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 LIST_HEAD(client_list);
98 #define CLIENT_REGISTERED XA_MARK_1
99 static DEFINE_XARRAY_FLAGS(clients, XA_FLAGS_ALLOC);
100 static DECLARE_RWSEM(clients_rwsem);
103 * If client_data is registered then the corresponding client must also still
106 #define CLIENT_DATA_REGISTERED XA_MARK_1
109 * struct rdma_dev_net - rdma net namespace metadata for a net
110 * @net: Pointer to owner net namespace
111 * @id: xarray id to identify the net namespace.
113 struct rdma_dev_net {
118 static unsigned int rdma_dev_net_id;
121 * A list of net namespaces is maintained in an xarray. This is necessary
122 * because we can't get the locking right using the existing net ns list. We
123 * would require a init_net callback after the list is updated.
125 static DEFINE_XARRAY_FLAGS(rdma_nets, XA_FLAGS_ALLOC);
127 * rwsem to protect accessing the rdma_nets xarray entries.
129 static DECLARE_RWSEM(rdma_nets_rwsem);
131 bool ib_devices_shared_netns = true;
132 module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
133 MODULE_PARM_DESC(netns_mode,
134 "Share device among net namespaces; default=1 (shared)");
136 * rdma_dev_access_netns() - Return whether a rdma device can be accessed
137 * from a specified net namespace or not.
138 * @device: Pointer to rdma device which needs to be checked
139 * @net: Pointer to net namesapce for which access to be checked
141 * rdma_dev_access_netns() - Return whether a rdma device can be accessed
142 * from a specified net namespace or not. When
143 * rdma device is in shared mode, it ignores the
144 * net namespace. When rdma device is exclusive
145 * to a net namespace, rdma device net namespace is
146 * checked against the specified one.
148 bool rdma_dev_access_netns(const struct ib_device *dev, const struct net *net)
150 return (ib_devices_shared_netns ||
151 net_eq(read_pnet(&dev->coredev.rdma_net), net));
153 EXPORT_SYMBOL(rdma_dev_access_netns);
156 * xarray has this behavior where it won't iterate over NULL values stored in
157 * allocated arrays. So we need our own iterator to see all values stored in
158 * the array. This does the same thing as xa_for_each except that it also
159 * returns NULL valued entries if the array is allocating. Simplified to only
160 * work on simple xarrays.
162 static void *xan_find_marked(struct xarray *xa, unsigned long *indexp,
165 XA_STATE(xas, xa, *indexp);
170 entry = xas_find_marked(&xas, ULONG_MAX, filter);
171 if (xa_is_zero(entry))
173 } while (xas_retry(&xas, entry));
177 *indexp = xas.xa_index;
178 if (xa_is_zero(entry))
182 return XA_ERROR(-ENOENT);
184 #define xan_for_each_marked(xa, index, entry, filter) \
185 for (index = 0, entry = xan_find_marked(xa, &(index), filter); \
187 (index)++, entry = xan_find_marked(xa, &(index), filter))
189 /* RCU hash table mapping netdevice pointers to struct ib_port_data */
190 static DEFINE_SPINLOCK(ndev_hash_lock);
191 static DECLARE_HASHTABLE(ndev_hash, 5);
193 static void free_netdevs(struct ib_device *ib_dev);
194 static void ib_unregister_work(struct work_struct *work);
195 static void __ib_unregister_device(struct ib_device *device);
196 static int ib_security_change(struct notifier_block *nb, unsigned long event,
198 static void ib_policy_change_task(struct work_struct *work);
199 static DECLARE_WORK(ib_policy_change_work, ib_policy_change_task);
201 static void __ibdev_printk(const char *level, const struct ib_device *ibdev,
202 struct va_format *vaf)
204 if (ibdev && ibdev->dev.parent)
205 dev_printk_emit(level[1] - '0',
208 dev_driver_string(ibdev->dev.parent),
209 dev_name(ibdev->dev.parent),
210 dev_name(&ibdev->dev),
214 level, dev_name(&ibdev->dev), vaf);
216 printk("%s(NULL ib_device): %pV", level, vaf);
219 void ibdev_printk(const char *level, const struct ib_device *ibdev,
220 const char *format, ...)
222 struct va_format vaf;
225 va_start(args, format);
230 __ibdev_printk(level, ibdev, &vaf);
234 EXPORT_SYMBOL(ibdev_printk);
236 #define define_ibdev_printk_level(func, level) \
237 void func(const struct ib_device *ibdev, const char *fmt, ...) \
239 struct va_format vaf; \
242 va_start(args, fmt); \
247 __ibdev_printk(level, ibdev, &vaf); \
253 define_ibdev_printk_level(ibdev_emerg, KERN_EMERG);
254 define_ibdev_printk_level(ibdev_alert, KERN_ALERT);
255 define_ibdev_printk_level(ibdev_crit, KERN_CRIT);
256 define_ibdev_printk_level(ibdev_err, KERN_ERR);
257 define_ibdev_printk_level(ibdev_warn, KERN_WARNING);
258 define_ibdev_printk_level(ibdev_notice, KERN_NOTICE);
259 define_ibdev_printk_level(ibdev_info, KERN_INFO);
261 static struct notifier_block ibdev_lsm_nb = {
262 .notifier_call = ib_security_change,
265 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
268 /* Pointer to the RCU head at the start of the ib_port_data array */
269 struct ib_port_data_rcu {
270 struct rcu_head rcu_head;
271 struct ib_port_data pdata[];
274 static void ib_device_check_mandatory(struct ib_device *device)
276 #define IB_MANDATORY_FUNC(x) { offsetof(struct ib_device_ops, x), #x }
277 static const struct {
280 } mandatory_table[] = {
281 IB_MANDATORY_FUNC(query_device),
282 IB_MANDATORY_FUNC(query_port),
283 IB_MANDATORY_FUNC(query_pkey),
284 IB_MANDATORY_FUNC(alloc_pd),
285 IB_MANDATORY_FUNC(dealloc_pd),
286 IB_MANDATORY_FUNC(create_qp),
287 IB_MANDATORY_FUNC(modify_qp),
288 IB_MANDATORY_FUNC(destroy_qp),
289 IB_MANDATORY_FUNC(post_send),
290 IB_MANDATORY_FUNC(post_recv),
291 IB_MANDATORY_FUNC(create_cq),
292 IB_MANDATORY_FUNC(destroy_cq),
293 IB_MANDATORY_FUNC(poll_cq),
294 IB_MANDATORY_FUNC(req_notify_cq),
295 IB_MANDATORY_FUNC(get_dma_mr),
296 IB_MANDATORY_FUNC(dereg_mr),
297 IB_MANDATORY_FUNC(get_port_immutable)
301 device->kverbs_provider = true;
302 for (i = 0; i < ARRAY_SIZE(mandatory_table); ++i) {
303 if (!*(void **) ((void *) &device->ops +
304 mandatory_table[i].offset)) {
305 device->kverbs_provider = false;
312 * Caller must perform ib_device_put() to return the device reference count
313 * when ib_device_get_by_index() returns valid device pointer.
315 struct ib_device *ib_device_get_by_index(const struct net *net, u32 index)
317 struct ib_device *device;
319 down_read(&devices_rwsem);
320 device = xa_load(&devices, index);
322 if (!rdma_dev_access_netns(device, net)) {
327 if (!ib_device_try_get(device))
331 up_read(&devices_rwsem);
336 * ib_device_put - Release IB device reference
337 * @device: device whose reference to be released
339 * ib_device_put() releases reference to the IB device to allow it to be
340 * unregistered and eventually free.
342 void ib_device_put(struct ib_device *device)
344 if (refcount_dec_and_test(&device->refcount))
345 complete(&device->unreg_completion);
347 EXPORT_SYMBOL(ib_device_put);
349 static struct ib_device *__ib_device_get_by_name(const char *name)
351 struct ib_device *device;
354 xa_for_each (&devices, index, device)
355 if (!strcmp(name, dev_name(&device->dev)))
362 * ib_device_get_by_name - Find an IB device by name
363 * @name: The name to look for
364 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
366 * Find and hold an ib_device by its name. The caller must call
367 * ib_device_put() on the returned pointer.
