \
191 for (index = 0, entry = xan_find_marked(xa, &(index), filter); \
193 (index)++, entry = xan_find_marked(xa, &(index), filter))
195 /* RCU hash table mapping netdevice pointers to struct ib_port_data */
196 static DEFINE_SPINLOCK(ndev_hash_lock);
197 static DECLARE_HASHTABLE(ndev_hash, 5);
199 static void free_netdevs(struct ib_device *ib_dev);
200 static void ib_unregister_work(struct work_struct *work);
201 static void __ib_unregister_device(struct ib_device *device);
202 static int ib_security_change(struct notifier_block *nb, unsigned long event,
204 static void ib_policy_change_task(struct work_struct *work);
205 static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
207 static void __ibdev_printk(const char *level, const struct ib_device *ibdev,
208 struct va_format *vaf)
210 if (ibdev && ibdev->dev.parent)
211 dev_printk_emit(level[1] - '0',
214 dev_driver_string(ibdev->dev.parent),
215 dev_name(ibdev->dev.parent),
216 dev_name(&ibdev->dev),
220 level, dev_name(&ibdev->dev), vaf);
222 printk("%s(NULL ib_device): %pV", level, vaf);
225 void ibdev_printk(const char *level, const struct ib_device *ibdev,
226 const char *format, ...)
228 struct va_format vaf;
231 va_start(args, format);
236 __ibdev_printk(level, ibdev, &vaf);
240 EXPORT_SYMBOL(ibdev_printk);
242 #define define_ibdev_printk_level(func, level) \
243 void func(const struct ib_device *ibdev, const char *fmt, ...) \
245 struct va_format vaf; \
248 va_start(args, fmt); \
253 __ibdev_printk(level, ibdev, &vaf); \
259 define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
260 define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
261 define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
262 define_ibdev_printk_level(ibdev_err, KERN_ERR);
263 define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
264 define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
265 define_ibdev_printk_level(ibdev_info, KERN_INFO);
267 static struct notifier_block ibdev_lsm_nb = {
268 .notifier_call = ib_security_change,
271 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
274 /* Pointer to the RCU head at the start of the ib_port_data array */
275 struct ib_port_data_rcu {
276 struct rcu_head rcu_head;
277 struct ib_port_data pdata[];
280 static void ib_device_check_mandatory(struct ib_device *device)
282 #define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
283 static const struct {
286 } mandatory_table[] = {
287 IB_MANDATORY_FUNC(query_device),
288 IB_MANDATORY_FUNC(query_port),
289 IB_MANDATORY_FUNC(query_pkey),
290 IB_MANDATORY_FUNC(alloc_pd),
291 IB_MANDATORY_FUNC(dealloc_pd),
292 IB_MANDATORY_FUNC(create_qp),
293 IB_MANDATORY_FUNC(modify_qp),
294 IB_MANDATORY_FUNC(destroy_qp),
295 IB_MANDATORY_FUNC(post_send),
296 IB_MANDATORY_FUNC(post_recv),
297 IB_MANDATORY_FUNC(create_cq),
298 IB_MANDATORY_FUNC(destroy_cq),
299 IB_MANDATORY_FUNC(poll_cq),
300 IB_MANDATORY_FUNC(req_notify_cq),
301 IB_MANDATORY_FUNC(get_dma_mr),
302 IB_MANDATORY_FUNC(dereg_mr),
303 IB_MANDATORY_FUNC(get_port_immutable)
307 device->kverbs_provider = true;
308 for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
309 if (!*(void **) ((void *) &device->ops +
310 mandatory_table[i].offset)) {
311 device->kverbs_provider = false;
318 * Caller must perform ib_device_put() to return the device reference count
319 * when ib_device_get_by_index() returns valid device pointer.
321 struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
323 struct ib_device *device;
325 down_read(&devices_rwsem);
326 device = xa_load(&devices, index);
328 if (!rdma_dev_access_netns(device, net)) {
333 if (!ib_device_try_get(device))
337 up_read(&devices_rwsem);
342 * ib_device_put - Release IB device reference
343 * @device: device whose reference to be released
345 * ib_device_put() releases reference to the IB device to allow it to be
346 * unregistered and eventually free.
348 void ib_device_put(struct ib_device *device)
350 if (refcount_dec_and_test(&device->refcount))
351 complete(&device->unreg_completion);
353 EXPORT_SYMBOL(ib_device_put);
355 static struct ib_device *__ib_device_get_by_name(const char *name)
357 struct ib_device *device;
360 xa_for_each (&devices, index, device)
361 if (!strcmp(name, dev_name(&device->dev)))
368 * ib_device_get_by_name - Find an IB device by name
369 * @name: The name to look for
370 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
372 * Find and hold an ib_device by its name. The caller must call
373 * ib_device_put() on the returned pointer.
375 struct ib_device *ib_device_get_by_name(const char *name,
376 enum rdma_driver_id driver_id)
378 struct ib_device *device;
380 down_read(&devices_rwsem);
381 device = __ib_device_get_by_name(name);
382 if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
383 device->ops.driver_id != driver_id)
387 if (!ib_device_try_get(device))
390 up_read(&devices_rwsem);
393 EXPORT_SYMBOL(ib_device_get_by_name);
395 static int rename_compat_devs(struct ib_device *device)
397 struct ib_core_device *cdev;
401 mutex_lock(&device->compat_devs_mutex);
402 xa_for_each (&device->compat_devs, index, cdev) {
403 ret = device_rename(&cdev->dev, dev_name(&device->dev));
406 "Fail to rename compatdev to new name %s\n",
407 dev_name(&device->dev));
411 mutex_unlock(&device->compat_devs_mutex);
415 int ib_device_rename(struct ib_device *ibdev, const char *name)
421 down_write(&devices_rwsem);
422 if (!strcmp(name, dev_name(&ibdev->dev))) {
423 up_write(&devices_rwsem);
427 if (__ib_device_get_by_name(name)) {
428 up_write(&devices_rwsem);
432 ret = device_rename(&ibdev->dev, name);
434 up_write(&devices_rwsem);
438 strlcpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
439 ret = rename_compat_devs(ibdev);
441 downgrade_write(&devices_rwsem);
442 down_read(&ibdev->client_data_rwsem);
443 xan_for_each_marked(&ibdev->client_data, index, client_data,
444 CLIENT_DATA_REGISTERED) {
445 struct ib_client *client = xa_load(&clients, index);
447 if (!client || !client->rename)
450 client->rename(ibdev, client_data);
452 up_read(&ibdev->client_data_rwsem);
453 up_read(&devices_rwsem);
457 int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
461 ibdev->use_cq_dim = use_dim;
466 static int alloc_name(struct ib_device *ibdev, const char *name)
468 struct ib_device *device;
474 lockdep_assert_held_write(&devices_rwsem);
476 xa_for_each (&devices, index, device) {
477 char buf[IB_DEVICE_NAME_MAX];
479 if (sscanf(dev_name(&device->dev), name, &i) != 1)
481 if (i < 0 || i >= INT_MAX)
483 snprintf(buf, sizeof buf, name, i);
484 if (strcmp(buf, dev_name(&device->dev)) != 0)
487 rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
492 rc = ida_alloc(&inuse, GFP_KERNEL);
496 rc = dev_set_name(&ibdev->dev, name, rc);
502 static void ib_device_release(struct device *device)
504 struct ib_device *dev = container_of(device, struct ib_device, dev);
507 WARN_ON(refcount_read(&dev->refcount));
508 if (dev->port_data) {
509 ib_cache_release_one(dev);
510 ib_security_release_port_pkey_list(dev);
511 rdma_counter_release(dev);
512 kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
517 xa_destroy(&dev->compat_devs);
518 xa_destroy(&dev->client_data);
519 kfree_rcu(dev, rcu_head);
522 static int ib_device_uevent(struct device *device,
523 struct kobj_uevent_env *env)
525 if (add_uevent_var(env, "NAME=%s", dev_name(device)))
529 * It would be nice to pass the node GUID with the event...