369 struct ib_device *ib_device_get_by_name(const char *name,
370 enum rdma_driver_id driver_id)
372 struct ib_device *device;
374 down_read(&devices_rwsem);
375 device = __ib_device_get_by_name(name);
376 if (device && driver_id != RDMA_DRIVER_UNKNOWN &&
377 device->ops.driver_id != driver_id)
381 if (!ib_device_try_get(device))
384 up_read(&devices_rwsem);
387 EXPORT_SYMBOL(ib_device_get_by_name);
389 static int rename_compat_devs(struct ib_device *device)
391 struct ib_core_device *cdev;
395 mutex_lock(&device->compat_devs_mutex);
396 xa_for_each (&device->compat_devs, index, cdev) {
397 ret = device_rename(&cdev->dev, dev_name(&device->dev));
400 "Fail to rename compatdev to new name %s\n",
401 dev_name(&device->dev));
405 mutex_unlock(&device->compat_devs_mutex);
409 int ib_device_rename(struct ib_device *ibdev, const char *name)
415 down_write(&devices_rwsem);
416 if (!strcmp(name, dev_name(&ibdev->dev))) {
417 up_write(&devices_rwsem);
421 if (__ib_device_get_by_name(name)) {
422 up_write(&devices_rwsem);
426 ret = device_rename(&ibdev->dev, name);
428 up_write(&devices_rwsem);
432 strlcpy(ibdev->name, name, IB_DEVICE_NAME_MAX);
433 ret = rename_compat_devs(ibdev);
435 downgrade_write(&devices_rwsem);
436 down_read(&ibdev->client_data_rwsem);
437 xan_for_each_marked(&ibdev->client_data, index, client_data,
438 CLIENT_DATA_REGISTERED) {
439 struct ib_client *client = xa_load(&clients, index);
441 if (!client || !client->rename)
444 client->rename(ibdev, client_data);
446 up_read(&ibdev->client_data_rwsem);
447 up_read(&devices_rwsem);
451 int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
455 ibdev->use_cq_dim = use_dim;
460 static int alloc_name(struct ib_device *ibdev, const char *name)
462 struct ib_device *device;
468 lockdep_assert_held_write(&devices_rwsem);
470 xa_for_each (&devices, index, device) {
471 char buf[IB_DEVICE_NAME_MAX];
473 if (sscanf(dev_name(&device->dev), name, &i) != 1)
475 if (i < 0 || i >= INT_MAX)
477 snprintf(buf, sizeof buf, name, i);
478 if (strcmp(buf, dev_name(&device->dev)) != 0)
481 rc = ida_alloc_range(&inuse, i, i, GFP_KERNEL);
486 rc = ida_alloc(&inuse, GFP_KERNEL);
490 rc = dev_set_name(&ibdev->dev, name, rc);
496 static void ib_device_release(struct device *device)
498 struct ib_device *dev = container_of(device, struct ib_device, dev);
501 WARN_ON(refcount_read(&dev->refcount));
502 if (dev->port_data) {
503 ib_cache_release_one(dev);
504 ib_security_release_port_pkey_list(dev);
505 rdma_counter_release(dev);
506 kfree_rcu(container_of(dev->port_data, struct ib_port_data_rcu,
511 xa_destroy(&dev->compat_devs);
512 xa_destroy(&dev->client_data);
513 kfree_rcu(dev, rcu_head);
516 static int ib_device_uevent(struct device *device,
517 struct kobj_uevent_env *env)
519 if (add_uevent_var(env, "NAME=%s", dev_name(device)))
523 * It would be nice to pass the node GUID with the event...
529 static const void *net_namespace(struct device *d)
531 struct ib_core_device *coredev =
532 container_of(d, struct ib_core_device, dev);
534 return read_pnet(&coredev->rdma_net);
537 static struct class ib_class = {
538 .name = "infiniband",
539 .dev_release = ib_device_release,
540 .dev_uevent = ib_device_uevent,
541 .ns_type = &net_ns_type_operations,
542 .namespace = net_namespace,
545 static void rdma_init_coredev(struct ib_core_device *coredev,
546 struct ib_device *dev, struct net *net)
548 /* This BUILD_BUG_ON is intended to catch layout change
549 * of union of ib_core_device and device.
550 * dev must be the first element as ib_core and providers
551 * driver uses it. Adding anything in ib_core_device before
552 * device will break this assumption.
554 BUILD_BUG_ON(offsetof(struct ib_device, coredev.dev) !=
555 offsetof(struct ib_device, dev));
557 coredev->dev.class = &ib_class;
558 coredev->dev.groups = dev->groups;
559 device_initialize(&coredev->dev);
560 coredev->owner = dev;
561 INIT_LIST_HEAD(&coredev->port_list);
562 write_pnet(&coredev->rdma_net, net);
566 * _ib_alloc_device - allocate an IB device struct
567 * @size:size of structure to allocate
569 * Low-level drivers should use ib_alloc_device() to allocate &struct
570 * ib_device. @size is the size of the structure to be allocated,
571 * including any private data used by the low-level driver.
572 * ib_dealloc_device() must be used to free structures allocated with
575 struct ib_device *_ib_alloc_device(size_t size)
577 struct ib_device *device;
579 if (WARN_ON(size < sizeof(struct ib_device)))
582 device = kzalloc(size, GFP_KERNEL);
586 if (rdma_restrack_init(device)) {
591 device->groups[0] = &ib_dev_attr_group;
592 rdma_init_coredev(&device->coredev, device, &init_net);
594 INIT_LIST_HEAD(&device->event_handler_list);
595 spin_lock_init(&device->event_handler_lock);
596 mutex_init(&device->unregistration_lock);
598 * client_data needs to be alloc because we don't want our mark to be
599 * destroyed if the user stores NULL in the client data.
601 xa_init_flags(&device->client_data, XA_FLAGS_ALLOC);
602 init_rwsem(&device->client_data_rwsem);
603 xa_init_flags(&device->compat_devs, XA_FLAGS_ALLOC);
604 mutex_init(&device->compat_devs_mutex);
605 init_completion(&device->unreg_completion);
606 INIT_WORK(&device->unregistration_work, ib_unregister_work);
610 EXPORT_SYMBOL(_ib_alloc_device);
613 * ib_dealloc_device - free an IB device struct
614 * @device:structure to free
616 * Free a structure allocated with ib_alloc_device().
618 void ib_dealloc_device(struct ib_device *device)
620 if (device->ops.dealloc_driver)
621 device->ops.dealloc_driver(device);
624 * ib_unregister_driver() requires all devices to remain in the xarray
625 * while their ops are callable. The last op we call is dealloc_driver
626 * above. This is needed to create a fence on op callbacks prior to
627 * allowing the driver module to unload.
629 down_write(&devices_rwsem);
630 if (xa_load(&devices, device->index) == device)
631 xa_erase(&devices, device->index);
632 up_write(&devices_rwsem);
634 /* Expedite releasing netdev references */
635 free_netdevs(device);
637 WARN_ON(!xa_empty(&device->compat_devs));
638 WARN_ON(!xa_empty(&device->client_data));
639 WARN_ON(refcount_read(&device->refcount));
640 rdma_restrack_clean(device);
641 /* Balances with device_initialize */
642 put_device(&device->dev);
644 EXPORT_SYMBOL(ib_dealloc_device);
647 * add_client_context() and remove_client_context() must be safe against
648 * parallel calls on the same device - registration/unregistration of both the
649 * device and client can be occurring in parallel.
651 * The routines need to be a fence, any caller must not return until the add
652 * or remove is fully completed.
654 static int add_client_context(struct ib_device *device,
655 struct ib_client *client)
659 if (!device->kverbs_provider && !client->no_kverbs_req)
662 down_write(&device->client_data_rwsem);
664 * Another caller to add_client_context got here first and has already
665 * completely initialized context.
667 if (xa_get_mark(&device->client_data, client->client_id,
668 CLIENT_DATA_REGISTERED))
671 ret = xa_err(xa_store(&device->client_data, client->client_id, NULL,
675 downgrade_write(&device->client_data_rwsem);
679 /* Readers shall not see a client until add has been completed */
680 xa_set_mark(&device->client_data, client->client_id,
681 CLIENT_DATA_REGISTERED);
682 up_read(&device->client_data_rwsem);
686 up_write(&device->client_data_rwsem);
690 static void remove_client_context(struct ib_device *device,
691 unsigned int client_id)
693 struct ib_client *client;
696 down_write(&device->client_data_rwsem);
697 if (!xa_get_mark(&device->client_data, client_id,
698 CLIENT_DATA_REGISTERED)) {
699 up_write(&device->client_data_rwsem);
702 client_data = xa_load(&device->client_data, client_id);
703 xa_clear_mark(&device->client_data, client_id, CLIENT_DATA_REGISTERED);
704 client = xa_load(&clients, client_id);
705 downgrade_write(&device->client_data_rwsem);
708 * Notice we cannot be holding any exclusive locks when calling the
709 * remove callback as the remove callback can recurse back into any
710 * public functions in this module and thus try for any locks those
713 * For this reason clients and drivers should not call the
714 * unregistration functions will holdling any locks.
716 * It tempting to drop the client_data_rwsem too, but this is required
717 * to ensure that unregister_client does not return until all clients
718 * are completely unregistered, which is required to avoid module
722 client->remove(device, client_data);
724 xa_erase(&device->client_data, client_id);
725 up_read(&device->client_data_rwsem);
728 static int alloc_port_data(struct ib_device *device)
730 struct ib_port_data_rcu *pdata_rcu;
733 if (device->port_data)
736 /* This can only be called once the physical port range is defined */
737 if (WARN_ON(!device->phys_port_cnt))
741 * device->port_data is indexed directly by the port number to make
742 * access to this data as efficient as possible.
744 * Therefore port_data is declared as a 1 based array with potential
745 * empty slots at the beginning.
747 pdata_rcu = kzalloc(struct_size(pdata_rcu, pdata,
748 rdma_end_port(device) + 1),
753 * The rcu_head is put in front of the port data array and the stored
754 * pointer is adjusted since we never need to see that member until
757 device->port_data = pdata_rcu->pdata;
759 rdma_for_each_port (device, port) {
760 struct ib_port_data *pdata = &device->port_data[port];
762 pdata->ib_dev = device;
763 spin_lock_init(&pdata->pkey_list_lock);
764 INIT_LIST_HEAD(&pdata->pkey_list);
765 spin_lock_init(&pdata->netdev_lock);
766 INIT_HLIST_NODE(&pdata->ndev_hash_link);
771 static int verify_immutable(const struct ib_device *dev, u8 port)
773 return WARN_ON(!rdma_cap_ib_mad(dev, port) &&
774 rdma_max_mad_size(dev, port) != 0);
777 static int setup_port_data(struct ib_device *device)
782 ret = alloc_port_data(device);
786 rdma_for_each_port (device, port) {
787 struct ib_port_data *pdata = &device->port_data[port];
789 ret = device->ops.get_port_immutable(device, port,
794 if (verify_immutable(device, port))
800 void ib_get_device_fw_str(struct ib_device *dev, char *str)
802 if (dev->ops.get_dev_fw_str)
803 dev->ops.get_dev_fw_str(dev, str);
807 EXPORT_SYMBOL(ib_get_device_fw_str);
809 static void ib_policy_change_task(struct work_struct *work)
811 struct ib_device *dev;
814 down_read(&devices_rwsem);
815 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
818 rdma_for_each_port (dev, i) {
820 int ret = ib_get_cached_subnet_prefix(dev,
825 "ib_get_cached_subnet_prefix err: %d, this should never happen here\n",
828 ib_security_cache_change(dev, i, sp);
831 up_read(&devices_rwsem);
834 static int ib_security_change(struct notifier_block *nb, unsigned long event,
837 if (event != LSM_POLICY_CHANGE)
840 schedule_work(&ib_policy_change_work);
841 ib_mad_agent_security_change();
846 static void compatdev_release(struct device *dev)
848 struct ib_core_device *cdev =
849 container_of(dev, struct ib_core_device, dev);
854 static int add_one_compat_dev(struct ib_device *device,
855 struct rdma_dev_net *rnet)
857 struct ib_core_device *cdev;
860 lockdep_assert_held(&rdma_nets_rwsem);
861 if (!ib_devices_shared_netns)
865 * Create and add compat device in all namespaces other than where it
866 * is currently bound to.
868 if (net_eq(read_pnet(&rnet->net),
869 read_pnet(&device->coredev.rdma_net)))
873 * The first of init_net() or ib_register_device() to take the
874 * compat_devs_mutex wins and gets to add the device. Others will wait
875 * for completion here.
877 mutex_lock(&device->compat_devs_mutex);
878 cdev = xa_load(&device->compat_devs, rnet->id);
883 ret = xa_reserve(&device->compat_devs, rnet->id, GFP_KERNEL);
887 cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
893 cdev->dev.parent = device->dev.parent;
894 rdma_init_coredev(cdev, device, read_pnet(&rnet->net));
895 cdev->dev.release = compatdev_release;
896 dev_set_name(&cdev->dev, "%s", dev_name(&device->dev));
898 ret = device_add(&cdev->dev);
901 ret = ib_setup_port_attrs(cdev);
905 ret = xa_err(xa_store(&device->compat_devs, rnet->id,
910 mutex_unlock(&device->compat_devs_mutex);
914 ib_free_port_attrs(cdev);
916 device_del(&cdev->dev);
918 put_device(&cdev->dev);
920 xa_release(&device->compat_devs, rnet->id);
922 mutex_unlock(&device->compat_devs_mutex);
926 static void remove_one_compat_dev(struct ib_device *device, u32 id)
928 struct ib_core_device *cdev;
930 mutex_lock(&device->compat_devs_mutex);
931 cdev = xa_erase(&device->compat_devs, id);
932 mutex_unlock(&device->compat_devs_mutex);
934 ib_free_port_attrs(cdev);
935 device_del(&cdev->dev);
936 put_device(&cdev->dev);
940 static void remove_compat_devs(struct ib_device *device)
942 struct ib_core_device *cdev;
945 xa_for_each (&device->compat_devs, index, cdev)
946 remove_one_compat_dev(device, index);
949 static int add_compat_devs(struct ib_device *device)
951 struct rdma_dev_net *rnet;
955 lockdep_assert_held(&devices_rwsem);
957 down_read(&rdma_nets_rwsem);
958 xa_for_each (&rdma_nets, index, rnet) {
959 ret = add_one_compat_dev(device, rnet);
963 up_read(&rdma_nets_rwsem);
967 static void remove_all_compat_devs(void)
969 struct ib_compat_device *cdev;
970 struct ib_device *dev;
973 down_read(&devices_rwsem);
974 xa_for_each (&devices, index, dev) {
975 unsigned long c_index = 0;
977 /* Hold nets_rwsem so that any other thread modifying this
978 * system param can sync with this thread.
980 down_read(&rdma_nets_rwsem);
981 xa_for_each (&dev->compat_devs, c_index, cdev)
982 remove_one_compat_dev(dev, c_index);
983 up_read(&rdma_nets_rwsem);
985 up_read(&devices_rwsem);
988 static int add_all_compat_devs(void)
990 struct rdma_dev_net *rnet;
991 struct ib_device *dev;
995 down_read(&devices_rwsem);
996 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
997 unsigned long net_index = 0;
999 /* Hold nets_rwsem so that any other thread modifying this
1000 * system param can sync with this thread.
1002 down_read(&rdma_nets_rwsem);
1003 xa_for_each (&rdma_nets, net_index, rnet) {
1004 ret = add_one_compat_dev(dev, rnet);
1008 up_read(&rdma_nets_rwsem);
1010 up_read(&devices_rwsem);
1012 remove_all_compat_devs();
1016 int rdma_compatdev_set(u8 enable)
1018 struct rdma_dev_net *rnet;
1019 unsigned long index;
1022 down_write(&rdma_nets_rwsem);
1023 if (ib_devices_shared_netns == enable) {
1024 up_write(&rdma_nets_rwsem);
1028 /* enable/disable of compat devices is not supported
1029 * when more than default init_net exists.
1031 xa_for_each (&rdma_nets, index, rnet) {
1036 ib_devices_shared_netns = enable;
1037 up_write(&rdma_nets_rwsem);
1042 ret = add_all_compat_devs();
1044 remove_all_compat_devs();
1048 static void rdma_dev_exit_net(struct net *net)
1050 struct rdma_dev_net *rnet = net_generic(net, rdma_dev_net_id);
1051 struct ib_device *dev;
1052 unsigned long index;
1055 down_write(&rdma_nets_rwsem);
1057 * Prevent the ID from being re-used and hide the id from xa_for_each.
1059 ret = xa_err(xa_store(&rdma_nets, rnet->id, NULL, GFP_KERNEL));
1061 up_write(&rdma_nets_rwsem);
1063 down_read(&devices_rwsem);
1064 xa_for_each (&devices, index, dev) {
1065 get_device(&dev->dev);
1067 * Release the devices_rwsem so that pontentially blocking
1068 * device_del, doesn't hold the devices_rwsem for too long.
1070 up_read(&devices_rwsem);
1072 remove_one_compat_dev(dev, rnet->id);
1075 * If the real device is in the NS then move it back to init.
1077 rdma_dev_change_netns(dev, net, &init_net);
1079 put_device(&dev->dev);
1080 down_read(&devices_rwsem);
1082 up_read(&devices_rwsem);
1084 xa_erase(&rdma_nets, rnet->id);
1087 static __net_init int rdma_dev_init_net(struct net *net)
1089 struct rdma_dev_net *rnet = net_generic(net, rdma_dev_net_id);
1090 unsigned long index;
1091 struct ib_device *dev;
1094 /* No need to create any compat devices in default init_net. */
1095 if (net_eq(net, &init_net))
1098 write_pnet(&rnet->net, net);
1100 ret = xa_alloc(&rdma_nets, &rnet->id, rnet, xa_limit_32b, GFP_KERNEL);
1104 down_read(&devices_rwsem);
1105 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
1106 /* Hold nets_rwsem so that netlink command cannot change
1107 * system configuration for device sharing mode.
1109 down_read(&rdma_nets_rwsem);
1110 ret = add_one_compat_dev(dev, rnet);
1111 up_read(&rdma_nets_rwsem);
1115 up_read(&devices_rwsem);
1118 rdma_dev_exit_net(net);
1124 * Assign the unique string device name and the unique device index. This is
1125 * undone by ib_dealloc_device.
1127 static int assign_name(struct ib_device *device, const char *name)
1132 down_write(&devices_rwsem);
1133 /* Assign a unique name to the device */
1134 if (strchr(name, '%'))
1135 ret = alloc_name(device, name);
1137 ret = dev_set_name(&device->dev, name);
1141 if (__ib_device_get_by_name(dev_name(&device->dev))) {
1145 strlcpy(device->name, dev_name(&device->dev), IB_DEVICE_NAME_MAX);
1147 ret = xa_alloc_cyclic(&devices, &device->index, device, xa_limit_31b,
1148 &last_id, GFP_KERNEL);
1153 up_write(&devices_rwsem);
1157 static void setup_dma_device(struct ib_device *device)
1159 struct device *parent = device->dev.parent;
1161 WARN_ON_ONCE(device->dma_device);
1162 if (device->dev.dma_ops) {
1164 * The caller provided custom DMA operations. Copy the
1165 * DMA-related fields that are used by e.g. dma_alloc_coherent()
1168 device->dma_device = &device->dev;
1169 if (!device->dev.dma_mask) {
1171 device->dev.dma_mask = parent->dma_mask;
1175 if (!device->dev.coherent_dma_mask) {
1177 device->dev.coherent_dma_mask =
1178 parent->coherent_dma_mask;
1184 * The caller did not provide custom DMA operations. Use the
1185 * DMA mapping operations of the parent device.