535 static const void *net_namespace(struct device *d)
537 struct ib_core_device *coredev =
538 container_of(d, struct ib_core_device, dev);
540 return read_pnet(&coredev->rdma_net);
543 static struct class ib_class = {
544 .name = "infiniband",
545 .dev_release = ib_device_release,
546 .dev_uevent = ib_device_uevent,
547 .ns_type = &net_ns_type_operations,
548 .namespace = net_namespace,
551 static void rdma_init_coredev(struct ib_core_device *coredev,
552 struct ib_device *dev, struct net *net)
554 /* This BUILD_BUG_ON is intended to catch layout change
555 * of union of ib_core_device and device.
556 * dev must be the first element as ib_core and providers
557 * driver uses it. Adding anything in ib_core_device before
558 * device will break this assumption.
560 BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
561 offsetof(struct ib_device, dev));
563 coredev->dev.class = &ib_class;
564 coredev->dev.groups = dev->groups;
565 device_initialize(&coredev->dev);
566 coredev->owner = dev;
567 INIT_LIST_HEAD(&coredev->port_list);
568 write_pnet(&coredev->rdma_net, net);
572 * _ib_alloc_device - allocate an IB device struct
573 * @size:size of structure to allocate
575 * Low-level drivers should use ib_alloc_device() to allocate &struct
576 * ib_device. @size is the size of the structure to be allocated,
577 * including any private data used by the low-level driver.
578 * ib_dealloc_device() must be used to free structures allocated with
581 struct ib_device *_ib_alloc_device(size_t size)
583 struct ib_device *device;
585 if (WARN_ON(size < sizeof(struct ib_device)))
588 device = kzalloc(size, GFP_KERNEL);
592 if (rdma_restrack_init(device)) {
597 device->groups[0] = &ib_dev_attr_group;
598 rdma_init_coredev(&device->coredev, device, &init_net);
600 INIT_LIST_HEAD(&device->event_handler_list);
601 spin_lock_init(&device->event_handler_lock);
602 mutex_init(&device->unregistration_lock);
604 * client_data needs to be alloc because we don't want our mark to be
605 * destroyed if the user stores NULL in the client data.
607 xa_init_flags(&device->client_data, XA_FLAGS_ALLOC);
608 init_rwsem(&device->client_data_rwsem);
609 xa_init_flags(&device->compat_devs, XA_FLAGS_ALLOC);
610 mutex_init(&device->compat_devs_mutex);
611 init_completion(&device->unreg_completion);
612 INIT_WORK(&device->unregistration_work, ib_unregister_work);
616 EXPORT_SYMBOL(_ib_alloc_device);
619 * ib_dealloc_device - free an IB device struct
620 * @device:structure to free
622 * Free a structure allocated with ib_alloc_device().
624 void ib_dealloc_device(struct ib_device *device)
626 if (device->ops.dealloc_driver)
627 device->ops.dealloc_driver(device);
630 * ib_unregister_driver() requires all devices to remain in the xarray
631 * while their ops are callable. The last op we call is dealloc_driver
632 * above. This is needed to create a fence on op callbacks prior to
633 * allowing the driver module to unload.
635 down_write(&devices_rwsem);
636 if (xa_load(&devices, device->index) == device)
637 xa_erase(&devices, device->index);
638 up_write(&devices_rwsem);
640 /* Expedite releasing netdev references */
641 free_netdevs(device);
643 WARN_ON(!xa_empty(&device->compat_devs));
644 WARN_ON(!xa_empty(&device->client_data));
645 WARN_ON(refcount_read(&device->refcount));
646 rdma_restrack_clean(device);
647 /* Balances with device_initialize */
648 put_device(&device->dev);
650 EXPORT_SYMBOL(ib_dealloc_device);
653 * add_client_context() and remove_client_context() must be safe against
654 * parallel calls on the same device - registration/unregistration of both the
655 * device and client can be occurring in parallel.
657 * The routines need to be a fence, any caller must not return until the add
658 * or remove is fully completed.
660 static int add_client_context(struct ib_device *device,
661 struct ib_client *client)
665 if (!device->kverbs_provider && !client->no_kverbs_req)
668 down_write(&device->client_data_rwsem);
670 * So long as the client is registered hold both the client and device
671 * unregistration locks.
673 if (!refcount_inc_not_zero(&client->uses))
675 refcount_inc(&device->refcount);
678 * Another caller to add_client_context got here first and has already
679 * completely initialized context.
681 if (xa_get_mark(&device->client_data, client->client_id,
682 CLIENT_DATA_REGISTERED))
685 ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
689 downgrade_write(&device->client_data_rwsem);
693 /* Readers shall not see a client until add has been completed */
694 xa_set_mark(&device->client_data, client->client_id,
695 CLIENT_DATA_REGISTERED);
696 up_read(&device->client_data_rwsem);
700 ib_device_put(device);
701 ib_client_put(client);
703 up_write(&device->client_data_rwsem);
707 static void remove_client_context(struct ib_device *device,
708 unsigned int client_id)
710 struct ib_client *client;
713 down_write(&device->client_data_rwsem);
714 if (!xa_get_mark(&device->client_data, client_id,
715 CLIENT_DATA_REGISTERED)) {
716 up_write(&device->client_data_rwsem);
719 client_data = xa_load(&device->client_data, client_id);
720 xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
721 client = xa_load(&clients, client_id);
722 up_write(&device->client_data_rwsem);
725 * Notice we cannot be holding any exclusive locks when calling the
726 * remove callback as the remove callback can recurse back into any
727 * public functions in this module and thus try for any locks those
730 * For this reason clients and drivers should not call the
731 * unregistration functions will holdling any locks.
734 client->remove(device, client_data);
736 xa_erase(&device->client_data, client_id);
737 ib_device_put(device);
738 ib_client_put(client);
741 static int alloc_port_data(struct ib_device *device)
743 struct ib_port_data_rcu *pdata_rcu;
746 if (device->port_data)
749 /* This can only be called once the physical port range is defined */
750 if (WARN_ON(!device->phys_port_cnt))
754 * device->port_data is indexed directly by the port number to make
755 * access to this data as efficient as possible.
757 * Therefore port_data is declared as a 1 based array with potential
758 * empty slots at the beginning.
760 pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
761 rdma_end_port(device) + 1),
766 * The rcu_head is put in front of the port data array and the stored
767 * pointer is adjusted since we never need to see that member until
770 device->port_data = pdata_rcu->pdata;
772 rdma_for_each_port (device, port) {
773 struct ib_port_data *pdata = &device->port_data[port];
775 pdata->ib_dev = device;
776 spin_lock_init(&pdata->pkey_list_lock);
777 INIT_LIST_HEAD(&pdata->pkey_list);
778 spin_lock_init(&pdata->netdev_lock);
779 INIT_HLIST_NODE(&pdata->ndev_hash_link);
784 static int verify_immutable(const struct ib_device *dev, u8 port)
786 return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
787 rdma_max_mad_size(dev, port) != 0);
790 static int setup_port_data(struct ib_device *device)
795 ret = alloc_port_data(device);
799 rdma_for_each_port (device, port) {
800 struct ib_port_data *pdata = &device->port_data[port];
802 ret = device->ops.get_port_immutable(device, port,
807 if (verify_immutable(device, port))
813 void ib_get_device_fw_str(struct ib_device *dev, char *str)
815 if (dev->ops.get_dev_fw_str)
816 dev->ops.get_dev_fw_str(dev, str);
820 EXPORT_SYMBOL(ib_get_device_fw_str);
822 static void ib_policy_change_task(struct work_struct *work)
824 struct ib_device *dev;
827 down_read(&devices_rwsem);
828 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
831 rdma_for_each_port (dev, i) {
833 int ret = ib_get_cached_subnet_prefix(dev,
838 "ib_get_cached_subnet_prefix err: %d, this should never happen here\n",
841 ib_security_cache_change(dev, i, sp);
844 up_read(&devices_rwsem);
847 static int ib_security_change(struct notifier_block *nb, unsigned long event,
850 if (event != LSM_POLICY_CHANGE)
853 schedule_work(&ib_policy_change_work);
854 ib_mad_agent_security_change();
859 static void compatdev_release(struct device *dev)
861 struct ib_core_device *cdev =
862 container_of(dev, struct ib_core_device, dev);
867 static int add_one_compat_dev(struct ib_device *device,
868 struct rdma_dev_net *rnet)
870 struct ib_core_device *cdev;
873 lockdep_assert_held(&rdma_nets_rwsem);
874 if (!ib_devices_shared_netns)
878 * Create and add compat device in all namespaces other than where it
879 * is currently bound to.