1187 WARN_ON_ONCE(!parent);
1188 device->dma_device = parent;
1190 /* Setup default max segment size for all IB devices */
1191 dma_set_max_seg_size(device->dma_device, SZ_2G);
1196 * setup_device() allocates memory and sets up data that requires calling the
1197 * device ops, this is the only reason these actions are not done during
1198 * ib_alloc_device. It is undone by ib_dealloc_device().
1200 static int setup_device(struct ib_device *device)
1202 struct ib_udata uhw = {.outlen = 0, .inlen = 0};
1205 setup_dma_device(device);
1206 ib_device_check_mandatory(device);
1208 ret = setup_port_data(device);
1210 dev_warn(&device->dev, "Couldn't create per-port data\n");
1214 memset(&device->attrs, 0, sizeof(device->attrs));
1215 ret = device->ops.query_device(device, &device->attrs, &uhw);
1217 dev_warn(&device->dev,
1218 "Couldn't query the device attributes\n");
1225 static void disable_device(struct ib_device *device)
1227 struct ib_client *client;
1229 WARN_ON(!refcount_read(&device->refcount));
1231 down_write(&devices_rwsem);
1232 xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
1233 up_write(&devices_rwsem);
1235 down_read(&clients_rwsem);
1236 list_for_each_entry_reverse(client, &client_list, list)
1237 remove_client_context(device, client->client_id);
1238 up_read(&clients_rwsem);
1240 /* Pairs with refcount_set in enable_device */
1241 ib_device_put(device);
1242 wait_for_completion(&device->unreg_completion);
1245 * compat devices must be removed after device refcount drops to zero.
1246 * Otherwise init_net() may add more compatdevs after removing compat
1247 * devices and before device is disabled.
1249 remove_compat_devs(device);
1253 * An enabled device is visible to all clients and to all the public facing
1254 * APIs that return a device pointer. This always returns with a new get, even
1257 static int enable_device_and_get(struct ib_device *device)
1259 struct ib_client *client;
1260 unsigned long index;
1264 * One ref belongs to the xa and the other belongs to this
1265 * thread. This is needed to guard against parallel unregistration.
1267 refcount_set(&device->refcount, 2);
1268 down_write(&devices_rwsem);
1269 xa_set_mark(&devices, device->index, DEVICE_REGISTERED);
1272 * By using downgrade_write() we ensure that no other thread can clear
1273 * DEVICE_REGISTERED while we are completing the client setup.
1275 downgrade_write(&devices_rwsem);
1277 if (device->ops.enable_driver) {
1278 ret = device->ops.enable_driver(device);
1283 down_read(&clients_rwsem);
1284 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1285 ret = add_client_context(device, client);
1289 up_read(&clients_rwsem);
1291 ret = add_compat_devs(device);
1293 up_read(&devices_rwsem);
1298 * ib_register_device - Register an IB device with IB core
1299 * @device:Device to register
1301 * Low-level drivers use ib_register_device() to register their
1302 * devices with the IB core. All registered clients will receive a
1303 * callback for each device that is added. @device must be allocated
1304 * with ib_alloc_device().
1306 * If the driver uses ops.dealloc_driver and calls any ib_unregister_device()
1307 * asynchronously then the device pointer may become freed as soon as this
1310 int ib_register_device(struct ib_device *device, const char *name)
1314 ret = assign_name(device, name);
1318 ret = setup_device(device);
1322 ret = ib_cache_setup_one(device);
1324 dev_warn(&device->dev,
1325 "Couldn't set up InfiniBand P_Key/GID cache\n");
1329 ib_device_register_rdmacg(device);
1331 rdma_counter_init(device);
1334 * Ensure that ADD uevent is not fired because it
1335 * is too early amd device is not initialized yet.
1337 dev_set_uevent_suppress(&device->dev, true);
1338 ret = device_add(&device->dev);
1342 ret = ib_device_register_sysfs(device);
1344 dev_warn(&device->dev,
1345 "Couldn't register device with driver model\n");
1349 ret = enable_device_and_get(device);
1350 dev_set_uevent_suppress(&device->dev, false);
1351 /* Mark for userspace that device is ready */
1352 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1354 void (*dealloc_fn)(struct ib_device *);
1357 * If we hit this error flow then we don't want to
1358 * automatically dealloc the device since the caller is
1359 * expected to call ib_dealloc_device() after
1360 * ib_register_device() fails. This is tricky due to the
1361 * possibility for a parallel unregistration along with this
1362 * error flow. Since we have a refcount here we know any
1363 * parallel flow is stopped in disable_device and will see the
1364 * NULL pointers, causing the responsibility to
1365 * ib_dealloc_device() to revert back to this thread.
1367 dealloc_fn = device->ops.dealloc_driver;
1368 device->ops.dealloc_driver = NULL;
1369 ib_device_put(device);
1370 __ib_unregister_device(device);
1371 device->ops.dealloc_driver = dealloc_fn;
1374 ib_device_put(device);
1379 device_del(&device->dev);
1381 dev_set_uevent_suppress(&device->dev, false);
1382 ib_device_unregister_rdmacg(device);
1383 ib_cache_cleanup_one(device);
1386 EXPORT_SYMBOL(ib_register_device);
1388 /* Callers must hold a get on the device. */
1389 static void __ib_unregister_device(struct ib_device *ib_dev)
1392 * We have a registration lock so that all the calls to unregister are
1393 * fully fenced, once any unregister returns the device is truely
1394 * unregistered even if multiple callers are unregistering it at the
1395 * same time. This also interacts with the registration flow and
1396 * provides sane semantics if register and unregister are racing.
1398 mutex_lock(&ib_dev->unregistration_lock);
1399 if (!refcount_read(&ib_dev->refcount))
1402 disable_device(ib_dev);
1404 /* Expedite removing unregistered pointers from the hash table */
1405 free_netdevs(ib_dev);
1407 ib_device_unregister_sysfs(ib_dev);
1408 device_del(&ib_dev->dev);
1409 ib_device_unregister_rdmacg(ib_dev);
1410 ib_cache_cleanup_one(ib_dev);
1413 * Drivers using the new flow may not call ib_dealloc_device except
1414 * in error unwind prior to registration success.
1416 if (ib_dev->ops.dealloc_driver) {
1417 WARN_ON(kref_read(&ib_dev->dev.kobj.kref) <= 1);
1418 ib_dealloc_device(ib_dev);
1421 mutex_unlock(&ib_dev->unregistration_lock);
1425 * ib_unregister_device - Unregister an IB device
1426 * @device: The device to unregister
1428 * Unregister an IB device. All clients will receive a remove callback.
1430 * Callers should call this routine only once, and protect against races with
1431 * registration. Typically it should only be called as part of a remove
1432 * callback in an implementation of driver core's struct device_driver and
1435 * If ops.dealloc_driver is used then ib_dev will be freed upon return from
1438 void ib_unregister_device(struct ib_device *ib_dev)
1440 get_device(&ib_dev->dev);
1441 __ib_unregister_device(ib_dev);
1442 put_device(&ib_dev->dev);
1444 EXPORT_SYMBOL(ib_unregister_device);
1447 * ib_unregister_device_and_put - Unregister a device while holding a 'get'
1448 * device: The device to unregister
1450 * This is the same as ib_unregister_device(), except it includes an internal
1451 * ib_device_put() that should match a 'get' obtained by the caller.
1453 * It is safe to call this routine concurrently from multiple threads while
1454 * holding the 'get'. When the function returns the device is fully
1457 * Drivers using this flow MUST use the driver_unregister callback to clean up
1458 * their resources associated with the device and dealloc it.
1460 void ib_unregister_device_and_put(struct ib_device *ib_dev)
1462 WARN_ON(!ib_dev->ops.dealloc_driver);
1463 get_device(&ib_dev->dev);
1464 ib_device_put(ib_dev);
1465 __ib_unregister_device(ib_dev);
1466 put_device(&ib_dev->dev);
1468 EXPORT_SYMBOL(ib_unregister_device_and_put);
1471 * ib_unregister_driver - Unregister all IB devices for a driver
1472 * @driver_id: The driver to unregister
1474 * This implements a fence for device unregistration. It only returns once all
1475 * devices associated with the driver_id have fully completed their
1476 * unregistration and returned from ib_unregister_device*().
1478 * If device's are not yet unregistered it goes ahead and starts unregistering
1481 * This does not block creation of new devices with the given driver_id, that
1482 * is the responsibility of the caller.
1484 void ib_unregister_driver(enum rdma_driver_id driver_id)
1486 struct ib_device *ib_dev;
1487 unsigned long index;
1489 down_read(&devices_rwsem);
1490 xa_for_each (&devices, index, ib_dev) {
1491 if (ib_dev->ops.driver_id != driver_id)
1494 get_device(&ib_dev->dev);
1495 up_read(&devices_rwsem);
1497 WARN_ON(!ib_dev->ops.dealloc_driver);
1498 __ib_unregister_device(ib_dev);
1500 put_device(&ib_dev->dev);
1501 down_read(&devices_rwsem);
1503 up_read(&devices_rwsem);
1505 EXPORT_SYMBOL(ib_unregister_driver);
1507 static void ib_unregister_work(struct work_struct *work)
1509 struct ib_device *ib_dev =
1510 container_of(work, struct ib_device, unregistration_work);
1512 __ib_unregister_device(ib_dev);
1513 put_device(&ib_dev->dev);
1517 * ib_unregister_device_queued - Unregister a device using a work queue
1518 * device: The device to unregister
1520 * This schedules an asynchronous unregistration using a WQ for the device. A
1521 * driver should use this to avoid holding locks while doing unregistration,
1522 * such as holding the RTNL lock.