881 if (net_eq(read_pnet(&rnet->net),
882 read_pnet(&device->coredev.rdma_net)))
886 * The first of init_net() or ib_register_device() to take the
887 * compat_devs_mutex wins and gets to add the device. Others will wait
888 * for completion here.
890 mutex_lock(&device->compat_devs_mutex);
891 cdev = xa_load(&device->compat_devs, rnet->id);
896 ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
900 cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
906 cdev->dev.parent = device->dev.parent;
907 rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
908 cdev->dev.release = compatdev_release;
909 dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
911 ret = device_add(&cdev->dev);
914 ret = ib_setup_port_attrs(cdev);
918 ret = xa_err(xa_store(&device->compat_devs, rnet->id,
923 mutex_unlock(&device->compat_devs_mutex);
927 ib_free_port_attrs(cdev);
929 device_del(&cdev->dev);
931 put_device(&cdev->dev);
933 xa_release(&device->compat_devs, rnet->id);
935 mutex_unlock(&device->compat_devs_mutex);
939 static void remove_one_compat_dev(struct ib_device *device, u32 id)
941 struct ib_core_device *cdev;
943 mutex_lock(&device->compat_devs_mutex);
944 cdev = xa_erase(&device->compat_devs, id);
945 mutex_unlock(&device->compat_devs_mutex);
947 ib_free_port_attrs(cdev);
948 device_del(&cdev->dev);
949 put_device(&cdev->dev);
953 static void remove_compat_devs(struct ib_device *device)
955 struct ib_core_device *cdev;
958 xa_for_each (&device->compat_devs, index, cdev)
959 remove_one_compat_dev(device, index);
962 static int add_compat_devs(struct ib_device *device)
964 struct rdma_dev_net *rnet;
968 lockdep_assert_held(&devices_rwsem);
970 down_read(&rdma_nets_rwsem);
971 xa_for_each (&rdma_nets, index, rnet) {
972 ret = add_one_compat_dev(device, rnet);
976 up_read(&rdma_nets_rwsem);
980 static void remove_all_compat_devs(void)
982 struct ib_compat_device *cdev;
983 struct ib_device *dev;
986 down_read(&devices_rwsem);
987 xa_for_each (&devices, index, dev) {
988 unsigned long c_index = 0;
990 /* Hold nets_rwsem so that any other thread modifying this
991 * system param can sync with this thread.
993 down_read(&rdma_nets_rwsem);
994 xa_for_each (&dev->compat_devs, c_index, cdev)
995 remove_one_compat_dev(dev, c_index);
996 up_read(&rdma_nets_rwsem);
998 up_read(&devices_rwsem);
1001 static int add_all_compat_devs(void)
1003 struct rdma_dev_net *rnet;
1004 struct ib_device *dev;
1005 unsigned long index;
1008 down_read(&devices_rwsem);
1009 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1010 unsigned long net_index = 0;
1012 /* Hold nets_rwsem so that any other thread modifying this
1013 * system param can sync with this thread.
1015 down_read(&rdma_nets_rwsem);
1016 xa_for_each (&rdma_nets, net_index, rnet) {
1017 ret = add_one_compat_dev(dev, rnet);
1021 up_read(&rdma_nets_rwsem);
1023 up_read(&devices_rwsem);
1025 remove_all_compat_devs();
1029 int rdma_compatdev_set(u8 enable)
1031 struct rdma_dev_net *rnet;
1032 unsigned long index;
1035 down_write(&rdma_nets_rwsem);
1036 if (ib_devices_shared_netns == enable) {
1037 up_write(&rdma_nets_rwsem);
1041 /* enable/disable of compat devices is not supported
1042 * when more than default init_net exists.
1044 xa_for_each (&rdma_nets, index, rnet) {
1049 ib_devices_shared_netns = enable;
1050 up_write(&rdma_nets_rwsem);
1055 ret = add_all_compat_devs();
1057 remove_all_compat_devs();
1061 static void rdma_dev_exit_net(struct net *net)
1063 struct rdma_dev_net *rnet = net_generic(net, rdma_dev_net_id);
1064 struct ib_device *dev;
1065 unsigned long index;
1068 down_write(&rdma_nets_rwsem);
1070 * Prevent the ID from being re-used and hide the id from xa_for_each.
1072 ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1074 up_write(&rdma_nets_rwsem);
1076 down_read(&devices_rwsem);
1077 xa_for_each (&devices, index, dev) {
1078 get_device(&dev->dev);
1080 * Release the devices_rwsem so that pontentially blocking
1081 * device_del, doesn't hold the devices_rwsem for too long.
1083 up_read(&devices_rwsem);
1085 remove_one_compat_dev(dev, rnet->id);
1088 * If the real device is in the NS then move it back to init.
1090 rdma_dev_change_netns(dev, net, &init_net);
1092 put_device(&dev->dev);
1093 down_read(&devices_rwsem);
1095 up_read(&devices_rwsem);
1097 xa_erase(&rdma_nets, rnet->id);
1100 static __net_init int rdma_dev_init_net(struct net *net)
1102 struct rdma_dev_net *rnet = net_generic(net, rdma_dev_net_id);
1103 unsigned long index;
1104 struct ib_device *dev;
1107 /* No need to create any compat devices in default init_net. */
1108 if (net_eq(net, &init_net))
1111 write_pnet(&rnet->net, net);
1113 ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1117 down_read(&devices_rwsem);
1118 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1119 /* Hold nets_rwsem so that netlink command cannot change
1120 * system configuration for device sharing mode.
1122 down_read(&rdma_nets_rwsem);
1123 ret = add_one_compat_dev(dev, rnet);
1124 up_read(&rdma_nets_rwsem);
1128 up_read(&devices_rwsem);
1131 rdma_dev_exit_net(net);
1137 * Assign the unique string device name and the unique device index. This is
1138 * undone by ib_dealloc_device.
1140 static int assign_name(struct ib_device *device, const char *name)
1145 down_write(&devices_rwsem);
1146 /* Assign a unique name to the device */
1147 if (strchr(name, '%'))
1148 ret = alloc_name(device, name);
1150 ret = dev_set_name(&device->dev, name);
1154 if (__ib_device_get_by_name(dev_name(&device->dev))) {
1158 strlcpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1160 ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1161 &last_id, GFP_KERNEL);
1166 up_write(&devices_rwsem);
1170 static void setup_dma_device(struct ib_device *device)
1172 struct device *parent = device->dev.parent;
1174 WARN_ON_ONCE(device->dma_device);
1175 if (device->dev.dma_ops) {
1177 * The caller provided custom DMA operations. Copy the
1178 * DMA-related fields that are used by e.g. dma_alloc_coherent()
1181 device->dma_device = &device->dev;
1182 if (!device->dev.dma_mask) {
1184 device->dev.dma_mask = parent->dma_mask;
1188 if (!device->dev.coherent_dma_mask) {
1190 device->dev.coherent_dma_mask =
1191 parent->coherent_dma_mask;
1197 * The caller did not provide custom DMA operations. Use the
1198 * DMA mapping operations of the parent device.
1200 WARN_ON_ONCE(!parent);
1201 device->dma_device = parent;
1203 /* Setup default max segment size for all IB devices */
1204 dma_set_max_seg_size(device->dma_device, SZ_2G);
1209 * setup_device() allocates memory and sets up data that requires calling the
1210 * device ops, this is the only reason these actions are not done during
1211 * ib_alloc_device. It is undone by ib_dealloc_device().
1213 static int setup_device(struct ib_device *device)
1215 struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1218 setup_dma_device(device);
1219 ib_device_check_mandatory(device);
1221 ret = setup_port_data(device);
1223 dev_warn(&device->dev, "Couldn't create per-port data\n");
1227 memset(&device->attrs, 0, sizeof(device->attrs));
1228 ret = device->ops.query_device(device, &device->attrs, &uhw);
1230 dev_warn(&device->dev,
1231 "Couldn't query the device attributes\n");
1238 static void disable_device(struct ib_device *device)
1242 WARN_ON(!refcount_read(&device->refcount));
1244 down_write(&devices_rwsem);
1245 xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1246 up_write(&devices_rwsem);
1249 * Remove clients in LIFO order, see assign_client_id. This could be
1250 * more efficient if xarray learns to reverse iterate. Since no new
1251 * clients can be added to this ib_device past this point we only need
1252 * the maximum possible client_id value here.
1254 down_read(&clients_rwsem);
1255 cid = highest_client_id;
1256 up_read(&clients_rwsem);
1259 remove_client_context(device, cid);
1262 /* Pairs with refcount_set in enable_device */
1263 ib_device_put(device);
1264 wait_for_completion(&device->unreg_completion);
1267 * compat devices must be removed after device refcount drops to zero.
1268 * Otherwise init_net() may add more compatdevs after removing compat
1269 * devices and before device is disabled.
1271 remove_compat_devs(device);
1275 * An enabled device is visible to all clients and to all the public facing
1276 * APIs that return a device pointer. This always returns with a new get, even
1279 static int enable_device_and_get(struct ib_device *device)
1281 struct ib_client *client;
1282 unsigned long index;
1286 * One ref belongs to the xa and the other belongs to this
1287 * thread. This is needed to guard against parallel unregistration.
1289 refcount_set(&device->refcount, 2);
1290 down_write(&devices_rwsem);
1291 xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1294 * By using downgrade_write() we ensure that no other thread can clear
1295 * DEVICE_REGISTERED while we are completing the client setup.
1297 downgrade_write(&devices_rwsem);
1299 if (device->ops.enable_driver) {
1300 ret = device->ops.enable_driver(device);
1305 down_read(&clients_rwsem);
1306 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1307 ret = add_client_context(device, client);
1311 up_read(&clients_rwsem);
1313 ret = add_compat_devs(device);
1315 up_read(&devices_rwsem);
1320 * ib_register_device - Register an IB device with IB core
1321 * @device:Device to register
1323 * Low-level drivers use ib_register_device() to register their
1324 * devices with the IB core. All registered clients will receive a
1325 * callback for each device that is added. @device must be allocated
1326 * with ib_alloc_device().
1328 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1329 * asynchronously then the device pointer may become freed as soon as this
1332 int ib_register_device(struct ib_device *device, const char *name)
1336 ret = assign_name(device, name);
1340 ret = setup_device(device);
1344 ret = ib_cache_setup_one(device);
1346 dev_warn(&device->dev,
1347 "Couldn't set up InfiniBand P_Key/GID cache\n");
1351 ib_device_register_rdmacg(device);
1353 rdma_counter_init(device);
1356 * Ensure that ADD uevent is not fired because it
1357 * is too early amd device is not initialized yet.
1359 dev_set_uevent_suppress(&device->dev, true);
1360 ret = device_add(&device->dev);
1364 ret = ib_device_register_sysfs(device);
1366 dev_warn(&device->dev,
1367 "Couldn't register device with driver model\n");
1371 ret = enable_device_and_get(device);
1372 dev_set_uevent_suppress(&device->dev, false);
1373 /* Mark for userspace that device is ready */
1374 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1376 void (*dealloc_fn)(struct ib_device *);
1379 * If we hit this error flow then we don't want to
1380 * automatically dealloc the device since the caller is
1381 * expected to call ib_dealloc_device() after
1382 * ib_register_device() fails. This is tricky due to the
1383 * possibility for a parallel unregistration along with this
1384 * error flow. Since we have a refcount here we know any
1385 * parallel flow is stopped in disable_device and will see the
1386 * NULL pointers, causing the responsibility to
1387 * ib_dealloc_device() to revert back to this thread.
1389 dealloc_fn = device->ops.dealloc_driver;
1390 device->ops.dealloc_driver = NULL;
1391 ib_device_put(device);
1392 __ib_unregister_device(device);
1393 device->ops.dealloc_driver = dealloc_fn;
1396 ib_device_put(device);
1401 device_del(&device->dev);
1403 dev_set_uevent_suppress(&device->dev, false);
1404 ib_device_unregister_rdmacg(device);
1405 ib_cache_cleanup_one(device);
1408 EXPORT_SYMBOL(ib_register_device);
1410 /* Callers must hold a get on the device. */
1411 static void __ib_unregister_device(struct ib_device *ib_dev)
1414 * We have a registration lock so that all the calls to unregister are
1415 * fully fenced, once any unregister returns the device is truely
1416 * unregistered even if multiple callers are unregistering it at the
1417 * same time. This also interacts with the registration flow and
1418 * provides sane semantics if register and unregister are racing.
1420 mutex_lock(&ib_dev->unregistration_lock);
1421 if (!refcount_read(&ib_dev->refcount))
1424 disable_device(ib_dev);
1426 /* Expedite removing unregistered pointers from the hash table */
1427 free_netdevs(ib_dev);
1429 ib_device_unregister_sysfs(ib_dev);
1430 device_del(&ib_dev->dev);
1431 ib_device_unregister_rdmacg(ib_dev);
1432 ib_cache_cleanup_one(ib_dev);
1435 * Drivers using the new flow may not call ib_dealloc_device except
1436 * in error unwind prior to registration success.
1438 if (ib_dev->ops.dealloc_driver) {
1439 WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1440 ib_dealloc_device(ib_dev);
1443 mutex_unlock(&ib_dev->unregistration_lock);
1447 * ib_unregister_device - Unregister an IB device
1448 * @device: The device to unregister
1450 * Unregister an IB device. All clients will receive a remove callback.
1452 * Callers should call this routine only once, and protect against races with
1453 * registration. Typically it should only be called as part of a remove
1454 * callback in an implementation of driver core's struct device_driver and
1457 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1460 void ib_unregister_device(struct ib_device *ib_dev)
1462 get_device(&ib_dev->dev);
1463 __ib_unregister_device(ib_dev);
1464 put_device(&ib_dev->dev);
1466 EXPORT_SYMBOL(ib_unregister_device);
1469 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1470 * device: The device to unregister
1472 * This is the same as ib_unregister_device(), except it includes an internal
1473 * ib_device_put() that should match a 'get' obtained by the caller.
1475 * It is safe to call this routine concurrently from multiple threads while
1476 * holding the 'get'. When the function returns the device is fully
1479 * Drivers using this flow MUST use the driver_unregister callback to clean up
1480 * their resources associated with the device and dealloc it.
1482 void ib_unregister_device_and_put(struct ib_device *ib_dev)
1484 WARN_ON(!ib_dev->ops.dealloc_driver);
1485 get_device(&ib_dev->dev);
1486 ib_device_put(ib_dev);
1487 __ib_unregister_device(ib_dev);
1488 put_device(&ib_dev->dev);
1490 EXPORT_SYMBOL(ib_unregister_device_and_put);
1493 * ib_unregister_driver - Unregister all IB devices for a driver
1494 * @driver_id: The driver to unregister
1496 * This implements a fence for device unregistration. It only returns once all
1497 * devices associated with the driver_id have fully completed their
1498 * unregistration and returned from ib_unregister_device*().
1500 * If device's are not yet unregistered it goes ahead and starts unregistering
1503 * This does not block creation of new devices with the given driver_id, that
1504 * is the responsibility of the caller.
1506 void ib_unregister_driver(enum rdma_driver_id driver_id)
1508 struct ib_device *ib_dev;
1509 unsigned long index;
1511 down_read(&devices_rwsem);
1512 xa_for_each (&devices, index, ib_dev) {
1513 if (ib_dev->ops.driver_id != driver_id)
1516 get_device(&ib_dev->dev);
1517 up_read(&devices_rwsem);
1519 WARN_ON(!ib_dev->ops.dealloc_driver);
1520 __ib_unregister_device(ib_dev);
1522 put_device(&ib_dev->dev);
1523 down_read(&devices_rwsem);
1525 up_read(&devices_rwsem);
1527 EXPORT_SYMBOL(ib_unregister_driver);
1529 static void ib_unregister_work(struct work_struct *work)
1531 struct ib_device *ib_dev =
1532 container_of(work, struct ib_device, unregistration_work);
1534 __ib_unregister_device(ib_dev);
1535 put_device(&ib_dev->dev);
1539 * ib_unregister_device_queued - Unregister a device using a work queue
1540 * device: The device to unregister
1542 * This schedules an asynchronous unregistration using a WQ for the device. A
1543 * driver should use this to avoid holding locks while doing unregistration,
1544 * such as holding the RTNL lock.