1524 * Drivers using this API must use ib_unregister_driver before module unload
1525 * to ensure that all scheduled unregistrations have completed.
1527 void ib_unregister_device_queued(struct ib_device *ib_dev)
1529 WARN_ON(!refcount_read(&ib_dev->refcount));
1530 WARN_ON(!ib_dev->ops.dealloc_driver);
1531 get_device(&ib_dev->dev);
1532 if (!queue_work(system_unbound_wq, &ib_dev->unregistration_work))
1533 put_device(&ib_dev->dev);
1535 EXPORT_SYMBOL(ib_unregister_device_queued);
1538 * The caller must pass in a device that has the kref held and the refcount
1539 * released. If the device is in cur_net and still registered then it is moved
1542 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
1548 mutex_lock(&device->unregistration_lock);
1551 * If a device not under ib_device_get() or if the unregistration_lock
1552 * is not held, the namespace can be changed, or it can be unregistered.
1553 * Check again under the lock.
1555 if (refcount_read(&device->refcount) == 0 ||
1556 !net_eq(cur_net, read_pnet(&device->coredev.rdma_net))) {
1561 kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
1562 disable_device(device);
1565 * At this point no one can be using the device, so it is safe to
1566 * change the namespace.
1568 write_pnet(&device->coredev.rdma_net, net);
1570 down_read(&devices_rwsem);
1572 * Currently rdma devices are system wide unique. So the device name
1573 * is guaranteed free in the new namespace. Publish the new namespace
1574 * at the sysfs level.
1576 ret = device_rename(&device->dev, dev_name(&device->dev));
1577 up_read(&devices_rwsem);
1579 dev_warn(&device->dev,
1580 "%s: Couldn't rename device after namespace change\n",
1582 /* Try and put things back and re-enable the device */
1583 write_pnet(&device->coredev.rdma_net, cur_net);
1586 ret2 = enable_device_and_get(device);
1589 * This shouldn't really happen, but if it does, let the user
1590 * retry at later point. So don't disable the device.
1592 dev_warn(&device->dev,
1593 "%s: Couldn't re-enable device after namespace change\n",
1596 kobject_uevent(&device->dev.kobj, KOBJ_ADD);
1598 ib_device_put(device);
1600 mutex_unlock(&device->unregistration_lock);
1606 int ib_device_set_netns_put(struct sk_buff *skb,
1607 struct ib_device *dev, u32 ns_fd)
1612 net = get_net_ns_by_fd(ns_fd);
1618 if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
1624 * Currently supported only for those providers which support
1625 * disassociation and don't do port specific sysfs init. Once a
1626 * port_cleanup infrastructure is implemented, this limitation will be
1629 if (!dev->ops.disassociate_ucontext || dev->ops.init_port ||
1630 ib_devices_shared_netns) {
1635 get_device(&dev->dev);
1637 ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
1638 put_device(&dev->dev);
1650 static struct pernet_operations rdma_dev_net_ops = {
1651 .init = rdma_dev_init_net,
1652 .exit = rdma_dev_exit_net,
1653 .id = &rdma_dev_net_id,
1654 .size = sizeof(struct rdma_dev_net),
1657 static int assign_client_id(struct ib_client *client)
1661 down_write(&clients_rwsem);
1663 * The add/remove callbacks must be called in FIFO/LIFO order. To
1664 * achieve this we assign client_ids so they are sorted in
1665 * registration order, and retain a linked list we can reverse iterate
1666 * to get the LIFO order. The extra linked list can go away if xarray
1667 * learns to reverse iterate.
1669 if (list_empty(&client_list)) {
1670 client->client_id = 0;
1672 struct ib_client *last;
1674 last = list_last_entry(&client_list, struct ib_client, list);
1675 client->client_id = last->client_id + 1;
1677 ret = xa_insert(&clients, client->client_id, client, GFP_KERNEL);
1681 xa_set_mark(&clients, client->client_id, CLIENT_REGISTERED);
1682 list_add_tail(&client->list, &client_list);
1685 up_write(&clients_rwsem);
1690 * ib_register_client - Register an IB client
1691 * @client:Client to register
1693 * Upper level users of the IB drivers can use ib_register_client() to
1694 * register callbacks for IB device addition and removal. When an IB
1695 * device is added, each registered client's add method will be called
1696 * (in the order the clients were registered), and when a device is
1697 * removed, each client's remove method will be called (in the reverse
1698 * order that clients were registered). In addition, when
1699 * ib_register_client() is called, the client will receive an add
1700 * callback for all devices already registered.
1702 int ib_register_client(struct ib_client *client)
1704 struct ib_device *device;
1705 unsigned long index;
1708 ret = assign_client_id(client);
1712 down_read(&devices_rwsem);
1713 xa_for_each_marked (&devices, index, device, DEVICE_REGISTERED) {
1714 ret = add_client_context(device, client);
1716 up_read(&devices_rwsem);
1717 ib_unregister_client(client);
1721 up_read(&devices_rwsem);
1724 EXPORT_SYMBOL(ib_register_client);
1727 * ib_unregister_client - Unregister an IB client
1728 * @client:Client to unregister
1730 * Upper level users use ib_unregister_client() to remove their client
1731 * registration. When ib_unregister_client() is called, the client
1732 * will receive a remove callback for each IB device still registered.
1734 * This is a full fence, once it returns no client callbacks will be called,
1735 * or are running in another thread.
1737 void ib_unregister_client(struct ib_client *client)
1739 struct ib_device *device;
1740 unsigned long index;
1742 down_write(&clients_rwsem);
1743 xa_clear_mark(&clients, client->client_id, CLIENT_REGISTERED);
1744 up_write(&clients_rwsem);
1746 * Every device still known must be serialized to make sure we are
1747 * done with the client callbacks before we return.
1749 down_read(&devices_rwsem);
1750 xa_for_each (&devices, index, device)
1751 remove_client_context(device, client->client_id);
1752 up_read(&devices_rwsem);
1754 down_write(&clients_rwsem);
1755 list_del(&client->list);
1756 xa_erase(&clients, client->client_id);
1757 up_write(&clients_rwsem);
1759 EXPORT_SYMBOL(ib_unregister_client);
1761 static int __ib_get_global_client_nl_info(const char *client_name,
1762 struct ib_client_nl_info *res)
1764 struct ib_client *client;
1765 unsigned long index;
1768 down_read(&clients_rwsem);
1769 xa_for_each_marked (&clients, index, client, CLIENT_REGISTERED) {
1770 if (strcmp(client->name, client_name) != 0)
1772 if (!client->get_global_nl_info) {
1776 ret = client->get_global_nl_info(res);
1777 if (WARN_ON(ret == -ENOENT))
1779 if (!ret && res->cdev)
1780 get_device(res->cdev);
1783 up_read(&clients_rwsem);
1787 static int __ib_get_client_nl_info(struct ib_device *ibdev,
1788 const char *client_name,
1789 struct ib_client_nl_info *res)
1791 unsigned long index;
1795 down_read(&ibdev->client_data_rwsem);
1796 xan_for_each_marked (&ibdev->client_data, index, client_data,
1797 CLIENT_DATA_REGISTERED) {
1798 struct ib_client *client = xa_load(&clients, index);
1800 if (!client || strcmp(client->name, client_name) != 0)
1802 if (!client->get_nl_info) {
1806 ret = client->get_nl_info(ibdev, client_data, res);
1807 if (WARN_ON(ret == -ENOENT))
1811 * The cdev is guaranteed valid as long as we are inside the
1812 * client_data_rwsem as remove_one can't be called. Keep it
1813 * valid for the caller.
1815 if (!ret && res->cdev)
1816 get_device(res->cdev);
1819 up_read(&ibdev->client_data_rwsem);
1825 * ib_get_client_nl_info - Fetch the nl_info from a client
1826 * @device - IB device
1827 * @client_name - Name of the client
1828 * @res - Result of the query
1830 int ib_get_client_nl_info(struct ib_device *ibdev, const char *client_name,
1831 struct ib_client_nl_info *res)
1836 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1838 ret = __ib_get_global_client_nl_info(client_name, res);
1839 #ifdef CONFIG_MODULES
1840 if (ret == -ENOENT) {
1841 request_module("rdma-client-%s", client_name);
1843 ret = __ib_get_client_nl_info(ibdev, client_name, res);
1845 ret = __ib_get_global_client_nl_info(client_name, res);
1854 if (WARN_ON(!res->cdev))
1860 * ib_set_client_data - Set IB client context
1861 * @device:Device to set context for
1862 * @client:Client to set context for
1863 * @data:Context to set
1865 * ib_set_client_data() sets client context data that can be retrieved with
1866 * ib_get_client_data(). This can only be called while the client is
1867 * registered to the device, once the ib_client remove() callback returns this
1870 void ib_set_client_data(struct ib_device *device, struct ib_client *client,
1875 if (WARN_ON(IS_ERR(data)))
1878 rc = xa_store(&device->client_data, client->client_id, data,
1880 WARN_ON(xa_is_err(rc));
1882 EXPORT_SYMBOL(ib_set_client_data);
1885 * ib_register_event_handler - Register an IB event handler
1886 * @event_handler:Handler to register
1888 * ib_register_event_handler() registers an event handler that will be
1889 * called back when asynchronous IB events occur (as defined in
1890 * chapter 11 of the InfiniBand Architecture Specification). This
1891 * callback may occur in interrupt context.
1893 void ib_register_event_handler(struct ib_event_handler *event_handler)
1895 unsigned long flags;
1897 spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
1898 list_add_tail(&event_handler->list,
1899 &event_handler->device->event_handler_list);
1900 spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
1902 EXPORT_SYMBOL(ib_register_event_handler);
1905 * ib_unregister_event_handler - Unregister an event handler
1906 * @event_handler:Handler to unregister
1908 * Unregister an event handler registered with
1909 * ib_register_event_handler().
1911 void ib_unregister_event_handler(struct ib_event_handler *event_handler)
1913 unsigned long flags;
1915 spin_lock_irqsave(&event_handler->device->event_handler_lock, flags);
1916 list_del(&event_handler->list);
1917 spin_unlock_irqrestore(&event_handler->device->event_handler_lock, flags);
1919 EXPORT_SYMBOL(ib_unregister_event_handler);
1922 * ib_dispatch_event - Dispatch an asynchronous event
1923 * @event:Event to dispatch
1925 * Low-level drivers must call ib_dispatch_event() to dispatch the
1926 * event to all registered event handlers when an asynchronous event
1929 void ib_dispatch_event(struct ib_event *event)
1931 unsigned long flags;
1932 struct ib_event_handler *handler;
1934 spin_lock_irqsave(&event->device->event_handler_lock, flags);
1936 list_for_each_entry(handler, &event->device->event_handler_list, list)
1937 handler->handler(handler, event);
1939 spin_unlock_irqrestore(&event->device->event_handler_lock, flags);
1941 EXPORT_SYMBOL(ib_dispatch_event);
1944 * ib_query_port - Query IB port attributes
1945 * @device:Device to query
1946 * @port_num:Port number to query
1947 * @port_attr:Port attributes
1949 * ib_query_port() returns the attributes of a port through the
1950 * @port_attr pointer.
1952 int ib_query_port(struct ib_device *device,
1954 struct ib_port_attr *port_attr)
1959 if (!rdma_is_port_valid(device, port_num))
1962 memset(port_attr, 0, sizeof(*port_attr));
1963 err = device->ops.query_port(device, port_num, port_attr);
1964 if (err || port_attr->subnet_prefix)
1967 if (rdma_port_get_link_layer(device, port_num) != IB_LINK_LAYER_INFINIBAND)
1970 err = device->ops.query_gid(device, port_num, 0, &gid);
1974 port_attr->subnet_prefix = be64_to_cpu(gid.global.subnet_prefix);
1977 EXPORT_SYMBOL(ib_query_port);
1979 static void add_ndev_hash(struct ib_port_data *pdata)
1981 unsigned long flags;
1985 spin_lock_irqsave(&ndev_hash_lock, flags);
1986 if (hash_hashed(&pdata->ndev_hash_link)) {
1987 hash_del_rcu(&pdata->ndev_hash_link);
1988 spin_unlock_irqrestore(&ndev_hash_lock, flags);
1990 * We cannot do hash_add_rcu after a hash_del_rcu until the
1994 spin_lock_irqsave(&ndev_hash_lock, flags);
1997 hash_add_rcu(ndev_hash, &pdata->ndev_hash_link,
1998 (uintptr_t)pdata->netdev);
1999 spin_unlock_irqrestore(&ndev_hash_lock, flags);
2003 * ib_device_set_netdev - Associate the ib_dev with an underlying net_device
2004 * @ib_dev: Device to modify
2005 * @ndev: net_device to affiliate, may be NULL
2006 * @port: IB port the net_device is connected to
2008 * Drivers should use this to link the ib_device to a netdev so the netdev
2009 * shows up in interfaces like ib_enum_roce_netdev. Only one netdev may be
2010 * affiliated with any port.
2012 * The caller must ensure that the given ndev is not unregistered or
2013 * unregistering, and that either the ib_device is unregistered or
2014 * ib_device_set_netdev() is called with NULL when the ndev sends a
2015 * NETDEV_UNREGISTER event.
2017 int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev,
2020 struct net_device *old_ndev;
2021 struct ib_port_data *pdata;
2022 unsigned long flags;
2026 * Drivers wish to call this before ib_register_driver, so we have to
2027 * setup the port data early.
2029 ret = alloc_port_data(ib_dev);
2033 if (!rdma_is_port_valid(ib_dev, port))
2036 pdata = &ib_dev->port_data[port];
2037 spin_lock_irqsave(&pdata->netdev_lock, flags);
2038 old_ndev = rcu_dereference_protected(
2039 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2040 if (old_ndev == ndev) {
2041 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2047 rcu_assign_pointer(pdata->netdev, ndev);
2048 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2050 add_ndev_hash(pdata);
2056 EXPORT_SYMBOL(ib_device_set_netdev);
2058 static void free_netdevs(struct ib_device *ib_dev)
2060 unsigned long flags;
2063 if (!ib_dev->port_data)
2066 rdma_for_each_port (ib_dev, port) {
2067 struct ib_port_data *pdata = &ib_dev->port_data[port];
2068 struct net_device *ndev;
2070 spin_lock_irqsave(&pdata->netdev_lock, flags);
2071 ndev = rcu_dereference_protected(
2072 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2074 spin_lock(&ndev_hash_lock);
2075 hash_del_rcu(&pdata->ndev_hash_link);
2076 spin_unlock(&ndev_hash_lock);
2079 * If this is the last dev_put there is still a
2080 * synchronize_rcu before the netdev is kfreed, so we
2081 * can continue to rely on unlocked pointer
2082 * comparisons after the put
2084 rcu_assign_pointer(pdata->netdev, NULL);
2087 spin_unlock_irqrestore(&pdata->netdev_lock, flags);
2091 struct net_device *ib_device_get_netdev(struct ib_device *ib_dev,
2094 struct ib_port_data *pdata;
2095 struct net_device *res;
2097 if (!rdma_is_port_valid(ib_dev, port))
2100 pdata = &ib_dev->port_data[port];
2103 * New drivers should use ib_device_set_netdev() not the legacy
2106 if (ib_dev->ops.get_netdev)
2107 res = ib_dev->ops.get_netdev(ib_dev, port);
2109 spin_lock(&pdata->netdev_lock);
2110 res = rcu_dereference_protected(
2111 pdata->netdev, lockdep_is_held(&pdata->netdev_lock));
2114 spin_unlock(&pdata->netdev_lock);
2118 * If we are starting to unregister expedite things by preventing
2119 * propagation of an unregistering netdev.
2121 if (res && res->reg_state != NETREG_REGISTERED) {
2130 * ib_device_get_by_netdev - Find an IB device associated with a netdev
2131 * @ndev: netdev to locate
2132 * @driver_id: The driver ID that must match (RDMA_DRIVER_UNKNOWN matches all)
2134 * Find and hold an ib_device that is associated with a netdev via
2135 * ib_device_set_netdev(). The caller must call ib_device_put() on the
2138 struct ib_device *ib_device_get_by_netdev(struct net_device *ndev,
2139 enum rdma_driver_id driver_id)
2141 struct ib_device *res = NULL;
2142 struct ib_port_data *cur;
2145 hash_for_each_possible_rcu (ndev_hash, cur, ndev_hash_link,
2147 if (rcu_access_pointer(cur->netdev) == ndev &&
2148 (driver_id == RDMA_DRIVER_UNKNOWN ||
2149 cur->ib_dev->ops.driver_id == driver_id) &&
2150 ib_device_try_get(cur->ib_dev)) {
2159 EXPORT_SYMBOL(ib_device_get_by_netdev);
2162 * ib_enum_roce_netdev - enumerate all RoCE ports
2163 * @ib_dev : IB device we want to query
2164 * @filter: Should we call the callback?
2165 * @filter_cookie: Cookie passed to filter
2166 * @cb: Callback to call for each found RoCE ports
2167 * @cookie: Cookie passed back to the callback
2169 * Enumerates all of the physical RoCE ports of ib_dev
2170 * which are related to netdevice and calls callback() on each
2171 * device for which filter() function returns non zero.
2173 void ib_enum_roce_netdev(struct ib_device *ib_dev,
2174 roce_netdev_filter filter,
2175 void *filter_cookie,
2176 roce_netdev_callback cb,
2181 rdma_for_each_port (ib_dev, port)
2182 if (rdma_protocol_roce(ib_dev, port)) {
2183 struct net_device *idev =
2184 ib_device_get_netdev(ib_dev, port);
2186 if (filter(ib_dev, port, idev, filter_cookie))
2187 cb(ib_dev, port, idev, cookie);
2195 * ib_enum_all_roce_netdevs - enumerate all RoCE devices
2196 * @filter: Should we call the callback?
2197 * @filter_cookie: Cookie passed to filter
2198 * @cb: Callback to call for each found RoCE ports
2199 * @cookie: Cookie passed back to the callback
2201 * Enumerates all RoCE devices' physical ports which are related
2202 * to netdevices and calls callback() on each device for which
2203 * filter() function returns non zero.