1546 * Drivers using this API must use ib_unregister_driver before module unload
1547 * to ensure that all scheduled unregistrations have completed.
1549 void ib_unregister_device_queued(struct ib_device *ib_dev)
1551 WARN_ON(!refcount_read(&ib_dev->refcount));
1552 WARN_ON(!ib_dev->ops.dealloc_driver);
1553 get_device(&ib_dev->dev);
1554 if (!queue_work(system_unbound_wq, &ib_dev->unregistration_work))
1555 put_device(&ib_dev->dev);
1557 EXPORT_SYMBOL(ib_unregister_device_queued);
1560 * The caller must pass in a device that has the kref held and the refcount
1561 * released. If the device is in cur_net and still registered then it is moved
1564 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1570 mutex_lock(&device->unregistration_lock);
1573 * If a device not under ib_device_get() or if the unregistration_lock
1574 * is not held, the namespace can be changed, or it can be unregistered.
1575 * Check again under the lock.
1577 if (refcount_read(&device->refcount) == 0 ||
1578 !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1583 kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1584 disable_device(device);
1587 * At this point no one can be using the device, so it is safe to
1588 * change the namespace.
1590 write_pnet(&device->coredev.rdma_net, net);
1592 down_read(&devices_rwsem);
1594 * Currently rdma devices are system wide unique. So the device name
1595 * is guaranteed free in the new namespace. Publish the new namespace
1596 * at the sysfs level.
1598 ret = device_rename(&device->dev, dev_name(&device->dev));
1599 up_read(&devices_rwsem);
1601 dev_warn(&device->dev,
1602 "%s: Couldn't rename device after namespace change\n",
1604 /* Try and put things back and re-enable the device */
1605 write_pnet(&device->coredev.rdma_net, cur_net);
1608 ret2 = enable_device_and_get(device);
1611 * This shouldn't really happen, but if it does, let the user
1612 * retry at later point. So don't disable the device.
1614 dev_warn(&device->dev,
1615 "%s: Couldn't re-enable device after namespace change\n",
1618 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1620 ib_device_put(device);
1622 mutex_unlock(&device->unregistration_lock);
1628 int ib_device_set_netns_put(struct sk_buff *skb,
1629 struct ib_device *dev, u32 ns_fd)
1634 net = get_net_ns_by_fd(ns_fd);
1640 if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1646 * Currently supported only for those providers which support
1647 * disassociation and don't do port specific sysfs init. Once a
1648 * port_cleanup infrastructure is implemented, this limitation will be
1651 if (!dev->ops.disassociate_ucontext || dev->ops.init_port ||
1652 ib_devices_shared_netns) {
1657 get_device(&dev->dev);
1659 ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1660 put_device(&dev->dev);
1672 static struct pernet_operations rdma_dev_net_ops = {
1673 .init = rdma_dev_init_net,
1674 .exit = rdma_dev_exit_net,
1675 .id = &rdma_dev_net_id,
1676 .size = sizeof(struct rdma_dev_net),
1679 static int assign_client_id(struct ib_client *client)
1683 down_write(&clients_rwsem);
1685 * The add/remove callbacks must be called in FIFO/LIFO order. To
1686 * achieve this we assign client_ids so they are sorted in
1687 * registration order.
1689 client->client_id = highest_client_id;
1690 ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1694 highest_client_id++;
1695 xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1698 up_write(&clients_rwsem);
1702 static void remove_client_id(struct ib_client *client)
1704 down_write(&clients_rwsem);
1705 xa_erase(&clients, client->client_id);
1706 for (; highest_client_id; highest_client_id--)
1707 if (xa_load(&clients, highest_client_id - 1))
1709 up_write(&clients_rwsem);
1713 * ib_register_client - Register an IB client
1714 * @client:Client to register
1716 * Upper level users of the IB drivers can use ib_register_client() to
1717 * register callbacks for IB device addition and removal. When an IB
1718 * device is added, each registered client's add method will be called
1719 * (in the order the clients were registered), and when a device is
1720 * removed, each client's remove method will be called (in the reverse
1721 * order that clients were registered). In addition, when
1722 * ib_register_client() is called, the client will receive an add
1723 * callback for all devices already registered.
1725 int ib_register_client(struct ib_client *client)
1727 struct ib_device *device;
1728 unsigned long index;
1731 refcount_set(&client->uses, 1);
1732 init_completion(&client->uses_zero);
1733 ret = assign_client_id(client);
1737 down_read(&devices_rwsem);
1738 xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1739 ret = add_client_context(device, client);
1741 up_read(&devices_rwsem);
1742 ib_unregister_client(client);
1746 up_read(&devices_rwsem);
1749 EXPORT_SYMBOL(ib_register_client);
1752 * ib_unregister_client - Unregister an IB client
1753 * @client:Client to unregister
1755 * Upper level users use ib_unregister_client() to remove their client
1756 * registration. When ib_unregister_client() is called, the client
1757 * will receive a remove callback for each IB device still registered.
1759 * This is a full fence, once it returns no client callbacks will be called,
1760 * or are running in another thread.
1762 void ib_unregister_client(struct ib_client *client)
1764 struct ib_device *device;
1765 unsigned long index;
1767 down_write(&clients_rwsem);
1768 ib_client_put(client);
1769 xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1770 up_write(&clients_rwsem);
1772 /* We do not want to have locks while calling client->remove() */
1774 xa_for_each (&devices, index, device) {
1775 if (!ib_device_try_get(device))
1779 remove_client_context(device, client->client_id);
1781 ib_device_put(device);
1787 * remove_client_context() is not a fence, it can return even though a
1788 * removal is ongoing. Wait until all removals are completed.
1790 wait_for_completion(&client->uses_zero);
1791 remove_client_id(client);
1793 EXPORT_SYMBOL(ib_unregister_client);
1795 static int __ib_get_global_client_nl_info(const char *client_name,
1796 struct ib_client_nl_info *res)
1798 struct ib_client *client;
1799 unsigned long index;
1802 down_read(&clients_rwsem);
1803 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1804 if (strcmp(client->name, client_name) != 0)
1806 if (!client->get_global_nl_info) {
1810 ret = client->get_global_nl_info(res);
1811 if (WARN_ON(ret == -ENOENT))
1813 if (!ret && res->cdev)
1814 get_device(res->cdev);
1817 up_read(&clients_rwsem);
1821 static int __ib_get_client_nl_info(struct ib_device *ibdev,
1822 const char *client_name,
1823 struct ib_client_nl_info *res)
1825 unsigned long index;
1829 down_read(&ibdev->client_data_rwsem);
1830 xan_for_each_marked (&ibdev->client_data, index, client_data,
1831 CLIENT_DATA_REGISTERED) {
1832 struct ib_client *client = xa_load(&clients, index);
1834 if (!client || strcmp(client->name, client_name) != 0)
1836 if (!client->get_nl_info) {
1840 ret = client->get_nl_info(ibdev, client_data, res);
1841 if (WARN_ON(ret == -ENOENT))
1845 * The cdev is guaranteed valid as long as we are inside the
1846 * client_data_rwsem as remove_one can't be called. Keep it
1847 * valid for the caller.
1849 if (!ret && res->cdev)
1850 get_device(res->cdev);
1853 up_read(&ibdev->client_data_rwsem);
1859 * ib_get_client_nl_info - Fetch the nl_info from a client
1860 * @device - IB device
1861 * @client_name - Name of the client
1862 * @res - Result of the query
1864 int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1865 struct ib_client_nl_info *res)
1870 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1872 ret = __ib_get_global_client_nl_info(client_name, res);
1873 #ifdef CONFIG_MODULES
1874 if (ret == -ENOENT) {
1875 request_module("rdma-client-%s", client_name);
1877 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1879 ret = __ib_get_global_client_nl_info(client_name, res);
1888 if (WARN_ON(!res->cdev))
1894 * ib_set_client_data - Set IB client context
1895 * @device:Device to set context for
1896 * @client:Client to set context for
1897 * @data:Context to set
1899 * ib_set_client_data() sets client context data that can be retrieved with
1900 * ib_get_client_data(). This can only be called while the client is
1901 * registered to the device, once the ib_client remove() callback returns this
1904 void ib_set_client_data(struct ib_device *device, struct ib_client *client,
1909 if (WARN_ON(IS_ERR(data)))
1912 rc = xa_store(&device->client_data, client->client_id, data,
1914 WARN_ON(xa_is_err(rc));
1916 EXPORT_SYMBOL(ib_set_client_data);
1919 * ib_register_event_handler - Register an IB event handler
1920 * @event_handler:Handler to register
1922 * ib_register_event_handler() registers an event handler that will be
1923 * called back when asynchronous IB events occur (as defined in
1924 * chapter 11 of the InfiniBand Architecture Specification). This
1925 * callback may occur in interrupt context.
1927 void ib_register_event_handler(struct ib_event_handler *event_handler)
1929 unsigned long flags;
1931 spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
1932 list_add_tail(&event_handler->list,
1933 &event_handler->device->event_handler_list);
1934 spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
1936 EXPORT_SYMBOL(ib_register_event_handler);
1939 * ib_unregister_event_handler - Unregister an event handler
1940 * @event_handler:Handler to unregister
1942 * Unregister an event handler registered with
1943 * ib_register_event_handler().
1945 void ib_unregister_event_handler(struct ib_event_handler *event_handler)
1947 unsigned long flags;
1949 spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
1950 list_del(&event_handler->list);
1951 spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
1953 EXPORT_SYMBOL(ib_unregister_event_handler);
1956 * ib_dispatch_event - Dispatch an asynchronous event
1957 * @event:Event to dispatch
1959 * Low-level drivers must call ib_dispatch_event() to dispatch the
1960 * event to all registered event handlers when an asynchronous event
1963 void ib_dispatch_event(struct ib_event *event)
1965 unsigned long flags;
1966 struct ib_event_handler *handler;
1968 spin_lock_irqsave(&event->device->event_handler_lock, flags);
1970 list_for_each_entry(handler, &event->device->event_handler_list, list)
1971 handler->handler(handler, event);
1973 spin_unlock_irqrestore(&event->device->event_handler_lock, flags);
1975 EXPORT_SYMBOL(ib_dispatch_event);
1978 * ib_query_port - Query IB port attributes
1979 * @device:Device to query
1980 * @port_num:Port number to query
1981 * @port_attr:Port attributes
1983 * ib_query_port() returns the attributes of a port through the
1984 * @port_attr pointer.
1986 int ib_query_port(struct ib_device *device,
1988 struct ib_port_attr *port_attr)
1993 if (!rdma_is_port_valid(device, port_num))
1996 memset(port_attr, 0, sizeof(*port_attr));
1997 err = device->ops.query_port(device, port_num, port_attr);
1998 if (err || port_attr->subnet_prefix)
2001 if (rdma_port_get_link_layer(device, port_num) != IB_LINK_LAYER_INFINIBAND)
2004 err = device->ops.query_gid(device, port_num, 0, &gid);
2008 port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
2011 EXPORT_SYMBOL(ib_query_port);
2013 static void add_ndev_hash(struct ib_port_data *pdata)
2015 unsigned long flags;
2019 spin_lock_irqsave(&ndev_hash_lock, flags);
2020 if (hash_hashed(&pdata->ndev_hash_link)) {
2021 hash_del_rcu(&pdata->ndev_hash_link);
2022 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2024 * We cannot do hash_add_rcu after a hash_del_rcu until the
2028 spin_lock_irqsave(&ndev_hash_lock, flags);
2031 hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
2032 (uintptr_t)pdata->netdev);
2033 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2037 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2038 * @ib_dev: Device to modify
2039 * @ndev: net_device to affiliate, may be NULL
2040 * @port: IB port the net_device is connected to
2042 * Drivers should use this to link the ib_device to a netdev so the netdev
2043 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2044 * affiliated with any port.
2046 * The caller must ensure that the given ndev is not unregistered or
2047 * unregistering, and that either the ib_device is unregistered or
2048 * ib_device_set_netdev() is called with NULL when the ndev sends a
2049 * NETDEV_UNREGISTER event.
2051 int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2054 struct net_device *old_ndev;
2055 struct ib_port_data *pdata;
2056 unsigned long flags;
2060 * Drivers wish to call this before ib_register_driver, so we have to
2061 * setup the port data early.
2063 ret = alloc_port_data(ib_dev);
2067 if (!rdma_is_port_valid(ib_dev, port))
2070 pdata = &ib_dev->port_data[port];
2071 spin_lock_irqsave(&pdata->netdev_lock, flags);
2072 old_ndev = rcu_dereference_protected(
2073 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2074 if (old_ndev == ndev) {
2075 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2081 rcu_assign_pointer(pdata->netdev, ndev);
2082 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2084 add_ndev_hash(pdata);
2090 EXPORT_SYMBOL(ib_device_set_netdev);
2092 static void free_netdevs(struct ib_device *ib_dev)
2094 unsigned long flags;
2097 if (!ib_dev->port_data)
2100 rdma_for_each_port (ib_dev, port) {
2101 struct ib_port_data *pdata = &ib_dev->port_data[port];
2102 struct net_device *ndev;
2104 spin_lock_irqsave(&pdata->netdev_lock, flags);
2105 ndev = rcu_dereference_protected(
2106 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2108 spin_lock(&ndev_hash_lock);
2109 hash_del_rcu(&pdata->ndev_hash_link);
2110 spin_unlock(&ndev_hash_lock);
2113 * If this is the last dev_put there is still a
2114 * synchronize_rcu before the netdev is kfreed, so we
2115 * can continue to rely on unlocked pointer
2116 * comparisons after the put
2118 rcu_assign_pointer(pdata->netdev, NULL);
2121 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2125 struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2128 struct ib_port_data *pdata;
2129 struct net_device *res;
2131 if (!rdma_is_port_valid(ib_dev, port))
2134 pdata = &ib_dev->port_data[port];
2137 * New drivers should use ib_device_set_netdev() not the legacy
2140 if (ib_dev->ops.get_netdev)
2141 res = ib_dev->ops.get_netdev(ib_dev, port);
2143 spin_lock(&pdata->netdev_lock);
2144 res = rcu_dereference_protected(
2145 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2148 spin_unlock(&pdata->netdev_lock);
2152 * If we are starting to unregister expedite things by preventing
2153 * propagation of an unregistering netdev.
2155 if (res && res->reg_state != NETREG_REGISTERED) {
2164 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2165 * @ndev: netdev to locate
2166 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2168 * Find and hold an ib_device that is associated with a netdev via
2169 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2172 struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2173 enum rdma_driver_id driver_id)
2175 struct ib_device *res = NULL;
2176 struct ib_port_data *cur;
2179 hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2181 if (rcu_access_pointer(cur->netdev) == ndev &&
2182 (driver_id == RDMA_DRIVER_UNKNOWN ||
2183 cur->ib_dev->ops.driver_id == driver_id) &&
2184 ib_device_try_get(cur->ib_dev)) {
2193 EXPORT_SYMBOL(ib_device_get_by_netdev);
2196 * ib_enum_roce_netdev - enumerate all RoCE ports
2197 * @ib_dev : IB device we want to query
2198 * @filter: Should we call the callback?
2199 * @filter_cookie: Cookie passed to filter
2200 * @cb: Callback to call for each found RoCE ports
2201 * @cookie: Cookie passed back to the callback
2203 * Enumerates all of the physical RoCE ports of ib_dev
2204 * which are related to netdevice and calls callback() on each
2205 * device for which filter() function returns non zero.
2207 void ib_enum_roce_netdev(struct ib_device *ib_dev,
2208 roce_netdev_filter filter,
2209 void *filter_cookie,
2210 roce_netdev_callback cb,
2215 rdma_for_each_port (ib_dev, port)
2216 if (rdma_protocol_roce(ib_dev, port)) {
2217 struct net_device *idev =
2218 ib_device_get_netdev(ib_dev, port);
2220 if (filter(ib_dev, port, idev, filter_cookie))
2221 cb(ib_dev, port, idev, cookie);
2229 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2230 * @filter: Should we call the callback?