2205 void ib_enum_all_roce_netdevs(roce_netdev_filter filter,
2206 void *filter_cookie,
2207 roce_netdev_callback cb,
2210 struct ib_device *dev;
2211 unsigned long index;
2213 down_read(&devices_rwsem);
2214 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED)
2215 ib_enum_roce_netdev(dev, filter, filter_cookie, cb, cookie);
2216 up_read(&devices_rwsem);
2220 * ib_enum_all_devs - enumerate all ib_devices
2221 * @cb: Callback to call for each found ib_device
2223 * Enumerates all ib_devices and calls callback() on each device.
2225 int ib_enum_all_devs(nldev_callback nldev_cb, struct sk_buff *skb,
2226 struct netlink_callback *cb)
2228 unsigned long index;
2229 struct ib_device *dev;
2230 unsigned int idx = 0;
2233 down_read(&devices_rwsem);
2234 xa_for_each_marked (&devices, index, dev, DEVICE_REGISTERED) {
2235 if (!rdma_dev_access_netns(dev, sock_net(skb->sk)))
2238 ret = nldev_cb(dev, skb, cb, idx);
2243 up_read(&devices_rwsem);
2248 * ib_query_pkey - Get P_Key table entry
2249 * @device:Device to query
2250 * @port_num:Port number to query
2251 * @index:P_Key table index to query
2252 * @pkey:Returned P_Key
2254 * ib_query_pkey() fetches the specified P_Key table entry.
2256 int ib_query_pkey(struct ib_device *device,
2257 u8 port_num, u16 index, u16 *pkey)
2259 if (!rdma_is_port_valid(device, port_num))
2262 return device->ops.query_pkey(device, port_num, index, pkey);
2264 EXPORT_SYMBOL(ib_query_pkey);
2267 * ib_modify_device - Change IB device attributes
2268 * @device:Device to modify
2269 * @device_modify_mask:Mask of attributes to change
2270 * @device_modify:New attribute values
2272 * ib_modify_device() changes a device's attributes as specified by
2273 * the @device_modify_mask and @device_modify structure.
2275 int ib_modify_device(struct ib_device *device,
2276 int device_modify_mask,
2277 struct ib_device_modify *device_modify)
2279 if (!device->ops.modify_device)
2282 return device->ops.modify_device(device, device_modify_mask,
2285 EXPORT_SYMBOL(ib_modify_device);
2288 * ib_modify_port - Modifies the attributes for the specified port.
2289 * @device: The device to modify.
2290 * @port_num: The number of the port to modify.
2291 * @port_modify_mask: Mask used to specify which attributes of the port
2293 * @port_modify: New attribute values for the port.
2295 * ib_modify_port() changes a port's attributes as specified by the
2296 * @port_modify_mask and @port_modify structure.
2298 int ib_modify_port(struct ib_device *device,
2299 u8 port_num, int port_modify_mask,
2300 struct ib_port_modify *port_modify)
2304 if (!rdma_is_port_valid(device, port_num))
2307 if (device->ops.modify_port)
2308 rc = device->ops.modify_port(device, port_num,
2312 rc = rdma_protocol_roce(device, port_num) ? 0 : -ENOSYS;
2315 EXPORT_SYMBOL(ib_modify_port);
2318 * ib_find_gid - Returns the port number and GID table index where
2319 * a specified GID value occurs. Its searches only for IB link layer.
2320 * @device: The device to query.
2321 * @gid: The GID value to search for.
2322 * @port_num: The port number of the device where the GID value was found.
2323 * @index: The index into the GID table where the GID was found. This
2324 * parameter may be NULL.
2326 int ib_find_gid(struct ib_device *device, union ib_gid *gid,
2327 u8 *port_num, u16 *index)
2329 union ib_gid tmp_gid;
2333 rdma_for_each_port (device, port) {
2334 if (!rdma_protocol_ib(device, port))
2337 for (i = 0; i < device->port_data[port].immutable.gid_tbl_len;
2339 ret = rdma_query_gid(device, port, i, &tmp_gid);
2342 if (!memcmp(&tmp_gid, gid, sizeof *gid)) {
2353 EXPORT_SYMBOL(ib_find_gid);
2356 * ib_find_pkey - Returns the PKey table index where a specified
2357 * PKey value occurs.
2358 * @device: The device to query.
2359 * @port_num: The port number of the device to search for the PKey.
2360 * @pkey: The PKey value to search for.
2361 * @index: The index into the PKey table where the PKey was found.
2363 int ib_find_pkey(struct ib_device *device,
2364 u8 port_num, u16 pkey, u16 *index)
2368 int partial_ix = -1;
2370 for (i = 0; i < device->port_data[port_num].immutable.pkey_tbl_len;
2372 ret = ib_query_pkey(device, port_num, i, &tmp_pkey);
2375 if ((pkey & 0x7fff) == (tmp_pkey & 0x7fff)) {
2376 /* if there is full-member pkey take it.*/
2377 if (tmp_pkey & 0x8000) {
2386 /*no full-member, if exists take the limited*/
2387 if (partial_ix >= 0) {
2388 *index = partial_ix;
2393 EXPORT_SYMBOL(ib_find_pkey);
2396 * ib_get_net_dev_by_params() - Return the appropriate net_dev
2397 * for a received CM request
2398 * @dev: An RDMA device on which the request has been received.
2399 * @port: Port number on the RDMA device.
2400 * @pkey: The Pkey the request came on.
2401 * @gid: A GID that the net_dev uses to communicate.
2402 * @addr: Contains the IP address that the request specified as its
2406 struct net_device *ib_get_net_dev_by_params(struct ib_device *dev,
2409 const union ib_gid *gid,
2410 const struct sockaddr *addr)
2412 struct net_device *net_dev = NULL;
2413 unsigned long index;
2416 if (!rdma_protocol_ib(dev, port))
2420 * Holding the read side guarantees that the client will not become
2421 * unregistered while we are calling get_net_dev_by_params()
2423 down_read(&dev->client_data_rwsem);
2424 xan_for_each_marked (&dev->client_data, index, client_data,
2425 CLIENT_DATA_REGISTERED) {
2426 struct ib_client *client = xa_load(&clients, index);
2428 if (!client || !client->get_net_dev_by_params)
2431 net_dev = client->get_net_dev_by_params(dev, port, pkey, gid,
2436 up_read(&dev->client_data_rwsem);
2440 EXPORT_SYMBOL(ib_get_net_dev_by_params);
2442 void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
2444 struct ib_device_ops *dev_ops = &dev->ops;
2445 #define SET_DEVICE_OP(ptr, name) \
2448 if (!((ptr)->name)) \
2449 (ptr)->name = ops->name; \
2452 #define SET_OBJ_SIZE(ptr, name) SET_DEVICE_OP(ptr, size_##name)
2454 if (ops->driver_id != RDMA_DRIVER_UNKNOWN) {
2455 WARN_ON(dev_ops->driver_id != RDMA_DRIVER_UNKNOWN &&
2456 dev_ops->driver_id != ops->driver_id);
2457 dev_ops->driver_id = ops->driver_id;
2460 WARN_ON(dev_ops->owner && dev_ops->owner != ops->owner);
2461 dev_ops->owner = ops->owner;
2463 if (ops->uverbs_abi_ver)
2464 dev_ops->uverbs_abi_ver = ops->uverbs_abi_ver;
2466 dev_ops->uverbs_no_driver_id_binding |=
2467 ops->uverbs_no_driver_id_binding;
2469 SET_DEVICE_OP(dev_ops, add_gid);
2470 SET_DEVICE_OP(dev_ops, advise_mr);
2471 SET_DEVICE_OP(dev_ops, alloc_dm);
2472 SET_DEVICE_OP(dev_ops, alloc_fmr);
2473 SET_DEVICE_OP(dev_ops, alloc_hw_stats);
2474 SET_DEVICE_OP(dev_ops, alloc_mr);
2475 SET_DEVICE_OP(dev_ops, alloc_mr_integrity);
2476 SET_DEVICE_OP(dev_ops, alloc_mw);
2477 SET_DEVICE_OP(dev_ops, alloc_pd);
2478 SET_DEVICE_OP(dev_ops, alloc_rdma_netdev);
2479 SET_DEVICE_OP(dev_ops, alloc_ucontext);
2480 SET_DEVICE_OP(dev_ops, alloc_xrcd);
2481 SET_DEVICE_OP(dev_ops, attach_mcast);
2482 SET_DEVICE_OP(dev_ops, check_mr_status);
2483 SET_DEVICE_OP(dev_ops, counter_alloc_stats);
2484 SET_DEVICE_OP(dev_ops, counter_bind_qp);
2485 SET_DEVICE_OP(dev_ops, counter_dealloc);
2486 SET_DEVICE_OP(dev_ops, counter_unbind_qp);
2487 SET_DEVICE_OP(dev_ops, counter_update_stats);
2488 SET_DEVICE_OP(dev_ops, create_ah);
2489 SET_DEVICE_OP(dev_ops, create_counters);
2490 SET_DEVICE_OP(dev_ops, create_cq);
2491 SET_DEVICE_OP(dev_ops, create_flow);
2492 SET_DEVICE_OP(dev_ops, create_flow_action_esp);
2493 SET_DEVICE_OP(dev_ops, create_qp);
2494 SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
2495 SET_DEVICE_OP(dev_ops, create_srq);
2496 SET_DEVICE_OP(dev_ops, create_wq);
2497 SET_DEVICE_OP(dev_ops, dealloc_dm);
2498 SET_DEVICE_OP(dev_ops, dealloc_driver);
2499 SET_DEVICE_OP(dev_ops, dealloc_fmr);
2500 SET_DEVICE_OP(dev_ops, dealloc_mw);
2501 SET_DEVICE_OP(dev_ops, dealloc_pd);
2502 SET_DEVICE_OP(dev_ops, dealloc_ucontext);
2503 SET_DEVICE_OP(dev_ops, dealloc_xrcd);
2504 SET_DEVICE_OP(dev_ops, del_gid);
2505 SET_DEVICE_OP(dev_ops, dereg_mr);
2506 SET_DEVICE_OP(dev_ops, destroy_ah);
2507 SET_DEVICE_OP(dev_ops, destroy_counters);
2508 SET_DEVICE_OP(dev_ops, destroy_cq);
2509 SET_DEVICE_OP(dev_ops, destroy_flow);
2510 SET_DEVICE_OP(dev_ops, destroy_flow_action);
2511 SET_DEVICE_OP(dev_ops, destroy_qp);
2512 SET_DEVICE_OP(dev_ops, destroy_rwq_ind_table);
2513 SET_DEVICE_OP(dev_ops, destroy_srq);
2514 SET_DEVICE_OP(dev_ops, destroy_wq);
2515 SET_DEVICE_OP(dev_ops, detach_mcast);
2516 SET_DEVICE_OP(dev_ops, disassociate_ucontext);
2517 SET_DEVICE_OP(dev_ops, drain_rq);
2518 SET_DEVICE_OP(dev_ops, drain_sq);
2519 SET_DEVICE_OP(dev_ops, enable_driver);
2520 SET_DEVICE_OP(dev_ops, fill_res_entry);
2521 SET_DEVICE_OP(dev_ops, get_dev_fw_str);
2522 SET_DEVICE_OP(dev_ops, get_dma_mr);
2523 SET_DEVICE_OP(dev_ops, get_hw_stats);
2524 SET_DEVICE_OP(dev_ops, get_link_layer);
2525 SET_DEVICE_OP(dev_ops, get_netdev);
2526 SET_DEVICE_OP(dev_ops, get_port_immutable);
2527 SET_DEVICE_OP(dev_ops, get_vector_affinity);
2528 SET_DEVICE_OP(dev_ops, get_vf_config);
2529 SET_DEVICE_OP(dev_ops, get_vf_stats);
2530 SET_DEVICE_OP(dev_ops, init_port);
2531 SET_DEVICE_OP(dev_ops, iw_accept);
2532 SET_DEVICE_OP(dev_ops, iw_add_ref);
2533 SET_DEVICE_OP(dev_ops, iw_connect);
2534 SET_DEVICE_OP(dev_ops, iw_create_listen);
2535 SET_DEVICE_OP(dev_ops, iw_destroy_listen);
2536 SET_DEVICE_OP(dev_ops, iw_get_qp);
2537 SET_DEVICE_OP(dev_ops, iw_reject);
2538 SET_DEVICE_OP(dev_ops, iw_rem_ref);
2539 SET_DEVICE_OP(dev_ops, map_mr_sg);
2540 SET_DEVICE_OP(dev_ops, map_mr_sg_pi);
2541 SET_DEVICE_OP(dev_ops, map_phys_fmr);
2542 SET_DEVICE_OP(dev_ops, mmap);
2543 SET_DEVICE_OP(dev_ops, modify_ah);
2544 SET_DEVICE_OP(dev_ops, modify_cq);
2545 SET_DEVICE_OP(dev_ops, modify_device);
2546 SET_DEVICE_OP(dev_ops, modify_flow_action_esp);
2547 SET_DEVICE_OP(dev_ops, modify_port);
2548 SET_DEVICE_OP(dev_ops, modify_qp);
2549 SET_DEVICE_OP(dev_ops, modify_srq);
2550 SET_DEVICE_OP(dev_ops, modify_wq);
2551 SET_DEVICE_OP(dev_ops, peek_cq);
2552 SET_DEVICE_OP(dev_ops, poll_cq);
2553 SET_DEVICE_OP(dev_ops, post_recv);
2554 SET_DEVICE_OP(dev_ops, post_send);
2555 SET_DEVICE_OP(dev_ops, post_srq_recv);
2556 SET_DEVICE_OP(dev_ops, process_mad);
2557 SET_DEVICE_OP(dev_ops, query_ah);
2558 SET_DEVICE_OP(dev_ops, query_device);
2559 SET_DEVICE_OP(dev_ops, query_gid);
2560 SET_DEVICE_OP(dev_ops, query_pkey);
2561 SET_DEVICE_OP(dev_ops, query_port);
2562 SET_DEVICE_OP(dev_ops, query_qp);
2563 SET_DEVICE_OP(dev_ops, query_srq);
2564 SET_DEVICE_OP(dev_ops, rdma_netdev_get_params);
2565 SET_DEVICE_OP(dev_ops, read_counters);
2566 SET_DEVICE_OP(dev_ops, reg_dm_mr);
2567 SET_DEVICE_OP(dev_ops, reg_user_mr);
2568 SET_DEVICE_OP(dev_ops, req_ncomp_notif);
2569 SET_DEVICE_OP(dev_ops, req_notify_cq);
2570 SET_DEVICE_OP(dev_ops, rereg_user_mr);
2571 SET_DEVICE_OP(dev_ops, resize_cq);
2572 SET_DEVICE_OP(dev_ops, set_vf_guid);
2573 SET_DEVICE_OP(dev_ops, set_vf_link_state);
2574 SET_DEVICE_OP(dev_ops, unmap_fmr);
2576 SET_OBJ_SIZE(dev_ops, ib_ah);
2577 SET_OBJ_SIZE(dev_ops, ib_cq);
2578 SET_OBJ_SIZE(dev_ops, ib_pd);
2579 SET_OBJ_SIZE(dev_ops, ib_srq);
2580 SET_OBJ_SIZE(dev_ops, ib_ucontext);
2582 EXPORT_SYMBOL(ib_set_device_ops);
2584 static const struct rdma_nl_cbs ibnl_ls_cb_table[RDMA_NL_LS_NUM_OPS] = {
2585 [RDMA_NL_LS_OP_RESOLVE] = {
2586 .doit = ib_nl_handle_resolve_resp,
2587 .flags = RDMA_NL_ADMIN_PERM,
2589 [RDMA_NL_LS_OP_SET_TIMEOUT] = {
2590 .doit = ib_nl_handle_set_timeout,
2591 .flags = RDMA_NL_ADMIN_PERM,
2593 [RDMA_NL_LS_OP_IP_RESOLVE] = {
2594 .doit = ib_nl_handle_ip_res_resp,
2595 .flags = RDMA_NL_ADMIN_PERM,
2599 static int __init ib_core_init(void)
2603 ib_wq = alloc_workqueue("infiniband", 0, 0);
2607 ib_comp_wq = alloc_workqueue("ib-comp-wq",
2608 WQ_HIGHPRI | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
2614 ib_comp_unbound_wq =
2615 alloc_workqueue("ib-comp-unb-wq",
2616 WQ_UNBOUND | WQ_HIGHPRI | WQ_MEM_RECLAIM |
2617 WQ_SYSFS, WQ_UNBOUND_MAX_ACTIVE);
2618 if (!ib_comp_unbound_wq) {
2623 ret = class_register(&ib_class);
2625 pr_warn("Couldn't create InfiniBand device class\n");
2626 goto err_comp_unbound;
2629 ret = rdma_nl_init();
2631 pr_warn("Couldn't init IB netlink interface: err %d\n", ret);
2637 pr_warn("Could't init IB address resolution\n");
2641 ret = ib_mad_init();
2643 pr_warn("Couldn't init IB MAD\n");
2649 pr_warn("Couldn't init SA\n");
2653 ret = register_blocking_lsm_notifier(&ibdev_lsm_nb);
2655 pr_warn("Couldn't register LSM notifier. ret %d\n", ret);
2659 ret = register_pernet_device(&rdma_dev_net_ops);
2661 pr_warn("Couldn't init compat dev. ret %d\n", ret);
2666 rdma_nl_register(RDMA_NL_LS, ibnl_ls_cb_table);
2667 roce_gid_mgmt_init();
2672 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2682 class_unregister(&ib_class);
2684 destroy_workqueue(ib_comp_unbound_wq);
2686 destroy_workqueue(ib_comp_wq);
2688 destroy_workqueue(ib_wq);
2692 static void __exit ib_core_cleanup(void)
2694 roce_gid_mgmt_cleanup();
2696 rdma_nl_unregister(RDMA_NL_LS);
2697 unregister_pernet_device(&rdma_dev_net_ops);
2698 unregister_blocking_lsm_notifier(&ibdev_lsm_nb);
2703 class_unregister(&ib_class);
2704 destroy_workqueue(ib_comp_unbound_wq);
2705 destroy_workqueue(ib_comp_wq);
2706 /* Make sure that any pending umem accounting work is done. */
2707 destroy_workqueue(ib_wq);
2708 flush_workqueue(system_unbound_wq);
2709 WARN_ON(!xa_empty(&clients));
2710 WARN_ON(!xa_empty(&devices));
2713 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_LS, 4);
2715 /* ib core relies on netdev stack to first register net_ns_type_operations
2716 * ns kobject type before ib_core initialization.
2718 fs_initcall(ib_core_init);
2719 module_exit(ib_core_cleanup);