2231 * @filter_cookie: Cookie passed to filter
2232 * @cb: Callback to call for each found RoCE ports
2233 * @cookie: Cookie passed back to the callback
2235 * Enumerates all RoCE devices' physical ports which are related
2236 * to netdevices and calls callback() on each device for which
2237 * filter() function returns non zero.
2239 void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2240 void *filter_cookie,
2241 roce_netdev_callback cb,
2244 struct ib_device *dev;
2245 unsigned long index;
2247 down_read(&devices_rwsem);
2248 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2249 ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2250 up_read(&devices_rwsem);
2254 * ib_enum_all_devs - enumerate all ib_devices
2255 * @cb: Callback to call for each found ib_device
2257 * Enumerates all ib_devices and calls callback() on each device.
2259 int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2260 struct netlink_callback *cb)
2262 unsigned long index;
2263 struct ib_device *dev;
2264 unsigned int idx = 0;
2267 down_read(&devices_rwsem);
2268 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2269 if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2272 ret = nldev_cb(dev, skb, cb, idx);
2277 up_read(&devices_rwsem);
2282 * ib_query_pkey - Get P_Key table entry
2283 * @device:Device to query
2284 * @port_num:Port number to query
2285 * @index:P_Key table index to query
2286 * @pkey:Returned P_Key
2288 * ib_query_pkey() fetches the specified P_Key table entry.
2290 int ib_query_pkey(struct ib_device *device,
2291 u8 port_num, u16 index, u16 *pkey)
2293 if (!rdma_is_port_valid(device, port_num))
2296 return device->ops.query_pkey(device, port_num, index, pkey);
2298 EXPORT_SYMBOL(ib_query_pkey);
2301 * ib_modify_device - Change IB device attributes
2302 * @device:Device to modify
2303 * @device_modify_mask:Mask of attributes to change
2304 * @device_modify:New attribute values
2306 * ib_modify_device() changes a device's attributes as specified by
2307 * the @device_modify_mask and @device_modify structure.
2309 int ib_modify_device(struct ib_device *device,
2310 int device_modify_mask,
2311 struct ib_device_modify *device_modify)
2313 if (!device->ops.modify_device)
2316 return device->ops.modify_device(device, device_modify_mask,
2319 EXPORT_SYMBOL(ib_modify_device);
2322 * ib_modify_port - Modifies the attributes for the specified port.
2323 * @device: The device to modify.
2324 * @port_num: The number of the port to modify.
2325 * @port_modify_mask: Mask used to specify which attributes of the port
2327 * @port_modify: New attribute values for the port.
2329 * ib_modify_port() changes a port's attributes as specified by the
2330 * @port_modify_mask and @port_modify structure.
2332 int ib_modify_port(struct ib_device *device,
2333 u8 port_num, int port_modify_mask,
2334 struct ib_port_modify *port_modify)
2338 if (!rdma_is_port_valid(device, port_num))
2341 if (device->ops.modify_port)
2342 rc = device->ops.modify_port(device, port_num,
2346 rc = rdma_protocol_roce(device, port_num) ? 0 : -ENOSYS;
2349 EXPORT_SYMBOL(ib_modify_port);
2352 * ib_find_gid - Returns the port number and GID table index where
2353 * a specified GID value occurs. Its searches only for IB link layer.
2354 * @device: The device to query.
2355 * @gid: The GID value to search for.
2356 * @port_num: The port number of the device where the GID value was found.
2357 * @index: The index into the GID table where the GID was found. This
2358 * parameter may be NULL.
2360 int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2361 u8 *port_num, u16 *index)
2363 union ib_gid tmp_gid;
2367 rdma_for_each_port (device, port) {
2368 if (!rdma_protocol_ib(device, port))
2371 for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2373 ret = rdma_query_gid(device, port, i, &tmp_gid);
2376 if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2387 EXPORT_SYMBOL(ib_find_gid);
2390 * ib_find_pkey - Returns the PKey table index where a specified
2391 * PKey value occurs.
2392 * @device: The device to query.
2393 * @port_num: The port number of the device to search for the PKey.
2394 * @pkey: The PKey value to search for.
2395 * @index: The index into the PKey table where the PKey was found.
2397 int ib_find_pkey(struct ib_device *device,
2398 u8 port_num, u16 pkey, u16 *index)
2402 int partial_ix = -1;
2404 for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2406 ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2409 if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2410 /* if there is full-member pkey take it.*/
2411 if (tmp_pkey & 0x8000) {
2420 /*no full-member, if exists take the limited*/
2421 if (partial_ix >= 0) {
2422 *index = partial_ix;
2427 EXPORT_SYMBOL(ib_find_pkey);
2430 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2431 * for a received CM request
2432 * @dev: An RDMA device on which the request has been received.
2433 * @port: Port number on the RDMA device.
2434 * @pkey: The Pkey the request came on.
2435 * @gid: A GID that the net_dev uses to communicate.
2436 * @addr: Contains the IP address that the request specified as its
2440 struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2443 const union ib_gid *gid,
2444 const struct sockaddr *addr)
2446 struct net_device *net_dev = NULL;
2447 unsigned long index;
2450 if (!rdma_protocol_ib(dev, port))
2454 * Holding the read side guarantees that the client will not become
2455 * unregistered while we are calling get_net_dev_by_params()
2457 down_read(&dev->client_data_rwsem);
2458 xan_for_each_marked (&dev->client_data, index, client_data,
2459 CLIENT_DATA_REGISTERED) {
2460 struct ib_client *client = xa_load(&clients, index);
2462 if (!client || !client->get_net_dev_by_params)
2465 net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2470 up_read(&dev->client_data_rwsem);
2474 EXPORT_SYMBOL(ib_get_net_dev_by_params);
2476 void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2478 struct ib_device_ops *dev_ops = &dev->ops;
2479 #define SET_DEVICE_OP(ptr, name) \
2482 if (!((ptr)->name)) \
2483 (ptr)->name = ops->name; \
2486 #define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2488 if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2489 WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2490 dev_ops->driver_id != ops->driver_id);
2491 dev_ops->driver_id = ops->driver_id;
2494 WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2495 dev_ops->owner = ops->owner;
2497 if (ops->uverbs_abi_ver)
2498 dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2500 dev_ops->uverbs_no_driver_id_binding |=
2501 ops->uverbs_no_driver_id_binding;
2503 SET_DEVICE_OP(dev_ops, add_gid);
2504 SET_DEVICE_OP(dev_ops, advise_mr);
2505 SET_DEVICE_OP(dev_ops, alloc_dm);
2506 SET_DEVICE_OP(dev_ops, alloc_fmr);
2507 SET_DEVICE_OP(dev_ops, alloc_hw_stats);
2508 SET_DEVICE_OP(dev_ops, alloc_mr);
2509 SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2510 SET_DEVICE_OP(dev_ops, alloc_mw);
2511 SET_DEVICE_OP(dev_ops, alloc_pd);
2512 SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2513 SET_DEVICE_OP(dev_ops, alloc_ucontext);
2514 SET_DEVICE_OP(dev_ops, alloc_xrcd);
2515 SET_DEVICE_OP(dev_ops, attach_mcast);
2516 SET_DEVICE_OP(dev_ops, check_mr_status);
2517 SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2518 SET_DEVICE_OP(dev_ops, counter_bind_qp);
2519 SET_DEVICE_OP(dev_ops, counter_dealloc);
2520 SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2521 SET_DEVICE_OP(dev_ops, counter_update_stats);
2522 SET_DEVICE_OP(dev_ops, create_ah);
2523 SET_DEVICE_OP(dev_ops, create_counters);
2524 SET_DEVICE_OP(dev_ops, create_cq);
2525 SET_DEVICE_OP(dev_ops, create_flow);
2526 SET_DEVICE_OP(dev_ops, create_flow_action_esp);
2527 SET_DEVICE_OP(dev_ops, create_qp);
2528 SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2529 SET_DEVICE_OP(dev_ops, create_srq);
2530 SET_DEVICE_OP(dev_ops, create_wq);
2531 SET_DEVICE_OP(dev_ops, dealloc_dm);
2532 SET_DEVICE_OP(dev_ops, dealloc_driver);
2533 SET_DEVICE_OP(dev_ops, dealloc_fmr);
2534 SET_DEVICE_OP(dev_ops, dealloc_mw);
2535 SET_DEVICE_OP(dev_ops, dealloc_pd);
2536 SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2537 SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2538 SET_DEVICE_OP(dev_ops, del_gid);
2539 SET_DEVICE_OP(dev_ops, dereg_mr);
2540 SET_DEVICE_OP(dev_ops, destroy_ah);
2541 SET_DEVICE_OP(dev_ops, destroy_counters);
2542 SET_DEVICE_OP(dev_ops, destroy_cq);
2543 SET_DEVICE_OP(dev_ops, destroy_flow);
2544 SET_DEVICE_OP(dev_ops, destroy_flow_action);
2545 SET_DEVICE_OP(dev_ops, destroy_qp);
2546 SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2547 SET_DEVICE_OP(dev_ops, destroy_srq);
2548 SET_DEVICE_OP(dev_ops, destroy_wq);
2549 SET_DEVICE_OP(dev_ops, detach_mcast);
2550 SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2551 SET_DEVICE_OP(dev_ops, drain_rq);
2552 SET_DEVICE_OP(dev_ops, drain_sq);
2553 SET_DEVICE_OP(dev_ops, enable_driver);
2554 SET_DEVICE_OP(dev_ops, fill_res_entry);
2555 SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2556 SET_DEVICE_OP(dev_ops, get_dma_mr);
2557 SET_DEVICE_OP(dev_ops, get_hw_stats);
2558 SET_DEVICE_OP(dev_ops, get_link_layer);
2559 SET_DEVICE_OP(dev_ops, get_netdev);
2560 SET_DEVICE_OP(dev_ops, get_port_immutable);
2561 SET_DEVICE_OP(dev_ops, get_vector_affinity);
2562 SET_DEVICE_OP(dev_ops, get_vf_config);
2563 SET_DEVICE_OP(dev_ops, get_vf_stats);
2564 SET_DEVICE_OP(dev_ops, init_port);
2565 SET_DEVICE_OP(dev_ops, iw_accept);
2566 SET_DEVICE_OP(dev_ops, iw_add_ref);
2567 SET_DEVICE_OP(dev_ops, iw_connect);
2568 SET_DEVICE_OP(dev_ops, iw_create_listen);
2569 SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2570 SET_DEVICE_OP(dev_ops, iw_get_qp);
2571 SET_DEVICE_OP(dev_ops, iw_reject);
2572 SET_DEVICE_OP(dev_ops, iw_rem_ref);
2573 SET_DEVICE_OP(dev_ops, map_mr_sg);
2574 SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
2575 SET_DEVICE_OP(dev_ops, map_phys_fmr);
2576 SET_DEVICE_OP(dev_ops, mmap);
2577 SET_DEVICE_OP(dev_ops, modify_ah);
2578 SET_DEVICE_OP(dev_ops, modify_cq);
2579 SET_DEVICE_OP(dev_ops, modify_device);
2580 SET_DEVICE_OP(dev_ops, modify_flow_action_esp);
2581 SET_DEVICE_OP(dev_ops, modify_port);
2582 SET_DEVICE_OP(dev_ops, modify_qp);
2583 SET_DEVICE_OP(dev_ops, modify_srq);
2584 SET_DEVICE_OP(dev_ops, modify_wq);
2585 SET_DEVICE_OP(dev_ops, peek_cq);
2586 SET_DEVICE_OP(dev_ops, poll_cq);
2587 SET_DEVICE_OP(dev_ops, post_recv);
2588 SET_DEVICE_OP(dev_ops, post_send);
2589 SET_DEVICE_OP(dev_ops, post_srq_recv);
2590 SET_DEVICE_OP(dev_ops, process_mad);
2591 SET_DEVICE_OP(dev_ops, query_ah);
2592 SET_DEVICE_OP(dev_ops, query_device);
2593 SET_DEVICE_OP(dev_ops, query_gid);
2594 SET_DEVICE_OP(dev_ops, query_pkey);
2595 SET_DEVICE_OP(dev_ops, query_port);
2596 SET_DEVICE_OP(dev_ops, query_qp);
2597 SET_DEVICE_OP(dev_ops, query_srq);
2598 SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2599 SET_DEVICE_OP(dev_ops, read_counters);
2600 SET_DEVICE_OP(dev_ops, reg_dm_mr);
2601 SET_DEVICE_OP(dev_ops, reg_user_mr);
2602 SET_DEVICE_OP(dev_ops, req_ncomp_notif);
2603 SET_DEVICE_OP(dev_ops, req_notify_cq);
2604 SET_DEVICE_OP(dev_ops, rereg_user_mr);
2605 SET_DEVICE_OP(dev_ops, resize_cq);
2606 SET_DEVICE_OP(dev_ops, set_vf_guid);
2607 SET_DEVICE_OP(dev_ops, set_vf_link_state);
2608 SET_DEVICE_OP(dev_ops, unmap_fmr);
2610 SET_OBJ_SIZE(dev_ops, ib_ah);
2611 SET_OBJ_SIZE(dev_ops, ib_cq);
2612 SET_OBJ_SIZE(dev_ops, ib_pd);
2613 SET_OBJ_SIZE(dev_ops, ib_srq);
2614 SET_OBJ_SIZE(dev_ops, ib_ucontext);
2616 EXPORT_SYMBOL(ib_set_device_ops);
2618 static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2619 [RDMA_NL_LS_OP_RESOLVE] = {
2620 .doit = ib_nl_handle_resolve_resp,
2621 .flags = RDMA_NL_ADMIN_PERM,
2623 [RDMA_NL_LS_OP_SET_TIMEOUT] = {
2624 .doit = ib_nl_handle_set_timeout,
2625 .flags = RDMA_NL_ADMIN_PERM,
2627 [RDMA_NL_LS_OP_IP_RESOLVE] = {
2628 .doit = ib_nl_handle_ip_res_resp,
2629 .flags = RDMA_NL_ADMIN_PERM,
2633 static int __init ib_core_init(void)
2637 ib_wq = alloc_workqueue("infiniband", 0, 0);
2641 ib_comp_wq = alloc_workqueue("ib-comp-wq",
2642 WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2648 ib_comp_unbound_wq =
2649 alloc_workqueue("ib-comp-unb-wq",
2650 WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2651 WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2652 if (!ib_comp_unbound_wq) {
2657 ret = class_register(&ib_class);
2659 pr_warn("Couldn't create InfiniBand device class\n");
2660 goto err_comp_unbound;
2663 ret = rdma_nl_init();
2665 pr_warn("Couldn't init IB netlink interface: err %d\n", ret);
2671 pr_warn("Could't init IB address resolution\n");
2675 ret = ib_mad_init();
2677 pr_warn("Couldn't init IB MAD\n");
2683 pr_warn("Couldn't init SA\n");
2687 ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2689 pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2693 ret = register_pernet_device(&rdma_dev_net_ops);
2695 pr_warn("Couldn't init compat dev. ret %d\n", ret);
2700 rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2701 roce_gid_mgmt_init();
2706 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2716 class_unregister(&ib_class);
2718 destroy_workqueue(ib_comp_unbound_wq);
2720 destroy_workqueue(ib_comp_wq);
2722 destroy_workqueue(ib_wq);
2726 static void __exit ib_core_cleanup(void)
2728 roce_gid_mgmt_cleanup();
2730 rdma_nl_unregister(RDMA_NL_LS);
2731 unregister_pernet_device(&rdma_dev_net_ops);
2732 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2737 class_unregister(&ib_class);
2738 destroy_workqueue(ib_comp_unbound_wq);
2739 destroy_workqueue(ib_comp_wq);
2740 /* Make sure that any pending umem accounting work is done. */
2741 destroy_workqueue(ib_wq);
2742 flush_workqueue(system_unbound_wq);
2743 WARN_ON(!xa_empty(&clients));
2744 WARN_ON(!xa_empty(&devices));
2747 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
2749 /* ib core relies on netdev stack to first register net_ns_type_operations
2750 * ns kobject type before ib_core initialization.
2752 fs_initcall(ib_core_init);
2753 module_exit(ib_core_cleanup);